Adaptive Thinking

EVOLUTION AND COGNITION

General Editor: Stephen Stich, Rutgers University

Published in the series

Simple Heuristics That Make Us Smart

Gerd Gigerenzer, Peter Todd, and
the ABC Research Group
Adaptive Thinking: Rationality in the Real World
Gerd Gigerenzer
Adaptive Thinking
Rationality in the Real World

Gerd Gigerenzer

OXFORD
UNIVERSITY PRESS
OXFORD
UNIVERSITY PRESS

Oxford New York

Athens Auckland Bangkok Bogota Buenos Aires Cape Town
Chennai Dar es Salaam Delhi Florence Hong Kong Istanbul Karachi
Kolkata Kuala Lumpur Madrid Melbourne Mexico City Mumbai
Nairobi Paris Sao Paulo Shanghai Singapore Taipei Tokyo Toronto Warsaw
and associated companies in
Berlin Ibadan

Copyright © 2000 by Oxford University Press

Published by Oxford University Press, Inc.
198 Madison Avenue, New York, New York 10016
First issued as an Oxford University Press paperback, 2002
Oxford is a registered trademark of Oxford University Press
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means,
electronic, mechanical, photocopying, recording, or otherwise,
without the prior permission of Oxford University Press.

Library of Congress Cataloging-in-Publication Data

Gigerenzer, Gerd.
Adaptive thinking : rationality in the real world / Gerd Gigerenzer.
p. cm.—(Evolution and cognition)
Includes bibliographical references and index.
ISBN 0-19-513622-5 (cloth); 0-19-515372-3 (pbk.)
1. Reasoning. 2. Thought and thinking. 3. Logic. I. Title. II. Series.
BC177.G53 2000
128'.33—dc21 99-052833

1 3 5 7 9 8 6 4 2
Printed in the United States of America
on acid-free paper
For Raine and Thalia
This page intentionally left blank
 PREFACE

Some years ago, I had lunch with a motley group of colleagues at Stanford,
mostly psychologists and economists, who were interested in decision making
in an uncertain world. We chewed our way through our sandwiches and
through the latest embellishments of the prisoner's dilemma, trading stories of
this or that paradox or stubborn irrationality. Finally, one economist concluded
the discussion with the following dictum: "Look," he said with conviction,
"either reasoning is rational or it's psychological."
This supposed opposition between the rational and the psychological has
haunted me ever since. For the economists and psychologists seated at the
picnic table with me that afternoon, it meant a division of labor. The heavenly
laws of logic and probability rule the realm of sound reasoning; psychology is
assumed to be irrelevant. Only if mistakes are made are psychologists called
in to explain how wrong-wired human minds deviate from these laws. Cher-
nobyl, U.S. foreign policy, and human disasters of many kinds have been as-
sociated with failures in logical thinking. Adopting this opposition, many text-
books present first the laws of logic and probability as the standard by which
to measure human thinking, then data about how people actually think. The
discrepancy between the two makes people appear to be irrational.
Adaptive Thinking offers a different story. I view the mind in relation to its
environment rather than in opposition to the laws of logic or probability. In a
complex and uncertain world, psychology is indispensable for sound reason-
ing; it is rationality's fuel rather than its brake. This book is about rethinking
rationality as adaptive thinking: to understand how minds cope with specific
environments, ecological and social. The chapters in this book elaborate the
idea that human thinking—from scientific creativity to simply understanding
what a positive HIV test means—"happens" partly outside of the mind. For
instance, new laboratory instruments can inspire scientists to create new meta-
phors and theories, and new ways of representing uncertainties can either
cloud or facilitate physicians' understanding of risks. In this sense, insight can
come from outside the mind.
viii PREFACE

The chapters provide both research programs and case studies. For instance,
the program of ecological rationality studies the mind in relation to its envi-
ronment, past and present. Bounded rationality stresses that sound reasoning
can be achieved by simple heuristics that do not follow the prescriptions of
logic and probability. Social rationality is a form of ecological rationality in
which the environment consists of conspecifics and that highlights the impor-
tance of domain-specific behavior and cognition in social environments.
Adaptive Thinking is a collection of what I consider the most important of
my papers on rationality, reasoning and rituals in the 1990s. I have rewritten,
updated, and shortened them to bring out the coherent story they tell as a
whole. The papers were originally addressed to different scientific communi-
ties. This book affords readers the opportunity, for the first time, to see how
the various theoretical endeavors and practical applications fit together.

Berlin G. G.
July 1999
 ACKNOWLEDGMENTS

People

My special thanks go to Daniel G. Goldstein, who coauthored the articles on

which Chapters 2 and 8 are based, and Ulrich Hoffrage, who coauthored the
articles on which Chapters 5, 6, and 7 are based. I have been extraordinarily
lucky to have had Dan and Ulrich first as students and now as colleagues—I
could have not wished for better companions. I would also like to thank Axel
Ebert and Heinz Kleinbolting, who coauthored the articles on which Chapters
5 and 7, respectively, are based.
The chapters in this book have profited from the comments and ideas of
many people: Hal R. Arkes, Peter Ayton, Paul B. Baltes, Peter Barker, Robert
Boyd, Valerie M. Chase, Michael Cole, Leda Cosmides, Denise D. Cummins,
Jean Czerlinski, Arnold Davidson, Frans de Waal, Juan Delius, Raphael Diep-
gen, Berna Eden, Ward Edwards, Ruma Falk, Christian Fiebach, Klaus Fiedler,
Baruch Fischhoff, Jennifer Freyd, Vittorio Girotto, William Goldstein, Daniel
G. Goldstein, Joel Goldstein, Adam S. Goodie, Howard E. Gruber, Kennneth R.
Hammond, John Hartigan, Reid Hastie, Wolfgang Hell, Ralph Hertwig, Miles
Hewstone, Johannes Honekopp, Ulrich Hoffrage, Janellen Huttenlocher, Je-
rome Kagan, Winfried Kain, Gideon Keren, Horst Kilcher, David Klahr, Gernot
Kleiter, Jonathan J. Koehler, Andreas Krebs, Elke M. Kurz, Paolo Legrenzi,
Sarah Lichtenstein, R. Duncan Luce, Albert Madansky, James S. Magnuson,
Brendan Maher, Ken Manktelow, Laura Martignon, David McNeill, Barbara
Mellers, Arthur Miller, Geoffrey F. Miller, John Monahan, David J. Murray, Da-
vid Over, Silvia Papai, John Payne, Jorg Quarder, Terry Regier, Frank Renke-
witz, Werner Schubo, Peter Sedlmeier, Anna Senkevitch, Simon Shaffer, Her-
bert A. Simon, Dan Sperber, Ursula M. Staudinger, Nancy Stein, Stephen M.
Stigler, Andreas Stoltze, Wolfgang Stroebe, Gerhard Strube, Patrick Suppes,
Zeno Swijtink, Peter M. Todd, John Tooby, Tom Trabasso, Amos Tversky, Elke
Weber, Andy Whiten, William Wimsatt, Norton Wise, Werner Wittmann, and
Alf Zimmer.
x ACKNOWLEDGMENTS

I am also very grateful to Donna Alexander, Valerie M. Chase, Dagmar Fecht,

Felix Guettler, Madeline Hoyt, Amy Johnson, Ulrich Kuhnert, Kathleen Much,
and Anita Todd for their support in the preparation of the manuscript.
Last but not least, I wish to thank Steve Stich and Peter Ohlin, who helped
bring this book into being.
There is one person to whom I owe so much that "thank you" is not the
right term: Lorraine Daston, who is my wife, emotional support, and intellec-
tual companion in the world of ideas.

Institutions

The research reported in this book was supported by fellowships and grants
from the Center for Advanced Study in the Behavioral Sciences, Stanford; the
Center for Interdisciplinary Research, Bielefeld, Germany; Deutsche For-
schungsgemeinschaft, Germany; Fonds zur Forderung der Wissenschaften,
Austria; the National Science Foundation; the Spencer Foundation; and the
University of Chicago School Mathematics Project Fund for Research in Math-
ematics Education. I am particularly grateful to the Max Planck Society, which
has provided outstanding research support since 1995.

Publishers

Chapter 1 is an edited version of G. Gigerenzer (1991), From tools to theories:

A heuristic of discovery in cognitive psychology, Psychological Review, 98,
254-267, reprinted with permission of the American Psychological Associa-
tion. Chapter 2 is a shortened and updated version of G. Gigerenzer and D. G.
Goldstein (1996), Mind as computer: The birth of a metaphor, Creativity Re-
search Journal, 9, 131-144, reprinted with permission of Lawrence Erlbaum
Associates. Chapter 3 is based on G. Gigerenzer (in press), Ideas in exile: The
struggles of an upright man, in K. R. Hammond and T. R. Stewart (eds.), The
essential Brunswik: Beginnings, explications, applications (New York: Oxford
University Press), reprinted with permission. Chapter 4 is an edited and short-
ened version of G. Gigerenzer (1998), Ecological intelligence: An adaptation
for frequencies, in D. D. Cummins and C. Allen (eds.), The evolution of mind
(pp. 9-29) (New York: Oxford University Press), reprinted with permission.
Chapter 5 is an edited version of G. Gigerenzer, U. Hoffrage, and A. Ebert
(1998), AIDS counseling for low-risk clients, AIDS CARE, 10, 197-211, re-
printed with permission of Carfax Publishers. Chapter 6 is based on parts of
G. Gigerenzer and U. Hoffrage (1995), How to improve Bayesian reasoning
without instruction: Frequency formats, Psychological Review, 102, 684-704,
and G. Gigerenzer and U. Hoffrage (1999), Overcoming difficulties in Bayesian
reasoning: A reply to Lewis & Keren and Mellers & McGraw, Psychological
Review, 106, 425-430, reprinted with permission of the American Psycho-
logical Association. Chapter 7 is a shortened version of G. Gigerenzer, U. Hof-
 ACKNOWLEDGMENTS Xi

frage, and H. Kleinbolting (1991), Probabilistic mental models: A Brunswikian

theory of confidence, Psychological Review, 98, 506-528, reprinted with per-
mission of the American Psychological Association. Chapter 8 is an updated
version of G. Gigerenzer and D. G. Goldstein (1996), Reasoning the fast and
frugal way: Models of bounded rationality, Psychological Review, 103, 650-
669, reprinted with permission of the American Psychological Association.
Chapters 9 and 10 are based on G. Gigerenzer (1996), Rationality: Why social
context matters, in P. B. Baltes, and U. M. Staudinger (eds.), Interactive minds:
Life-span perspectives on the social foundation of cognition (pp. 319-346)
(Cambridge: Cambridge University Press), reprinted with permission. Chapter
11 is a shortened version of G. Gigerenzer (1997), The modularity of social
intelligence, in A. Whiten & R. W. Byrne (eds.), Machiavellian intelligence II
(pp. 264-288) (Cambridge: Cambridge University Press), reprinted with per-
mission. Chapter 12 is based on G. Gigerenzer (1991), How to make cognitive
illusions disappear. Beyond "heuristics and biases," in W. Stroebe & M. Hew-
stone (eds.), European Review of Social Psychology, 2, 83-115, reprinted with
permission of John Wiley & Sons. Chapter 13 is a slightly shortened version
of G. Gigerenzer (1993), The Superego, the Ego, and the Id in statistical rea-
soning, in G. Keren & C. Lewis (eds.), A handbook for data analysis in the
behavioral sciences: Methodological issues (pp. 311-339). Hillsdale, NJ:
Lawrence Erlbaum Associates, reprinted with permission. Chapter 14 is an
edited version of G. Gigerenzer (1998), Surrogates for theories, Theory and
Psychology, 8, 195-204, reprinted with permission of Sage Publications.
This page intentionally left blank
 CONTENTS

I Where Do New Ideas Come From? 1

1. From tools to theories: A heuristic of discovery 3
2. Mind as computer: The social origin of a metaphor 26
3. Ideas in exile: The struggles of an upright man 44

II Ecological Rationality 57
4. Ecological intelligence 59
5. AIDS counseling for low-risk clients 77
6. How to improve Bayesian reasoning without instruction 92

III Bounded Rationality 125

7. Probabilistic mental models 129
8. Reasoning the fast and frugal way 166

IV Social Rationality 199

9. Rationality: Why social context matters 201
10. Domain-specific reasoning: Social contracts and cheating
detection 211
11. The modularity of social intelligence 226

V Cognitive Illusions and Statistical Rituals 237

12. How to make cognitive illusions disappear 241
13. The Superego, the Ego, and the Id in statistical reasoning 267
14. Surrogates for theories 289

References 297
Name Index 329
Subject Index 337
This page intentionally left blank
 I
WHERE DO NEW IDEAS
COME FROM?

I wrote "From Tools to Theories" in one of the cabinlike offices at the Cen-
ter for Advanced Study in Palo Alto in 1990. That was in the good old days
when the offices had no telephones, e-mail, or other communication facilita-
tors to interrupt one's thoughts. In the meantime, the Center, like you and I,
has surrendered to technology. Chapter 1 is about the impact of new tech-
nologies on creative thinking—an impact of a productive rather than a dis-
ruptive kind. New tools can suggest new scientific ideas and metaphors
about nature, society, and the mind. When this happens, we can trace dis-
coveries back to the changing technological environment in which they
evolved rather than attributing them to some mystical process inside a scien-
tist's head. In this sense, new j'nsights can come from outside the mind.
Two influential tools fueled the cognitive revolution: new statistical tech-
niques and the computer. Both started as tools for data processing and ended
up as theories of mind. The power of tools to inspire new theories derives
from changes both in the technological environment (new tools) and in the
social environment in which a scientist works (the community of tool users).
The social environment is influential in several ways. First, it affects the
pragmatic use of a tool (of which there are many), which then leaves its
mark on the new theories of mind. Second, entrenchment of the tool in the
research community is an important precondition for its final acceptance as
a model of mind. Finally, new social organizations can inspire the creation
of tools in the first place, as evidenced by the invention of the machine com-
puter. Babbage's computer was modeled after a new social organization of
work, namely, the division of labor in large-scale manufacturing. The social
origin of the computer illustrates how a metaphor can cut both ways: First
computers were modeled after minds, and later minds were modeled after
computers.
Computers and statistics have both been used to fulfill the timeless long-
ing to replace judgment by the application of content-blind, mechanical
rules. Such mechanization has become an ideal in many professions, includ-
2 WHERE DO NEW IDEAS COME FROM?

ing copyediting. For instance, the copyeditor at Psychological Review who

worked on "From Tools to Theories" had the rule that a single author speak-
ing in the first person should not use "we" but "I." One sentence in my orig-
inal manuscript read: "Good ideas are hard to come by, and we should be
grateful for those few we have, whatever their lineage." I lost this argument;
every time I reinstated the "we," the copyeditor took it out again, until it
eventually was replaced by "one."
The three chapters in this section elaborate and extend ideas developed
in two earlier books, Cognition as Intuitive Statistics (1987, with D. J. Mur-
ray) and Measurement and Modeling in Psychology (1981, published in
German). "From Tools To Theories" won the American Association for the
Advancement of Science (AAAS) Prize for Behavioral Science Research in
1991. It shows how statistical methods have turned into metaphors of mind.
The tools-to-theories heuristic, one of several heuristics of discovery, can ex-
plain the emergence of a broad range of theories and shed light on their lim-
itations and potentials.
 1

From Tools to Theories

A Heuristic of Discovery

scientific inquiry can be viewed as "an ocean, continuous everywhere and

without a break or division" (Leibniz, 1690/1951, p. 73). Hans Reichenbach
(1938) nonetheless divided this ocean into two great seas, the context of dis-
covery and the context of justification. Philosophers, logicians, and mathe-
maticians claimed justification as a part of their territory and dismissed the
context of discovery as none of their business, or even as "irrelevant to the
logical analysis of scientific knowledge" (Popper, 1935/1959, p. 31). Their sun
shines over one part of the ocean and has been enlightening about matters of
justification, but the other part of the ocean still remains in a mystical darkness
where imagination and intuition reign, or so it is claimed. Popper, Braithwaite,
and others ceded the dark part of the ocean to psychology and, perhaps, so-
ciology, but few psychologists have fished in these waters. Most did not dare
or care.
The discovery versus justification distinction has oversimplified the under-
standing of scientific inquiry. For instance, in the debate over whether the
context of discovery is relevant to understanding science, both sides in the
controversy have construed the question as whether the earlier stage of dis-
covery should be added to the later justification stage (Nickles, 1980). Con-
ceiving the two-context distinction as a temporal distinction (first discovery,
then justification), however, can be misleading because justification procedures
(checking and testing) and discovery processes (having new ideas) take place
during all temporal stages of inquiry. In fact, the original distinction drawn by
Reichenbach in 1938 did not include this temporal simplification; his was not
even a strict dichotomy (see Curd, 1980). I believe that the prevailing inter-
pretation of the two contexts as conceptually distinct events that are in one
and only one temporal sequence has misled many into trying to understand
discovery without taking account of justification.
In this chapter, I argue that discovery can be understood by heuristics (not
a logic) of discovery. I propose a heuristic of discovery that makes use of meth-
ods of justification, thereby attempting to bridge the artificial distinction be-

3
4 WHERE DO NEW IDEAS COME FROM?

tween the two. Furthermore, I attempt to demonstrate that this discovery heu-
ristic may be of interest not only for an a posteriori understanding of theory
development but also for understanding limitations of present-day theories and
research programs and for the further development of alternatives and new
possibilities. The discovery heuristic that I call the tools-to-theories heuristic
(see Gigerenzer & Murray, 1987) postulates a close connection between the
light and the dark parts of Leibniz's ocean: Scientists' tools for justification
provide the metaphors and concepts for their theories.
The power of tools to shape, or even to become, theoretical concepts is an
issue largely ignored in both the history and philosophy of science. Inductivist
accounts of discovery, from Bacon to Reichenbach and the Vienna Circle, focus
on the role of data but do not consider how the data are generated or processed.
Nor do the numerous anecdotes about discoveries—Newton watching an apple
fall in his mother's orchard while pondering the mystery of gravitation; Galton
taking shelter from a rainstorm during a country outing when discovering cor-
relation and regression toward mediocrity; and the stories about Fechner, Ke-
kule, Poincare, and others that link discovery to the three B's: beds, bicycles,
and bathrooms. What unites these anecdotes is the focus on the vivid but
prosaic circumstances; they report the setting in which a discovery occurs,
rather than analyzing the process of discovery.
The question Is there a logic of discovery? and Popper's (1935/1959) con-
jecture that there is none have misled many into assuming that the issue is
whether there exists a logic of discovery or only idiosyncratic personal and
accidental reasons that explain the "flash of insight" of a particular scientist
(Nickles, 1980). I do not think that formal logic and individual personality are
the only alternatives, nor do I believe that either of these is a central issue for
understanding discovery.
The process of discovery can be shown, according to my argument, to pos-
sess more structure than thunderbolt guesses but less definite structure than a
monolithic logic of discovery, of the sort Hanson (1958) searched for, or a
general inductive hypothesis-generation logic (e.g., Reichenbach, 1938). The
present approach lies between these two extremes; it looks for structure be-
yond the insight of a genius but does not claim that the tools-to-theories heu-
ristic is (or should be) the only account of scientific discovery. The tools-to-
theories heuristic applies neither to all theories in science nor to all cognitive
theories; it applies to a specific group of cognitive theories developed during
the last three or four decades, after the so-called cognitive revolution.
Nevertheless, similar heuristics have promoted discovery in physics, phys-
iology, and other areas. For instance, it has been argued that once the me-
chanical clock became the indispensable tool for astronomical research, the
universe itself came to be understood as a kind of mechanical clock, and God
as a divine watchmaker. Lenoir (1986) showed how Faraday's instruments for
recording electric currents shaped the understanding of electrophysiological
processes by promoting concepts such as "muscle current" and "nerve cur-
rent."
 FROM TOOLS TO THEORIES 5

Thus, this discovery heuristic boasts some generality both within cognitive
psychology and within science, but this generality is not unrestricted. Because
there has been little research in how tools of justification influence theory
development, the tools-to-theories heuristic may be more broadly applicable
than I am able to show in this chapter. If my view of heuristics of discovery
as a heterogeneous bundle of search strategies is correct, however, this implies
that generalizability is, in principle, bounded.
What follows has been inspired by Herbert Simon's notion of heuristics of
discovery but goes beyond his attempt to model discovery with programs such
as BACON that attempt to induce scientific laws from data (discussed later).
My focus is on the role of the tools that process and produce data, not the data
themselves, in the discovery and acceptance of theories.

How Methods of Justification Shape Theoretical Concepts

My general thesis is twofold:

1. Discovery. New scientific tools, once entrenched in a scientist's daily
practice, suggest new theoretical metaphors and concepts.
2. Acceptance. Once proposed by an individual scientist (or a group),
the new theoretical metaphors and concepts are more likely to be
accepted by the scientific community if their members are also users
of the new tools.
By tools I mean both analytical and physical methods that are used to eval-
uate given theories. Analytical tools can be either empirical or nonempirical.
Examples of analytical methods of the empirical kind are tools for data pro-
cessing, such as statistics; examples of the nonempirical kind are normative
criteria for the evaluation of hypotheses, such as logical consistency. Examples
of physical tools of justification are measurement instruments, such as clocks.
In this chapter, I focus on analytical rather than physical tools of justification,
and among these, on techniques of statistical inference and hypothesis testing.
My topic is theories of mind and how social scientists discovered them after
the emergence of new tools for data analysis rather than of new data.
In this context, the tools-to-theories heuristic consists in the discovery of
new theories by changing the conception of the mind through the analogy of
the statistical tool. The result can vary in depth from opening new general
perspectives, albeit mainly metaphorical, to sharp discontinuity in specific
cognitive theories caused by the direct transfer of scientists' tools into theories
of mind.
A brief history follows. In American psychology, the study of cognitive pro-
cesses was suppressed in the early 20th century by the allied forces of oper-
ationalism and behaviorism. The operationalism and the inductivism of the
Vienna Circle, as well as the replacement of the Wundtian experiment by ex-
perimentation with treatment groups (Danziger, 1990), paved the way for the
6 WHERE DO NEW IDEAS COME FROM?

institutionalization of inferential statistics in American experimental psychol-

ogy between 1940 and 1955 (Gigerenzer, 1987a; Toulmin & Leary, 1985). In
experimental psychology, inferential statistics became almost synonymous
with scientific method. Inferential statistics, in turn, provided a large part of
the new concepts for mental processes that have fueled the so-called cognitive
revolution since the 1960s. Theories of cognition were cleansed of terms such
as restructuring and insight, and the new mind has come to be portrayed as
drawing random samples from nervous fibers, computing probabilities, cal-
culating analyses of variance (ANOVA), setting decision criteria, and perform-
ing utility analyses.
After the institutionalization of inferential statistics, a broad range of cog-
nitive processes, conscious and unconscious, elementary and complex, were
reinterpreted as involving "intuitive statistics." For instance, Tanner and Swets
(1954) assumed in their theory of signal detectability that the mind "decides"
whether there is a stimulus or only noise, just as a statistician of the Neyman-
Pearson school decides between two hypotheses. In his causal attribution the-
ory, Harold H. Kelley (1967) postulated that the mind attributes a cause to an
effect in the same way as behavioral scientists have come to do, namely by
performing an ANOVA and testing null hypotheses. These two influential the-
ories show the breadth of the new conception of the "mind as an intuitive
statistician." They also exemplify cognitive theories that were suggested not
by new data but by new tools of data analysis.
In what follows, I present evidence for three points. First, the discovery of
theories based on the conception of the mind as an intuitive statistician caused
discontinuity in theory rather than being merely a new, fashionable language:
It radically changed the kind of phenomena reported, the kind of explanations
looked for, and even the kind of data that were generated. This first point
illustrates the profound power of the tools-to-theories heuristic to generate
quite innovative theories. Second, I provide evidence for the "blindness" or
inability of researchers to discover and accept the conception of the mind as
an intuitive statistician before they became familiar with inferential statistics
as part of their daily routine. The discontinuity in cognitive theory is closely
linked to the preceding discontinuity in method, that is, to the institutionali-
zation of inferential statistics in psychology. Third, I show how the tools-to-
theories heuristic can help to define the limits and possibilities of current cog-
nitive theories that investigate the mind as an intuitive statistician.

Discontinuity in Cognitive Theory Development

What has been called the "cognitive revolution" is more than the overthrow
of behaviorism by mentalist concepts. These concepts have been continuously
part of scientific psychology since its emergence in the late 19th century, even
coexisting with American behaviorism during its heyday (Lovie, 1983). The
cognitive revolution did more than revive the mental; it has changed what the
mental means, often dramatically. One source of this change is the tools-to-
 FROM TOOLS TO THEORIES 7

theories heuristic, with its new analogy of the mind as an intuitive statistician.
To show the discontinuity within cognitive theories, I briefly discuss two areas
in which an entire statistical technique, not only a few statistical concepts,
became a model of mental processes: (a) stimulus detection and discrimination
and (b) causal attribution.
What intensity must a 440-Hz tone have to be perceived? How much heavier
than a standard stimulus of 100 g must a comparison stimulus be in order for
a perceiver to notice a difference? How can the elementary cognitive processes
involved in those tasks, known today as stimulus detection and stimulus dis-
crimination, be explained? Since Herbart (1834), such processes have been
explained by using a threshold metaphor: Detection occurs only if the effect
an object has on the nervous system exceeds an absolute threshold, and dis-
crimination between two objects occurs if the excitation from one exceeds that
from another by an amount greater than a differential threshold. E. H. Weber
and G. T. Fechner's laws refer to the concept of fixed thresholds; Titchener
(1896) saw in differential thresholds the long-sought-after elements of mind
(he counted approximately 44,000); and classic textbooks, such as Brown and
Thomson's (1921) and Guilford's (1954), document methods and research.
Around 1955, the psychophysics of absolute and differential thresholds was
revolutionized by the new analogy between the mind and the statistician. W. P
Tanner and others proposed a "theory of signal detectability" (TSD), which
assumes that the Neyman-Pearson technique of hypothesis testing describes
the processes involved in detection and discrimination. Recall that in Neyman-
Pearson statistics, two sampling distributions (hypotheses H0 and HJ and a
decision criterion (which is a likelihood ratio) are defined, and then the data
observed are transformed into a likelihood ratio and compared with the de-
cision criterion. Depending on which side of the criterion the data fall, the
decision "reject H0 and accept H," or "accept H0 and reject H," is made. In
straight analogy, TSD assumes that the mind calculates two sampling distri-
butions for noise and signal plus noise (in the detection situation) and sets a
decision criterion after weighing the cost of the two possible decision errors
(Type I and Type II errors in Neyman-Pearson theory, now called false alarms
and misses}. The sensory input is transduced into a form that allows the brain
to calculate its likelihood ratio, and depending on whether this ratio is smaller
or larger than the criterion, the subject says "no, there is no signal" or "yes,
there is a signal." Tanner (1965) explicitly referred to his new model of the
mind as a "Neyman-Pearson" detector, and, in unpublished work, his flow-
charts included a drawing of a homunculus statistician performing the uncon-
scious statistics in the brain (Gigerenzer & Murray, 1987, pp. 49-53).
The new analogy between mind and statistician replaced the century-old
concept of a fixed threshold by the twin notions of observer's attitudes and
observer's sensitivity. Just as the Neyman-Pearson technique distinguishes be-
tween a subjective part (e.g., selection of a criterion dependent on cost-benefit
considerations) and a mathematical part, detection and discrimination became
understood as involving both subjective processes, such as attitudes and cost-
benefit considerations, and sensory processes. Swets, Tanner, and Birdsall
8 WHERE DO NEW IDEAS COME FROM?

(1964, p. 52) considered this link between attitudes and sensory processes to
be the main thrust of their theory. The analogy between technique and mind
made new research questions thinkable, such as How can the mind's decision
criterion be manipulated? A new kind of data even emerged: Two types of
error were generated in the experiments, false alarms and misses, just as the
statistical theory distinguishes two types of error.
As far as I can tell, the idea of generating these two kinds of data was not
common before the institutionalization of inferential statistics. The discovery
of TSD was not motivated by new data; rather, the new theory motivated a
new kind of data. In fact, in their seminal article, Tanner and Swets (1954,
p. 401} explicitly admitted that their theory "appears to be inconsistent with
the large quantity of existing data on this subject" and proceeded to criticize
the "form of these data."
The Neyman-Pearsonian technique of hypothesis testing was subsequently
transformed into a theory of a broad range of cognitive processes, ranging from
recognition in memory (e.g., Murdock, 1982; Wickelgren & Norman, 1966) to
eyewitness testimony (e.g., Birnbaum, 1983) to discrimination between ran-
dom and nonrandom patterns (e.g., Lopes, 1982).
My second example concerns theories of causal reasoning. In Europe, Albert
Michotte (1946/1963), Jean Piaget (1930), the gestalt psychologists, and others
had investigated how certain temporospatial relationships between two or
more visual objects, such as moving dots, produced phenomenal causality. For
instance, the participants were made to perceive that one dot launches, pushes,
or chases another. After the institutionalization of inferential statistics, Harold
H. Kelley (1967) proposed in his "attribution theory" that the long-sought laws
of causal reasoning are in fact the tools of the behavioral scientist: R. A.
Fisher's ANOVA. Just as the experimenter has come to infer a causal relation-
ship between two variables from calculating an ANOVA and performing an F
test, the person-in-the-street infers the cause of an effect by unconsciously
doing the same calculations. By the time Kelley discovered the new meta-
phor for causal inference, about 70% of all experimental articles already used
ANOVA (Edgington, 1974).
The theory was accepted quickly in social psychology; Kelley and Michaela
(1980) reported there were more than 900 references in one decade. The vision
of the Fisherian mind radically changed the understanding of causal reasoning,
the problems posed to participants, and the explanations looked for. I list a
few discontinuities that reveal the "fingerprints" of the tool, (a) ANOVA needs
repetitions or numbers as data in order to estimate variances and covariances.
Consequently, the information presented to the participants in studies of causal
attribution consists of information about the frequency of events (e.g., Mc-
Arthur, 1972), which played no role in either Michotte's or Piaget's work, (b)
Whereas Michotte's work still reflects the broad Aristotelian conception of four
causes (see Gavin, 1972), and Piaget (1930) distinguished 17 kinds of causality
in children's minds, the Fisherian mind concentrates on the one kind of causes
for which ANOVA is used as a tool (similar to Aristotle's "material cause"), (c)
In Michotte's view, causal perception is direct and spontaneous and needs no
 FROM TOOLS TO THEORIES 9

inference, as a consequence of largely innate laws that determine the organi-

zation of the perceptual field. ANOVA, in contrast, is used in psychology as a
technique for inductive inferences from data to hypotheses, and the focus in
Kelley's attribution theory is consequently on the data-driven, inductive side
of causal perception.
The latter point illustrates that the specific use of a tool, that is, its practical
context rather than its mathematical structure, can also shape theoretical con-
ceptions of mind. To elaborate on this point, assume that Harold Kelley had
lived one-and-a-half centuries earlier than he did. In the early 19th century,
significance tests (similar to those in ANOVA) were already being used by
astronomers (Swijtink, 1987), but they used their tests to reject data, so-called
outliers, and not to reject hypotheses. At least provisionally, the astronomers
assumed that the theory was correct and mistrusted the data, whereas the
ANOVA mind, following the current statistical textbooks, assumes the data to
be correct and mistrusts the theories. So, to a nineteenth-century Kelley, the
mind's causal attribution would have seemed expectation driven rather than
data driven: The statistician homunculus in the mind would have tested the
data and not the hypothesis.
As is well documented, most of causal attribution research after Kelley took
the theoretical stand that attribution is a "lay version of experimental design
and analysis" (Jones & McGillis, 1976, p. 411), and elaboration of the theory
was in part concerned with the kind of intuitive statistics in the brain. For
instance, Ajzen and Fishbein (1975) argued that the homunculus statistician
is Bayesian rather than Fisherian.
These two areas—detection and discrimination, and causal reasoning—may
be sufficient to illustrate some of the fundamental innovations in the explan-
atory framework, in the research questions posed, and in the kind of data
generated. The spectrum of theories that model cognition after statistical in-
ference ranges from auditive and visual perception to recognition in memory
and from speech perception to thinking and reasoning. It reaches from the
elementary, physiological end to the global, conscious end of the continuum
called cognitive. I give one example for each end. (a) Luce (1977) viewed the
central nervous system (CNS) as a statistician who draws a random sample
from all activated fibers, estimates parameters of the pulse rate, aggregates this
estimate into a single number, and uses a decision criterion to arrive at the
final perception. This conception has led to new and interesting questions; for
instance, How does the CNS aggregate numbers? and What is the shape of the
internal distributions? (b) The 18th-century mathematicians Laplace and Con-
dorcet used their "probability of causes" to model how scientists reason (Das-
ton, 1988). Recently, Massaro (1987) proposed the same statistical formula as
an algorithm of pattern recognition, as "a general algorithm, regardless of the
modality and particular nature of the patterns" (p. 16).
The degree to which cognitive theories were shaped by the statistical tool
varies from theory to theory. On the one hand, there is largely metaphorical
use of statistical inference. An example is Gregory's (1974) hypothesis-testing
view of perception, in which he reconceptualized Helmholtz's "unconscious
10 WHERE DO NEW IDEAS COME FROM?

inferences" as Fisherian significance testing: "We may account for the stability
of perceptual forms by suggesting that there is something akin to statistical
significance which must be exceeded by the rival interpretation and the rival
hypothesis before they are allowed to supersede the present perceptual hy-
pothesis" (p. 528). In his theory of how perception works, Gregory also ex-
plained other perceptual phenomena, using Bayesian and Neyman-Pearsonian
statistics as analogies, thus reflecting the actual heterogeneous practice in the
social sciences. Here, a new perspective, but no quantitative model, is gener-
ated. On the other hand, there are cognitive theories that propose quantitative
models of statistical inference that profoundly transform qualitative concepts
and research practice. Examples are the various TSDs of cognition mentioned
earlier and the theory of adaptive memory as statistical optimization by An-
derson and Milson (1989).
To summarize: The tools-to-theories heuristic can account for the discovery
and acceptance of a group of cognitive theories in apparently unrelated sub-
fields of psychology, all of them sharing the view that cognitive processes can
be modeled by statistical hypothesis testing. Among these are several highly
innovative and influential theories that have radically changed our under-
standing of what cognitive means.

Before the Institutionalization of Inferential Statistics

There is an important test case for the present hypotheses (a) that familiarity
with the statistical tool is crucial to the discovery of corresponding theories of
mind and (b) that the institutionalization of the tool within a scientific com-
munity is crucial for the broad acceptance of those theories. That test case is
the era before the institutionalization of inferential statistics. Theories that con-
ceive of the mind as an intuitive statistician should have a very small likeli-
hood of being discovered and even less likelihood of being accepted. The two
strongest tests are cases in which (a) someone proposed a similar conceptual
analogy and (b) someone proposed a similar probabilistic (formal) model. The
chances of theories of the first kind being accepted should be small, and the
chances of a probabilistic model being interpreted as "intuitive statistics"
should be similarly small. I know of only one case each, which I analyze after
defining first what I mean by the phrase "institutionalization of inferential
statistics."
Statistical inference has been known for a long time but not used as theories
of mind. In 1710, John Arbuthnot proved the existence of God using a signif-
icance test; as mentioned earlier, astronomers used significance tests in the
19th century; G. T. Fechner's (1897) statistical text Kollektivmasslehre included
tests of hypotheses; W. S. Cosset (using the pseudonym Student) published the
t test in 1908; and Fisher's significance testing techniques, such as ANOVA, as
well as Neyman-Pearsonian hypothesis-testing methods, have been available
since the 1920s (see Gigerenzer et al., 1989). Bayes's rule has been known
 FROM TOOLS TO THEORIES 11

since 1763. Nonetheless, there was little interest in these techniques in exper-
imental psychology before 1940 (Rucci & Tweney, 1980).
The statisticians' conquest of new territory in psychology started in the
1940s. By 1942, Maurice Kendall could comment on the statisticians' expan-
sion: "They have already overrun every branch of science with a rapidity of
conquest rivalled only by Attila, Mohammed, and the Colorado beetle" (p. 69).
By the early 1950s, half of the psychology departments in leading American
universities offered courses on Fisherian methods and had made inferential sta-
tistics a graduate program requirement. By 1955, more than 80% of the experi-
mental articles in leading journals used inferential statistics to justify conclu-
sions from the data (Sterling, 1959). Editors of major journals made significance
testing a requirement for articles submitted and used the level of significance as
a yardstick for evaluating the quality of an article (e.g., Melton, 1962).
I therefore use 1955 as a rough date for the institutionalization of the tool
in curricula, textbooks, and editorials. What became institutionalized as the
logic of statistical inference was a mixture of ideas from two opposing camps,
those of R. A. Fisher on the one hand and Jerzy Neyman and Egon S. Pearson
(the son of Karl Pearson) on the other (see Chapter 13).

Discovery and Rejection of the Analogy

The analogy between the mind and the statistician was first proposed before
the institutionalization of inferential statistics, in the early 1940s, by Egon
Brunswik at Berkeley (e.g., Brunswik, 1943). As Leary (1987) has shown, Brun-
swik's probabilistic functionalism was based on a very unusual blending of
scientific traditions, including the probabilistic world view of Hans Reichen-
bach and members of the Vienna Circle and Karl Pearson's correlational sta-
tistics.
The important point here is that in the late 1930s, Brunswik changed his
techniques for measuring perceptual constancies, from calculating (nonstatis-
tical) "Brunswik ratios" to calculating Pearson correlations, such as functional
and ecological validities. In the 1940s, he also began to think of the organism
as "an intuitive statistician," but it took him several years to spell out the
analogy in a clear and consistent way.
The analogy is this: The perceptual system infers its environment from un-
certain cues by (unconsciously) calculating correlation and regression statis-
tics, just as the Brunswikian researcher does when (consciously) calculating
the degree of adaptation of a perceptual system to a given environment. Brun-
swik's intuitive statistician was a statistician of the Karl Pearson school, like
the Brunswikian researcher. Brunswik's intuitive statistician was not well
adapted to the psychological science of the time, however, and the analogy
was poorly understood and generally rejected.
Brunswik's analogy came too early to be comprehended and accepted by
his colleagues of the experimental community; it came before the institution-
alization of statistics as the indispensable method of scientific inference, and
12 WHERE DO NEW IDEAS COME FROM?

it came with the "wrong" statistical model, correlational statistics. Correlation

was an indispensable method not in experimental psychology but rather in its
rival discipline, known as the Galton-Pearson program, or, as Lee Cronbach
(1957) put it, the "Holy Roman Empire" of "correlational psychology" (p. 671).
The schism between the two scientific communities had been repeatedly
taken up in presidential addresses before the American Psychological Asso-
ciation (Cronbach, 1957; Dashiell, 1939) and had deeply affected the values
and the mutual esteem of psychologists (Thorndike, 1954). Brunswik could
not persuade his colleagues from the experimental community to consider the
statistical tool of the competing community as a model of how the mind works.
Ernest Hilgard (1955), in his rejection of Brunswik's perspective, did not mince
words: "Correlation is an instrument of the devil" (p. 228).
Brunswik, who coined the metaphor of "man as intuitive statistician," did
not survive to see the success of his analogy. It was accepted only after statis-
tical inference became institutionalized in experimental psychology and with
the new institutionalized tools rather than (Karl) Pearsonian statistics serving
as models of mind. Only in the mid-1960s, however, did interest in Brun-
swikian models of mind emerge (e.g., Hammond, Stewart, Brehmer, & Stein-
mann, 1975).
The tendency to accept the statistical tools of one's own scientific com-
munity (here, the experimental psychologists) rather than those of a competing
community as models of mind is not restricted to Brunswik's case. For ex-
ample, Fritz Heider (1958, pp. 123, 297), whom Harold Kelley credited for
having inspired his ANOVA theory, had repeatedly suggested factor analysis—
another indispensable tool of the correlational discipline—as a model of causal
reasoning. Heider's proposal met with the same neglect by the American ex-
perimental community as did Brunswik's correlational model. Kelley replaced
the statistical tool that Heider suggested by ANOVA, the tool of the experi-
mental community. It seems to be more than a mere accident that both Brun-
swik and Heider came from a similar, German-speaking tradition, where no
comparable division into two communities with competing methodological
imperatives existed.

Probabilistic Models without the Intuitive Statistician

My preceding point is that the statistical tool was accepted as a plausible anal-
ogy of cognitive processes only after its institutionalization in experimental
psychology. My second point is that although some probabilistic models of
cognitive processes were advanced before the institutionalization of inferential
statistics, they were not interpreted using the metaphor of the mind as intuitive
statistician. The distinction I draw is between probabilistic models that use
the metaphor and ones that do not. The latter kind is illustrated by models
that use probability distributions for perceptual judgment, assuming that var-
iability is caused by lack of experimental control, measurement error, or other
factors that can be summarized as experimenter's ignorance. Ideally, if the ex-
perimenter had complete control and knowledge (such as Laplace's superin-
 FROM TOOLS TO THEORIES 13

telligence), all probabilistic terms could be eliminated from the theory. This
does not hold for a probabilistic model that is based on the metaphor. Here,
the probabilistic terms model the ignorance of the mind rather than that of the
experimenter. That is, they model how the homunculus statistician in the brain
comes to terms with a fundamentally uncertain world. Even if the experi-
menter had complete knowledge, the theories would remain probabilistic be-
cause it is the mind that is ignorant and needs statistics.
The key example is represented in L. L. Thurstone, who in 1927 formulated
a model for perceptual judgment that was formally equivalent to the present-
day TSD. But neither Thurstone nor his followers recognized the possibility
of interpreting the formal structure of their model in terms of the intuitive
statistician. Like TSD, Thurstone's model had two overlapping normal distri-
butions, which represented the internal values of two stimuli and which spec-
ified the corresponding likelihood ratios, but it never occurred to Thurstone
to include in his model the conscious activities of a statistician, such as the
weighing of the costs of the two errors and the setting of a decision criterion.
Thus neither Thurstone nor his followers took the—with hindsight—small step
to develop the "law of comparative judgment" into TSD. When Duncan Luce
(1977) reviewed Thurstone's model 50 years later, he found it hard to believe
that nothing in Thurstone's writings showed the least awareness of this small
but crucial step. Thurstone's perceptual model remained a mechanical, albeit
probabilistic, stimulus-response theory without a homunculus statistician in
the brain. The small conceptual step was never taken, and TSD entered psy-
chology by an independent route.
To summarize: There are several kinds of evidence for a close link between
the institutionalization of inferential statistics in the 1950s and the subsequent
broad acceptance of the metaphor of the mind as an intuitive statistician: (a)
the general failure to accept, and even to understand, Brunswik's intuitive
statistician before the institutionalization of the tool and (b) the case of Thur-
stone, who proposed a probabilistic model that was formally equivalent to one
important present-day theory of intuitive statistics but was never interpreted
in this way; the analogy was not yet seen. Brunswik's case illustrates that tools
may act on two levels: First, new tools may suggest new cognitive theories to
a scientist. Second, the degree to which these tools are institutionalized within
the scientific community to which the scientist belongs can prepare (or hinder)
the acceptance of the new theory. This close link between tools for justification
on the one hand and discovery and acceptance on the other reveals the arti-
ficiality of the discovery-justification distinction. Discovery does not come first
and justification afterward. Discovery is inspired by justification.

How Heuristics of Discovery May Help in Understanding

Limitations and Possibilities of Current Research Programs

In this section I argue that the preceding analysis of discovery is of interest

not only for a psychology of scientific discovery and creativity (e.g., Gardner,
14 WHERE DO NEW IDEAS COME FROM?

1988; Gruber, 1981; Tweney, Doherty, & Mynatt, 1981) but also for the eval-
uation and further development of current cognitive theories. The general
point is that institutionalized tools like statistics do not come as pure mathe-
matical (or physical) systems but with a practical context attached. Features
of this context in which a tool has been used may be smuggled Trojan-horse
fashion into the new cognitive theories and research programs. One example
was mentioned earlier: The formal tools of significance testing have been used
in psychology as tools for rejecting hypotheses, with the assumption that the
data are correct, whereas in other fields and at other times the same tools were
used as tools for rejecting data (outliers), with the assumption that the hy-
potheses were correct. The latter use of statistics is practically extinct in ex-
perimental psychology (although the problem of outliers routinely emerges)
and therefore also absent in theories that liken cognitive processes to signifi-
cance testing. In cases like these, analysis of discovery may help to reveal blind
spots associated with the tool and, as a consequence, new possibilities for
cognitive theorizing.
I illustrate this potential in more detail using examples from the "judgment
under uncertainty" program of Daniel Kahneman, Amos Tversky, and others
(see Kahneman & Tversky, 1982). This stimulating research program emerged
from the earlier research on human information processing by Ward Edwards
and his coworkers. In Edwards's work, the dual role of statistics as a tool and
a model of mind is again evident: Edwards, Lindman, and Savage (1963) pro-
posed Bayesian statistics for scientific hypothesis evaluation and considered
the mind as a reasonably good, albeit conservative, Bayesian statistician (e.g.,
Edwards, 1966). The judgment-under-uncertainty program also investigates
reasoning as intuitive statistics but focuses on so-called errors in probabilistic
reasoning. In most of the theories based on the metaphor of the intuitive stat-
istician, statistics or probability theory is used both as normative and as de-
scriptive of a cognitive process (e.g., both as the optimal and the actual mech-
anism for speech perception and human memory; see Massaro, 1987, and
Anderson & Milson, 1989, respectively). This is not the case in the judgment-
under-uncertainty program; here, statistics and probability theory are used
only in the normative function, whereas actual human reasoning has been
described as "biased," "fallacious," or "indefensible" (on the rhetoric, see
Lopes, 1991).
In the following, I first point out three features of the practical use of the
statistical tool (as opposed to the mathematics). Then I show that these features
reemerge in the judgment-under-uncertainty program, resulting in severe lim-
itations on that program. Finally, I suggest how this hidden legacy of the tool
could be eliminated to provide new impulses and possibilities for the research
program.
The first feature is an assumption that can be called "There is only one
statistics." Textbooks on statistics for psychologists (usually written by non-
mathematicians) generally teach statistical inference as if there existed only
one logic of inference. Since the 1950s and 1960s, almost all texts teach a
mishmash of R. A. Fisher's ideas tangled with those of Jerzy Neyman and Egon
 FROM TOOLS TO THEORIES 15

S. Pearson, but without acknowledgment. The fact that Fisherians and

Neyman-Pearsonians could never agree on a logic of statistical inference is not
mentioned in the textbooks, nor are the controversial issues that divide them.
Even alternative statistical logics for scientific inference are rarely discussed.
For instance, Fisher (1955) argued that concepts such as Type II error, power,
the setting of a level of significance before the experiment and its interpretation
as a long-run frequency of errors in repeated experiments are concepts inap-
propriate for scientific inference—at best they could be applied to technology
(his pejorative example was Stalin's). Neyman, for his part, declared that some
of Fisher's significance tests are "worse than useless" (because their power is
less than their size; see Hacking, 1965, p. 99). I know of no textbook written
by psychologists for psychologists that mentions and explains this and other
controversies about the logic of inference. Instead, readers are presented with
an intellectually incoherent mix of Fisherian and Neyman-Pearsonian ideas,
but a mix presented as a seamless, uncontroversial whole: the logic of scientific
inference (for more details, see Chapter 13).
The second assumption that became associated with the tool during its in-
stitutionalization is "There is only one meaning of probability." For instance,
Fisher and Neyman-Pearson had different interpretations of what a level of
significance means. Fisher's was an epistemic interpretation, that is, that the
level of significance indicates the confidence that can be placed in the partic-
ular hypothesis under test, whereas Neyman's was a strictly frequentist and
behavioristic interpretation, which claimed that a level of significance does
not refer to a particular hypothesis but to the relative frequency of wrongly
rejecting the null hypothesis if it is true in the long run. Although the textbooks
teach both Fisherian and Neyman-Pearsonian ideas, these alternative views of
what a probability (such as a level of significance) could mean are generally
neglected—not to speak of the many other meanings that have been proposed
for the formal concept of probability.
Third and last, the daily practice of psychologists assumes that statistical
inference can be applied mechanically without checking the underlying as-
sumptions of the model. The importance of checking whether the assumptions
of a particular statistical model hold in a given application has been repeatedly
emphasized, particularly by statisticians. The general tendency in psycholog-
ical practice (and other social sciences) has been to apply the test anyhow, as
a kind of ritual of justification required by journals but poorly understood by
authors and readers alike.
These features of the practical context, in which the statistical tool has been
used, reemerge at the theoretical level in current cognitive psychology, just as
the tools-to-theories heuristic would lead one to expect.

Example 1: There Is Only One Statistics,

Which Is Normative

Tversky and Kahneman (1974) described their judgment-under-uncertainty

program as a two-step procedure. First, participants are confronted with a rea-
16 WHERE DO NEW IDEAS COME FROM?

soning problem, and their answers are compared with the so-called normative
or correct answer, supplied by statistics and probability theory. Second, the
deviation between the participant's answer and the so-called normative an-
swer, also called a bias of reasoning, is attributed to some heuristic of reason-
ing.
One implicit assumption at the heart of this research program says that
statistical theory provides exactly one answer to the real-world problems pre-
sented to the participants. If this were not true, the deviation between partic-
ipants' judgments and the "normative" answer would be an inappropriate ex-
planandum, because there are as many different deviations as there are
statistical answers. Consider the following problem:
A cab was involved in a hit-and-run accident at night. Two companies,
the Green and the Blue, operate in the city. You are given the following
data:

(i) 85% of the cabs in the city are Green and 15% are Blue, (ii) A witness
identified the cab as a Blue cab. The court tested his ability to identify
cabs under the appropriate visibility conditions. When presented with a
sample of cabs (half of which were Blue and half of which were Green),
the witness made correct identifications in 80% of the cases and erred
in 20% of the cases.

Question: What is the probability that the cab involved in the accident
was Blue rather than Green? (Tversky & Kahneman, 1980, p. 62)
The authors inserted the values specified in this problem into Bayes's rule
and calculated a probability of .41 as the "correct" answer, and, despite criti-
cism, they have never retreated from that claim. They saw in the difference
between this value and the participants' median answer of .80 an instance of
a reasoning error, known as neglect of base rates. But alternative statistical
solutions to the problem exist.
Tversky and Kahneman's reasoning is based on one among many possible
Bayesian views—which the statistician I. J. Good (1971), not all too seriously,
once counted up to 46,656. For instance, using the classical principle of in-
difference to determine the Bayesian prior probabilities can be as defensible
as Tversky and Kahneman's use of base rates of "cabs in the city" for the
relevant priors, but it leads to a probability of .80 instead of .41 (Levi, 1983).
Or, if Neyman-Pearson theory is applied to the cab problem, solutions range
between .28 and .82, depending on the psychological theory about the wit-
ness's criterion shift—the shift from witness testimony at the time of the ac-
cident to witness testimony at the time of the court's test (Birnbaum, 1983;
Gigerenzer & Murray, 1987, pp. 167-174).
There may be more arguable answers to the cab problem, depending on
what statistical or philosophical theory of inference one uses and what as-
sumptions one makes. Indeed, the range of possible statistical solutions is
about the range of participants' actual answers. The point is that none of these
statistical solutions is the only correct answer to the problem, and therefore it
 FROM TOOLS TO THEORIES 17

makes little sense to use the deviation between a participant's judgment and
one of these statistical answers as the psychological explanandum.
Statistics is an indispensable tool for scientific inference, but, as Neyman
and Pearson (1928, p. 176) pointed out, in "many cases there is probably no
single best method of solution." Rather, several such theories are legitimate,
just as "Euclidean and non-Euclidean geometries are equally legitimate" (Ney-
man, 1937, p. 336). My point is this: The idee fixe that statistics speaks with
one voice has reappeared in research on intuitive statistics. The highly inter-
esting judgment-under-uncertainty program could progress beyond the present
point if (a) participants' judgments rather than deviations between judgments
and a so-called normative solution are considered as the data to be explained
and if (b) various statistical models are proposed as competing hypotheses of
problem-solving strategies rather than one model being proposed as the general
norm for rational reasoning. The willingness of many researchers to accept the
claim that statistics speaks with one voice is the legacy of the institutionalized
tool, not of statistics per se.
Note the resulting double standard: Many researchers on intuitive statistics
argue that their participants should draw inferences from data to hypotheses
by using Hayes's rule, although they themselves do not. Rather, the researchers
use the institutionalized mixture of Fisherian and Neyman-Pearsonian statis-
tics to draw their inferences from data to hypotheses.

Example 2: There Is Only One Interpretation

of Probability

Just as there are alternative logics of inference, there are alternative interpre-
tations of probability that have been part of the mathematical theory since its
inception in the mid-17th century (Daston, 1988; Hacking, 1975). Again, both
the institutionalized tool and the recent cognitive research on probabilistic
reasoning exhibit the same blind spot concerning the existence of alternative
interpretations of probability. For instance, Lichtenstein, Fischhoff, and Phil-
lips (1982) have reported and summarized research on a phenomenon called
overconfidence. Briefly, participants were given questions such as "Absinthe
is (a) a precious stone or (b) a liqueur"; they chose what they believed was the
correct answer and then were asked for a confidence rating in their answer,
for example, 90% certain. When people said they were 100% certain about
individual answers, they had in the long run only about 80% correct answers;
when they were 90% certain, they had in the long run only 75% correct an-
swers; and so on. This discrepancy was called overconfidence bias and was
explained by general heuristics in memory search, such as confirmation biases,
or general motivational tendencies, such as a so-called illusion of validity.
My point is that two different interpretations of probability are compared:
degrees of belief in single events (i.e., that this answer is correct) and relative
frequencies of correct answers in the long run. Although 18th-century math-
ematicians, like many of today's cognitive psychologists, would have had no
problem in equating the two, most mathematicians and philosophers since
18 WHERE DO NEW IDEAS COME FROM?

then have. For instance, according to the frequentist point of view, the term
probability, when it refers to a single event, "has no meaning at all" (Mises,
1928/1957, p. 11) because probability theory is about relative frequencies in
the long run. Thus, for a frequentist, probability theory does not apply to
single-event confidences, and therefore no confidence judgment can violate
probability theory. To call a discrepancy between confidence and relative fre-
quency a bias in probabilistic reasoning would mean comparing apples and
oranges. Moreover, even subjectivists would not generally think of a discrep-
ancy between confidence and relative frequency as a bias (see Kadane & Lich-
tenstein, 1982, for a discussion of conditions). For a subjectivist such as Bruno
de Finetti, probability is about single events, but rationality is identified with
the internal consistency of probability judgments. As de Finetti (1931/1989,
p. 174) emphasized: "However an individual evaluates the probability of a
particular event, no experience can prove him right, or wrong; nor in general,
could any conceivable criterion give any objective sense to the distinction one
would like to draw, here, between right and wrong."
Nonetheless, the literature on overconfidence is largely silent on even the
possibility of this conceptual problem (but see Keren, 1987). The question
about research strategy is whether to use the deviation between degrees of
belief and relative frequencies (again considered as a bias) as the explanandum
or to accept the existence of several meanings of probability and to investigate
the kind of conceptual distinctions that untutored people make. Almost all
research has been done within the former research strategy. And, indeed, if
the issue were a general tendency to overestimate one's knowledge, as the term
overconfidence suggests—for instance, as a result of general strategies of mem-
ory search or motivational tendencies—then asking people for degrees of belief
or for frequencies should not matter.
But it does. In a series of experiments (Gigerenzer, Hoffrage, & Kleinbolting,
1991; see also May, 1987), participants were given several hundred questions
of the absinthe type and were asked for confidence judgments after every ques-
tion was answered (as usual). In addition, after each 50 (or 10, 5, and 2) ques-
tions, they were asked how many of those questions they believed they had
answered correctly; that is, frequency judgments were requested. This design
allowed comparison both between their confidence in their individual answers
and true relative frequencies of correct answers, and between judgments of
relative frequencies and true relative frequencies. Comparing frequency judg-
ments with the true frequency of correct answers showed that overestimation
or overconfidence disappeared in 80% to 90% of the participants, depending
on experimental conditions. Frequency judgments were precise or even
showed underestimation. Ironically, after each frequency judgment, partici-
pants went on to give confidence judgments (degrees of belief) that exhibited
what has been called overconfidence.
As in the preceding example, a so-called bias of reasoning disappears if a
controversial norm is dropped and replaced by several descriptive alternatives,
statistical models, and meanings of probability, respectively. Thus probabilities
 FROM TOOLS TO THEORIES 19

for single events and relative frequencies seem to refer to different meanings
of confidence in the minds of the participants. This result is inconsistent with
previous explanations of the alleged bias by deeper cognitive deficiencies (e.g.,
confirmation biases) and has led to the theory of probabilistic mental models,
which describes mechanisms that generate different confidence and frequency
judgments (see Chapter 7). Untutored intuition seems to be capable of making
conceptual distinctions of the sort statisticians and philosophers make (e.g.,
Cohen, 1986; Lopes, 1981; Teigen, 1983). And it suggests that the important
research questions to be investigated are How are different meanings of prob-
ability cued in everyday language? and How does this affect judgment?, rather
than How can the alleged bias of overconfidence be explained by some general
deficits in memory, cognition, or personality?
The same conceptual distinction can help to explain other kinds of judg-
ments under uncertainty. For instance, Tversky and Kahneman (1982a, 1983)
used a personality sketch of a character named Linda that suggested she was
a feminist. Participants were asked which is more probable: (a) that Linda is
a bank teller or (b) that Linda is a bank teller and active in the feminist move-
ment. Most participants chose Alternative b, which Tversky and Kahneman
(1982a) called a "fallacious" belief, to be explained by their hypothesis that
people use a limited number of heuristics—in the present case, representa-
tiveness (the similarity between the description of Linda and the alternatives
a and b). Participants' judgments were called a conjunction fallacy because the
probability of a conjunction of events (bank teller and active in the feminist
movement) cannot be greater than the probability of one of its components.
As in the example just given, this normative interpretation neglects two
facts. First, in everyday language, words like probable legitimately have several
meanings, just as "if. . . then" and "or" constructions do. The particular mean-
ing seems to be automatically cued by content and context. Second, statisti-
cians similarly have alternative views of what probability is about. In the con-
text of some subjectivist theories, choosing Alternative b truly violates the
rules of probability; but for a frequentist, judgments of single events such as
in the Linda problem have nothing to do with probability theory: As the stat-
istician G. A. Barnard (1979, p. 171) objected, they should be treated in the
context of psychoanalysis, not probability.
Again, the normative evaluation explicit in the term conjunction fallacy
is far from being uncontroversial, and progress in understanding reasoning
may be expected by focusing on people's judgments as explanandum rather
than on their deviations from a so-called norm. As in the previous example,
if problems of the Linda type are rephrased as involving frequency judgments
(e.g., "How many out of 100 cases that fit the description of Linda are [a]
bank tellers and [b] bank tellers and active in the feminist movement?"), then
the so-called conjunction fallacy decreases from 77% to 27%, as Fiedler (1988)
showed. "Which alternative is more probable?" is not the same as "Which
alternative is more frequent?" in the Linda context. Tversky and Kahneman
(1983) found similar results, but they maintained their normative claims and
20 WHERE DO NEW IDEAS COME FROM?

treated the disappearance of the phenomenon merely as an exception to the

rule (p. 293).

Example 3: Commitment to Assumptions versus

Neglect of Them

It is a commonplace that the validity of a statistical inference is to be measured

against the validity of the assumptions of the statistical model for a given
situation. In the actual context of justification, however, in psychology and
probably beyond, there is little emphasis on pointing out and checking crucial
assumptions. The same neglect is a drawback in some Bayesian-type proba-
bility revision studies. Kahneman and Tversky's (1973) famous engineer-
lawyer study is a case in point (see also Mueser, Cowan, & Mueser, 1999). In
the study, a group of students was told that a panel of psychologists had made
personality descriptions of 30 engineers and 70 lawyers, that they (the stu-
dents) would be given 5 of these descriptions, chosen at random, and that their
task was to estimate for each description the probability that the person de-
scribed was an engineer. A second group received the same instruction and
the same descriptions but was given inverted base rates, that is, 70 engineers
and 30 lawyers. Kahneman and Tversky found that the mean probabilities
were about the same in the two groups and concluded that base rates were
ignored. They explained this alleged bias in reasoning by postulating that peo-
ple use a general heuristic, called representativeness, which means that people
generally judge the posterior probability simply by the similarity between a
description and their stereotype of an engineer.
Neither Kahneman and Tversky's (1973) study nor any of the follow-up
studies checked whether the participants were committed to or were aware of
a crucial assumption that must hold in order to make the given base rates
relevant: the assumption that the descriptions have been randomly drawn from
the population. If not, the base rates are irrelevant. There have been studies,
such as Kahneman and Tversky's (1973) "Tom W." study, in which participants
were not even told whether the descriptions were randomly sampled. In the
engineer-lawyer study, participants were so informed (in only one word), but
the information was false. Whether a single word is sufficient to direct the
attention of participants toward this crucial information is an important ques-
tion in itself, because researchers cannot assume that in everyday life, people
are familiar with situations in which profession guessing is about randomly
selected people. Thus many of the participants may not have been committed
to the crucial assumption of random selection.
In a controlled replication (Gigerenzer, Hell, & Blank, 1988), a simple
method was used to make participants aware of this crucial assumption: Par-
ticipants themselves drew each description (blindly) out of an urn and gave
their probability judgments. This condition made base-rate neglect largely dis-
appear; once the participants were committed to the crucial assumption of
random sampling, their judgments were closer to Bayesian predictions than to
base rate neglect. This finding indicates that theories of intuitive statistics have
 FROM TOOLS TO THEORIES 21

to deal with how the mind analyzes the structure of a problem (or environ-
ment) and how it infers the presence or absence of crucial statistical assump-
tions—just as the practicing statistician has to first check the structure of a
problem in order to decide whether a particular statistical model can be ap-
plied. Checking structural assumptions precedes statistical calculations (see
also Cohen, 1982; Einhorn & Hogarth, 1981; Ginossar & Trope, 1987).
My intention here is not to criticize this or that specific experiment, but
rather to draw attention to the hidden legacy that tools bequeath to theories.
The general theme is that some features of the practical context in which a
tool has been used (to be distinguished from its mathematics) have reemerged
and been accepted in a research program that investigates intuitive statistics,
impeding progress. Specifically, the key problem is a simplistic conception of
normativeness that confounds one view about probability with the criterion
for rationality.
Although I have dwelt on the dangerous legacy that tools hand on to the-
ories, I do not mean to imply that a theory that originates in a tool is ipso facto
a bad theory. The history of science, not just the history of psychology, is
replete with examples to the contrary. Good ideas are hard to come by, and
one should be grateful for those few that one has, whatever their lineage. But
knowing that lineage can help to refine and criticize the new ideas. In those
cases in which the tools-to-theories heuristic operates, this means taking a
long, hard look at the tools—and the statistical tools of social scientists are
overdue for such a skeptical inspection.

Discussion

New technologies have been a steady source of metaphors of mind: "In my

childhood we were always assured that the brain was a telephone switchboard.
('What else could it be?')," recalled John Searle (1984, p. 44). The tools-to-
theories heuristic is more specific than general technology metaphors. Scien-
tists' tools, not just any tools, are used to understand the mind. Holograms are
not social scientists' tools, but computers are, and part of their differential
acceptance as metaphors of mind by the psychological community may be a
result of psychologists' differential familiarity with these devices in research
practice.
The computer, serial and parallel, would be another case study for the tools-
to-theories heuristic—a case study that is in some aspects different. For in-
stance, John von Neumann (1958) and others explicitly suggested the analogy
between the serial computer and the brain. But the main use of computers in
psychological science was first in the context of justification: for processing
data; making statistical calculations; and as an ideal, endlessly patient exper-
imental subject. Recently, the computer metaphor and the statistics metaphors
of mind have converged, both in artificial intelligence and in the shift toward
massively parallel computers simulating the interaction between neurons.
22 WHERE DO NEW IDEAS COME FROM?

Herbert A. Simon's Heuristics of Discovery and

the Tools-to-Theories Heuristic

Herbert A. Simon (1973) and his coworkers (e.g., Langley, Simon, Bradshaw,
& Zytkow, 1987) explicitly reconsidered the possibility of a logic of discovery.
For example, a series of programs called BACON has "rediscovered" quanti-
tative empirical laws, such as Kepler's third law of planetary motion. How
does BACON discover a law? Basically, BACON starts from data and analyzes
them by applying a group of heuristics until a simple quantitative law can be
fitted to the data. Kepler's law, for instance, can be rediscovered by using heu-
ristics such as "If the values of two numerical terms increase together, then
consider their ratio" (Langley et al., 1987, p. 66). Such heuristics are imple-
.mented as production rules.
What is the relation between heuristics used in programs like BACON and
the tools-to-theories heuristics? First, the research on BACON was concerned
mainly with the ways in which laws could be induced from data. BACON's
heuristics work on extant data, whereas the tools-to-theories heuristic works
on extant tools for data generation and processing and describes an aspect of
discovery (and acceptance) that goes beyond data. As I argued earlier, new
data can be a consequence of the tools-to-theories heuristic, rather than the
starting point to which it is applied. Second, what can be discovered seems to
have little overlap. For Langley et al. (1987), discoveries are of two major
kinds: quantitative laws such as Kepler's law and qualitative laws such as
taxonomies using clustering methods. In fact, the heuristics of discovery pro-
posed in that work are similar to the statistical methods of exploratory data
analysis (Tukey, 1977). It is this kind of intuitive statistics that serves as the
analogy to discovery in Simon's approach. In contrast, the tools-to-theories
heuristic can discover new conceptual analogies, new research programs, and
new data. It cannot—at least not directly—derive quantitative laws by sum-
marizing data, as BACON's heuristics can.
The second issue, What can be discovered?, is related to the first, that is,
to Simon's approach to discovery as induction from data, as "recording in a
parsimonious fashion, sets of empirical data" (Simon, 1973, p. 475). More re-
cently, Simon and Kulkarni (1988) went beyond that data-centered view of
discovery and made a first step toward characterizing the heuristics used by
scientists for planning and guiding experimental research. Although Simon
and Kulkarni did not explore the potential of scientists' tools for suggesting
theoretical concepts (and their particular case study may not invite this), the
tools-to-theories heuristic can complement this recent, broader program to un-
derstand discovery. Both Simon's heuristics and the tools-to-theories heuristic
go beyond the inductive probability approach to discovery (such as Reichen-
bach's). The approaches are complementary in their focus on aspects of dis-
covery, but both emphasize the possibility of understanding discovery by ref-
erence to heuristics of creative reasoning, which go beyond the merely
personal and accidental.
 FROM TOOLS TO THEORIES 23

The Tools-to-Theories Heuristic beyond

Cognitive Psychology

The examples of discovery I give in this chapter are modest instances com-
pared with the classical literature in the history of science treating the contri-
bution of a Copernicus or a Darwin. But in the narrower context of recent
cognitive psychology, the theories I have discussed count as among the most
influential. In this more prosaic context of discovery, the tools-to-theories heu-
ristic can account for a group of significant theoretical innovations. And, as I
have argued, this discovery heuristic can both open and foreclose new avenues
of research, depending on the interpretations attached to the statistical tool.
My focus is on analytical tools of justification, and I have not dealt with phys-
ical tools of experimentation and data processing. Physical tools, once familiar
and considered indispensable, also may become the stuff of theories. This
holds not only for the hardware (like the software) of the computer, but also
for theory innovation beyond recent cognitive psychology. Smith (1986) argued
that Edward C. Tolman's use of the maze as an experimental apparatus trans-
formed Tolman's conception of purpose and cognition into spatial character-
istics, such as cognitive maps. Similarly, he argued that Clark L. Hull's fasci-
nation with conditioning machines has shaped Hull's thinking of behavior as
if it were machine design. With the exception of Danziger's (1985, 1987) work
on changing methodological practices in psychology and their impact on the
kind of knowledge produced, however, there seems to exist no systematic re-
search program on the power of familiar tools to shape new theories in psy-
chology.
But the history of science beyond psychology provides some striking in-
stances of scientists' tools, both analytical and physical, that ended up as the-
ories of nature. Hackmann (1979), Lenoir (1986), and Wise (1988) have ex-
plored how scientific instruments shaped the theoretical concepts of, among
others, Emil DuBois-Reymond and William Thomson (Lord Kelvin).
The case of Adolphe Quetelet illustrates nicely how the tools-to-theories
heuristic can combine with an interdisciplinary exchange of theories. The sta-
tistical error law (normal distribution) was used by astronomers to handle ob-
servational errors around the true position of a star. Quetelet (1842/1969), who
began as an astronomer, transformed the astronomer's tool for taming error into
a theory about society: The true position of a star turned into 1'homme moyen,
or the ideal average person within a society, and observational errors turned
into the distribution of actual persons (with respect to any variable) around
1'homme moyen—actual persons now being viewed as nature's errors. Quete-
let's social error theory was in turn seminal in the development of statistical
mechanics; Ludwig Boltzmann and James Clerk Maxwell in the 1860s and
1870s reasoned that gas molecules might behave as Quetelet's humans do;
erratic and unpredictable as individuals, but regular and predictable when
considered as a collective (Porter, 1986). By this strange route of discovery—
from astronomer's tool to a theory of society, and from a theory of society to
24 WHERE DO NEW IDEAS COME FROM?

a theory of a collective of gas molecules—the deterministic Newtonian view

of the world was finally overthrown and replaced by a statistical view of nature
(see Gigerenzer et al., 1989). Thus there seems to exist a broader, interdisci-
plinary framework for the tools-to-theories heuristic proposed here, which has
yet to be explored.

Discovery Reconsidered
Let me conclude with some reflections on how the present view stands in
relation to major themes in scientific discovery.

Data-tO-Theories Reconsidered Should psychologists continue to tell their stu-

dents that new theories originate from new data, if only because "little is
known about how theories come to be created," as J. R. Anderson remarked
in the introduction to his Cognitive Psychology (1980, p. 17)? Holton (1988)
noted the tendency among physicists to reconstruct discovery with hindsight
as originating from new data, even if this is not the case. His most prominent
example is Einstein's special theory of relativity, which was and still is cele-
brated as an empirical generalization from Michelson's experimental data by
such eminent figures as R. A. Millikan and H. Reichenbach, as well as by the
textbook writers. As Holton demonstrated with firsthand documents, the role
of Michelson's data in the discovery of Einstein's theory was slight, a conclu-
sion shared by Einstein himself.
Similarly, with respect to more modest discoveries, I argue that a group of
recent cognitive theories did not originate from new data, but in fact often
created new kinds of data. Tanner and Swets (1954) are even explicit that their
theory was inconsistent with the extant data. Numerical probability judgments
have become the stock-in-trade data of research on inductive thinking since
Edwards's (1966) work, whereas this kind of dependent variable was still un-
known in Humphrey's (1951) review of research on thinking.
The strongest claim for an inductive view of discovery came from the Vi-
enna Circle's emphasis on sensory data (reduced to the concept of "pointer
readings"). Carnap (1928/1969), Reichenbach (1938), and others focused on
what they called the rational reconstruction of actual discovery rather than on
actual discovery itself, in order to screen out the merely irrational and psy-
chological. For instance, Reichenbach reconstructed Einstein's special theory
of relativity as being "suggested by closest adherence to experimental facts,"
a claim that Einstein rejected, as mentioned earlier (see Holton, 1988, p. 296).
It seems fair to say that all attempts to logically reconstruct discovery in sci-
ence have failed in practice (Blackwell, 1983, p. 111). The strongest theoretical
disclaimer concerning the possibility of a logic of discovery came from Popper,
Hempel, and other proponents of the hypothetico-deductive account, resulting
in the judgment that discovery, not being logical, occurs irrationally. Theories
are simply "guesses guided by the unscientific" (Popper, 1959, p. 278). In con-
trast, I have dealt with guesses that are guided by the scientific, by tools of
justification. Induction from data and irrational guesses are not exhaustive of
 FROM TOOLS TO THEORIES 25

scientific discovery, and the tools-to-theories heuristic explores the field be-
yond.

Scientists' Practice Reconsidered The tools-to-theories heuristic is about sci-

entists' practice, that is, the analytical and physical tools used in the conduct
of experiments. This practice has a long tradition of neglect. The very philos-
ophers who called themselves logical empiricists had, ironically, little interest
in the empirical practice of scientists. Against their reduction of observation
to pointer reading, Kuhn (1970) emphasized the theory ladenness of observa-
tion. Referring to perceptual experiments and gestalt switches, he said: "Sci-
entists see new and different things when looking with familiar instruments
in places they have looked before" (p. 111). Both the logical empiricists and
Kuhn were highly influential on psychology (see Toulmin & Leary, 1985), but
neither's view has emphasized the role of tools and experimental conduct.
Their role in the development of science has been grossly underestimated until
recently (Danziger, 1985; Lenoir, 1988).
Through the lens of theory, it has been said, growth of knowledge can be
understood. But there is a recent move away from a theory-dominated account
of science that pays attention to what really happens in the laboratories. Hack-
ing (1983) argued that experimentation has a life of its own and that not all
observation is theory laden. Galison (1987) analyzed modern experimental
practice, such as in high-energy physics, focusing on the role of the fine-
grained web of instruments, beliefs, and practice that determine when a fact
is considered to be established and when experiments end. Both Hacking and
Galison emphasized the role of the familiarity experimenters have with their
tools, and the importance and relative autonomy of experimental practice in
the quest for knowledge. This is the broader context in which the present tools-
to-theories heuristic stands: the conjecture that theory is inseparable from in-
strumental practices.
In conclusion, my argument is that discovery in recent cognitive psychology
can be understood beyond mere inductive generalizations or lucky guesses.
More than that, I argue that for a considerable group of cognitive theories,
neither induction from data nor lucky guesses played an important role.
Rather, these innovations in theory can be accounted for by the tools-to-
theories heuristic. So can conceptual problems and possibilities in current the-
ories. Scientists' tools are not neutral. In the present case, the mind has been
re-created in their image.
 2
Mind as Computer
The Social Origin of a Metaphor

Have philosophers of science spent too little time inside the laboratories to
be drawn in by the glamour of technology? Tools, after all, fascinate scientists.
New tools can directly, rather than through new data, inspire new theories.
This chapter extends the thesis of a tools-to-theories heuristic from statistical
tools to the computer.1 Recall that the thesis is twofold:

1. Discovery. New scientific tools, once entrenched in a scientist's daily

practice, suggest new theoretical metaphors and concepts.
2. Acceptance. Once proposed by an individual scientist (or a group),
the new theoretical metaphors and concepts are more likely to be
accepted by the scientific community if their members are also users
of the new tools.

This chapter is divided into two parts. In the first part, we argue that a
conceptual divorce between intelligence and calculation circa 1800, motivated
by a new social organization of work, made mechanical computation (and ul-
timately the computer) conceivable. The tools-to-theories heuristic comes into
play in the second part. When computers finally became standard laboratory
tools in the 20th century, the computer was proposed, and with some delay
accepted, as a model of mind. Thus we travel in a full circle from mind to
computer and back.

The work on which this chapter is based was coauthored with D. G. Goldstein.
1. Although we are only dealing with theories of mind, this does not imply that the
tools-to-theories heuristic is not applicable in the analysis of other scientific domains.
Schaffer (1992) provided several examples from the history of electromagnetism, in
which theories stemmed from tools. For instance, in 1600, the court physician William
Gilbert described the Earth as a vast spherical magnet. This new idea stemmed from the
tool he had invented (a magnet, the small terrella) and subsequently used as an analogy
to understand the world. This projection had consequences. Gilbert inferred that, be-
cause his terrella rotated, so did the Earth. The tool proved Copernicanism.

26
 MIND AS COMPUTER 27

From Mind to Computer

"Well, Babbage, what are you dreaming about?" to which I replied, "I
am thinking that all these tables (pointing to the logarithms) might be
calculated by machinery." (Charles Babbage, 1812/1994, p. 31)

The president of the Astronomical Society of London, Henry Colebrooke

(1825), summed up the significance of Babbage's work: "Mr. Babbage's inven-
tion puts an engine in place of the computer" (p. 510). This seems a strange
statement about the man who is now praised for having invented the computer.
But, at Babbage's time, the computer was a human being—in this case, some-
one who was hired for exhaustive calculations of astronomical and naviga-
tional tables.
How did Babbage (1791-1871) ever arrive at the idea of putting a mechan-
ical computer in place of a human one? A divorce between intelligence and
calculation, as Daston (1994) argued, made it possible for Babbage to conceive
this idea.
In the Enlightenment, calculation was not considered a rote, mechanical
thought process. In contrast, philosophers of the time held that intelligence
and even moral sentiment were in their essence forms of calculation. Calcu-
lation was the opposite of the habitual and the mechanical, remote from the
realm of menial labor. For Condillac, d'Alembert, Condorcet, and other En-
lightenment philosophers, the healthy mind worked by constantly taking apart
ideas and sensations into their minimal elements, then comparing and rear-
ranging these elements into novel combinations and permutations. Thought
was a combinatorial calculus, and great thinkers were proficient calculators.
In the eulogies of great mathematicians, for instance, prodigious mental reck-
oning was a favorite topic—Gauss's brilliant arithmetic was perhaps the last
of these stock legends. Calculation was the essence of moral sentiment, too.
Even self-interest and greed (as opposed to dangerous passions), by their nature
of being calculations, were at least predictable and thereby thought to reinforce
the orderliness of society (Daston, 1988, 1994).

The Computer as a Factory of Workers

By the turn of the nineteenth century, calculation was shifting from the com-
pany of hommes eclaires and savants to that of the unskilled work force.
Extraordinary mental arithmetic became associated with the idiot savant and
the sideshow attraction. Calculation became seen as dull, repetitive work,
best performed by patient minds that lacked imagination. Women ultimately
staffed the "bureaux de calculs" in major astronomical and statistical projects
(despite their earlier being accused of vivid imaginations and mental rest-
lessness; see Daston, 1992). Talent and genius ceased to be virtuoso com-
binatorics and permutations and turned into romantic, unanalyzable creations.
Thereby, the stage became set for the neoromanticism in twentieth-century
philosophy of science that declared creativity as mystical and the context of
28 WHERE DO NEW IDEAS COME FROM?

discovery as "irrelevant to the logical analysis of scientific knowledge" (Pop-

per, 1959, p. 31).
Daston (1994) and Schaffer (1994) argued that one force in this transfor-
mation was the introduction of large-scale division of labor in manufacturing,
as evidenced in the automatic system of the English machine-tool industry and
in the French government's large-scale manufacturing of logarithmic and trig-
onometric tables for the new decimal system in the 1790s. French engineer
Gaspard Riche de Prony organized the French government's titanic project for
the calculation of 10,000 sine values to the unprecedented precision of 25
decimal places and some 200,000 logarithms to 14 or 15 decimal places during
the French Revolution. Inspired by Adam Smith's praise of the division of
labor, Prony organized the project in a hierarchy of tasks. At the top was a
handful of excellent mathematicians, including Adrien Legendre and Lazare
Carnot, who devised the formulae; in the middle were 7 or 8 persons trained
in analysis; at the bottom were 70 or 80 unskilled persons who knew only the
rudiments of arithmetic and who performed millions of additions and sub-
tractions. These "manufacturing" methods, as Prony called them, pushed cal-
culation away from intelligence and toward work. The terms work and me-
chanical have been linked in both England and France since the middle of the
nineteenth century. Work concerned the body but not the mind; in large-scale
manufacturing, each worker did only one thing his or her whole life.
After it was shown that elaborate calculation could be carried out by an
assemblage of unskilled workers, each knowing very little about the large com-
putation, it became possible for Babbage to conceive of replacing these workers
with machinery. Babbage's view of the computer bore a great resemblance to
a factory of unskilled human workers. When Babbage talked about the parts
of his "Analytical Engine," the arithmetic computation and the storage of num-
bers, he called these the "mill" and the "store," respectively (Babbage, 1812/
1994, p. 23). The metaphor came from the textile industry, in which yarns were
brought from the store to the mill, were woven into fabric, and were then sent
back to the store. In the Analytical Engine, numbers were brought from the
store to the arithmetic mill for processing, and the results were returned to the
store. Commenting on this resemblance, Lady Lovelace said, "We may say
most aptly that the Analytical Engine weaves algebraic patterns just as the
Jaquard loom weaves flowers and leaves" (Babbage, 1812/1994, p. 27).2 In his
chapter on the "division of mental labor," Babbage explicitly referred to the
French government's program for the computation of new decimal tables as
the inspiration and foundation of a general science of machine intelligence.
Let us summarize the argument. During the Enlightenment, calculation was
the distinctive activity of the scientist and the genius and the very essence of

2. The Jaquard loom, a general-purpose device loaded with a set of punched cards,
could be used to weave infinite varieties of patterns. Factories in England were equipped
with hundreds of these machines, and Babbage was one of the "factory tourists" of the
1830s and 1840s.
 MIND AS COMPUTER 29

the mental life. New ideas and insights were assumed to be the product of the
novel combinations and permutations of ideas and sensations. In the first de-
cades of the nineteenth century, numerical calculation was separated from the
rest of intelligence and demoted to one of the lowest operations of the human
mind. After calculation became the repetitive task of an army of unskilled
workers, Babbage could envision mechanical computers replacing human
computers. Pools of human computers and Babbage's mechanical computer
manufactured numbers in the same way as the factories of the day manufac-
tured their goods.3

The Computer as a Brain

Babbage's dream that all tables of logarithms could be calculated by a machine,

however, did not turn into a reality during his lifetime. He never completed
any of the three machines he had started to build. Modern computers, such as
the ENIAC and the ED VAC at the University of Pennsylvania, came about dur-
ing and after World War II. Did the fathers of computer science see the mind
as a computer? We argue that the contemporary analogy stating that the mind
is a computer was not yet established before the "cognitive revolution" of the
1960s. As far as we can tell, two groups were willing to draw a parallel be-
tween the human and the computer, but neither used the computer as a theory
of mind. One group, which tentatively compared the nervous system and the
computer, is represented by Hungarian mathematician John von Neumann
(1903-1957). The other group, which investigated the idea that machines
might be capable of thought, is represented by English mathematician and
logician Alan Turing (1912-1954).
Von Neumann, known as the father of the modern computer, wrote about
the possibility of an analogy between the computer and the human nervous
system. It seems that von Neumann's reading of Warren McCulloch and Walter
Pitts's (1943) paper, "A Logical Calculus of the Ideas Immanent in Nervous
Activity," triggered his interest in information processing in the human brain
soon after the paper was published (Aspray, 1990). McCulloch and Pitts's paper
starts with the statement that, because of the all-or-none character of the ner-
vous system, neural events can be represented by means of propositional logic.
The McCulloch-Pitts model did not deal with the structure of neurons, which
were treated as "black boxes." The model was largely concerned with the
mathematical rules governing the input and output of signals. In a 1945 report
on ED VAC (the Electronic Discrete Variable Computer), von Neumann de-
scribed the computer as being built from McCulloch and Pitts's idealized neu-
rons rather than from vacuum tubes, electromechanical relays, or median -

3. Calculation became dissociated and opposed not only to the human intellect but
also to moral impulse. Madame de Stael, for instance, used the term calcul only in
connection with the "egoism and vanity" of those opportunists who exploited the
French Revolution for their own advantage and selfishness (Daston, 1994).
30 WHERE DO NEW IDEAS COME FROM?

ical switches. Understanding the computer in terms of the human nervous

system appeared strange to many, including the chief engineers of the ENIAC
project, Eckert and Mauchly (Aspray, 1990, p. 173). But, von Neumann hoped
that his theory of natural and artificial automata would improve understanding
of the design both of computers and of the human nervous system. His last
work (for the Silliman lectures), neither finished nor delivered due to illness,
was largely concerned with pointing out similarities between the nervous sys-
tem and the computer—between the neuron and the vacuum tube—but added
cautionary notes on their differences (von Neumann, 1958).
What was the reception of von Neumann's tentative analogy between the
nervous system and the computer? His intellectual biographer, Aspray (1990,
p. 181), concluded that psychologists and physiologists were less than enthu-
siastic about the McCulloch-Pitts model; Seymor Papert spoke of "a hostile or
indifferent world" (McCulloch, 1965, p. xvii), and McCulloch himself admitted
the initial lack of interest in their work (p. 9).

The Computer as a Mind

I believe that at the end of the century the use of words and general
educated opinion will have altered so much that one will be able to
speak of machines thinking without expecting to be contradicted. (Alan
Turing, 1950, p. 442)
Von Neumann and others looked for a parallel between the machine and the
human on the level of hardware. Turing (1950), in contrast, thought the ob-
servation that both the modern digital computer and the human nervous sys-
tem are electrical was based on a "very superficial similarity" (p. 439). He
pointed out that the first digital computer, Babbage's Analytical Engine, was
purely mechanical (as opposed to electrical) and that the important similarities
to the mind are in function rather than in hardware. Turing discussed the
question of whether machines can think rather than the question of whether
the mind is like a computer. Thus he was looking in a direction opposite that
in which psychologists were looking after the cognitive revolution, and con-
sequently he did not propose any theories of mind. For example, the famous
Turing test is about whether a machine can imitate a human mind but not vice
versa. Turing argued that it would be impossible for a human to imitate a
computer, as evidenced by the human's inability to perform complex numer-
ical calculations quickly. Turing also discussed the question of whether a com-
puter could be said to have free will, a property of humans. Many years later,
cognitive psychologists, under the assumptions that the mind is a computer
and that computers lack free will, pondered the question of whether humans
could be said to have free will. A similar story to this is that Turing (1947/
1969) contemplated teaching machines to be intelligent using the same prin-
ciples used to teach children. The analogy of the computer as a mind was
reversed again after the cognitive revolution, as McCorduck (1979) pointed out,
when Massachusetts Institute of Technology (MIT) psychologists tried to teach
children with the very methods that had worked for computers.
 MIND AS COMPUTER 31

Turing anticipated much of the new conceptual language and even the very
problems Allen Newell and Herbert Simon later attempted to address, as we
see in the second part of this chapter. With amazing prophecy, Turing sug-
gested that many intellectual issues can be translated into the form "Find a
number n such that. . ."; that is, he suggested that searching is the key concept
for problem solving and that Whitehead and Russell's (1935) Principia Math-
ematica might be a good start for demonstrating the power of the machine
(McCorduck, 1979, p. 57).
Not only did Turing's life end early and under tragic circumstances, but his
work had practically no influence on artificial intelligence in Britain until the
mid-1960s (McCorduck, 1979, p. 68). Neither von Neumann nor his friends
were persuaded to look beyond similarities between cells and diodes to func-
tional similarities between humans and computers.
To summarize, we have looked at two groups who compared humans and
computers before the cognitive revolution. One of these groups, represented
by von Neumann, spoke tentatively about the computer as a brain but warned
about taking the analogy too far. The other group, represented by Turing, asked
whether the computer has features of the human mind but not vice versa—
that is, this group did not attempt to design theories of mind through the
analogy of the tool.
Before the second half of the century, the mind was not yet a computer.
However, a new incarnation of the Enlightenment view of intelligence as a
combinatorial calculus was on the horizon.

From Computer to Mind

The computer is a member of an important family of artifacts called

symbol systems, or more explicitly, physical symbol systems . . . The hy-
pothesis is that a physical symbol system . . . has the necessary and suf-
ficient means for general intelligent action. (Herbert Simon, 1969, p. 26)

What has been called in retrospect the cognitive revolution in American psy-
chology of the 1960s is more than an overthrow of behaviorism by mentalist
concepts. The cognitive revolution did more than revive the mental; it changed
its meaning. One source of this change is the projection of new tools (i.e.,
statistics and computers) into the mind. We refer to this heuristic of discovery
as the tools-to-theories heuristic. The two new classes of theories that emerged
and that partially overlap pictured the new mind as an "intuitive statistician"
or a "computer program."
In this section, we see how a tools-to-theories explanation accounts for the
new conception of the mind as a computer, focusing on the discovery and
acceptance of Simon and Newell's brand of information-processing psychol-
ogy. We try to reconstruct the discovery of Newell and Simon's (1972)
information-processing model of mind and its (delayed) acceptance by the psy-
chological community in terms of the tools-to-theories heuristic.
32 WHERE DO NEW IDEAS COME FROM?

Discovery

Babbage's mechanical computer was preceded by human computers. Similarly,

Newell and Simon's first computer program, the "Logic Theorist" (LT), was
preceded by a human computer. Before the LT was up and running, Newell
and Simon reconstructed their computer program out of human components
(Simon's wife, children, and several graduate students) in order to see if it
would work. Newell wrote up the subroutines of the LT program on index
cards:
To each member of the group, we gave one of the cards, so that each
person became, in effect, a component of the LT computer program—a
subroutine that performed some special function, or a component of its
memory. It was the task of each participant to execute his or her subrou-
tine, or to provide the contents of his or her memory, whenever called
by the routine at the next level above that was then in control.
So we were able to simulate the behavior of the LT with a computer
constructed of human components. . . . The actors were no more respon-
sible . . . than the slave boy in Plato's Meno, but they were successful in
proving the theorems given them. (Simon, 1991, p. 207)
The parallels to Prony's bureaux de calculs and the large-scale manufac-
turing of the new factories of the early nineteenth century are striking. At
essence is a division of labor, in which the work is done by a hierarchy of
humans—each requiring little skill and repeating the same routine again and
again. Complex processes are achieved by an army of workers who never see
but a little piece of the larger picture.4
However, between Prony's human computer and Simon's human computer
is an important difference. Prony's human computer and Babbage's mechanical
computer (modeled after it) performed numerical calculations. Simon's human
computer did not. Simon's humans matched symbols, applied rules to sym-
bols, and searched through lists of symbols—in short, performed what is now
generally known as symbol manipulation.
The reader will recall from the first part of this chapter that the divorce
between intelligence and numerical calculation made it possible for Babbage
to replace the human computer with a mechanical one. In the twentieth cen-
tury, intelligence and calculation are still divorced. Given this divorce and the
early conception of the computer as a fancy number cruncher, it is no wonder
that the computer never suggested itself as a theory of mind. We argue that an

4. The Manhattan Project at Los Alamos, where the atomic bomb was constructed,
housed another human computer. Although the project could draw on the best technol-
ogy available, in the early 1940s mechanical calculators (e.g., the typewriter-sized Mar-
chant calculator) could only add, subtract, multiply, and, with some difficulty, divide.
Richard Feynman and Nicholas Metropolis arranged a pool of people (mostly scientists'
wives, who were getting paid three-eighths of the scientists' salary), each of whom re-
petitively performed a small calculation (e.g., cubing a number) and passed the result
on to another person, who incorporated it into yet another computation (Gleick, 1992).
 MIND AS COMPUTER 33

important precondition for the view of mind as a computer is the realization

that computers are symbol-manipulation devices in addition to being numer-
ical calculators. Newell and Simon were among the first to realize this. In
interviews with Pamela McCorduck (1979), Newell recalled, "I've never used
a computer to do any numerical processing in my life" (p. 129). Newell's first
use of the computer at RAND Corporation—a prehistoric card-programmed
calculator hooked up to a line printer—was printing symbols representing air-
planes for each sweep of a radar antenna.
The symbol-manipulating nature of the computer was important to Simon
because it corresponded to some of his earlier views on the nature of intelli-
gence:

The metaphor I'd been using, of a mind as something that took some
premises and ground them up and processed them into conclusions, be-
gan to transform itself into a notion that a mind was something which
took some program inputs and data and had some processes which op-
erated on the data and produced output, (cited in McCorduck, 1979,
p. 127)

It is interesting to note that 20 years after seeing the computer as a symbol-

manipulating device, Newell and Simon came forth with the explicit hypoth-
esis that a physical symbol system is necessary and sufficient for intelligence.
The Logic Theorist generated proofs for theorems in symbolic logic—spe-
cifically, the first 25 or so theorems in Whitehead and Russell's (1935) Principia
Mathematica. It even managed to find a proof more elegant than the corre-
sponding one in the Principia Mathematica.
In the summer of 1958, psychology was given a double dose of the new
school of information-processing psychology. One dose was the publication of
Newell, Shaw, and Simon's (1958) Psychological Review article, "Elements of
a Theory of Human Problem Solving"; the other dose was the Research Train-
ing Institute on the Simulation of Cognitive Processes at the RAND Corpora-
tion, which we discuss later.
The Psychological Review article is an interesting document of the transi-
tion between the view that the LT is a tool for proving theorems in logic (the
artificial intelligence view) and an emerging view that the LT is a model of
human reasoning (the information-processing view). In fact, Newell et al.
(1958) went back and forth between both views, explaining that "the program
of LT was not fashioned directly as a theory of human behavior; it was con-
structed in order to get a program that would prove theorems in logic" (p. 154);
later, they wrote that the LT "provides an explanation for the processes used
by humans to solve problems in symbolic logic" (p. 163). The evidence pro-
vided for projecting the machine into the mind is mainly rhetorical. For in-
stance, Newell et al. spent several pages arguing for the resemblance between
the methods of LT and concepts (e.g., "set," "insight," "hierarchy") described
in the earlier psychological literature on human problem solving.
In all fairness, despite Newell et al.'s claim, the resemblance to these earlier
concepts as they were used in the work of Karl Duncker, Wolfgang Kohler, and
34 WHERE DO NEW IDEAS COME FROM?

others is slight. New discoveries, by definition, clash with what has come be-
fore, but it is often a useful strategy to hide the amount of novelty and to claim
historical continuity. When Tanner and Swets (1954) proposed (in the Psycho-
logical Review four years earlier) that another scientific tool (i.e., the Neyman-
Pearsonian techniques of hypothesis testing) would model the cognitive pro-
cesses of stimulus detection and discrimination, their signal-detection model
also clashed with earlier notions, such as the notion of a sensory threshold.
Tanner and Swets, however, chose not to conceal this schism between the old
and the new theories, explicitly stating that their new theory "appears to be
inconsistent with the large quantity of existing data on this subject" (p. 401).
As we argued before, there is a different historical continuity in which Newell
and Simon's ideas stand—the earlier Enlightenment view of intelligence as a
combinatorial calculus.

Conceptual Change

Newell et al. (1958) tried to emphasize the historical continuity of what was
to become their new information-processing model of problem solving, as did
Miller, Galanter, and Pribram (1960) in their Plans and the Structure of Be-
havior when they linked their version of Newell and Simon's theory to many
great names such as William James, Frederic Bartlett, and Edward Tolman. We
believe that these early claims for historical continuity served as protection:
George Miller, who was accused by Newell and Simon of having stolen their
ideas and gotten them all wrong, said, "I had to put the scholarship into the
book, so they would no longer claim that those were their ideas. As far as I
was concerned they were old familiar ideas" (Baars, 1986, p. 213). In contrast
to this rhetoric, here we emphasize the discontinuity introduced by the trans-
formation of the new tool into a theory of mind.

The New Mind

What was later called the "new mental chemistry" pictured the mind as a
computer program:

The atoms of this mental chemistry are symbols, which are combinable
into larger and more complex associational structures called lists and list
structures. The fundamental "reactions" of the mental chemistry employ
elementary information processes that operate upon symbols and symbol
structures: copying symbols, storing symbols, retrieving symbols, input-
ting and outputting symbols, and comparing symbols. (Simon, 1979,
p. 363)

This atomic view is certainly a major conceptual change in the views about
problem solving compared to the theories of Kohler, Wertheimer, and Duncker,
 MIND AS COMPUTER 35

but it bears much resemblance to the combinatorial view of intelligence of the

Enlightenment philosophers.5
The different physical levels of a computer lead to Newell's cognitive
hierarchy, which separates the knowledge level, symbol level, and register-
transfer levels of cognition. The seriality of 1971-style computers is actually
embedded in Newell's cognitive theory (Arbib, 1993).
One of the major concepts in computer programming that made its way into
the new models of the mind is the decomposition of complexity into simpler
units, such as the decomposition of a program into a hierarchy of simpler
subroutines or into a set of production rules. On this analogy, the most com-
plex processes in psychology, such as scientific discovery, can be explained
through simple subprocesses. Thus the possibility of the logic of scientific
discovery, the existence of which Karl Popper so vehemently disclaimed, has
returned in the analogy between computer and mind (Langley, Simon, Brad-
shaw, & Zytkow, 1987).
The first general statement of Newell and Simon's new vision of mind ap-
peared in their 1972 book, Human Problem Solving. Newell and Simon argued
for the idea that higher level cognition proceeds much like the behavior of a
production system—a formalism from computer science (and before that sym-
bolic logic) that had never before been used in psychological modeling. Newell
and Simon (1972) wrote of the influence of programming concepts on their
models:

Throughout the book we have made use of a wide range of organizational

techniques known to the programming world: explicit flow control, sub-
routines, recursion, iteration statements, local naming, production sys-
tems, interpreters, and so on. . . . We confess to a strong premonition that
the actual organization of human programs closely resembles the pro-
duction system organization, (p. 803)

Here we do not attempt to probe the depths of how Newell and Simon's
ideas of information processing changed theories of mind; the commonplace
usage of computer terminology in the cognitive psychological literature since
1972 is a reflection of this. How natural it seems for present-day psychologists
to speak of cognition in terms of encoding, storage, retrieval, executive pro-
cesses, algorithms, and computational cost.

5. In fact, the new view was directly inspired by 19th-century mathematician George
Boole (1854/1958), who, in the very spirit of the Enlightenment mathematicians such as
the Bernoullis and Laplace, set out to derive the laws of logic, algebra, and probability
from what he believed to be the laws of human thought. Boole's algebra culminated in
Whitehead and Russell's (1935) Principia Mathematica, describing the relation between
mathematics and logic, and in Claude Shannon's seminal work (his master's thesis at
MIT in 1940), which used Boolean algebra to describe the behavior of relay and switch-
ing circuits (McCorduck, 1979, p. 41).
36 WHERE DO NEW IDEAS COME FROM?

New Experiments, New Data

The tools-to-theories heuristic implies that new theories need not be a conse-
quence of new experiments and new data. Instead, new tools can transform
the kinds of experiments performed and data collected. This consequence of
the tools-to-theories heuristic is also known to have happened when statistical
tools turned into theories of mind (and around the same time).
One such case is the revolution of psychophysics through a new tool called
Neyman-Pearsonian hypothesis testing (see Chapter 1). The new theory of
mind inspired by this tool is known as signal-detection theory. The Neyman-
Pearson technique deals with two kinds of errors, Type I and Type II (or false
alarms and misses). When Tanner and Swets (1954) projected the tool into
the mind, stimulus detection and discrimination—earlier understood in terms
of "thresholds"—then became seen as a decision between two competing
hypotheses based on a criterion that balances the probability of two kinds of
errors. Consequently, the avalanche of experiments on auditory and visual
detection and discrimination that followed their proposal kept track of both
kinds of error in participants' judgments. The important point is that earlier
experiments, such as the classical works of Fechner and Thurstone, paid
attention to only one kind of error (Gigerenzer, 1994b). What happened is that
the new statistical tool inspired a new theory of mind, which in turn changed
the kind of data generated in research. In this way, Tanner and Swets were
able, in good conscience, to discard the years of contradicting results that
preceded them.
A similar story is to be told with the conceptual change brought about by
Newell and Simon—it mandated a new type of experiment that in turn in-
volved new kinds of subjects, data, and justification. In academic psychology
of the day, the standard experimental design, modeled after the statistical
methods of Ronald Fisher, involved many subjects and randomized treatment
groups. The 1958 Psychological Review article used the same terminology of
design of the experiment and subject but radically changed their meanings.
There were no longer groups of human or animal subjects. There was only one
subject—an inanimate being, Logic Theorist. There was no longer an experi-
ment in which data are generated by either observation or measurement. Ex-
periment took on the meaning of simulation.
In this new kind of experiment, the data were of an unforeseen type—com-
puter printouts of the intermediate results of the program. These new data, in
turn, required new methods of hypothesis testing. How did Newell and Simon
tell if their program was doing what minds do? There were two methods. For
Newell and Simon, simulation was a form of justification itself: A theory that
is coded as a working computer program shows that the processes it describes
are, at the very least, sufficient to perform the task, or, in the more succinct
words of Simon (1992a), "A running program is the moment of truth" (p. 155).
Furthermore, a stronger test of the model is made by comparing the output of
the computer to the think-aloud protocols of human subjects.
 MIND AS COMPUTER 37

Although all of this was a methodological revolution in the experimental

practice of the time, some important parallels exist between the new
information-processing approach and the turn-of-the-century German ap-
proach to studying mental processes. These parallels concern the analysis of
individual subjects (rather than group means), the use of think-aloud proce-
dures, and the status of the subject. In the early German psychology, as well
as in American psychology of the time (until about the 1930s), the unit of
analysis was the individual person, not the average of a group (Danziger, 1990).
The two most prominent kinds of data in early German psychology were re-
action times and introspective reports. Introspective reports have been
frowned on ever since the inception of American behaviorism, but think-aloud
protocols, their grandchildren, are back (as are reaction times). Furthermore,
in the tradition of the Leipzig (Wundt) and Wiirzburg (Kiilpe) schools, the
subject was more prestigious and important than the experimenter. Under the
assumption that the thought process is introspectively penetrable, the subject,
not the experimenter, was assumed to provide the theoretical description of
the thought process. In fact, the main experimental contribution of Kiilpe, the
founder of the Wiirzburg school, was to serve as a subject, and it was often
the subject who published the article. In the true spirit of these schools, Newell
and Simon put their subject, the LT, as a coauthor of a paper submitted to the
Journal of Symbolic Logic. Regrettably, the paper was rejected (as it contained
no new results from the point of view of modern logic), and the LT never tried
to publish again.

Acceptance

The second dose of information processing administered to psychology (after

the 1958 Psychological Review article) was the Research Training Institute
on the Simulation of Cognitive Processes at the RAND Corporation, organized
by Newell and Simon. At the institute, lectures and seminars were conducted;
IPL-IV programming was taught; and the LT, the General Problem Solver, and
the EPAM model of memory were demonstrated on the RAND computer. In
attendance were some scientists who would eventually develop computer-
simulation methods of their own—including George Miller, Robert Abelson,
Bert Green, and Roger Shepard.
An early but deceptive harbinger of acceptance for the new information-
processing theory was the publication, right after the summer institute, of
Plans and the Structure of Behavior (Miller et al., 1960). Despite the afore-
mentioned 1959 dispute with Newell and Simon over the ownership and va-
lidity of the ideas within, this book drew a good deal of attention from all of
psychology.
It would seem the table was set for the new information-processing psy-^
chology; however, it did not take hold. Simon (1991, p. 232) complained of
the psychological community, which took only a "cautious interest" in Newell
38 WHERE DO NEW IDEAS COME FROM?

and Simon's ideas. The "acceptance" part of the tools-to-theories thesis can
explain this: Computers were not yet entrenched in the daily routine of psy-
chologists, as we show here.

No Familiar Tools, No Acceptance

We take two institutions as case studies to demonstrate the part of the tools-
to-theories hypothesis that concerns acceptance—the Harvard University Cen-
ter for Cognitive Studies and Carnegie-Mellon University (CMU). The former
never came to embrace fully the new information-processing psychology; the
latter did but after a considerable delay. Tools-to-theories might explain both
phenomena.
George Miller, the cofounder of the Center for Cognitive Studies, was cer-
tainly a proponent of the new information-processing psychology. As we said,
Miller et al.'s (1960) Plans and the Structure of Behavior was so near to Newell
et al.'s (1958) ideas that it was at first considered a form of theft, although the
version of the book that did see the presses is filled with citations recognizing
Newell et al. Given Miller's enthusiasm, one might expect the center, partially
under Miller's leadership, to blossom into information-processing research. It
never did. Looking at the 1963-1969 annual reports (Harvard University Cen-
ter for Cognitive Studies, 1963, 1964, 1966, 1968, 1969), we found only a few
symposia or papers dealing with computer simulation.
Although the center had a PDP-4C Computer and the reports anticipated
the possibility of using it for cognitive simulation, as late as 1969 it never
happened. The reports mention that the computer served to run experiments,
demonstrate the feasibility of computer research, and draw visitors to the lab-
oratory. However, difficulties involved in using the tool were considerable. The
PDF saw 83 hours of use on an average week in 1965-1966, but 56 of these
were spent on debugging and maintenance. In the annual reports are several
remarks of the type, "It is difficult to program computers. . . . Getting a program
to work may take months." The center even turned out a 1966 technical report
entitled Programmanship, or How to Be One-Up on a Computer without Ac-
tually Ripping out Its Wires.
What might have kept the Harvard computer from becoming a metaphor
of the mind was that the researchers could not integrate this tool into their
everyday laboratory routine. The tool even turned out to be a steady source
of frustration. As tools-to-theories suggests, this lack of entrenchment in every-
day practice accounted for the lack of acceptance of the new information-
processing psychology. Simon (1979) took notice of this:
Perhaps the most important factors that impeded the diffusion of the new
ideas, however, were the unfamiliarity of psychologists with computers
and the unavailability on most campuses of machines and associated
software (list processing programming languages) that were well adapted
to cognitive simulation. The 1958 RAND Summer Workshop, mentioned
earlier, and similar workshops held in 1962 and 1963, did a good deal
to solve the first problem for the 50 or 60 psychologists who participated
 MIND AS COMPUTER 39

in them; but workshop members often returned to their home campuses

to find their local computing facilities ill-adapted to their needs, (p. 365)

At CMU, Newell, Simon, a new information-processing-enthusiastic de-

partment head, and a very large National Institute of Mental Health (NIMH)
grant were pushing "the new IP [information processing] religion" (H. A.
Simon, personal communication, 1994). Even this concerted effort failed to
proselytize the majority of researchers within their own department. This again
indicates that entrenchment of the new tool in everyday practice was an im-
portant precondition for the spread of the metaphor of the mind as a computer.

Acceptance of Theory Follows Familiarity with Tool

At CMU in the late 1950s, the first doctoral theses involving computer simu-
lation of cognitive processes were being written (H. A. Simon, personal com-
munication, 1994). But this was not representative of the national state of af-
fairs. In the mid-1960s, a small number of psychological laboratories were built
around computers, including those of CMU, Harvard, Michigan, Indiana, MIT,
and Stanford (Aaronson, Grupsmith, & Aaronson, 1976, p. 130). As indicated
by the funding history of NIMH grants for cognitive research, the amount of
computer-using research tripled over the next decade. In 1967, only 15% of
the grants being funded had budget items related to computers (e.g., program-
mer salaries, hardware, supplies); by 1975, this figure had increased to 46%.
The late 1960s saw a turn toward mainframe computers that lasted until the
late 1970s, when the microcomputer started its invasion of the laboratory. In
the 1978 Behavioral Research Methods & Instrumentation conference, micro-
computers were the issue of the day (Castellan, 1981, p. 93). By 1984, the
journal Behavioral Research Methods &- Instrumentation appended the word
Computers to its title to reflect the broad interest in the new tool. By 1980, the
cost of computers had dropped an order of magnitude from what it was in
1970 (Castellan, 1981, 1991). During the last two decades, computers have
become the indispensable research tool of the psychologist.
After the tool became entrenched in everyday laboratory routine, a broad
acceptance of the view of the mind as a computer followed. In the early 1970s,
information-processing psychology finally caught on at CMU. Every CMU-
authored article in the proceedings of the 1973 Carnegie Symposium on Cog-
nition mentions some sort of computer simulation. For the rest of the psycho-
logical community, which was not as familiar with the tool, the date of broad
acceptance was years later. Simon (1979) estimated that, from about 1973 to
1979, the number of active research scientists working in the information-
processing vein had "probably doubled or tripled" (p. 390).
This does not mean that the associated methodology became accepted as
well. It clashed too strongly with the methodological ritual that was institu-
tionalized during the 1940s and 1950s in experimental psychology. We use the
term ritual here for the mechanical practice of a curious mishmash between
Fisher's and Neyman-Pearson's statistical techniques, which was taught to psy-
40 WHERE DO NEW IDEAS COME FROM?

chologists as the sine qua non of scientific method (see Chapter 13). Most
psychologists assumed, as the textbooks told them, that there is only one way
to do good science. But their own heroes—Fechner, Wundt, Pavlov, Kb'hler,
Bartlett, Piaget, Skinner, and Luce, to name a few—had never used this "rit-
ual." Some had used experimental practices that resembled the newly pro-
posed methods used to study the mind as computer.

Pragmatics

Some of our experimental colleagues have objected to our earlier analysis of

how statistical tools turned into theories of mind. They have argued that tools
are irrelevant in discovery and that our tools-to-theories examples are merely
illustrations of psychologists' being quick to realize that the mathematical
structure of a tool (e.g., ANOVA) is precisely that of the mind. It is not easy to
convince someone who believes (in good Neoplatonic fashion) that today's
theory of mind exactly fits the nature of the mind—that such a splendid theory
might mirror something other than reality pure and simple. If it were true that
tools have no role in discovery and that the new theories just happen to mirror
the mathematical structure of the tool, then the pragmatics of the use of a
tool—which is independent of the mathematical structure—would find no
place in the new theories. In this section, however, we provide evidence that
not only the new tool but also its pragmatic uses are projected into the mind.
The tools-to-theories heuristic cannot be used to defend a spurious Neoplaton-
ism.
One example is Kelley's (1967) causal attribution theory, which postulates
that the mind draws a causal inference in the same way social scientists do,
by using Fisher's ANOVA. As described in Chapter 1, the pragmatics, in ad-
dition to the mathematics of ANOVA, were projected into the mind. The prac-
tical use of a tool is generally undetermined by its mathematical structure. The
mathematics of significance testing, as in ANOVA, has been used both for re-
jecting hypotheses based on data and for rejecting data (e.g., outliers in astro-
nomical observations) based on hypotheses. Scientists have to get rid of both
bad hypotheses and bad data. In the psychological laboratories, however,
ANOVA was (and is) used almost exclusively for rejecting hypotheses based
on data. Dubious data, in contrast, were (and still are) dealt with informally.
When Kelley projected ANOVA into the mind, this specific, practical use (i.e.,
rejecting hypotheses) was projected along with it. In sharp contrast to earlier
theoretical accounts, such as Michotte's and Piaget's, causal inference was seen
as data driven, as an inductive inference from data to causes. Kelley's new
mind used the tool in the same way the researcher uses the tool—to trust the
data (the information given) and to mistrust the hypotheses. The inductive
view of causal attribution became one of the classic topics of social psychology,
even to the point of defining the field.
 MIND AS COMPUTER 41

The same process of projecting pragmatic aspects of the use of a tool into
a theory can be shown for the view of the mind as a computer. One example
is Levelt's (1989) model of speaking. The basic unit in Levelt's model, which
he called the "processing component," corresponds to the computer program-
mer's concept of a subroutine. We argue that Levelt's model not only borrowed
the subroutine as a tool but also borrowed the practical aspects of how sub-
routines are used and constructed in computer programming.
A subroutine (or "subprocess") is a group of computer instructions (usually
serving a specific function) that are separated from the main routine of a com-
puter program. It is common for subroutines to perform often needed func-
tions, such as extracting cube roots or rounding numbers. There is a major
pragmatic issue involved in writing subroutines that centers on the "principle
of isolation" (Simon & Newell, 1986). The issue is whether subroutines should
be black boxes or not. According to the principle of isolation, the internal
workings of the subroutine should remain a mystery to the main program, and
the outside program should remain a mystery to the subroutine. Black-box
subroutines have become known as program modules, perfect for the divide-
and-conquer strategy programmers often use to tackle large problems. To the
computer, however, it makes no difference whether subroutines are isolated or
not. Subroutines that are not isolated work just as well as those that are. The
only real difference between the two types of subroutine is psychological. Sub-
routines that violate the principle of isolation are more difficult for the pro-
grammer to read, write, debug, maintain, and reuse. For this reason, introduc-
tory texts on computer programming stress the principle of isolation as the
very essence of good programming style.
The principle of isolation—a pragmatic feature of using subroutines as a
programming tool—has a central place in Levelt's model, in which the pro-
cessing components are "black boxes" that exemplify Fodor's notion of infor-
mational encapsulation (Levelt, 1989, p. 15). In this way, Levelt's psychological
model embodies a maxim of good computer programming—the principle of
isolation. That this practical aspect of the use of the tool shaped a theory of
speaking is not an evaluation of the quality of the theory. Our point concerns
origins, not validity. However, this pragmatic feature of subroutines has not
always served the model well. Kita (1993) and Levinson (1992) have attacked
Levelt's model at its Achilles' heel—its insistence on isolated processing com-
ponents.
To summarize the second part of this chapter, we started with the separation
between intelligence and calculation and argued that the realization that com-
puters can do more than arithmetic was an important precondition for the view
of the mind as a computer. Newell and Simon seem to have been the first who
tried to understand the mind in terms of a computer program, but the accep-
tance of their information-processing view was delayed until the psychologists
became used to computers in their daily laboratory routine. We have argued
that, along with the tool, its pragmatic use has been projected into theories of
mind. Now that the metaphor is in place, many find it difficult to see how the
42 WHERE DO NEW IDEAS COME FROM?

mind could be anything else: To quote Philip Johnson-Laird (1983): "The com-
puter is the last metaphor; it need never be supplanted" (p. 10).

Social Computers

The tools-to-theories heuristic can reverse the commonly assumed fixed tem-
poral order between discovery and justification—discovery first, justification
second. New tools for justification enter the laboratory first, new theories fol-
low. In the case of Babbage's computer, the tool itself was modeled after a new
social system, the organization of work in large-scale manufacturing. The
model for the machine computer was a social computer.
The argument was that economic changes—the large-scale division of labor
in manufacturing and in the "bureaux de calculs"—went along with the break-
down of the Enlightenment conception of the mind, in which calculation was
the distinctive essence of intelligence. Once calculation was separated from
the rest of intelligence and relegated to the status of a dull and repetitive task,
Babbage could envision replacing human computers with mechanical ones.
Both human and mechanical computers manufactured numbers as the factories
of the day manufactured goods. In the twentieth century, the technology be-
came available to make Babbage's dream a reality. Computers became indis-
pensable scientific tools for everything from number crunching to simulation.
Our focus was on the work by Herbert Simon and Allen Newell and their
colleagues, who proposed the tool as a theory of mind. Their proposal reunited
mere calculation with what was now called "symbol processing," returning to
the Enlightenment conception of mind. After personal computers found a
place in nearly every psychological laboratory, broad acceptance of the meta-
phor of the mind as computer followed.6
A question remains: Why was the digital computer used as a model of the
individual mind rather than of the social organization of many minds? As the
social roots of the idea behind Babbage's computer shows, there is nothing
inherently individualistic about the business of computation. We can only
speculate that it was the traditional focus of psychological research on indi-
viduals that suggested the analogy between the computer and the individual
mind and that in less individualistic disciplines the computer would have had
a better chance of becoming a model of social organization. In fact, anthro-
pologist Ed Hutchins (1995) has proposed using the digital computer as a
model of how social groups make decisions, for instance, how a crew on a
large ship solves the problem of navigation. Here the computer is used to

6. The reconstruction of the path "from mind to computer and back" also provides
an explanation for one widespread type of resistance to the computer metaphor of mind.
The post-Enlightenment divorce between intelligence and calculation still holds to this
day, and, for those who still associate the computer with mere calculation (as opposed
to symbol processing), the mind-as-a-computer is a contradiction in itself.
 MIND AS COMPUTER 43

model the division of labor in the storing, processing, and exchange of infor-
mation among members of a social group. This notion of distributed intelli-
gence completes the circle traveled by the computer metaphor. Once modeled
after the social organization of human work, the computer has now become a
model of the social organization of human work.
 3

Ideas in Exile
The Struggles of an Upright Man

I he sparkling intellectual atmosphere of early twentieth-century Vienna pro-

duced Wittgenstein, Popper, Neurath, and Gb'del—in addition to a string of
other great thinkers. Among them was Karl Biihler, who, when he founded the
Vienna Psychological Institute in 1922, was one of the foremost psychologists
in the world. Egon Brunswik began to study psychology in Vienna in 1923
and soon became an active participant in Buhler's famous Wednesday evening
discussion group; on Thursdays he went to Moritz Schlick's Thursday evening
discussion group (Leary, 1987). Schlick was the founder and leading member
of the European school of positivist philosophers known as the Vienna Circle.
In 1927, Brunswik submitted his doctoral thesis to Biihler and Schlick, the
same two advisors to whom Karl Popper submitted his thesis a year later.
The intellectual tension between Wednesday and Thursday evenings was
vibrant. The logical positivist doctrine of the Vienna Circle posited that the
relation between scientific language and its sense-data referents should and
could be unambiguous. Biihler, in contrast, had shown that the relation be-
tween perceptual cues and their objects, as well as between words and their
objects, was irreducibly ambiguous. Brunswik sided with Buhler. He did try,
though, to resolve the tension by adopting the position of Hans Reichenbach,
the leader of the Berlin school of logical positivism, who argued that all knowl-
edge is probabilistic.
Influenced by Buhler's biologically motivated concern with the success of
organisms in their world, Brunswik's research in the 1920s and 1930s aimed
at studying "perceptual achievement" in the presence of ambiguous cues. The
three traditional perceptual constancies—size, shape, and color—were the pro-
totype for achievement, that is, how accurate perception is when aspects of
the environment change. Brunswik extended the question of how well an or-
ganism infers size, shape, and color under varying context variables (such as
illumination) to the more general problem of studying the invariance of the
perception of one characteristic of an object when the others vary. For instance,
he studied how the perceived size of coins changed when their value and the

44
 IDEAS IN EXILE 45

number of coins were varied—coins higher in value appeared to be larger in

size and greater in number than those of lesser value (Brunswik, 1934, p. 147).
In Brunswik's terms, what we see are perceptual compromises that he attrib-
uted to the learning of cues from experience (e.g., coins of higher value actually
do tend to be larger in size). He manipulated up to four variables simultane-
ously in factorial designs (he had not yet developed the idea of representative
design) and measured how the perception of each variable depended on the
values of the others. This Vienna program of "multidimensional psychophys-
ics" measured the context-dependency of judgment (for an introduction see
Gigerenzer & Murray, 1987, pp. 61-81). In contrast, its independence from con-
text was assumed in the one-dimensional psychophysics associated with G. T.
Fechner and S. S. Stevens, in which one studied a variable in isolation (such
as perceived size), held everything else constant, and then compared the per-
ceived with the actual size to obtain the psychophysical function.
In the early 1930s, Brunswik was far ahead of mainstream psychophysics
in the study of context dependency. This is not to say that there was no room
for theoretical development in his multidimensional psychophysics; for in-
stance, Brunswik treated his two explanatory concepts, perceptual compro-
mises and cue learning, as equivalent, whereas these actually are different and
can lead to contradictory predictions about the effect of context (Gigerenzer &
Murray, 1987, pp. 70-74). However, his Vienna program had virtually no im-
pact on the future of psychophysics, except for a few scattered studies.
One reason for this lack of influence was that the Vienna Psychological
Institute's program was destroyed soon after Brunswik accepted a position at
Berkeley in 1937. In early 1938, the Nazis entered Vienna and arrested and
dismissed Biihler because of his political views, which were considered dan-
gerous to the "peace and public order of the (Philosophical) Faculty" (Ash,
1987, p. 157). Eventually, he fled to the United States, but no one offered the
once celebrated Karl Biihler an adequate position; his brilliant career crumbled
in exile. Schlick had died a few years earlier from gunshot wounds inflicted
by a deranged student, and the political pressure of fascism caused the Vienna
Circle to disband, with many of its members fleeing to the United States. Brun-
swik had to start practically from scratch at Berkeley.

Brunswik in the Plural

Unlike the Vienna program, Brunswik's Berkeley program—probabilistic func-

tionalism—is well known. It is so well known, in fact, that there is not just
one Brunswik, but several. One is the Brunswik absorbed by contemporary
psychology: he-was-one-of-us. These good-natured colleagues spell his name
"Brunswick," confuse his term ecological validity with generalizability from
the laboratory to the environment and representative design with the repre-
sentativeness heuristic. In their friendly embrace, Brunswik comes out a fore-
runner and guardian of today's status quo. No conflict surfaces, all is quiet,
nothing must be questioned.
46 WHERE DO NEW IDEAS COME FROM?

There is a more sophisticated image, in which Brunswik's ideas basically

boil down to three correlations and one unorthodoxy. The correlations are
functional validities, ecological validities, and cue utilization coefficients, and
the heresy is representative design—the frightening idea of sacrificing experi-
mental control and, possibly even worse, of leaving one's laboratory to study
people in their real-world environments. Correlations are fine, the unortho-
doxy is repugnant. This view gets some work done, but it cuts right through
the middle of Brunswik's intellectual heart.
There is a third view of Brunswik: opposition by neglect. This is not an
active opposition against an intellectual enemy; Brunswik does not seem to
have notable intellectual enemies, unlike many other scholars. The opposition
takes the form of silence and a lack of understanding of what the fuss is all
about. For instance, in his Sensation and Perception in the History of Exper-
imental Psychology (1942), Edwin G. Boring, the dean of the history of psy-
chology and an arch-determinist, covered Brunswik's work in Vienna, which
encompassed experimental control in multidimensional designs. But after
Brunswik had fleshed out his probabilistic functionalism and representative
design, he was not even mentioned in Boring's A History of Experimental Psy-
chology (1957) and History, Psychology, and Science (1963). As Ken Hammond
(1980, p. 9) reported, Boring's verdict was "Brunswik was a brilliant man who
wasted his life." Informed neglect can be as toxic to new ideas as an unin-
formed embrace.
In the following, I describe what I think of when I think of Brunswik. I do
not think of correlations; I think of the struggles of an upright man.

Intellectual Integrity

What impresses me deeply is Brunswik's uncompromising intellectual sincer-

ity: the courage to think through the consequences of one's ideas carefully and
to speak out in public even when the scientific community does not want to
listen and makes one pay a price for these standards. And Brunswik paid
dearly. Brunswik's personal struggle was, in my view, about maintaining his
intellectual integrity in a scientific community in which his ideas fell on hos-
tile ground held by ignorant troops. Great thinkers often learn, to their surprise,
that new ideas are less than welcome.
What were these new ideas that inspired so much hostility? Brunswik's
probabilistic functionalism can be summarized in the following concepts:
achievement, ambiguity of cues, vicarious functioning, and representative de-
sign. That is, an organism needs to make inferences about its environment to
adjust, survive, and reproduce (achievement); the proximal cues available to
it to make these inferences about its environment are uncertain (ambiguity);
the organism processes ambiguous cues by substituting or combining them
(vicarious functioning); in order to study achievement and vicarious function-
ing, researchers need to use representative designs. This is Brunswik's linking
 IDEAS IN EXILE 47

of biological purpose, environment, cognitive process, and research method-

ology.
The hostile ground itself was a minefield of dogmas: determinism, the Co-
lumbia Bible, and Fisher's experimental design. Determinism was a fading, but
still strong, dogma, and the other two were newly emerging dogmas.

Refinancing Determinism

In their struggle to get psychology recognized as a science, many of Brunswik's

fellow psychologists in America maintained an old-fashioned ideal no longer
characteristic of modern science, from evolutionary biology to quantum phys-
ics. This ideal demanded certain knowledge and universal laws, as Newtonian
mechanics had purported to deliver. As an example of this longing for cer-
tainty, Edwin Boring declared as late as 1963 that "determinism reigns" (p. 14).
The two debates of Brunswik's program, which were then published in the
Psychological Review in 1943 and 1955, illustrated the way Brunswik's prob-
abilism collided with the leading experimental psychologists' belief in deter-
minism (see Gigerenzer, 1987b). Probabilism was interpreted as a confession
of failure. For instance, Clark Hull (1943) declared in the first debate in Chi-
cago that he and Kurt Lewin believed in uniform laws of behavior that corre-
spond to correlations of 1.00. Because the effort to isolate deterministic laws
is laborious and time-consuming, "all of us may as well give it up, as Brunswik
seems already to have done" (p. 204).
Twelve years later, in the second debate in Berkeley, David Krech (1955)
confronted Brunswik with his personal confession of faith:

I have always made it a cardinal principle to live beyond my income.

And although I have yet to find a one-to-one correlation in psychology
. . . I am always ready to make another promissory note and promise that
if you bear with us we will find uniform laws. . . . And if I can't pay off
on my first promissory note I will come seeking refinancing. . . . I have
faith that despite our repeated and inglorious failures we will someday
come to a theory which is able to give a consistent and complete de-
scription of reality. But in the meantime, I repeat, you must bear with
us. (p. 230)
Refinancing went on for some time. The fixation on uniform laws of behav-
ior was one of the reasons why many of the commentators did not understand
the nature of Brunswik's probabilism—which was located neither in the en-
vironment, as Krech and Hilgard interpreted Brunswik, nor in the organism,
as Hull seemed to do, but rather in the relationship between the organism and
the environment.
The dogma of determinism did not survive Brunswik very long, but the
next two methodological faiths did. They are still entrenched in the minds of
most experimental psychologists—and in their hearts, because these method-
ologies have been taught as if they were moral principles.
48 WHERE DO NEW IDEAS COME FROM?

The Columbia Bible

The Henry Holt publishing company advertised in 1938: "THE BIBLE is OUT."
Robert S. Woodworth had finally published his long-awaited Experimental
Psychology. This textbook, which was known popularly as the "Columbia Bi-
ble," narrowed the many existing practices of experimentation (see Danziger,
1990) to one and only one legitimate form: vary an independent variable (or a
few), hold all the conditions constant, and observe the effect on the dependent
variable. In Brunswik's copy of the Columbia Bible (which Ken Hammond so
kindly lent me), on page 2, the passage "all the conditions constant, except for
one" is underlined twice and Brunswik's pencil notation "imposs[ible]!" is in
the margin. It is not without irony that Brunswik taught courses for years using
Woodworth's textbook, as the notes in his copy indicate. An estimated 100,000
North American psychology majors and graduate students learned what ex-
perimental research is from the bible and its revised edition (Woodworth &
Schlosberg, 1954). The book was translated into many languages and widely
used around the world (Evans, 1990). It was enormously successful; many
psychologists can no longer envision more than one experimental method in
science.
In this book, Woodworth excluded correlation methods and individual dif-
ferences from the domain of experimental psychology. The bible separated the
murky waters of correlation, which obscure the causes of behavior, from the
bright sun of experimentation, where cause and effect can be distinguished
clearly. The result was a strange institutional partition into "two disciplines
of scientific psychology" (Cronbach, 1957), the "Tiny Little Island" of exper-
imental psychology and the "Holy Roman Empire" of correlational psychology.
Brunswik's probabilistic functionalism fit into neither of these disciplines.
His intellectual vision was one of coherence between theory and methodology:
to start with the purpose or function (achievement in natural environments)
and a subject matter (vicarious functioning of perception and judgment) and
to choose a matching methodology (representative design).
There is no such intellectual vision behind the creation of the two "scien-
tific disciplines." Each was, and still is, a historically arbitrary collection of
purpose, subject matter, and method that have no necessary logical or psycho-
logical affinity to each other (Gigerenzer, 1987b). For instance, there is no psy-
chological reason why the study of intelligence is linked with individual dif-
ferences and correlations, whereas the study of thinking is linked to general
laws and experiments. Nor is there a reason why one group should rarely read
or cite the other group's work. Like most ordinary humans who bond with
their peers, psychologists in one camp looked down on their colleagues in the
other camp, declaring their adversaries' methods inferior and their purpose of
little scientific interest and public value. The correlation between psycholo-
gists' esteem for their colleagues in one camp and their colleagues in the other
camp was —.80 (Thorndike, 1954)—alas, a substantial, but not perfect, corre-
lation.
 IDEAS IN EXILE 49

Brunswik found himself and his ideas exiled from his discipline. Ernest
Hilgard (1955), an eminent experimental psychologist, put his lack of regard
for Brunswik's methods in no uncertain terms: "Correlation is an instrument
of the devil" (p. 228). But methods per se are neither good nor bad; the ques-
tion is whether they match a theory or not. Brunswik's intellectual integrity
demanded that he think for himself, deciding what the proper method was,
rather than just climbing on the bandwagon. The tragedy is that he found
himself in a no-man's-land between the two newly established disciplines.

Fisher's Straitjacket
B. F. Skinner once told me that he had thought of dedicating one of his books
to "the statisticians and scientific methodologists with whose help this book
would have never been completed." He had second thoughts, and, in fact,
dedicated the book to those who actually were helpful, "to the pigeon staff."
Skinner had had in mind those statisticians who imposed Sir Ronald Fisher's
doctrine that the design of an experiment must match the statistical method,
such as analysis of variance.
Fisher's randomized control group experiments were tailor-made to Wood-
worth's ideal of experimentation, and analysis of variance allowed one to study
more than one independent variable. Skinner's resistance arose when research-
ers started to use Fisher's method compulsively rather than in a thoughtful
way, that is, as a tool, which is—like all tools—useful only in specific situa-
tions. Editors began to make what they believed was good scientific method a
sine qua non for publication: factorial designs, large numbers of participants,
and small p values.
Statistical thinking became replaced by a mindless ritual performed in the
presence of any set of data (see Chapter 13). Skinner confessed to me that he
once tried a factorial design with some two dozen animals. But only once. He
lost experimental control because he could not keep so many animals at the
same level of deprivation, and the magnitude of error in his data increased.
Why increase error just to have a method that measures error?
The Skinnerians escaped the emerging pressure of editors to publish studies
with large numbers of animals by founding a new journal in 1958, the Journal
of the Experimental Analysis of Behavior. Brunswik, however, had no follow-
ing with which he could found his own journal. Like Skinner, he remarked
drolly that "our ignorance of Fisher's work on factorial design and its mathe-
matical evaluation . . . paid off" (1956, p. 102). As almost all great psycholo-
gists did, he analyzed individuals rather than comparing group means, and he
continued to employ his own nonfactorial representative designs. But he also
sometimes felt that he should make concessions, for instance, when he per-
formed "a routine analysis of variance for the factorially orthodox part of our
experiment" (1956, p. 106).
In Brunswik's struggle with Fisher's ideas, unlike Skinner's, a classic con-
troversy repeated itself. Karl Pearson, who, with Francis Galton, founded cor-
50 WHERE DO NEW IDEAS COME FROM?

relation methods, was involved in a terrible intellectual and personal feud with
Fisher. This fight between these towering statisticians repeated itself in psy-
chology between the proponents of their respective tools. Just at the time when
Brunswik adopted Pearson's correlation methods around 1940, Fisherian meth-
ods began to spread. By 1955, when Brunswik died, Fisherian methods had
overrun, conquered, and redefined every branch of experimental psychology.
Then the newly institutionalized tools evolved into new theories of mind.
When Brunswik's vision of the mind as an intuitive statistician finally became
a great success in experimental psychology, the mind's intuitive statistician
was not of the Karl Pearson school, as Brunswik had imagined. Rather, the
homunculus statistician used the new laboratory tools, such as analysis of
variance. For instance, according to Harold Kelley's (1967) causal attribution
theory, the mind attributes a cause to an effect in the same way as researchers
have come to do—by calculating an intuitive version of analysis of variance
(see Chapter 1). Brunswik had never been able to persuade his colleagues from
experimental psychology that the mind would use the techniques of the com-
peting discipline of correlational psychology.

The Price of Intellectual Integrity

Woodworth's bible had excommunicated Brunswik from experimental psy-
chology, and the institutionalization of Fisher's methods as the sine qua non
of scientific method set Brunswik's ideas outside the realm of what was con-
sidered proper scientific method. Brunswik must have soon realized that the
edifice he had erected had become, as Ken Hammond (1966) expressed it suc-
cinctly, a significant landmark that "was virtually empty; there were visitors,
it is true, but no one stayed" (p. v). Although Brunswik, unlike the exiled
Buhler, had chosen freely to leave Vienna for the United States, he found his
ideas in exile. Unlike in Vienna, at Berkeley he seems not to have had a group
of students who worked on his ideas, nor did his working atmosphere support
the philosophical and interdisciplinary spirit that continued to enhance his
writings. But there was no way back; the Vienna program and the Vienna
Circle had been destroyed, and Brunswik himself had moved beyond multi-
dimensional psychophysics. What is one to do if one has lost the old compan-
ions and failed to enlist new ones? The obvious easy choice would have been
to conform to the new Zeitgeist, but the option of surrendering his ideas seems
never to have occurred to Brunswik. It is easy to be true to one's ideas if
everyone applauds—I admire Brunswik's intellectual integrity because, in his
case, only very, very few applauded. Standing upright must have been difficult,
lonely, and depressing.

Do the Ideas Matter?

American psychology would hardly remember Brunswik's ideas had not one
of his students, Ken Hammond, kept his memory alive to the present day. But
 IDEAS IN EXILE 51

is the memory of Egon Brunswik of more than historical interest? Are his ideas
still exiled, and if so, does it matter?

Representative Sampling

Brunswik (1956) sadly reported that his success in persuading fellow research-
ers to shift to representative sampling of stimuli is "very slow going and hard
to maintain" (p. 39). He complained that his colleagues practiced "double stan-
dards" by being concerned with the sampling of participants but not of stim-
ulus objects. Representative sampling of stimuli is one aspect of the more gen-
eral notion of representative design.
It would be an error to introduce representative sampling as a new dogma
to replace current methodological dogmas. The point is to choose the appro-
priate sampling method for the problem under discussion. For instance, rep-
resentative sampling of objects from a class is indispensable if one wants to
make general statements about the degree of "achievement," or its flip side,
the fallacies of perception and judgment concerning this class of objects. But
if the purpose is testing competing models of cognitive strategies and flat max-
ima obscure the discriminability of strategies, then using selected stimuli that
discriminate between the strategies may be the only choice (see Rieskamp &
Hoffrage, 1999).
Is the idea of representative sampling of any relevance for present-day re-
search? Imagine Brunswik browsing through recent textbooks on cognitive psy-
chology and looking for what we have discovered about achievement in judg-
ment—now more fashionably labeled fallacies and cognitive illusions. It
would catch his eye that the stimuli used in the demonstrations of fallacies
were typically selected rather than representative: the five letters in Tversky
and Kahneman's (1973) study from which the availability heuristic was con-
cluded; the personality sketches in Kahneman and Tversky's (1973) engineer-
lawyer study from which base-rate neglect was concluded; and the general-
knowledge questions from which the overconfidence bias was concluded
(Lichtenstein, Fischhoff, & Phillips, 1982), among others. Brunswik would
have objected that if one wants to measure achievement or demonstrate fal-
lacies in a reference class of objects, one needs to take a representative (or
random) sample of these objects. If not, one can "demonstrate" almost any
level of performance by selecting those objects for which performance is at its
worst (or at its best). In fact, when one uses representative (rather than se-
lected) samples in these three studies, performance greatly improves: The
errors in estimating the frequency of letters largely disappear (Sedlmeier,
Hertwig, & Gigerenzer, 1998); the estimated probabilities that a person is an
engineer approach Bayes's rule (Gigerenzer et al., 1988); and the over-
confidence bias completely disappears (Chapter 7; Juslin, Olsson, & Winman,
1998). These celebrated cognitive illusions, attributed to the participants, are
in part due to the selected sampling done by the experimenters.
These examples illustrate that representative sampling of stimuli is still a
blind spot in some areas of research. In survey research, it would be a mistake
52 WHERE DO NEW IDEAS COME FROM?

to present the odd views of a few selected citizens as public opinion; that the
same applies to stimulus objects is still not commonly acknowledged. Unre-
flectively selected samples can produce apparently general phenomena that
occupy us for years and then finally dissolve into an issue of mere sampling.

Natural Sampling
Imagine Brunswik looking at the studies on Bayesian reasoning, which
emerged about 10 years after his death. When he learned that people neglect
base rates he might have been surprised because his rats did not (Brunswik,
1939). His rats were not perfect, but they were sensitive to the difference of
the base rates of reinforcement in the two sides of a T-maze and to the ratio
as well. Sensitized by the frequentist Reichenbach, Brunswik's eye would have
caught an essential difference between his study and the base-rate studies of
the 1970s and 1980s: His rats learned the base rates from actual experienced
frequencies, whereas the humans in almost all studies that reported base-rate
neglect could not; they were presented summary information in terms of prob-
abilities or percentages. Rats would not understand probabilities, and humans
have only recently in their evolution begun to struggle with this representation
of uncertainty. Does representation matter? Christensen-Szalanski and Beach
(1982) presented base rates in terms of actual frequencies, sequentially en-
countered, and reported that base-rate neglect largely disappeared. This pro-
cess of sampling instances from a population sequentially is known as natural
sampling. Natural sampling is the everyday equivalent—for rats and humans
alike—of the representative sampling done by scientific experimenters. When
observed frequencies are based on natural sampling—that is, on raw (rather
than normalized) counts of events made in an ecological (rather than experi-
mental) setting—then one can show that Bayesian computations become sim-
pler than with probabilities, and people have more insight into Bayesian prob-
lems (Chapter 4).

Structure of Environments
A most important insight I gained from Brunswik's writings is the relevance
of the structure of information in environments to the study of judgment. Brun-
swik tentatively proposed measuring environmental structure by ecological va-
lidities and measuring these in turn by correlation coefficients. Brunswik,
though, almost as much as Skinner, hesitated to look into the black box, and
so he failed to see the important connection between the structure of environ-
ments and that of mediation. Adaptive mental strategies can exploit certain
structures. For instance, if there is a match between the structure of the en-
vironment and that of a strategy, a simple heuristic that processes much less
information than multiple regression can nevertheless make as many (or more)
accurate inferences about its environment (Martignon & Hoffrage, 1999). Her-
bert Simon had emphasized the link between cognitive processes and envi-
ronmental structure in his famous 1956 Psychological Review article on
 IDEAS IN EXILE 53

bounded rationality. However, in recent years, bounded rationality has been

reduced to cognitive limitations, and the structure of environments has been
largely forgotten as an indispensable part of understanding bounded rational-
ity, sometimes even by Simon himself (e.g., 1987). The study of the structure
of environments is still in its infancy.
Much of psychology after the cognitive revolution is about what is in our
heads: Which logic does human reasoning embody? How many primary emo-
tions should we distinguish? It is little concerned with what cognition, emo-
tion, and behavior are for and how they relate to the structure of environments,
both physical and social. Brunswik's focus on achievement, in contrast, is
functional, focusing on the accuracy of perception and judgment. Accuracy is
not the only goal; to be able to act quickly, to come in first, or to establish
social relations of trust and cooperation also exemplify achievement in a
broader sense.
The structure of environments is essential for understanding cognition and
behavior in terms of adaptation, because adaptations are relative to (past) en-
vironments. To flesh out the Darwinian aspect of Brunswikian psychology, one
needs to distinguish between past and present environments, between ecolog-
ical validities in past and in present environments, and between social envi-
ronments composed of conspecifics (where cues are actually signals) and other
environments (e.g., physical environments in which humans do not cooperate
or bargain with their inhabitants). For instance, smooth skin in female humans
may have been a highly valid cue for reproductive capability during most of
human evolution, signaling good health (Buss, 1987). In current environments
with abundant medical technology, the ecological validity of smooth skin may
have decreased to almost nil, but men's proximal mechanisms, cognitive and
emotional, may still rely on such cues. A Darwinian psychology is a historical
psychology, one that looks into the past to learn about the present (e.g., Cos-
mides & Tooby, 1992).
Brunswik repeatedly alluded to Darwin, and the notions of function,
achievement, and environmental structure all relate to evolution by natural
selection. He, however, never developed or carried these ideas any further.
Neo-Brunswikians have done little to develop the Darwinian fragment, con-
sistent with the prevailing anxiety about evolution in the American psychology
establishment. Given that even Pope John Paul II finally announced in the
Quarterly Review of Biology (1997) that evolution (of the body, not of the spirit)
is a plausible hypothesis, more psychologists might find the courage to think
about what we can learn from modern evolutionary theory—even if some still
continue to consider such thoughts politically incorrect.

Models of Vicarious Functioning

Schlick's Thursday evening discussion groups seem to have had a lasting effect
on Brunswik. The methodological objectivity of the Vienna Circle helped
Brunswik to focus his work on the measurement of objective achievement
rather than on cognitive processes ("mediation"). He hesitated to speak about
54 WHERE DO NEW IDEAS COME FROM?

the unobservable process of mediation, and even in 1937 still declared that
psychology is a science of "what" rather than of "how." The question of how
mediation works should be studied only insofar as it throws light on the ques-
tion of what an organism achieves. Only later did Brunswik (e.g., 1957) grant
a place, though only a second place, to the study of cognitive processes.
Given his reluctance to open the black box, I am not sure how Brunswik
would look at the process models of vicarious functions that were inspired by
his ideas: multiple regression models on the one hand (e.g., Hammond,
Hursch, & Todd, 1964) and the theory of probabilistic mental models (PMM
theory) and the fast and frugal lens model on the other (Chapter 8). When
Brunswik coined the metaphor of the "intuitive statistician," he tentatively
suggested that the process of vicarious functioning might be like multiple re-
gression (Doherty & Kurz, 1996). Brunswik's measurement tool turned into a
theory of cognitive processes. In the neo-Brunswikian revival, multiple regres-
sion became the model of vicarious functioning, and, unfortunately, it remains
so. Ken Hammond, like Brunswik, has had second thoughts, but by and large,
the tool has become part of the message. It structures our thinking about Bruns-
wik.
Brunswik's reluctance to think about processes may explain why his ex-
amples for vicarious functioning vacillated back and forth between two differ-
ent processes, substitution and combination. Some of his examples—such as
Hull's habit family hierarchy and the psychoanalytic substitution mechanism
in which one cause can manifest itself as various symptoms—referred to sub-
stitution without combination, others to the combination of cues. The fast and
frugal lens model, based on PMM theory, assumes substitution without com-
bination, emphasizing that judgments need to be made quickly and- on the
basis of limited knowledge (see Gigerenzer & Kurz, in press). Here Egon Bruns-
wik meets Herbert Simon, creating models of bounded rationality in which
simple cognitive heuristics exploit environmental structures.

A Love of History, Philosophy, and Methodology

Just as the human species has a history, so do our theories and methods. Not
knowing where they come from can blind one to understanding why one pro-
pounds a particular theory or uses a specific method. Nevertheless, looking
down at history is symptomatic for much of current psychology. Brunswik had
written about the history of his field and had published in philosophical jour-
nals; possibly it is just that background that helped him to see that there are
differences between methodologies and that one actually needs to make in-
formed choices. Many researchers do not seem to make these choices; rather,
they take on the methodological practice of their field and then defend it as if
it were religious dogma. If one reads Brunswik, one finds a constant stream of
thought about methodology, from preferring matching tasks over numerical
response tasks in order to minimize the confounding of perception with judg-
ment to the larger program of representative design. In contrast, the enthusiasm
with which some methods have been mechanically applied as general-purpose
 IDEAS IN EXILE 55

tools—factor analysis, multidimensional scaling, and analysis of variance,

among others—springs from ignorance of history, philosophy, and the meth-
odologies of other scientific disciplines. Methodology is an art, not a science
of compulsive hand washing.
This is not to say that every psychologist must be a master of history, but
history can protect one against confusing present-day methodological conven-
tions with the sine qua non of scientific research.

The Search for Objectivity in the Twilight of Uncertainty

John Locke (1690/1959) remarked that "God . . . has afforded us only with the
twilight of probability; suitable, I presume, to that state of mediocrity and pro-
bationership he has been pleased to place us in here. . . . " Buhler's psychology
opened the door for Brunswik to the twilight of uncertainty, and the Vienna
Circle inspired him to search for objective knowledge behind that door. What
Brunswik found there: that we know. What he was looking for is more: not
answers, but the right questions. From him, one can learn to rethink that which
is taken for granted. I have.
Yet there is another, deeper message in the work of Egon Brunswik: the
value of the struggle for intellectual integrity—daring to think ideas through,
with all the consequences, and remaining true to them even if they are con-
demned to exile. Kant's final two words in his lovely essay on the Enlighten-
ment capture the essence of this struggle: sapere aude, that is, have the courage
to know.
This page intentionally left blank
 II
ECOLOGICAL RATIONALITY

I he tools-to-theories heuristic generates new ideas about the workings of

the mind by looking outside of the mind. So does the program of "ecological
rationality." Whereas the first draws on the scientist's laboratory environ-
ment, the second draws on people's natural environments, past and present.
Ecological rationality refers to the study of how cognitive strategies exploit
the representation and structure of information in the environment to make
reasonable judgments and decisions. The importance of studying the link be-
tween mind and its environment was emphasized by Egon Brunswik, who
compared mind and environment to two married people who have to come
to terms with each other by mutual adaptation. This couple should not be
divorced, as often is done, in research. More recently, Roger Shepard (1990,
p. 213) expressed the same insight: "We may look into that window [on the
mind] as through a glass darkly, but what we are beginning to discern there
looks very much like a reflection of the world." This is well expressed, but
how can ecological rationality be used as a tool of research?
For instance, a puzzling phenomenon in human judgment is the apparent
neglect of base rates. Even animals do not seem to neglect base rates. Previ-
ous attempts to explain base-rate neglect focused on the shoddy software in
the human mind that might cause this cognitive illusion—could the culprit
be shortcomings in memory, motivation, or computational capabilities? Note
that this approach entails looking exclusively inside the head for an explana-
tion. The program of ecological rationality suggests a different question: In
what environments, past or present, would neglect of base rates be rational?
The answer is, when information is acquired through natural sampling,
which yields simple counts (not normalized by base rates; see the three
chapters in this section). During most of their history—before the advent of
probability theory—humans, like animals, have acquired information about
uncertainties and risks through natural sampling of event frequencies rather
than in terms of probabilities or percentages. The important point is that un-
der conditions of natural sampling, one can make perfectly rational infer-
ences without paying attention to base rates.
58 ECOLOGICAL RATIONALITY

The first two chapters in this section illustrate the practical relevance of
this argument for criminal law, medical diagnosis, AIDS counseling, and
other professions concerned with uncertainties and risks. Should evidence of
wife battering be admissible in the trial of a man accused of murdering his
wife? How many 40-year-old women with a positive mammogram in routine
screening actually have breast cancer? How likely is it that a man with a
positive HIV test actually has the virus? Earlier studies have documented
that many experts—and most patients and jurors—do not understand how to
answer these questions, possibly because they neglect base rates or are con-
fused by probabilities. I show that the notion of ecological rationality leads
to a simple method for helping experts and laypersons alike. One can restore
the representation of uncertainty that humans have encountered throughout
their evolution by translating probabilities back into natural frequencies—the
outcome of natural sampling. This change can turn innumeracy into insight.
The final chapter in this section is theoretical and experimental rather
than applied. It defines the concepts of natural sampling, natural frequen-
cies, and reports experimental evidence for the impact of various external
representations on statistical thinking. The mental strategies or shortcuts
people use, not only their numerical estimates of risks, turn out to be a func-
tion of the external representation of numbers we choose.
Ecological rationality can refer to the adaptation of mental processes to
the representation of information, as in this section. It also can refer to the
adaptation of mental processes to the structure of information in an environ-
ment, as illustrated in the section on bounded rationality and, in more de-
tail, in Simple Heuristics that Make Us Smart (Gigerenzer, Todd, & the ABC
Research Group, 1999). In both cases, it is important to distinguish between
past and present environments, particularly when we are studying humans,
who change their environments rapidly. Studying how past environments
differ from present environments reminds us that an ecological perspective
has an evolutionary and historical dimension. Here we go beyond Brun-
swik's metaphor of the married couple, which focuses on the adaptation be-
tween the mind and its current spouse while forgetting its previous mar-
riages. The program of ecological rationality is a research heuristic, not a
foolproof recipe—just as new laboratory tools do not always lead to good
theories for mental processes.
 4

Ecological Intelligence

When I left a restaurant in a charming town in Tuscany one night, I looked

for my yellow-green rented Renault 4 in the parking lot. There was none.
Instead, there was a blue Renault 4 sitting in the lot, the same model but
the wrong color. I still feel my fingers hesitating to put my key into the lock
of this car, but the lock opened. I drove the car home. When I looked out
the window the next morning, there was a yellow-green Renault 4 standing
in bright sunlight outside. What had happened? My color-constancy system
did not work with the artificial light at the parking lot. Color constancy, an
impressive adaptation of the human perceptual system, normally allows us
to see the same color under changing illuminations, under the bluish light
of day as well as the reddish light of the setting sun. Color constancy, however,
fails under certain artificial lights, such as sodium or mercury vapor lamps,
which were not present in the environment when mammals evolved (Shepard,
1992).
Human color vision is adapted to the spectral properties of natural sunlight.
More generally, our perceptual system has been shaped by the environment in
which our ancestors evolved, the environment often referred to as the "envi-
ronment of evolutionary adaptiveness," or EEA (Tooby & Cosmides, 1992).
Similarly, human morphology, physiology, and the nervous and immune sys-
tems show exquisite adaptations. The tubular form of the bones maximizes
strength and flexibility while minimizing weight; bones are, pound for pound,
stronger than solid steel bars, and the best man-made heart valves cannot yet
match the way natural valves open and close (Nesse & Williams, 1995). Like
color constancy, however, these systems can be fooled and may break down
when stable, long-term properties of the environment to which they were
adapted change.
In this chapter, I propose that human reasoning processes, like those of
color constancy, are designed for information that comes in a format that
was present in the EEA. I will focus on a class of inductive reasoning pro-
cesses technically known as Bayesian inference, specifically a simple version

59
60 ECOLOGICAL RATIONALITY

thereof in which an organism infers from one or a few indicators which of

two events is true.

Bayesian Inference

David Eddy (1982) asked physicians to estimate the probability that a woman
has breast cancer given that she has a positive mammogram on the basis of the
following information:
The probability that a patient has breast cancer is 1% (the physician's
prior probability).
If the patient has breast cancer, the probability that the radiologist will
correctly diagnose it is 80% (sensitivity or hit rate).
If the patient has a benign lesion (no breast cancer), the probability that
the radiologist will incorrectly diagnose it as cancer is 9.6% (false pos-
itive rate).
QUESTION: What is the probability that a patient with a positive mam-
mogram actually has breast cancer?
Eddy reported that 95 out of 100 physicians estimated the probability of
breast cancer after a positive mammogram to be about 75%. The inference from
an observation (positive test) to a disease, or more generally, from data D to a
hypothesis H, is often referred to as "Bayesian inference," because it can be
modeled by Bayes's rule:

Equation 1 shows how the probability p(H I D) that the woman has breast
cancer (H) after a positive mammogram (D) is computed from the prior prob-
ability p(H) that the patient has breast cancer, the sensitivity p(D\H), and the
false positive rate p(D I -H] of the mammography test. The probability p(H\ D)
is called the "posterior probability." The symbol —H stands for "the patient
does not have breast cancer." Equation 1 is Bayes's rule for binary hypotheses
and data. The rule is named after Thomas Bayes (1702 [?]-1761), an English
dissenting minister, to whom this solution of the problem of how to make an
inference from data to hypothesis (the so-called inverse problem; see Daston,
1988) is attributed.1 The important point is that Equation 1 results in a prob-
ability of 7.8%, not 75% as estimated by the majority of physicians. In other

1. As we know from Stephen M. Stigler's Law of Eponymy, no scientific discovery

is named after its original discoverer, and Bayes's rule seems to be no exception to this
law (Stigler, 1983).
 ECOLOGICAL INTELLIGENCE 61

words, the probability that the woman has breast cancer is one order of mag-
nitude smaller than estimated.
This result, together with an avalanche of studies reporting that laypeople's
reasoning does not follow Bayes's rule either, has (mis-)led many to believe
that Homo sapiens would be inept to reason the Bayesian way. Listen to some
influential voices: "In his evaluation of evidence, man is apparently not a con-
servative Bayesian: he is not Bayesian at all" (Kahneman & Tversky, 1972,
p. 450). "Tversky and Kahneman argue, correctly, I think, that our minds are
not built (for whatever reason) to work by the rules of probability" (Gould,
1992, p. 469).2 The literature of the last 25 years has reiterated again and again
the message that people are bad reasoners, neglect base rates most of the time,
neglect false positive rates, and are unable to integrate base rate, hit rate, and
false positive rate the Bayesian way (for a review see Koehler, 1996). Proba-
bility problems such as the mammography problem have become the stock-in-
trade of textbooks, lectures, and party entertainment. It is guaranteed fun to
point out how dumb others are. And aren't they? There seem to be many
customers eager to buy the message of "inevitable illusions" wired into our
brains (Piattelli-Palmarini, 1994).

Ecological Bayesian Inference: An Adaptation for Frequencies

Back to color constancy. If a human visual system enters an environment il-

luminated by sodium vapor lamps, its color-constancy algorithms will fail.
This does not mean, however, that human minds are not built to work by color-
constancy algorithms. Similarly, if a human reasoning system enters an envi-
ronment in which statistical information is formatted differently from that en-
countered in the environment in which humans evolved, the reasoning
algorithms may fail. But this does not imply that human minds are not built
to reason the Bayesian way. The issue is not whether nature has equipped our
minds with good or with bad statistical software, as the "optimists" versus
"pessimists" discussion about human rationality suggests (Jungermann, 1983).
The issue I address here is the adaptation of mental algorithms to their envi-
ronment. By "mental algorithms," I mean induction mechanisms that perform
classification, estimation, or other forms of uncertain inferences, such as de-
ciding what color an object is or inferring whether a person has a disease.
For which information formats have mental algorithms been designed?
What matters for an algorithm that makes inductive inferences is the format of
numerical information. Eddy presented information (about the prevalence of
breast cancer, the sensitivity, and the false positive rate of the test) in terms of
probabilities and percentages, just as most experimenters did who found hu-

2. For a critical discussion of these interpretations, see Cohen (1981), Gigerenzer

(1994a, 1996a), Gigerenzer and Murray (1987, chap. 5), and Lopes (1991); for a reply,
see Kahneman and Tversky (1996).
62 ECOLOGICAL RATIONALITY

mans making irrational judgments. What was the format of the numerical in-
formation humans encountered during their evolution? We know too little
about these environments, for instance, about the historically normal condi-
tions of childbirth, or how strong a factor religious doctrines were, and most
likely, these varied considerably between societies. But concerning the format
of numerical information, I believe we can be as certain as we ever can be—
probabilities and percentages were not the way organisms encountered infor-
mation. Probabilities and percentages are quite recent forms of representations
of uncertainty. Mathematical probability emerged in the mid-seventeenth cen-
tury (Hacking, 1975), and the concept of probability itself did not gain prom-
inence over the primitive notion of "expectation" before the mid-eighteenth
century (Daston, 1988). Percentages became common notations only during the
nineteenth century, after the metric system was introduced during the French
Revolution (mainly, though, for interest and taxes rather than for representing
uncertainty). Only in the second half of the twentieth century did probabilities
and percentages become entrenched in the everyday language of Western coun-
tries as representations of uncertainty. To summarize, probabilities and per-
centages took millennia of literacy and numeracy to evolve as a format to
represent degrees of uncertainty. In what format did humans acquire numerical
information before that time?
I propose that the original format was natural frequencies, acquired by nat-
ural sampling. Let me explain what this means by a parallel to the mammog-
raphy problem, using the same numbers. Think about a physician in an illit-
erate society. Her people have been afflicted by a new, severe disease. She has
no books nor statistical surveys; she must rely solely on her experience. For-
tunately, she discovered a symptom that signals the disease, although not with
certainty. In her lifetime, she has seen 1,000 people, 10 of whom had the dis-
ease. Of those 10, eight showed the symptom; of the 990 not afflicted, 95 did.
Thus there were 8 + 95 = 103 people who showed the symptom, and only 8
of these had the disease. Now a new patient appears. He has the symptom.
What is the probability that he actually has the disease?
The physician in the illiterate society does not need a pocket calculator to
estimate the Bayesian posterior probability. All she needs to do is to keep track
of the number of symptom and disease cases (8) and the number of symptom
and no-disease cases (95). The probability that the new patient actually has
the disease can be "seen" easily from these frequencies:

Equation 2 is Bayes's rule for natural frequencies, in which a is the number

of cases with symptom and disease and b is the number of cases having the
symptom but lacking the disease. The chance that the new patient has the
disease is less than 8 out of 100, or 8%. Our physician who learns from ex-
perience cannot be fooled as easily into believing that the chances are about
75%, as many of her contemporary colleagues did.
 ECOLOGICAL INTELLIGENCE 63

The comparison between Equations 1 and 2 reveals an important theoretical

result: Bayesian reasoning is computationally simpler (in terms of the number
of operations performed, such as additions and multiplications) when the in-
formation is in natural frequencies (Equation 2) rather than in probabilities
(Equation 1) (see Kleiter, 1994). As Equation 2 shows, the base rates of event
frequencies (such as 10 in 1,000) need not be kept in memory; they can be
ignored as they are implicit in the frequencies a and b.
Let me be clear how the terms "natural sampling" and "natural frequencies"
relate. Natural sampling is the sequential process of updating event frequencies
from experience. A foraging organism who, day after day, samples potential
resources for food and learns the frequencies with which a cue (e.g., the pres-
ence of other species) predicts food performs natural sampling by updating the
frequencies a and b from observation to observation. Natural sampling is dif-
ferent from systematic experimentation, in which the sample sizes (the base
rates) of each treatment group are fixed in advance. For instance, in a clinical
experiment, one might select 100 patients with cancer and 100 without cancer
and then perform tests on these groups. By fixing the base rates, the frequencies
obtained in such experimental designs no longer carry information about the
base rates. This is not to say that controlled sampling in systematic experi-
ments is useless; it just serves a different purpose. Brunswik's (1955) method
of "representative sampling" in a natural environment is an example of ap-
plying the idea of natural sampling to experimental design.
Natural frequencies report the final tally of a natural sampling process.
There is more than one way to present the final tally. In the case of the phy-
sician in the illiterate society, I specified the total number of observations
(1,000), the frequency of the disease, and the frequencies a and b of hits and
false positives, respectively: "In her lifetime, she has seen 1,000 people, 10 of
whom had the disease. Of those 10, eight showed the symptom; of the 990 not
afflicted, 95 did." This is a straightforward translation of the base rates, hit
rates, and false positive rates into natural frequencies. Alternatively, one can
communicate the frequencies a and b alone: "In her lifetime, she has seen 8
people with symptom and disease, and 95 people with symptom and no dis-
ease." The former natural frequencies use a standard menu ("standard" be-
cause slicing up the information in terms of base rate, hit rate, and false pos-
itive rate is deeply entrenched today), the latter use a short menu (see Chapter
6). Both lead to the same result.
Natural frequencies must not to be confused with a representation in terms
of relative frequencies (e.g., a base rate of .01, a hit rate of .80, and a false
positive rate of .096). Relative frequencies are, like probabilities and percent-
ages, normalized numbers that no longer carry information about the natural
base rates. Relative frequencies, probabilities, and percentages are to human
reasoning algorithms (that do Bayesian-type inference) as sodium vapor lamps
are to human color-constancy algorithms. This analogy has, like every analogy,
its limits. For instance, humans can be taught, although with some mental
agony, to reason by probabilities, but not, I believe, to maintain color constancy
under sodium vapor illumination.
64 ECOLOGICAL RATIONALITY

Note that the total number of observations—communicated only when nat-

ural frequencies are expressed in the standard menu—need not be the actual
total number of observations. It can be any convenient number such as 100 or
1,000. The computational simplicity of natural frequencies holds indepen-
dently of whether the actual or a convenient number is used. For example, if
the actual sample size was 5,167 patients, one can nevertheless represent the
information in the same way as above. "For every 1,000 patients we expect 10
who have breast cancer, and 8 out of these 10 will test positive."3
The hypothesis that mental algorithms were designed for natural frequen-
cies is consistent with (a) a body of studies that report that humans can monitor
frequencies fairly accurately (Barsalou & Ross, 1986; Hintzman & Block, 1972;
Jonides & Jones, 1992), (b) the thesis that humans process frequencies (almost)
automatically, that is, without or with little effort, awareness, and interference
with other processes (Hasher & Zacks, 1984), (c) the thesis that probability
learning and transfer derive from frequency learning (Estes, 1976), and (d) de-
velopmental studies on counting in children and animals (e.g., Gallistel & Gel-
man, 1992). This is not to say that humans and animals count all possible
events equally well, nor could they. A conceptual mechanism must first decide
what the units of observation are so that a frequency encoding mechanism can
count them. This preceding conceptual process is not dealt with by the hy-
pothesis that mental algorithms are designed for natural frequencies (but see
the connection proposed by Erase, Cosmides, & Tooby, 1998).
Thus my argument has two parts: evolutionary (and developmental) pri-
macy of natural frequencies and ease of computation. First, mental algorithms,
from color constancy to inductive reasoning, have evolved in an environment
with fairly stable characteristics. If there are mental algorithms that perform
Bayesian-type inferences from data to hypotheses, these are designed for nat-
ural frequencies acquired by natural sampling, and not for probabilities or
percentages. Second, when numerical information is represented in natural
frequencies, Bayesian computations reduce themselves to a minimum. Both
parts of the argument are necessary. For instance, the computational part could
be countered by hypothesizing that there might be a single neuron in the hu-
man mind that almost instantaneously computes Equation 1 on the basis of
probability information. The evolutionary part of the argument makes it un-
likely that such a neuron has evolved that computes using an information
format that was not present in the environment in which our ancestors
evolved.
This argument has testable consequences. First, laypeople—that is, persons
with no professional expertise in diagnostic inference—are more likely to rea-
son the Bayesian way when the information is presented in natural frequencies
than in a probability format. This effect should occur without any instruction
in Bayesian inference. Second, experts such as physicians who make diagnos-

3. However, there is a price to be paid if one replaces the actual with a convenient
sample size. One can no longer compute second-order probabilities (Kleiter, 1994).
 ECOLOGICAL INTELLIGENCE 65

tic inferences on a daily basis should, despite their experience, show the same
effect. Third, the "inevitable illusions" (Piattelli-Palmarini, 1994), such as
base-rate neglect, should become evitable by using natural frequencies. Finally,
natural frequencies should provide a superior vehicle for teaching Bayesian
inference. In what follows, I report tests of these predictions and several ex-
amples drawn from a broad variety of everyday situations.
This is not to say that probabilities are useless or perverse. In mathematics,
they play their role independent of whether or not they suit human reasoning,
just as Riemannian and other non-Euclidean geometries play their roles in-
dependent of the fact that human spatial reasoning is Euclidean.

Breast Cancer

Eddy (1982) provides only a scant, one-paragraph description of his study of

physicians' institutions and refers to a study by Casscells, Schoenberger, and
Grayboys (1978) that showed similar results. Both studies used a probability
format. Would natural frequencies make any difference to experts such as
physicians? Ulrich Hoffrage and I tested 48 physicians in Munich, Germany,
on the mammography problem. These physicians had an average professional
experience of 14 years. Twenty-four physicians read the information in a
probability format as in Eddy's study, the other 24 read the same information
in natural frequencies. Physicians were always asked for a single-event prob-
ability (as in Eddy's study) when the information was in probabilities; they
were always asked for a frequency judgment when the information was in
natural frequencies. The two formats of the mammography problem are
shown in Table 4.1. Each physician got four diagnostic problems (including
the mammography problem), two in a probability format and two in natural
frequencies (the details are in Gigerenzer, 1996b; Hoffrage & Gigerenzer,
1996, 1998).
Consider first the thinking of one typical physician, a 59-year-old director
of a university clinic, whom I call Dr. Average. He spent 30 minutes on the
four problems and another 15 minutes discussing the results with the inter-
viewer. As a high-status physician, he was not used to having his diagnostic
intuitions being tested, and he became visibly nervous. He first got the mam-
mography problem in the probability format (Table 4.1) and commented, "I
never inform my patients about statistical data. I would tell the patient that
mammography is not so exact, and I would, in any case, perform a biopsy."
He estimated the probability of breast cancer after a positive mammogram as
80% + 10% = 90%. That is, he added the sensitivity to the false positive rate
(this is an unusual strategy). Nervously, he remarked: "Oh, what nonsense. I
can't do it. You should test my daughter, she studies medicine." Dr. Average
was as helpless with the second problem, Bechterev's disease, in a probability
format. Here he estimated the posterior probability by multiplying the base
rate by the sensitivity, a common strategy used by laypeople ("joint occur-
rence," see Tables 6.3 and 6.4 in Chapter 6).
66 ECOLOGICAL RATIONALITY

When Dr. Average saw the first problem in a frequency format, his ner-
vousness subsided. "That's so easy," he remarked with relief, and came up
with the Bayesian answer, as he did with the second problem in a frequency
format. Dr. Average's reasoning turned Bayesian the moment the information
was in frequencies, despite his never having heard of, or at least not remem-
bering, Bayes's rule. In the words of a 38-year-old gynecologist faced with the
mammography problem in a frequency format: "A first grader could do that.
Wow, if someone can't solve this . . . !"
Consider now all the physicians' diagnostic inferences concerning breast
cancer. Do natural frequencies foster insight in them?
In the probability format, only 2 out of 24 physicians (8%) came up with
the Bayesian answer. The median estimate of the probability of breast cancer
after a positive mammogram was 70%, consistent with Eddy's findings. With
natural frequencies, however, 11 out of 24 physicians (46%) responded with
the Bayesian answer. Across all four diagnostic problems, similar results were
obtained—10% Bayesian responses in the probability format and 46% with
natural frequencies (Figure 4.1).

Table 4.1 The mammography problem: Probability format and

natural frequencies
To facilitate early detection of breast cancer, women are encouraged from a particular
age on to participate at regular intervals in routine screening, even if they have no
obvious symptoms. Imagine you use mammography to conduct such a breast cancer
screening in a certain region. For symptom-free women age 40 to 50 who participate
in screening using mammography, the following information is available for this re-
gion:

Probability format
The probability that one of these women has breast cancer is 1 %.
If a woman has breast cancer, the probability is 80% that she will have a positive
mammogram.
If a woman does not have breast cancer, the probability is 10% that she will still have
a positive mammogram.
Imagine a woman (age 40 to 50, no symptoms) who has a positive mammogram in
your breast cancer screening. What is the probability that she actually has breast can-
cer? %

Natural frequencies
Ten out of every 1,000 women have breast cancer.
Of these 10 women with breast cancer, 8 will have a positive mammogram.
Of the remaining 990 women without breast cancer, 99 will still have a positive mam-
mogram.
Imagine a sample of women (age 40 to 50, no symptoms) who have positive mammo-
grams in your breast cancer screening. How many of these women do actually have
breast cancer? out of
 ECOLOGICAL INTELLIGENCE 67

Figure 4.1 How to foster diagnostic insight in physicians. Each of 48 physi-

cians got information about four standard screening tests, two in a probabil-
ity format and two in natural frequencies. Table 4.1 shows the two forms of
information representation for mammography screening. The four diagnostic
tasks were to infer the presence of (1) breast cancer from a positive mammo-
gram, (2) colorectal cancer from a positive hemoccult test, (3) Bechterev's
disease (ankylosing spondylitis) from a positive HL-antigen-B27 test, and
(4) phenylketonuria from a positive Guthrie test. In each diagnostic task,
the physicians reasoned more often consistent with Bayes's rule when the
numerical information was in natural frequencies. Probabilities tended to
cloud their minds.

The lesson of these results is not to blame physicians' or laypeople's minds

when they stumble over probabilities. Rather, the lesson is to represent infor-
mation in textbooks, in curricula, and in physician—patient interactions in nat-
ural frequencies that correspond to the way information was encountered in
the environment in which human minds evolved.

Colorectal Cancer

The hemoccult test is a widely used and well-known test for colorectal cancer.
Windeler and Kobberling (1986) report that just as physicians overestimated
the (posterior) probability that a patient has colorectal cancer if the hemoccult
test is positive, they also overestimated the base rate of colorectal cancer, the
sensitivity (hit rate), and the false positive rate of the test. Windeler and Kob-
berling asked these physicians about probabilities and percentages. Would nat-
ural frequencies improve physicians' estimates of what a positive test tells
about the presence of colorectal cancer? The 48 physicians in the study re-
ported previously were given the best available estimates for the base rate,
sensitivity, and false positive rate, as published in Windeler and Kobberling
68 ECOLOGICAL RATIONALITY

(1986). The following is a shortened version of the full text (structured like
the mammography problem in Table 4.1) given to the physicians. In the prob-
ability format, the information was:
The probability that a person has colorectal cancer is 0.3%.
If a person has colorectal cancer, the probability that the test is positive
is 50%.
If a person does not have colorectal cancer, the probability that the test
is positive is 3%.
What is the probability that a person who tests positive actually has
colorectal cancer?
When one inserts these values in Bayes's rule (Equation 1), the resulting
probability is 4.8%. In natural frequencies, the information was:
30 out of every 10,000 people have colorectal cancer.
Of these 30 people with colorectal cancer, 15 will test positive.
Of the remaining 9,970 people without colorectal cancer, 300 will still
test positive.
Imagine a group of people who test positive. How many of these will
actually have colorectal cancer?
When the information was in the probability format, only 1 out of 24 phy-
sicians (4%) could find the Bayesian answer, or anything close to it. The me-
dian estimate was one order of magnitude higher, namely 47%. When the
information was presented in natural frequencies, 16 out of 24 physicians
(67%) came up with the Bayesian answer (details are in Gigerenzer, 1996b;
Hoffrage & Gigerenzer, 1998).

Wife Battering

Alan Dershowitz, the Harvard law professor who advised the defense in the
first O. J. Simpson trial, claimed repeatedly that evidence of abuse and batter-
ing should not be admissible in a murder trial. In his best-seller, Reasonable
Doubts: The Criminal Justice System and the O. J. Simpson Case (1996), Der-
showitz says: "The reality is that a majority of women who are killed are killed
by men with whom they have a relationship, regardless of whether their men
previously battered them. Battery, as such, is not a good independent predictor
of murder" (p. 105). Dershowitz stated on U.S. television in March 1995 that
only about one-tenth of 1% of wife batterers actually murder their wives. In
response to Dershowitz, I. J. Good, a distinguished professor emeritus of sta-
tistics at the Virginia Polytechnic Institute, published an article in Nature to
correct for the possible misunderstandings of what that statement implies for
the probability that O. J. Simpson actually murdered his wife in 1994 (Good,
1995). Good's argument is that the relevant probability is not the probability
 ECOLOGICAL INTELLIGENCE 69

that a husband murders his wife if he batters her. Instead, the relevant prob-
ability is the probability that a husband has murdered his wife if he battered
her and if she was actually murdered by someone. More precisely, the relevant
probability is not p(G\Bai] but p(G\Bat and M), in which G stands for "the
husband is guilty" (that is, did the murder in 1994), Bat means that "the hus-
band battered his wife," and M means that "the wife was actually murdered
by somebody in 1994."
My point concerns the way Good presents his argument, not the argument
itself. Good presented the information in single-event probabilities and odds
(rather than in natural frequencies). I will first summarize Good's argument as
he made it. I hope I can demonstrate that you the reader—unless you are a
trained statistician or exceptionally smart with probabilities—will be confused
and have some difficulty following it. Thereafter, I will present the same ar-
gument in natural frequencies, and confusion should turn into insight. Let's
see.

Good's Argument in Conditional Probabilities

Good bases his calculations of p(G I Bat and M) on the odds version of Bayes's
rule:

posterior odds = prior odds X likelihood ratio

which in the present case is:

p(G\Batand M) _ p(GI Bat) p(M\ G and Bat)

p(-G\Bat and M) ~ p(-G\Baf] * p(M\ -G and Bat)
where — G stands for "the husband is not guilty."
The following six equations (Good-1 to Good-6) show Good's method of
explaining to the reader how to estimate p(G\Bat and M). Good starts with
Dershowitz's figure of one-tenth of 1%, arguing that if the husband commits
the murder, the probability is at least 1/10 that he will do it in 1994:4

Therefore, the prior odds (O) are:

Furthermore, the probability of a woman being murdered given that her

husband has murdered her (whether he is a batterer or not) is unity:

4. Good possibly assumed that the average wife batterer is married less than 10 years.
Good also made a second calculation assuming a value of p(G\Bat) that is half as large.
70 ECOLOGICAL RATIONALITY

Because there are about 25,000 murders per year in the U.S. population of
about 250,000,000, Good estimates the probability of a battered woman being
murdered, but not by her husband, as:

From Equations Good-3 and Good-4, it follows that the likelihood ratio is about
10,000/1; therefore, the posterior odds can be calculated:

That is, the probability that a murdered, battered wife was killed by her hus-
band is:

Good's point is that "most members of a jury or of the public, not being
familiar with elementary probability, would readily confuse this with
P(G\Bat), and would thus be badly misled by Dershowitz's comment" (Good,
1995, p. 541). He adds that he sent a copy of this note to both Dershowitz and
the Los Angeles Police Department, reminding us that Bayesian reasoning
should be taught at the precollege level.
Good's persuasive argument, I believe, could have been understood more
easily by his readers and the Los Angeles Police Department if the information
had been presented in natural frequencies rather than in the single-event prob-
abilities and odds in the six equations. As with breast cancer and colorectal
cancer, one way to represent information in natural frequencies is to start with
a concrete sample of individuals divided into subclasses, in the same way it
would be experienced by natural sampling. Here is a frequency version of
Good's argument.

Good's Argument in Natural Frequencies

Think of 10,000 battered married women. Within one year, at least one will
be murdered by her husband. Of the remaining 9,999 who are not killed by
their husbands, one will be murdered by someone else. Thus we expect at
least two battered women to be murdered, at least one by her husband and
one by someone else. Therefore, the probability p(G\Bat and M) that a mur-
dered, battered woman was killed by her husband is at least Vz.
This probability is not to be confounded with the probability that O. J.
Simpson is guilty; a jury must take into account much more evidence than
battering. But the probability shows that abuse-and-battering is a good predic-
tor of the husband's (or boyfriend's) guilt, disproving Dershowitz's assertion to
the contrary.
In natural frequencies, Good's argument is short and transparent. My con-
jecture is that more ordinary people, including employees of the Los Angeles
Police Department and jurors, could understand and communicate the argu-
ment if the information were represented in natural frequencies rather than in
probabilities or odds.
 ECOLOGICAL INTELLIGENCE 71

In legal jargon, evidence of wife battering is probative, not prejudicial. This

analysis is consistent with the impressive transcultural evidence about homi-
cide accumulated by Daly and Wilson (1988). The typical function of wife
battering seems to be to exert proprietary rights over the sexuality and repro-
ductivity of women, as well as threats against infidelity. Battering can "spill
over" into killing, and killing is the tip of a huge iceberg of wife abuse.

AIDS Counseling

Under the headline, "A False HIV Test Caused 18 Months of Hell," the Chicago
Tribune (3/5/93) published the following letter and response:
Dear Ann Landers: In March 1991,1 went to an anonymous testing center
for a routine HIV test. In two weeks, the results came back positive.
I was devastated. I was 20 years old and doomed. I became severely
depressed and contemplated a variety of ways to commit suicide. After
encouragement from family and friends, I decided to fight back.
My doctors in Dallas told me that California had the best care for HIV
patients, so I packed everything and headed west. It took three months
to find a doctor I trusted. Before this physician would treat me, he in-
sisted on running more tests. Imagine my shock when the new results
came back negative. The doctor tested me again, and the results were
clearly negative.
I'm grateful to be healthy, but the 18 months I thought I had the virus
changed my life forever. I'm begging doctors to be more careful. I also
want to tell your readers to be sure and get a second opinion. I will
continue to be tested for HIV every six months, but I am no longer ter-
rified.
David in Dallas

Dear Dallas: Yours is truly a nightmare with a happy ending, but don't
blame the doctor. It's the lab that needs to shape up. The moral of your
story is this: Get a second opinion. And a third. Never trust a single test.
Ever.
Ann Landers
David does not mention what his Dallas doctors told him about the chances
that he actually had the virus after the positive test, but he seems to have
inferred that a positive test meant that he had the virus, period. In fact, when
we studied AIDS counselors in Germany, we found that many doctors and
social workers (erroneously) tell their low-risk clients that a positive HIV test
implies that the virus is present (see Chapter 5). These counselors know that
a single ELISA (enzyme-linked immunoabsorbent assay) test can produce a
false positive, but they erroneously assume that the whole series of ELISA and
Western blot tests would wipe out every false positive. How could a doctor
have explained the actual risk to David and spared him the nightmare?
I do not have HIV statistics for Dallas, so I will use German figures for
illustration. (The specific numbers are not the point here.) In Germany, the
72 ECOLOGICAL RATIONALITY

prevalence of HIV infections in heterosexual men between the ages of 20 and

30 who belong to no known risk group can be estimated as about 1 in 10,000,
or 0.01%. The corresponding base rate for homosexual men is estimated at
about 1.5%. The hit rate (sensitivity) of the typical test series (repeated ELISA
and Western blot tests) is estimated at about 99.8%. The estimates of the false
positive rate vary somewhat; a reasonable estimate seems to be 0.01%. Given
these values, and assuming that David was at the time of the routine HIV test
a heterosexual man with low-risk behavior, what is the probability that he
actually had the virus after testing positive? If his physician had actually given
David these probabilities, David nevertheless might not have understood what
to conclude.
But the physician could have communicated the information in natural fre-
quencies. She might have said, "Your situation is the following: Think of
10,000 heterosexual men like you. We expect one to be infected with the virus,
and he will, with practical certainty, test positive. From the 9,999 men who
are not infected, one additional individual will test positive. Thus we get two
individuals who test positive, but only one of them actually has the virus. This
is your situation. The chances that you actually have the virus after the positive
test are about 1 in 2, or 50%." If the physician had explained the risk in this
way, David might have understood that there was, as yet, no reason to contem-
plate suicide or to move to California.
We do not know what risk group David was in. Whatever the statistics are,
however, most people of average intelligence can understand the risk of HIV
after a positive test when the numbers are represented by a counselor in nat-
ural frequencies.
Ann Landers's answer—don't blame the doctor, blame the lab—however,
overlooks the fact that despite whatever possible reasons there may be for false
positives (such as the presence of cross-reacting antibodies or blood samples
being confused in the lab), a doctor should inform the patient that false posi-
tives occur, and about how frequently they occur. What information do pro-
fessional AIDS counselors actually provide? How do they communicate risks?
To find this out, a brave student of mine went as a client to 20 public health
centers to take 20 HIV tests. The results will be reported in Chapter 5.

Expert Witnesses

Evidentiary problems such as the evaluation of eyewitness testimony consti-

tuted one of the first domains of probability theory (Gigerenzer et al., 1989,
chap. 1). Statisticians have taken the stand as expert witnesses for almost a
century now: In the Dreyfus case in the late nineteenth century in France, or
more recently, in People vs. Collins in California (Koehler, 1992). The convic-
tions in both cases were ultimately reversed and the statistical arguments dis-
credited. Part of the problem seems to have been that the statistical arguments
were couched not in natural frequencies but in probabilities that confused both
the prosecution who were making the arguments and the jury and the judges
 ECOLOGICAL INTELLIGENCE 73

who tried to understand the arguments. I will explain this point with the case
of a chimney sweep who was accused of having committed a murder in Wup-
pertal, Germany (Schrage, n.d.).
The Rheinischer Merkur (No. 39, 1974) reported:
On the evening of July 20, 1972, the 40-year-old Wuppertal painter Wil-
helm Fink and his 37-year-old wife Ingeborg took a walk in the woods
and were attacked by a stranger. The husband was hit by three bullets
in the throat and the chest, and fell down. Then the stranger attempted
to rape his wife. When she defended herself and, unexpectedly, the shot-
down husband got back on his feet to help her, the stranger shot two
bullets into the wife's head and fled.
Three days later, 20 kilometers from the scene of the crime, a forest ranger
discovered the car of Werner Wiegand, a 25-year-old chimney sweep who used
to spend his weekends in the vicinity. The husband, who had survived, at first
thought he recognized the chimney sweep in a photo. Later, he grew less cer-
tain and began to think that another suspect was the murderer. When the other
suspect was found innocent, however, the prosecution came back to the chim-
ney sweep and put him on trial. The chimney sweep had no previous convic-
tions and denied being the murderer. The Rheinischer Merkur described the
trial:
After the experts had testified and explained their "probability theories,"
the case seemed to be clear: Wiegand, despite his denial, must have been
the murderer. Dr. Christian Rittner, a lecturer at the University of Bonn,
evaluated the traces of blood as follows: 17.29% of German citizens share
Wiegand's blood group, traces of which have been found underneath the
fingernails of the murdered woman; 15.69% of German share [her] blood
group that was also found on Wiegand's boots; based on a so-called
"cross-combination" the expert subsequently calculated an overall prob-
ability of 97.3% that Wiegand "can be considered the murderer." And
concerning the textile fiber traces which were found both on Wiegand's
clothes and on those of the victim. . . . Dr. Ernst Rohm from the Munich
branch of the State Crime Department explained: "The probability that
textile microfibers of this kind are transmitted from a human to another
human who was not in contact with the victim is at most 0.06%. From
this results a 99.94% certainty for Wiegand being the murderer."
Both expert witnesses agreed that, with a high probability, the chimney
sweep was the murderer. These expert calculations, however, collapsed when
the court discovered that the defendant was in his hometown, 100 kilometers
away from the scene of the crime at the time of the crime.
So what was wrong with the expert calculations? One can dispel the con-
fusion in court by representing the uncertainties in natural frequencies. Let
us assume that the blood underneath the fingernails of the victim was indeed
the blood of the murderer, that the murderer carried traces of the victim's
blood (as the expert witnesses assumed), and that there were 10 million men
in Germany who could have committed the crime (these and the following
74 ECOLOGICAL RATIONALITY

figures are from Schrage, n.d., but the specific figures do not matter for my
argument). Let us assume further that on one of every 100 of these men a close
examination would find microscopic traces of foreign blood, that is, on 100,000
men. Of these, some 15,690 men (15.69%) will carry traces from blood that
is of the victim's blood type. Of these 15,690 men, some 2,710 (17.29%) will
also have the blood type that was found underneath the victim's fingernails
(here, I assume independence between the two pieces of evidence). Thus there
are some 2,710 men (including the murderer) who might appear guilty based
on the two pieces of blood evidence. The chimney sweep is one of these men.
Therefore, given the two pieces of blood evidence, the probability that the
chimney sweep is the murderer is about 1 in 2,710, and not 97.3%, as the
first expert witness testified.
The same frequency method can be applied to the textile traces. Let us
assume that the second expert witness was correct when he said that the prob-
ability of the chimney sweep carrying the textile trace, if he were not the
murderer, would be at most 0.06%. Let us assume as well that the murderer
actually carries that trace. Then some 6,000 of the 10 million would carry this
textile trace, and only one of them would be the murderer. Thus the probability
that the chimney sweep was the murderer, given the textile fiber evidence, was
about 1 in 6,000, and not 99.94%, as the second expert witness testified.
What if one combines both the blood and the textile evidence together,
which seems not to have happened at the trial? In this case, one of the 2,710
men who satisfy both pieces of blood type evidence would be the murderer,
and he would show the textile traces. Of the remaining innocent men, we
expect one or two (0.06%) to also show the textile traces (assuming mutual
independence of the three pieces of evidence). Thus there would be two or
three men who satisfy all three types of evidence. One of them is the murderer.
Therefore, the probability that the chimney sweep was the murderer, given the
two pieces of blood sample evidence and the textile evidence, would be be-
tween .3 and .5. This probability would not be beyond reasonable doubt.

Teaching Statistical Reasoning

The teaching of statistical reasoning is, like that of reading and writing, part
of forming an educated citizenship. Our technological world, with its abun-
dance of statistical information, makes the art of dealing with uncertain infor-
mation particularly relevant. Reading and writing is taught to every child in
modern Western democracies, but statistical thinking is not (Shaughnessy,
1992). The result has been termed "innumeracy" (Paulos, 1988). But can sta-
tistical reasoning be taught? Previous studies that attempted to teach Bayesian
inference, mostly by corrective feedback, had little or no training effect (e.g.,
Peterson, DuCharme, & Edwards, 1968; Schaefer, 1976). This result seems to
be consistent with the view that the mind does not naturally reason the Bay-
esian way. However, the argument developed in this chapter suggests a "nat-
 ECOLOGICAL INTELLIGENCE 75

ural" method of teaching: Instruct people how to represent probability infor-

mation in natural frequencies.
Peter Sedlmeier and I designed a tutorial program that teaches Bayesian
reasoning, based on the assumption that cognitive algorithms have evolved for
dealing with natural frequencies (Sedlmeier, 1997; Sedlmeier & Gigerenzer, in
press). The goal of this tutorial is to teach participants how to reason the Bay-
esian way when the information is represented in probabilities, as is usually
the case in newspapers, medical textbooks, and other information sources. The
computerized tutorial instructs participants in how to represent the probability
information in terms of natural frequencies, rather than teaching them how to
insert probabilities into Bayes's rule (Equation 1). The tutorial consists of two
parts. In the first part, participants are shown how to translate probability in-
formation into natural frequencies, visually aided by a frequency tree (or a
frequency grid); the method is illustrated by two medical problems, one of
them the mammography problem. In the second part, participants solve eight
other problems, with step-by-step guidance on what to do as well as step-by-
step feedback. If participants have difficulties, the system provides immediate
help that ensures that every participant solves all problems correctly. To see
how effective this "representation training" is, we compared it with standard
"rule training" in which people are taught how to insert probabilities into
Bayes's rule. Both training procedures were computerized, and both were sup-
ported with the same visual aids (frequency trees). There were three criteria:
immediate learning effect after training, transfer to new questions, and stability
over time. Stability over time has proven to be the most difficult to obtain, in
experimental studies as well as in teaching practice. For instance, many who
teach statistical reasoning report that students often successfully perform in
the final exam, but a few weeks later they have already forgotten most of what
they learned.
In three studies, the immediate effect of the representation training was
always larger than that of the rule training, by 10 percentage points or more.
Transfer was about the same. The most striking difference was obtained in
stability. For instance, in the study with the longest interval—people were
called back three months after training—the median performance in the group
that had received the representation training was a strong 100%. That is, all
problems were solved—even after three months. In contrast, the performance
in the group that had received the rule training was 57%, reflecting the well-
known steep forgetting curve. Teaching how to translate probabilities into
frequencies seems to have a more lasting effect than teaching rules for proc-
essing probabilities.
Thus there is evidence that (what I take to be) the natural format of infor-
mation in the environment in which humans evolved can be used to teach
people how to deal with probability information. This may be good news both
for instructors who plan to design precollege curricula that teach young people
how to infer risks in a technological world and for those unfortunate souls
among us charged with teaching undergraduate statistics.
76 ECOLOGICAL RATIONALITY

Conclusions

Information needs representation. If a representation is recurrent and stable

during human evolution, one can expect that mental algorithms are designed
to operate on this representation. In this chapter, I applied this argument to
the understanding of human inferences under uncertainty. The thesis is that
mental algorithms were designed for natural frequencies, the recurrent format
of information until very recently. I have dealt with a specific class of infer-
ences that correspond to a simple form of Bayesian inferences, in which one
of several possible states is inferred from one or a few cues. Here, mental
computations are simpler when information is encountered in the same form
as in the environment in which our ancestors evolved, rather than in the mod-
ern form of probabilities or percentages. The evidence from a broad variety of
everyday situations and laboratory experiments shows that natural frequencies
can make human minds more insightful.
 5

AIDS Counseling for Low-Risk Clients

hornier Senator Lawton Chiles of Florida reported at an AIDS conference in

1987 that of 22 blood donors in Florida who were notified that they tested
HIV-positive with the ELISA test, seven committed suicide. In the same med-
ical text that reported this tragedy, the reader is informed that "even if the
results of both AIDS tests, the ELISA and WB (Western blot), are positive, the
chances are only 50-50 that the individual is infected" (Stine, 1996, pp. 333,
338). Situations like this can occur when people with low-risk behavior, such
as blood donors, test positive. The discrepancy between what clients believe
a positive HIV test means and what it actually does mean seems to have cost
human lives in addition to the toll the disease itself has taken. One of the goals
of AIDS counseling is to explain the actual risk to the client. This chapter
deals with pretest HIV counseling of low-risk clients concerning the meaning
of a positive HIV test in German public AIDS counseling centers. We address
three questions: What information do counselors communicate to the client
concerning the chances of an HIV infection given a positive test? Is this infor-
mation communicated in a way the client can understand? How can the com-
munication and the accuracy of the information be improved?

Counseling Clients with Low-Risk Behavior

We are interested in the counseling received by members of the largest pop-

ulation group: heterosexuals who do not engage in risky behavior, such as IV-
drug use. These people take HIV tests for various reasons: voluntarily, because
they want to find out whether they are infected before getting married, having
children, or for other reasons; or involuntarily, because they are immigrants,
applicants for health or life insurance, military personnel, blood donors, or
members of other groups that are required by law to take the test. The Swedish

The work on which this chapter is based was coauthored with U. Hoffrage and
A. Ebert.

77
78 ECOLOGICAL RATIONALITY

government, for instance, has encouraged voluntary testing to the point that
"people who are unlikely to be infected are the ones who take the test, in
droves" (Mansson, 1990). Involuntary testing is a legal possibility in several
countries, one that insurers exploit to protect themselves against losses. For
instance, in 1990, Bill Clinton (then governor of Arkansas) had to take an HIV
test to get his life insurance renewed. People with low-risk behavior may be
subjected to HIV tests not only involuntarily but also unknowingly. For in-
stance, large companies in Bombay have reportedly subjected their employees
to blood tests without telling them that they were being tested for AIDS; when
a test was positive, the employee was fired.
Counseling people at low risk requires paying particular attention to false
positives, that is, to the possibility that the client has a positive HIV test even
though he or she is not infected with the virus. The lower the prevalence of
HIV in a group, the larger the proportion of false positives among those who
test positive. In other words, if a client with high-risk behavior tests positive,
the probability that he actually is infected with HIV is very high, but if some-
one with low-risk behavior tests positive, this probability may be as low as
50%, as indicated previously. If clients are not informed about this fact, they
tend to believe that a positive test means that they are infected with absolute
certainty. The case of the young man from Dallas described in the previous
chapter is one example. If he had committed suicide, as the blood donors in
the Florida case did, we might never have found out that his test was a false
positive. Emotional pain and lives can be saved if counselors inform the clients
about the possibility of false positives.1
We do not know of any study that has investigated what AIDS counselors
tell their clients about the meaning of a positive test. We pondered long over
the proper methodology, such as sending questionnaires to counselors or ask-
ing them to participate in paper-and-pencil tests. However, we decided against
questionnaires and similar methods because they are open to the criticism that
they tell us little about actual counseling sessions. For instance, these methods
have been criticized for not allowing physicians to pose their own questions
to get further information, to use their own estimates of the relevant statistical
information rather than those provided by the experimenter, and for removing
the element of actual concern for the patient, because either the patient is
fictional or the case was resolved years ago.
In the end, we decided to take a direct route. One of us went as a client to
20 counseling sites and took a series of counseling sessions and HIV tests. We

1. In their review of suicidal behavior and HIV infection, Catalan and Pugh (1995)
conclude that "suicidal ideas, completed suicide and deliberate self-harm are not un-
common in people with HIV-infection" (p. 119). However, as they themselves point out,
the evidence is far from conclusive: Many reports are anecdotal or involve few cases,
results vary between countries, and methodological problems make matching with com-
parison groups difficult (e.g., Marzuk & Perry, 1993; Pugh et al., 1993). A recent pro-
spective cohort study that controlled for several factors found a 1.35-fold increase in
suicides in HIV-positives relative to HIV-negatives, whereas earlier studies reported a 7-
to 36-fold increase in risk for HIV-positives (Dannenberg et al., 1996).
 AIDS COUNSELING FOR LOW-RISK CLIENTS 79

were interested in one important issue that AIDS counselors have to explain
to the client: What does a positive test result mean? To answer this question,
it is useful to know: (a) the base rate of HIV in heterosexual men with low
risk, which is referred to as the prevalence, (b) the probability that the test is
positive if the client is infected, which is referred to as the sensitivity (or hit
rate) of the test, and (c) the probability that the test is positive if the client is
not infected, which is known as the false positive rate (or 1 — specificity).
From this information, one can estimate what a positive test actually means,
that is, the probability of being infected if one tests positive, also known as
the positive predictive value (PPV). Let us first get the best estimates for these
values from the literature.

Prevalence

Germany has a relatively small number of reported AIDS cases. The cumula-
tive number by the end of 1995 was 13,665, as compared with some 30,000
in Italy, 38,000 in France, and more than 500,000 in the USA (World Health
Organization, 1996). Thus one can assume that the prevalence of HIV is also
comparatively low. The client in our study was 27 years old, a German het-
erosexual male who did not engage in risky behavior. What is the prevalence
of the HIV virus in 20- to 30-year-old heterosexual men in Germany who do
not engage in risky behavior? A reasonable estimate is about one in 10,000 or
0.01%.2 This figure is in the range of the prevalence of HIV in blood donors
in the United States (a group with low prevalence within the United States),

2. This is a crude estimate, given that there seem to be no published figures for the
prevalence of HIV for 20- to 30-year-old men with low-risk behavior in Germany. This
value is based on two approximations. One is to estimate the unknown prevalence by
the known prevalence in first-time male blood donors (as opposed to repeat donors, who
are a highly selected group). The proportion of HIV-positives in some 130,000 first-time
male blood donors (1986-1991, state of Baden-Wurttemberg) was 1.5 in 10,000 (Maurer
et al., 1993). For comparison, the proportion among repeat donors was one order of
magnitude smaller, about 1.2 in 100,000 (Maurer et al., 1993). Because false positives
occur, the proportion of men actually infected is smaller than 1.5 in 10,000. This esti-
mate is crude in several respects; for instance, it does not differentiate by age group and
assumes that men with low-risk behavior are comparable to first-time blood donors.
A second way to estimate the unknown prevalence is by the proportion of HIV-
positives who report infection through heterosexual contact. Dietz et al. (1994, p. 1998)
found that 3.8% of HIV-positives reported that they were infected by heterosexual con-
tact, as opposed to homosexual/bisexual behavior, injecting drug use, and other risks
(for similar figures see Gliick et al., 1990; Hoffman-Valentin, 1991; Schering, 1992). In
1994, when our study was begun, the number of HIV-positives in Germany was about
65,000, of which some 29% were in the 20- to 30-year-old age group. If one assumes
that the figure of 3.8% also holds for this age group, this results in an estimated 700
HIV-positives in this age group reporting infection through heterosexual contact. Be-
cause in 1994 there were an estimated 6 million German men between 20 and 30 who
were in no known risk group (a total of 6,718,500 men minus an estimated 11% who
belong to one or more of the known risk groups; see Statistisches Bundesamt, 1994), the
proportion of HIV-positives who report infection through heterosexual contact can be
estimated as 1.2 in 10,000.
80 ECOLOGICAL RATIONALITY

which has been estimated at one in 10,000 (Busch, 1994, p. 229) or two in
10,000 (George & Schochetman, 1994, p. 90).

Sensitivity and Specificity

HIV testing typically involves the following sequence. If the first test, ELISA,
is negative, the client is notified that he or she is HIV-negative. If positive, at
least one more ELISA (preferably from a different manufacturer) is conducted.
If the result is again positive, then the more expensive and time-consuming
Western blot test is performed. If the Western blot is also positive, then the
client is notified of being HIV-positive, and sometimes a second blood sample
is also tested. Thus two errors can occur. First, a client who is infected is
notified that he is HIV-negative. The probability of this error (false negative)
is the complement of the sensitivity of the ELISA test. The estimates for the
sensitivity typically range between 98% and 99.8% (Eberle et al., 1988; George
& Schochetman, 1994; Schwartz et al., 1990; Spielberg et al., 1989; Tu et al.,
1992; Wilber, 1991). Second, a client who is not infected is notified of being
HIV-positive. The probability of this second error (false positive) is the com-
plement of the combined specificity of the ELISA and Western blot tests. Al-
though all surveys agree that false positives do occur, the quantitative estimates
vary widely.3 This is in part due to the fact that what constitutes a positive
Western blot test has not been standardized (various agencies use different
reagents, testing methods, and test-interpretation criteria), that the ELISAs and
the Western blot tests are not independent (that is, one cannot simply multiply
the individual false positive rates of the tests to calculate the combined false
positive rate), and that the higher the prevalence in a group, the lower the
specificity seems to be for this group (Wittkowski, 1989). For instance, 20 sam-
ples—half with HIV antibodies and half without (the laboratories were not
informed which samples were which)—were sent in 1990 to each of 103 lab-
oratories in six World Health Organization (WHO) regions (Snell et al., 1992).

Both ways to estimate the unknown prevalence give consistent numbers; neverthe-
less, they should only be taken as rough approximations. Because not all of these HIV-
positives have the virus (due to false positives), we need to correct these numbers down-
ward. A prevalence of about 1 in 10,000 seems to be a reasonable estimate for the
unknown prevalence of the HIV virus in 20- to 30-year-old heterosexual German men
with low-risk behavior.
3. Among the reasons for false positives are the presence of cross-reacting antibodies
(Stine, 1996); false positive reactions with nonspecifically "sticky" IgM antibodies (Ep-
stein, 1994, p. 56); false positives from samples placed in the wrong wells; and contam-
ination of wells containing negative specimens by positive samples from adjacent wells.
In addition, heat-treated, lipemic, and hemolyzed sera may cause false positives; false
positive results have been reported to occur in 19% of hemophilia patients and in 13%
of alcoholic patients with hepatitis (George & Schochetman, 1994, p. 69). People who
have liver disease, have received a blood transfusion or gamma globulin within six
weeks of the test, or have received vaccines for influenza and hepatitis B may test false
positive as well (Stine, 1996, p. 333).
 AIDS COUNSELING FOR LOW-RISK CLIENTS 81

About 70 different combinations of tests were applied. Of the samples without

HIV antibodies, 1.3% were incorrectly classified as positive. A combined spec-
ificity of only 98.7%, as in this blind proficiency testing, however, is an un-
usually low estimate. Most of the estimates in the literature are considerably
higher, usually higher than 99.9% (Burke et al., 1988; Eberle et al., 1988; Peichl-
Hoffman, 1991; Tu et al., 1992). For instance, the German Red Cross achieved
for first-time blood donors a combined specificity of 99.98% (Wittkowski, 1989).
From the figures published, a reasonable estimate for the combined specificity
(two ELISAs, one Western blot, one blood sample) seems to be about 99.99%.
That is, the false positive rate is about one in 10,000. This is an estimate, and
more accurate numbers may be available from future research.

Positive Predictive Value

What the client needs to understand is the probability of being infected with
HIV if he tests positive. The predictive value of a positive test (PPV) can be
calculated from the prevalence p(HIV}, the sensitivity p(pos\HIV), and the
false positive rate p(pos I no HIV):

where p(no HIV] equals 1 — p(HIV). Equation 1 is known as Bayes's rule. This
rule expresses the important fact that the smaller the prevalence, the smaller
the probability that a client is infected if the test is positive. What is the pre-
dictive value of a positive test (repeated ELISA and Western blot, one blood
sample) for a 20- to 30-year-old heterosexual German man who does not engage
in risky behavior? Inserting the previous estimates—a prevalence of 0.01%, a
sensitivity of 99.8%, and a specificity of 99.99%—into Bayes's rule, the PPV
results in 0.50, or 50%.
An estimated PPV of about 50% for heterosexual men who do not engage
in risky behavior is consistent with the report of the Enquete Committee of the
German Bundestag, which estimated the PPV for low-risk people as "less than
50%" (Deutscher Bundestag, 1990, p. 121).

How to Communicate the Positive Predictive Value

Even if a counselor understands this formula, ordinary people rarely do. More-
over, we know from paper-and-pencil studies in the United States and in Ger-
many that even experienced physicians have great difficulties when asked to
infer the PPV from probability information. But we also have seen in Chapter
4 that physicians' performance can be substantially improved, by a factor of
more than four, if the information is presented in natural frequencies rather
than in terms of probabilities or percentages.
How would a counselor communicate information in natural frequencies?
She might explain to the patient the meaning of a positive test in the following
way: "Imagine 10,000 heterosexual men like you being tested. One has the
82 ECOLOGICAL RATIONALITY

virus, and he will with practical certainty test positive. Of the remaining non-
infected men, one will also test positive (the false positive rate of 0.01%). Thus
we expect that two men will test positive, and only one of them has HIV. This
is the situation you are in if you test positive; the chance of having the virus
is one out of two, or 50%."
This simple method can be applied whatever the relevant numbers are as-
sumed to be. If the prevalence is two in 10,000, the PPV would be two out of
three, or 67%. The numbers can be adjusted; the point is that clients can
understand more easily if the counselor communicates in natural frequencies
than in probabilities. With a frequency representation the client can "see" how
the PPV depends on the prevalence. If the prevalence of HIV among German
homosexuals is about 1.5%, then the counselor might explain: "Think of
10,000 homosexual men like you. About 150 have the virus, and they all will
likely test positive. Of the remaining noninfected men, one will also test pos-
itive. Thus we expect that 151 men will test positive, and 150 of them have
HIV. This is the situation you are in if you test positive; the chance of having
the virus is 150 out of 151, or 99.3%."
In general, the PPV is the number of true positives (TP) divided by the
number of true positives plus false positives (FP):

The comparison between Equations 1 and 2 shows that natural frequencies

make mental computations easier. Physicians and laypeople alike can under-
stand risks better when the information is communicated in natural frequen-
cies rather than in probabilities or percentages.

Public HIV Counseling

Some 300 German public health centers ("Gesundheitsamter") offer free HIV
tests and AIDS counseling for the general public. By 1990, these centers had
hired 315 counselors, 43% of whom were physicians, 22% social workers, and
7% psychologists. The rest had various professional training (Fischer, 1990).
As in other countries, counseling before testing is designed to make sure that
the client understands the testing procedure, the risks for HIV infection, and
the meaning of either a positive or negative test (Ward, 1994). The report
of the Enquete Committee of the German Bundestag (Deutscher Bundestag,
1990, p. 122) directs the counselor explicitly to perform a "quantitative and
qualitative assessment of the individual risk" and to "explain the reliability of
the test result" before a test is taken. If the client decides to take a test, ano-
nymity is guaranteed in all German states (unlike in the United States, where
in 25 states the patient's name is reported; Stine, 1996, p. 346). Counseling
requires both social tact and knowledge about the uncertainties involved in
testing, and the fact that in 1990 about 37% of clients tested at publicly funded
clinics in the United States failed to return for their test results suggests that
counseling is not always successful (Doll & Kennedy, 1994).
 AIDS COUNSELING FOR LOW-RISK CLIENTS 83

What information concerning the meaning of a positive test do counselors

in German public health centers give a client with low-risk behavior? How is
this information communicated (e.g., in probabilities or in natural frequen-
cies)?

Method

Counseling Centers

The "client" visited 20 public AIDS counseling centers in 20 German cities,

including large cities such as Berlin, Hamburg, and Munich. Two additional
counseling centers were visited in a pilot study that was conducted to design
the details of the interview. The 20 counseling centers were distributed over
nine German states in former West Germany. Of the 20 counselors, 14 were
physicians and 6 were social workers; 12 were female and 8 were male.
The client first contacted the health centers by telephone and made an ap-
pointment. He could visit two centers in short sequence, followed by a break
of at least two weeks to allow the hematomas from the perforation of the veins
in his arms to heal. These breaks were necessary, otherwise signs in the arm
might have suggested to the counselor that the client was a drug addict.
Investigating AIDS counselors' performance without their knowledge raises
ethical problems. We consulted the Ethics Committee of the German Associ-
ation of Psychology, which informed us that in the present case the expected
utility of the results of the study could justify deceiving the counselors. Public
counseling is public behavior; nevertheless, in deference to the Ethics Com-
mittee's interpretation of German privacy laws, we decided not to tape the
sessions. Moreover, we protect the anonymity of the counselors. We apologize
to all of the counselors for having used this covert method but believe that the
results of this study justify the approach by revealing what can be improved
in future AIDS counseling.

The Interview
The client asked the counselor the following questions (unless the counselor
provided the information unprompted):
1. Sensitivity of the HIV test. If one is infected with HIV, is it possible
to have a negative test result? How reliable does the test identify a
virus if the virus is present?
2. False positives. If one is not infected with HIV, is it possible to have
a positive test result? How reliable is the test with respect to a false
positive result?
3. Prevalence of HIV in heterosexual men. How frequent is the virus in
my risk group, that is, heterosexual men, 20 to 30 years old, with no
known risk such as drug use?
4. Predictive value of a positive test. What is the probability that men in
my risk group actually do have HIV after a positive test?
84 ECOLOGICAL RATIONALITY

5. Window period. How much time has to pass between infection and
test, so that antibodies can be detected?
The pilot study indicated a tendency in counselors to provide vague and
noninformative answers, such as, "Don't worry; the test is very reliable; trust
me." It also indicated that if the client asked for clarification more than twice,
the counselors were likely to become upset and angry, experiencing the client's
insistence on clarification as a violation of social norms of communication.
Based on these pilot sessions, the interview included the following scheme for
clarifying questions: If the counselor's answer was a quantitative estimate (a
number or a range) or if the counselor said that he or she could not (or did
not want to) give a more precise answer, then the client went on to the next
question. If the answer was qualitative (e.g., "fairly certain") or if the counselor
misunderstood or avoided answering the question, then the client asked for
further clarification and, if necessary, repeated this request for clarification one
more time. If, after the third attempt, there was still no success, the client did
not push further and went on to the next question. When the client needed
to ask for clarification concerning the prevalence of HIV (Question 3), he always
repeated his specific risk group; when asking for clarification concerning the
PPV (Question 4), he always referred to the specific prevalence in his risk group.
As mentioned previously, when the client asked for the prevalence of HIV
in his risk group, he specified this group as "heterosexual men, 20 to 30 years
old, with no known risk such as drug use." When counselors asked for more
information, which happened in only 11 of the sessions, the client explained
that he was 27 years old, monogamous, and that neither his current nor his
(few) previous sexual partners used drugs or engaged in risky behavior. In two
of these 11 cases, the client was given a detailed questionnaire to determine
his risk; in one of these cases the counselor did not look at the questionnaire,
and the client still had it in his hands when he left the site.
The client was trained in simulated sessions to use a coding system (ques-
tion number; number of repetitions of a question; the counselor's answer at
each repetition; e.g., "1; 2; 99.9%") that allowed him to write down the rele-
vant information in shorthand during the counseling or, if the session was very
brief, to rehearse the code in memory and write it down immediately after the
counseling session.
After the counseling session, the client took the HIV test, except for three
cases (in two he would have had to wait several hours to take the test, and in
one case the counselor suggested that the client might first consider it over-
night before making the decision of whether or not to take the test).

Results

Four counseling sessions are shown, for illustration, in Table 5.1. The client's
questions are abbreviated (e.g., sensitivity?], and the information provided by
the counselor directly follows the question. The counselors' answers to the
client's clarifying questions are preceded by a dash in subsequent lines.
 Table 5.1 Four sample counseling sessions
Session 1. The counselor is a female social worker
Sensitivity? False negatives really never occur. Although, if I think about the literature,
there were reports about such cases.
—I don't know exactly how many.
—It happened only once or twice.
False positives? No, because the test is repeated; it is absolutely certain.
—If there are antibodies, the test identifies them unambiguously and with absolute
certainty.
—No, it is absolutely impossible that there are false positives; by repeating the test
it is absolutely certain.
Prevalence? I can't tell you this exactly.
—Between about 1 of 500 to I of 1,000.
Positive predictive value? As I have now told you repeatedly, the test is absolutely cer-
tain.

Session 2. The counselor is a male physician

Sensitivity? When there are enough antibodies, then the test identifies them in every
case. Two tests are performed; the first test is in its fourth generation and is tuned
to be very specific and sensitive. Nevertheless, it is tuned in a way that it is more
likely to identify positives than negatives.
—99.8% sensitivity and specificity. But we repeat the test, and when it comes out
positive, then the result is as solid as cast iron.
False positives? With certainty, they don't occur; if there are false results, then only
false negatives, occurring when the antibodies have not formed.
—If you take the test here, including a confirmatory test, it is extremely certain: in
any case the specificity is 99.7%. This is as solid as cast iron. We exclude confu-
sions by using two tests.
Prevalence? The classification of individuals into risk groups is by now outdated,
therefore one cannot look at it that way.
—I don't remember this. There is a trend of the virus spreading in the general pub-
lic. Statistics are of no use for the individual case!
Positive predictive value? As I already have said: extremely certain, 99.8%.

Session 3. The counselor is a female physician

Sensitivity? The test is very, very reliable, that is, about 99.98%.
False positives? The test will be repeated. After the first test, one does not speak of
positive, but only of reactive. When all tests are performed, then the result is
sure.
—It is hard to say how many false positives occur.
—How many precisely? I would have to look up the literature to see if I could find
this information there.
Prevalence? That depends on the region.
—Of the circa 67,000 infected people [in Germany], 9% are heterosexual.
—In Munich we have 10,000 infected people, that is, 1% of the population. But
these are only numbers, which tell you nothing about whether you have the virus
or not.
Positive predictive value? As I already have mentioned, the result is 99.98% sure. If
you get a positive result, you can trust it.
—continued
86 ECOLOGICAL RATIONALITY

Table 5.1 (continued)

Session 4: The counselor is a female social worker

Sensitivity? Very, very reliable.
—No, not absolutely sure, such a thing doesn't exist in medicine, because it may be
possible that the virus cannot be identified.
—Close to 100%; I don't know exactly.
False positives? They exist but are extremely rare.
—In the order of a tenth of a percent. Probably less. However, in your risk group,
compared to high-risk groups, false positives are proportionally more frequent
[than true positives].
—I don't know the exact value.
Prevalence? With the contacts you had, the infection is unlikely.
—Generally one can't say. In our own institution, among some 10,000 tests in the
last seven years, there were only three or four heterosexuals, non-drug addicts, or
similar non-risk-group persons who tested positive.
Positive predictive value? As mentioned, the test is not 100% sure. If the test confuses
the [HIV] antibodies with others, then other methods such as repeated tests do not
help. And if someone like you does not have a real risk, then I could imagine that
even 5% to 10% of those who get a positive result will have gotten a false posi-
tive result.
When the client asked a question, he did not use the technical terms shown here (such as sensitivity)
but the wording specified in the "interview" section. The answer of the counselor is given after each
question. The following lines (beginning with a '—') are the counselor's responses to the client's
request for clarification.

Sensitivity and Window Period

Nineteen of 20 counselors gave the client information concerning sensitivity.
(The twentieth refused to give any information concerning sensitivity, speci-
ficity, and the predictive value before the test result was obtained. When the
client picked up the test result, he got no information either.) Most counselors
gave the client realistic information concerning the sensitivity (Table 5.2).
However, five of the 19 counselors incorrectly informed the client that it would
be impossible to get a false negative result, except during the window period.

Table 5.2 Information provided by the counselors

100% certainty > 99.9% > 99% > 90% Range

Sensitivity 5 (of 19) 5 6 3 90-100%
Specificity 13 (of 19) 3 3 0 99.7-100%
Prevalence — — — — 0.0075-6%
PPV 10 (of 18) 5 1 2 90-100%
Note: Not all counselors provided numerical estimates. The verbal assertion "absolutely certain" is
treated here as equivalent to 100% certain; verbal assertions such as "almost absolutely certain" and
"very, very certain" are classified as > 99%, and assertions such as "very reliable" are classified as
> 90%.
 AIDS COUNSELING FOR LOW-RISK CLIENTS 87

Fifteen counselors provided information concerning the window period when

asked for the sensitivity. The median estimate for the window period was 12
weeks.

False Positives

Thirteen of 19 counselors informed the client incorrectly that false positives

do not occur (e.g., Session 1). Eleven of these explained that the reason is that
repeated testing with ELISA and Western blot eliminates all false positives.
Five of these 13 counselors told the client that false positives had occurred in
the 1980s but no longer today, and two said that false positives would occur
only in foreign countries, such as France, but not in Germany. In addition to
these 13 counselors, three other counselors first suggested that false positives
would not occur, but became less certain when the client repeated his question
and admitted the possibility of false positives (e.g., Sessions 2 and 3). Only
the three remaining counselors informed the client right away about the ex-
istence of false positives. One of them (Session 4) was the only counselor who
informed the client about the important fact that the proportion of false posi-
tives to true positives is particularly high in heterosexuals such as the client.

Prevalence

The question concerning the prevalence of HIV in heterosexual men with low-
risk behavior produced the most uncertainty among the counselors. Sixteen of
20 (all counselors responded) expressed uncertainty or ignorance or argued that
the prevalence for heterosexual men with low-risk behavior cannot be deter-
mined (e.g., because of unreported cases) or that it would be of no use for the
individual case (e.g., Session 2). Several counselors searched for publications
in response to the client's question but found only irrelevant statistics, such as
the large number of HIV-positives in West Berlin: "The Wall was the best con-
dom for East Berlin," one counselor answered. Twelve counselors provided
numerical estimates, with a median of 0.1%. The variability of the estimates
was considerable (Table 5.2), including the extreme estimate that in people
such as the client an HIV infection is "less probable than winning the lottery
three times" (we have not included this value in Table 5.2). Four counselors
asserted that information concerning prevalence is of little or no use: "But
statistics don't help us in the individual case—and we also have no precise
data" (see also Sessions 2 and 3). Two counselors said that they have problems
remembering numbers or reasoning with numbers; for instance: "I have diffi-
culties reasoning with statistical information. It's about groups and the transfer
is problematic. It reminds me of playing the lottery. The probability of getting
all six correct is very small; nevertheless, every week someone wins."

Positive Predictive Value

Recall that under the currently available estimates, only some 50% of heter-
osexual German men with low-risk behavior actually have HIV if they test
88 ECOLOGICAL RATIONALITY

positive. The information provided by the counselors was quite different. Half
of the counselors (10 of 18; two repeatedly ignored this question) told the client
that if he tested positive it was absolutely certain (100%) that he has HIV (Table
5.2 and Session 1). Five told him that the probability is 99.9% or higher (e.g.,
Session 3). Thus, if the client had tested positive and trusted the information
provided by these 15 counselors, he might indeed have contemplated suicide,
as many have before (Stine, 1996).
How did the counselors arrive at this inflated estimate of the predictive
value? They seemed to have two lines of thought. A total of eight counselors
confused the sensitivity with the PPV(a confusion also reported by Eddy, 1982,
and Elstein, 1988); that is, they gave the same number for the sensitivity and
the PPV (e.g., Sessions 2 and 3). Three of these eight counselors explained
that, except for the window period, the sensitivity is 100% and therefore the
PPV was also 100%. Another five counselors reasoned by the second strategy.
They (erroneously) assumed that false positives would be eliminated through
repeated testing and concluded from this (consistently) that the PPV is 100%.
For both groups, the client's question concerning the PPV must have appeared
as one they had already answered. In fact, more than half of the counselors
(11 of 18) explicitly introduced their answers with a phrase such as, "As I
have already said . . ." (e.g., Sessions 1-3). Consistent with this observation,
the answers to the question concerning the PPV came rather quickly, and the
client did not need to ask for clarification as often as before. The average num-
ber of questions asked by the client on the PPV was only 1.8, compared to 2.4,
2.4, and 2.5 for sensitivity, specificity, and prevalence, respectively.
Table 5.2 lists two counselors who provided estimates of the PPV in the
correct direction (between 99% and 90%). Only one of these (Session 4), how-
ever, arrived at this estimate by reasoning that the proportion of false positives
among all positives increases when the prevalence decreases. She was also the
only one who explained to the client that there are reasons for false positives
that cannot be eliminated by repeated testing, such as that the test reacts to
antibodies that it confuses with HIV antibodies. The second counselor first
asserted that after a positive test an HIV infection is "completely certain," but
when the client asked what "completely certain" meant, the physician had
second thoughts and said that the PPV is "at least in the upper 90s" and "I
can't be more exact."

How Is the Information Communicated?

There was not a single counselor who communicated the information in nat-
ural frequencies, the representation physicians and laypeople can understand
best. Except for the prevalence of HIV, all numerical information was com-
municated to the client in terms of percentages. The four sessions in Table 5.1
illustrate this fact. As a consequence, clients will most likely not understand,
and several counselors also seemed not to understand the numbers they were
 AIDS COUNSELING FOR LOW-RISK CLIENTS 89

communicating. This can be inferred from the fact that several counselors gave
the client inconsistent pieces of information but seemed not to notice.
Two examples illustrate this disturbing fact. One physician told the client
that the prevalence of HIV in men such as the client is 0.1% or slightly higher,
and the sensitivity, specificity, and the PPV are each 99.9%. To see that this
information is contradictory, we represent it in natural frequencies. Imagine
1,000 men taking an HIV test. One of these men (0.1%) is infected, and he will
test positive with practical certainty. Of the remaining uninfected men, one
will also test positive (because the specificity is assumed to be 99.9%, which
implies a false positive rate of 0.1%). Thus two test positive, and one of them
is infected. Therefore, the odds of being infected with HIV are 1 to 1 (50%),
and not 999 to 1 (99.9%). (Even if the physician assumed a prevalence of 0.5%,
the odds are 5 to 1 rather than 999 to 1.)
Next consider the information the client received in Session 2. Assume for
the prevalence (which the counselor did not provide) the median estimate of
the other counselors, namely 0.1%. Again imagine 1,000 men. One has the
virus, and he will test positive with practical certainty (the counselor's esti-
mated sensitivity: 99.8%). Of the remaining uninfected men, three will also
test positive (the counselor's estimated specificity: 99.7%). Thus we expect
four to test positive, one of whom actually has the virus. Therefore, the prob-
ability of being infected if the test is positive is 25% (one in four), not 99.8%
as the counselor told the client.
If the counselors had been trained to represent information in natural fre-
quencies, these inconsistencies could have been easily detected. But the coun-
selors seem to have had no training in how to represent and communicate
information concerning risk. A hypothetical session in which an "ideal" coun-
selor uses natural frequencies is given here. Because the client did not find
such a counselor, the following session is fictional:
Sensitivity? The test will be positive in about 998 of 1,000 persons with
an HIV infection. Depending on circumstances, such as the specific tests
used, this estimate can vary slightly.
False positives? About 1 in 10,000. False positives can be largely reduced
by repeated testing (ELISA and Western blot), but not completely elim-
inated. Among the reasons for false positives are . . .
Prevalence? About 1 in 10,000 German heterosexual men with low-risk
behavior is HIV infected.
Positive predictive value? Think about 10,000 heterosexual men like you.
One is infected, and he will test positive with practical certainty. Of the
remaining noninfected men, one will also test positive. Thus we expect
that two men will test positive, and only one of them has HIV. This is
the situation you are in if you test positive. Your chance of having the
virus is about 1 in 2.
Do the brochures available in AIDS centers, a source from which the coun-
selors might draw, provide help in understanding what a positive test means
90 ECOLOGICAL RATIONALITY

when prevalence is low? We studied 78 different brochures from the 20 cen-

ters, some of them handed to the client by counselors, ranging from
publications of the federal government to reports from the local counseling
sites. These brochures contained a flood of useful pieces of information, such
as facts concerning the media through which the virus can and cannot be
transmitted, but very little about what a positive test means when prevalence
is low. In particular, there was no information about the prevalence in men
and women with no risky behavior. The most frequently available literature
was a series of ten issues edited by the Federal Center for Health Education
(Bundeszentrale fur gesundheitliche Aufklarung, 1988-1993). In the first issue,
the problem of false positives is mentioned, and it is remarked that the re-
peated test is "reliable." In the second issue, false positives are again briefly
mentioned, and the reasonable recommendation is made that people who have
no known risk and nevertheless test positive should head for a second test, for
instance in a public counseling site. (One might wonder what those counselors
who believe that a positive result is absolutely certain will tell such a person.)
The third issue promises "in the near future" antibody tests that identify HIV-
1 and HIV-2 infections with certainty—in contradiction to the statement of the
Enquete Committee of the German Bundestag (Deutscher Bundestag, 1988,
p. 79) that there will be no absolute certainty in identifying HIV, as there is
none with other viral infections. Nor do the remaining issues provide quanti-
tative estimates of the uncertainties involved with the test. Estimates are only
provided for the window period, which is irrelevant for the number of false
positives. In no case is an attempt made to explain to the reader the relation
between prevalence, sensitivity, false positives, and the positive predictive
value in an understandable way. Thus, based on these brochures, neither the
counselor nor the client can learn what a positive test means when prevalence
is low.

Conclusions

This study shows, for a sample of public AIDS counseling centers in Germany,
that counselors were not prepared to explain to a man with low-risk behavior
what it would mean if he tested positive for HIV. This is not to say that the
counselors were generally ignorant; on the contrary, several counselors gave
long and sophisticated lectures concerning immunodiagnostic techniques, the
nature of proteins, and the pathways of infection. But when it came to ex-
plaining to the client the risk of being infected if he tests positive, there was
a lack of information as well as a lack of knowledge of how to communicate
risks.
The key problems identified in this study are:

1. All counselors communicated information in terms of probabilities

and percentages rather than in a format that helps the clients (and
themselves) to understand, such as natural frequencies.
 AIDS COUNSELING FOR LOW-RISK CLIENTS 91

2. Only one of 20 counselors (Session 4) explained the fact that the lower
the prevalence, the higher the proportion of false positives among
positive tests.
3. A majority of counselors incorrectly assured the client that false pos-
itives would never occur. Counselors had a simple, deterministic ex-
planation: False positives would be eliminated through repeated test-
ing (and similarly, false negatives would be eliminated after the
window period).
4. Half of the counselors asserted incorrectly that if a low-risk person
tests positive, it is absolutely certain (100%) that he is infected with
the virus. Counselors arrived at this erroneous judgment by one of
two strategies. One group confused the sensitivity of the test with the
PPV. A second group assumed that there are no false positives be-
cause of repeated tests, which implies that a positive test indicates an
infection with absolute certainty.
We do not know how representative these results are for AIDS counseling
of low-risk client groups in other centers in Germany or in other countries.
This study seems to be the first one of this kind, but there is no reason to
believe that the sample of counseling centers visited is not representative of
Germany (precisely, former West Germany). The lesson of this study is the
importance of teaching counselors how to explain to clients in simple terms
the risks involved. The counselors need rough estimates of false positives,
sensitivity, and the prevalence of HIV in various risk groups. Then they can
be taught to communicate this information in an understandable way. Exper-
imental evidence suggests that the most efficient and simple method is to train
counselors to represent the relevant information in natural frequencies and to
communicate it to the client in the same way.4 Such training takes little time,
and is cost-effective, and participants do not show the usual decay of what
they had learned over time (Chapter 4).
The competence to explain in simple language what a positive result means
is certainly not all that a counselor needs to be able to do, but it is an important
part. Proper information may prevent self-destructive reactions in clients.
These reactions are avoidable, an unnecessary toll on top of the one the disease
itself takes from humankind.

4. There is also experimental evidence that the error made most often by AIDS coun-
selors in this study, confusing the sensitivity with the PPV of the test, is markedly re-
duced (from 19% to 5% of all diagnostic inferences) when information is represented
in terms of frequencies rather than probabilities (Hoffrage & Gigerenzer, 1998).
 6
How to Improve Bayesian Reasoning
without Instruction

Is the mind, by design, predisposed against performing Bayesian inference?

The classical probabilists of the Enlightenment, including Condorcet, Poisson,
and Laplace, equated probability theory with the common sense of educated
people, who were known then as "hommes eclaires." Laplace (1814/1951) de-
clared that "the theory of probability is at bottom nothing more than good sense
reduced to a calculus which evaluates that which good minds know by a sort of
instinct, without being able to explain how with precision" (p. 196). The avail-
able mathematical tools, in particular the rules of Bayes and Bernoulli, were
seen as descriptions of actual human judgment (Daston, 1981, 1988). However,
the years of political upheaval during the French Revolution prompted Laplace,
unlike earlier writers such as Condorcet, to issue repeated disclaimers that prob-
ability theory, because of the interference of passion and desire, could not ac-
count for all relevant factors in human judgment. The Enlightenment view—that
the laws of probability are the laws of the mind—moderated as it was through
the French Revolution, had a profound influence on nineteenth- and twentieth-
century science. This view became the starting point for seminal contributions
to mathematics, as when George Boole (1854/1958) derived the laws of algebra,
logic, and probability from what he believed to be the laws of thought. It also be-
came the basis of vital contributions to psychology, as when Piaget and Inhelder
(1951/1975) added an ontogenetic dimension to their Enlightenment view of
probabilistic reasoning. And it became the foundation of contemporary notions
of rationality in philosophy and economics.
Ward Edwards and his colleagues (Edwards, 1968; Phillips & Edwards,
1966; and earlier Rouanet, 1961) were the first to test experimentally whether
human inference follows Bayes's rule. Edwards concluded that inferences, al-
though "conservative," were usually proportional to those calculated from
Bayes's rule. In the 1970s and 1980s, proponents of the heuristics-and-biases
program, however, arrived at the opposite conclusion: that people systemati-
The work on which this chapter is based was coauthored with U. Hoffrage.

92
 IMPROVING BAYESIAN REASONING 93

cally neglect base rates in Bayesian inference problems. "The genuineness, the
robustness, and the generality of the base-rate fallacy are matters of established
fact" (Bar-Hillel, 1980, p. 215). Bayes's rule, like Bernoulli's theorem, was no
longer thought to describe the workings of the mind. But passion and desire
were no longer blamed as the causes of the disturbances. The new claim was
stronger. The discrepancies were taken as tentative evidence that "people do
not appear to follow the calculus of chance or the statistical theory of predic-
tion" (Kahneman & Tversky, 1973, p. 237). It was proposed that as a result of
"limited information-processing abilities" (Lichtenstein, Fischhoff, & Phillips,
1982, p. 333), people are doomed to compute the probability of an event by
crude, nonstatistical rules such as the "representativeness heuristic."
Here is the problem. There are contradictory claims as to whether people
naturally reason according to Bayesian inference. The two extremes are rep-
resented by the Enlightenment probabilists and by proponents of the
heuristics-and-biases program. Their conflict cannot be resolved by finding fur-
ther examples of good or bad reasoning; text problems generating one or the
other can always be designed. Our particular difficulty is that after more than
two decades of research, we still know little about the cognitive processes
underlying human inference, Bayesian or otherwise. This is not to say that
there have been no attempts to specify these processes. For instance, it is un-
derstandable that when the "representativeness heuristic" was first proposed
in the early 1970s to explain base-rate neglect, it was only loosely denned. Yet
at present, representativeness remains a vague and ill-defined notion. For some
time it was hoped that factors such as "concreteness," "vividness," "causal-
ity," "salience," "specificity," "extremeness," and "relevance" of base-rate in-
formation would be adequate to explain why base-rate neglect seemed to come
and go (e.g., Ajzen, 1977; Bar-Hillel, 1980; Borgida & Brekke, 1981). However,
these factors have led neither to an integrative theory nor even to specific
models of underlying processes (Hammond, 1990; Koehler, 1996; Lopes, 1991;
Scholz, 1987).
Some have suggested that there is perhaps something to be said for both
sides, that the truth lies somewhere in the middle: Maybe the mind does a
little of both Bayesian computation and quick-and-dirty inference. This com-
promise avoids the polarization of views but makes no progress on the theo-
retical front.
Both views, however, are based on an incomplete analysis: They focus on
cognitive processes, Bayesian or otherwise, without making the connection
between what we will call a cognitive algorithm and an information format.
We (a) provide a theoretical framework that specifies why frequency formats
should improve Bayesian reasoning and (b) present two studies that test
whether they do. Our goal is to lead research on Bayesian inference out of the
present conceptual cul-de-sac and to shift the focus from human errors to hu-
man engineering (see Edwards & von Winterfeldt, 1986): how to help people
reason the Bayesian way without even teaching them.
94 ECOLOGICAL RATIONALITY

Algorithms Are Designed for Information Formats

Our argument centers on the intimate relationship between a cognitive algo-

rithm and an information format. This point was made in a more general form
by the physicist Richard Feynman. In his classic The Character of Physical Law,
Feynman (1967) placed great emphasis on the importance of deriving different
formulations for the same physical law, even if they are mathematically equiv-
alent (e.g., Newton's law, the local field method, and the minimum principle).
Different representations of a physical law, Feynman reminded us, can evoke
varied mental pictures and thus assist in making new discoveries: "Psycho-
logically they are different because they are completely unequivalent when you
are trying to guess new laws" (p. 53). We agree with Feynman. The assertion
that mathematically equivalent representations can make a difference to human
understanding is the key to our analysis of intuitive Bayesian inference.
We use the general term information representation and the specific terms
information format and information menu to refer to external representations,
recorded on paper or on some other physical medium. Examples are the var-
ious formulations of physical laws included in Feynman's book and the Feyn-
man diagrams. External representations need to be distinguished from the in-
ternal representations stored in human minds, whether the latter are
prepositional (e.g., Pylyshyn, 1973) or pictorial (e.g., Kosslyn & Pomerantz,
1977). In this article, we do not make specific claims about internal represen-
tations, although our results may be of relevance to this issue.
Consider numerical information as an example of external representation.
Numbers can be represented in Roman, Arabic, and binary systems, among
others. These representations can be mapped one to one onto each other and
are in this sense mathematically equivalent. But the form of representation can
make a difference for an algorithm that does, say, multiplication. The algo-
rithms of our pocket calculators are tuned to Arabic numbers as input data
and would fail badly if one entered binary numbers. Similarly, the arithmetic
algorithms acquired by humans are designed for particular representations
(Stigler, 1984). Contemplate for a moment long division in Roman numerals.
Our general argument is that mathematically equivalent representations of
information entail algorithms that are not necessarily computationally equiv-
alent (although these algorithms are mathematically equivalent in the sense
that they produce the same outcomes; see Larkin & Simon, 1987; Marr, 1982).
This point has an important corollary for research on inductive reasoning.
Suppose we are interested in figuring out what algorithm a system uses. We
will not detect the algorithm if the representation of information we provide
the system does not match the representation with which the algorithm works.
For instance, assume that in an effort to find out whether a system has an
algorithm for multiplication, we feed that system Roman numerals. The ob-
servation that the system produces mostly garbage does not entail the conclu-
sion that it lacks an algorithm for multiplication.
In the previous two chapters, we applied this argument to statistical think-
ing—more precisely, to an elementary form of Bayesian inference. These chap-
 IMPROVING BAYESIAN REASONING 95

ters illustrate how one can help experts turn innumeracy into insight. In this
chapter, we will explore in more depth the differences between natural fre-
quencies and probabilities and extend this analysis to other representations.

Natural Sampling

Natural sampling is the process of encountering instances in a population se-

quentially. The outcome of natural sampling is natural frequencies. Figure 6.la
illustrates the natural frequencies observed in a sample of 1,000 women, in
which 10 have breast cancer and 990 do not.
There are two ways to arrive at frequencies that are not natural frequencies.
The first is through systematic sampling, in which the base rates are fixed

Figure 6.1 (a) Natural frequencies; (b) absolute frequencies that are not natu-
ral frequencies (obtained by systematic sampling or by normalizing natural
frequencies with respect to base rates); (c) relative frequencies or probabili-
ties. H = hypothesis; D = data.
96 ECOLOGICAL RATIONALITY

before any observations are made. Systematic sampling is typically employed

in experiments that test the effect of a treatment. Figure 6.1b shows an example
in which 1,000 women with and 1,000 women without breast cancer were
tested. Note that Figure 6.1b contains absolute frequencies that were not ob-
tained through natural sampling. Unlike natural frequencies, they do not con-
tain information about the base rates of women with and without cancer.
There is a second way to arrive at frequencies that are not natural frequen-
cies. This is by normalizing natural frequencies with respect to the base rates,
that is, by setting the base rates to the same value, such as 1,000 (Figure G.lb)
or 1.0 (Figure 6.1c). For instance, consider the 10 women with cancer in Figure
6.la (left column), of whom 8 test positive. Normalizing this natural frequency
results in 800 of 1,000 (Figure 6.lb), or .8 (Figure 6.1c). Normalized natural
frequencies, like absolute frequencies obtained through systematic sampling,
thus have the base-rate information filtered out of them.
Natural frequencies, such as those in Figure 6.la, result from the most com-
mon form of direct observation (outside systematic experimentation in sci-
ence). Young children can count events from an early age but do not under-
stand fractions and other kinds of normalized counts until much later in their
development (Dehaene, 1997). Examples of natural frequencies in science can
be found in medical and epidemiological screening data from which base rates,
hit rates, and false positive rates are derived. The qualifier "natural" in the
terms natural sampling and natural frequencies emphasizes that they are based
on observations made in an ecological (rather than an experimental) setting
and on raw (rather than normalized) counts of events.
The important point is that natural frequencies facilitate Bayesian compu-
tations. The reason for this is that they carry information about base rates,
whereas normalized frequencies and probabilities do not (Kleiter, 1994). If in-
formation is presented in normalized values, one has to multiply these by the
base rates in order to bring the base rates "back in." Natural frequencies need
not be multiplied in this way.
To summarize, natural sampling yields natural frequencies, which carry in-
formation about base rates and thereby facilitate Bayesian computations. Sys-
tematic sampling and normalization do not lead to natural frequencies. Sys-
tematic sampling does not capture information about base rates because they
are fixed before the observations are made. Normalization entails discarding
the base-rate information after the observations are made.
How exactly do natural frequencies facilitate Bayesian computations? Con-
sider first the standard probability format (Table 6.1) that is one interpretation
of Figure 6.1c (another one is in terms of relative frequencies). Here the infor-
mation is represented in terms of single-event probabilities: All information—a
base rate of 1%, a hit rate of 80%, and a false alarm rate of 9.6%—is in the
form of probabilities attached to a single person, and the task is to estimate a
single-event probability. The probabilities are expressed as percentages; alter-
natively, they can be represented as numbers between zero and one. We refer
to this representation (base rate, hit rate, and false alarm rate expressed as
single-event probabilities) as the standard probability format.
 Table 6.1 Information formats and menus for the mammography problem
Format and menu Description of problem

Standard probability format The probability of breast cancer is 1% for a woman

at age forty who participates in routine screening.
If a woman has breast cancer, the probability is 80%
that she will get a positive mammogram.
If a woman does not have breast cancer, the proba-
bility is 9.6% that she will also get a positive
mammogram.
A woman in this age group had a positive mam-
mogram in a routine screening. What is the proba-
bility that she actually has breast can-
cer? %
Standard frequency format 10 out of every 1,000 women at age forty who par-
ticipate in routine screening have breast cancer.
8 out of every 10 women with breast cancer will get
a positive mammogram.
95 out of every 990 women without breast cancer
will also get a positive mammogram.
Here is a new representative sample of women at
age forty who got a positive mammogram in rou-
tine screening. How many of these women do you
expect to actually have breast cancer?
out of
Short probability format The probability that a woman at age forty will get a
positive mammogram in routine screening is
10.3%.
The probability of breast cancer and a positive
mammogram is 0.8% for a woman at age forty
who participates in routine screening.
A woman in this age group had a positive mam-
mogram in a routine screening. What is the proba-
bility that she actually has breast can-
cer? %
Short frequency format 103 out of every 1,000 women at age forty get a pos-
itive mammogram in routine screening.
8 out of every 1,000 women at age forty who partici-
pate in routine screening have breast cancer and a
positive mammogram.
Here is a new representative sample of women at
age forty who got a positive mammogram in rou-
tine screening. How many of these women do you
expect to actually have breast cancer?
out of
98 ECOLOGICAL RATIONALITY

What is the algorithm needed to calculate the Bayesian posterior probability

p(cancer I positive) from the standard probability format? Again, we use the
symbols H and — H for the two hypotheses or possible outcomes (breast cancer
and no breast cancer) and D for the data obtained (positive mammogram). As
shown in Chapter 4, a Bayesian algorithm for computing the posterior proba-
bility p(H\ D) with the values given in the standard probability format amounts
to solving the following equation:

The result is 0.078. However, physicians, college students, and staff at Harvard
Medical School all have equally great difficulties with this and similar medical
problems and typically estimate the posterior probability p(cancer I positive)
to be between 70% and 80%, rather than 7.8% (Chapter 4).
The experimenters who have amassed the apparently damning body of ev-
idence that humans fail to meet the norms of Bayesian inference have usually
given their research participants information in the standard probability format
(or its variant, in which one or more of the three percentages are relative fre-
quencies; see below). Studies on the cab problem (Bar-Hillel, 1980), the light-
bulb problem (Lyon & Slovic, 1976), and various disease problems (Casscells
et al., 1978; Eddy, 1982; Hammerton, 1973) are examples. Results from these
and other studies have generally been taken as evidence that the human mind
does not reason with Bayesian algorithms. Yet this conclusion is not war-
ranted, as explained before. One would be unable to detect a Bayesian algo-
rithm within a system by feeding it information in a representation that does
not match the representation with which the algorithm works.
In the last few decades, the standard probability format has become a com-
mon way to communicate information ranging from medical and statistical
textbooks to psychological experiments. But we should keep in mind that it is
only one of many mathematically equivalent ways of representing information;
it is, moreover, a recently invented notation. Neither the standard probability
format nor Equation 1 was used in Bayes's (1763) original essay. As Figure 6.2
shows, with natural frequencies one does not need a pocket calculator to es-
timate the Bayesian posterior. All one needs is the number of cases that had
both the symptom and the disease (here, 8) and the number of symptom cases
(here, 8 + 95). A Bayesian algorithm for computing the posterior probability
p(H\D) from the frequency format (see Figure 6.2, left side) requires solving
the following equation:

where d&h (data and hypothesis) is the number of cases with symptom and
disease, and d8t—his the number of cases having the symptom but lacking the
 IMPROVING BAYESIAN REASONING 99

Figure 6.2 Bayesian inference and information representation (natural sam-

pling of frequencies and standard probability format).

disease. One does not even need to keep track of the base rate of the disease.
A medical student who struggles with single-event probabilities presented in
medical textbooks may on the other hand have to rely on a calculator and end
up with little understanding of the result (see Figure 6.2, right side).1 Hence-
forth, when we use the term frequency format, we always refer to natural fre-
quencies as defined by the natural sampling tree in Figure 6.2.
Comparison of Equations 1 and 2 leads to our first theoretical result:
Result 1: Computational demands. Bayesian algorithms are computa-
tionally simpler when information is encoded in a frequency format
rather than a standard probability format.
By "computationally simpler" we mean that (a) fewer operations (multiplica-
tion, addition, or division) need to be performed in Equation 2 than Equation
1, and (b) the operations can be performed on natural numbers (absolute fre-
quencies) rather than fractions (such as percentages).
Equations 1 and 2 are mathematically equivalent formulations of Bayes's
rule. Both produce the same result, p(H\D) = .078. Equation 1 is a standard
version of Bayes's rule in today's textbooks in the social sciences, whereas

1. This clinical example illustrates that the standard probability format is a conven-
tion rather than a necessity. Clinical studies often collect data that have the structure of
frequency trees as in Figure 6.2. Such information can always be represented in fre-
quencies as well as probabilities.
100 ECOLOGICAL RATIONALITY

Equation 2 corresponds to Thomas Bayes's (1763) original "Proposition 5" (see

Earman, 1992).
Equation 2 implies three further (not independent) theoretical results con-
cerning the estimation of a Bayesian posterior probability p(H\ D] in frequency
formats (Kleiter, 1994).
Result 2: Attentional demands. Only two kinds of information need to
be attended to in natural sampling: the absolute frequencies d&h and
d&—h (or, alternately d&h and d, where d is the sum of the two frequen-
cies).
An organism does not need to keep track of the whole tree in Figure 6.2 but
only of the two pieces of information contained in the bold circles. These are
the hit and false alarm frequencies (not to be confused with hit and false alarm
rates).
Result 3: Base rates need not be attended to.
Neglect of base rates is perfectly rational in natural sampling. For instance,
one does not need to pay attention to the base rate of the disease (10 out of
1,000; see Figure 6.2).
Result 4: Posterior distributions can be computed.
Natural frequencies can carry more information than probabilities. Information
about the sample size allows inference beyond single-point estimates, such as
the computation of posterior distributions, confidence intervals for posterior
probabilities, and second-order probabilities (Kleiter, 1994; Sahlin, 1993). In
this chapter, however, we focus only on single-point estimation.
For the design of the experiments reported herein, it is important to note
that the Bayesian algorithms (Equations 1 and 2) work on the final tally of
frequencies (see Figure 6.2), not on the sequential record of updated frequen-
cies. Thus, the same four results still hold even if nothing but the final tally
is presented to the participants in an experiment.

Information Format and Menu

We propose to distinguish two aspects of information representation, infor-

mation format and information menu. The standard probability format has a
probability format, whereas a frequency format is obtained by natural sam-
pling. However, as the second result (attentional demands) shows, there is
another difference. The standard probability format displays three pieces of
information, whereas two are sufficient in natural sampling. We use the term
information menu to refer to the manner in which information is segmented
into pieces within any format. The standard probability format displays the
three pieces p(H], p(D\H), and p(D\ —H] (often called base rate, hit rate, and
false alarm rate, respectively). We refer to this as the standard menu. Natural
 IMPROVING BAYESIAN REASONING 101

sampling yields a more parsimonious menu with only two pieces of informa-
tion, d&h and dlk—h (or alternatively, d&/i and d). We call this the short menu.
So far we have introduced the probability format with a standard menu and
the frequency format with a short menu. However, information formats and
menus can be completely crossed. For instance, if we replace the probabilities
in the standard probability format with frequencies, we get a standard menu
with a frequency format, or the standard frequency format. Table 6.1 uses the
mammography problem to illustrate the four versions that result from crossing
the two menus with the two formats. All four displays are mathematically
equivalent in the sense that they lead to the same Bayesian posterior proba-
bility. In general, within the same format information can be divided into var-
ious menus; within the same menu, it can be represented in a range of formats.
To transform the standard probability format into the standard frequency
format, we simply replaced 1% with "10 out of 1,000," "80%" with "8 out of
10," and so on (following the tree in Figure 6.2) and phrased the task in terms
of a frequency estimate. All else went unchanged. Note that whether the fre-
quency format actually carries information about the sample size (e.g., that
there were exactly 1,000 women) or not (as in Table 6.1, where it is said "in
every 1,000 women") makes no difference for Results 1 to 3 because these
relate to single-point estimates only (unlike Result 4).
What are the Bayesian algorithms needed to draw inferences from the two
new format-menu combinations? The complete crossing of formats and menus
leads to two important results. A Bayesian algorithm for the short probability
format, that is, the probability format with a short menu (as in Table 6.1),
amounts to solving the following equation:

This version of Bayes's rule is equivalent to Equation 1. The algorithm for

computing p(H\ D) from Equation 3, however, is computationally simpler than
the algorithm for computing p(H\D) from Equation 1.
What Bayesian computations are needed for the standard frequency format?
Equation 2 specifies the computations for both the standard and short menus
in frequency formats. The same algorithm is sufficient for both menus. In the
standard frequency format of the mammography problem, for instance, the
expected number of actual breast cancer cases among positive tests is com-
puted as 8/(8 + 95). Thus we have the following two important theoretical
results concerning formats (probability vs. frequency) and menus (standard vs.
short):
Result 5: With a probability format, the Bayesian computations are sim-
pler in the short menu than in the standard menu.

Result 6: With a frequency format, the Bayesian computations are the

same for the two menus.
102 ECOLOGICAL RATIONALITY

If the two pieces of information in the short menu are d&h and d, as in
Table 6.1, rather than dlkh and d&.—h, then the Bayesian computations are even
simpler because the sum in the denominator is already computed.

Relative Frequencies
Several studies of Bayesian inference have used standard probability formats
in which one, two, or all three pieces of information were presented as relative
frequencies rather than as single-event probabilities—although the task still
was to estimate a single-event probability (e.g., Tversky & Kahneman's, 1982b,
cab problem). For instance, in the following version of the mammography
problem, all information is represented in relative frequencies (in %).
Relative frequency version (standard menu)
1% of women at age forty who participate in routine screening have
breast cancer. 80% of women with breast cancer will get positive mam-
mograms. 9.6% of women without breast cancer will also get positive
mammograms. A woman in this age group had a positive mammogram
in a routine screening. What is the probability that she actually has breast
cancer? %
Is the algorithm needed for relative frequencies computationally equivalent to
the algorithm for natural frequencies? The relative frequency format does not
display the natural frequencies needed for Equation 2. Rather, the numbers are
the same as in the probability format, making the Bayesian computation the
same as in Equation 1. This yields the following result:
Result 7: Algorithms for relative frequency versions are computationally
equivalent to those for the standard probability format.
We tested several implications of Results 1 through 7 (except Result 4) in
the studies reported below.

The Format of the Single-Point Estimate

Whether estimates relate to single events or frequencies has been a central
issue within probability theory and statistics since the decline of the classical
interpretation of probability in the 1830s and 1840s. The question has polar-
ized subjectivists and frequentists, additionally subdividing frequentists into
moderate frequentists, such as R. A. Fisher, and strong frequentists, such as J.
Neyman. A single-point estimate can be interpreted as a probability or a fre-
quency. For instance, clinical inference can be about the probability that a
particular person has cancer or about the frequency of cancer in a new sample
of people. Foraging (Simon, 1956; Stephens & Krebs, 1986) provides an excel-
lent example of a single-point estimate reasonably being interpreted as a fre-
quency. The foraging organism is interested in making inferences that lead to
satisfying results in the long run. Will it more often find food if it follows Cue
X or Cue Y? Here the single-point estimate can be interpreted as an expected
 IMPROVING BAYESIAN REASONING 103

frequency for a new sample. In the experimental research of the past two de-
cades, participants were almost always required to estimate a single-event
probability. But this need not be. In the experiments reported herein, we asked
people both for single-event probability and frequency estimates.
To summarize, mathematically equivalent information need not be com-
putationally and psychologically equivalent. We have shown that Bayesian
algorithms can depend on information format and menu, and we derived sev-
eral specific results for when algorithms are computationally equivalent and
when they are not.

Cognitive Strategies for Bayesian Inference

How might the mind draw inferences that follow Bayes's rule? Surprisingly,
this question seems rarely to have been posed. Psychological explanations typ-
ically were directed at "irrational" deviations between human inference and
the laws of probability; the "rational" seems not to have demanded an expla-
nation in terms of cognitive processes. The cognitive account of probabilistic
reasoning by Piaget and Inhelder (1951/1975), as one example, stops at the
precise moment the adolescent turns "rational," that is, reaches the level of
formal operations.
We propose three classes of cognitive strategies for Bayesian inference: first,
the algorithms corresponding to Equations 1 through 3; second, physical an-
alogs of Bayes's rule, as anticipated by Bayes's (1763) billiard table; and third,
shortcuts that simplify the Bayesian computations in Equations 1 through 3.

Physical Analogs

We illustrate physical analogs and shortcuts by drawing on actual performance

from the studies reported here, in which none of the participants was familiar
with Bayes's rule. The German measles problem (in standard probability for-
mat and with the numerical information given in Study 2) serves as our ex-
ample.
German measles during early pregnancy can cause severe prenatal dam-
age in the child. Therefore, pregnant women are routinely tested for
German measles infection. In one such test, a pregnant woman is found
to be infected. In order best to advise this woman what to do, the phy-
sician first wants to determine the probability of severe prenatal damage
in the child if a mother has German measles during early pregnancy. The
physician has the following information: The probability of severe pre-
natal damage in a child is 0.5%. The probability that a mother had
German measles during early pregnancy if her child has severe prenatal
damage is 40%. The probability that a mother had German measles dur-
ing early pregnancy if her child does not have severe prenatal damage
is 0.01%. What is the probability of severe prenatal damage in the child
if the mother has German measles during early pregnancy? %
104 ECOLOGICAL RATIONALITY

Figure 6.3 A physical analog developed by a brilliant student. The "beam

cut" is illustrated for the German measles problem in the standard probabil-
ity format. H = severe prenatal damage in the child, D = mother had
German measles in early pregnancy. The information is p(H) = 0.5%,
p(D\H] = 40%, and p(D\ -H} = 0.01%. The task is to infer p(H\D).

The "beam analysis" (see Figure 6.3) is a physical analog of Bayes's rule de-
veloped by one of our research participants. This student represented the class
of all possible outcomes (child has severe prenatal damage and child does not
have severe prenatal damage) by a beam. He drew inferences (here, about the
probability that the child has severe prenatal damage) by cutting off two pieces
from each end of the beam and comparing their size. His algorithm was as
follows:
Step 1: Base rate cut. Cut off a piece the size of the base rate from the
right end of the beam.
Step 2: Hit rate cut. From the right part of the beam (base rate piece),
cut off a proportion p(D I H].
Step 3: False alarm cut. From the left part of the beam, cut off a propor-
tion p(D I -H).
Step 4: Comparison. The ratio of the right piece to both pieces is the
posterior probability.
This algorithm amounts to Bayes's rule in the form of Equation 1.
 IMPROVING BAYESIAN REASONING 105

Shortcuts: Probability Format

Shortcuts exploit the structure of environments. Used in the right situation,
they are ecologically rational—they can ignore information and reduce com-
putations without substantial loss of accuracy. We have observed in our ex-
periments three elementary shortcuts and several combinations thereof. For
instance, by ignoring small "slices," one can simplify the computation without
much loss of accuracy, which is easily compensated for by the fact that less
computation means a reduced chance of computational errors. We illustrate
these shortcuts using the beam analysis (see Figure 6.3). However, these short-
cuts are not restricted to physical analogs, and they were used by many of our
participants.

Rare-Event Shortcut Rare events—that is, outcomes with small base rates, such
as severe prenatal damage—enable simplification of the Bayesian inference
with little reduction in accuracy. If an event is rare, that is, if p(H] is very
small, and p(—H) is therefore close to 1.0, then p(D\ —H)p(—H) can be ap-
proximated by p{D\ —H}. That is, instead of cutting the proportion p(D\ —H]
off the left part of the beam (Step 3), it is sufficient to cut a piece of absolute
size p(D\ —H). The rare-event shortcut (see Figure 6.3) is as follows:
IF the event is rare,
THEN simplify Step 3: Cut a piece of absolute size p(D\ —H}.
This shortcut corresponds to the approximation

The shortcut works well for the German measles problem, where the base rate
of severe prenatal damage is very small, p(H] = .005. The shortcut estimates
p(H\ D] as .9524, whereas Bayes's rule gives .9526. It also works with the mam-
mography problem, where it generates an estimate of .077, compared with .078
from Bayes's rule.

Big Hit-Rate Shortcut Large values of p(D I H) (such as high sensitivities in

medical tests; that is, excellent hit rates) allow one to skip Step 2 with little
loss of accuracy. If p[D\ H) is very large, then the p(H] piece is practically the
same size as the piece one obtains from cutting all but a tiny sliver from the
p(H) piece. The big hit-rate shortcut is then as follows:
IF p(D I H ) is very large,
THEN skip Step 2.
This shortcut corresponds to the approximation

The big hit-rate shortcut would not work as well as the rare-event shortcut in
the German measles problem because p ( D \ H ) is only .40. Nevertheless, the
shortcut estimate is only a few percentage points removed from that obtained
106 ECOLOGICAL RATIONALITY

with Bayes's rule (.980 instead of .953). The big hit-rate shortcut works well,
to offer one instance, in medical diagnosis tasks where the hit rate of a test is
high (e.g., around .99 as in HIV tests).

Comparison Shortcut If one of the two pieces obtained in Steps 2 and 3 is

small relative to the other, then the comparison in Step 4 can be simplified
with little loss of accuracy. For example, German measles in early pregnancy
and severe prenatal damage in the child occur more frequently than do German
measles and no severe damage. More generally, if D&H cases are much more
frequent than D&—H cases (as in the German measles problem), or vice versa
(as in the mammography problem), then only two pieces (rather than three)
need to be related in Step 4. The comparison shortcuts for these two cases are
as follows:
IF D&-H occurs much more often than D&H,
THEN simplify Step 4: Take the ratio ofD&H (right piece) to D&-H (left
piece) as the posterior probability.
This shortcut corresponds to the approximation

Note that the right side of this approximation is equivalent to the posterior
odds ratio p(H\D]/p(—H\D). Thus the comparison shortcut estimates the pos-
terior probability by the posterior odds ratio.
IF D&H occurs much more often than D&—H,
THEN simplify Step 4: Take the ratio ofD&-H (left piece) to D&H (right
piece) as the complement of the posterior probability.
This shortcut corresponds to the approximation

The comparison shortcut estimates p(H\D] as .950 in the German measles

problem, whereas Bayes's rule gives .953. The comparison shortcut is simpler
when the D&—H cases are the more frequent ones, which is typical for medical
diagnosis, where the number of false alarms is much larger than the number
of hits, as in mammography.

Multiple Shortcuts Two or three shortcuts can be combined, which results in

a large computational simplification. What we call the quick-and-clean short-
cut combines all three. Its conditions include a rare event, a large hit rate, and
many D&—H cases compared with D&H cases (or vice versa). The quick-and-
clean shortcut is as follows:
IF an event H is rare, p(D\H] high, and D&-H cases much more frequent
than D&'H cases,
THEN simply divide the base rate by the false alarm rate.
 IMPROVING BAYESIAN REASONING 107

This shortcut corresponds to the approximation

The conditions of the quick-and-clean shortcut seem to be not infrequently

satisfied. Consider routine HIV testing: According to present law, the U.S. im-
migration service makes an HIV test a condition sine qua non for obtaining a
green card. Mr. Quick has applied for a green card and wonders what a positive
test result (in the first ELISA test) indicates. The information available is a base
rate of .002, a hit rate of .99, and a false alarm rate of .02; all three conditions
for the quick-and-clean shortcut are thus satisfied. Mr. Quick computes .002/
.02 — .10 as an estimate of the posterior probability of actually being infected
with the HIV virus if he tests positive. Bayes's rule results in .09. The shortcut
is therefore an excellent approximation. Alternately, if D&H cases are more
frequent, then the quick-and-clean shortcut is to divide the false alarm rate by
the base rate and to use this as an estimate for 1— p(H\ D). In the mammography
and German measles problems, where the conditions are only partially satis-
fied, the quick-and-clean shortcut still leads to surprisingly good approxima-
tions. The posterior probability of breast cancer is estimated at .01/.096, which
is about .10 (compared with .078), and the posterior probability of severe pre-
natal damage is estimated as .98 (compared with .953).

Shortcuts: Frequency Format

Does the standard frequency format invite the same shortcuts? Consider the
inference about breast cancer from a positive mammogram, as illustrated in
Figure 6.2. Would the rare-event shortcut facilitate the Bayesian computations?
In the probability format, the rare-event shortcut uses p(D I —H) to approximate
p(-H]p(D I — H); in the frequency format, the latter corresponds to the absolute
frequency 95 (i.e., d&—h) and no approximation is needed. Thus a rare-event
shortcut is of no use and would not simplify the Bayesian computation in
frequency formats. The same can be shown for the big hit-rate shortcut for the
same reason. The comparison shortcut, however, can be applied in the fre-
quency format:
IF d&—h occurs much more often than d&h,
THEN compute d&h/d&-h.
The condition and the rationale are the same as in the probability format.
To summarize, we proposed three classes of cognitive strategies underlying
Bayesian inference: (a) algorithms that satisfy Equations 1 through 3; (b) phys-
ical analogs that work with operations such as "cutting" instead of multiplying
(Figure 6.3); and (c) three shortcuts that can exploit environmental structures.

Predictions

We now derive several predictions from the theoretical results obtained. The
predictions specify conditions that do and do not make people reason the
108 IMPROVING BAYESIAN REASONING

Bayesian way. The predictions should hold independently of whether the cog-
nitive strategies follow Equations 1 through 3, whether they are physical an-
alogs of Bayes's rule, or whether they include shortcuts.
Prediction 1: Frequency formats elicit a substantially higher proportion
of Bayesian inferences than probability formats.
This prediction is derived from Result 1, which states that the Bayesian algo-
rithm is computationally simpler in frequency formats.2
Prediction 2: Probability formats elicit a larger proportion of Bayesian
inferences for the short menu than for the standard menu.
This prediction is deduced from Result 5, which states that with a probability
format, the Bayesian computations are simpler in the short menu than in the
standard menu.
Prediction 3: Frequency formats elicit the same proportion of Bayesian
inferences for the two menus.
This prediction is derived from Result 6, which states that with a frequency
format, the Bayesian computations are the same for the two menus.
Prediction 4: Relative frequency formats elicit the same (small) propor-
tion of Bayesian inferences as probability formats.
This prediction is derived from Result 7, which states that the Bayesian algo-
rithms are computationally equivalent in both formats.

Operational Criteria for Identifying Cognitive Strategies

The data we obtained for each of several thousand problem solutions were
composed of a participant's (a) probability or frequency estimate and (b) on-
line protocol ("write aloud" protocol) of his or her reasoning. Data type (a)
allowed for an outcome analysis, as used exclusively in most earlier studies
on Bayesian inference, whereas data type (b) allowed additionally for a process
analysis.

Double Check: Outcome and Process

We classified an inferential process as a Bayesian inference only if (a) the

estimated probability or frequency was exactly the same as the value calcu-

2. At the point when we introduced Result 1, we had dealt solely with the standard
probability format and the short frequency format. However, Prediction 1 also holds
when we compare formats across both menus. This is the case because (a) the short
menu is computationally simpler in the frequency than in the probability format, be-
cause the frequency format involves calculations with natural numbers and the proba-
bility format with fractions, and (b) with a frequency format, the Bayesian computations
are the same for the two menus (Result 6).
 ECOLOGICAL RATIONALITY 109

lated from applying Bayes's rule to the information given (outcome criterion),
and (b) the on-line protocol specified that one of the Bayesian computations
defined by Equations 1 through 3 or one (or several) of the shortcuts was used,
either by means of calculation or physical representation (process criterion).
We applied the same strict criteria to identify non-Bayesian cognitive strate-
gies.

Outcome: Strict Rounding Criterion By the phrase "exactly the same" in the
outcome criterion, we mean the exact probability or frequency, with exceptions
made for rounding up or down to the next full percentage point (e.g., in the
German measles problem, where rounding the probability of 95.3% down or
up to a full percentage point results in 95% or 96%). If, for example, the on-
line protocol showed that a participant in the German measles problem had
used the rare-event shortcut and the answer was 95% or 96% (by rounding),
this inferential process was classified as a Bayesian inference. Estimates below
or above were not classified as Bayesian inferences: If, for example, another
participant in the same problem used the big hit-rate shortcut (where the con-
dition for this shortcut is not optimally satisfied) and accordingly estimated
98%, this was not classified as a Bayesian inference. Cases of the latter type
ended up in the category of "less frequent strategies." This example illustrates
the strictness of the joint criteria. The strict rounding criterion was applied to
the frequency format in the same way as to the probability format.
When a participant answered with a fraction—such as that resulting from
Equation 3—without performing the division, this was treated as if he or she
had performed the division. We did not want to evaluate basic arithmetic
skills. Similarly, if a participant arrived at a Bayesian equation but made a
calculation error in the division, we ignored the calculation error.

Process: "Write Aloud" Protocols Statistical reasoning often involves pictorial

representations as well as computations. Neither are easily expressed verbally,
as in "think aloud" methods. Pictorial representations and computations con-
sequently are usually expressed in the form of drawings and by writing down
equations and calculations. We designed a "write aloud" technique for tracking
the reasoning process without asking the participant to talk aloud either during
or after the task.
The "write aloud" method consisted of the following steps. First, partici-
pants were instructed to record their reasoning unless merely guessing the
answer. We explained that a protocol may contain a variety of elements, such
as diagrams, pictures, calculations, or whatever other tools one may use to find
a solution. Each problem was on a separate page, which thus allowed ample
space for notes, drawings, and calculations. Second, after a participant had
completed a problem, he or she was asked to indicate whether the answer was
based on a calculation or on a guess. Third, when a "write aloud" protocol
was unreadable or the process that generated the probability estimate was un-
clear, and the participant had indicated that the given result was a calculation,
then he or she was interviewed about the particular problem after completing
 Table 6.2 Information given and Bayesian solutions for the 15 problems in Study 1
Task: Estimate p(H\D] Information (standard frequency format)" Bayesb

H D H D\H D\ -H p(H\D)

Breast cancer Mammogram positive 10 1,000 8 10 95 990 7.77

Prenatal damage in child German measles in
mother 21 10,000 10 21 50 10,000 16.70
Blue cab Eyewitness says "blue" 15 100 12 15 17 85 41.38
AIDS HIV test positive 100 1,000,000 100 100 1,000 1,000,000 9.09
Heroin addict Fresh needle prick 10 100,000 10 10 190 100,000 5.00
Pregnant Pregnancy test positive 20 1,000 19 20 5 980 79.17
Car accident Driver drunk 100 10,000 55 100 500 9,900 9.91
Bad posture in child Heavy books carried
daily 50 1,000 20 50 190 950 9.52
Accident on way to Child lives in urban
school area 30 1,000 27 30 388 970 6.51
Committing suicide Professor 240 1,000,000 36 240 120,000 1,000,000 0.03
Red ball Marked with star 400 500 300 400 25 100 92.31
Choosing course in eco-
nomics Career oriented 300 1,000 210 300 350 700 37.50
Active feminist Bank teller 5,000 100,000 20 5,000 2,000 95,000 0.99
Pimp Wearing a Rolex 50 1,000,000 40 50 500 1,000,000 7.41
Admission to school Particular placement
test result 360 1,000 270 360 128 640 67.84
a. The representation of the information is shown only for the standard frequency format (frequency format and standard menu). The other representations (see Table
6.1) can be derived from this. The two numbers for each piece of information are connected by an "out of" relation; for example, the information concerning Hin the
first problem should be read as "10 out of 1,000."
b. Probabilities are expressed as percentages.
 IMPROVING BAYESIAN REASONING 111

all tasks. This happened only a few times. If a participant could not immedi-
ately identify what his or her notes meant, we did not inquire further.
The "write aloud" method avoids two problems associated with retrospec-
tive verbal reports: that memory of the cognitive strategies used may have
faded by the time of a retrospective report (Ericsson & Simon, 1984) and that
participants may have reported how they believe they ought to have thought
rather than how they actually thought (Nisbett & Wilson, 1977).
We used the twin criteria of outcome and process to cross-check outcome
by process and vice versa. The outcome criterion prevents a shortcut from
being classified as a Bayesian inference when the precondition for the shortcut
is not optimally satisfied. The process criterion protects against the opposite
error, that of inferring from a probability judgment that a person actually used
Bayesian reasoning when he or she did not.
We designed two studies to identify the cognitive strategies and test the
predictions. Study 1 was designed to test Predictions 1,2, and 3.

Study 1: Information Formats and Menus

Method
Sixty students, 21 men and 39 women from ten disciplines (predominantly
psychology) from the University of Salzburg, Austria, were paid for their par-
ticipation. The median age was 21 years. None of the participants was familiar
with Bayes's rule. Participants were studied individually or in small groups
of 2 or 3 (in two cases, 5). On the average, students worked 73 min in the
first session (range = 25-180 min) and 53 min in the second (range = 30-120
min).
We used two formats, probability and frequency, and two menus, standard
and short. The two formats were crossed with the two menus, so four versions
were constructed for each problem. There were 15 problems, including the
mammography problem (Eddy, 1982; see Table 6.1), the cab problem (Tversky
& Kahneman, 1982b), and a short version of Ajzen's (1977) economics problem.
The four versions of each problem were constructed in the same way as ex-
plained before with the mammography problem (see Table 6.1).3 In the fre-
quency format, participants were always asked to estimate the frequency of "h
out of d"; in the probability format, they were always asked to estimate the
probability p(H\D}. Table 6.2 shows for each of the 15 problems the infor-
mation given in the standard frequency format; the information specified in
the other three versions can be derived from that.

3. If the Y number in "X out of Y" was large and odd, such as 9,950, we rounded
the number to a close, more simple number, such as 10,000. The German measles prob-
lem is an example. This made practically no difference for the Bayesian calculation and
was meant to prevent participants from being puzzled by odd Y numbers.
112 ECOLOGICAL RATIONALITY

Participants were randomly assigned to two groups, with the members of

both answering each of the 15 problems in two of the four versions. One group
received the standard probability format and the short frequency format; the
other, the standard frequency format and the short probability format. Each
participant thus worked on 30 tasks. There were two sessions, 1 week apart,
with 15 problems each. Formats and menus were distributed equally over the
sessions. The two versions of one problem were always given in different ses-
sions. The order of the problems was determined randomly, and two different
random orders were used within each group.

Results

Bayesian Reasoning
Predition 1: Frequency formats elicit a substantially higher proportion of
Bayesian inferences than probability formats.
Do frequency formats foster Bayesian reasoning? Yes. Frequency formats elic-
ited a substantially higher proportion of Bayesian inferences than probability
formats: 46% in the standard menu and 50% in the short menu. Probability
formats, in contrast, elicited 16% and 28%, for the standard menu and the
short menu, respectively. These proportions of Bayesian inferences were ob-
tained by the strict joint criteria of process and outcome and held fairly stable
across 15 different inference problems. Note that 50% Bayesian inferences
means 50% of all answers, and not just of those answers where a cognitive
strategy could be identified. The percentage of identifiable cognitive strategies
across all formats and menus was 84%.
Figure 6.4 shows the proportions of Bayesian inferences for each of the 15
problems. The individual problems mirror the general result. For each prob-
lem, the standard probability format elicited the smallest proportion of Baye-
sian inferences. Across formats and menus, in every problem Bayesian infer-
ences were the most frequent.
The comparison shortcut was used quite aptly in the standard frequency
format, that is, only when the precondition of this shortcut was satisfied to a
high degree. It was most often used in the suicide problem, in which the ratio
between DikH cases and D&—H cases was smallest (Table 6.2), that is, in which
the precondition was best satisfied. Here, 9 out of 30 participants used the
comparison shortcut (and 5 participants used the Bayesian algorithm without
a shortcut). In all 20 instances where the shortcut was used, 17 satisfied the
strict outcome criterion, and the remaining 3 were accurate to within 4 per-
centage points.
Because of the strict rounding criterion, the numerical estimates of the par-
ticipants using Bayesian reasoning can be directly read from Table 6.2. For
instance, in the short frequency version of the mammography problem, 43.3%
of participants (see Figure 6.4) came up with a frequency estimate of 8 out of
103 (or another value equivalent to 7.8%, or between 7% and 8%).
 IMPROVING BAYESIAN REASONING 113

Figure 6.4 Proportion of Bayesian inferences in the 15 problems of Study 1.

Standard probability = probability format with standard menu; short fre-
quency = frequency format with short menu; and so on.

The empirical result in Figure 6.4 is consistent with the theoretical result
that frequency formats can be handled by Bayesian algorithms that are com-
putationally simpler than those required by probability formats.
Prediction 2: Probability formats elicit a larger proportion of Bayesian
inferences for the short menu than for the standard menu.
The percentages of Bayesian inferences in probability formats were 16% and
28% for the standard menu and the short menu, respectively. Prediction 2
holds for each of the 15 problems (Figure 6.4).
Prediction 3: The proportion of Bayesian inferences elicited by the fre-
quency format is independent of the menu.
The effect of the menu largely, but not completely, disappeared in the fre-
quency format. The short menu elicited 3.7 percentage points more Bayesian
strategies than the standard menu. The residual superiority of the short menu
could have the following cause: Result 2 (attentional demands) states that in
natural sampling it is sufficient for an organism to monitor either the frequen-
cies d&h and d or d&h and d&—h. We have chosen the former pair for the
114 ECOLOGICAL RATIONALITY

short menus in our studies and thus reduced the Bayesian computation by one
step, that of adding up d&h and d!k—h to d, which was part of the Bayesian
computation in the standard but not the short menu. This additional compu-
tational step is consistent with the small difference in the proportions of Baye-
sian inferences found between the two menus in the frequency formats.
How does the impact of format on Bayesian reasoning compare with that
of menu? The effect of the format was about three times larger than that of the
menu (29.9 and 21.6 percentage points difference compared with 12.1 and 3.7).
Equally striking, the largest percentage of Bayesian inferences in the two prob-
ability menus (28%) was considerably smaller than the smallest in the two
frequency menus (46%).

Non-Bayesian Reasoning We found three major non-Bayesian cognitive strat-

egies (see Table 6.3).
Joint Occurrence. The most frequent non-Bayesian strategy was a computa-
tion of the joint occurrence of D and H. Depending on the menu, this in-
volved calculating p(H]p(D\H), or simply "picking" p(H&D) (or the corre-
sponding values for the frequency format). Joint occurrence does not neglect
base rates; it neglects the false alarm rate in the standard menu and p(D] in

Table 6.3 Cognitive strategies in Study 1

Information format and menu

Probability Frequency
UiOgmuve r onnai /o OI
strategy equivalent Standard Short Standard Short Total total

Bayesian P(H\D] 69 126 204 221 620 34.9

Joint occurrence p(H&D) 39 97 20 97 253 14.3
Adjusted joint oc-
currence p(H&D) ±.05 64 55 119 6.7
Fisherian p(D\H) 67 36 103 5.8
Adjusted Fisherian p(D\H\ ±.05 32 19 51 2.9
Multiply all p(D)p(H&D) 79 12 91 5.1
Likelihood sub- p(D\H]-
traction p(D\ -H) 30 4 34 1.9
Base rate only p(H) 6 13 19 1.1
Less frequent
strategies (<1%
of total) 71 32 60 29 192 10.8
Not identified 119 52 89 32 292 16.5
Total 433 450 445 446 1,774" 100.0

Note: Numbers are absolute frequencies.

a. The sum of total answers is 1,774 rather than 1,800 (60 participants times 30 tasks) because of
some participants' refusals to answer and a few missing data.
 IMPROVING BAYESIAN REASONING 115

the short menu. Joint occurrence always underestimates the Bayesian poste-
rior unless p(D) = I. From participants' "write aloud" protocols, we learned
about a variant, which we call adjusted joint occurrence, in which the partic-
ipant starts with joint occurrence and adjusts it slightly (5 or fewer percentage
points).
Fisherian. Not all statisticians are Bayesians. Ronald A. Fisher, who invented
the analysis of variance and promoted significance testing, certainly was not.
In Fisher's (1955) theory of significance testing, an inference from data D to a
null hypothesis H0 is based solely on p(D\H0), which is known as the "exact
level of significance." The exact level of significance ignores base rates and
false alarm rates. With some reluctance, we labeled the second most frequent
non-Bayesian strategy—picking p(DlH) and ignoring everything else—"Fish-
erian." Our hesitation lay in the fact that it is one thing to ignore everything
else besides p(D\H], as Fisher's significance testing method does, and quite
another thing to confuse p(D\H] with p(HID). For instance, a p value of 1%
is often erroneously believed to mean, by both researchers and some statistical
textbook authors (see Chapter 13), that the probability of the null hypothesis
being true is 1%. Thus the term Fisherian refers to this widespread misinter-
pretation rather than to Fisher's actual ideas (we hope that Sir Ronald would
forgive us).
There exist several related accounts of the strategy for inferring p(H\D]
solely on the basis of p(D\H) Included in these are the tendency to infer "cue
validity" from "category validity" (Medin, Wattenmaker, & Michalski, 1987)
and the related thesis that people have spontaneous access to sample spaces
that correspond to categories (e.g., cancer) rather than to features associated
with categories (Gavanski & Hui, 1992). Unlike the Bayesian algorithms and
joint occurrence, the Fisherian strategy is menu specific: It cannot be elicited
from the short menu. We observed from participants' "write aloud" protocols
the use of a variant, which we call adjusted Fisherian, in which the participant
started with p(D I H) and then adjusted this value slightly (5 or fewer percent-
age points) in the direction of some other information.
Likelihood Subtraction. Jerzy Neyman and Egon S. Pearson challenged
Fisher's null-hypothesis testing. They argued that hypothesis testing is a de-
cision between (at least) two hypotheses that is based on a comparison of the
probability of the observed data under both, which they construed as the like-
lihood ratio p(D\H) I p(D\ -H}. We observed a version of the Neyman-Pear-
son method, the likelihood subtraction strategy, which computes p(D\H) —
p(D\ —H}. As in Neyman-Pearson hypotheses testing, this strategy makes no
use of prior probabilities and thus neglects base-rate information. The cogni-
tive strategy is menu specific (it can only be elicited by the standard menu)
and occurred predominantly in the probability format. In Robert Nozick's ac-
count, likelihood subtraction, also known as AR, is considered a measure of
evidential support (see Schum, 1994), and McKenzie (1994) has simulated the
performance of this and other non-Bayesian strategies.
116 ECOLOGICAL RATIONALITY

Others. There were cases of multiply all in the short menu (the logic of which
escaped us) and a few cases of base rate only in the standard menu (a pro-
portion similar to that reported in Gigerenzer, Hell, & Blank, 1988). We iden-
tified a total of 10.8% other strategies; these are not described here because
each was used in fewer than 1% of the solutions.

Summary of Study 1

The standard probability format—the information representation used in most

earlier studies—elicited 16% Bayesian inferences. When information was pre-
sented in a frequency format, this proportion jumped to 46% in the standard
menu and 50% in the short menu. The results of Study 1 are consistent with
Predictions 1,2, and 3. Frequency formats, in contrast to probability formats,
"invite" Bayesian reasoning, a result that is consistent with the computational
simplicity of Bayesian algorithms entailed by frequencies. Two of the three
major classes of non-Bayesian strategies our participants used—Fisherian and
likelihood subtraction—mimic statistical inferential algorithms used and dis-
cussed in the literature.

Study 2: Cognitive Strategies for Probability Formats

In this study we concentrated on probability and relative frequency rather than

on frequency formats. Thus, we explored cognitive strategies in the two for-
mats used by almost all previous studies on base-rate neglect. Our goal was to
test Prediction 4 and to provide another test of Prediction 2.
We used two formats, probability and relative frequency, and three menus:
standard, short, and hybrid. The hybrid menu displayed p(H), p(D\H), and
p(D), or the respective relative frequencies. The first two pieces come from the
standard menu, the third from the short menu. With the probability format and
the hybrid menu, a Bayesian algorithm amounts to solving the following equation:

The two formats and the three menus were mathematically interchangeable
and always entailed the same posterior probability. However, the Bayesian al-
gorithm for the short menu is computationally simpler than that for the stan-
dard menu, and the hybrid menu is in between; therefore the proportion of
Bayesian inferences should increase from the standard to the hybrid to the
short menu (extended Prediction 2). In contrast, the Bayesian algorithms for
the probability and relative frequency formats are computationally equivalent;
thus there should be no difference between these two formats (Prediction 4).

Method
Fifteen students from the fields of biology, linguistics, English studies, German
studies, philosophy, political science, and management at the University of
 IMPROVING BAYESIAN REASONING 117

Konstanz, Germany, served as participants. Eight were men, and 7 were

women; the median age was 22 years. They were paid for their participation
and studied in one group. None was familiar with Bayes's rule.
We used 24 problems, half from Study 1 and the other half new.4 For each
of the 24 problems, the information was presented in three menus, which re-
sulted in a total of 72 tasks. Each participant performed all 72 tasks. We ran-
domly assigned half of the problems to the probability format and half to the
relative frequency format; each participant thus answered half of the problems
in each format. All probabilities and relative frequencies were stated in per-
centages. The questions were always posed in terms of single-event probabil-
ities.
Six 1-hour sessions were scheduled, spaced equally over a 3-week interval.
In each session, 12 tasks were performed. Participants received the 72 tasks in
different orders, which were determined as follows: (a) Tasks that differed only
in menu were never given in the same session, and (b) the three menus were
equally frequent in every session. Within these two constraints, the 72 tasks
were randomly assigned to six groups of 12 tasks each, with the 12 tasks within
each group randomly ordered. These six groups were randomly assigned to
the six sessions for each participant. Finally, to control for possible order ef-
fects within the three (two) pieces of information (Kroznick, Li, & Lehman,
1990), we determined the order randomly for each participant.
The procedure was the same as in Study 1, except that we had participants
do an even larger number of inference problems and that we did not use the
"write aloud" instruction. However, participants could (and did) spontane-
ously "write aloud." After a student had completed all 72 tasks, he or she
received a new booklet. This contained copies of a sample of 6 tasks the stu-
dent had worked on, showing the student's probability estimates, notes, draw-
ings, calculations, and so forth. Attached to each task was a questionnaire in
which the student was asked, "Which information did you use for your esti-
mates?" and "How did you derive your estimate from the information? Please
describe this process as precisely as you can." Thus, in Study 2, we had only
limited "write aloud" protocols and after-the-fact interviews available. A spe-
cial prize of 25 deutsche marks was offered for the person with the best per-
formance.

Results
We could identify cognitive strategies in 67% of 1,080 probability judgments.
Table 6.4 shows the distribution of the cognitive strategies for the two formats
as well as for the three menus.

4. Study 2 was performed before Study 1 but is presented here second because it
builds on the central Study 1. In a few cases the numerical information in the problems
(e.g., German measles problem) was different in the two studies.
 Table 6.4 Cognitive strategies in Study 2
Information format Information menu

Cognitive Formal Relative % of

strategy equivalent frequency Probability Standard Hybrid Short Total total

Joint occurance p(H&D) 91 88 46 31 102 179 16.6

Bayesian p(H\D) 60 66 23 40 63 126 11.7
Fisherian p(D\H) 46 45 41 50 91 8.4
Adjusted Fisherian p(D\H)±.05 20 29 20 29 49 4.5
Multiply all p(D)p(H&D] 11 27 3 35 38 3.5
False alarm comple-
ment 1 - p(D\-H] 17 20 37 37 3.4
Likelihood subtraction p(D\H] -
p(D\-H) 19 9 28 28 2.6
Base rate only p(H) 14 10 14 10 24 2.2
Total negatives 1 - p(D) 10 7 9 8 17 1.6
Positives times base
rate p(D)p(H) 7 7 14 14 1.3
Postives times hit rate p(D)p(D\H) 4 9 13 13 1.2
Hit rate minus base rate p(D\H) -p(H) 6 5 3 8 11 1.0
Less frequent strategies
(<1% of total) 60 37 37 34 26 97 9.0
Not identified 175 181 111 119 126 356 33.0
Total 540 540 360 360 360 1,080 100.0
Note: Numbers are absolute frequencies.
 IMPROVING BAYESIAN REASONING 119
Bayesian Reasoning
Prediction 4: Relative frequency formats elicit the same (small) propor-
tion of Bayesian inferences as probability formats.
Table 6.4 shows that the number of Bayesian inferences is not larger for the
relative frequency format (60) than for the probability format (66). Consistent
with Prediction 4, the numbers are about the same. More generally, Bayesian
and non-Bayesian inferences were spread about equally between the two for-
mats. Therefore, we do not distinguish probability and relative frequency for-
mats in our further analysis.
Prediction 2 (extended to three menus): The proportion of Bayesian in-
ferences elicited by the probability format is lowest for the standard
menu followed in ascending order by the hybrid and short menus.
Study 2 allows for a second test of Prediction 2, now with three menus. Bay-
esian inferences almost doubled from the standard to the hybrid menu and
almost tripled in the short menu (Table 6.4). Thus the prediction holds again.
In Study 1, the standard probability menu elicited 16% Bayesian inferences,
as opposed to 28% for the short menu. In Study 2, the corresponding per-
centages of Bayesian inferences in probability formats were generally lower,
6.4% and 17.5%. What remained unchanged, however, was the difference be-
tween the two menus, about 12 percentage points, which is consistent with
Prediction 2.

Non-Bayesian Reasoning Study 2 replicated the three major classes of non-

Bayesian strategies identified in Study 1: joint occurrence, Fisherian, and like-
lihood subtraction. There was also a simpler variant of the last, the false alarm
complement strategy, which computes 1 - p(D I — H) and is a shortcut for like-
lihood subtraction when diagnosticity (the hit rate) is high. The other new
strategies—"total negatives," "positives times base rate," "positives times hit
rate," and "hit rate minus base rate"—were only or predominantly elicited by
the hybrid menu and seemed to us to be trial and error calculations. They
seem to have been used in situations in which the participants had no idea of
how to reason from the probability or relative frequency format and tried some-
how to integrate the information (such as by multiplying everything).

Are Individual Inferences Menu Dependent? Each participant worked on each

problem in three different menus. This allows us to see to what extent the
cognitive strategies and probability estimates of each individual were stable
across menus. The degree of menu dependence (the sensitivity of strategies
and estimates to changes in menu) in probability formats was striking. The
number of times the same strategy could be used across the three menus is
some number between 0 and 360 (24 problems times 15 participants). The
actual number was only 16 and consisted of 10 Bayesian and 6 joint occurrence
strategies. Thus in 96% of the 360 triples, the cognitive strategy was never the
same across the three menus.
120 ECOLOGICAL RATIONALITY

The reasoning of two participants, Rudiger and Oliver, illustrates this de-
pendence of thought on menu. They try to solve the German measles problem,
in which the task is to estimate the probability p(H I D ) of severe prenatal dam-
age in the child (H) if the mother had German measles during pregnancy (D).
In the standard menu, the information (probability expressed in percentages)
was p(H) = 0.5%, p(D\H) = 40%, and p(D\-H) = 0.01%; in the hybrid
menu, p(H) = 0.5%, p(D\H] = 40%, and p(D] = 0.21%; and in the short
menu, p(D] = 0.21% and p(D&H) = 0.2%.
Rudiger, age 22, management. In the standard menu, Rudiger focused
on p(D\H], explaining that because a child of an infected mother is at
such high risk (40%), his estimate would accordingly be high. He ad-
justed p(D\H) by 5% and estimated the posterior probability of severe
prenatal damage as 35% (adjusted Fisherian). In the hybrid menu, he
picked the base rate and estimated the same probability as 0.5% with
the argument that p(D I H ) and p(D] "are without significance" (base rate
only]. In the short menu, he picked p(H8tD) and estimated 0.2% because
"this is the information that specifies the probability of severe damage
in the child. The percentage of infected mothers, however, is irrelevant"
(joint occurrence).
Oliver, age 22, German literature. In the standard menu, Oliver stated
that the "correlation between not having damage and nevertheless
having measles," as he paraphrased p(D\ —H), was the only relevant in-
formation. He calculated 1 - p(D\ -H) = 99.99% and rounded to 100%,
which was his estimate (false alarm complement). In the hybrid menu,
he concluded that the only relevant information was the base rate of
severe prenatal damage, and his estimate consequently dropped to 0.5%
(base rate only). In the short menu, he determined the proportion of
severe damage and measles in all cases with German measles, which led
him to the Bayesian answer of 95.3%.
The thinking of Rudiger and Oliver illustrates how strongly cognitive strategies
can depend on the representation of information, resulting in estimates that
may vary as much as from 0.5% to 100% (as in Oliver's case). These cases also
reveal how helpless and inconsistent participants were when information was
represented in a probability or relative frequency format.

The Effect of Extensive Practice

With 72 inference problems per participant, Study 2 can answer the question
of whether mere practice (without feedback or instruction) increased the pro-
portion of Bayesian inferences. There was virtually no increase during the first
three sessions, which comprised 36 tasks. Only thereafter did the proportion
increase—from .04, .07, and .14 (standard, hybrid, and short menus, respec-
tively) in the first three sessions to .08, .14, and .21 in Sessions 4 through 6.
Thus extensive practice seems to be needed to increase the number of Bayesian
responses. In Study 1, with "only" 30 problems per participant, the proportion
increased slightly from .30 in the first session to .38 in the second. More gen-
 IMPROVING BAYESIAN REASONING 121

erally, with respect to all cognitive strategies, we found that when information
was presented in a frequency format, our participants became more consistent
in their use of strategies with time and practice, whereas there was little if any
improvement over time with probability formats.

Summary of Study 2

Our theoretical results were that the computational complexity of Bayesian

algorithms varied between the three probability menus but not between the
probability and relative frequency formats. Empirical tests showed that the
actual proportion of Bayesian inferences followed this pattern; the proportion
strongly increased across menus but did not differ between the probability and
the relative frequency formats, which is consistent with Predictions 2 and 4.

General Discussion

We return to our initial question: Is the mind, by design, predisposed against

performing Bayesian inference? The conclusion of 30 years of heuristics-and-
biases research would suggest as much. This previous research, however, has
consistently neglected Feynman's (1967) insight that mathematically equiva-
lent information formats need not be psychologically equivalent. An evolu-
tionary point of view suggests that the mind is tuned to frequency formats,
which is the information format humans encountered long before the advent
of probability theory. We have combined Feynman's insight with the evolu-
tionary argument and explored the computational implications: "Which com-
putations are required for Bayesian inference by a given information format
and menu?" Mathematically equivalent representations of information can en-
tail computationally different Bayesian algorithms. We have argued that infor-
mation representation affects cognitive strategies in the same way. We deduced
four novel predictions concerning when information formats and menus make
a difference and when they do not. Data from more than 2,800 individual
problem solutions are consistent with the predictions. Frequency formats made
many participants' inferences strictly conform (in terms of outcome and pro-
cess) to Bayes's rule without any teaching or instruction. These results were
found for a number of inferential problems, including classic demonstrations
of non-Bayesian inference such as the cab problem and the mammography
problem.
The results of the 15 problems in Study 1 constitute most of the data avail-
able today about Bayesian inference with frequency information. We know of
only a few studies that have looked at Bayesian inference through frequency
formats. Christensen-Szalanski and Beach (1982) sequentially presented symp-
tom and disease information for 100 patients and asked participants to estimate
^(disease I positive). Thus, their format was mixed: natural sampling of fre-
quencies with a single-event probability judgment (see also Gavanski & Hui,
1992). The means from the natural sampling condition conformed better to
122 ECOLOGICAL RATIONALITY

Bayes's rule than those from the standard probability version; however, only
means—and not individual judgments or processes—were analyzed. Cosmides
and Tooby (1996) constructed a dozen or so versions of the medical problem
presented by Casscells et al. (1978). They converted, piece by piece, probability
information into frequencies and showed how this increases, at the same pace,
the proportion of Bayesian answers. They reported that when the frequency
format was mixed—that is, when the information was represented in frequen-
cies but the single-point estimate was a single-event probability or vice versa—
the effect of the frequency format was reduced by roughly half. Their results
are consistent with our theoretical framework.
At the beginning of this chapter, we contrasted the belief of the Enlighten-
ment probabilists that the laws of probability theory were the laws of the mind
(at least for hommes eclaires) with the belief of the proponents of the
heuristics-and-biases program that the laws of probability are not the laws of
the mind. We side with neither view, nor with.those who have settled some-
where in between the two extremes. Both views are based on an incomplete
analysis: They focus on cognitive strategies, good or bad, without making the
connection between a strategy and the information format it has been designed
for. Through exploration of the computational consequences of an evolutionary
argument, a novel theoretical framework for understanding intuitive Bayesian
inference has emerged.
Why have so many experimental studies used the standard probability for-
mat? Part of the reason may be historical accident. There is nothing in Bayes's
rule that dictates whether the mathematical probabilities pertain to single
events or to frequencies, nor is the choice of format and menus specified by
the formal rules of probability. Thomas Bayes himself seemed not to have sided
with either single-event probabilities or frequencies. Like his fellow Enlight-
enment probabilists, he blurred the distinction between warranted degrees of
belief and objective frequencies by trying to combine the two (Barman, 1992).
Thus the experimental research on Bayesian inference could as well have
started with frequency representations, if not for the historical accident that it
became tied to Savage's (1954) agenda of bringing singular events back into
the domain of probability theory. For instance, if psychological research had
been inspired by behavioral ecology, foraging theory, or other ecological ap-
proaches to animal behavior in which Bayes's rule figures prominently (e.g.,
Stephens & Krebs, 1986), then the information format used in human studies
might have been frequencies from the very beginning.
We would like to emphasize that our results hold for an elementary form
of Bayesian inference, with binary hypotheses and data. Pregnancy tests, mam-
mograms, HIV tests, and the like are everyday examples where this elementary
form of inference is of direct relevance. However, there exist other situations
in which hypotheses, data, or both are multinomial or continuous and where
there is not only one datum, but several. Massaro (1998), for instance, conjec-
tured that when there are two or more pieces of evidence or cues—such as
two medical test results—a representation in natural frequencies would no
longer help to improve Bayesian reasoning. Krauss, Martignon, and Hoffrage
 IMPROVING BAYESIAN REASONING 123

(1999), in contrast, have shown that natural sampling can be generalized to

two cues and observed about the same improvement in Bayesian reasoning
with two cues as in the experiments reported here. However, beyond some
number of cues, Bayesian calculations, even using natural frequencies, will
become extremely complicated or intractable. Here, models of bounded ra-
tionality can take over where intuitive Bayesian reasoning leaves off. The fol-
lowing two chapters deal with fast and frugal heuristics that can survive in
complex real-world environments by virtue of their simplicity and robustness.
This page intentionally left blank
 Ill
BOUNDED RATIONALITY

I rying to change a university is like trying to move a cemetery—this meta-

phor is easily understood. But why would Herbert A. Simon have pro-
posed a pair of scissors as a metaphor for bounded rationality? To under-
stand the power of human intelligence, one needs to analyze the match
between cognitive strategies and the structure of environments. Together
they are like a pair of scissors, each blade of little use on its own but ef-
fective in concert with the other. The program of "bounded rationality" is
the third research strategy explored in this book. Models of bounded ra-
tionality address the following question: How do people make decisions in
the real world, where time is short, knowledge lacking, and other re-
sources limited?
The program of bounded rationality confronts us with two challenging
questions. First, what are the simple heuristics that people use given scarce
resources (the cognitive blade)? Second, what structures of information can
these heuristics exploit, that is, in what environments do they succeed and
fail? The second question concerns the ecological rationality of heuristics
(the fit between the cognitive and environmental blades), which emphasizes
the interface between bounded and ecological rationality. By analyzing the
match between heuristic and environment, we can predict how fast, frugal,
and accurate a heuristic will be.
The two chapters in this section focus on the first question, although they
also contain the seeds of an answer to the second question, which is ex-
plored in depth in two books, Simple Heuristics That Make Us Smart (Giger-
enzer, Todd, & the ABC Research Group, 1999) and Bounded Rationality:
The Adaptive Toolbox (Gigerenzer & Selten, 2000). Chapter 7 contains the
essence (but not the name) of the "Take The Best" heuristic, which makes
decisions on the basis of only one good reason and ignores the rest. The sur-
prisingly good performance of this heuristic, first demonstrated in Chapter 8,
initially elicited one or both of two reactions: "It cannot be!" or "We knew it
all along!"
126 BOUNDED RATIONALITY

I have been asked, How did you discover these fast and frugal heuristics?
In fact, most of the credit goes to my students and colleagues in our interdis-
ciplinary research group, the Center for Adaptive Behavior and Cognition.
We benefited from a healthy mixture of persistence and luck. The discovery
of the recognition heuristic and the less-is-more effect illustrates how insight
can emerge from failure to accomplish something else. In an article on prob-
abilistic mental model (PMM) theory (Chapter 7), Ulrich Hoffrage, Heinz
Kleinbolting, and I derived a bold prediction about the "hard-easy" effect,
in which people are overconfident about their ability to solve hard but not
easy questions. The prediction was that the hard-easy effect would disap-
pear when both kinds of questions are representatively sampled.
After the article appeared, we designed an experiment to test the predic-
tion. We needed a hard and an easy set of questions and a domain from
which to draw representative samples. At that time, I was teaching at the
University of Salzburg, a cheerful architectural blend of postmodern Bau-
haus and Austrian Empire-style marble and gold. Common sense dictated
that German cities, about which students there knew a lot, would be an easy
set, and American cities, about which they knew comparatively little, would
be a hard set. How could it be any other way? We drew 100 random pairs of
cities from the 75 largest German cities, such as Bielefeld and Heidelberg,
and another 100 random pairs of the 75 largest American cities, such as San
Diego and San Antonio. The task was to judge which of two cities has the
larger population. When we saw the results, we could not believe our eyes.
These German-speaking students gave slightly more accurate answers for the
American cities (76.0%) than for the German cities (75.6%). How could they
have made as many correct judgments in a domain about which they knew
little as in one about which they knew a lot?
Salzburg has excellent restaurants. One night our research group had din-
ner at one of them to mourn the failed experiment—we could not test the
prediction because we failed to generate a hard and an easy set of questions.
As we tried in vain to make sense of the counterintuitive result, our col-
league Anton Kiihberger politely remarked: "Why don't you look in your
PMM paper? The answer is there." What an embarrassing moment: He was
right. Our paper said that having heard of one city and not of the other is a
cue that the first city is larger—a fast and frugal strategy that we later named
the "recognition heuristic." How could this heuristic explain our puzzling
result? Because most of the students in Salzburg had heard of all the largest
German cities, they could not use the recognition heuristic in that set. But
when it came to American cities, many of which they had never heard, they
could use it. By exploiting the wisdom in missing knowledge, the recogni-
tion heuristic can lead to highly accurate judgments when lack of recogni-
tion is not random but systematically correlated with the criterion (here,
population size). Lack of recognition can be highly informative.
So the less-is-more effect was discovered accidentally when it ruined an
experiment. After I left the architecturally playful University of Salzburg for
the stolidly gothic University of Chicago, I met Daniel Goldstein, with whom
 BOUNDED RATIONALITY 127

I began to study the recognition heuristic and the less-is-more effect syste-
matically (Goldstein & Gigerenzer, 1999). Meanwhile, others succeeded in
testing and confirming our original prediction about the hard-easy effect
(Juslin, 1993; Juslin, Winman, & Olsson, 2000; Klayman et al, 1999). This
story illustrates how scientific discovery can come by failing to do one thing
and yet achieving another.
This page intentionally left blank
 7

Probabilistic Mental Models

Do people think they know more than they really do? In the last 20 years,
cognitive psychologists have amassed a large and apparently damning body of
experimental evidence on overconfidence in knowledge, evidence that is in
turn part of an even larger and more damning literature on so-called cognitive
biases. The cognitive bias research claims that people are naturally prone to
making mistakes in reasoning and memory, including the mistake of over-
estimating their knowledge. In this chapter, we propose a new theoretical
model for confidence in knowledge based on the more charitable assumption
that people are good judges of the reliability of their knowledge, provided that
the knowledge is representatively sampled from a specified reference class. We
claim that this model both predicts new experimental results (that we have
tested) and explains a wide range of extant experimental findings on confi-
dence, including some perplexing inconsistencies.
Moreover, it is the first theoretical framework to integrate the two most striking
and stable effects that have emerged from confidence studies—the overconfidence
effect and the hard—easy effect—and to specify the conditions under which these
effects can be made to appear, disappear, and even invert. In most recent studies
(including our own, reported herein), participants are asked to choose between
two alternatives for each of a series of general-knowledge questions. Here is atyp-
ical example: "Which city has more inhabitants? (a) Hyderabad or (b) Islamabad."
Participants choose what they believe to be the correct answer and then are di-
rected to specify their degree of confidence (usually on a 50%-100% scale) that
their answer is indeed correct. After the participants answer many questions of
this sort, the responses are sorted by confidence level, and the relative frequencies
of correct answers in each confidence category are calculated. The overconfidence
effect occurs when the confidence judgments are larger than the relative frequen-
cies of the correct answers; the hard-easy effect occurs when the degree of over-
confidence increases with the difficulty of the questions, where the difficulty is
measured by the percentage of correct answers.
The work on which this chapter is based was coauthored with U. Hoffrage and
H. Kleinbolting.

129
130 BOUNDED RATIONALITY

Both effects seem to be stable. Fischhoff (1982) reviewed the attempts to

eliminate overconfidence by numerous "debiasing methods," such as giving
rewards, clarifying instructions, warning participants in advance about the
problem, and using better response modes—all to no avail. He concluded that
these manipulations "have so far proven relatively ineffective," and that over-
confidence was "moderately robust" (p. 440). Von Winterfeldt and Edwards
(1986, p. 539) agreed that "overconfidence is a reliable, reproducible finding."
Yet these robust phenomena still await a theory. In particular, we lack a com-
prehensive theoretical framework that explains both phenomena, as well as
the various exceptions reported in the literature, and integrates the several
local explanatory attempts already advanced. That is the aim of this chapter.
It consists of three parts: (a) an exposition of the proposed theory of probabi-
listic mental models (PMM theory), including predictions of new experimental
findings based on the theory; (b) a report of our experimental tests confirming
these predictions; and (c) an explanation of apparent anomalies in previous
experimental results by means of PMMs.

PMM Theory

This theory deals with spontaneous confidence—that is, with an immediate

reaction, not the product of long reflection. Figure 7.1 shows a flow chart of
the processes that generate confidence judgments in two-alternative general-
knowledge tasks. There are two strategies. When presented with a two-
alternative confidence task, the participant first attempts to construct what we
call a local mental model (local MM) of the task. This is a solution by memory
and elementary logical operations. If this fails, a PMM is constructed that goes
beyond the structure of the task in using probabilistic information from a nat-
ural environment.
For convenience, we illustrate the theory using a problem from the follow-
ing experiments: "Which city has more inhabitants? (a) Heidelberg or (b)
Bonn." As explained earlier, the participants' task is to choose a or b and to
give a numerical judgment of their confidence (that the answer chosen is cor-
rect).

Local MM

We assume that the mind first attempts a direct solution that could generate
certain knowledge by constructing a local MM. For instance, a participant may
recall from memory that Heidelberg has a population between 100,000 and
200,000, whereas Bonn has more than 290,000 inhabitants. This is already
sufficient for the answer "Bonn" and a confidence judgment of 100%. In gen-
eral, a local MM can be successfully constructed if (a) precise figures can be
retrieved from memory for both alternatives, (b) intervals that do not overlap
can be retrieved, or (c) elementary logical operations, such as the method of
exclusion, can compensate for missing knowledge. Figure 7.2 illustrates a sue-
Figure 7.1 Cognitive processes in solving a two-alternative general-knowledge task. MM = mental model;
PMM — probabilistic mental model.
132 BOUNDED RATIONALITY

Figure 7.2 Local mental model of a two-alternative general-knowledge task.

cessful local MM for the previous example. Now consider a task in which the
target variable is not quantitative (such as the number of inhabitants) but is
qualitative: "If you see the nationality letter P on a car, is it from Poland or
Portugal?" Here, either direct memory about the correct answer or the method
of exclusion is sufficient to construct a local MM. The latter is illustrated by
a participant reasoning "Since I know that Poland has PL it must be Portugal"
(Allwood & Montgomery, 1987, p. 370).
The structure of the task must be examined to define more generally what
is referred to as a local MM. The task consists of two objects, a and b (alter-
natives), and a target variable t. First, a local MM of this task is local; that is,
only the two alternatives are taken into account, and no reference class of
objects is constructed (see the following discussion). Second, it is direct; that
is, it contains only the target variable (e.g., number of inhabitants), and no
probability cues are used. Third, no inferences besides elementary operations
of deductive logic (such as exclusion) occur. Finally, if the search is successful,
the confidence in the knowledge produced is evaluated as certain. In these
respects, our concept of a local MM is similar to what Johnson-Laird (1983,
pp. 134-142) called a "mental model" in syllogistic inference.
A local MM simply matches the structure of the task; there is no use of the
probability structure of an environment and, consequently, no frame for in-
ductive inference as in a PMM. Because memory can fail, the "certain" knowl-
edge produced can sometimes be incorrect. These failures contribute to the
amount of overconfidence to be found in 100%-confident judgments.

PMM

Local MMs are of limited success in general-knowledge tasks and in most nat-
ural environments, although they seem to be sufficient for solving some syl-
logisms and other problems of deductive logic (see Johnson-Laird, 1983). If no
local MM can be activated, it is assumed that a PMM is constructed next. A
PMM solves the task by inductive inference, and it does so by putting the
specific task into a larger context. A PMM connects the specific structure of
the task with a probability structure of a corresponding natural environment
(stored in long-term memory). In our example, a natural environment could
be the class of all cities in Germany with a set of variables defined on this
 PROBABILISTIC MENTAL MODELS 133

class, such as the number of inhabitants. This task selects the number of in-
habitants as the target and the variables that covary with this target as the cues.
A PMM is different from a local MM in several respects. First, it contains
a reference class of objects that includes the objects a and b. Second, it uses
a network of variables in addition to the target variable for indirect inference.
Thus it is neither local nor direct. These two features also change the third
and fourth aspects of a local MM. Probabilistic inference is part of the cognitive
process, and uncertainty is part of the outcome.

Reference Class
The term reference class refers to the class of objects or events that a PMM
contains. In our example, the reference class "all cities in Germany" may be
generated. To generate a reference class means to generate a set of objects
known from a person's natural environment that contains objects a and b.
The reference class determines which cues can function as probability cues
for the target variable and what their cue validities are. For instance, a valid
cue in the reference class "all cities in Germany" would be the soccer-team
cue; that is, whether a city's soccer team plays in the German soccer Bundes-
liga, in which the 18 best teams compete. Cities with more inhabitants are
more likely to have a team in the Bundesliga. The soccer-team cue would not
help in the Hyderabad-Islamabad task, which must be solved by a PMM con-
taining a different reference class with different cues and cue validities.

Probability Cues
A PMM for a given task contains a reference class, a target variable, probability
cues, and cue validities. A variable is a probability cue Q (for a target variable
in a reference class R) if the probability p(a) of a being correct is different from
the conditional probability of a being correct, given that the values of a and b
differ on C,. If the cue is a binary variable such as the soccer-team cue, this
condition can be stated as follows:

where aQb signifies the relation of a and b on the cue C, (e.g., a has a soccer
team in the Bundesliga, but b does not) and p(a I aCtb; R] is the cue validity of
Ct in R.
Thus cue validities are thought of as conditional probabilities, following
Rosch (1978) rather than Brunswik (1955), who defined his "cue utilizations"
as Pearson correlations. Conditional probabilities need not be symmetric as
correlations are. This allows the cue to be a better predictor for the target than
the target is for the cue, or vice versa. Cue validity is a concept in the PMM,
whereas the corresponding concept in the environment is ecological validity
(Brunswik, 1955), which is the true relative frequency of any city having more
inhabitants than any other one in R if aCtb. For example, consider the reference
class all cities in Germany with more than 100,000 inhabitants. The ecological
134 BOUNDED RATIONALITY

validity of the soccer-team cue here is .91 (calculated for 1988/1989 for what
then was West Germany). That is, if one checked all pairs in which one city
a has a team in the Bundesliga but the other city b does not, one would find
that in 91% of these cases city a has more inhabitants.

Vicarious Functioning

Probability cues are generated, tested, and, if possible, activated. We assume

that the order in which cues are generated is not random; in particular, we
assume that the order reflects the hierarchy of cue validities. For the reference
class all cities in Germany, the following cues are examples that can be gen-
erated: (a) the soccer-team cue; (b) whether one city is a state capital and the
other is not (state capital cue); (c) whether one city is located in the Ruhrgebiet,
the industrial center of Germany, and the other in largely rural Bavaria (in-
dustrial cue); (d) whether the letter code that identifies a city on a license plate
is shorter for one city than for the other (large cities are usually abbreviated
by only one letter, smaller cities by two or three; license plate cue); and (e)
whether one has heard of one city and not of the other (recognition cue). Con-
sider now the Heidelberg-Bonn problem again. The first probability cue is
generated and tested to see whether it can be activated for that problem. Be-
cause neither of the two cities has a team in the Bundesliga, the first cue does
not work.
In general, with a binary cue and the possibility that the participant has no
knowledge, there are nine possibilities (see Figure 7.3). In only two of these
can a cue be activated. In all other cases, the cue is useless (although one could
further distinguish between the four known—unknown cases and the three re-
maining cases). If a cue cannot be activated, then a further cue is generated

Figure 7.3 Two conditions in which a cue can be activated.

 PROBABILISTIC MENTAL MODELS 135

and tested. In the Heidelberg-Bonn task, none of the five cues cited earlier can
in fact be activated. Finally, one cue may be generated that can be activated,
such as whether one city is the capital of the country and the other is not
(capital cue). This cue has a small probability of being activated—a small ac-
tivation rate in R (because it applies only to pairs that include Bonn]—and it
does not have a particularly high cue validity in R because it is well known
that Bonn is not exactly London or Paris.
The Heidelberg-Bonn problem illustrates that probability cues may have
small activation rates in R, and as a consequence, several cues may have to be
generated and tested before one is found that can be activated. The capital cue
that can be activated for the Heidelberg-Bonn comparison may fail for the next
problem, for instance a Heidelberg-Gottingen comparison. Cues can substitute
for one another from problem to problem, a process that Brunswik (1955]
called "vicarious functioning."

End of Cue Generation and Testing Cycle

If (a) the number of problems is large or other kinds of time pressure apply
and (b) the activation rate of cues is rather small, then one can assume that
the cue generation and testing cycle ends after the first cue that can be acti-
vated has been found. Both conditions seem to be typical for general-
knowledge questions. For instance, even when participants were explicitly in-
structed to produce all possible reasons for and against each alternative, they
generated only about three on the average and four at most (Koriat, Lichten-
stein, & Fischhoff, 1980). If no cue can be activated, we assume that choice is
made randomly, and "confidence 50%" is chosen.

Choice of Answer and Confidence Judgment

Choice of answer and confidence judgment are determined by the cue validity.
Choice follows the rule:

If a is chosen, the confidence that a is correct is given by the cue validity:

Note that the assumption that confidence equals cue validity is not arbitrary;
it is both rational and simple in the sense that good calibration is to be ex-
pected if cue validities correspond to ecological validities. This holds true even
if only one cue is activated.
Thus choice and confidence are inferred from the same activated cue. Both
are expressions of the same conditional probability. Therefore, they need not
be generated in the temporal sequence choice followed by confidence. The
latter is, of course, typical for actual judgments and often enforced by the
instructions in confidence studies.
136 BOUNDED RATIONALITY

Confidence in the Long Run and Confidence

in Single Events

Until now, only confidence in single events—such as the answer "Bonn" is

correct—has been discussed. Confidence in one's knowledge can also be ex-
pressed with respect to sequences of answers or events, such as "How many
of the last 50 questions do you think you answered correctly?" This distinction
is parallel to that between probabilities of single events and relative frequen-
cies in the long run—a distinction that is fundamental to all discussions on
the meaning of probability (see Gigerenzer et al., 1989). Probabilities of single
events (confidences) and relative frequencies are not the same for many
schools of probability, and we argue that they are not evaluated by the same
cognitive processes either.
Consider judgments of frequency. General-knowledge tasks that involve a
judgment of the frequency of correct answers (frequency tasks) can rarely be
answered by constructing a local MM. The structure of the task contains one
sequence of N questions and answers, and the number of correct answers is
the target variable. Only limiting cases, such as small N (i.e., if only a few
questions are asked) combined with the belief that all answers were correct,
may allow one to solve this task by a local MM. Again, to construct a local
MM of the task means that the mental model consists of only the local se-
quence of total N answers (no reference class), and because one attempts to
solve the task by direct access to memory about the target variable, no network
of probability cues is constructed.
Similarly, a PMM of a frequency task is different from a PMM of a confi-
dence task. A confidence task about city size in Germany has "cities in Ger-
many" as a reference class; however, a task that involves judgments of fre-
quencies of correct answers in a series of N questions about city size has a
different reference class: Its reference class will contain series of similar ques-
tions in similar testing situations. Because the target variable also differs (num-
ber of correct answers instead of number of inhabitants), the PMM of a fre-

Table 7.1 Probabilistic mental models for confidence task versus frequency
task: Differences between target variables, reference classes, and probability
cues
PMM Confidence task Frequency task

Target variable Number of inhabitants Number of correct answers

Reference class Cities in Germany Sets of general-knowledge questions
in similar testing situations
Probability cues For example, soccer-team For example, base rates of previous
cue or state capital cue performance or average
confidence in N answers
Note: For illustration, questions of the Heidelberg-Bonn type are used. PMM = probabilistic mental
model.
 PROBABILISTIC MENTAL MODELS 137

quency task will also contain different cues and cue validities. For instance,
base rates of performance in earlier general knowledge or similar testing sit-
uations could serve as a probability cue for the target variable. Again, our basic
assumption is that a PMM connects the structure of the task with a known
structure of the participant's environment.
Table 7.1 summarizes the differences between PMMs that are implied by
the two different tasks. Note that in our account, both confidences in a single
event and judgments of frequency are explained by reference to experienced
frequencies. However, these frequencies relate to different target variables and
reference classes. We use this assumption to predict systematic differences
between these kinds of judgments.

Adaptive PMMs and Representative Sampling

A PMM is an inductive device that uses the "normal" life conditions in known
environments as the basis for induction. How well does the structure of prob-
ability cues defined on R in a PMM represent the actual structure of probability
cues in the environment? This question is also known as that of "proper cog-
nitive adjustment" (Brunswik, 1964, p. 22). If the order of cue validities
roughly corresponds to that of the ecological validities, then the PMM is well
adapted to a known environment. In Brunswik's view, cue validities are
learned by observing the frequencies of co-occurrences in an environment.
A large literature exists that suggests that (a) memory is often (but not al-
ways) excellent in storing frequency information from various environments
and (b) the registering of event occurrences for frequency judgments is a fairly
automatic cognitive process requiring very little attention or conscious effort
(e.g., Gigerenzer, 1984; Hasher, Goldstein, & Toppino, 1977; Howell & Burnett
1978; Sedlmeier, Hertwig, & Gigerenzer, 1998; Zacks, Hasher, & Sanft, 1982).
Hasher and Zacks (1979) concluded that frequency of occurrence, spatial lo-
cation, time, and word meaning are among the few aspects of the environment
that are encoded automatically and that encoding of frequency information is
"automatic at least in part because of innate factors" (p. 360). In addition,
Hintzman, Nozawa, and Irmscher (1982) proposed that frequencies are stored
in memory in a nonnumerical analog mode.
Whatever the mechanism of frequency encoding, we use the following as-
sumption for deriving our predictions: If participants had repeated experience
with a reference class, a target variable, and cues in their environment, we
assume that cue validities correspond well to ecological validities. (This holds
true for the average in a group of participants, but individual idiosyncrasies
in learning the frequency structure of the environment may occur.) This is a
bold assumption made in ignorance of potential deviations between specific
cue validities and ecological validities. If such deviations existed and were
known, predictions by PMM theory could be improved. The assumption, how-
ever, derives support from both the literature on automatic frequency process-
ing and a large body of neo-Brunswikian research on the correspondence be-
tween ecological validities and cue utilization (the latter of which corresponds
138 BOUNDED RATIONALITY

to our cue validities; e.g., Arkes & Hammond, 1986; Armelius, 1979; Brehmer
& Joyce, 1988; MacGregor & Slovic, 1986).
Note that this adaptiveness assumption does not preclude that individuals
(as well as the average participant) err. Errors can occur even if a PMM is
highly adapted to a given environment. For instance, if an environment is
changing or is changed in the laboratory by an experimenter, an otherwise
well-adapted PMM may be suboptimal in a predictable way.
Brunswik's notion of "representative sampling" is important here. If a per-
son experienced a representative sample of objects from a reference class, one
can expect his or her PMM to be better adapted to an environment than if he
or she happened to experience a skewed, unrepresentative sample.
Representative sampling is also important in understanding the relation be-
tween a PMM and the task. If a PMM is well adapted, but the set of objects
used in the task (questions) is not representative of the reference class in the
environment, performance in tasks will be systematically suboptimal.
To avoid confusion with terms such as calibration, we will use the term
adaptation only when we are referring to the relation between a PMM and a
corresponding environment—not, however, for the relation between a PMM
and a task.

Predictions

A concrete example can help motivate our first prediction. Two of our col-
leagues, K and O, are eminent wine tasters. K likes to make a gift of a bottle
of wine from his cellar to Friend O, on the condition that O guesses what
country or region the grapes were grown in. Because O knows the relevant
cues, O can usually pick a region with some confidence. O also knows that K
sometimes selects a quite untypical exemplar from his ample wine cellar to
test Friend O's limits. Thus, for each individual wine, O can infer the proba-
bility that the grapes ripened in, say, Portugal as opposed to South Africa with
considerable confidence from his knowledge about cues. In the long run, how-
ever, O nevertheless expects the relative frequency of correct answers to be
lower because K occasionally selects unusual items.
Consider tests of general knowledge, which share an important feature with
the wine-tasting situation: Questions are selected to be somewhat difficult and
sometimes misleading. This practice is common and quite reasonable for test-
ing people's limits, as in the wine-tasting situation. Indeed, there is apparently
not a single study on confidence in knowledge in which a reference class has
been defined and a representative (or random) sample of general-knowledge
questions has been drawn from this population. For instance, consider the
reference class "metropolis" and the geographical north-south location as the
target variable. A question like "Which city is farther north? (a) New York or
(b) Rome" is likely to appear in a general-knowledge test (almost everyone gets
it wrong), whereas a comparison between Berlin and Rome is not.
 PROBABILISTIC MENTAL MODELS 139

The crucial point is that confidence and frequency judgments refer to dif-
ferent kinds of reference classes. A set of questions can be representative with
respect to one reference class and, at the same time, selected with respect to
the other class. Thus, a set of 50 general-knowledge questions of the city type
may be representative for the reference class "sets of general-knowledge ques-
tions" but not for the reference class "cities in Germany" (because city pairs
have been selected for being difficult or misleading). Asking for a confidence
judgment summons up a PMM on the basis of the reference class "cities in
Germany"; asking for a frequency judgment summons up a PMM on the basis
of the reference class "sets of general-knowledge questions." The first predic-
tion can now be stated.
1. Typical general-knowledge tasks elicit both overconfidence and ac-
curate frequency judgments.
By "typical" general-knowledge tasks we refer to a set of questions that is
representative for the reference class "sets of general-knowledge questions."
This prediction is derived in the following way: If (a) PMMs for confidence
tasks are well adapted to an environment containing a reference class R (e.g.,
all cities in Germany) and (b) the actual set of questions is not representative
for R but selected for difficult pairs of cities, then confidence judgments exhibit
overconfidence. Condition A is part of our theory (the simplifying assumption
we just made), and Condition B is typical for the general-knowledge questions
used in studies on confidence as well as in other testing situations.
If (a) PMMs for frequency-of-correct-answer tasks are well adapted with
respect to an environment containing a reference class R (e.g., the set of all
general-knowledge tests experienced earlier), and (b) the actual set of questions
is representative for R, then frequency judgments are expected to be accurate.
Again, Condition A is part of our theory, and Condition B will be realized in
our experiments by using a typical set of general-knowledge questions.
Taken together, the prediction is that the same person will exhibit over-
confidence when asked for her confidence that a particular answer is correct
and accurate estimates when asked for a judgment of the frequency of correct
answers. This prediction is shown by the two points on the left side of Figure
7.4. This prediction cannot be derived from any of the previous accounts of
overconfidence.
To introduce the second prediction, we return to the wine-tasting story.
Assume that K changes his habit of selecting unusual wines from his wine
cellar and instead buys a representative sample of French red wines and lets
O guess from what region they come. However, K does not tell O about the
new sampling technique. O's average confidence judgments will now be close
to the proportion of correct answers. In the long run, O nevertheless expects
the proportion of correct answers to be smaller, still assuming the familiar
testing situation in which wines were selected, not randomly sampled. Thus
O's frequency judgments will show underestimation.
Consider now a set of general-knowledge questions that is a random sample
from a defined reference class in the participant's natural environment. We use
140 BOUNDED RATIONALITY

Figure 7.4 Predicted differences between confidence and frequency judgments-

the confidence-frequency effect.

the term natural environment to denote a knowledge domain familiar to the

participants in the study. This is a necessary (although not sufficient) condition
to assume that PMMs are, on the average, well adapted. In the experiments
reported herein, we used West German participants and the reference class
"all cities with more than 100,000 inhabitants in West Germany." (The study
was conducted before the unification of Germany.) The second prediction is
about this situation:
2. If the set of general-knowledge tasks is randomly sampled from a nat-
ural environment, we expect overconfidence to be zero, but frequency
judgments to exhibit underestimation.
Derivation is as before: If PMMs for confidence tasks are well adapted with
respect to R, and the actual set of questions is a representative sample from
R, then overconfidence is expected to disappear. If PMMs for frequency-of-
correct-answers tasks are well adapted with respect to R', and the actual set
of questions is not representative for R', then frequency judgments are ex-
pected to be underestimations of true frequencies.
Again, this prediction cannot be derived from earlier accounts. Figure 7.4
shows Predictions 1 and 2. The predicted differences between confidence and
frequency judgments are referred to as the confidence-frequency effect.
Testing these predictions also allows for testing the assumption of well-
adapted PMMs for the confidence task. Assume that PMMs are not well
 PROBABILISTIC MENTAL MODELS 141

adapted. Then a representative sample of city questions should not generate

zero overconfidence but rather over- or underconfidence, depending on
whether cue validities overestimate or underestimate ecological validities.
Similarly, if PMMs for frequency judgments are not well adapted, frequency
judgments should deviate from true frequencies in typical general-knowledge
tasks. Independent of the degree of adaptation, however, the confidence—fre-
quency effect should emerge, but the curves in Figure 7.4 would be transposed
upward or downward.
We turn now to the standard way in which overconfidence has been dem-
onstrated in previous research, comparing confidence levels with relative fre-
quencies of correct answers at each confidence level. This standard compari-
son runs into a conceptual problem well known in probability theory and
statistics: A discrepancy between subjective probabilities in single events (i.e.,
the confidence that a particular answer is correct) and relative frequencies in
the long run is not a bias in the sense of a violation of probability theory, as
is clear from several points of view within probability theory. For instance, for
a frequentist such as Richard von Mises (1928/1957), probability theory is
about frequencies (in the long run), not about single events. According to this
view, the common interpretation of overconfidence as a bias is based on com-
paring apples with oranges. What if that conceptual problem is avoided and,
instead, the relative frequency of correct answers in each confidence category
is compared with the estimated relative frequency in each confidence category?
PMM theory makes an interesting prediction for this situation, following the
same reasoning as for the frequency judgments in Predictions 1 and 2 (which
were estimated frequency-of-correct answers in a series of N questions,
whereas estimated relative frequencies in each confidence category are the
concern here):
3. Comparing estimated relative frequencies with true relative frequen-
cies of correct answers makes overestimation disappear.
More precisely, if the set of general-knowledge questions is selected, over- or
underestimation is expected to be zero; if the set is randomly sampled, un-
derestimation is expected. Thus PMM theory predicts that the distinction be-
tween confidence and relative frequency is psychologically real, in the sense
that participants do not believe that a confidence judgment of X% implies a
relative frequency of X%, and vice versa. We know of no study on over-
confidence that has investigated this issue. Most have assumed instead that
there is, psychologically, no difference.
Prediction 4 concerns the hard-easy effect, which says that overconfidence
increases when questions get more difficult (e.g., Lichtenstein & Fischhoff,
1977). The effect refers to confidence judgments only, not to frequency judg-
ments. In our account, the hard-easy effect is not simply a function of dif-
ficulty. Rather, it is a function of difficulty and a separate dimension, selected
versus representative sampling. (Note that the terms hard and easy refer to
the relative difficulty of two samples of items, whereas the terms selected
and representative refer to the relation between one sample and a reference
142 BOUNDED RATIONALITY

class in the person's environment.) PMM theory specifies conditions under

which the hard—easy effect occurs, disappears, and is reversed. A reversed
hard-easy effect means that overconfidence decreases when questions are
more difficult.
In Figure 7.5, the line descending from H to E represents a hard—easy effect:
Overconfidence in the hard set is larger than in the easy set. The important
distinction (in addition to hard vs. easy) is whether a set was obtained by
representative sampling or was selected. For instance, assume that PMMs are
well adapted and that two sets of tasks differing in percentage correct (i.e.,
in difficulty) are both representative samples from their respective reference
classes. In this case, one would expect all points to be on the horizontal zero-
overconfidence line in Figure 7.5 and the hard-easy effect to be zero. More
generally:
4. // two sets, hard and easy, are generated by the same sampling pro-
cess (representative sampling or same deviation from representative),
the hard-easy effect is expected to be zero.

Figure 7.5 Predicted reversal of the hard-easy effect. H = hard; E = easy.

 PROBABILISTIC MENTAL MODELS 143

If sampling deviates in both the hard and the easy set equally from represen-
tative sampling, points will lie on a horizontal line parallel to the zero-
overconfidence line.
Now consider the case that the easy set is selected from a corresponding
reference class (e.g., general-knowledge questions), but the hard set is a rep-
resentative sample from another reference class (denoted as H' in Figure 7.5).
One then would predict a reversal of the hard-easy effect, as illustrated in
Figure 7.5 by the double line from E to H'.

5. Given two sets, one a representative sample from a reference class in

a natural environment, the other selected from another reference class
for being difficult, if the representative set is harder than the selected
set, then the hard-easy effect is reversed.

In the next section, Predictions 1, 2, and 3 are tested in two experiments; in

the Explaining Anomalies in the Literature section, Predictions 4 and 5 are
checked against results in the literature.

Experiment 1

Method
Two sets of questions were used, which we refer to as the representative and
the selected set. The representative set was determined in the following way.
We used as a reference class in a natural environment (an environment known
to our participants) the set of all cities in West Germany with more than
100,000 inhabitants. There were 65 cities (Statistisches Bundesamt, 1986).
From this reference class, a random sample of 25 cities was drawn, and all
pairs of cities in the random sample were used in a complete paired compar-
ison to give 300 pairs. No selection occurred. The target variable was the num-
ber of inhabitants, and the 300 questions were of the following kind: "Which
city has more inhabitants? (a) Solingen or (b) Heidelberg." We chose city ques-
tions for two reasons. First, and most important, this content domain allowed
for a precise definition of a reference class in a natural environment and for
random sampling from this reference class. The second reason was for com-
parability. City questions have been used in earlier studies on overconfidence
(e.g., Keren, 1988; May, 1987).
In addition to the representative set, a typical set of 50 general-knowledge
questions, as in previous studies, was used. Two examples are "Who was born
first? (a) Buddha or (b) Aristotle" and "When was the zipper invented? (a)
before 1920 or (b) after 1920."
After each answer, the participant gave a confidence judgment (that this
particular answer was correct). Two kinds of frequency judgments were used.
First, after each block of 50 questions, the participant estimated the number
of correct answers among the 50 answers given. Because there were 350 ques-
144 BOUNDED RATIONALITY

tions, every participant gave seven estimates of the number of correct answers.
Second, after the participants answered all questions, they were given an en-
larged copy of the confidence scale used throughout the experiment and were
asked for the following frequency judgment: "How many of the answers that
you classified into a certain confidence category are correct? Please indicate
for every category your estimated relative frequency of correct answers."
In Experiment 1, we also introduced two of the standard manipulations in
the literature. The first was to inform and warn half of our participants of the
overconfidence effect, and the second was to offer half of each group a mon-
etary incentive for good performance. Both are on a list of "debiasing" methods
known as being relatively ineffective (Fischhoff, 1982), and both contributed
to the view that overconfidence is a robust phenomenon. If PMM theory is
correct, the magnitude of effects resulting from the manipulations in this
chapter—confidence versus frequency judgment and selected versus represen-
tative sampling—should be much larger than those resulting from the two stan-
dard "debiasing" manipulations.

Participants Participants were 80 students (43 men and 37 women) at the

University of Konstanz who were paid for participation. Participants were
tested in small groups of a maximum of 12 persons.

Design and Procedure This was a 2 X 2 X 2 design with representative-

selected set varied within participants and warning-no warning about over-
confidence and monetary incentive-no incentive as independent variables var-
ied between participants. Half of the participants answered the representative
set first; the other half, the selected set. Order of questions was determined
randomly in both sets.
The confidence scale consisted of seven categories, 50%, 51%-60%, 61%—
70%, 71%-80%, 81%-90%, 91%-99%, and 100% confident. The 50%- and
100%-confidence values were introduced as separate categories because pre-
vious research showed that participants often tend to use these particular val-
ues. Participants were told first to mark the alternative that seemed to be the
correct one, and then to indicate with a second cross their confidence that the
answer was correct. If they only guessed, they should cross the 50% category;
if they were absolutely certain, they should cross the 100% category. We ex-
plained that one of the alternatives was always correct. In the warning con-
dition, participants received the following information: "Most earlier studies
found a systematic tendency to overestimate one's knowledge; that is, there
were many fewer answers correct than one would expect from the confidence
ratings given. Please keep this warning in mind." In the incentive condition,
participants were promised 20 German marks (or a bottle of French cham-
pagne) for the best performance in the group, in addition to the payment that
everyone received (7.50 marks).
To summarize, 350 questions were presented, with a confidence judgment
after each question, a frequency judgment after each 50 questions, and a judg-
 PROBABILISTIC MENTAL MODELS 145

ment of relative frequencies of correct answers in each confidence category at

the end.
For comparison with the literature on calibration, we used the following
measure:

where n is the total number of answers, nt is the number of times the confi-
dence judgment pf was used, and ff is the relative frequency of correct answers
for all answers assigned confidence p,. / is the number of different confidence
categories used (/ = 7), and p and/ are the overall mean confidence judgment
and percentage correct, respectively. A positive difference is called overcon-
fidence. For convenience, we report over- and underconfidence in percentages
(X 100).

Results
Prediction 1 PMM theory predicts that in the selected set (general-knowledge
questions), people show overestimation in confidence judgments (overconfi-
dence) and, simultaneously, accurate frequency judgments.
The open-circle curve in Figure 7.6 shows the relation between judgments
of confidence and the true relative frequency of correct answers in the selected
set—that is, the set of mixed general-knowledge questions. The relative fre-
quency of correct answers (averaged over all participants) was 72.4% in the
100%-confidence category, 66.3% in the 95% category, 58.0% in the 85% cat-
egory, and so on. The curve is far below the diagonal (calibration curve) and
similar to the curves reported by Lichtenstein, Fischhoff, and Phillips (1982,
Figure 2). It replicates and demonstrates the well-known overconfidence effect.
Percentage correct was 52.9, mean confidence was 66.7, and overconfidence
was 13.8.
Participants' frequency judgments, however, are fairly accurate, as Table 7.2
(last row) shows. Each entry is averaged over the 20 participants in each con-
dition. For instance, the figure —1.8 means that, on average, participants in
this condition underestimated the true number of correct answers by 1.8. Av-
eraged across the four conditions, we get —1.2, which means that participants
missed the true frequency by an average of only about 1 correct answer in the
set of 50 questions. Quite accurate frequency judgments coexist with over-
confidence. The magnitude of this confidence—frequency effect found is shown
in Figure 7.7 (left side). PMM theory predicts this systematic difference be-
tween confidence and frequency judgments, within the same person and the
same general-knowledge questions.

Prediction 2 PMM theory predicts that in the representative set (city ques-
tions) people show zero overconfidence and, at the same time, underestimation
in frequency judgments.
Figure 7.6 Calibration curves for three sets. Overconfidence appears when
questions are selected (open circles) but disappears when questions are rep-
resentative (black squares). The matched set controls for the different content
of the two sets. Here, questions are selected from the representative set to
match the difficulty of the selected set, and overconfidence is again pro-
duced.

Table 7.2 Mean differences between estimated and true frequencies of correct
answers
No warning- Incentive Warning Warning
Set no incentive only only and incentive

Representative
1-50 -9.9 -9.4 -8.8 -8.7
51-100 -9.5 -10.4 -12.0 -11.3
101-150 -9.9 -10.9 -10.9 -9.9
151-200 -6.7 -6.7 -9.4 -5.9
201-250 -9.8 -9.8 -8.0 -5.3
251-300 -9.5 -10.8 -9.4 -9.1
Average -9.2 -9.7 -9.7 -8.4
Selected -1.8 -0.6 -2.7 0.3

Note: Negative signs denote underestimation of true number of correct answers.

 PROBABILISTIC MENTAL MODELS 147

Figure 7.7 Confidence-frequency effect in representative and selected sets.

Frequency judgments are long-run frequencies, N = 50.

The solid-square curve in Figure 7.6 shows the relation between confidence
and percentage correct in the representative set—that is, the city questions.
For instance, percentage correct in the 100%-confidence category was 90.8%,
instead of 72.4%. Overconfidence disappeared (—0.9%). Percentage correct
and mean confidence were 71.7 and 70.8, respectively.
The confidence curve for the representative set is similar to a regression
curve for the estimation of relative frequencies by confidence, resulting in un-
derconfidence in the left part of the confidence scale, overconfidence in the
right, and zero overconfidence on the average.
Table 7.2 shows the differences between estimated and true frequencies for
each block of 50 items and each of the conditions, respectively. Again, each
entry is averaged over the 20 participants in each condition. For instance,
participants who were given neither information nor incentive underestimated
their true number of correct answers by 9.9 (on the average) in the first 50
items of the representative set. Table 7.2 shows that the values of the mean
differences were fairly stable over the six subsets, and, most important, they
are, without exception, negative (i.e., underestimation).
148 BOUNDED RATIONALITY

The following is an illustration at the individual level: Participant 1 esti-

mated 28, 30, 23, 25, 23, and 23, respectively, for the six subsets, compared
with 40, 38, 40, 36, 35, and 32 correct solutions, respectively. An analysis of
individual judgments confirmed average results. Among the 80 participants,
71 underestimated the number of correct answers, whereas only 8 participants
overestimated it (frequency judgments were missing for 1 participant). Inci-
dentally, 7 of these 8 participants were male. In the selected set, for compar-
ison, 44 participants underestimated and 35 participants overestimated the
number of correct answers, and 1 participant got it exactly right.
We have attributed the emergence and disappearance of overconfidence to
selection versus use of a representative set. One objection to this analysis is
that the difference between the open-circle and the solid-square curve in Fig-
ure 7.6 is confounded with a difference in the content of both sets. The selected
set includes a broad range of general-knowledge questions, whereas the do-
main of the representative set (cities) is necessarily more restricted. To check
for this possible confound, we determined the item difficulties for each of the
50 general-knowledge questions and selected a subset of 50 city questions that
had the same item difficulties. If the difference in Figure 7.6 is independent
of content but results from the selection process, this "matched" subset of city
questions should generate the same calibration curves showing overconfidence
as the selected set of general-knowledge questions did. Figure 7.6 shows that
this is the case (open-square curve). Both content domains produce the same
results if questions are selected.
To summarize, in the representative set, overestimation disappears in con-
fidence judgments, and zero-overconfidence coexists with frequency judg-
ments that show large underestimation. Results confirm Prediction 2. Figure
7.7 (right side) shows the magnitude of the confidence-frequency effect found.
No previous theory of confidence can predict the results depicted in Figure
7.7.

Prediction 3 PMM theory predicts that overestimation will disappear if the

relative frequencies of correct answers (percentage correct) in each confidence
category are compared with the estimated relative frequencies. Because partic-
ipants estimated percentage correct for all confidence judgments—that is, in-
cluding both the selected and the representative set—we expect not only that
overestimation will disappear (the prediction from the selected set) but also
that it will turn into underestimation (the prediction from the representative
set).
The solid line in Figure 7.8 shows the results for Experiment 1: Estimated
relative frequencies are well calibrated and show underestimation in five out
of seven confidence categories. Overestimation of one's knowledge disappears.
The only exception is the 100%-confidence category. The latter is the confi-
dence category that contains all solutions by local MMs, and errors in memory
or elementary logical operations may account for the difference. Figure 7.8 is
a "frequentist" variant of the calibration curve of Figure 7.6. Here, true per-
centage correct is compared with estimated percentage correct, rather than
 PROBABILISTIC MENTAL MODELS 149

Figure 7.8 Calibration curves for judgments of percentage correct in confi-

dence categories. Values are averaged across both sets of questions.

with confidence. For instance, in the 100%-confidence category, true and es-
timated percentage correct were 88.8% and 93.0%, respectively.
Averaged across experimental conditions, the ratio between estimated fre-
quency in the long run and confidence value is fairly constant, around .87, for
confidence ratings between 65% and 95%. It is highest in the extreme cate-
gories (see Table 7.3).
To summarize, participants explicitly distinguished between confidence in
single answers and the relative frequency of correct answers associated with

Table 7.3 Estimated and true percentage correct in each confidence category
(summarized over the representative and the selected sets)
No. of % correct
f~t £• J
Confidence confidence /~\ 1 J
Uver-/under-
category judgments Estimated True estimation

100 5,166 93.0 88.8 4.2

91-99 1,629 82.7 81.6 1.1
81-90 2,534 73.1 74.6 -1.5
71-80 2,950 64.3 70.1 -5.8
61-70 3,506 57.3 65.6 -8.3
51-60 4,036 53.7 63.3 -9.6
50 8,178 49.8 56.3 -6.5
X or M 27,999 64.8 69.1 -4.2
150 BOUNDED RATIONALITY

a confidence judgment. This result is implied by PMM theory, according to

which different reference classes are cued by confidence and frequency tasks.
As stated in Prediction 3, overestimation disappeared. However, the magnitude
of underestimation was not, as might be expected, as pronounced as in the
frequency judgments dealt with in Predictions 1 and 2. Except for this finding,
results conformed well to Prediction 3. Note that no previous theory of con-
fidence in knowledge we are aware of makes this conceptual distinction and
that prediction. Our results contradict much of what has been assumed about
how the untutored mind understands the relation between confidence and rel-
ative frequency of correct answers.

Warning about Overconfidence and Monetary Incentive Mean confidence judg-

ments were indistinguishable between participants warned of overconfidence
and those uninformed. If a monetary incentive was announced, overconfidence
was more pronounced with incentive than without incentive in five categories
(65%-100%) and less pronounced in the 50% category, with an average in-
crease of 3.6 percentage points.
The monetary incentive effect resulted from the incentive/no-warning
group, in which confidence judgments were higher than in all three other
groups (but we found the same percentage correct in all groups). One reason
for this interaction could be that we did not specify in the instructions a cri-
terion for best performance. If warned of overconfidence, participants could
easily infer that the incentive was for minimizing overconfidence. If not
warned, at least some participants could also have attempted to maximize per-
centage correct. None of these attempts, however, was successful, consistent
with PMM theory and earlier studies (e.g., Fischhoff, Slovic, & Lichtenstein,
1977). The effort to raise the percentage correct seems to have raised confi-
dence instead, an outcome that cannot be accounted for by PMM theory. The
size of this effect, however, was small compared with both the confidence-
frequency effect and that of selected versus representative sampling.
To summarize, neither warning of overconfidence nor associated monetary
incentive decreased overconfidence or increased percentage correct, replicat-
ing earlier findings that knowledge about overconfidence is not sufficient to
change confidence. An incentive that participants seem to have interpreted as
rewarding those who maximize the percentage correct, however, increased
confidence.

Order of Presentation and Sex Which set (representative vs. selected) was
given first had no effect on confidences, neither in Experiment 1 nor in Ex-
periment 2. Arkes, Christensen, Lai, and Blumer (1987) found an effect of the
difficulty of one set of items on the confidence judgments for a second set
when participants received feedback for their performance in the first set. In
our experiment, however, no feedback was given. Thus, participants had no
reason to correct their confidence judgments, such as by subtracting a constant
value. Sex differences in degree of overconfidence in knowledge have been
claimed by both philosophy and folklore. Our study, however, showed no sig-
 PROBABILISTIC MENTAL MODELS 151

nificant differences between the sexes in either overconfidence or calibration,

in either Experiment 1 or in Experiment 2. (The men's confidence judgments
were on the average 5 percentage points higher than women's, but so was their
percentage correct. This replicates Lichtenstein and Fischhoff's, 1981, findings
about students at the University of Oregon.)
To summarize, as predicted by PMM theory, we can experimentally make
overconfidence (overestimation) appear, disappear, and invert. Experiment 1
made our participants consistently switch back and forth among these re-
sponses. The key to this finding is a pair of concepts that have been neglected
by the main previous explanations of confidence in one's knowledge—confi-
dence versus frequency judgment and representative versus selected sampling.

Experiment 2

We tried to replicate the results and test several objections. First, to strengthen
the case against PMM theory, we instructed the participants both verbally and
in written form that confidence is subjective probability, and that among all
cases where a subjective probability of X% was chosen, X% of the answers
should be correct. Several authors have argued that such a frequentist instruc-
tion could enhance external calibration or internal consistency (e.g., Kahne-
man & Tversky, 1982; May, 1987). According to PMM theory, however, confi-
dence is already inferred from frequency (with or without this instruction)—
but from frequencies of co-occurrences between, say, number of inhabitants
and several cues, and not from base rates of correct answers in similar testing
situations (see Table 7.1). Thus, in our view, the preceding caution will be
ineffective because the base rate of correct answers is not a probability cue
that is defined on a reference class such as cities in Germany.
Second, consider the confidence-frequency effect. We have shown that this
new effect is implied by PMM theory. One objection might be that the differ-
ence between confidence and frequency judgments is an artifact of the re-
sponse function, just as overconfidence has sometimes been thought to be.
Consider the following interpretation of overconfidence. If (a) confidence is
well calibrated but (b) the response function that transforms confidence into a
confidence judgment differs from an identity function, then (c) overconfidence
or underconfidence "occurs" on the response scale. Because an identity func-
tion has not been proven, Anderson (1986), for instance, denoted the over-
confidence effect and the hard—easy effect as "largely meaningless" (p. 91):
They might just as well be response function artifacts.
A similar objection could be made against the interpretation of the confi-
dence-frequency effect within PMM theory. Despite the effect's stability across
selected and representative sets, it may just reflect a systematic difference be-
tween response functions for confidence and frequency judgments. This con-
jecture can be rephrased as follows: If (a) the difference between "internal"
confidence and frequency impression is zero, but (b) the response functions
that transform both into judgments differ systematically, then (c) a confidence-
152 BOUNDED RATIONALITY

frequency effect occurs on the response scales. We call this the response-
function conjecture.
How can this conjecture be tested? According to PMM theory, the essential
basis on which both confidence and frequency judgments are formed is the
probability cues, not response functions. We assumed earlier that frequency
judgments are based mainly on base rates of correct answers in a reference
class of similar general-knowledge test situations. If we make another cue
available, then frequency judgments should change. In particular, if we make
the confidence judgments more easily retrievable from memory, these can be
used as additional probability cues, and the confidence—frequency effect
should decrease. This was done in Experiment 2 by introducing frequency
judgments in the short run, that is, frequency judgments for a very small num-
ber of questions. Here, confidence judgments can be more easily retrieved from
memory than they could in the long run. Thus, if PMM theory is correct, the
confidence-frequency effect should decrease in the short run. If the issue were,
however, different response functions, then the availability of confidence judg-
ments should not matter because confidence and frequency impression are
assumed to be identical in the first place. Thus, if the conjecture is correct, the
confidence-frequency effect should be stable.
In Experiment 2, we varied the length N of a series of questions from the
long-run condition N = 50 in Experiment 1 to the smallest possible short run
of AT- 2.
Third, in Experiment 1 we used a response scale ranging from 50% to 100%
for confidence judgments but a full-range response scale for frequency judg-
ments ranging from 0 to 50 correct answers (which corresponds to 0% to
100%). Therefore one could argue that the confidence—frequency effect is an
artifact of the different ranges of the two response scales. Assume that (a) there
is no difference between internal confidence and frequency, but (b) because
confidence judgments are limited to the upper half of the response scale,
whereas frequency judgments are not, (c) the confidence-frequency effect re-
sults as an artifact of the half-range response scale in confidence judgments
We refer to this as the response-range conjecture. It can be backed up by at
least two hypotheses.

1. Assume that PMM theory is wrong and participants indeed use base rates
of correct answers as a probability cue for confidence in single answers.
Then confidence shouldfoe considerably lower. If participants anticipate
misleading questions, eVen confidences lower than 50% are reasonable to
expect on this conjecture. Confidences below 50%, however, cannot be
expressed on a scale with a lower boundary at 50%, whereas they can at
the frequency scale. Effects of response range such as those postulated in
range-frequency theory (Parducci, 1965) or by Schonemann (1983) may
enforce the distorting effect of the half-range format. In this account, both
the overconfidence effect and the confidence-frequency effect are gener-
ated by a response-scale effect. With respect to overconfidence, this con-
jecture has been made and has claimed some support (e.g., May, 1986,
1987; Ronis & Yates, 1987). We call this the base rate hypothesis.
 PROBABILISTIC MENTAL MODELS 153

2. Assume that PMM theory is wrong in postulating that choice and confi-
dence are essentially one process and that the true process is a temporal
sequence: choice, followed by search for evidence, followed by confi-
dence judgment. Koriat et al. (1980), for instance, proposed this se-
quence. Assume further, contrary to Koriat, that the mind is "Popper-
ian," searching for disconfirming rather than for confirming evidence to
determine the degree of "corroboration" of an answer. If the participant
is successful in retrieving disconfirming evidence from memory but is
not allowed to change the original answer, confidence judgments less
than 50% will result. Such disconfirmation strategies, however, can
hardly be detected using a 50%-100% format, whereas they could in a
full-scale format. We call this the disconfirmation strategy hypothesis.
To test the response-range conjecture, half of the participants in Experiment
2 were given full-range response scales, whereas the other half received the
response scales used in Experiment 1.

Method

Participants Ninety-seven new participants at the University of Konstanz

(not enrolled in psychology) were paid for their participation. There were 59
males and 38 females. As in Experiment 1, participants were tested in small
groups of no more than 7.

Design and procedure This was a 4 X 2 X 2 design, with length of series (50,
10, 5, and 2) and response scale (half range vs. full range) varied between
participants and type of knowledge questions (selected vs. representative set)
varied within participants.
The procedure and the materials were like that in Experiment 1, except for
the following. We used a new random sample of 21 (instead of 25) cities. This
change decreased the number of questions in the representative set from 300
to 210. As mentioned earlier, we explicitly instructed the participants to in-
terpret confidences as frequencies of correct answers: "We are interested in
how well you can estimate subjective probabilities. This means, among all the
answers where you give a subjective probability of X%, there should be X%
of the answers correct." This calibration instruction was orally repeated and
emphasized to the participants.
The response scale contained the means (50%, 55%, 65%, . . . , 95%, 100%)
of the intervals used in Experiment 1 rather than the intervals themselves to
avoid the problematic assumption that means would represent intervals. End-
points were marked absolutely certain that the alternative chosen is correct
(100%), both alternatives equally probable (50%), and, for the full-range scale,
absolutely certain that the alternative chosen is incorrect (0%). In the full-
range scale, one reason for using confidences between 0% and 45% was ex-
plained in the following illustration: "If you think after you have made your
choice that you would have better chosen the other alternative, do not change
your choice, but answer with a probability smaller than 50%."
154 BOUNDED RATIONALITY

After each set of AT = 50 (10, 5, or 2) answers, participants gave a judgment

of the number of correct answers. After having completed 50 + 210 = 260
confidence judgments and 5, 26, 52, or 130 frequency judgments (depending
on the participant's group), participants in both response-scale conditions were
presented the same enlarged copy of the 50%—100% response scale and asked
to estimate the relative frequency of correct answers in each confidence cate-
gory.

Results
Response-Range Conjecture We tested the conjecture that the systematic dif-
ference in confidence and frequency judgments stated in Predictions 1 and 2
(confidence-frequency effect) and shown in Experiment 1 resulted from the
availability of only a limited response scale for confidence judgments (50% to
100%).
Forty-seven participants were given the full-range response scale for con-
fidence judgments. Twenty-two of these never chose confidences below 50%;
the others did. The number of confidence judgments below 50% was small.
Eleven participants used them only once (in altogether 260 judgments), 5 did
twice, and the others 3 to 7 times. There was one outlier, a participant who
used them 67 times. In total, participants gave a confidence judgment smaller
than 50% for only 1.1% of their answers (excluding the outlier: 0.6%). If the
response-range conjecture had been correct, participants would have used con-
fidence judgments below 50% much more frequently.
In the representative set, overconfidence was 3.7% (SEM — 1.23) in the full-
range scale condition and 1.8% (SEM = 1.15) in the half-range condition. In
the selected set, the corresponding values were 14.4 (SEM = 1.54) and 16.4
(SEM = 1.43). Averaging all questions, we got slightly larger overconfidence in
the full-range condition (mean difference = 1.2). The response-range conjec-
ture, however, predicted a strong effect in the opposite direction. Frequency
judgments were essentially the same in both conditions. Hence, the confi-
dence-frequency effect can also be demonstrated when both confidence and
frequency judgments are made on a full-range response scale.
To summarize, there was (a) little use of confidences below 50% and (b) no
decrease of overconfidence in the full-range condition. These results contradict
the response-range conjecture.
A study by Ronis and Yates (1987) seems to be the only other study that
has compared the full-range and the half-range format in two-alternative choice
tasks, but it did not deal with frequency judgments. These authors also re-
ported that only about half their participants used confidence judgments below
50%, although they did so more frequently than our participants. Ronis and
Yates concluded that confidences below 50% had only a negligible effect on
overconfidence and calibration (pp. 209-211). Thus results in both studies are
consistent. The main difference is that Ronis and Yates seem to consider only
"failure to follow the instructions" and "misusing the probability scale"
(p. 207) as possible explanations for confidence judgments below 50%. In con-
 PROBABILISTIC MENTAL MODELS 155

trast, we argue that there are indeed plausible cognitive mechanisms—the base
rate and disconfirmation strategy hypotheses—that imply these kind of judg-
ments, although they would contradict PMM theory.
Both Experiment 2 and the Ronis and Yates (1987) study do not rule out,
however, a more fundamental conjecture that is difficult to test. This argument
is that internal confidence (not frequency) takes a verbal rather than a numer-
ical form and that it is distorted on any numerical probability rating scale, not
just on a 50%-100% response scale. Zimmer (1983, 1986) argued that verbal
expressions of uncertainty (such as "highly improbably" and "very likely") are
more realistic, more precise, and less prone to overconfidence and other so-
called judgmental biases than are numerical judgments of probability. Zim-
mer's fuzzy-set modeling of verbal expressions, like models of probabilistic
reasoning that dispense with the Kolmogoroff axioms (e.g., Cohen, 1989; Ky-
burg, 1983; Shafer, 1978), remains a largely unexplored source of alternative
accounts of confidence.
For the remaining analysis, we do not distinguish between the full-range
and the half-range response format. For combining the data, we receded an-
swers like "alternative a, 40% confident" as "alternative b, 60% confident,"
following Ronis and Yates (1987).

Predictions 1 and 2: Confidence-Frequency Effect The question is whether the

confidence—frequency effect can be replicated under the explicit instruction
that subjective probabilities should be calibrated to frequencies of correct an-
swers in the long run. Calibration curves in Experiment 2 were similar to those
in Figure 7.6 and are not shown here for this reason. Figure 7.7 shows that the
confidence-frequency effect replicates. In the selected set, mean confidence
was 71.6%, and percentage correct was 56.2. Mean estimated number of cor-
rect answers (transformed into percentages) in the series of N = 50 was 52.0%.
As stated in Prediction 1, overconfidence in single answers coexists with fairly
accurate frequency judgments, which once again show slight underestimation.
In the representative set, mean confidence was 78.1% and percentage cor-
rect was 75.3%. Mean estimated number of correct answers per 50 answers
was 63.5%. As forecasted in Prediction 2, overconfidence largely disappeared
(2.8 percentage points), and frequency judgments showed underestimation
(—11.8 percentage points).
An individual analysis produced similar results. The confidence-frequency
effect (average confidence higher than average frequency judgment) held for
82 (83) participants in the selected (representative) set (out of 97). Answering
the selected set, 92 respondents showed overconfidence, and 5 showed un-
derconfidence. In the representative set, however, 60 exhibited overconfidence
and 37 exhibited underconfidence.

Prediction 3: Estimated Percentage Correct in Confidence Categories After the

participants answered the 260 general-knowledge questions, they were asked
what percentage they thought they had correct in each confidence category. As
shown by the dashed line in Figure 7.8, results replicated well. Average esti-
156 BOUNDED RATIONALITY

mated percentage correct differed again from confidence and was close to the
actual percentage correct.
Despite the instruction not to do so, our participants still distinguished be-
tween a specific confidence value and the corresponding percentage of correct
responses. Therefore confidence and hypothesized percentage correct should
not be used as synonyms. As suggested by this experiment, an instruction
alone cannot override the cognitive processes at work.
In the 100%-confidence category, for instance, 67 participants gave esti-
mates below 100%. In a postexperimental interview, we pointed out to them
that these judgments imply that they assumed they had not followed the cal-
ibration instruction. Most explained that in each single case, they were in fact
100% confident. But they also knew that, in the long run, some answers would
nonetheless be wrong, and they did not know which ones. Thus they did not
know which of the 100% answers they should correct. When asked how they
made the confidence judgments, most answered by giving examples of prob-
ability cues, such as "I know that this city is located in the Ruhrgebiet (in-
dustrial belt), and most cities there are rather large." Interviews provided ev-
idence for several probability cues, but no evidence that base rate expectations,
as reported in frequency judgments, were also used in confidence judgments.

Response-Function Conjecture: Frequency Judgments in the Short and Long

Runs We tested the conjecture that the confidence-frequency effect stated in
Predictions 1 and 2 and shown in Experiment 1 might be due to different
response functions for confidence and frequency judgments, rather than to dif-
ferent cognitive processes, as postulated by PMM theory. If the conjecture were
true, the availability of confidence judgments in the short run should not
change the confidence—frequency effect (see the previous discussion).

Figure 7.9 Decrease of the confidence-frequency effect in short runs (N = 50,

10, 5, and 2). Values are differences between mean confidence and estimated
percentage correct in a series of length N, Values are averaged across all
questions.
 PROBABILISTIC MENTAL MODELS 157

Contrary to the response-function conjecture, the length of series showed

an effect on the judgments of frequency of correct answers. Figure 7.9 shows
the extent of the disappearance of the confidence-frequency effect in the short
run. The curve shows that the effect decreased from N = 50 to N = 2, averaged
across both sets of items. The decrease was around 12 percentage points, an
amount similar in the selected set (from 18.9% to 6.9%) and in the represen-
tative set (from 15.7% to 3.3%).
The breakdown of the confidence-frequency effect in the short run is in-
consistent with the objection that the effect can be reduced to a systematic
difference in response functions. This result is, however, consistent with the
notion that the shorter the run, the more easily are confidence judgments avail-
able from memory, and, thus, the more they can be used as probability cues
for the true number of correct answers.

Discussion

Our starting point was the overconfidence effect, reported in the literature as
a fairly stable cognitive illusion in evaluating one's general knowledge and
attributed to general principles of memory search, such as confirmation bias
(Koriat et al., 1980), to general motivational tendencies such as fear of invalid-
ity (Mayseless & Kruglanski, 1987), to insensitivity to task difficulty (see von
Winterfeldt & Edwards, 1986, p. 128), and to wishful thinking and other "def-
icits" in cognition, motivation, and personality. Our view, in contrast, proposes
that one evaluates one's knowledge by probabilistic mental models. In our
account, the main deficit of most cognitive and motivational explanations is
that they neglect the structure of the task and its relation to the structure of a
corresponding environment known to the participants. If people want to search
for confirming evidence or to believe that their answers are more correct than
they are because of some need, wish, or fear, then overestimation of accuracy
should express itself independently of whether they judge single answers or
frequencies, a selected or representative sample of questions, and hard or easy
questions.
Our experiments also do not support the explanation of overconfidence and
the hard—easy effect by assuming that participants are insensitive to task dif-
ficulty: In frequency tasks we have shown that participants' judgments of their
percentage correct in the long run are in fact close to actual percentage correct,
although confidences are not. Overconfidence does not imply that participants
are not aware of task difficulty. At least two more studies have shown that
estimated percentage correct can correspond closely to true percentage correct
in general-knowledge tasks. Allwood and Montgomery (1987) asked their par-
ticipants to estimate how difficult each of 80 questions was for their peers and
found that difficulty ratings (M = 57%) were more realistic (percentage correct
= 61%) than confidence judgments (M = 74%). May (1987) asked her partic-
ipants to estimate their percentage of correct answers after they completed an
158 BOUNDED RATIONALITY

experiment with two-alternative questions. She found that judgments of per-

centage correct accorded better with the true percentage correct than did con-
fidences.
In our account, overconfidence results from one of two causes, or both: (a)
a PMM for a task is not properly adapted to a corresponding environment (e.g.,
cue validities do not correspond to ecological validities), or (b) the set of ob-
jects used is not a representative sample from the corresponding reference
class in the environment but is selected for difficulty. If a is the true cause,
using a representative sample from a known environment should not eliminate
overconfidence. If b is true, it should. In both experiments, overconfidence in
knowledge about city populations was eliminated, as implied by b. Thus ex-
perimental results are consistent with both PMM theory and the assumption
that individual PMMs are on the average well adapted to the city environment
we used. Overconfidence resulted from a set of questions that was selected for
difficulty. Underconfidence, conversely, would result from questions selected
to be easy.
The foregoing comments do not mean that overestimation of knowledge is
just an artifact of selected questions. If it were, then judgments of frequency
of correct answers should show a similar degree of overestimation. What we
have called the confidence-frequency effect shows that this is not the case.
Several authors have proposed that judgments in the frequency mode are
more accurate, realistic, or internally consistent than probabilities for single
events (e.g., Teigen, 1974, p. 62; Tversky & Kahneman, 1983). Our account is
different. PMM theory states conditions under which mean judgments of con-
fidence are systematically larger than judgments of relative frequency. PMM
theory does not, however, imply that frequency judgments are generally better
calibrated. On the contrary, frequency judgments may be miscalibrated for the
same reasons as confidence judgments. The set of tasks may not be represen-
tative for the reference class from which the inferences are made.
The experimental control of overestimation—how to make overestimation
appear, disappear, and invert—gives support to PMM theory. These predic-
tions, however, do not exhaust the inferences that can be derived from PMM
theory.

Explaining Anomalies in the Literature

In this section, we explain a series of apparently inconsistent findings and

integrate these into PMM theory.

Ronis and Yates (1987) We have mentioned that the Ronis and Yates study is
the only other study that tested a full-range response scale for two-alternative
tasks. The second purpose of that study was to compare confidence judgments
in situations in which the participant knows that the answers are known to
the experimenter (general-knowledge questions) with outcomes of upcoming
basketball games, in which answers are not yet known. In all three (response-
scale) groups, percentage correct was larger for general-knowledge questions
 PROBABILISTIC MENTAL MODELS 159

Figure 7.10 Reversal of the hard-easy effect in Ronis and Yates (1987) and
Keren (1988).

than for basketball predictions. Given this result, what would current theories
predict about overconfidence? The insensitivity hypothesis proposes that peo-
ple are largely insensitive to percentage correct (see von Winterfeldt & Ed-
wards, 1986, p. 128). This implies that overconfidence will be larger in the
more difficult (hard) set: the hard-easy effect. (The confirmation bias and mo-
tivational explanations are largely mute on the difficulty issue.) PMM theory,
in contrast, predicts that overconfidence will be larger in the easier set (hard-
easy effect reversal; see Prediction 5) because general-knowledge questions
(the easy set) were selected and basketball predictions were not; only with
clairvoyance could one select these predictions for percentage correct.
In fact, Ronis and Yates (1987) reported an apparent anomaly: three hard-
easy effect reversals. In all groups, overconfidence was larger for the easy
general-knowledge questions than for the hard basketball predictions (Figure
7.10). Prediction 5 accounts for these observed reversals of the hard-easy ef-
fect.

Koridt et al. (1980) Experiment 2 of Koriat et al.'s study provided a direct test
of the confirmation bias explanation of overconfidence. The explanation is this:
(a) participants first choose an answer based on their knowledge, then (b) they
selectively search for confirming memory (or for evidence discontinuing the
alternative not chosen), and (c) this confirming evidence generates over-
confidence. Between the participants' choice of an answer and their confidence
judgment, the authors asked them to give reasons for the alternative chosen.
160 BOUNDED RATIONALITY

Three groups of participants were asked to write down one confirming reason,
one disconfirming reason, or one of each, respectively. Reasons were given for
half of the general-knowledge questions; otherwise, no reasons were given
(control condition). If the confirmation bias explanation is correct, then asking
for a contradicting reason (or both reasons) should decrease overconfidence
and improve calibration. Asking for a confirming reason, however, should
make no difference "since those instructions roughly simulate what people
normally do" (Koriat et al, 1980, p. 111).
What does PMM theory predict? According to PMM theory, choice and con-
fidence are inferred from the same activated cue. This cue is by definition a
confirming reason. Therefore, the confirming-reason and the no-reason (con-
trol) tasks engage the same cognitive processes. The difference is only that in
the former the supporting reason is written down. Similarly, the disconfirming-
reason and both-reason tasks involve the same cognitive processes. Further-
more, PMM theory implies that there is no difference between the two pairs
of tasks.
This result is shown in Table 7.4. In the first row we have the no-reason
and confirming-reason tasks, which are equivalent. Here, only one cue is ac-
tivated, which is confirming. There is no disconfirming cue. Now consider the
second row, the disconfirming-reason and both-reason tasks, which are again
equivalent. Both tasks are solved if one additional cue, which is disconfirming,
can be activated. Thus, for PMM theory, the cue generation and testing cycle
is started again, and cues are generated according to the hierarchy of cue va-
lidities and tested as to whether they can be activated for the problem at hand.
The point is that the next cue that can be activated may turn out to be either
confirming or disconfirming.
For simplicity, assume that the probability that the next activated cue turns
out to be confirming or disconfirming is the same. If it is disconfirming, the
cycle is stopped, and two cues in total have been activated, one confirming

Table 7.4 Predictions of PMM theory for the effects of asking for a
disconfirming reason
Cues activated
Predicted change in
Task No. of cues activated CON DIS Probability confidence

No; CON 1 CON 1 0

DIS; both 2 DIS CON 1 1 .5 Decrease

DIS; both 3 DIS CON 2 1 .25 Increase

DIS; both 4 DIS CON 3 1 .125 Increase

DIS; both >4 >3 1 .125 Increase

Note: CON = confirming reason; DIS = disconfirming reason; No = no reason; both = both reasons.
 PROBABILISTIC MENTAL MODELS 161

and one discontinuing. This stopping happens with probability .5, and it de-
creases both confidence and overconfidence. (Because the second cue activated
has a lower cue validity, however, confidence is not decreased below 50%.) If
the second cue activated is again confirming, a third has to be activated, and
the cue generation and testing cycle is entered again. If the third cue is dis-
confirming, the cycle stops with two confirming cues and one disconfirming
cue activated, as shown in the third row of Table 7.4. This stopping is to be
expected with probability .25. Because the second cue has higher cue validity
than the third, disconfirming, cue, overall an increase in confidence and over-
confidence is to be expected. If the third cue is again confirming, the same
procedure is repeated. Here and in all subsequent cases confidence will
increase. As shown in Table 7.4, the probabilities of an increase sum up to
.5 (.25 + .125 + .125), which is the same as the probability of a decrease.
Thus PMM theory leads to the prediction that, overall, asking for a discon-
firming reason will not change confidence or overconfidence. As just shown,
the confirmation-bias hypothesis, in contrast, predicts that asking for a discon-
firming reason should decrease confidence and overconfidence.
What were the results of the Koriat study? In both crucial conditions,
disconfirming reason and both reasons, the authors found only small and
nonsignificant decreases of overconfidence (2% and 1%, respectively) and
similar small improvements in calibration (.006 each, significant only in the
disconfirming-reason task). These largely insignificant differences are consis-
tent with the prediction by PMM theory that asking for a disconfirming reason
makes no difference and are inconsistent with the confirmation-bias ex-
planation. Further evidence comes from a replication of the Koriat study by
Fischhoff and MacGregor (1982), who reported zero effects of disconfirming
reasons.
To summarize, the effects on confidence of giving confirming and discon-
firming reasons in the Koriat study can be both explained by and integrated
into PMM theory. There is no need to postulate a confirmation bias.

Dawes (1980) Overconfidence has been attributed to people's tendency to

"overestimate the power of our 'intellect' as opposed to that of our coding
abilities." Such overestimation "has been reinforced by our realization that we
have developed a technology capable of destroying ourselves" (Dawes, 1980,
p. 328). Dawes proposed that overconfidence is characteristic for general-
knowledge questions but absent in perceptual tasks; he designed a series of
experiments to test this proposal. PMM theory, however, gives no special treat-
ment to perceptual tasks. On the contrary, it predicts overconfidence if per-
ceptual tasks are selected for perceptual illusions—that is, for being mislead-
ing—whereas zero overconfidence is to be expected if tasks are not selected.
Pictures in textbooks on visual illusions are probably the set of items that
produces the most extreme overconfidence yet demonstrated. Nevertheless, in
a natural environment, perception is generally reliable.
Dawes reported inconsistent results. When perceptual stimuli were syste-
matically constructed from a Square X Circle matrix, as in the area task, and
162 BOUNDED RATIONALITY

no selection for stimuli that generated perceptual illusions took place, over-
confidence was close to zero (perception of areas of squares is quite well
adapted in adults; see Gigerenzer & Richter, 1990). This result is predicted by
both accounts. The anomaly arises with the second perceptual task used—
judging which of two subsequent tones is longer. If the second tone was longer,
Dawes reported almost perfect calibration, but if the first tone was longer, par-
ticipants exhibited large overconfidence.
PMM theory predicts that in the inconsistent acoustic task, perceptual stim-
uli have been selected (albeit unwittingly) for a perceptual illusion. This is in
fact the case. From the literature on time perception, we know that of two
subsequently presented tones, the tone more recently heard appears to be
longer. This perceptual illusion is known as the negative presentation effect
(e.g., Fraisse, 1964; Sivyer & Finlay, 1982). It implies a smaller percentage of
correct answers in the condition in which the tone presented first was longer,
because this tone is perceived to be shorter. A decrease in percentage correct
in turn increases overconfidence. In Dawes's (1980) experiments, this is exactly
the inconsistent condition in which overconfidence occurred. Thus, from the
perspective we propose, this inconsistent result can be reconciled.

Keren (1988) A strict distinction between perceptual judgment and intellec-

tual judgment cannot be derived from many views of perception, such as
signal-detection theory (Tanner & Swets, 1954). Reflecting on this fact, Keren
(1988) proposed a slightly modified hypothesis: The more perception-like a
task is, the less overconfident and the better calibrated participants will be.
"As a task requires additional higher processing and transformation of the
original sensory input, different kinds of possible cognitive distortions may
exist (such as inappropriate inferences) that may limit the ability to accurately
monitor our higher cognitive processes" (Keren, 1988, p. 99).
Keren (1988, Experiment 1) used general-knowledge questions and two
kinds of perceptual tasks, one of them more difficult than the general-
knowledge task, the other less difficult. Keren tested the hypothesis that con-
fidence judgments in perceptual tasks are better calibrated than in general-
knowledge tasks. He could not support it, however. Instead, he found an
anomaly: The comparison between the general-knowledge task and the more
difficult perceptual task reversed the hard-easy effect (see Figure 7.10). As
derived in Prediction 5, this puzzling reversal is implied by PMM theory if
the Landolt rings used in the more difficult perceptual task were not selected
for perceptual illusions, as seems to be the case (Keren, 1988, p. 100).
Note that the kind of general-knowledge questions used (population of cities
or countries, and distances between cities) would easily permit defining a ref-
erence class in a known environment and obtaining representative samples.
But no representative sample of general-knowledge questions was generated.
This lack makes the other predictions from PMM theory coincide with Keren's
(1988): overconfidence in general-knowledge, and zero overconfidence in the
two perceptual tasks. Results show this outcome, except for the large-gap Lan-
dolt rings condition, which generated considerable underconfidence. PMM
 PROBABILISTIC MENTAL MODELS 163

theory cannot account for the latter, nor can the notion of degree of perception-
likeness.
A second perceptual task was letter identification. In Experiment 3, Keren
(1988) used two letter-identification tasks, which were identical except that
the exposure time of the letters to be recognized was either short or long. Mean
percentages correct were 63.5 for short and 77.2 for long exposures. According
to earlier explanations, such as participants' insensitivity to task difficulty, a
hard-easy effect should result. According to PMM theory, however, the hard-
easy effect should be zero, because both tasks were generated by the same
sampling process (Prediction 4). In fact, Keren (1988, p. 112) reported that in
both tasks, overconfidence was not significantly different from zero. Prediction
4 accounts for this disappearance of the hard-easy effect in a situation in
which differences in percentage correct were large.

The Brunswikian Perspective

PMM theory draws heavily on the Brunswikian notions of a natural environ-
ment known to an individual, reference classes in this environment, and rep-
resentative sampling from a reference class. We went beyond the Brunswikian
focus on achievement (rather than process) by providing a theoretical frame-
work of the processes that determine choice, confidence, and frequency judg-
ment.
Choice and confidence are a result of a cue-testing and activation cycle,
which is analogous to Newell and Simon's (1972) postulate that "problem solv-
ing takes place by search in a problem space" (p. 809). Furthermore, the em-
phasis on the structure of the task in PMM theory is similar to Newell and
Simon's proposition that "the structure of the task determines the possible
structures of the problem space" (p. 789). Unlike PMM theorists, however,
Newell and Simon also assumed in the tasks they studied (cryptarithmetic,
logic, and chess) a relatively simple mapping between the external structure
of the task and the internal representation in a problem space (see Allport,
1975). Although it is cued by the task structure, we assume that a PMM (the
functional equivalent of a problem space) has a large surplus structure (the
reference class and the cues), which is taken from a known structure in
the problem solver's natural environment. The emphasis on the structure of
everyday knowledge or environment (as distinguished from the task environ-
ment) has been most forcefully defended by Brunswik. Although Newell and
Simon (1972, p. 874) called Brunswik and Tolman "the real forerunners" of
their work, they seem not to distinguish clearly between the notions of a prob-
abilistic everyday environment and a task environment. This theory is an at-
tempt to combine both views. Brunswik's focus on achievement (during his
behavioristic phase; see Leary, 1987) corresponds more closely to the part of
research on probabilistic judgment that focuses on calibration, rather than on
the underlying cognitive processes.
The importance of the cognitive representation of the task was studied by
the Wurzburg school and emphasized in Gestalt theoretical accounts of think-
164 BOUNDED RATIONALITY

ing (e.g., Duncker, 1935/1945), and this issue has regained favor (e.g., Brehmer,
1988; Hammond, Stewart, Brehmer, & Steinmann, 1975). In their review, Ein-
horn and Hogarth (1981) emphasized that "the cognitive approach has been
concerned primarily with how tasks are represented. The issue of why tasks
are represented in particular ways has not yet been addressed" (p. 57). PMM
theory addresses this issue. Different tasks, such as confidence and frequency
tasks, cue different reference classes and different probability cues from known
environments. It is these environments that provide the particular represen-
tation, the PMM, of a task.
Many parts of PMM theory need further expansion, development, and test-
ing. Open issues include the following: (a) What reference class is activated?
For city comparisons, this question has a relatively clear answer, but in gen-
eral, more than one reference class can be constructed to solve a problem, (b)
Are cues always generated according to their rank in the cue validity hierar-
chy? Alternative models of cue generation could relax this strong assumption,
assuming, for instance, that the first cue generated is the cue activated in the
last problem. The latter would, however, decrease the percentage of correct
answers, (c) What are the conditions under which we may expect PMMs to be
well adapted? There exists a large body of neo-Bmnswikian research that, in
general, indicates good adaptation but also points out exceptions (e.g., Arme-
lius, 1979; Bjorkman, 1987; Brehmer & Joyce, 1988; Hammond & Wascoe,
1980). (d) What are the conditions under which cue substitution without cue
integration is superior to multiple cue integration? PMM theory assumes a pure
cue substitution model—a cue that cannot be activated can be replaced by any
other cue—without integration of two or more cues. We focused on the sub-
stitution and not the integration aspect of Brunswik's vicarious functioning
(see Gigerenzer & Murray, 1987, pp. 66-81), in contrast to the multiple regres-
sion metaphor of judgment. Despite its simplicity, the substitution model pro-
duces zero overconfidence and a large number of correct answers, if the PMM
is well adapted. There may be more reasons for simple substitution models.
Armelius and Armelius (1974), for instance, reported that participants were
well able to use ecological validities, but not the correlations between cues. If
the latter is the case, then multiple cue integration may not work well.

Conclusions

We conjecture that confidence in one's knowledge of the kind studied here—

immediate and spontaneous rather than a product of long-term reflection—is
largely determined by the structure of the task and the structure of a corre-
sponding, known environment in a person's long-term memory. We provided
experimental evidence for this hypothesis by showing how changes in the task
(confidence vs. frequency judgment) and in the relationship between task and
environment (selected vs. representative sampling) can make the two stable
effects reported in the literature—overconfidence and the hard-easy effect—
emerge, disappear, and invert at will. We have demonstrated a new phenom-
 PROBABILISTIC MENTAL MODELS 165

enon, the confidence-frequency effect. One cannot speak of a general over-

confidence bias anymore, in the sense that it relates to deficient processes of
cognition or motivation. In contrast, participants seem to be able to make fine
conceptual distinctions—confidence versus frequency—of the same kind as
probabilists and statisticians do. Earlier attempts postulating general deficien-
cies in information processing or motivation cannot account for the experi-
mental results predicted by PMM theory and confirmed in two experiments.
PMM theory seems to be the first theory in this field that gives a coherent
account of these various effects by focusing on the relation between the struc-
ture of the task, the structure of a corresponding environment, and a PMM.
 8
Reasoning the Fast and Frugal Way

O rganisms make inductive inferences. Darwin (1872/1965) observed that peo-

ple use facial cues, and such as eyes that waver and lids that hang low, to infer
a person's guilt. Male toads, roaming through swamps at night, use the pitch
of a rival's croak to infer its size when deciding whether to fight (Krebs &
Davies, 1987). Stockbrokers must make fast decisions about which of several
stocks to trade or invest when only limited information is available. The list
goes on. Inductive inferences are typically based on uncertain cues: The eyes
can deceive, and so can a tiny toad with a deep croak in the darkness.
How does an organism make inferences about unknown aspects of the en-
vironment? There are three directions in which to look for an answer. From
Pierre Laplace to George Boole to Jean Piaget, many scholars have defended
the now classical view that the laws of human inference are the laws of prob-
ability and statistics (and to a lesser degree logic, which does not deal as easily
with uncertainty). Indeed, the Enlightenment probabilists derived the laws of
probability from what they believed to be the laws of human reasoning (Das-
ton, 1988). Following this time-honored tradition, much contemporary re-
search in psychology, behavioral ecology, and economics assumes standard
statistical tools to be the normative and descriptive models of inference and
decision making. Multiple regression, for instance, is both the economist's uni-
versal tool (McCloskey, 1985) and a model of inductive inference in multiple-
cue learning (Hammond, 1990) and clinical judgment (Brehmer, 1994); Bayes's
rule is a model of how animals infer the presence of predators or prey (Ste-
phens & Krebs, 1986) as well as of human reasoning and memory (Anderson,
1990). This Enlightenment view that probability theory and human reasoning
are two sides of the same coin crumbled in the early nineteenth century but
has remained strong in psychology and economics.
In the past 25 years, this stronghold came under attack by proponents of
the heuristics-and-biases program, who concluded that human inference is sys-
tematically biased and error prone, suggesting that the laws of inference are

The work on which this chapter is based was coauthored with D. G. Goldstein.

166
 REASONING THE FAST AND FRUGAL WAY 167

quick-and-dirty heuristics and not the laws of probability (Tversky & Kahne-
man, 1974). This second perspective appears diametrically opposed to the clas-
sical rationality of the Enlightenment, but this appearance is misleading. It has
retained the normative kernel of the classical view. For example, a discrepancy
between the dictates of classical rationality and actual reasoning is what de-
fines a reasoning error in this program. Both views accept the laws of proba-
bility and statistics as normative, but they disagree about whether humans can
live up to these norms.
Many experiments have been conducted to test the validity of these two
views, identifying a host of conditions under which the human mind appears
more rational or irrational. But most of this work has dealt with simple situ-
ations, such as Bayesian inference with binary hypotheses, one single piece of
binary data, and all the necessary information conveniently laid out for the
participant (Chapter 6). In many real-world situations, however, there are mul-
tiple pieces of information, which are not independent, but redundant. Here,
Bayes's rule and other "rational" algorithms quickly become mathematically
complex and computationally intractable, at least for ordinary human minds.
These situations make neither of the two views look promising. If one were to
apply the classical view to such complex real-world environments, this would
suggest that the mind is a supercalculator like a Laplacean demon (Wimsatt,
1976)—carrying around the collected works of Kolmogoroff, Fisher, or Ney-
man—and simply needs a memory jog, like the slave in Plato's Meno. On the
other hand, the heuristics-and-biases view of human irrationality would lead
us to believe that humans are hopelessly lost in the face of real-world com-
plexity, given their supposed inability to reason according to the canon of
classical rationality, even in simple laboratory experiments.
There is a third way to look at inference, focusing on the psychological
and ecological rather than on logic and probability theory. This view ques-
tions classical rationality as a universal norm and thereby questions the very
definition of "good" reasoning on which both the Enlightenment and the
heuristics-and-biases views were built. Herbert Simon, possibly the best
known proponent of this third view, proposed looking for models of bounded
rationality instead of classical rationality. Simon (1956, 1982) argued that in-
formation-processing systems typically need to satisfies rather than to opti-
mize. Satisficing, a blend of sufficing and satisfying, is a word of Scottish or-
igin, which Simon uses to characterize strategies that successfully deal with
conditions of limited time, knowledge, or computational capacities. His con-
cept of satisficing postulates, for instance, that an organism would choose the
first object (a mate, perhaps) that satisfies its aspiration level—instead of the
intractable sequence of taking the time to survey all possible alternatives, es-
timating probabilities and utilities for the possible outcomes associated with
each alternative, calculating expected utilities, and choosing the alternative
that scores highest.
Let us stress that Simon's notion of bounded rationality has two sides,
one cognitive and one ecological. As early as in Administrative Behavior
(1945), he emphasized the cognitive limitations of real minds as opposed to
168 BOUNDED RATIONALITY

the omniscient Laplacean demons of classical rationality. As early as in his

Psychological Review article titled "Rational Choice and the Structure of the
Environment" (1956), Simon emphasized that minds are adapted to real-
world environments. The two go in tandem: "Human rational behavior is
shaped by a scissors whose two blades are the structure of task environments
and the computational capabilities of the actor" (Simon, 1990, p. 7). For the
most part, however, theories of human inference have focused exclusively
on the cognitive side, equating the notion of bounded rationality with the
statement that humans are limited information processors, period. In a
Procrustean-bed fashion, bounded rationality became almost synonymous
with heuristics and biases, thus paradoxically underpinning classical ration-
ality as the normative standard for both biases and bounded rationality (for
a discussion of this confusion see Lopes, 1992). Simon's insight that the
minds of living systems should be understood relative to the environment
in which they evolved, rather than to the tenets of classical rationality, has
had little impact so far in research on human inference. Simple psychological
strategies that were observed in human inference, reasoning, or decision mak-
ing were often discredited without a fair trial, because they looked so stupid
by the norms of classical rationality. For instance, when Keeney and Raiffa
(1993) discussed the lexicographic ordering strategy they had observed in
practice—a procedure related to the models of bounded rationality we pro-
pose in this chapter—they concluded that this procedure "is naively simple"
and "will rarely pass a test of 'reasonableness' " (p. 78). They did not report
such a test. We do.
Initially, the concept of bounded rationality was only vaguely defined, often
as that which is not classical economics, and one could "fit a lot of things into
it by foresight and hindsight," as Simon (1992b, p. 18) himself put it. We wish
to do more than oppose the Laplacean demon view. We strive to come up with
something positive that could replace this unrealistic view of mind. What are
these simple, intelligent heuristics capable of making near-optimal inferences?
How fast and how accurate are they? In this chapter, we propose a class of
heuristics that exhibit bounded rationality in both of Simon's senses. These
"fast and frugal heuristics" operate with simple psychological principles that
satisfy the constraints of limited time, knowledge, and computational might,
rather than those of classical rationality. At the same time, they are designed
to be fast and frugal without a significant loss of inferential accuracy, because
they can exploit the structure of environments.
The chapter is organized as follows. We begin by describing the task the
cognitive heuristics are designed to address, the Take The Best heuristic, and
the real-world environment on which its performance will be tested. Next, we
report on a competition in which the heuristic competes with "rational" strat-
egies in making inferences about a real-world environment. The "rational"
strategies start with an advantage: They use more time, information, and com-
putational might to make inferences. Finally, we study variants of the heuristic
that make faster inferences and get by with even less knowledge.
 REASONING THE FAST AND FRUGAL WAY 169

The Task

We deal with inferential tasks in which a choice must be made between two
alternatives on a quantitative dimension. Consider the following example:
Which city has a larger population? (a) Hamburg (b) Cologne.
Two-alternative-choice tasks occur in various contexts in which inferences
need to be made with limited time and knowledge, such as in decision making
and risk assessment during driving (e.g., exit the highway now or stay on);
treatment-allocation decisions (e.g., who to treat first in the emergency room:
the 80-year-old heart attack victim or the 16-year-old car accident victim); and
financial decisions (e.g., whether to buy or sell in the trading pit). Inference
concerning population demographics, such as city populations of the past,
present, and future (e.g., Brown & Siegler, 1993), is of importance to people
working in urban planning, industrial development, and marketing. Popula-
tion demographics, which is better understood than, say, the stock market, will
serve us later as a "drosophila" environment that allows us to analyze the
behavior of heuristics.
We study two-alternative-choice tasks in situations in which a person has
to make an inference based solely on knowledge retrieved from memory. We
refer to this as inference from memory, as opposed to inference from givens.
Inference from memory involves search in declarative knowledge and has been
investigated in studies of, inter alia, confidence in general knowledge (e.g.,
Juslin, 1994; Sniezek & Buckley, 1993); the effect of repetition on belief (e.g.,
Hertwig, Gigerenzer, & Hoffrage, 1997); hindsight bias (e.g., Fischhoff, 1977);
quantitative estimates of area and population of nations (Brown & Siegler,
1993); and autobiographic memory of time (Huttenlocher, Hedges, & Prohaska,
1988). Studies of inference from givens, on the other hand, involve making
inferences from information presented by an experimenter (e.g., Hammond,
Hursch, & Todd, 1964). In the tradition of Ebbinghaus's nonsense syllables,
attempts are often made here to prevent individual knowledge from having an
impact on the results by using problems about hypothetical referents instead
of actual ones. For instance, in celebrated judgment and decision-making tasks,
such as the "cab" problem and the "Linda" problem, all the relevant infor-
mation is provided by the experimenter, and individual knowledge about cabs
and hit-and-run accidents, or feminist bank tellers, is considered of no rele-
vance (Gigerenzer & Murray, 1987). As a consequence, limited knowledge or
individual differences in knowledge play a small role in inference from givens.
In contrast, the heuristics proposed in this chapter perform inference from
memory, they use limited knowledge as input, and as we will show, they can
actually profit from a lack of knowledge.
Assume that a person does not know or cannot deduce the answer to the
Hamburg-Cologne question but needs to make an inductive inference from
related real-world knowledge. How is this inference derived? How can we
predict choice (Hamburg or Cologne) from a person's state of knowledge?
170 BOUNDED RATIONALITY

Theory

The cognitive heuristics we propose are realizations of a framework for mod-

eling inferences from memory, the theory of probabilistic mental models
(Chapter 7). The theory of probabilistic mental models assumes that inferences
about unknown states of the world are based on probability cues. The theory
relates three visions: (a) Inductive inference needs to be studied with respect
to natural environments, as emphasized by Brunswik and Simon; (b) inductive
inference is carried out by satisficing algorithms, as emphasized by Simon;
and (c) inductive inferences are based on frequencies of events in a reference
class, as proposed by Reichenbach and other frequentist statisticians. The the-
ory of probabilistic mental models accounts for choice and confidence, but
only choice is addressed in this chapter.
The major thrust of the theory is that it replaces the canon of classical
rationality with simple, plausible psychological mechanisms of inference—
mechanisms that a mind can actually carry out under limited time and knowl-
edge and that could have possibly arisen through evolution. Most traditional
models of inference, from linear multiple regression models to Bayesian mod-
els to neural networks, try to find some optimal integration of all information
available: Every bit of information is taken into account, weighted, and com-
bined in a computationally expensive way. The family of heuristics based on
PMM theory does not implement this classical ideal. Search in memory for
relevant information is reduced to a minimum, and there is no integration (but
rather a substitution) of pieces of information. These boundedly rational heu-
ristics dispense with the fiction of the omniscient Laplacean demon, who has
all the time and knowledge to search for all relevant information, to compute
the weights and covariances, and then to integrate all this information into an
inference.

Limited Knowledge
A PMM is an inductive device that uses limited knowledge to make fast in-
ferences. Different from mental models of syllogisms and deductive inference
(Johnson-Laird, 1983), which focus on the logical task of truth preservation
and where knowledge is irrelevant (except for the meanings of connectives
and other logical terms), PMMs perform intelligent guesses about unknown
features of the world, based on uncertain indicators. To make an inference
about which of two objects, a or b, has a higher value, knowledge about a
reference class R is searched, with a, b € R. In our example, knowledge about
the reference class "cities in Germany" could be searched. The knowledge
consists of probability cues Ct(i = I,. . . , n) and the cue values ai and bf of
the objects for the j'th cue. For instance, when making inferences about pop-
ulations of German cities, the fact that a city has a professional soccer team in
the major league (Bundesliga] may come to a person's mind as a potential cue.
That is, when considering pairs of German cities, if one city has a soccer team
 REASONING THE FAST AND FRUGAL WAY 171

Figure 8.1 Illustration of bounded search through limited knowledge. Objects

a, b, and c are recognized; object d is not. Cue values are positive (+) or
negative (—); missing knowledge is shown by question marks. Cues are or-
dered according to their validities. To infer whether a > b, the Take The
Best heuristic looks up only the cue values in the shaded space; to infer
whether b > c, search is bounded to the dotted space. The other cue values
are not looked up.

in the major league and the other does not, then the city with the team is
likely, but not certain, to have the larger population.
Limited knowledge means that the matrix of objects by cues has missing
entries (i.e., objects, cues, or cue values may be unknown). Figure 8.1 models
the limited knowledge of a person. She has heard of three German cities, a, b,
and c, but not of d (represented by three positive and one negative recognition
values). She knows some facts (cue values) about these cities with respect to
five binary cues. For a binary cue, there are two cue values, positive (e.g., the
city has a soccer team) or negative (it does not). Positive refers to a cue value
that signals a higher value on the target variable (e.g., having a soccer team is
correlated with a large population). Unknown cue values are shown by a ques-
tion mark. Because she has never heard of d, all cue values for object d are,
by definition, unknown.
People rarely know all the information on which an inference could be
based, that is, knowledge is limited. We model limited knowledge in two re-
spects: A person can have (a) incomplete knowledge of the objects in the ref-
erence class (e.g., she recognizes only some of the cities), (b) limited knowledge
of the cue values (facts about cities), or (c) both. For instance, a person who
does not know all of the cities with soccer teams may know some cities with
positive cue values (e.g., Munich and Hamburg certainly have teams), many
with negative cue values (e.g., Heidelberg and Potsdam certainly do not have
teams), and several cities for which cue values will not be known.

The Take The Best Heuristic

The first fast and frugal heuristic presented is called Take The Best, because
its policy is "take the best, ignore the rest." It is the basic heuristic in the
172 BOUNDED RATIONALITY

PMM framework. Variants that work faster or with less knowledge are de-
scribed later. We explain the steps of Take The Best for binary cues (the heu-
ristic can be easily generalized to many valued cues), using Figure 8.1 for
illustration.
Take The Best assumes a rank order of cues according to their subjective
validities (as in Figure 8.1). We call the highest ranking cue (that discriminates
between the two alternatives) the best cue. The heuristic is shown in the form
of a flow diagram in Figure 8.2.

Step 0: Recognition Heuristic The recognition heuristic is invoked when the

mere recognition of an object is a predictor of the target variable (e.g., popu-
lation). The recognition heuristic states the following: If only one of the two
objects is recognized, then choose the recognized object. If neither of the two
objects is recognized, then choose randomly between them. If both of the ob-
jects are recognized, then proceed to Step 1.
Example: If a person in the knowledge state shown in Figure 8.1 is asked
to infer which of city a and city d has more inhabitants, the inference will be
city a, because the person has never heard of city d before.

Step 7. Search Rule Choose the cue with the highest validity that has not yet
been tried for this choice task. Look up the cue values of the two objects.

Step 2. Stopping Rule If one object has a positive cue value and the other does
not (i.e., either a negative or an unknown value; see Figure 8.3), then stop

Figure 8.2 Flow diagram of the Take The Best heuristic.

 REASONING THE FAST AND FRUGAL WAY 173

search and go on to Step 3. Otherwise go back to Step 1 and search for another
cue. If no further cue is found, then guess.

Step 3. Decision Rule (one-reason decision making) Predict that the object with
the positive cue value has the higher value on the criterion.
Examples: Suppose the task is judging which of city a or b is larger (Figure
8.1). Both cities are recognized (Step 0), and search for the best cue results in
a positive and a negative cue value for Cue 1 (Step 1). The cue discriminates,
and search is terminated (Step 2). The person makes the inference that city a
is larger (Step 3).
Suppose now the task is judging which city b or c is larger. Both cities are
recognized (Step 0), and search for the cue values results in a negative cue
value on object b for Cue 1, but the corresponding cue value for object c is
unknown (Step 1). The cue does not discriminate, so search is continued (Step
2). Search for the next cue results in positive and negative cue values for Cue
2 (Step 1). This cue discriminates and search is terminated (Step 2). The person
makes the inference that city b is larger (Step 3).
The features of this heuristic are (a) search extends through only a portion
of the total knowledge in memory (as shown by the shaded and dotted parts
of Figure 8.1) and is stopped immediately when the first discriminating cue is
found, (b) the algorithm does not attempt to integrate information but uses
one-reason decision making instead, and (c) the total amount of information
processed is contingent on each task (pair of objects) and varies in a predict-
able way among individuals with different knowledge. This fast and compu-
tationally simple heuristic is a model of bounded rationality rather than of
classical rationality. There is a close parallel with Simon's concept of "satis-
ficing": Take The Best stops search after the first discriminating cue is found,
just as Simon's satisficing algorithm stops search after the first option that
meets an aspiration level.
The heuristic is hardly a standard statistical tool for inductive inference: It
does not use all available information, it is non-compensatory and nonlinear,

Figure 8.3 Stopping rule. A cue discriminates between two alternatives if one
has a positive cue value and the other does not. The four discriminating
cases are shaded. If a cue discriminates, search is stopped.
174 BOUNDED RATIONALITY

and variants of it can violate transitivity. Thus it differs from standard linear
tools for inference such as multiple regression, as well as from nonlinear neu-
ral networks that are compensatory in nature. Take The Best is noncompen-
satory because only the best discriminating cue determines the inference or
decision; no combination of other cue values can override this decision. In this
way, the heuristic does not conform to the classical economic view of human
behavior (e.g., Becker, 1976), where, on the assumption that all aspects can be
reduced to one dimension (e.g., money), there always exists a trade-off between
commodities or pieces of information. That is the heuristic violates the Ar-
chimedian axiom, which implies that for any multidimensional object a (a^
a2, . . . ,aj preferred to b (b1, b2, . . . ,bn), where a1 dominates ba, this preference
can be reversed by taking multiples of any one or a combination of b2, b3, . . . ,
bn. As we discuss, variants of this heuristic also violate transitivity, one of the
cornerstones of classical rationality (McClennen, 1990).

Empirical Evidence
Despite their flagrant violation of traditional standards of rationality, Take The
Best and PMM theory have been successful in integrating various extant phe-
nomena in inference from memory and predicting novel phenomena. These
include conditions under which overconfidence occurs, disappears, and in-
verts to underestimation (Gigerenzer, 1993b; Juslin, 1993, 1994; Juslin, Win-
man, & Persson, 1995; but see Griffin & Tversky, 1992), and those in which the
hard-easy effect occurs, disappears, and inverts—predictions that have been
experimentally confirmed by Hoffrage (1994) and by Juslin (1993).
Fast and frugal heuristics allow for predictions of individual choices, in-
cluding individual differences based on each person's knowledge. Broder (in
press) reported that when search for information is costly, about 65% of the
participants' choices were consistent with Take The Best, compared to fewer
than 10% with a linear strategy. (For similar results, see Rieskamp & Hoffrage,
1999.) Hoffrage and Hertwig (1999) showed that a memory updating model
with Take The Best could correctly predict some 75% of all individual occur-
rences of hindsight bias. Goldstein and Gigerenzer (1999) showed that the rec-
ognition heuristic predicted individual participants' choices in about 90% to
100% of all cases, even when participants were taught information that sug-
gested doing otherwise (negative cue values for the recognized objects). Among
the evidence for the empirical validity of Take The Best are the tests of a bold
prediction, the less-is-more effect, which postulates conditions under which
people with little knowledge make better inferences than those who know
more. This surprising prediction has been experimentally confirmed. For in-
stance, U.S. students make slightly more correct inferences about German city
populations (about which they know little) than about U.S. cities, and vice
versa for German students (Gigerenzer, 1993b; Goldstein & Gigerenzer, 1999;
Hoffrage, 1994). The recognition heuristic has been successfully applied to
stock investment (Borges et al., 1999); on rumor-based stock market trading,
see DiFonzo (1994). Other species also practice one-reason decision making
 REASONING THE FAST AND FRUGAL WAY 175

closely resembling Take The Best, such as when female guppies choose be-
tween males on the basis of an order of cues (Dugatkin, 1996). For general
reviews, see Gigerenzer et al. (1999) and McClelland and Bolger (1994).
The reader familiar with the original heuristic presented in Gigerenzer et
al. (1991, see Chapter 7) will have noticed that we simplified the stopping
rule.1 In the present version, search is already terminated if one object has a
positive cue value and the other does not, whereas in the earlier version,
search was terminated only when one object had a positive value and the other
a negative one (cf. Figure 7.3 in Chapter 7 with Figure 8.3 in this chapter).
This change follows empirical evidence that participants tend to use this faster,
simpler stopping rule (Hoffrage, 1994).
This chapter does not attempt to provide further empirical evidence. For
the moment, we assume that the model is descriptively valid and investigate
how accurate this fast and frugal heuristic is in drawing inferences about un-
known aspects of a real-world environment. Can a heuristic based on simple
psychological principles that violate the norms of classical rationality make a
fair number of accurate inferences?

The Environment

We tested the performance of Take The Best on how accurately it made infer-
ences about a real-world environment. The environment was the set of all
cities in Germany with more than 100,000 inhabitants (83 cities after German
reunification), with population as the target variable. The model of the envi-
ronment consisted of 9 binary ecological cues and the actual 9 X 8 3 cue values.
The full model of the environment is shown in Gigerenzer and Goldstein
(1996a).
Each cue has an associated validity that is indicative of its predictive power.
The ecological validity of a cue is the relative frequency with which the cue
correctly predicts the target, defined with respect to the reference class (e.g.,
all German cities with more than 100,000 inhabitants). For instance, if one
checks all pairs in which one city has a soccer team but the other city does
not, one finds that in 87% of these cases, the city with the team also has the
higher population. This value is the ecological validity of the soccer team cue.
The validity vi of the rth cue is

vi= p[t(a) > t(b) \ a1 is positive and b1 is negative],

where t(a] and t(b] are the values of objects a and b on the target variable t
and p is a probability measured as a relative frequency in R.
The ecological validity of the nine cues ranged over the whole spectrum:
from .51 (only slightly better than chance) to 1.0 (certainty), as shown in Table

1. Also, we now use the term stopping rule instead of activation rule.
176 BOUNDED RATIONALITY

Table 8.1 Cues, ecological validities, and discrimination rates

Ecological Discrimination
Cue validity rate

National capital (Is the city the national

capital?) 1.00 .02
Exposition site (Was the city once an expo-
sition site?) .91 .25
Soccer team (Does the city have a team in
the major league?) .87 .30
Intercity train (Is the city on the Intercity
line?) .78 .38
State capital (Is the city a state capital?) .77 .30
License plate (Is the abbreviation only one
letter long?) .75 .34
University (Is the city home to a univer-
sity?) .71 .51
Industrial belt (Is the city in the industrial
belt?) .56 .30
East Germany (Was the city formerly in
East Germany?) .51 .27

8.1. A cue with a high ecological validity, however, is often not useful if its
discrimination rate is small.
Table 8.1 also shows the discrimination rates for each cue. The discrimi-
nation rate of a cue is the relative frequency with which the cue discriminates
between any two objects from the reference class. The discrimination rate is a
function of the distribution of the cue values and the number N of objects in
the reference class. Let the relative frequencies of the positive and negative
cue values be x and y, respectively. Then the discrimination rate di of the j'th
cue is

as an elementary calculation shows. Thus, if AT is very large, the discrimination

rate is approximately 2xiyi.2 The larger the ecological validity of a cue, the
better the inference. The larger the discrimination rate, the more often a cue
can be used to make an inference. In the present environment, ecological va-
lidities and discrimination rates are negatively correlated. The redundancy of

2. For instance, if N = 2 and one cue value is positive and the other negative (x, =
Yi = .5), d; = 1.0. If N increases, with x, and yt held constant, then dt decreases and
converges to 2x,y,.
 REASONING THE FAST AND FRUGAL WAY 177

cues in the environment, as measured by pairwise correlations between cues,

ranges between -.25 and .54, with an average absolute value of .19.3

The Competition

The question of how well a fast and frugal heuristic performs in a real-world
environment has rarely been posed in research on inductive inference. The
present simulations seem to be the first to test how well one-reason decision
making does compared with standard integration strategies, which require
more knowledge, time, and computational power. This question is important
for Simon's postulated link between the cognitive and the ecological: If the
simple psychological principles in fast and frugal heuristics are tuned to ec-
ological structures, these heuristics should not fail outright. We propose a com-
petition between various inferential strategies. The contest will go to the strat-
egy that scores the highest proportion of correct inferences (accuracy) using
the smallest number of cues (frugality).

Simulating Limited Knowledge

We simulated people with varying degrees of knowledge about cities in Ger-
many. Limited knowledge can take two forms. One is limited recognition of
objects in the reference class. The other is limited knowledge about the cue
values of recognized objects. To model limited recognition knowledge, we sim-
ulated people who recognized between 0 and 83 German cities. To model lim-
ited knowledge of cue values, we simulated 6 basic classes of people, who
knew 0%, 10%, 20%, 50%, 75%, or 100% of the cue values associated with
the objects they recognized. Combining the two sources of limited knowledge
resulted in 6 X 84 types of people, each having different degrees and kinds of
limited knowledge. Within each type of people, we created 500 simulated in-

3. There are various other measures of redundancy besides pairwise correlation. The
important point is that whatever measure of redundancy one uses, the resultant value
does not have the same meaning for all strategies. For instance, all that counts for Take
The Best is what proportion of correct inferences the second cue adds to the first in the
cases where the first cue does not discriminate, how much the third cue adds to the first
two in the cases where they do not discriminate, and so on. If a cue discriminates, search
is terminated, and the degree of redundancy in the cues that were not included in the
search is irrelevant. Integration strategies, in contrast, integrate all information and, thus,
always work with the total redundancy in the environment (or knowledge base). For
instance, when deciding among objects a, b, c, and d in Figure 8.1, the cue values of
Cues 3, 4, and 5 do not matter from the point of view of Take The Best (because search
is terminated before reaching Cue 3). However, the values of Cues 3, 4, and 5 affect the
redundancy of the ecological system, from the point of view of all integration algorithms.
The lesson is that the degree of redundancy in an environment depends on the kind of
strategy that operates on the environment. One needs to be cautious in interpreting
measures of redundancy without reference to a strategy.
178 BOUNDED RATIONALITY

dividuals, who differed randomly from one another in the particular objects
and cue values they knew. All objects and cue values known were determined
randomly within the appropriate constraints, that is, a certain number of ob-
jects known, a certain total percentage of cue values known, and the validity
of the recognition heuristic (as explained in the following paragraph).
The simulation needed to be realistic in the sense that the simulated people
could invoke the recognition heuristic. Therefore, the sets of cities the simu-
lated people knew had to be carefully chosen so that the recognized cities were
larger than the unrecognized ones a certain percentage of the time. We per-
formed a survey to get an empirical estimate of the actual covariation between
recognition of cities and city populations. Let us define the recognition validity
a to be the probability, in a reference class, that one object has a greater value
on the target variable than another, in the cases where one object is recognized
and the other is not:
a = p[t(a) > t(b)a is positive and br is negative],
where t(a) and t(b] are the values of objects a and b on the target variable t, ar
and br are the recognition values of a and b, and p is a probability measured
as a relative frequency in R.
In a pilot study of 26 undergraduates at the University of Chicago, we found
that the cities they recognized (within the 83 largest in Germany) were larger
than the cities they did not recognize in about 80% of all possible comparisons.
We incorporated this value into our simulations by choosing sets of cities (for
each knowledge state, i.e., for each number of cities recognized) where the
known cities were larger than the unknown cities in about 80% of all cases.
Thus the cities known by the simulated individuals had the same relationship
between recognition and population as did those of the human individuals.
Let us first look at the performance of Take The Best.

Testing the Take The Best Heuristic

We tested how well individuals using Take The Best did at answering real-
world questions such as, Which city has more inhabitants: (a) Heidelberg or
(b) Bonn? Each of the 500 simulated individuals in each of the 6 X 84 types
was tested on the exhaustive set of 3,403 city pairs, resulting in a total of 500
X 6 X 84 X 3,403 tests, that is, about 858 million.
The curves in Figure 8.4 show the average proportion of correct inferences
for each proportion of objects and cue values known. The x axis represents
the number of cities recognized, and the y axis shows the proportion of correct
inferences that Take The Best drew. Each of the 6 X 84 points that make up
the six curves is an average proportion of correct inferences taken from 500
simulated individuals, who each made 3,403 inferences.
When the proportion of cities recognized was zero, the proportion of correct
inferences was at chance level (.5). When up to half of all cities were recog-
nized, performance increased at all levels of knowledge about cue values. The
maximum percentage of correct inferences was around 77%. The striking re-
 REASONING THE FAST AND FRUGAL WAY 179

Figure 8.4 Correct inferences about the population of German cities (two-
alternative-choice tasks) by Take The Best. Inferences are based on actual in-
formation about the 83 largest cities and nine cues for population (see text).
Limited knowledge of the simulated individuals is varied across two dimen-
sions: (a) the number of cities recognized (x axis) and (b) the percentage of
cue values known (the six curves).

suit was that this maximum was not achieved when individuals knew all cue
values of all cities, but rather when they knew less. This result shows the
ability of the heuristic to exploit limited knowledge, that is, to do best when
not everything is known. Thus, Take The Best produces the less-is-more effect.
At any level of limited knowledge of cue values, learning more German cities
will eventually cause a decrease in the proportion correct. Take, for instance,
the curve where 75% of the cue values were known and the point where the
simulated participants recognized about 60 German cities. If these individuals
learned about the remaining German cities, their proportion correct would de-
crease. The rationale behind the less-is-more effect is the recognition heuristic,
and it can be understood best from the curve that reflects 0% of total cue values
known. Here, all decisions are made on the basis of the recognition heuristic,
or by guessing. On this curve, the recognition heuristic comes into play most
when half of the cities are known, so it takes on an inverted-U shape. When
half the cities are known, the recognition heuristic can be activated most often,
that is, for roughly 50% of the questions. Because we set the recognition va-
lidity in advance, 80% of these inferences will be correct. In the remaining
half of the questions, when recognition cannot be used (either both cities are
recognized or both cities are unrecognized), then the organism is forced to
guess and only 50% of the guesses will be correct. Using the 80% effective
180 BOUNDED RATIONALITY

recognition validity half of the time and guessing the other half of the time,
the organism scores 65% correct, which is the peak of the bottom curve. The
mode of this curve moves to the right with increasing knowledge about cue
values. Note that even when a person knows everything, all cue values of all
cities, there are states of limited knowledge in which the person would make
more accurate inferences. We are not going to discuss the conditions of this
counterintuitive effect and the supporting experimental evidence here (see
Goldstein & Gigerenzer, 1999). Our focus is on how much better integration
strategies can do in making inferences.

Integration Strategies

We asked several colleagues in the fields of statistics and economics to devise

decision strategies that would do better than Take The Best. The five integra-
tion strategies we simulated and pitted against Take The Best in a competition
were among those suggested by our colleagues. These competitors include
"proper" and "improper" linear models (Dawes, 1979; Lovie & Lovie, 1986).
These strategies, in contrast to Take The Best, embody two classical principles
of rational inference: (a) complete search—they use all available information
(cue values)—and (b) complete integration—they combine all these pieces of
information into a single value. In short, we refer in this chapter to strategies
that satisfy these principles as "rational" (in quotation marks) strategies.

Contestant 1: Tallying Let us start with a simple integration strategy: tallying

of positive evidence (Goldstein, 1994). In this strategy, the number of positive
cue values for each object is tallied across all cues (i — 1, . . . , n], and the
object with the largest number of positive cue values is chosen. Integration
strategies are not based (at least explicitly) on the recognition heuristic. For
this reason, and to make the integration strategies as strong as possible, we
allow all the integration strategies to make use of recognition information (the
positive and negative recognition values; see Figure 8.1). Integration strategies
treat recognition as a cue, like the nine ecological cues in Table 8.1. That is,
in the competition, the number of cues (n) is thus equal to 10 (because rec-
ognition is included). The decision criterion for tallying is the following:

then choose city a.

then choose city b.

then guess.

The assignments of a, and bt are the following:

 REASONING THE FAST AND FRUGAL WAY 181

1 if the Mi cue value is positive

a,, b,• = 0 if the ith cue value is negative
0 if the ith cue value is unknown.
Let us compare cities a and b from Figure 8.1. By tallying the positive cue
values, a would score 2 points and b would score 3. Thus, tallying would
choose b to be the larger, in opposition to Take The Best, which would infer
that a is larger. Variants of tallying, such as the frequency-of-good-features
heuristic, have been discussed in the decision literature (Alba & Marmorstein,
1987; Payne, Bettman, & Johnson, 1993).

Contestant 2: Weighted Tallying Tallying treats all cues alike, independent of

cue validity. Weighted tallying of positive evidence is identical with tallying,
except that it weights each cue according to its ecological validity, v;. The
ecological validities of the cues appear in Table 8.1. We set the validity of the
recognition cue to .8, which is the empirical average determined by the pilot
study. The decision rule is as follows:

then choose city a.

then choose city b.

then guess.

Note that weighted tallying needs more information than either tallying or Take
The Best, namely, quantitative information about ecological validities. In the
simulation, we provided the real ecological validities to give this strategy a
good chance.
Calling again on the comparison of objects a and b from Figure 8.1, let us
assume that the validities would be .8 for recognition and .9, .8, .7, .6, .51 for
Cues 1 through 5. Weighted tallying would thus assign 1.7 points to a and 2.3
points to b. Thus weighted tallying would also choose b to be the larger.
Both tallying strategies treat negative information and missing information
identically. That is, they consider only positive evidence. The following strat-
egies distinguish between negative and missing information and integrate both
positive and negative information.

Contestant 3: Unit-Weight Linear Model The unit-weight linear model is a spe-

cial case of the equal-weight linear model (Huber, 1989) and has been advo-
cated as a good approximation of weighted linear models (Dawes, 1979; Ein-
horn & Hogarth, 1975). The decision criterion for unit-weight integration is the
same as for tallying; only the assignment of a, and b, differs:
1 if the ith cue value is positive
a,, bjr = — 1 if the Mi cue value is negative
0 if the ith cue value is unknown.
182 BOUNDED RATIONALITY

Comparing objects a and b from Figure 8.1 would involve assigning 1.0
points to a and 1.0 points to b and, thus, choosing randomly. This simple linear
model corresponds to Model 2 in Einhorn and Hogarth (1975, p. 177) with the
weight parameter set equal to 1.

Contestant 4: Weighted Linear Model This model is like the unit-weight linear
model except that the values of a, and fo, are multiplied by their respective
ecological validities. The decision criterion is the same as with weighted tal-
lying. The weighted linear model (or some variant of it) is often viewed as an
optimal rule for preferential choice, under the idealization of independent di-
mensions or cues (e.g., Keeney & Raiffa, 1993; Payne et al., 1993). Comparing
objects a and b from Figure 8.1 would involve assigning 1.0 points to a and
0.8 points to b and, thus, choosing a to be the larger.

Contestant 5: Multiple Regression The weighted linear model reflects the dif-
ferent validities of the cues but not the dependencies between cues. Multiple
regression creates weights that reflect the covariances between predictors or
cues and is commonly seen as an "optimal" linear way to integrate various
pieces of information into an estimate (e.g., Brunswik, 1955; Hammond, 1966).
Neural networks using the delta rule determine their "optimal" weights by the
same principles as multiple regression does (Stone, 1986). The delta rule car-
ries out the equivalent of a multiple linear regression from the input patterns
to the targets.
The weights for the multiple regression could simply be calculated from
the full information about the nine ecological cues. To make multiple regres-
sion an even stronger competitor, we also provided information about which
cities the simulated individuals recognized. Thus the multiple regression used
nine ecological cues and the recognition cue to generate its weights. Because
the weights for the recognition cue depend on which cities are recognized, we
calculated 6 X 500 X 84 sets of weights: one for each simulated individual.
Unlike any of the other strategies, regression had access to the actual city pop-
ulations (even for those cities not recognized by the hypothetical person) in
the calculation of the weights.4 During the quiz, each simulated person used
the set of weights provided to it by multiple regression to estimate the popu-
lations of the cities in the comparison.
There was a missing-values problem in computing these 6 X 84 X 500 sets
of regression coefficients, because most simulated individuals did not know
certain cue values, for instance, the cue values of the cities they did not rec-
ognize. We strengthened the performance of multiple regression by substituting

4. We cannot claim that these integration strategies are the best ones, nor can we
know a priori which small variations will succeed in our bumpy real-world environ-
ment. An example: After we had completed the simulations, we learned that regressing
on the ranks of the cities does slightly better than regressing on the city populations.
The key issue is what are the structures of environments in which particular strategies
and variants thrive.
 REASONING THE FAST AND FRUGAL WAY 183

unknown cue values with the average of the cue values the person knew for
the given cue.5 This was done both in creating the weights and in using these
weights to estimate populations. Unlike cross-validation procedures in which
weights are estimated from one half of the data and inferences based on these
weights are made for the other half, the regression strategy had access to all
the information (except, of course, the unknown cue values)—more informa-
tion than was given to any of the competitors. In the competition, multiple
regression and, to a lesser degree, the weighted linear model approximate the
ideal of the Laplacean demon.

Results
Frugality Take The Best is designed to enable quick decision making. Com-
pared with the integration strategies, how frugal is it, measured by the amount
of information searched in memory? For instance, in Figure 8.1, Take The Best
would look up four cue values (including the recognition cue values) to infer
that a is larger than b. None of the integration strategies use limited search;
thus they always look up all cue values.
Figure 8.5 shows the number of cue values retrieved from memory by Take
The Best for various levels of limited knowledge. Take The Best reduces search
in memory considerably. Depending on the knowledge state, this heuristic
needed to search for between 2 (the number of recognition values) and 20 (the
maximum possible cue values: Each city has nine cue values and one recog-
nition value). For instance, when a person recognized half of the cities and
knew 50% of their cue values, then, on average, only about 4 cue values (that
is, one fifth of all possible) were searched for. The average across all simulated
participants was 5.9, which was less than a third of all available cue values.

Accuracy Given that it searches only for a limited amount of information, how
accurate is Take The Best, compared with the integration strategies? We ran
the competition for all states of limited knowledge shown in Figure 8.4. We
first report the results of the competition in the case where each strategy
achieved its best performance: when 100% of the cue values were known.
Figure 8.6 shows the results of the simulations, carried out in the same way
as those in Figure 8.4.
To our surprise, Take The Best drew as many correct inferences as any of
the other strategies, and more than some. The curves for Take The Best, mul-
tiple regression, weighted tallying, and tallying are so similar that there are
only slight differences among them. Weighted tallying performed about as well
as tallying, and the unit-weight linear model performed about as well as the
weighted linear model—demonstrating that the previous finding that weights

5. If no single cue value was known for a given cue, the missing values were sub-
stituted by .5. This value was chosen because it is the midpoint of 0 and 1, which are
the values used to stand for negative and positive cue values, respectively.
184 BOUNDED RATIONALITY

Figure 8.5 Frugality: Number of cue values looked up by Take The Best and
by the competing integration strategies (see text), depending on the number
of objects recognized (0-83) and the percentage of cue values known.

may be chosen in a fairly arbitrary manner, as long as they have the correct
sign (Dawes, 1979), is generalizable to tallying. The two integration strategies
that make use of both positive and negative information, unit-weight and
weighted linear models, made considerably fewer correct inferences. By look-
ing at the lower-left and upper-right corners of Figure 8.6, one can see that all
competitors do equally well with a complete lack of knowledge or with com-
plete knowledge. They differ when knowledge is limited. Note that some strat-
egies can make more correct inferences when they do not have complete
knowledge: a demonstration of the less-is-more effect mentioned earlier.
What was the result of the competition across all levels of limited knowl-
edge? Table 8.2 shows the result for each level of limited knowledge of cue val-
ues, averaged across all levels of recognition knowledge. (Table 8.2 reports also
the performance of two variants of Take The Best, which we discuss later: the
Minimalist and Take The Last.) The values in the 100% column of Table 8.2 are
the values in Figure 8.6 averaged across all levels of recognition. Take The Best
made as many correct inferences as one of the competitors (weighted tallying)
and more than the others. Because it was also the most frugal, we judged the
competition goes to Take The Best as the highest performing, overall.
To our knowledge, this is the first time that it has been demonstrated that
a fast and frugal heuristic, that is, Take The Best, can draw as many correct
inferences about a real-world environment as integration strategies, across all
Figure 8.6 Results of the competition. The curve for Take The Best is identi-
cal with the 100% curve in Figure 8.4. The results for proportion correct
have been smoothed by a running median smoother, to lessen visual noise
between the lines.

Table 8.2 Results of the competition: Average percentage of correct inferences

Percentage of cue values known

Strategy 10 20 50 75 100 Average

Take the Best 62.1 63.5 66.3 67.8 69.1 65.8

Weighted tallying 62.1 63.5 66.3 67.9 69.3 65.8
Regression 62.5 63.5 65.7 67.4 69.4 65.7
Tallying 62.0 63.3 65.9 67.6 69.1 65.6
Weighted linear model 62.3 62.7 62.3 61.9 62.5 62.3
Unit-weight linear model 62.1 62.2 62.1 62.0 62.2 62.1

Minimalist 61.9 63.1 65.0 66.1 67.4 64.7

Take The Last 61.9 63.0 64.6 65.8 67.5 64.5

Note: Values are rounded; averages are computed from the unrounded values. Bottom two heuristics
are variants of Take The Best.
186 BOUNDED RATIONALITY

states of limited knowledge. The dictates of classical rationality would have

led one to expect the integration strategies to do substantially better than the
fast and frugal heuristic.
Two results of the simulation can be derived analytically. First and most
obvious is that if knowledge about objects is zero, then all strategies perform
at a chance level. Second, and less obvious, is that if all objects and cue values
are known, then tallying produces as many correct inferences as the unit-
weight linear model. This is because, under complete knowledge, the score
under the tallying strategy is an increasing linear function of the score arrived
at in the unit-weight linear model.6 The equivalence between tallying and unit-
weight linear models under complete knowledge is an important result. It is
known that unit-weight linear models can sometimes perform about as well as
proper linear models (i.e., models with weights that are chosen in an optimal
way, such as in multiple regression; see Dawes, 1979). The equivalence implies
that under complete knowledge, merely counting pieces of positive evidence
can work as well as proper linear models. This result clarifies one condition
under which searching only for positive evidence, a strategy that has some-
times been labeled confirmation bias or positive test strategy, can be a reason-
able and efficient inferential strategy (Klayman & Ha, 1987; Tweney & Walker,
1990).
Why do the unit-weight and weighted linear models perform markedly
worse under limited knowledge of objects? The reason is the simple and bold
recognition heuristic. Strategies that do not exploit the recognition heuristic
in environments where recognition is strongly correlated with the target vari-
able pay the price of a considerable number of wrong inferences. The unit-
weight and weighted linear models use recognition information and integrate
it with all other information but do not follow the recognition heuristic, that
is, they sometimes choose unrecognized cities over recognized ones. Why is
this? In the environment, there are more negative cue values than positive
ones, and most cities have more negative cue values than positive ones. From
this it follows that when a recognized object is compared with an unrecognized
object, the (weighted) sum of cue values of the recognized object will often be
smaller than that of the unrecognized object (which is —1 for the unit-weigh
model and —.8 for the weighted linear model). Here the unit-weight and
weighted linear models often make the inference that the unrecognized object
is the larger one, due to the overwhelming negative evidence for the recognized
object. Such inferences contradict the recognition heuristic. Tallying, in con-
trast, has the recognition heuristic built in implicitly. Because tallying ignores
negative information, the tally for an unrecognized object is always 0 and,

6. The proof for this is as follows. The tallying score t for a given object is the number
n+ of positive cue values, as defined above. The score u for the unit-weight linear model
is n+ - n~, where n~ is the number of negative cue values. Under complete knowledge,
n — n+ + n~, where n is the number of cues. Thus, t = n+, and u = n+ — n~. Because
n~ — n — n+, by substitution into the formula for u, we find that u — n+ — (n — n+) =
2t - n.
 REASONING THE FAST AND FRUGAL WAY 187

thus, is always smaller than the tally for a recognized object, which is at least
1 (for tallying, or .8 for weighted tallying, due to the positive value on the
recognition cue). Thus tallying always arrives at the inference that a recognized
object is larger than an unrecognized one.
Note that this explanation of the different performances puts the full weight
in a psychological principle (the recognition heuristic) explicit in Take The
Best, as opposed to the statistical issue of how to find optimal weights in a
linear function. To test this explanation, we reran the simulations for the unit-
weight and weighted linear models under the same conditions but replaced
the recognition cue with the recognition heuristic. The simulation showed that
the recognition heuristic accounts for all the difference.

Can Heuristics Get by with Even Less Time and Knowledge?

Take The Best produced a surprisingly high proportion of correct inferences,

compared with more computationally expensive integration strategies. Making
correct inferences despite limited knowledge is an important adaptive feature of
a heuristic, but being right is not the only thing that counts. In many situations,
time is limited, and acting fast can be as important as being correct. For instance,
if you are driving on an unfamiliar highway and you have to decide in an instant
what to do when the road forks, your problem is not necessarily making the best
choice, but simply making a quick choice. Pressure to be quick is also character-
istic for certain types of verbal interactions, such as press conferences, in which
a fast answer indicates competence, or commercial interactions, such as having
telephone service installed, where the customer has to decide in a few minutes
which of a dozen calling features to purchase. These situations entail the dual
constraints of limited knowledge and limited time. Take The Best is already
faster and more frugal than the integration strategies, because it performs only a
limited search and does not compute weighted sums of cue values. Can it be
made even faster? It can, if search is guided by the recency of cues in memory
rather than by cue validity.

The Take The Last Heuristic

Take The Last first tries the cue that discriminated the last time. If this cue
does not discriminate, the heuristic then tries the cue that discriminated the
time before last, and so on. The algorithm differs from Take The Best in Step
1, which is now reformulated as Step I'.

Step T. Search Rule If there is a record of which cues stopped search in pre-
vious problems, choose the cue that stopped search in the most recent problem
and that has not yet been tried. Look up the cue values of the two objects.
Otherwise try a random cue and build up such a record.
Thus, in Step 2, the algorithm goes back to Step 1'. Variants of this search
principle have been studied as the "Einstellung effect" in the water jar exper-
188 BOUNDED RATIONALITY

iments (Luchins & Luchins, 1994), in which the solution strategy of the most
recently solved problem is tried first on the subsequent problem. This effect
has also been noted in physicians' generation of diagnoses for clinical cases
(Weber, Bockenholt, Hilton, & Wallace, 1993).
This heuristic does not need a rank order of cues according to their valid-
ities; all that needs to be estimated is the direction in which a cue points. The
rank order of cue validities is replaced by a memory of which cues were last
used. Note that such a record can be built up independently of any knowledge
about the structure of an environment and neither needs, nor uses, any feed-
back about whether inferences are right or wrong.

The Minimalist Heuristic

Can reasonably accurate inferences be achieved with even less knowledge?

What we call the Minimalist heuristic needs neither information about the rank
ordering of cue validities nor the discrimination history of the cues. In its
ignorance, the heuristic picks cues in a random order. The algorithm differs
from Take The Best in Step 1, which is now reformulated as Step 1":

Step 1". Search Rule Draw a cue randomly (without replacement) and look up
the cue values of the two objects.
The Minimalist does not necessarily speed up search, but it tries to get by
with even less knowledge than any other strategy.

Results
Frugality How frugal are the heuristics? The simulations showed that for each
of the two variant heuristics, the relationship between amount of knowledge
and the number of cue values looked up had the same form as for Take The
Best (Figure 8.5). That is, unlike the integration strategies, the curves are con-
cave and the number of cues searched for is maximal when knowledge of cue
values is lowest. The average number of cue values looked up was lowest for
Take The Last (5.3) followed by the Minimalist (5.6) and Take The Best (5.9).
As knowledge becomes more and more limited (on both dimensions: recog-
nition and cue values known), the difference in frugality becomes smaller and
smaller. The reason why the Minimalist looks up fewer cue values than Take
The Best is that cue validities and cue discrimination rates are negatively cor-
related (Table 8.1); therefore, randomly chosen cues tend to have larger dis-
crimination rates than cues chosen by cue validity.

Accuracy What is the price to be paid for speeding up search or reducing the
knowledge of cue orderings and discrimination histories to nothing? We tested
the performance of the two heuristics in the same environment as all other
strategies. Figure 8.7 shows the proportion of correct inferences that the Min-
imalist achieved. For comparison, the performance of Take The Best with
100% of cue values known is indicated by a dotted line. Note that the Mini-
 REASONING THE FAST AND FRUGAL WAY 189

Figure 8.7 Performance of Minimalist. For comparison, the performance of

Take The Best (TTB) is shown as a dotted line, for the case in which 100%
of cue values are known.

malist performed surprisingly well. The maximum difference appeared when

knowledge was complete and all cities were recognized. In these circum-
stances, the Minimalist did about 4 percentage points worse than Take The
Best. On average, the proportion of correct inferences was only 1.1 percentage
points less than the best strategies in the competition (Table 8.2).
The performance of Take The Last is similar to Figure 8.7, and the average
number of correct inferences is shown in Table 8.2. Take The Last was more
frugal but scored slightly lower than the Minimalist. Take The Last has an
interesting ability, which fooled us in an earlier series of tests, in which we
used a systematic (as opposed to a random) method for presenting the test
pairs, starting with the largest city and pairing it with all others, and so on.
An integration strategy such as multiple regression cannot "find out" that it is
being tested in this systematic way, and its inferences are accordingly inde-
pendent of the sequence of presentation. However, Take The Last found out
and won this first round of the competition, outperforming the other compet-
itors by some 10 percentage points. How did it exploit systematic testing?
Recall that it tries, first, the cue that discriminated the last time. If this cue
does not discriminate, it proceeds with the cue that discriminated the time
before, and so on. In doing so, when testing is systematic in the way described,
it tends to find, for each city that is being paired with all smaller ones, the
group of cues for which the larger city has a positive value. Trying these cues
190 BOUNDED RATIONALITY

first increases the chances of finding a discriminating cue that points in the
right direction (toward the larger city). We learned our lesson and reran the
whole competition with randomly ordered pairs of cities.

Discussion

The competition showed a surprising result: The Take The Best heuristic drew
as many correct inferences about unknown features of a real-world environ-
ment as any of the integration strategies, and more than some of them. Two
further simplifications of the heuristic—Take The Last (replacing knowledge
about the rank orders of cue validities with a memory of the discrimination
history of cues) and Minimalist (dispensing with both)—showed a compara-
tively small loss in correct inferences, and only when knowledge about cue
values was high.
To the best of our knowledge, this is the first inference competition between
fast and frugal heuristics and "rational" strategies in a real-world environment.
The result is of importance for encouraging research that focuses on the power
of simple psychological mechanisms, that is, on the design and testing of mod-
els of bounded rationality. The result is also of importance as an existence
proof that cognitive strategies capable of successful performance in a real-
world environment do not need to satisfy the classical norms of rational in-
ference. The classical norms may be sufficient but are not necessary for good
inference in real environments.

Bounded Rationality
In this section, we discuss the fundamental psychological mechanism postu-
lated by the PMM family of heuristics: one-reason decision making. We discuss
how this mechanism exploits the structure of environments in making fast
inferences that differ from those arising from standard models of rational rea-
soning.

One-Reason Decision Making What we call one-reason decision making is a

specific form of bounded rationality. The inference, or decision, is based on a
single, good reason. There is no compensation between cues. One-reason de-
cision making is probably the most challenging feature of the PMM family of
heuristics. As we mentioned before, it is a design feature that is not present
in those models that depict human inference as an optimal integration of all
information available (implying that all information has been looked up in the
first place), including linear multiple regression and nonlinear neural net-
works. One-reason decision making means that each choice is based exclu-
sively on one reason (i.e., cue), but this reason may be different from decision
to decision. This allows for highly context-sensitive modeling of choice. One-
reason decision making is not compensatory. Compensation is, after all, the
 REASONING THE FAST AND FRUGAL WAY 191

cornerstone of classical rationality, assuming that all commodities can be com-

pared and everything has its price. Compensation assumes commensurability.
However, human minds do not trade everything; some things are supposed to
be without a price (Elster, 1979). For instance, if a person must choose between
two actions that might help him or her get out of deep financial trouble, and
one involves killing someone, then no amount of money or other benefits might
compensate for the prospect of bloody hands. He or she takes the action that
does not involve killing a person, whatever other differences exist between the
two options. More generally, hierarchies of ethical and moral values are often
noncompensatory: True friendship, military honors, and doctorates are sup-
posed to be without a price.
Noncompensatory strategies—such as lexicographic, conjunctive, and dis-
junctive rules—have been discussed in the literature, and some empirical ev-
idence has been reported (e.g., Einhorn, 1970; Fishburn, 1988). The closest
relative to the PMM family of heuristics is the lexicographic rule. Most evi-
dence for lexicographic processes seems to come from studies on decision un-
der risk (for a recent summary, see Lopes, 1995). However, despite empirical
evidence, noncompensatory lexicographic strategies have often been dis-
missed at face value because they violate the tenets of classical rationality
(Keeney & Raiffa, 1993; Lovie & Lovie, 1986). The PMM family is both more
general and more specific than the lexicographic rule. It is more general be-
cause only Take The Best uses a lexicographic procedure in which cues are
ordered according to their validity, whereas the variant heuristics do not. It is
more specific, because several other psychological principles are integrated
with the lexicographic rule in Take The Best, such as the recognition heuristic
and the rules for confidence judgment (which are not dealt with in this chap-
ter; see Chapter 7).
Serious models that comprise noncompensatory inferences are hard to find.
One of the few examples is in Breiman, Friedman, Olshen, and Stone (1993),
who reported a simple, noncompensatory algorithm with only 3 binary, or-
dered cues, which classified heart attack patients into high- and low-risk
groups and was more accurate than standard statistical classification methods
that used up to 19 variables. The practical relevance of this noncompensatory
classification algorithm is obvious: In the emergency room, the physician can
quickly obtain the measures on one, two, or three variables and does not need
to perform any computations because there is no integration. This group of
statisticians constructed decision trees that approach the task of classification
(and estimation) much as Take The Best handles two-alternative choice. Rel-
evance theory (Sperber, Cara, & Girotto, 1995) postulates that people generate
consequences from rules according to accessibility and stop this process when
expectations of relevance are met. Although relevance theory has not been as
formalized, we see its stopping rule as parallel to that of Take The Best. Finally,
optimality theory (Legendre, Raymond, & Smolensky, 1993; Prince & Smolen-
sky, 1991) proposes that hierarchical noncompensation explains how the gram-
mar of a language determines which structural description of an input best
192 BOUNDED RATIONALITY

satisfies well-formedness constraints. Optimality theory (which is actually a

satisficing theory) applies the same inferential principles as PMM theory to
phonology and morphology.

Recognition Heuristic The recognition heuristic is a version of ignorance-based

decision making that exploits a lack of knowledge. The very fact that one does
not know is used to make accurate inferences. The recognition heuristic is an
intuitively plausible principle that seems not to have been used until now in
models of bounded rationality. However, it has long been used to good advan-
tage by humans and other animals. For instance, advertisement techniques as
used by Benetton put all their effort into making sure that every customer
recognizes the brand name, with no effort made to inform about the product
itself. The idea behind this is that recognition is a strong force in customers'
choices. One of our dear (and well-read) colleagues, after seeing a draft of this
chapter, explained to us how he makes inferences about which books are worth
acquiring. If he finds a book about a great topic but does not recognize the
name of the author, he makes the inference that it is probably not worth buy-
ing. If, after an inspection of the references, he does not recognize most of the
names, he concludes the book is not even worth reading. The recognition heu-
ristic is also known as one of the rules that guide food preferences in animals.
For instance, rats choose the food that they recognize having eaten before (or
having smelled on the breath of fellow rats) and avoid novel foods (Gallistel,
Brown, Carey, Gelman, & Keil, 1991).
The empirical validity of the recognition heuristic for inferences about un-
known city populations, as used in the present simulations, can be directly
tested in several ways. First, participants are presented pairs of cities, among
them critical pairs in which one city is recognized and the other unrecognized,
and their task is to infer which one has more inhabitants. The recognition
heuristic predicts the recognized city. In our empirical tests, participants fol-
lowed the recognition heuristic in roughly 90% to 100% of all cases (Goldstein
& Gigerenzer, 1999). Second, participants are taught a cue, its ecological valid-
ity, and the cue values for some of the objects (such as whether a city has a
soccer team or not). Subsequently, they are tested on critical pairs of cities,
one recognized and one unrecognized, where the recognized city has a nega-
tive cue value (which indicates lower population). The second test is a harder
test for the recognition heuristic than the first one and can be made even harder
by using more cues with negative cue values for the recognized object, and by
other means. Tests of the second kind have been performed, and participants
still followed the recognition heuristic more than 90% of the time, providing
evidence for its empirical validity (Goldstein & Gigerenzer, 1999).
The recognition heuristic is a useful strategy in domains where recognition
is a predictor of a target variable, such as whether a food contains a toxic
substance. In cases where recognition does not predict the target, the PMM
algorithms can still perform the inference, but without the recognition heuris-
tic (i.e., Step 0 is canceled).
 REASONING THE FAST AND FRUGAL WAY 193

Limited Search Both one-reason decision making and the recognition heuristic
realize limited search by defining stopping rules. Integration strategies, in con-
trast, do not provide any model of stopping rules and implicitly assume ex-
haustive search (although they may provide rules for tossing out some of the
variables in a lengthy regression equation). Stopping rules are crucial for mod-
eling inference under limited time, as in Simon's examples of satisficing, where
search among alternatives terminates when a certain aspiration level is met.

Nonlinearity Linearity is a mathematically convenient tool that has domi-

nated the theory of rational choice since its inception in the mid-seventeenth
century. The assumption is that the various components of an alternative add
up independently to its overall estimate or utility. In contrast, nonlinear infer-
ence does not operate by computing linear sums of (weighted) cue values.
Nonlinear inference has many varieties, including simple principles such as
in the conjunctive and disjunctive algorithms (Einhorn, 1970) and highly com-
plex ones such as in nonlinear multiple regression and neural networks. Take
The Best and its variants belong to the family of simple nonlinear models. One
advantage of simple nonlinear models is transparency; every step in the heu-
ristic can be followed through, unlike fully connected neural networks with
numerous hidden units and other free parameters.
Our competition revealed that the unit-weight and weighted versions of the
linear models lead to about equal performance, consistent with the finding that
the choice of weights, provided the sign is correct, often does not matter much
(Dawes, 1979). In real-world domains, such as in the prediction of sudden infant
death from a linear combination of eight variables (Carpenter, Gardner, Mc-
Weeny, & Emery, 1977), the weights can be varied across a broad range without
decreasing predictive accuracy: a phenomenon known as the "flat maximum ef-
fect" (Lovie & Lovie, 1986; von Winterfeldt & Edwards, 1982). The competition
in addition showed that the flat maximum effect extends to tallying, with unit-
weight and weighted tallying performing about equally well. The performance
of Take The Best showed that the flat maximum can extend beyond linear mod-
els: Inferences based solely on the best cue can be as accurate as any weighted or
unit-weight linear combination of all cues.
Most research in psychology and economics has preferred linear models for
description, prediction, and prescription (Edwards, 1954, 1962; Lopes, 1994;
von Winterfeldt & Edwards, 1982). Historically, linear models such as analysis
of variance and multiple regression originated as tools for data analysis in
psychological laboratories and were subsequently projected by means of the
"tools-to-theories heuristic" into theories of mind (Chapter 1). From the suf-
ficiently good fit of linear models in many judgment studies, it has been in-
terpreted that humans in fact might combine cues in a linear fashion. However,
whether this can be taken to mean that humans actually use linear models is
controversial (Hammond & Summers, 1965; Hammond & Wascoe, 1980). For
instance, within a certain range, data generated from the (nonlinear) law of
falling bodies can be fitted well by a linear regression. For the data in this
194 BOUNDED RATIONALITY

study, a multiple linear regression resulted in R2 = .87, which means that a

linear combination of the cues can predict the target variable quite well. But
the simpler, nonlinear Take The Best could match this performance. Thus good
fit of a linear model does not rule out simpler models of inference.
Shepard (1967) reviewed the empirical evidence for the claim that humans
integrate information by linear models. He distinguished between the percep-
tual transformation of raw sensory inputs into conceptual objects and prop-
erties and the subsequent inference based on conceptual knowledge. He con-
cluded that the perceptual analysis integrates the responses of the vast number
of receptive elements into concepts and properties by complex nonlinear rules
but once this is done, "there is little evidence that they can in turn be juggled
and recombined with anything like this facility" (Shepard, 1967, p. 263). Al-
though our minds can take account of a host of different factors, and although
we can remember and report doing so, "it is seldom more than one or two that
we consider at any one time" (Shepard, 1967, p. 267). In Shepard's view, there
is little evidence for integration, linear or otherwise, in what we term infer-
ences from memory—even without constraints of limited time and knowledge.
A further kind of evidence does not support linear integration as a model of
memory-based inference. People often have great difficulties in handling cor-
relations between cues (e.g., Armelius & Armelius, 1974), whereas integration
models such as multiple regression need to handle intercorrelations. To sum-
marize, for memory-based inference, there seems to be little empirical evi-
dence for the view of the mind as a Laplacean demon equipped with the com-
putational powers to perform multiple regressions. But this need not be taken
as bad news. The beauty of the nonlinear heuristics is that they can match the
demon's performance with less searching, less knowledge, and less computa-
tional might.

Intransitivity Transitivity is a cornerstone of classical rationality. It is one of

the few tenets that the Anglo-American school of Ramsey and Savage shares
with the competing Franco-European school of Allais (Fishburn, 1991). If we
prefer a to b and b to c, then we should also prefer a to c. The linear strategies
in our competition always produce transitive inferences (except for ties, where
the algorithmrandomly guessed), and city populations are, in fact, transitive.
The PMM family includes heuristics that do not violate transitivity (such as
Take The Best), and others that do (e.g., Minimalist). The Minimalist randomly
selects a cue on which to base the inference, therefore intransitivities can re-
sult. Table 8.2 shows that in spite of these intransitivities, overall performance
of the heuristic is only about 1 percentage point lower than that of the best
transitive strategy and a few percentage points better than some transitive strat-
egies.
An organism that used Take The Best with a stricter stopping rule (actually,
the original version found in Gigerenzer et al., 1991; see Chapter 7) could also
be forced into making intransitive inferences. The stricter stopping rule is that
search is only terminated when one positive and one negative cue value (but
not one positive and one unknown cue value) are encountered. Figure 8.8
 REASONING THE FAST AND FRUGAL WAY 195

Figure 8.8 Limited knowledge and a stricter stopping rule can produce intran-
sitive inferences.

illustrates a state of knowledge in which this stricter stopping rule gives the
result that a dominates b, b dominates c, and c dominates a.7
Biological systems, for instance, can exhibit systematic intransitivities
based on incommensurability between two systems on one dimension (Gilpin,
1975; Lewontin, 1968). Imagine three species: a, b, and c. Species a inhabits
both water and land; species b inhabits both water and air. Therefore, the two
only compete in water, where species a defeats species b. Species c inhabits
land and air, so it only competes with b in the air, where it is defeated by b.
Finally, when a and c meet, it is only on land, and here, c is in its element
and defeats a. A linear model that estimates some value for the combative
strength of each species independently of the species with which it is com-
peting would fail to capture this nontransitive cycle.

Inferences without Estimation Einhorn and Hogarth (1975) noted that in the
unit-weight model "there is essentially no estimation involved in its use"
(p. 177), except for the sign of the unit weight. A similar result holds for the
heuristics reported here. Take The Best does not need to estimate regression
weights; it only needs to estimate a rank ordering of cue validities. Take The
Last and the Minimalist involve essentially no estimation (except for the sign
of the cues). The fact that there is no estimation problem has an important
consequence: An organism can use as many cues as it has experienced, without
being concerned about whether the size of the sample experienced is suffi-
ciently large to generate reliable estimates of weights.

7. Note that missing knowledge is necessary for intransitivities to occur. If all cue
values are known, no intransitive inferences can possibly result. Take The Best with the
stricter stopping rule allows precise predictions about the occurrence of intransitivities
over the course of knowledge acquisition. For instance, imagine a person whose knowl-
edge is described by Figure 8.8, except that she does not know the value of Cue 2 for
object c. This person would make no intransitive judgments comparing objects a, b, and
c. If she were to learn that object c had a negative cue value for Cue 2, she would produce
an intransitive judgment. If she learned one piece more, namely, the value of Cue 1 for
object c, then she would no longer produce an intransitive judgment. The prediction is
that transitive judgments should turn into intransitive ones and back during learning.
Thus intransitivities do not simply depend on the amount of limited knowledge but also
on what knowledge is missing.
196 BOUNDED RATIONALITY

Cue Redundancy and Performance Einhorn and Hogarth (1975) suggested that
unit-weight models can be expected to perform approximately as well as
proper linear models when (a) R2 from the regression model is in the moderate
or low range (around .5 or smaller) and (b) predictors (cues) are correlated.
Are these two criteria necessary, sufficient, or both to explain the performance
of Take The Best? Take The Best and its variants certainly can exploit cue
redundancy: If cues are highly correlated, one cue can do the job.
We have already seen that in the present environment, R2 = .87, which is
in the high rather than the moderate or low range. As mentioned earlier, the
pairwise correlations between the nine ecological cues ranged between -.25
and .54, with an absolute average value of .19. Thus, despite a high R2 and
only moderate-to-small correlation between cues, the heuristics performed
quite successfully. Their excellent performance in the competition can be ex-
plained only partially by cue redundancy, because the cues were only mod-
erately correlated. High cue redundancy, thus, does seem sufficient but is not
necessary for the successful performance of the heuristics.
A New Perspective on the Lens Model Ecological theorists such as Brunswik
(1955) emphasized that the cognitive system is designed to find many path-
ways to the world, substituting missing cues with whatever cues happen to be
available. Brunswik labeled this ability vicarious functioning, in which he saw
the most fundamental principle of a science of perception and cognition. His
proposal to model this adaptive process by linear multiple regression has in-
spired a long tradition of neo-Brunswikian research (Brehmer, 1994; Ham-
mond, 1990), although the empirical evidence for mental multiple regression
is still controversial (e.g., Brehmer & Brehmer, 1988). However, vicarious func-
tioning need not be equated with linear regression. The PMM family of heu-
ristics provides an alternative, nonadditive model of vicarious functioning, in
which cue substitution operates without integration. This offers a new per-
spective on Brunswik's lens model. In a fast and frugal lens model, the first
discriminating cue that passes through inhibits any other rays passing through
and determines judgment (Gigerenzer & Kurz, in press). Noncompensatory vi-
carious functioning is consistent with some of Brunswik's original examples,
such as the substitution of behaviors in Hull's habit-family hierarchy, and the
alternative manifestation of symptoms according to the psychoanalytic writ-
ings of Frenkel-Brunswik (see Gigerenzer & Murray, 1987, chap. 3).
It has been reported sometimes that teachers, physicians, and other profes-
sionals claim that they use seven or so criteria to make judgments (e.g., when
grading papers or making a differential diagnosis) but that experimental tests
showed that they in fact often used only one criterion (Shepard, 1967). At first
glance, this seems to indicate that those professionals make outrageous claims.
But it need not be. If experts' vicarious functioning works according to the
PMM heuristics, then they are correct in saying that they use many predictors,
but the decision is made by only one at any time.
What Counts as Good Reasoning? Much of the research on reasoning in the
last decades has assumed that sound reasoning can be reduced to principles
 REASONING THE FAST AND FRUGAL WAY 197

of internal consistency, such as additivity of probabilities, conformity to truth-

table logic, and transitivity. For instance, research on the Wason selection task,
the "Linda" problem, and the "cab" problem has evaluated reasoning almost
exclusively by some measure of internal consistency (Gigerenzer, 1995,1996a).
Cognitive strategies, however, need to meet more important constraints than
internal consistency: (a) They need to be psychologically plausible, (b) they
need to be fast, and (c) they need to make accurate inferences in real-world
environments. In real time and real environments, the possibility that a heu-
ristic (e.g., the Minimalist) can make intransitive inferences does not mean
that it will make them all the time or that this feature will significantly hurt
its accuracy. What we have not addressed in this chapter are constraints on
human reasoning that emerge from the fact that Homo sapiens is a social an-
imal. For instance, some choices (e.g., who to treat first in an emergency room)
need to be justified (Tetlock, 1992). Going with the single best reason, the
strategy of Take The Best, has an immediate appeal for justification and can
be more convincing and certainly easier to communicate than some compli-
cated weighting of cues.

Future Research Among the questions that need to be addressed in future

research are the following. First, how can we generalize the present heuristics
from two-alternative-choice tasks to other inferential tasks, such as classifica-
tion and estimation? The reported success of the classification and regression
tree models (Breiman et al., 1993), which use a form of one-reason decision
making, is an encouraging sign that what we have shown here for two-
altOernative-choice tasks might be generalizable. In fact, Berretty, Todd, and
Martignon (1999) recently proposed fast and frugal heuristics for classification,
as have Hertwig, Hoffrage, and Martignon (1999) for estimation. Second, what
is the structure of real-world environments that allows simple heuristics to
perform so well? We need to develop a conceptual language that can capture
important aspects of the structure of environments that simple cognitive heu-
ristics can exploit. For instance, Martignon and Hoffrage (1999) have identified
two structures that Take The Best can take advantage of: noncompensatory
information and scarce information. For more on these new results, see Gig-
erenzer, Todd, and the ABC Research Group (1999).

Can Reasoning Be Rational and Psychological?

At the beginning of this chapter, we pointed out the common opposition be-
tween the rational and the psychological, which emerged in the nineteenth
century after the breakdown of the classical interpretation of probability. Since
then, rational inference is commonly reduced to logic and probability theory,
and psychological explanations are called on when things go wrong. This di-
vision of labor is, in a nutshell, the basis on which much of the current re-
search on reasoning and decision making under uncertainty is built.
We believe that after 40 years of toying with the notion of bounded ration-
ality, it is time to overcome the opposition between the rational and the
198 BOUNDED RATIONALITY

psychological and to reunite the two. The PMM family of heuristics provides
precise computational models that attempt to do so. They differ from the En-
lightenment's unified view of the rational and psychological, in that they focus
on simple psychological mechanisms that operate under constraints of limited
time and knowledge and are supported by empirical evidence. The single most
important result in this chapter is that simple psychological mechanisms can
yield about as many (or more) correct inferences more quickly and with less
information than integration strategies that embody classical properties of ra-
tional inference. The demonstration that a fast and frugal heuristic won the
competition defeats the widespread view that only "rational" strategies can be
accurate. Models of inference do not have to forsake accuracy for simplicity.
The mind can have it both ways.
 IV
SOCIAL RATIONALITY

I he study of human thinking is deeply suspicious of introducing anything

genuinely social into the world of "pure" rationality. As in much of cogni-
tive science, most researchers have fallen in love with syllogisms, first-order
logic, probability theory, and other systems that abstract and distract from
the semantics and pragmatics of thinking, not to mention social motives.
In 1987, I spent the fall at the Department of Psychology at Harvard. One
day, I asked my colleagues whether they knew of a theory of thinking that
starts with a social motive rather than a logical structure. Sheldon White an-
swered by handing me a copy of a 1985 dissertation he had supervised. The
280-page work, entitled "Deduction or Darwinian algorithms?" and written
by Leda Cosmides, dealt with the Wason selection task—a classical stock-in-
trade of research on reasoning. It set forth the first theory applied to the se-
lection task that began with an adaptive problem, cheating detection in so-
cial exchange, rather than with material conditionals, truth tables, and the
like. I read the dissertation in one long sitting and designed a critical test of
the theory using perspective change. The result is reported in Chapter 10.
Social rationality is a specific form of ecological rationality, one in which
the environment consists of other humans (or, more generally, conspecifics).
The program of social rationality explains human judgment and decision
making in terms of the structure of social environments. Chapter 9 illustrates
how behaviors that look irrational from an individualistic point of view can
turn out to be well adapted to a specific social environment. Social environ-
ments foster different strategies than physical environments, such as imita-
tion instead of deliberation, and demand attention to information that is
unique to social interaction, such as cues that could reveal that one is being
cheated or bluffed.
In this section, I argue for the domain specificity of mental strategies and
the heuristic value of evolutionary theory in guessing what the relevant do-
mains are. Domain-specific processes allow organisms to solve adaptive
problems quickly with specialized tools, such as specific emotions. There are
200 SOCIAL RATIONALITY

several ways in which the mind might implement such a division of labor.
For instance, some have proposed mental modules for intuitive physics,
mathematics, and biology—a view that turns academic subjects into do-
mains. An evolutionary perspective suggests that a different division of labor
has evolved, one directed at solving important adaptive problems, such as
attachment development, mate search, parenting, social exchange, coalition
formation, and maintaining and upsetting dominance hierarchies. The mod-
ule dedicated to solving each of these problems needs to integrate motiva-
tion, perception, thinking, emotion, and behavior into a functional unit. This
is not to say that domain-specific modules are encapsulated and dissociated;
they are probably as coordinated as the violins, violas, cellos, oboes, and
French horns in an orchestra, or the liver, kidneys, lungs, and heart in a hu-
man body.
The idea of modules specialized for certain adaptive problems conflicts
with the compartmentalization of psychology. Today's areas of specialization
are defined in terms of faculties, such as memory, thinking, decision making,
intelligence, motivation, and emotion. These faculties have become institu-
tionalized in modern university curricula and grant agencies. They deter-
mine the professional self-perception of our colleagues, what they read and
what they ignore, their departmental alliances, and the hiring of professors.
If you ask a psychologist at a conference what she is doing, you will proba-
bly get an answer such as "I am a cognitive psychologist," "I do emotions,"
"I am a judgment and decision-making person," or "My field is motivation."
Evolutionary thinking is an antidote to this faculty view of the mind. Adap-
tive problems and their modern equivalents, such as foraging and dieting
and social exchange and markets, demand the orchestration of these facul-
ties, not their segregation.
In my opinion, the partitioning of psychological research into faculties is
one of the greatest barriers to progress. Research on modularity forces us to
reconsider the borders that have gone unquestioned for many decades. Re-
thinking rationality means rethinking the organization of the fields that study
it. Most interesting problems do not respect today's disciplinary boundaries.
Nor should we.
 9
Rationality
Why Social Context Matters

I want to argue against an old and beautiful dream. It was Leibniz's dream,
but not his alone. Leibniz (1677/1951) hoped to reduce rational reasoning to
a universal calculus, which he termed the Universal Characteristic. The plan
was simple: to establish characteristic numbers for all ideas, which would
reduce every question to calculation. Such a rational calculus would put an
end to scholarly bickering; if a dispute arose, the contending parties could
settle it quickly and peacefully by sitting down and calculating. For some time,
the Enlightenment probabilists believed that the mathematical theory of prob-
ability had made this dream a reality. Probability theory rather than logic be-
came the flip side of the newly coined rationality of the Enlightenment, which
acknowledged that humankind lives in the twilight of probability rather than
the noontime sun of certainty, as John Locke expressed it. Leibniz guessed
optimistically of the Universal Characteristic that "a few selected persons
might be able to do the whole thing in five years" (Leibniz, 1677/1951, p. 22).
By around 1840, however, mathematicians had given up as thankless and even
antimathematical the task of reducing rationality to a calculus (Daston, 1988).
Psychologists and economists have not.
Contemporary theories embody Leibniz's dream in various forms. Piaget
and Inhelder's (1951/1975) theory of cognitive development holds that, by
roughly age 12, human beings begin to reason according to the laws of prob-
ability theory; Piaget and Inhelder thus echo the Enlightenment conviction that
human rationality and probability theory are two sides of the same coin. Neo-
classical economic theories center on the assumption that Jacob Bernoulli's
expected utility maximization principle or its modern variants, such as sub-
jective expected utility, define rationality in all contexts. Similarly, neo-
Bayesians tend to claim that the formal machinery of Bayesian statistics defines
rational inferences in all contexts. In cognitive psychology, formal axioms and
rules—consistency, transitivity, and Bayes's rule, for example, as well as entire
statistical techniques—figure prominently in recent theories of mind and war-
rant the rationality of cognition (Chapter 1).

201
202 SOCIAL RATIONALITY

All these theories have been criticized as descriptively incomplete or in-

adequate, most often by showing that principles from logic or probability the-
ory (such as consistency) are systematically violated in certain contexts. Piaget
himself wondered why adults outside of Geneva seemed not to reach the level
of formal operations. But even critics have generally retained the beautifully
simple principles drawn from logic and probability theory as normative, albeit
not descriptively valid—that is, as definitions of how we should reason. In this
chapter, I will address the question of whether these principles are indeed
normative: sufficient for defining rational behavior.
My discussion will challenge one central assumption in the modern vari-
ants of Leibniz's dream: that formal axioms and rules of choice can define
rational behavior without referring to factors external to choice behavior. To
the contrary, I will argue that these principles are incomplete as behavioral
norms in the sense that their normative validity depends on the social context
of the behavior, such as social objectives, values, and motivations.
The point I wish to defend is that formal axioms and rules cannot be im-
posed as universal yardsticks of rationality independent of social objectives,
norms, and values; they can, however, be entailedby certain social objectives,
norms, and values. Thus I am not arguing against axioms and rules, only
against their a priori imposition as context-independent yardsticks of ration-
ality.
Leibniz's dream was of a formal calculus of reasonableness that could be
applied to everything. Modern variants tend to go one step further and assume
that the calculus of rationality has already been found and can be imposed in
all contexts. I will focus only on the social context in this chapter, arguing that
the idea of imposing a context-independent, general-purpose rationality is a
limited and confused one. The several examples that follow seek to demon-
strate that only by referring to something external to the rules or axioms, such
as social objectives, values, and norms, can we decide whether an axiom or
choice rule entails rational behavior.

Consistency: Property Alpha

Internal consistency of choice figures prominently as a basic requirement for

human rationality in decision theory, behavioral economics, game theory, and
cognitive theories. It is often seen as the requirement of rational choice. One
basic condition of the internal consistency of choice is known as "Property
Alpha," also called the "Chernoff condition" or "independence of irrelevant
alternatives" (Sen, 1993). The symbols S and T denote two (nonempty) sets of
alternatives, and x (S] denotes that alternative x is chosen from the set S:
Property Alpha:

Property Alpha demands that if x is chosen from S, and x belongs to a subset

T of S, then x must be chosen from Tas well. The following two choices would
be inconsistent in the sense that they violate Property Alpha:
 RATIONALITY 203

1. x is chosen given the options {x, y\.

2. y is chosen given the options {x, y, z}.
Property Alpha is violated here because x is chosen when the two alternatives
{x, 7} are offered, but y is chosen when z is added to the menu. (Choosing x
is interpreted here as a rejection of y, not as a choice that results from mere
indifference.) It may indeed appear odd and irrational that someone who
chooses x and rejects y when offered the choice set {x, y} would choose y and
reject x when offered the set {x, y, z}.
Property Alpha formulates consistency exclusively in terms of the internal
consistency of choice behavior with respect to sets of alternatives. No reference
is made to anything external to choice—for instance, intentional states such
as people's social objectives, values, and motivations. This exclusion of every-
thing psychological beyond behavior is in line with Samuelson's (1938) pro-
gram of freeing theories of behavior from any traces of utility and from the
priority of the notion of "preference." As Little (1949) commented on the un-
derlying methodological program, Samuelson's "revealed preference" formu-
lation "is scientifically more respectable [since] if an individual's behavior is
consistent, then it must be possible to explain the behavior without reference
to anything other than behavior" (p. 90). Sen (1993) has launched a forceful
attack on internal consistency, as defined by Property Alpha and similar prin-
ciples, and what follows is based on his ideas and examples.
The last apple. At a dinner party, a fruit basket is passed around. When
it reaches Mr. Polite, one apple is left in the basket. Dining alone, Mr.
Polite would face no dilemma; he would take it. In company, however,
he must choose between the apple (7) or nothing (x). He decides to be-
have decently and go without (x). If the basket had contained another
apple (z), he could reasonably have chosen y over x without violating
standards of good behavior. Choosing x over y from the choice set {x, 7}
and choosing 7 over x from the choice set {x, y, z} violates Property
Alpha, even though there is nothing irrational about Mr. Polite's behavior
given his scruples in social interaction. If he had not held to such values
of politeness, then Property Alpha might have been entailed. But it can-
not be imposed independent of his values.
Waiting for dinner. Consider a second example. Mr. Pleasant is invited
to his colleague's home on Sunday at 9:00 P.M. On arriving, he takes a
seat in the living room, and his host offers him crackers and nuts (7).
Mr. Pleasant decides to take nothing (x), because he is hungry for a sub-
stantial meal and does not want to fill up before dinner. After a while,
the colleague's wife comes in with tea and cake (z). The menu has
thereby been extended to {x, y, z}, but there is a larger implication: The
new option z also has destroyed an illusion about what the invitation
included. Now Mr. Pleasant chooses the crackers and nuts (7) over noth-
ing (x). Again, this preference reversal violates Property Alpha. Given
the guest's expectations of what the invitation entailed, however, there
is nothing irrational about his behavior.
Tea at night. Here is a final example similar to the last. Mr. Sociable has
met a young artist at a party. When the party is over, she invites him to
204 SOCIAL RATIONALITY

come back to her place for tea. He chooses to have tea with her (x) over
returning home (7). The young lady then offers him a third choice—to
share some cocaine at her apartment (z). This extension of the choice set
may quite reasonably affect Mr. Sociable's ranking of x and y. Depending
on his objectives and values, he may consequently choose to go home
(7)-
All three examples seek to illustrate the same,point: Property Alpha will or
will not be entailed depending on the social objectives, values, and expecta-
tions of the individual making the choice. To impose Property Alpha as a
general yardstick of rational behavior independent of social objectives or other
factors external to choice behavior seems fundamentally flawed.
The conclusion is not that consistency is an invalid principle; rather, con-
sistency, as defined by Property Alpha or similar principles, is indeterminate.
The preceding examples illustrate different kinds of indeterminateness. With
respect to the last apple, social values define what the alternatives in the choice
set are and, thereby, what consistency is about. If there are many apples in the
basket, the choice is between "apple" and "nothing." If a single apple remains
and one does not share the values of Mr. Polite, the alternatives are the same;
for Mr. Polite, however, they become "last apple" and "nothing." In the dinner
and tea examples, one learns something new about the old alternatives when
a new choice is introduced. The fresh option provides new information—that
is, it reduces uncertainty about the old alternatives.
To summarize the argument: Consistency, as defined by Property Alpha,
cannot be imposed on human behavior independent of something external to
choice behavior, such as social objectives and expectations. Social concerns
and moral views (e.g., politeness), as well as inferences from the menu offered
(learning from one option as to what others may involve), determine whether
internal consistency is or is not entailed.

Maximizing: Choice under Uncertainty

The birth year of the mathematical theory of probability is usually dated at

1654, when the French mathematicians Blaise Pascal and Pierre Fermat ex-
changed letters about gambling problems. One of these problems, the double-
six problem, was posed by the Chevalier de Mere, a notorious gambler. From
his rich experience, de Mere knew it was advantageous to bet on the occur-
rence of at least one six in a series of four dice tosses. Therefore, he reasoned,
it must be as advantageous to bet on the occurrence of at least one double-six
in 24 tosses of a pair of dice. Fortune, however, disappointed him. Was he
unlucky but right, or unlucky because he was wrong?
The mathematical theory of probability that emerged from the exchange
between Pascal and Fermat showed how to compute the expectation of an
event. Because the probability of not getting a six with a fair die is 5/6, the
probability of not getting a six in four throws is (5/6)4, which is .482. Thus, de
Mere was right that it is advantageous to bet on the occurrence of at least one
 RATIONALITY 205

six in four tosses, which is .518. In the same way, the answer to his puzzle
can be calculated. The probability of not getting a double six in one toss of a
pair of dice is 35/36, therefore the probability of not getting at least one double
six in 24 tosses is (35/S6) 24 , which is .509. Thus de Mere was unlucky because
his reasoning was wrong.
In general terms, maximizing expected utility means maximizing the prod-
uct between probabilities and utilities. In the simple case of a choice between
two options, x and y (e.g., no six and at least one six), with probabilities p(x)
and p(y], and utilities U(x] and U(y), one can maximize the gain according to
the rule:
Maximizing expected utility:
Choose x if p(x]U(x) > p(y)U(y).
For de Mere, the utilities were equal, because he and his gambling partner
bet the same amount of money on x and y, respectively. All this seems to be
straightforward once the mathematical theory is in place.
Now consider the following situation, which seems to be, formally, essen-
tially equivalent. The choice set is again {x, y}. Choosing x will lead to a re-
inforcement with a probability p(x) — .80, whereas choosing y will only lead
to the same reinforcement with a probability p(y] = .20. That is, the utilities
of the outcomes (reinforcements) are the same, but their probabilities differ. It
is easy to see that when the choice is repeated n times, the expected number
of reinforcements will be maximized if an organism always chooses x:
Maximizing with equal utilities:
Always choose x if p(x) > p(y).
Consider a hungry rat in a T-maze where reinforcement is obtained at the left
end in 80% of cases and at the right end in 20% of cases. The rat will maximize
reinforcement if it always turns left. Imagine students who watch the rat run-
ning and predict on which side the reinforcement will appear in each trial.
They will also maximize their number of correct predictions by always saying
"left." But neither rats nor students seem to maximize. Under a variety of
experimental conditions, organisms choose both alternatives with relative fre-
quencies that roughly match the probabilities (Gallistel, 1990):
Probability matching:
Choose x with probability p(x);
choose y with probability p(y).
In the preceding example, the expected rate of reinforcements is 80% for max-
imizing, but only 68% for probability matching (this value is calculated by
.802 + .202 = .68). The conditions of the seemingly irrational behavior of prob-
ability matching are discussed in the literature (e.g., Brunswik, 1939; Estes,
1976; Gallistel, 1990).
Violations of maximizing by probability matching pose a problem for a
context-independent account of rational behavior in animals and humans.
What looks irrational for an individual, however, can be optimal for a group.
206 SOCIAL RATIONALITY

Again, the maximizing principle does not capture the distinction between the
individual in social isolation and in social interaction. Under natural condi-
tions of foraging, there will not be just one rat but many who compete to
exploit food resources. If all choose to forage in the spot where previous ex-
perience suggests food is to be found in greatest abundance, then each may
get only a small share. The one mutant organism that sometimes chooses the
spot with less food would be better off. Natural selection will favor those ex-
ceptional individuals who sometimes choose the less attractive alternative.
Thus maximizing is not always an evolutionarily stable strategy in situations
of competition among individuals. Given certain assumptions, probability
matching may in fact be an evolutionarily stable strategy, one that does not
tend to create conditions that select against it (Fretwell, 1972; Gallistel, 1990).
To summarize the argument: The maximization rule cannot be imposed on
behavior independent of social context. Whether an organism performs in iso-
lation or in the context of other organisms can determine, among other things,
whether maximization is entailed as an optimal choice rule.

Betting against the Probabilities

Mr. Smart would like to invest the $10,000 in his savings account in the hope
of increasing his capital. After some consideration, he opts to risk the amount
in a gamble with two possible outcomes, x and y. The outcomes are determined
by a fair roulette wheel with 10 equal sections, 6 of them white (x) and 4 black
(y). Thus the probability p(x) of obtaining white is .6, and the probability p(y)
of obtaining black is .4. The rules of the game are that he has to bet all his
money ($10,000) either on black or on white. If Mr. Smart guesses the outcome
correctly, his money will be doubled; otherwise, he will lose three-quarters of
his investment. Could it ever be advantageous for Mr. Smart to bet on black?
If Mr. Smart bets on white, his expectation is $20,000 with a probability of
.6, and $2,500 with a probability of .4. The expected value E(x) is (.6 X
$20,000) + (.4 X $2,500) = $13,000. But if he bets on black, the expected value
E(y) is only (.4 X $20,000) + (.6 X $2,500) = $9,500. Betting on white would
give him an expectation larger than the sum he invests. Betting on black, on
the other hand, would result in an expectation lower than the sum he invests.
A maximization of the expected value implies betting on white:

Maximizing expected value:

Choose x if E(x) > E(y),

where E(x) = £ p(x)V(x). The principle of maximizing the expected value (or
subjective variants such as expected utility) is one of the cornerstones of clas-
sical definitions of rationality. Mr. Smart would be a fool to bet on black,
wouldn't he?
 RATIONALITY 207

Let me apply the same argument again. The principle of maximizing the
expected value does not distinguish between the individual in social isolation
and in social interaction. If many individuals face the same choice, could it
be to the benefit of the whole group that some sacrifice themselves and bet on
black? Let us first look at an example from biology.
Cooper (1989; Cooper & Kaplan, 1982) discussed conditions under which
it is essential for the survival of the group that some individuals bet against
the probabilities and do not, at the individual level, maximize their expected
value. Consider a hypothetical population of organisms whose evolutionary
fitness (measured simply by the finite rate of increase in their population)
depends highly on protective coloration. Each winter predators pass through
the region, decimating those within the population that can be spotted against
the background terrain. If the black soil of the organisms' habitat happens to
be covered with snow at the time, the best protective coloration is white; oth-
erwise, it is black. The probability of snow when predators pass through is .6,
and protectively colored individuals can expect to survive the winter in num-
bers sufficient to leave an average of two surviving offspring each, whereas the
conspicuous ones can expect an average of only 0.25 offspring each. This ex-
ample assumes a simple evolutionary model with asexual breeding (each off-
spring is genetically identical to its parent), seasonal breeding (offspring are
produced only in spring), and semelparous breeding (each individual produces
offspring only once in a lifetime at the age of exactly one year).

Adaptive Coin-Flipping
Suppose two genotypes, W and WB, are in competition within a large popu-
lation. Individuals of genotype W always have white winter coloration; that is,
W is a genotype with a uniquely determined phenotypic expression. Genotype
WB, in contrast, gives rise to both white and black individuals, with a ratio of
5 to 3. Thus, 3 out of 8 individuals with genotype WB are "betting" on the
low probability of no snow. Each of these individuals' expectation to survive
and reproduce is smaller than that of all other individuals in both W and WB.
How will these two genotypes fare after 1,000 generations (1,000 years)? We
can expect that there was snow cover in about 600 winters, exposed black soil
in about 400 winters. Then the number of individuals with genotype W will
be doubled 600 times and reduced to one-fourth 400 times. If n is the original
population size, the population size after 1,000 years is

That is, genotype W will have been wiped out with practical certainty after
1,000 years. How does genotype WB do? In the 600 snowy winters, 5/s of the
population will double in number and three-eighths will be reduced to 25%,
with corresponding proportions for the 400 winters without snow. The number
of individuals after 1,000 years is then
208 SOCIAL RATIONALITY

Thus genotype WB is likely to win the evolutionary race easily.1 (The large
estimated number is certainly an overestimation, however, because it does not
take account of such other constraints as food resources.) The reason why WB
has so much better a chance of survival than W is that a considerable propor-
tion of the WB individuals do not maximize their individual expectations but
"bet" on small probabilities.
This violation of individual maximization has been termed "adaptive coin-
flipping" (Cooper & Kaplan, 1982), meaning that individuals are genetically
programmed to "flip coins" to adopt phenotypic traits. Thus the phenotype is
ultimately determined by the nature of the coin-flipping process, rather than
uniquely specified by the genotype.2
Back to Mr. Smart. Assume he won and wants to try again. So do his nu-
merous brothers, sisters, and cousins, who all are willing to commit their entire
investment capital to this gamble. The game is offered every week, and the
rules are as before: Each person's choice is every week to bet all his or her
investment capital either on black or on white (no hedging of bets). If everyone
wanted solely to maximize his or her individual good, his or her money would
be better invested in white than in black, because the chances to double one's
assets are 60% for white compared with only 40% for black. Investing in black
would appear irrational. But we know from our previous calculations that
someone who invests all his or her money every week in white will, with a
high probability, lose every dollar of assets in the long run.
If Mr. Smart and his extended family, however, acted as one community
rather than as independent individuals—that is, create one investment capital
fund in which they share equally—they can quickly increase their capital with
a high probability. Every week they would need to instruct three-eighths of
their members to invest in black, the rest in white. This social sharing is es-
sentially the same situation as the "adaptive coin-flipping" example. Thus Mr.
Smart's betting on black needs to be judged against his motivation: If he is

1. I have reported the numbers only for the most likely event (i.e., 600 snowy winters
out of 1,000 winters). If one looks at all possible events, one finds that those in which
W would result in a larger population size than WB are extremely rare (Cooper, 1989).
Nevertheless, the expected value is larger for W than for WB, due to the fact that in
those very few cases in which W results in a larger population size, this number is
astronomically large. The reader who is familiar with the St. Petersburg paradox will
see a parallel (Wolfgang Hell has drawn my attention to this fact). The parallel is best
illustrated in Lopes's (1981) simulations of businesses selling the St. Petersburg gamble.
Although these businesses sold the gamble far below its expected value, most nonethe-
less survived with great profits.
2. Adaptive coin-flipping is a special case of a general phenomenon: In variable en-
vironments (in which the time scale of variation is greater than the generation time of
the organism, as in the example given), natural selection does not maximize expected
individual fitness but geometric mean fitness (Gillespie, 1977).
 RATIONALITY 209

cooperating with others for their common interest, then betting on the wrong
side of a known probability is part of an optimal strategy. If he is not cooper-
ating but, rather, investing for his own immediate benefit, then betting on black
is the fastest way to ruin.
This example, like the preceding ones, attempts to illustrate that a rule such
as maximizing the expected value cannot be imposed on behavior without
consideration of the social context. Is this context a single individual wagering
all his or her assets at once or a population that risks their collective assets or
offspring at regular intervals? It makes all the difference, because individual
maximization can lead to the extinction of the genotype.

Conclusions

These examples show that general principles such as consistency and maximiz-
ing are insufficient for capturing rationality. I have argued that there is no way of
determining whether a behavioral pattern is consistent or maximizes without
first referring to something external to choice behavior (Sen, 1993). The external
factor investigated in this chapter is the social context of choice behavior, in-
cluding objectives, motivations, and values. I am not arguing against consis-
tency, maximization, or any given rule per se but against the a priori imposition
of a rule or axiom as a requirement for rationality, independent of the social con-
text of judgment and decision and, likewise, of whether the individual operates
in isolation or within a social context (see Chapter 12; Elster, 1990).
One way to defend general principles against this argument would be to
say that maximization poses no restrictions on what individuals maximize, be
it their own good (utilities) or the fitness of their genotype. Switching from
individual goals to genotypic fitness can save the concept of maximization.
Such a defense would imply, however, that maximization cannot be imposed
independent of the motivations and goals built into living systems, which is
precisely the point I have asserted. By the same token, to claim that consis-
tency poses no restrictions on whatever consistency is about would destroy
the very idea of behavioral consistency, because Property Alpha would as a
result be open to any external interpretation and would no longer impose any
constraint on choice.
More generally, the formal principles of logic, probability theory, rational
choice theory, and other context-independent principles of rationality are often
rescued and defended by post hoc justifications. Post hoc reasoning typically
uses the social objectives, values, and motivations of organisms to make room
for exceptions or to reinterpret the alternatives in axioms or rules until they
are compatible with the observed result. Contemporary neoclassical econom-
ics, for instance, provides little theoretical basis for specifying the content and
shape of the utility function; it thus affords many degrees of freedom for fitting
any phenomenon to the theory (Simon, 1986). In Elster's (1990) formulation,
a theory of rationality can fail through indeterminacy (rather than through
inadequacy) to the extent that it fails to yield unique predictions.
210 SOCIAL RATIONALITY

The challenge is to go beyond general-purpose principles of rationality that

allow context to slip in through the back door. What would a theory of rea-
soning that lets social context in through the front door look like? In the next
chapter, I present and discuss one example of such a "front door" theory.
Inspired by evolutionary theory, it uses the theory of reciprocal altruism to
define social objectives and motivations underlying human reasoning.
 10

Domain-Specific Reasoning
Social Contracts and Cheating Detection

What counts as human rationality? Reasoning processes that embody content-

independent formal theories, such as propositional logic, or reasoning pro-
cesses that are well designed for solving specific adaptive problems? What
would a theoretical framework look like that starts with an adaptive problem
rather than with logical structure? In this chapter, I first introduce the challenge
of domain specificity and then apply a domain-specific theory for a specific
adaptive problem, social exchange, to a reasoning task that has puzzled re-
searchers for several decades.

Are Mental Processes Domain Specific?

The psychological flip side of Leibniz's dream of a universal calculus of rea-

sonableness is the assumption that there is one—or at most a few—universal
mechanisms that govern all of reasoning, learning, memory, inference, imita-
tion, imagery, and so on. I will call these assumed mechanisms general-
purpose mechanisms because they have no features specialized for processing
particular kinds of content. For instance, when Piaget started to work on men-
tal imagery and memory, he did not expect and search for processes different
from logical thinking. Rather, he attempted to demonstrate that at each stage
in development, imagery and memory express the same logical structure as
the one he had found in his earlier studies on children's thinking (Gruber &
Voneche, 1977). Similarly, B. F. Skinner's laws of operant behavior were de-
signed to be general purpose: to hold true for all stimuli and responses (the
assumption of the equipotentiality of stimuli).
John Garcia's anomalous findings (e.g., Garcia & Koelling, 1966) challenged
not only the notion of the equipotentiality of stimuli but also the law of con-
tiguity, which postulates the necessity of immediate reinforcement, indepen-
dent of the nature of the stimulus and response. For instance, when the taste
of flavored water is repeatedly paired with an electric shock immediately after

211
212 SOCIAL RATIONALITY

tasting, rats have great difficulty learning to avoid the flavored water. Yet in
just one trial the rat can learn to avoid the flavored water when it is followed
by experimentally induced nausea, even when the nausea occurs 2 hours later:

From the evolutionary view, the rat is a biased learning machine de-
signed by natural selection to form certain CS—US [conditioned stimu-
lus-unconditioned stimulus] associations rapidly but not others. From a
traditional learning viewpoint, the rat was an unbiased learner able to
make any association in accordance with the general principles of con-
tiguity, effect, and similarity. (Garcia y Robertson & Garcia, 1985, p. 25)

Garcia's evolutionary challenge, however, was not welcomed by mainstream

neobehaviorists. In 1965, after 10 years of research, he openly pointed out the
clash between the data and the ideal of general-purpose mechanisms—and his
manuscripts suddenly began to be rejected by the editors of the APA (American
Psychological Association) journals. This pattern continued for the next 13
years until, in 1979, Garcia was awarded the APA's Distinguished Scientific
Contribution Award (Lubek & Apfelbaum, 1987). By then, stimulus equipoten-
tiality was driven out of behaviorism but had found new fertile ground in
cognitive psychology.
The view that psychological mechanisms such as those described in the
laws of operant behavior are designed for specific classes of stimuli rather than
being general purpose is known as domain specificity (e.g., Hirschfeld & Gel-
man, 1994a), biological preparedness (Seligman & Hager, 1972), or, in biology,
special-design theories (Williams, 1966).
Mainstream cognitive psychology, however, still tries to avoid domain spec-
ificity. The senses, language, and emotions have occasionally been accepted as
domain-specific adaptations (Fodor, 1983). But the "central" cognitive pro-
cesses that define the rationality of Homo sapiens—reasoning, inference, judg-
ment, and decision making—have not. Even such vigorous advocates of do-
main specificity as Fodor (1983) have held so-called central processes to be
general purpose. Research on probabilistic, inductive, and deductive reasoning
tends to define good reasoning exclusively in terms of formal axioms and rules
similar to those discussed in Chapter 9. Mental logic, Johnson-Laird's mental
models, and Piaget's formal operations all are examples of the hope that rea-
soning can be understood without reference to its content.
A glance at textbooks on cognitive psychology reveals how we have bottle-
fed our students on the idea that whenever thinking is the object of our in-
vestigation, content does not matter. Typically, a chapter on "deductive rea-
soning" teaches prepositional logic and violations thereof by human reasoning,
whereas a chapter on "probabilistic reasoning" teaches the laws of probability
theory and violations thereof by human reasoning. Similarly, "fallacies" of
reasoning are defined against formal structure—the base-rate fallacy, the con-
junction fallacy, and so on. Content is merely illustrative and cosmetic, as it
is in textbooks of logic. Whether a problem concerns white and black swans,
blue and green taxicabs, or artists and beekeepers does not seem to matter.
Content has not yet assumed a life of its own. For the most part, it is seen only
 DOMAIN-SPECIFIC REASONING 213

as a disturbing factor that sometimes facilitates and sometimes hinders formal,

rational reasoning.
Is there an alternative? In what follows, I shall describe a domain-specific
theory of cognition that relates reasoning to the evolutionary theory of recip-
rocal altruism (Cosmides & Tooby, 1992). This theory turns the traditional ap-
proach upside down. It does not start out with a general-purpose principle
from logic or probability theory or a variant thereof; it takes social objectives
as fundamental, which in turn makes content fundamental, because social ob-
jectives have specific contents. Traditional formal principles of rationality are
not imposed; they can be entailed or not, depending on the social objectives.

Cheating Detection in Social Contracts

One feature that sets humans and some other primates apart from almost all
animal species is the existence of cooperation among genetically unrelated
individuals within the same species, known as reciprocal altruism or coop-
eration. The thesis that such cooperation has been practiced by our ancestors
since ancient times, possibly for at least several million years, is supported by
evidence from several sources. First, our nearest relatives in the hominid line,
chimpanzees, also engage in certain forms of sophisticated cooperation (de
Waal & Luttrell, 1988), and in more distant relatives, such as macaques and
baboons, cooperation can still be found (e.g., Packer, 1977). Second, coopera-
tion is both universal and highly elaborated across human cultures, from
hunter-gatherers to technologically advanced societies. Finally, paleoanthro-
pological evidence also suggests that cooperation is extremely ancient (e.g.,
Tooby & DeVore, 1987).
Why altruism? Kin-related helping behavior, such as that by the sterile
worker castes in insects, which so troubled Darwin, has been accounted for by
generalizing "Darwinian fitness" to "inclusive fitness"—that is, to the number
of surviving offspring an individual has plus the individual's effect on the
number of offspring produced by its relatives (Hamilton, 1964). But why re-
ciprocal altruism, which involves cooperation among two or more nonrelated
individuals? The now-classic answer draws on the economic concept of trade
and its analogy to game theory (Axelrod, 1984; Williams, 1966). If the repro-
ductive benefit of being helped is greater than the cost of helping, then indi-
viduals who engage in reciprocal helping can outreproduce those who do not,
causing the helping design to spread. A vampire bat, for instance, will die if
it fails to find food for two consecutive nights, and there is high variance in
food-gathering success. Food sharing allows the bats to reduce this variance,
and the best predictor of whether a bat, having foraged successfully, will share
its food with a hungry nonrelative is whether the nonrelative has shared food
with the bat in the past (Wilkinson, 1990).
But "always cooperate" would not be an evolutionarily stable strategy. This
can be seen using the analogy of the prisoner's dilemma. If a group of individ-
uals always cooperates, then individuals who always defect—that is, who take
214 SOCIAL RATIONALITY

the benefit but do not reciprocate—can invade and outreproduce the cooper-
ators. Where the opportunity for defecting (or cheating) exists, indiscriminate
cooperation would eventually be selected out. "Always defect" would not be
an evolutionarily stable strategy, either. A group of individuals who always
defect can be invaded by individuals who cooperate in a selective (rather than
indiscriminate) way. A simple heuristic for selective cooperation is "Cooperate
on the first move; for subsequent moves, do whatever your partner did on the
preceding move" (a strategy known as Tit For Tat). There are several rules in
addition to Tit For Tat that lead to cooperation with other "selective cooper-
ators" and exclude or retaliate against cheaters (Axelrod, 1984).
The important point is that selective cooperation would not work without
a cognitive heuristic for detecting cheaters—or, more precisely, a heuristic for
directing an organism's attention to information that could reveal that it (or its
group) is being cheated (Cosmides & Tooby, 1992). Neither indiscriminate co-
operation nor indiscriminate cheating demands such a heuristic. In vampire
bats, who exchange only one thing—regurgitated blood—such a heuristic can
be restricted to a sole commodity. Cheating, or more generally noncooperation,
would mean, "That other bat took my blood when it had nothing, but it did
not share blood with me when I had nothing." In humans, who exchange many
goods (including such abstract forms as money), a cheating-detection heuristic
needs to work on a more general level of representation—in terms, for exam-
ple, of "benefits" and "costs."
To summarize, cooperation between two or more individuals for their mu-
tual benefit is a solution to a class of important adaptive problems, such as the
sharing of scarce food when foraging success is highly variable. Rather than
being indiscriminate, cooperation needs to be selective, requiring a cognitive
heuristic that directs attention to information that can reveal cheating. This
evolutionary account of cooperation, albeit still general, can be applied to a
specific puzzle in the psychology of reasoning.

The Selection Task

In 1966, Peter Wason invented the "selection task" to study reasoning about
conditionals. This was to become one of the most extensively researched sub-
jects in cognitive psychology during the following decades. The selection task
involves four cards and a conditional statement in the form "If P then Q." One
example is, "If there is a 'D' on one side of the card, then there is a '3' on the
other side." The four cards are placed on a table so that the participant can read
only the information on the side facing upward. For instance, the four cards
may read "D," "E," "3," and "4." The participant's task is to indicate which of
the four cards need(s) to be turned over to find out whether the statement has
been violated. Table 10.1 shows two examples of selection tasks, each with a
different content: a numbers-and-letters rule and a transportation rule.
Because the dominant approach has been to impose prepositional logic as
a general-purpose standard of rational reasoning in the selection task (inde-
 DOMAIN-SPECIFIC REASONING 215

Table 10.1 Two selection tasks

Numbers-and-letters rule
If there is a "D" on one side of the card, then there is a "3" on the other side.
Each of the following cards has a letter on one side and a number on the other. In-
dicate only the card(s) you definitely need to turn over to see if the rule has been
violated.
D E 3 4

Transportation rule
If a person goes in to Boston, then he takes the subway.
The cards below have information about four Cambridge residents. Each card repre-
sents one person. One side of the card tells where the person went and the other
side tells how the person got there. Indicate only the card(s) you definitely need to
turn over to see if the rule has been violated.
SUBWAY ARLINGTON CAB BOSTON

pendent, of course, of the content of the conditional statements), it is crucial

to recall that, according to propositional logic, a conditional "If P then Q" can
be violated only by "P & not-Q." That is, the logical falsity of a material con-
ditional is defined within propositional logic in the following way:
Logical falsity:
"If P then Q" is logically false if and only if "P & not-Q."
Thus the "P" and "not-Q" cards, and no others, must be selected, because only
these can reveal "P & not-Q" instances. In the numbers-and-letters rule, these
cards correspond to the "D" and "4" cards and, in the transportation problem,
to the "Boston" and "cab" cards.
Wason's results showed, however, that human inferences did not generally
follow propositional logic. An avalanche of studies has since confirmed this,
reporting that, with numbers-and-letters rules, only about 10% of the partici-
pants select both the "P" and "not-Q" cards, whereas most select the "P" card
and the "Q" card or only the "P" card. It was soon found that the selections
were highly dependent on the content of the conditional statement. This was
labeled the "content effect." For instance, about 30% to 40% of participants
typically choose the "P" and "not-Q" cards in the transportation problem (Cos-
mides, 1989). These results seem to be inconsistent with Piaget's claim that
adults should have reached the stage of formal operations (Legrenzi & Murino,
1974; Wason, 1968; Wason & Johnson-Laird, 1970).
Within a decade it was clear that these results—a low overall proportion of
"P & not-Q" answers and the content effect—contradicted the model of human
reasoning provided by propositional logic. One might expect that propositional
logic was then abandoned; but it was abandoned only as a descriptive model
of reasoning. Propositional logic was, however, retained as the normative,
216 SOCIAL RATIONALITY

content-independent yardstick of good reasoning, and actual human reasoning

was blamed as irrational. The experimental manipulations were evaluated, as
is still the case today, in terms of whether or not they "facilitated logical rea-
soning." Much effort was directed at explaining participants' apparent irra-
tionality, including their "incorrigible conviction that they are right when they
are, in fact, wrong" (Wason, 1981, p. 356). It was proposed that the mind runs
with deficient mental software—for example, confirmation bias, matching bias,
and availability heuristic—rather than by propositional logic. Yet these pro-
posals were as general purpose as propositional logic; they could be applied
to any content. It seems fair to say that these vague proposals have not led to
an understanding of what participants do in the selection task.
Only since the mid-1980s have a few dissidents dared to design theories
that start with the content of the conditional statement rather than with prop-
ositional logic (Cheng & Holyoak, 1985; Cosmides, 1989; Cummins, 1998;
Light, Girotto, & Legrenzi, 1990; Over & Manktelow, 1993). I will concentrate
here exclusively on Cosmides's proposal, which takes the evolutionary theory
of cooperation as its starting point (for a different evolutionary account, see
Klix, 1993).
Cosmides's (1989) central point is that selective cooperation demands the
ability to detect cheaters. This ability presupposes several others, including
that of distinguishing different individuals, recognizing when a reciprocation
(social contract) is offered, and computing costs and benefits, all of which I
ignore here (Cosmides & Tooby, 1992). Being cheated in a social contract of
the type
If you take the benefit, then you have to pay the cost
means that the other party has exhibited the following behavior:
Benefit taken and cost not paid.
The evolutionary perspective suggests that humans, who belong to one of
the few species practicing reciprocal altruism since time immemorial, have
evolved a cognitive system for directing attention to information that could
reveal cheaters. That is, once a cognitive system has classified a situation as
one of cooperation, attention will be directed to information that could reveal
"benefit taken and cost not paid." Note that cheating detection in social con-
tracts is a domain-specific mechanism; it would not apply if a conditional
statement is coded as a threat, such as "If you touch me, then I'll kill you."
But how does this help us to understand the content effect in the selection
task?
The thesis is that the cheating-detection mechanism required by the theory
of reciprocal altruism guides reasoning in the selection task:
If the conditional statement is coded as a social contract, and the sub-
ject is cued in to the perspective of one party in the contract, then
attention is directed to information that can reveal being cheated.
 DOMAIN-SPECIFIC REASONING 217

In other words, a participant should select those cards that correspond to

"benefit taken" and "cost not paid," whatever the cards' logical status is. This
application of the theory of reciprocal altruism to an unresolved issue in hu-
man reasoning is, of course, a bold thesis.

Cheating Detection in the Selection Task

The numbers-and-letters rule and the transportation rule are not social con-
tracts. There are not two partners who have engaged in a contract; the rules
are descriptive ones rather than obligations, permissions, or other contracts
with mutual costs and benefits. Therefore, social contract theory is mute on
these problems. Cosmides (1989), however, showed that if a rule expressed a
social contract, then the percentage of "benefit taken" and the "costs not paid"
selections was very high. But this result can also be consistent with competing
accounts that do not invoke reciprocal altruism, so we need to look more
closely at tests that differentiate between competing accounts. Below is a sam-
ple of tests with that aim.

What Guides Reasoning: Availability or Cheater Detection?

The major account of the content effect in the 1970s and 1980s was variously
called "familiarity" and "availability" (Manktelow & Evans, 1979; Pollard,
1982), without ever being precisely defined. The underlying idea is that the
more familiar a statement is, the more often a participant may have experi-
enced associations between the two propositions in a conditional statement,
including those that are violations ("benefit taken" and "cost not paid") of the
conditional statement. In this view, familiarity makes violations more "avail-
able" in memory, and selections may simply reflect availability. According to
this conjecture, therefore, familiarity and not social contracts accounts for se-
lecting the "benefit taken" and "cost not paid" cards. If familiarity were indeed
the guiding cognitive principle, then unfamiliar social contracts should not
elicit the same results. However, Cosmides (1989), Gigerenzer and Hug (1992),
and Platt and Griggs (1993) showed that social contracts with unfamiliar prop-
ositions elicit the same high number of "benefit taken" and "cost not paid"
selections, in contradiction to the availability account.

Social Contracts without Cheater Detection

Are People Simply Good at Reasoning

about Social Contracts?

The game-theoretic models for the evolution of cooperation require, as argued

earlier, some mechanism for detecting cheaters in order to exclude them from
218 SOCIAL RATIONALITY

Table 10.2 A social contract

Overnight rule
If someone stays overnight in the cabin, then that person must bring along a bundle
of wood from the valley.
The cards below have information about four hikers. Each card represents one per-
son. One side of the card tells whether the person stayed overnight in the cabin,
and the other side tells whether he or she carried a bundle of wood. Indicate only
the card(s) you definitely need to turn over to see if the rule has been violated.
STAYS
DOES NOT STAY CARRIED NO
OVERNIGHT IN CARRIED WOOD
OVERNIGHT WOOD
THE CABIN

the benefits of cooperation. The second conjecture, however, rejects any role
of cheating detection in the selection task, claiming that people are, for some
reason, better at reasoning about social contracts than about numbers-and-
letters problems. Social contracts may be more "interesting" or "motivating,"
or people may have some "mental model" for social contracts that affords
"clear" thinking. Although this alternative is nebulous, it needs to be taken
into account; in her tests, Cosmides (1989) never distinguished between social
contracts and cheating detection.
But one can experimentally disentangle social contracts from cheating de-
tection. Klaus Hug and I also used social contracts but varied whether the
search for violations constituted looking for cheaters or not (Gigerenzer & Hug,
1992). For instance, consider the following social contract: "If someone stays
overnight in the cabin, then that person must bring along a bundle of wood
from the valley" (Table 10.2). This was presented in one of two context stories.
The "cheating" version explained that a cabin high in the Swiss Alps serves
as an overnight shelter for hikers. Because it is cold and firewood is not oth-
erwise available at this altitude, the Swiss Alpine Club has made the rule that
each hiker who stays overnight in the cabin must bring along a bundle of
firewood from the valley. The participants were cued to the perspective of a
guard who checks whether any of four hikers has violated the rule. The four
hikers were represented by four cards (Table 10.2).
In the "no-cheating" version, the participants were cued to the perspective
of a member of the German Alpine Association, visiting the same cabin in the
Swiss Alps to find out how it is managed by the local Alpine Club. He observes
people carrying firewood into the cabin, and a friend accompanying him sug-
gests that the Swiss may have the same overnight rule as the Germans, namely,
"If someone stays overnight in the cabin, then that person must bring along a
bundle of wood from the valley." That this is also the Swiss Alpine Club's rule
is not the only possible explanation; alternatively, only its members (who do
not stay overnight in the cabin), and not the hikers, might bring firewood. The
participants were now in the position of an observer who checks information
to find out whether the social contract suggested by his friend actually holds.
 DOMAIN-SPECIFIC REASONING 219

This observer does not represent a party in a social contract. The participants
instruction was the same as in the cheating version.
Thus, in the cheating scenario, the observation "benefit taken and cost not
paid" means that the party represented by the guard is being cheated; in the
no-cheating scenario, the same observation suggests only that the Swiss Alpine
Club never made the supposed rule in the first place.
Assume as true the conjecture that what matters is only that the rule is a
social contract, making the game-theoretic model (which requires a cheating
mechanism) irrelevant. Because in both versions the rule is always the same
social contract, such a conjecture implies that there should be no difference
in the selections observed. In the overnight problem, however, 89% of the
participants selected "benefit taken" and "cost not paid" when cheating was
at stake, compared with 53% in the no-cheating version (Figure 10.1). Simi-
larly, the averages across all four test problems used were 83% and 45%, re-
spectively, consistent with the game-theoretic account of cooperation (Giger-
enzer & Hug, 1992).

Figure 10.1 The absence of the possibility of being cheated reduces the "bene-
fit taken and cost not paid" selections (which coincide with the P & not-Q
selections), even when all rules are social contracts. From Gigerenzer and
Hug (1992).
220 SOCIAL RATIONALITY

The "availability theory," in contrast, predicts no difference between the

cheating and the no-cheating versions because the rules are identical and the
context stories are similar in content and length. Pragmatic reasoning schema
theory (Cheng & Holyoak, 1985) also predicts no difference, because the rules
are in both versions permission or obligation rules, and cheating detection
motivations and perspectives do not enter this theory.
These results support the assumption of a cheater-detection mechanism that
guides information selection, as postulated by social contract theory. The study
also shows that the concepts of a social contract rule and a cheater-detection
mechanism can be experimentally separated.

Perspective Change

Do Social Contracts Simply Facilitate Logical Reasoning?

In most of Cosmides's tests, the predicted "benefit taken" and "cost not paid"
selections corresponded to the truth conditions of conditionals in preposi-
tional logic. Thus a third conjecture would be that social contracts may some-
how facilitate logical reasoning, which we tested by deducing predictions from
the cheating-detection hypothesis that contradicted prepositional logic (Gig-
erenzer & Hug, 1992). The key to these tests is that cheating detection is prag-
matic and perspectival, whereas prepositional logic is aperspectival. For in-
stance, in the day-off problem in Table 10.3, one group of participants was
cued to the perspective of an employee, in which case cheating detection and
prepositional logic indeed predict the same cards. For a second group of par-
ticipants, we switched the perspective from employee to employer but held
everything else constant (the conditional statement, the four cards, and the
instruction shown in Table 10.3). For the employer, being cheated means "did
not work on the weekend and did get a day off"; that is, in this perspective
participants should select the "did not work on the weekend" and the "did

Table 10.3 A social contract with perspective change

Day-off rule
If an employee works on the weekend, then that person gets a day off during the
week.
The cards below have information about four employees. Each card represents one
person. One side of the card tells whether the person worked on the weekend, and
the other side tells whether the person got a day off during the week. Indicate only
the card(s) you definitely need to turn over to see if the rule has been been violated.
WORKED ON DID NOT WORK
DID GET A DAY DID NOT GET A
THE ON THE
OFF DAY OFF
WEEKEND WEEKEND
 DOMAIN-SPECIFIC REASONING 221

get a day off" cards, which correspond to the "not-P" and "Q" cards. (Note
that "not-P & Q" selections have rarely been observed in selection tasks.) Thus
perspective change can play cheating detection against general-purpose logic.
The two competing predictions are: If the cognitive system attempts to detect
instances of "benefit taken and cost not paid" in the other party's behavior,
then a perspective switch implies switching card selections; if the cognitive
system reasons according to propositional logic, however, pragmatic perspec-
tives are irrelevant and there should be no switch in card selections.
The results showed that when the perspective was changed, the cards se-
lected also changed in the predicted direction (Figure 10.2). The effects were
strong and robust across the three rules tested (Gigerenzer & Hug, 1992). For
instance, in the employee perspective of the day-off problem, 75% of the par-
ticipants had selected "worked on the weekend" and "did not get a day off,"
but only 2% had selected the other pair of cards. In the employer perspective,
this 2% (who had selected "did not work on the weekend" and "did get a day
off") rose to more than 60%. The result is consistent with the thesis that at-
tention is directed toward information that could reveal oneself (or one's
group) as being cheated in a social contract but is inconsistent with the claim
that reasoning is directed by propositional logic independent of content.
Thus social contracts do not simply facilitate logical reasoning. I believe
that the program of reducing context merely to an instrument for "facilitating"
logical reasoning is misguided. My point is the same as for Property Alpha
(Chapter 9). Reasoning consistent with propositional logic is entailed by some
perspectives (e.g., the employee's), but is not entailed by other perspectives
(e.g., the employer's).
Two additional conjectures can be dealt with briefly. First, several authors
have argued that the cheating-detection thesis can be invalidated because "log-
ical facilitation" (large proportions of "P & not-Q" selections) has also been
found in some conditional statements that were not social contracts (e.g.,
Cheng & Holyoak, 1989; Politzer & Nguyen-Xuan, 1992). This conjecture mis-
construes the thesis in two respects. The thesis is not about "logical facilita-
tion"; the conjunction "benefit taken and cost not paid" is not the same as the
logical conjunction "P & not-Q," as we have seen. Furthermore, a domain-
specific theory makes, by definition, no prediction about performance outside
its own domain; it can only be refuted within that domain.
The second conjecture also tries to reduce the findings to propositional
logic, pointing out that a conditional that states a social contract is generally
understood as a biconditional "if and only if." In this case all four cards can
reveal logical violations and need to be turned over. However, it is not true
that four-card selections are frequent when cheating detection is at stake. We
found in about half of the social contract problems (12 problems, each an-
swered by 93 students) that not a single participant had selected all four cards;
for the remaining problems, the number was very small. Only when cheating
detection was excluded (the no-cheating versions) did four-card selections in-
crease to a proportion of about 10% (Gigerenzer & Hug, 1992). There is, then,
222 SOCIAL RATIONALITY

Figure 10.2 Social contracts in which both parties have the option to cheat
allow us to test whether reasoning about social contracts follows aperspecti-
val propositional logic (that is, the hypothesis that the conditional rule is in-
terpreted as a material conditional) or the pragmatic and domain-specific
goal of cheating detection. Results show that in both perspectives (Party A
and Party B; e.g., employer and employee), participants search for informa-
tion that could reveal that their party is being cheated, whether this informa-
tion corresponds to P & not-Q (as the material conditional would suggest) or
to not-P & Q. This result also does not support the hypothesis that partici-
pants interpret the rule as a biconditional, which implies that they would
have to check all four cards. Four-card selections were rare. From Gigerenzer
and Hug (1992).

no evidence that participants follow propositional logic even if we assume that

they interpret the implication as a biconditional. In other words, people were
not looking for cheaters per se but only for evidence of being cheated them-
selves. Our participants were not reasoning with a Kantian moral but with a
Machiavellian intelligence.
Logical reductionism cannot explain how the mind infers that a particular
conditional should be understood as a material implication, a biconditional,
or something else. This inference is accomplished, I believe, by coding the
 DOMAIN-SPECIFIC REASONING 223

specific content of the conditional statement as an instance of a larger domain,

such as social contract, threat, and warning (Fillenbaum, 1977).

Conclusions

The evolutionary theory of cooperation illustrates how to begin constructing

a theory of cognition situated in social interaction. The idea is to start with a
specific design that a cognitive system requires for social interaction, rather
than with a general-purpose, formal system—in other words, to start with the
functional and see what logic it entails, rather than to impose some logic a
priori. The virtues of this approach are as evident as its unresolved questions.
Among these questions are: How can we precisely describe the "Darwinian
algorithms" that determine when a social contract is in place? How does the
mind infer that the conditional statement "If you touch me, then I'll kill you"
does not imply a social contract but a threat? What are the cues coding this
specific statement into the domain of "threats" rather than "social contracts"?
Once a statement is categorized into a particular domain, what distribution of
attention is implicated by that domain? In a threat, for example, attention
needs to be directed to information that can reveal being bluffed or double-
crossed rather than cheated (Fillenbaum, 1977). The challenge is to design
theoretical proposals for reasoning and inference processes in other domains
of human interaction beyond cooperation in social contracts.
To approach reasoning as situated in social interaction is to assume that the
cognitive system (a) generalizes a specific situation as an instance of a larger
domain and (b) reasons about the specific situation by applying a domain-
specific cognitive module. This raises two questions about the nature of the
domains and the design of the modules.

What Are the Domains and at What Level of Abstraction

Are They Located?
Imagine a vertical dimension of abstraction, in which the specific problem
corresponds to the lowest level of abstraction, and some formal representation
of the problem, stripped of any content and context, to the highest. Two dia-
metrically opposed views correspond to the ends of this continuum of abstrac-
tion. First, it may be argued that the cognitive system operates at the lowest
level of abstraction, guided by familiarity and availability of instances in mem-
ory (e.g., Griggs & Cox, 1982). Second, it may be argued that the cognitive
system generalizes the specific problem to the highest level of abstraction (e.g.,
prepositional logic), performs some logical operations on this formal represen-
tation, and translates the result back into application to the specific problem.
Variants of the latter view include Piaget's theory of formal operations and
mental logics.
224 SOCIAL RATIONALITY

The primary challenge of domain specificity is to find a level of abstraction

between the two extremes, where some content is stripped but an adequate
amount retained. For instance, the level of social contracts and cheating de-
tection could turn out to be too abstract, because cheating may assume differ-
ent forms (e.g., in contracts in which both or only one side can be cheated;
Gigerenzer & Hug, 1992), requiring different procedures of cheating detection.
In contrast, the notions of social contracts and cheating detection may not be
abstract enough, needing to be stripped of some content and placed at the more
general level of social regulations, such as obligations, permissions, and other
kinds of deontic reasoning (Cheng & Holyoak, 1985, 1989; Over & Manktelow,
1993). This focus on level of abstraction parallels Rosch's (1978) concern with
"basic level objects."

What Is the Design of a Domain-Specific

Cognitive Module?

A cognitive module organizes the processes—such as inference, emotion, and

the distribution of attention—that have been evolved and learned to handle a
domain. To classify a specific situation as an instance of a given domain, a
cognitive module must be connected to an inferential mechanism. For in-
stance, David Premack and others assume that humans and primates first clas-
sify an encounter as an instance of either social interaction (in the broadest
sense) or interaction with the nonliving world. There is evidence that the cues
used for this inference involve motion patterns, analyzed by cognitive systems
to classify objects in the world as "self-propelled" or not; this analysis is rem-
iniscent of Fritz Heider's and Albert Michotte's work (Premack, 1990; Sperber,
1994; Thines, Costall, & Butterworth, 1991). Cognitive modules dealing with
something external that has been coded as "self-propelled" attend to infor-
mation such as whether it is friend or enemy, prey or predator. For a module
that deals with inanimate things, no attention needs to be directed to infor-
mation of this kind. Domain-specific modules can thus distribute attention in
a more focused way than a domain-general mechanism. The challenge now
before us is to come up with rich and testable models about the design of
cognitive modules.

Toward a Social Rationality

Researchers in several disciplines are converging on a domain-specific pro-

gram of studying reasoning and inference situated in social interaction. Pri-
matologists have joined philosophers and psychologists in studying "social
intelligence" (Kummer et al., 1997) and "Machiavellian intelligence" (Byrne
& Whiten, 1988). Linguists and philosophers have begun to reinterpret the
conclusions of experimental research, in particular the so-called fallacies and
biases, by arguing that the interaction between participant and experimenter
is constrained by conversational rather than formal axioms (e.g., Adler, 1991;
Grice, 1975; Sperber & Wilson, 1986). Social psychologists have tested some
 DOMAIN-SPECIFIC REASONING 225

of these proposals experimentally, concluding among other things that perva-

sive reasoning "biases" may not reflect universal shortcomings of the human
mind but instead the application of Gricean conversational principles that con-
flict with what formal logic seems to dictate (e.g., Hertwig & Gigerenzer, 1999;
Hilton, 1995; Schwarz et al., 1991). Similarly, Tetlock's (1992) concept of "ac-
countability" models the social side of decision making by emphasizing that
people do not simply choose the better alternative, but in certain social inter-
actions, choose the alternative they can better justify. Developmental psychol-
ogists have departed from Piaget's general-purpose processes and investigate
the domain-specific processes and their change during development (Hirsch-
feld & Gelman, 1994a). The convergence of these approaches promises a new
vision of reasoning and rationality situated in social context.
I can only hope that this chapter will inspire some readers to rethink the
imposition of formal axioms or rules as "rational," independent of context.
The challenging alternative is to put the psychological and the social first—
and then to examine what formal principles these entail. We need less Aristotle
and more Darwin in order to understand the messy business of how to be
rational in the uncertain world of interacting human beings. And we may have
to abandon a dream. Leibniz's vision of a sovereign calculus, the Universal
Characteristic, was a beautiful one. If only it had proved true.
 11
The Modularity of Social Intelligence

In a "protected threat," a baboon induces a dominant member of its group to

attack a third one. The baboon appeases the dominant member, whom it uses
as a tool to threaten the target, and maneuvers to prevent the target from doing
the same (Kummer, 1988). This "social tool use" is mastered by baboons at
puberty, whereas chimpanzees are adult before they learn to use a stone as a
tool for cracking hard nuts (Boesch & Boesch, 1984). Primates appear to ma-
nipulate social objects with more ease and sophistication than physical tools.
Observations such as these have suggested that primate intelligence is de-
signed primarily for the social rather than the physical and have led to the
Machiavellian intelligence hypothesis (Whiten & Byrne, 1988) or social intel-
ligence hypothesis (Kummer et al., 1997). The social intelligence hypothesis
assumes that social intelligence preceded (evolutionary headstart), influences,
and qualitatively differs from nonsocial intelligence. The term Machiavellian
intelligence emphasizes the besting of rivals for personal gain over coopera-
tion, whereas the term social intelligence (which is the more general term) is
neutral on the balance between exploitation and cooperation.
The social intelligence hypothesis is both stimulating and vague. It is stim-
ulating because it reminds us that whenever psychologists study intelligence
and learning in humans or animals, it is almost invariably about inanimate
objects: symbols, sticks, and bananas. It is vague because the nature of the
intelligence it invokes is largely unclear, and as a consequence, the mecha-
nisms of social intelligence have not yet been specified. The modular version
of the social intelligence hypothesis I propose assumes that social intelligences
come in the plural, as do nonsocial intelligences. The notion of modularity is
nothing new, but, as the introduction to a recent book on this topic illustrates,
it means many things to many people (Hirschfeld & Gelman, 1994b). There
exist several important approaches to modularity (e.g., Baron-Cohen, 1995;
Leslie, 1994; Millikan, 1984; Pinker, 1994). It will soon become clear which
ones inspired my own ideas. I will now describe the modular organization of
social intelligence. Reader be warned: from here on is speculation.

226
 THE MODULARITY OF SOCIAL INTELLIGENCE 227

The thesis that social intelligence is modular is motivated by two reasons:

the shortsightedness of natural selection and the combinatorial explosion of
intelligent systems. Natural selection works without a big plan but results in
specific adaptations accumulated over generations. Thus it seems unlikely that
natural selection designed a general-purpose intelligence that embodies, say,
the Piagetian formal operations or a Bayesian inference machine. Even if this
had happened, such a general-purpose intelligence runs into the problem of
combinatorial explosion, as evidenced by the frame problem in artificial in-
telligence: Unless the infinite possibilities to combine elements and relations
in a general-purpose system are drastically reduced by semantic constraints,
an organism would be paralyzed and unable to react in time. For instance,
unless attention is constrained to specific types of interactors and interactions,
and semantic structure is a priori built in that tells the organism what to learn,
what to look for, and what to ignore, an intelligent organism would be unable
to perform even the most elementary tasks: to detect predators, prey, and mates
and be fast enough to survive and reproduce.
I propose the following assumptions about the nature of modular social
intelligence:

A Module for Social Intelligence Is a Faculty,

Not a General Factor

Intelligence is often assumed to be of one kind, one general ability that helps
an organism cope with all situations—such as Francis Gallon's "natural abil-
ity," Spearman's general intelligence factor g, and the numberless definitions
that start with "intelligence is the general ability to . . ." The thesis that intel-
ligence is a unified general ability has been created only recently, in the mid-
19th century, by Francis Galton, Herbert Spencer, and Hippolyte Taine, among
others (Daston, 1992). The idea of one general intelligence was motivated by
Darwin's theory of evolution (Galton was Darwin's cousin) and seemed to pro-
vide the missing continuum between animals and humans, as well as between
human races, and last but not least, between men and women.
Such a unified general ability was alien to the earlier faculty psychology,
which dated back to Aristotle. Faculty psychology posited a collection of fac-
ulties and talents in the mind, such as imagination, memory, and judgment.
These faculties organized an intricate division of mental labor, and no single
one nor their sum coincided with our concept of intelligence (Daston, 1992).
Faculty psychology was revived, in the language of factor analysis, in the late
1930s when L. L. Thurstone claimed about seven primary mental abilities. In
the second half of the 20th century, the mind has become again a crowded
place. Evidence has been announced for dozens of factors of intelligence, and
Guilford and Hoepfner (1971) even claimed the confirmation of some 98 factors
of cognitive ability (see Carroll, 1982). Cognitive psychologists who use ex-
periments rather than IQ tests also divide up cognition in terms of faculties
(but you will not catch one using that term): deductive reasoning, inductive
228 SOCIAL RATIONALITY

reasoning, problem solving, memory, attention, judgment and decision making,

and so forth.
A modular organization of intelligence assumes, similar to the earlier fac-
ulty psychology before Galton, several intelligences that have special designs,
and not one general-purpose intelligence. The general thesis is that social in-
telligence is an "adaptive toolbox" with several special tools rather than one
single general-purpose hammer (Gigerenzer & Todd, 1999).

A Module for Social Intelligence Is Domain Specific,

Not Domain General

Intelligence modules, however, are not like Thurstone's primary mental abili-
ties and faculties such as reasoning. I distinguish between two types of fac-
ulties: domain specific and domain general. Faculties such as deductive rea-
soning, memory, and numerical ability (as well as such factors as "fluid" and
"crystallized" intelligence) are assumed to treat any content identically, that
is, to operate in a domain-general way. The laws of memory, for instance, in
this view, are not about what is memorized; they are formulated without ref-
erence to content. Fodor (1983) called these domain-general faculties "hori-
zontal" as opposed to "vertical" domain-specific faculties. The modularity of
social intelligence, I propose, is vertical.
The doctrine of domain-general mechanisms flourished in Skinner's behav-
iorism, before it was generally rejected following experimental work by John
Garcia and others (see Chapter 10). Learning through imitation (rather than
reinforcement) is also reported to be domain specific. Rhesus monkeys, for
instance, reared in the laboratory show no fear toward venomous snakes. How-
ever, one will show fear if it sees another monkey exhibiting a fear reaction
toward snakes. Yet the monkey does not become afraid of just any stimulus:
If it sees another monkey emit a fear reaction toward a flower, it does not
acquire a fear of flowers (Cosmides & Tooby, 1994b; Mineka & Cook, 1988).
Learning by imitation of others, like learning by association, is simultaneously
enabled and constrained by specific "expectations" of what to avoid, what to
fear, or more generally, what causal connections to establish. Without domain-
specific mechanisms, an organism would not "know" what to look for, nor
which of the infinite possible causal connections to check. Such an organism
would be paralyzed by data analysis like the quantophrenic researcher who
measures everything one can think of, computes correlation matrices of di-
nosaurian dimensions, and blindly searches for significant correlations. De-
spite the available evidence to the contrary, Skinner's ideal of domain gener-
ality has survived the cognitive revolution and is flourishing in present-day
conceptions of the mind.
Domain generality is possibly the most influential and suspect idea in 20th-
century psychology. Psychologists love to organize their field by horizontal
faculties such as attention, memory, perception, problem solving, and judg-
 THE MODULARITY OF SOCIAL INTELLIGENCE 229

ment and decision making. Terms such as these organize the chapter structure
of textbooks, the specialties of scientific journals, the divisional structure in
grant agencies, and the self-identity of numerous colleagues. Psychologists
tend to identify with horizontal faculties, not with domains.
I propose, in contrast, that modules for social intelligence are domain spe-
cific. How should we think about these modules? Fodor (1983), a vehement
proponent of modularity, has argued that modularity is restricted to input sys-
tems (the senses) and language, whereas central processes such as reasoning
are domain general. I term this the "weak modularity thesis." In his view,
modules are specifically designed mechanisms for voice recognition in con-
specifics, for face recognition in conspecifics, and color perception, among
others. I disagree with Fodor's opposition between modular sensory processes
(and language) and general-purpose central processes. Social intelligence in-
volves both perceptual processes and mechanisms for reasoning and inductive
inference. For instance, assume there is a module for social contracts, that is,
a module that enables cooperation between unrelated conspecifics for their
mutual benefit. Such a module would need to incorporate both "central" pro-
cesses, such as cost-benefit computations and search algorithms for informa-
tion that could reveal that one is being cheated, and sensory processes such
as face recognition. Without both "peripheral" and "central" mechanisms, nei-
ther social contracts nor cheating detection would be possible.
What I call the "strong modularity thesis" postulates that modules include
central processes as well as sensory mechanisms (and language). The function
of modules is not tied to "peripheral" as opposed to "central" processes.
Rather, their function is to solve specific problems of adaptive significance and
to do this quickly. A problem of adaptive significance can be described as an
evolutionarily recurrent problem whose solution promoted reproduction (Cos-
mides & Tooby, 1994a, b; Miller & Todd, 1995, 1998). Candidates include co-
alition forming and cooperation, foraging, predator avoidance, navigation,
mate selection, and rearing children. To solve such problems, modules need
to combine "peripheral" and "central" processes. Thus the domains (more
precisely, the "proper" domains; see next section) of modules are important
adaptive problems and not just perceptual (plus language) tasks.

A Module for Social Intelligence Has a Proper

and an Actual Domain

Assume there is a social intelligence module designed for handling social con-
tracts in a hunter-gatherer society. The proper domain of the module may have
been the exchange of food for the mutual benefit of both parties involved in
the contract (because food sharing is not too common among animals, an al-
ternative hypothesis would be that the proper domain concerned social ser-
vices such as alliance formation; see Harcourt & de Waal, 1992). Generations
later, currency has been developed, and the module's representation of possi-
230 SOCIAL RATIONALITY

ble benefits and costs exchanged in a social contract needs to be expanded to

tokens that can be exchanged for benefits. Soon economic systems will be
invented in which the exchange of hard currency is no longer the norm, and
benefits and costs become more and more abstract. The actual domain of the
module has shifted from the exchange of grain and meat to buying futures and
options. The mechanisms of the module, however, perform largely the same
task: a routine that leads individuals to enter into social contracts (sharing)
when resources (such as meat) are highly variable and scarce; a representation
of what the benefits and costs are for oneself and for one's kin; perceptual
algorithms and a memory that allows identification of the partner by face,
voice, or name recognition; and a search heuristic that looks for information
that could reveal that one is being cheated.
Thus the proper domain of a module is that for which the module ac-
tually evolved; the actual domain is one to which the module is transferred
or extended, following changes in the environment (Sperber, 1994). By being
transferred to new domains, the mechanisms of the modules themselves may
change; adaptations (modules) can become exaptations for new adaptations
(Gould & Vrba, 1982). The distinction between proper and actual domain is
a matter of degree rather than of kind, and the actual domain is most likely
larger than the proper domain. Modules for social intelligence in humans
seem to differ from those in primates in that they have larger actual domains
and in that the actual domains may have less overlap with the proper do-
main.
The actual domains of modules for human social intelligence can extend
beyond the human. Anthropomorphism is social intelligence reaching out be-
yond Homo sapiens. Anthropomorphism has counted as a scientific sin since
the 17th century, and earlier as a theological sin; nevertheless human intelli-
gence cannot resist projecting human social categories, intentions, and morals
onto nonhumans (Mitchell et al., 1997). Darwin himself practiced empathic
anthropomorphism (but not anthropocentrism) in particular with respect to
dogs; animal rights activists often invoke the same sentiment. Anthropomor-
phism of a less empathic nature extends to phylogenetically distant species:
"Rape" in scorpion flies and "ant slavery" are examples. In opposition to all
this, behaviorism values purifying scientific language, and a story told about
the Columbia University philosopher Sidney Morgenbesser illustrates this. Af-
ter B. F. Skinner gave a talk at Columbia, Sidney stood up and said: "Professor
Skinner, I always tried to understand what the essence of behaviorism is. Now
I think I know. Behaviorism is the denial of anthropomorphism for humans."
Social intelligence can reach beyond animals and still create powerful meta-
phors. Dawkins's (1976) "selfish gene" lives in a world of savage competition,
exploitation, and deceit. Physicists, chemists, and astronomers certainly cen-
sor similar anthropomorphic descriptions in press, but in scientists' informal
and private conversations, intentions are frequently attributed to particles and
matter (Atran, 1990).
 THE MODULARITY OF SOCIAL INTELLIGENCE 231

A Module Is Activated by a Triggering Algorithm

Assume there is a simple social organism with two modules for social intel-
ligence: One deals with social contracts, the other with threats. Thus this or-
ganism knows only two ways to deal with conspecifics: to engage with them
in the exchange of certain goods to their mutual benefit and to threaten indi-
viduals to get one's way (and react when others do so). As simple as the social
intelligence of this organism is, the organism needs to decide when to activate
the social contract module and when the threat module. All modules cannot
be activated at the same time because the very advantage of modularity is to
focus attention and to prevent combinatorial explosion. For instance, the social
contract module focuses attention on information that can reveal that the or-
ganism is being cheated, whereas this information is of no relevance for a
threat module. A threat module needs to attend to information that can reveal,
for instance, whether the other side is bluffing or whether high-status individ-
uals are present who could be used for "protected threat" (Kummer, 1988).
How is one of the two modules activated? I assume that there is a triggering
algorithm that attends to a small set of cues whose presence signals either
threat or social contract. These signals can include facial expressions, gestures,
body movements, and verbal statements. Assume that the organisms do have
language. A simple algorithm can quickly recognize whether a verbal state-
ment of the type "if you do X, then I do Y" is a threat or a social contract. If
Y is a negative consequence for me, and follows X in time, then I am being
threatened. If Y is a benefit for me and the temporal sequence can be either
way, then I am being offered a social contract. I call such simple heuristics "trig-
gering algorithms" because their function is to activate a module that can focus
attention, emotion, and behavioral responses so that fast reaction is possible.
Triggering algorithms can err, that is, not activate the appropriate module,
such as mistaking a serious threat for pretend play. The likelihood of triggering
the wrong module may increase when there are more than two modules, but
redundancy in cues, such as verbal cues, facial cues, and gestures, may reduce
errors.

Modules Are Hierarchically Organized

When a mind has not just two but a large number of modules, a single trig-
gering algorithm may be too slow to discriminate between all possibilities si-
multaneously. In such a socially more intelligent mind, modules can be hier-
archically connected by triggering algorithms, as in a sequential decision tree.
Hierarchical organization corresponds to the idea that species started out with
a few modules, to which more specialized modules were added later in phy-
logeny, and to Wimsatt's (1986) notion of generative entrenchment.
Assume I march through a forest at night. Visibility is poor, a storm is com-
ing up, and I suddenly see the contours of a large dark object that seems to
232 SOCIAL RATIONALITY

move slowly. A triggering algorithm needs to decide quickly whether the object
is "self-propelled" (animal or human) or not (plant or physical object; Premack
& Premack, 1994). According to the Premacks, this decision is based on the
object's motion pattern. Recall the demonstrations by Fritz Heider in which
the motion patterns of two points in two-dimensional space make us "see" the
points as animate or inanimate, chasing, hunting, hurting, or supporting one
another (e.g., Heider & Simmel, 1944). These are beautiful demonstrations, but
they include no descriptions of the algorithms that make us see all these social
behaviors. How could the first triggering algorithm work? A simple algorithm
would analyze only external movements (such as the direction and accelera-
tion of the object) and not internal movements (the relative movement of the
body parts). For instance, if a motion pattern centers on my own position, such
as an object that circles around me or speeds up toward me, the algorithm
infers a self-propelled object. Moreover, it infers a self-propelled object that
takes some interest in me. Motion patterns that center around the object's own
center of gravity, in contrast, indicate that the object is a plant (e.g., a tree).
Now, if the motion pattern indicates that the object is self-propelled, the trig-
gering algorithm may activate a module for unrecognized self-propelled ob-
jects. This module will immediately set the organism into a state of physio-
logical and emotional arousal, initiate behavioral routines such as stopping
and preparing to run away, and activate a second, more specialized triggering
algorithm whose task is to decide whether the self-propelled object is animal
or human. Assume that this second triggering algorithm infers from shape and
motion information that the object is human. A module for social encounters
with unknown humans is subsequently activated, which initiates a search for
individual recognition in memory and may initiate an appeal for voice contact
in order to find out whether the other is friend or enemy, is going to threaten
or help me, and so on. This is pure speculation, but one might work out the
mechanisms of a hierarchical organization along these lines.
Modules that are hierarchically organized can act quickly, as only a few
branches of the combinatorial tree need to be traveled. For instance, if the first
triggering algorithm had indicated that the unknown object was not self-
propelled, then all subsequent information concerning whether it is human or
animal, friend or enemy, or predator or prey could have been ignored.

An Intelligence Module Works with Fast

and Frugal Heuristics

There are two views about the machinery of intelligent behavior. The classical
view is that the laws of probability theory or logic define intelligent processes:
Intelligent agents have rules such as Bayes's rule, the law of large numbers,
transitive inference, and consistency built in. This was the view of the Enlight-
enment mathematicians, to which Jean Piaget added an ontogenetic dimension,
and it still is a dominant view in economics, cognitive psychology, artificial
intelligence, and optimal foraging theory. For instance, Bayes's rule has been
 THE MODULARITY OF SOCIAL INTELLIGENCE 233

proposed as a normative or descriptive model of how animals infer the pres-

ence of predators and prey (Stephens & Krebs,-1986), as well as how humans
reason, categorize, and memorize (Anderson, 1990). The problem with this
view is that in any sufficiently rich environment, Bayesian computations be-
come mathematically so complex that one needs to assume that minds are
"Laplacean demons" that have unlimited computational power, time, and
knowledge. To find a way out of this problem, researchers often make unreal-
istic assumptions about the structure of natural environments, namely as-
sumptions that reduce the Bayesian computations. Anderson (1990), for in-
stance, found himself forced to make the assumption that environmental
features would be generally independent, in order to save the fiction of the
Bayesian homunculus in the mind (Gigerenzer, 1991c). Despite their psycho-
logical implausibility, Laplacean demons are bustling in contemporary theories
of the mind: as models of choice, categorization, estimation, and inference,
among others. The rationale seems to be this: Cognition is rational, Bayes's
rule defines rationality, ergo cognition works with Bayes's rule. The same ra-
tionale seems to hold for other statistical tools that turned into theories of
mind, such as multiple regression, Neyman-Pearson decision techniques, and
analysis of variance (Chapter 1).
The second view, the one I propose here, is that modules of social intelli-
gence, including the triggering algorithms, work with fast and frugal strategies
instead of the costly "optimal" algorithms. There are several reasons that favor
simple and specifically designed heuristics rather than expensive and general
ones. First, there is, in fact, no single method of inference—statistical or log-
ical—that works best in all real-world contexts. Second, as mentioned before,
in real-world situations, "optimal" computations can quickly become so com-
plex that one is forced to make highly simplifying assumptions about the en-
vironment. Third, algorithms for social intelligence need to work under con-
straints of limited time and knowledge—for instance, one may not have the
time to search for further information. Fourth, the means and ends of social
intelligence are broader than consistency (coherence) and accuracy—the ac-
cepted norms of logic and statistics. Social intelligence can involve being in-
consistent (e.g., adaptive unpredictability may be optimal in competitive sit-
uations: the opponent will be unable to predict one's behavior), taking high
risks in trying to come out first (that is, options with low probabilities rather
than those that maximize expected value), responding quickly rather than ac-
curately (e.g., to make too long a pause in a conversation in order to think of
the best answer can be embarrassing and seen as impolite), and making deci-
sions that one can justify or defend afterward (Todd, 2000a, b).
The key argument against fast and frugal heuristics is that their simplicity
raises the suspicion that they are really bad. But this need not be the case:
Simple principles can be quick-and-clean rather than quick-and-dirty. Con-
sider a simple heuristic for choice called "Take The Best" (Gigerenzer & Gold-
stein, 1996). Take The Best infers which of two alternatives scores higher on
some criterion, such as which of two food items is more dangerous or which
of two cities has a higher population. Take The Best bases decisions on one
234 SOCIAL RATIONALITY

single reason, namely on the first good reason on which two alternatives differ.
The first good reason can be simply that the individual does not recognize (has
never heard of) one of the two alternatives. This "recognition heuristic" seems
to operate in domains in which recognition is correlated with the variable that
needs to be inferred. For instance, rats who can choose between food that they
recognize and food that is new to them do not accept the new food unless they
have smelled it on the breath of a fellow rat (Gallistel et al., 1991). The sur-
prising result is that simple heuristics such as Take The Best can make as
accurate inferences about real-world environments as costly statistical algo-
rithms of the Laplacean demon type (Martignon & Laskey, 1999).
If short-sighted evolution has equipped us with adaptive heuristics rather
than with the collected works of logic and probability theory, this result in-
dicates that we need to rethink human rationality. The challenges are to un-
derstand what these heuristics are and to describe the structure of the envi-
ronments in which they can perform well and in which they cannot. My
proposal is that both the triggering algorithms and the mechanisms of the mod-
ule can be modeled as fast and frugal heuristics (Gigerenzer et al., 1999).

Modules Combine Cognitive, Emotional, and Motivational

Tools to Guide Inference and Behavior

In each module, various cognitive, emotional, behavioral, and motivational

processes are wired together. A social contract module, for instance, includes
perceptual machinery to recognize different individuals; a long-term memory
that stores the history of past exchanges with other individuals in order to
know when to cooperate, when to defect, and when to punish for defection;
knowledge about what constitutes a benefit and what a cost for oneself; and
emotional reactions such as anger that signal to others that one will go ruth-
lessly after cheaters (Cosmides & Tooby, 1992; Gigerenzer, 1995). Intelligent
modules can use cognitive, emotional, or motivational processes vicariously
as means to achieve a goal. For instance, emotional reactions (such as disgust
in matters of food) can substitute for learning from experience when events
are too rare or too deadly for individual learning.
The very challenge of a modular concept of social intelligence is that it
crosses the established boundaries of horizontal faculties. Thereby, the notion
of modularity questions the institutionalized subdisciplines of psychology. It
opens up a conception of intelligence that integrates cognition with, rather
than setting it apart from, other adaptive functions such as motivations and
emotions.

Summary

In 1976, Nick Humphrey wrote a stimulating essay on the social function of

intellect that provided "the single most important seed" (Byrne & Whiten,
 THE MODULARITY OF SOCIAL INTELLIGENCE 235

1988, p. 1) for Machiavellian intelligence. Humphrey's paper contains thought-

provoking ideas about social intelligence but also the seed of a sterile research
methodology. In the postscript we are told that his "central thesis" demands
"that there should be a positive correlation across species between 'social com-
plexity' and 'individual intelligence' " (Humphrey, 1976/1988, p. 26). A labo-
ratory test of social skill, Humphrey proposes, is urgently needed, which
"ought, if I am right, to double as a test of 'high-level intelligence' " (p. 26).
This proposal, as innocent as it looks, steers toward repeating a grave error
that has swamped research on (nonsocial) intelligence. The error is to start
with no theory but with seductive everyday concepts such as social skill and
creative intelligence, then go on to design tests to measure these vague con-
cepts, and to pray to heaven that these tests miraculously transform loose
thinking into precise mechanisms of intelligence.1 Despite prayers that were
backed up by statistics, some 90 years of factor analyzing and correlating IQ
tests has not noticeably increased our understanding of the mechanisms of
human intelligence. I fear that Humphrey's proposal to look for correlations
between some tests for social complexity, social skill, and individual intelli-
gence will be doomed to the same failure.
The alternative is to start boldly and theoretically. The challenge is to design
the possible mechanisms of social intelligence and to test these by means of
experiment, observation, and simulation. The mechanisms of a modular social
intelligence that I have outlined, as speculative as they are, can serve as a start.
We will have to take up some hard questions. What are the domains (proper
and actual) for a given species? What is the mechanism of a module? What is
the algorithm that triggers a module? If we join forces, we can do it.

1. The notion of test intelligence, left undefined in its content, has had many faces,
including the moral and the social (Daston, 1992). For instance, the creators of the first
intelligence tests, Binet and Simon (1914), asked questions about social skills, such as
"Why should one judge a person by his acts rather than by his words?" In the first
edition of the Stanford-Binet Test, Louis Terman (1916) expressed the intimate link
between lack of intelligence and morally inappropriate behavior in no uncertain terms:
"Every feeble-minded woman is a prostitute." In the 1937 revision of the text (with M. A.
Merrill), this sentence was deleted. Piece by piece, IQ tests became pure and puritan.
This page intentionally left blank
 V
COGNITIVE ILLUSIONS AND
STATISTICAL RITUALS

The "discovery" of cognitive illusions was not the first assault on human
rationality. Sigmund Freud's attack is probably the best known: According to
him, the unconscious wishes and desires of the human id are a steady
source of intrapsychical conflict that manifests itself in all kinds of irrational
fears, beliefs, and behavior. But the cognitive-illusion assault is stronger than
the psychoanalytic one. It does not need to invoke a conflict between ra-
tional judgment and unconscious wishes and desires to explain humans' ap-
parent irrationality: Judgment is itself fundamentally deficient. Homo sap-
iens appears to be a misnomer. During the last few decades, cognitive
illusions have become fodder for classroom demonstrations and textbooks.
Isn't it fun to show how dumb everyone else is, and after all, aren't they?
In the spring of 1990, I gave a talk in the Department of Psychology at
Stanford University entitled "Beyond heuristics and biases: How to make
cognitive illusions disappear." The first chapter in this section is an updated
version of this talk. It became the fountainhead of an ongoing, heated debate
over the nature of human rationality and the litany of sins people seem to
commit routinely against reason (e.g., Gigerenzer, 1996a; Kahneman & Tver-
sky, 1996), the so-called "rationality wars" (Samuels, Stich, & Bishop, in
press). Cognitive illusions have been linked to perceptual illusions, suggest-
ing that they are "inevitable illusions" (Piattelli-Palmarini, 1994). The politi-
cal implications of this view are not hard to see. Given the message that or-
dinary citizens are unable to estimate uncertainties and risks, one might
conclude that a government would be well advised to keep these nitwits out
of important decisions regarding new technologies and environmental risks.
In Chapter 12,1 criticize the narrow norms that make humans look irrational
and show how to make inevitable illusions "evitable."
Research on cognitive illusions is but one example of the more general phe-
nomenon of replacing statistical thinking with narrow, simplistic norms. The
statistical practices institutionalized in many social and medical sciences are
another case in point; they have little to do with statistical thinking and instead
238 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

promote statistical rituals. Textbooks teach our students the equivalent of com-
pulsive hand washing, the result being confusion and anxiety.
I once asked a well-known author who was busy preparing the latest edi-
tion of his best-selling statistical text for psychologists why he promoted the
usual incoherent mishmash of Fisherian and Neyman-Pearsonian prescrip-
tions for testing hypotheses (see Chapter 13). He did not try to deny the
problem, but he told me whom to blame for it. First, there was his publisher,
who had insisted that he supply a statistical cookbook and take out anything
that hinted at the existence of alternative tools for statistical inference,
which he did. Next, there were his fellow researchers, who did not aim to
truly understand statistics but to get their papers published. Finally, he
passed the blame on to the editors who demanded a statistical ritual and to
his university administration, which determined salary increases by counting
the number of papers published. When I asked him in what statistical meth-
ods he himself believed, he said that deep in his heart he was a Bayesian. I
was shocked. What a Faustian pact—an author successfully sells a method
in which he does not believe, which students and researchers then naively
mistake for the moral guidelines of doing science.
This is not to say that the many textbook authors who borrowed from
him are as aware of the confusion behind the ritual of null hypothesis test-
ing as he was. In my experience, many authors are innocent because igno-
rant, which is one way to maintain one's intellectual integrity. I wrote Chap-
ter 13 as an antidote to mindless statistics for both students and future
textbook writers.
The larger social and intellectual background for this chapter can be
found in The empire of chance: How probability changed science and every-
day life (Gigerenzer et al., 1989). This book was written by an interdiscipli-
nary group of scholars who studied the probabilistic revolution in the sci-
ences at the Center for Interdisciplinary Research in Bielefeld, a place where
nothing could distract us from work. Geoffrey Loftus reviewed the book in
Contemporary Psychology in 1991. When he became the editor of Memory &
Cognition in 1993, in his editorial statement he asked researchers to stop
submitting manuscripts with legions of p-values, t-values, and F-values and
instead present sound descriptive statistics, such as figures with error bars
around means. I admire him for having the courage to stand up against
mindless null hypothesis testing. A few years later, I asked Geoffrey how his
crusade was going. To Geoffrey's surprise, the resistance was coming from
the researchers. Most of them insisted on going through with the ritual. As
this case illustrates, editors alone cannot be blamed for psychologists' con-
tinued reliance on misguided statistical procedures. I might add that I have
never had a problem publishing experimental papers without null hypothe-
sis tests.
The story of null hypothesis testing in psychology is reminiscent of Hans
Christian Andersen's tale of the emperor's new clothes. In a sense, the proce-
dure has no clothes: Its outcome, the p-value, does not inform the reader
about the size of the effect, the probability that the null hypothesis is true,
 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS 239

the probability that the alternative hypothesis is true, or the probability that
the result is replicable. Nevertheless, studies in the United States, Great Brit-
ain, and Germany indicate that some 80% to 90% of academic psychologists
"see" one or more of these attractive garments on the p-value.
I have been asked what we should do instead of null hypothesis testing.
The answer is: not a new ritual. Chapter 6 illustrated one alternative. First,
the data is tested against multiple hypotheses—alternative models of cogni-
tive strategies—rather than against one null hypothesis. Second, multiple hy-
potheses are tested against the judgments of each individual participant
rather than against the average across individuals, thus enabling detection of
multiple strategies. Third, hypotheses are tested against both outcome and
process data rather than outcome data only.
The most important thing is to define candidate hypotheses before start-
ing the business of hypothesis testing. Null hypothesis testing encourages
theoretical laziness. To use it, one does not need to specify one's research
hypothesis or a substantive alternative except "chance." This scant require-
ment allows surrogates for theories to grow like weeds (Chapter 14). We
need statistical thinking, not statistical rituals.
This page intentionally left blank
 12
How to Make Cognitive Illusions Disappear

Social psychology was transformed by the "cognitive revolution." Cognitive

imperialism has been both praised and lamented. But a second revolution has
transformed most of the sciences so fundamentally that it is now hard to see
that it could have been different before. It has made concepts such as proba-
bility, chance, and uncertainty indispensable for understanding nature, society,
and the mind. This sweeping conceptual change has been called the "proba-
bilistic revolution" (Kriiger, Daston, & Heidelberger, 1987; Kriiger, Gigerenzer,
& Morgan, 1987). The probabilistic revolution differs from the cognitive rev-
olution in its genuine novelty and its interdisciplinary scope. Statistical me-
chanics, Mendelian genetics, Brownian motion, radioactive decay, random
drift, randomized experimental design, statistical inference—these are some of
the fruits of that transformation. Social psychology was no exception. It cur-
rently bears the marks of both the cognitive revolution and the probabilistic
revolution.
Probabilistic and statistical concepts were piggybacked onto cognitive con-
cepts. Some of the most popular theories and research programs owed their
genesis to an analogy between social cognition and "intuitive statistics." In
1967, for instance, Harold Kelley proposed that the layperson attributes a cause
to an effect in the same way as a statistician of the Fisherian school would, by
(unconsciously) calculating an analysis of variance (ANOVA). Research on the
ANOVA mind soon became mainstream social psychology (Kelley & Michaela,
1980). It is well documented in the history of science that statistics trans-
formed almost everything it touched. So has causal attribution (Chapter 1). Just
as statistical calculations are those of an individual statistician, attribution and
social cognition were investigated as the calculations of individual minds, con-
firming the individualism in social psychology (Newcombe & Rutter, 1982).
More recently, Bayesian statistics, rather than Fisherian statistics, has been
used as a yardstick to evaluate social cognition, and as measured by this new
yardstick, many people's judgments seemed to be flawed by fallacies and errors
in statistical reasoning. "Hot" motivational terms were replaced by the "cold"

241
242 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

cognitive language of intuitive statistics. Self-serving perceptions and attribu-

tions, ethnocentric beliefs, and many types of human conflict were analyzed
as passionless information-processing errors, due to basic shortcomings in in-
tuitive statistical reasoning (e.g., Borgida & Brekke, 1981; Nisbett & Ross, 1980;
Sherman, Judd, & Park, 1989). Social cognitive psychologists started to study
(what they believed to be) errors in probabilistic reasoning, such as the base-
rate fallacy, the conjunction fallacy, and overconfidence bias, and adopted the
explanatory language of Kahneman and Tversky's "heuristics," such as rep-
resentativeness and availability. Some, such as Strack (1988), even pointed to
Kahneman and Tversky's heuristics as primary evidence of the end of the "cri-
sis" of social psychology and of new, rising confidence and decisive progress
in the field.

Heuristics and Biases

The "heuristics and biases" program of Kahneman, Tversky, and others has
generated two main results concerning judgment under uncertainty: (1) a list
of so-called biases, fallacies, or errors in probabilistic reasoning, such as the
base-rate fallacy and the conjunction fallacy, and (2) explanations of these bi-
ases in terms of cognitive heuristics such as representativeness. Table 12.1
gives a taste of the conclusions drawn from this program.
Kahneman and Tversky (1982) see the study of systematic errors in proba-
bilistic reasoning, also called "cognitive illusions," as similar to that of visual
illusions. "The presence of an error of judgment is demonstrated by comparing
people's responses either with an established fact (e.g., that the two lines are
equal in length) or with an accepted rule of arithmetic, logic, or statistics"
(p. 493). Their distinction between "correct" and "erroneous" judgments un-
der uncertainty has been echoed by many social psychologists: "We follow
conventional practice by using the term 'normative' to describe the use of a
rule when there is a consensus among formal scientists that the rule is appro-
priate for the particular problem" (Nisbett & Ross, 1980, p. 13).
Social psychology is not the only area in which the "heuristics and biases"
program has made strong inroads. Experimental demonstrations of "fallacious"
judgments have entered law (e.g., Saks & Kidd, 1980), economics (e.g., Frey,
1990), management science (e.g., Bazerman, 1990), medical diagnosis (e.g.,
Casscells, Schoenberger, & Grayboys, 1978), behavioral auditing (see Shanteau,
1989), philosophy (e.g., Stich, 1990), and many other fields. There is no doubt
that understanding judgment under uncertainty is essential in all these fields.
It is the achievement of the "heuristics and biases" program to have finally
established this insight as a central topic of psychology. Earlier pioneers who
studied intuitive statistics (Hofstatter, 1939; Peirce & Jastrow, 1884; Wendt,
1966) had little impact. Even Ward Edwards and his colleagues (e.g., Edwards,
1968), who started the research from which Kahneman and Tversky's "heuris-
 HOW TO MAKE COGNITIVE ILLUSIONS DISAPPEAR 243

Table 12.1 A sample of conclusions from the heuristics-and-biases program

In making predictions and judgments under uncertainty, people do not appear to fol-
low the calculus of chance or the statistical theory of prediction. Instead, they rely on
a limited number of heuristics which sometimes yield reasonable judgments and some-
times lead to severe and systematic errors.
Daniel Kahneman &- Amos Tversky, 1973, p. 237
It appears that people lack the correct programs for many important judgmental tasks.
. . . We have not had the opportunity to evolve an intellect capable of dealing concep-
tually with uncertainty.
Paul Slavic, Baruch Fischhoff, & Sarah Lichtenstein, 1976, p. 174
The genuineness, the robustness, and the generality of the base-rate fallacy are matters
of established fact.
Maya Bar-Hillel, 1980, p. 215
The biases of framing and overconfidence just presented suggest that individuals are
generally affected by systematic deviations from rationality.
Max Bazerman & M. A. Neale, 1986, p. 317
[Overconfidence bias] has proved so robust that it is hard to acquire much insight into
the psychological processes underlying it.
Baruch Fischhoff, 1988, p. 172
. . . mental illusions should be considered the rule rather than the exception.
Richard H. Thaler, 1991, p. 4
[We are] a species that is uniformly probability-blind, from the humble janitor to the
Surgeon General. . . . We should not wait until A. Tversky and D. Kahneman receive a
Nobel prize for economics. Our self-liberation from cognitive illusions ought to start
even sooner.
Massimo Piattelli-Palmarini, 1991, p. 35

tics and biases" program emerged, had no comparable influence on cognitive

and social psychology.
Despite its influence, I will argue that the "heuristics and biases" program
is merely an important transitional stage, which must be transformed if long-
term progress is to be made. I will review some serious shortcomings of that
research program and show how they can be overcome.
In this chapter I do three things. First, I discuss the validity of the normative
yardstick that is used to define people's judgments as systematic errors in prob-
abilistic reasoning. I will argue that most so-called errors or cognitive illusions
are, contrary to the assertions in the literature, in fact not violations of prob-
ability theory. In their normative claims, Tversky and Kahneman, and social
psychologists following in their footsteps, have neglected conceptual distinc-
tions that are fundamental to probability and statistics. Second, I show that if
we pay attention to these conceptual distinctions, we can make apparently
stable "cognitive illusions" disappear, reappear, or even invert. Third, the in-
teresting fact that intuitive reasoning is highly sensitive to conceptual distinc-
tions made by statisticians (but ignored by many psychologists) leads to a re-
vised understanding of judgment under uncertainty.
244 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

Why Biases Are Not Biases

In the "heuristics and biases" program, a bias or error in probabilistic reason-

ing is defined as a systematic discrepancy between a person's judgment and a
norm. What is that norm? It is often referred to as "the normative theory of
prediction" (Kahneman & Tversky, 1973, p. 243), as the "normative principles
of statistical prediction" (Ajzen, 1977, p. 304), or simply as an "accepted rule"
of statistics. Many have understood this rhetoric to imply that there exists
precisely one "correct" answer to the cab problem, engineer-lawyer problem,
Linda problem, and other problems posed to participants—an answer sanc-
tioned by the authority of the eminent mathematicians, probabilists, and stat-
isticians of this century. The claim that all these problems have one correct
answer is crucial. If they did not have one and only one answer, it would make
little sense first to identify "errors" and "cognitive illusions" and then to use
these cognitive illusions to understand the principles of inductive reasoning,
in the way that visual illusions are used to understand the principles of normal
perception. This two-step program, identifying errors and explaining them, in
analogy to perceptual research, is the basic idea behind the heuristics-and-
biases program (Kahneman & Tversky, 1982, p. 493).
But what does the "heuristics and biases" investigation of judgment under
uncertainty have to do with probability and statistics? The short answer to this
question is: all too little. The probabilistic rules against which cognitive and
social psychologists have measured the proficiency of their participants are in
fact a highly (and, I shall argue, often misleadingly) selected sample of those
routinely used, consulted, and discussed by working probabilists and statisti-
cians. When claiming "errors" and "fallacies," cognitive and social psychol-
ogists have largely ignored conceptual and technical distinctions fundamental
to probability and statistics.
What in the heuristics-and-biases literature is called the "normative theory
of probability" or the like is in fact a very narrow kind of neo-Bayesian view
that is shared by some theoretical economists and cognitive psychologists and
to a lesser degree by practitioners in business, law, and artificial intelligence.
It is not shared by proponents of the frequentist view of probability that dom-
inates today's statistics departments, nor by proponents of many other views;
it is not even shared by all Bayesians, as I shall show shortly. By this narrow
standard of "correct" probabilistic reasoning, the most distinguished proba-
bilists and statisticians of our century—figures of the stature of Richard von
Mises and Jerzy Neyman—would be guilty of "biases" in probabilistic reason-
ing.1 Let me illustrate this point with some of the best-known demonstrations
of "fallacies."

1. Despite the widespread rhetoric of a single "normative theory of prediction," it

should be kept in mind that the problem of inductive reasoning still has no universal
solution (the "scandal of philosophy") but many competing ones. The controversies
 HOW TO MAKE COGNITIVE ILLUSIONS DISAPPEAR 245

Overconfidence Bias

Confidence in general knowledge is typically studied with questions of the

following kind:
Which city has more inhabitants?
(a) Hyderabad, (b) Islamabad
How confident are you that your answer is correct?
50% / 60% / 70% / 80% / 90% / 100%
The participant chooses what he or she believes is the correct answer and
then rates his or her confidence that the answer is correct. After many partic-
ipants answer many questions, the experimenter counts how many answers in
each of the confidence categories were actually correct. The typical finding is
that in all the cases in which participants said, "I am 100% confident that my
answer is correct," the relative frequency of correct answers was only about
80%; in all the cases in which they said, "I am 90% confident that my answer
is correct," the relative frequency of correct answers was only about 75%, and
so on (for an overview, see Lichtenstein, Fischhoff, & Phillips, 1982). This
systematic discrepancy between confidence and relative frequency is termed
"overconfidence."
Little has been achieved in explaining this "bias." A common proposal is
to explain "biases" by other, deeper mental flaws. For instance, Koriat, Lich-
tenstein, and Fischhoff (1980) proposed that the overconfidence bias is caused
by a "confirmation bias." Their explanation was this: After one alternative is
chosen, the mind searches for further information that confirms the answer

between the Fisherians, the Neyman-Pearsonians, and the Bayesians are evidence of this
unresolved rivalry. For the reader who is not familiar with the fundamental issues, two
basic themes may help introduce the debate (for more, see Hacking, 1965). The first
issue relevant for our topic is whether probability is additive (that is, satisfies the Kol-
mogorov axioms, e.g., that the probabilities of all possible events sum up to 1) or not.
The above-mentioned points of view (including that of the heuristics-and-biases pro-
gram) subscribe to additivity, whereas L. J. Cohen's (e.g., 1982) Baconian probabilities
are nonadditive (for more on nonadditive theories, see Shafer, 1976). In my opinion,
Cohen correctly criticizes the normative claims in the heuristics-and-biases program
insofar as not all uses of "probability" that refer to single events must be additive—but
this does not imply that Baconian probability is the only alternative, nor that one should
assume, as Cohen did, that all minds reason rationally (or at least are competent to do
so) in all situations. I do not deal with this issue in this chapter (but see Gigerenzer,
1991d). The second fundamental issue is whether probability theory is about relative
frequencies in the long run or (also) about single events. For instance, the question "What
is the relative frequency of women over 60 who have breast cancer?" refers to frequen-
cies, whereas "What is the probability that Ms. Young has breast cancer?" refers to a
single event. Bayesians usually assume that (additive) probability theory is about single
events, whereas frequentists hold that statements about single cases have nothing to do
with probability theory (they may be dealt with by cognitive psychology, but not by
probability theory).
246 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

given, but not for information that could falsify it. This selective information
search artificially increases confidence. The key idea in this explanation is that
the mind is not a Popperian. Despite the popularity of the confirmation bias
explanation in social psychology, there is little or no support for this hypoth-
esis in the case of confidence judgments (see Chapter 7).
As with many "cognitive illusions," overconfidence bias seems to be a ro-
bust fact waiting for a theory. This "fact" was quickly generalized to account
for human disasters of many kinds, such as deadly accidents in industry (Spet-
tell & Liebert, 1986), confidence in clinical diagnosis (Arkes, 1981), and short-
comings in management and negotiation (Bazerman, 1990) and in the legal
process (Saks & Kidd, 1980), among others.

The Normative Issue Is overconfidence bias really a "bias" in the sense of a

violation of probability theory? Let me rephrase the question: Has probability
theory been violated if one's degree of belief (confidence) in a single event (i.e.,
that a particular answer is correct) is different from the relative frequency of
correct answers one generates in the long run? The answer is "no." It is in fact
not a violation according to several interpretations of probability.
Let us look first at the now dominant school of probability: the frequentists
(the frequentist interpretation of probability has been dominant since about
1840; see Daston, 1988; Porter, 1986). Most readers of this chapter will have
been trained in the frequentist tradition and, for instance, will have been taught
that the probabilities of Type I and Type II errors are long-run frequencies of
errors in repeated experiments, not probabilities of single outcomes or hy-
potheses. For a frequentist like the mathematician Richard von Mises, the term
"probability," when it refers to a single event, "has no meaning at all for us"
(1928/1957, p. 11). For predictions of single events, as studied in present-day
overconfidence research, he put the issue in crystal-clear terms: "Our proba-
bility theory has nothing to do with questions such as: 'Is there a probability
of Germany being at some time in the future involved in a war with Liberia?' "
(p. 9). In this view, probability theory is about frequencies, not about single
events. To compare the two means comparing apples with oranges.
Even the major opponents of the frequentists—subjectivists such as Bruno
de Finetti—would not generally think of a discrepancy between confidence
and relative frequency as a "bias," albeit for different reasons. For a subjectiv-
ist, probability is about single events, but rationality is identified with the
internal consistency of subjective probabilities. As de Finetti emphasized,
"however an individual evaluates the probability of a particular event, no ex-
perience can prove him right, or wrong; nor, in general, could any conceivable
criterion give any objective sense to the distinction one would like to draw,
here, between right and wrong" (1931/1989, p. 174).
Other theories and interpretations of probability are also at odds with the
claim that overconfidence is a bias, that is, a violation of probability theory.
But I will stop here and summarize the normative issue. A discrepancy be-
tween confidence in single events and relative frequencies in the long run is
not an error or a violation of probability theory from many experts' points of
 HOW TO MAKE COGNITIVE ILLUSIONS DISAPPEAR 247

view. It only looks like it from a narrow interpretation of probability that blurs
the distinction between single events and frequencies fundamental to proba-
bility theory. (The choice of the word "overconfidence" for the discrepancy
put the "fallacy" message into the term itself.)

How to Make the Cognitive Illusion Disappear If there are any robust cognitive
biases at all, overconfidence in one's knowledge would seem to be a good
candidate. "Overconfidence is a reliable, reproducible finding" (von Winter-
feldt & Edwards, 1986, p. 539). "Can anything be done? Not much" (Edwards
& von Winterfeldt, 1986, p. 656). "Debiasing" methods, such as warning the
participants of the overconfidence phenomenon before the experiment and of-
fering them money to avoid it, have had little or no effect (Fischhoff, 1982).
Setting the normative issue straight has important consequences for under-
standing confidence judgments. Let us go back to the metaphor of the mind as
an intuitive statistician. I now take the term "statistician" to refer to a statis-
tician of the dominant school in this (and in the last) century, not one adopting
the narrow perspective some psychologists and economists have suggested.
Assume that the mind is a frequentist. Like a frequentist, the mind should be
able to distinguish between single-event confidences and frequencies in the
long run.
This view has testable consequences. Ask people for their estimated fre-
quencies of correct answers and compare them with true frequencies of correct
answers, instead of comparing the latter frequencies with confidences. We are
now comparing apples with apples. Ulrich Hoffrage, Heinz Kleinbolting, and
I carried out such experiments. Participants answered several hundred ques-
tions of the Islamabad-Hyderabad type (see above), and, in addition, estimated
their frequencies of correct answers.
Table 12.2 (top row) shows the usual "overconfidence bias" when single-
event confidences are compared with actual relative frequencies of correct
answers. In both experiments, the difference was around 13 to 15 percentage
points, which is a large discrepancy. After each set of 50 general knowledge

Table 12.2 How to make the overconfidence bias disappear

Experiment 1 Experiment 2
Difference between (n = 80) (n = 97)

Mean confidence and relative

frequency of correct answers
("overconfidence bias") +13.8 +15.4
Estimated frequency and
frequency of correct answers — 2.4 —4.2

Note: To make values for frequency and confidence judgments comparable, all frequen-
cies were transformed to relative frequencies. Values shown are differences multiplied
by a factor of 100. Positive values denote "overconfidence" (Gigerenzer, Hoffrage, &
Kleinbolting, 1991).
248 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

questions, we asked the same participants, "How many of these 50 questions

do you think you got right?" Comparing their estimated frequencies with actual
frequencies of correct answers made "overconfidence" disappear. Table 12.2
(second row) shows that estimated frequencies were practically identical with
actual frequencies, with even a small tendency toward underestimation. The
"cognitive illusion" was gone. Similar results were obtained when participants
estimated the relative frequencies of correct answers in each confidence cate-
gory. In all cases in which participants said they were "100% (90%,
80%,. . . ) confident," they estimated that, in the long run, they had a lower
percentage of answers correct, and their estimates were close to the true rela-
tive frequencies of correct answers (May, 1987, reported similar results). Elim-
inating the experimenter's normative confusion between single events and fre-
quencies made the participants' "overconfidence bias" disappear.
The general point is (i) a discrepancy between probabilities of single events
(confidences) and long-run frequencies need not be framed as an "error" and
called "overconfidence bias," and (ii) judgments need not be "explained" by
a flawed mental program at a deeper level, such as "confirmation bias." Rather,
people seem to be able intuitively to make conceptual distinctions similar to
those that professional statisticians make. How they do it can be accounted for
by the theory of "probabilistic mental models" (PMM), which explains both
confidence and frequency judgments in terms of frequentist probability cues
(Chapter 7). PMM theory is a frequentist theory of judgment and uncertainty;
it can predict overconfidence, good calibration, and underestimation within
the same participant.

Conjunction Fallacy

The original demonstration of the "conjunction fallacy" was with problems of

the following kind (Tversky & Kahneman, 1983, p. 299):

Linda is 31 years old, single, outspoken and very bright. She majored in
philosophy. As a student, she was deeply concerned with issues of dis-
crimination and social justice, and also participated in antinuclear dem-
onstrations.

Participants were asked which of two alternatives was more probable:

Linda is a bank teller (T)

Linda is a bank teller and is active in the feminist movement (T&F)

Eighty-five percent of the participants chose T&F in the Linda problem (see
Table 12.3). Tversky and Kahneman, however, argued that the "correct" an-
swer is T, because the probability of a conjunction of two events, such as T&F,
can never be greater than that of one of its constituents. They explained this
"fallacy" as induced by the representativeness heuristic. They assumed that
judgments were based on the match (similarity, representativeness) between
the description of Linda and the two alternatives T and T&F. That is, since
 HOW TO MAKE COGNITIVE ILLUSIONS DISAPPEAR 249

Table 12.3 Linda problem: How to make the conjunction

fallacy disappear
Single-event
Linda problem probability Frequency

Tver sky & Kahneman (1983)

Which is more probable? 85 —
Probability ratings 82 —
Probability ratings T* 57 —
Betting 56

Fiedler (1988)
Exp. 1 91 22
Exp. 2 83 17

Hertwig & Gigerenzer (1999)

Studies 1 and 3 83 0
Studies 2 and 4 88 13

Note: Numbers are violations (in %) of the conjunction rule. The various versions
of the Linda problem are (i) which is more probable (see text), (ii) probability
ratings on a 9-point scale, (iii) probability ratings using the alternative "Linda is
a bank teller whether or not she is active in the feminist movement" (T*) instead
of "Linda is a bank teller" (T), (iv) hypothetical betting, that is, participants were
asked "If you could win $10 by betting on an event, which of the following would
you choose to bet on?" Fiedler asked participants to rank order T, T&F, and other
alternatives with respect to their probability. In his first frequency version the
population size was always 100, in the second it varied. Hertwig and Gigerenzer
asked participants to rank order T, T&F, and F with respect to their probability,
or estimate their frequency. Tversky and Kahneman (1983, p. 309) had reported
a facilitating effect of frequency judgments for a different problem.

Linda was described as if she were a feminist and T&F contains the term "fem-
inist," people believe that T&F is more probable.
This alleged demonstration of human irrationality in the Linda problem has
been widely publicized in psychology, philosophy, economics, and beyond.
Stephen J. Gould (1992, p. 469) put the message clearly:
I am particularly fond of [the Linda] example, because I know that the
[conjunction] is least probable, yet a little homunculus in my head con-
tinues to jump up and down, shouting at me, "but she can't just be a
bank teller; read the description." . . . Why do we consistently make this
simple logical error? Tversky and Kahneman argue, correctly I think, that
our minds are not built (for whatever reason) to work by the rules of
probability.
I suggest that Gould should have had more trust in the rationality of his ho-
munculus.

The Normative Issue Is the "conjunction fallacy" a violation of probability

theory, as has been claimed in the literature? Has a person who chooses T&F
as the more probable alternative violated probability theory? Again, the answer
250 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

is "no." Choosing T&F is not a violation of probability theory, and for the same
reason given previously. For a frequentist, this problem has nothing to do with
probability theory. Participants were asked for the probability of a single event
(that Linda is a bank teller), not for frequencies. For instance, the statistician
Barnard (1979) commented thus on subjective probabilities for single events:
"If we accept it as important that a person's subjective probability assessments
should be made coherent, our reading should concentrate on the works of
Freud and perhaps Jung rather than Fisher and Neyman" (p. 171).
Note that problems that are claimed to demonstrate the "conjunction fal-
lacy" are structurally different from "confidence" problems. In the former, sub-
jective probabilities (that Linda is a bank teller or a bank teller and a feminist)
are compared with one another; in the latter, they are compared with frequen-
cies.
To summarize the normative issue, what is called the "conjunction fallacy"
looks like a violation of some subjective theories of probability, including Bay-
esian theory. It is not, however, a violation of a major view of probability, the
frequentist conception.

How to Make the Cognitive Illusion Disappear What if the mind were a fre-
quentist? If the untutored mind is as sensitive to the distinction between single
cases and frequencies as a statistician of the frequentist school is, then we
should expect dramatically different judgments if we pose the above problem
in a frequentist mode, such as the following:
There are 100 persons who fit the description above (i.e., Linda's).
How many of them are:
(a) bank tellers
(b) bank tellers and active in the feminist movement.
Participants are now asked for frequency judgments rather than for single-
event probabilities. If the mind solves the Linda problem by using a represen-
tativeness heuristic, changes in information representation should not matter
because they do not change the degree of similarity. The description of Linda
is still more representative of (or similar to) the conjunction "teller and femi-
nist" than of "teller." Participants therefore should still exhibit the conjunction
fallacy. Table 12.3, however, shows that with frequency judgments, the "con-
junction fallacy" largely disappears. The effect is dramatic, from some 80% to
90% conjunction violations in probability judgments to 10% to 20% in fre-
quency judgments, with one study even reporting 0%.
What accounts for this striking effect of frequency judgments? Hertwig and
Gigerenzer (1999) analyzed how participants understood the phrase "which is
more probable?", for instance, by asking them to paraphrase the problem to
another person who is not a native speaker of the language in which the prob-
lem was presented. The results indicate that most participants did not under-
stand "probability" in the sense of mathematical probability but as one of the
many other legitimate meanings that are listed in, for example, the Oxford
English Dictionary (e.g., meaning credibility, typicality, or that there is evi-
 HOW TO MAKE COGNITIVE ILLUSIONS DISAPPEAR 251

dence). The term frequency, unlike probability, narrows down the spectrum of
possible interpretations to meanings that follow mathematical probability.
The results in Table 12.3 are consistent with the earlier work by Inhelder
and Piaget (1969), who showed children a box containing wooden beads, most
of them brown, but a few white. They asked the children, "Are there more
wooden beads or more brown beads in this box?" By the age of eight, a majority
of children responded that there were more wooden beads, indicating that they
understand conjunctions (class inclusions). Note that Inhelder and Piaget
asked children for frequency judgments, not probability judgments.

Base-Rate Fallacy
Among all cognitive illusions, the "base-rate fallacy" has probably received
the most attention. The neglect of base rates seems in direct contradiction to
the widespread belief that judgments are unduly affected by stereotypes (Land-
man & Manis, 1983), and for this and other reasons it has generated a great
deal of interesting research on the limiting conditions for the "base-rate fal-
lacy" in attribution and judgment (e.g., Ajzen, 1977; Borgida & Brekke, 1981).
For instance, in their review, Borgida and Brekke argue for the pervasiveness
of the "base-rate fallacy" in everyday reasoning about social behavior, ask the
question "Why are people susceptible to the base-rate fallacy?" (1981, p. 65),
and present a list of conditions under which the "fallacy" is somewhat re-
duced, such as "vividness," "salience," and "causality" of base-rate informa-
tion.
My analysis is different. Again I first address the normative claims that
people's judgments are "fallacies" using two examples that reveal two different
aspects of the narrow understanding of good probabilistic reasoning in much
of this research.
The first is from Casscells, Schoenberger, and Grayboys (1978, p. 999) and
presented by Tversky and Kahneman (1982b, p. 154) to demonstrate the gen-
erality of the phenomenon:
If a test to detect a disease whose prevalence is 1/1000 has a false positive
rate of 5%, what is the chance that a person found to have a positive
result actually has the disease, assuming you know nothing about the
person's symptoms or signs?
Sixty students and staff at Harvard Medical School answered this medical
diagnosis problem. Almost half of them judged the probability that the person
actually had the disease to be 0.95 (modal answer), the average answer was
0.56, and only 18% of participants responded 0.02. The latter was considered
to be the correct answer. Note the enormous variability in judgments. Little
has been achieved in explaining how people make these judgments and why
the judgments are so strikingly variable.

The Normative Issue But do statistics and probability give one and only one
"correct" answer to that problem? The answer is again "no." And for the same
252 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

reason, as the reader will already have guessed. As in the case of confidence
and conjunction judgments, participants were asked for the probability of a
single event, that is, that "a person found to have a positive result actually has
the disease." If the mind is an intuitive statistician of the frequentist school,
such a question has no necessary connection to probability theory. Further-
more, even for a Bayesian, the medical diagnosis problem has several possible
answers. One piece of information necessary for a Bayesian calculation is miss-
ing: the test's long-run frequency of correctly diagnosing persons who have the
disease (admittedly a minor problem if we can assume a high "true positive
rate"). A more serious difficulty is that the problem does not specify whether
or not the person was randomly drawn from the population to which the base
rate refers. Clinicians, however, know that patients are usually not randomly
selected—except in screening and large survey studies—but rather "select"
themselves by exhibiting symptoms of the disease. In the absence of random
sampling, it is unclear what to do with the base rates specified. The modal
response, 0.95, would follow from applying the Bayesian principle of indif-
ference (i.e., same prior probabilities for each hypothesis), whereas the answer
0.02 would follow from using the specified base rates and assuming random
sampling. In fact, the range of actual answers corresponds quite well to the
range of possible solutions.

How to Make the Cognitive Illusion Disappear The literature overflows with
assertions of the generality and robustness of the "base-rate fallacy," such as:
"the base-rate effect appears to be a fairly robust phenomenon that often results
from automatic or unintentional cognitive processes" (Landman & Manis,
1983, p. 87); and "many (possibly most) subjects generally ignore base rates
completely" (Pollard & Evans, 1983, p. 124; see also Table 12.1). Not only are
the normative claims often simplistic and, therefore, misleading, but so too are
the robustness assertions.
What happens if we do something similar as for the "overconfidence bias"
and the "conjunction fallacy," that is, rephrase the medical diagnosis problem
in a frequency format? Cosmides and Tooby (1996) did so. They compared the
original problem (above) with a frequency format, in which the same infor-
mation was given:

One out of 1000 Americans has disease X. A test has been developed to
detect when a person has disease X. Every time the test is given to a
person who has the disease, the test comes out positive. But sometimes
the test also comes out positive when it is given to a person who is
completely healthy. Specifically, out of every 1000 people who are per-
fectly healthy, 50 of them test positive for the disease.

Imagine that we have assembled a random sample of 1000 Americans.

They were selected by a lottery. Those who conducted the lottery had
no information about the health status of any of these people. How many
people who test positive for the disease will actually have the disease?
out of
 HOW TO MAKE COGNITIVE ILLUSIONS DISAPPEAR 253

In this frequentist version of the medical diagnosis problem, both the in-
formation and the question are phrased in terms of frequencies. (In addition,
the two pieces of information missing in the original version [see above] are
supplied. In numerous other versions of the medical diagnosis problem, Cos-
mides and Tooby showed that the striking effect [see Table 12.4] on partici-
pants' reasoning is mainly due to the transition from a single-event problem
to a frequency format, and only to a lesser degree to the missing information.)
Participants were Stanford University undergraduates.
If the question was rephrased in natural frequencies, as shown above, then
the Bayesian answer of 0.02—that is, the answer "one out of 50 (or 51)"—was
given by 76% of the participants. The "base-rate fallacy" disappeared. By com-
parison, the original single-event version elicited only 12% Bayesian answers
in Cosmides and Tooby's study. Chapter 6 provides an explanation for this
effect.
Cosmides and Tooby identified one condition in which almost every par-
ticipant found the Bayesian answer of 0.02. Participants received the frequen-
tist version of the medical diagnosis problem (except that it reported a random
sample of "100 Americans" instead of "1000 Americans"), and in addition a
page with 100 squares (10 X 10). Each of these squares represented one Amer-
ican. Before the frequentist question "How many people who test positive . . . "
was put, participants were asked to (i) circle the number of people who will
have the disease and (ii) to fill in squares to represent people who will test
positive. After that, 23 out of 25 participants came up with the Bayesian an-
swer (see frequency format, pictorial, in Table 12.4).
All three examples point in the same direction: The mind acts as if it were
a frequentist; it distinguishes between single events and frequencies in the long
run—just as probabilists and statisticians do. Despite the fact that researchers
in the "heuristics and biases" program routinely ignore this distinction fun-
damental to probability theory when they claim to have identified "errors," it

Table 12.4 How to make the "base-rate fallacy" disappear: The

medical diagnosis problem
Bayesian
answers
Medical diagnosis problem N (%)

Original single-event version 60 18

(Casscells, Schoenberger, & Grayboys,
1978)
Single-event version, replication 25 12
(Cosmides & Tooby, 1996)
Frequency format 50 76
(Cosmides & Tooby, 1996)
Frequency format, pictorial 25 92
(Cosmides & Tooby, 1996)
254 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

would be foolish to label these judgments "fallacies." These results not only
point to a truly new understanding of judgment under uncertainty, but they
also seem to be relevant for teaching statistical reasoning.

Selected versus Random Sampling: More on the

Base-Rate Fallacy

Another conceptual distinction routinely used by probabilists and statisticians

is that between random sampling and selected sampling. Again, little attention
has been given to that distinction when intuitive statistical reasoning is inves-
tigated. The original medical diagnosis problem is silent about whether the
patient was randomly selected from the population. That this crucial infor-
mation is missing is not atypical. For instance, in the "Tom W." problem (Kah-
neman & Tversky, 1973), no information is given about how the personality
sketch of Tom W. was selected, whether randomly or not. The same holds for
the personality sketches of Gary W. and Barbara T. in Ajzen's (1977) base-rate
studies.
But the issue is not necessarily resolved simply by asserting random sam-
pling verbally in the problem. Consider the following famous demonstration
of base-rate neglect in which random sampling is actually mentioned. A group
of students had to solve the engineer-lawyer problem (Kahneman & Tversky,
1973, pp. 241-242):
A panel of psychologists have interviewed and administered personality
tests to 30 engineers and 70 lawyers, all successful in their respective
fields. On the basis of this information, thumbnail descriptions of the 30
engineers and 70 lawyers have been written. You will find on your forms
five descriptions, chosen at random from the 100 available descriptions.
For each description, please indicate your probability that the person
described is an engineer, on a scale from 0 to 100.
Two of these thumbnail descriptions were:
Jack is a 45-year-old man. He is married and has four children. He is
generally conservative, careful, and ambitious. He shows no interest in
political and social issues and spends most of his free time on his many
hobbies which include home carpentry, sailing, and mathematical puz-
zles. The probability that Jack is one of the 30 engineers in the sample
of 100 is %.

Dick is a 30-year-old man. He is married with no children. A man of

high ability and high motivation, he promises to be quite successful in
his field. He is well liked by his colleagues.
A second group of students received the same instructions and the same
descriptions, but were told that the base rates were 70 engineers and 30 law-
yers (as opposed to 30 engineers and 70 lawyers). Kahneman and Tversky
found that the mean response in both groups of students was for the most part
the same, and concluded that base rates were largely ignored. Their explana-
 HOW TO MAKE COGNITIVE ILLUSIONS DISAPPEAR 255

tion was that participants use a representativeness heuristic, that is, they judge
the probability by the similarity (representativeness) between a description and
their stereotype of an engineer. Kahneman and Tversky believed that their
participants were violating "one of the basic principles of statistical predic-
tion," the integration of prior probability with specific evidence by Bayes's
rule. The result was given much weight: "The failure to appreciate the rele-
vance of prior probability in the presence of specific evidence is perhaps one
of the most significant departures of intuition from the normative theory of
prediction" (p. 243).2

The Normative Issue The phrase "the normative theory of prediction," or

probability, is standard rhetoric in the "heuristics and biases" program. But
what is this normative theory? Certainly it is not frequentist—to name, for
example, only the most popular theory of probability. So let us infer what the
authors mean by "the normative theory" from what they want their partici-
pants to do. This seems to be simply to apply a formula—Bayes's rule—to the
engineer-lawyer problem. But there is more to good probabilistic reasoning
than applying formulas mechanically. There are assumptions to be checked
(see Mueser, Cowan, & Mueser, 1999). Is the structure of the problem the same
as the structure of the statistical model underlying the formula?
One important structural assumption is random sampling. If the descrip-
tions of Jack, Dick, and the others were not randomly sampled but selected,
the base rates of engineers and lawyers specified were indeed irrelevant. In
fact, the descriptions were made up and not randomly sampled from a popu-
lation with the base rates specified—although the participants were told the
contrary. Whether the single word "random" in the instruction is enough to
commit participants to this crucial structural assumption is a problem in it-
self—particularly because we cannot assume that people are familiar with sit-
uations in which profession guessing is about randomly drawn people. For
instance, both in the United States and in Germany there is a popular TV
program in which a panel of experts guesses the profession of a candidate,
who answers only "yes" or "no" to their questions. Here, the experts would
perform badly if they started out with the known base rates of professions, say
in the United States, and revised them according to Bayes's rule. The candi-
dates were selected, not randomly drawn.

2. The terms "prior probabilities" and "base rates" are frequently used interchange-
ably in the psychological literature. But these concepts are not identical. It is the prior
probabilities that are fed into Bayes's rule, and these priors may be informed by base
rates. Base rates are just one piece of information among several that a person can con-
sider relevant for making up her prior probabilities. Equating prior probabilities with
one particular kind of base-rate information would be a narrow understanding of Bayes-
ian reasoning. Such reasoning might be defensible in those situations in which one
knows very little, but not in real-life situations in which one can base judgments on rich
knowledge.
256 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

Random Sampling Increases Use of Base Rates One way to understand partic-
ipants' judgments is to assume that the engineer-lawyer problem activates ear-
lier knowledge associated with profession guessing, which can be used as an
inferential framework—a "mental model"—to solve the problem.3 But, as I
have argued, we cannot expect random sampling to be part of this mental
model. If my analysis is correct, then base-rate use can be increased if we take
care to commit the participants to the crucial property of random sampling—
that is, break apart their mental models and insert the new structural assump-
tion. In contrast, if the true explanation is that participants rely on the repre-
sentativeness heuristic, then the participants should continue to neglect base
rates.
There is a simple method of making people aware of random sampling in
the engineer-lawyer problem, which we used in a replication of the original
study (Gigerenzer, Hell, & Blank, 1988). The participants themselves drew each
description (blindly) out of an urn, unfolded the description, and gave their
probability judgments. There was no need to tell them about random sampling
because they did it themselves. This condition increased the use of base rates.
Participants' judgments were closer to Bayesian predictions than to base-rate
neglect. When we used, for comparison, the original study's version of the
crucial assumption—as a one-word assertion—neglect of base rates appeared
again (although less intensely than in Kahneman and Tversky's study).

Worthless Specific Evidence and the Base-Rate Fallacy The description of

"Dick" (see above) is a particularly interesting case. It was constructed to be
totally uninformative for distinguishing engineers from lawyers. Kahneman
and Tversky (1973) reported that the median probabilities were the same (0.50)
in both base-rate groups. That people neglect base rates even when only
"worthless specific evidence" is given has been taken by the authors to dem-
onstrate the "strength" (p. 242) of the representativeness heuristic. This strik-
ing result led to many a speculation:
The fact that the base rate is ignored even when the individuating infor-
mation is useless (for example, the target is 'ambitious' or 'well liked')
suggests that the preference for specific-level evidence is so great that
the base rate or high-level default information is not even retrieved once
the subject tries to make a prediction on the basis of the specific infor-
mation. (Holland et al., 1986, p. 218)
Such statements need to be corrected. First, if the crucial structural as-
sumption of random sampling is made clear, base rates are no longer ignored
in participants' judgments about the "uninformative" description of Dick, just
as for "informative" descriptions such as Jack's. Second, and equally striking,

3. I use the term "mental model" in a sense that goes beyond Johnson-Laird's (1983).
As in the theory of probabilistic mental models (Chapter 7), a mental model is an infer-
ential framework that generalizes the specific task to a reference class (and probability
cues defined on it) that a person knows from his or her environment.
 HOW TO MAKE COGNITIVE ILLUSIONS DISAPPEAR 257

I show that even with Kahneman and Tversky's original "verbal assertion
method," that is, a one-word assertion of random sampling, there is in fact no
support for the claim that judgments about an uninformative description are
guided by a general representativeness heuristic—contrary to assertions in the
literature.
Table 12.5 lists all studies of the uninformative description "Dick" that I
am aware of—all replications of Kahneman and Tversky's (1973) verbal asser-
tion method. The two base-rate groups were always 30% and 70% engineers.
According to Kahneman and Tversky's argument, the difference between the
two base-rate groups should approach the difference between the two base
rates, that is, 40% (or somewhat less, if the description of Dick was not per-
ceived as totally uninformative by the participants). The last column shows
their result mentioned above, a zero difference, which we (Gigerenzer, Hell, &
Blank, 1988) could closely replicate. Table 12.5 also shows, however, that sev-
eral studies found substantial mean differences up to 37%, which comes very
close to the actual difference between base-rate groups.
Seen together, the studies seem to be as inconsistent as it is possible to be:
Every result between zero difference (base-rate neglect) and the actual base-
rate difference has been obtained. This clearly contradicts the rhetoric of ro-
bustness and generality of the base-rate fallacy, such as: "Regardless of what
kind of information is presented, subjects pay virtually no attention to the base
rate in guessing the profession of the target" (Holland et al., 1986, p. 217). And
it contradicts the explanation of the so-called fallacy: the proposed general
representativeness heuristic.

Table 12.5 How to make the "base-rate fallacy" disappear: The uninformative
description "Dick" in the engineer—lawyer problem
"Dick" Mean
encountered difference
No. of first (relative between base
Study descriptions frequency) rate groups8

Gigerenzer, Hell, & Blank

(1988)b 6 1/6 1.2
Kahneman & Tversky (1973)c 5 1/5 0.0
Wells & Harvey (1978) 2 1/2 18.0
Ginosssar & Trope (1987)d 1 1 24.0
Ginossar & Trope (1980)c 1 1 31.0
Gigerenzer, Hell, & Blank (1988)e 1 1 37.0

a. Entries are {p70 (E I D)—p 30 (E I D)} X 100, where p70 (E I D) is the mean probability judgment that
"Dick" is an engineer, given the description and the 70% base rate.
b. Order of descriptions systematically varied.
c. Medians (no means reported).
d. Three descriptions were used, but "Dick" was always encountered first.
e. Separate analysis for all participants who encountered "Dick" first.
258 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

How to explain these apparently inconsistent results? Table 12.5 gives us a

clue. There is a striking correlation between the number of descriptions each
participant read and judged and the mean difference between base-rate groups.
The key variable seems to be the relative frequency with which participants
encountered "Dick" first, which is a direct function of the number of descrip-
tions. In all studies in which only Dick was used (i.e., the number of descrip-
tions was 1), or in which a separate analysis was performed for all participants
who encountered Dick first, there is a strong base-rate effect. If Dick and one
informative description (Jack) were used, as in Wells and Harvey (1978), then
the base-rate effect is in between, because of averaging across participants who
encountered Dick either before or after the informative description. Thus Table
12.5 supports the following conclusions. (1) Contrary to claims in the litera-
ture, participants did make use of the base rates if only uninformative infor-
mation ("Dick") was presented. (2) The neglect of base rates occurred only in
a specific condition, that is, when the participants had encountered one or
more informative descriptions before they judged "Dick"—in other words,
when "Dick" occurred in the second, third, fourth, or later position. (3) The
more descriptions a participant encountered, the less often "Dick" was in the
first position, and—because of averaging across positions (such as in Kahne-
man and Tversky's study)—the smaller the difference was between base-rate
groups.
This result should not occur if the intuitive statistician operates with a rep-
resentativeness heuristic. Again, an explanatory framework using mental mod-
els based on knowledge about a particular domain (here, profession guessing)
seems to be superior. If an informative description is encountered first, a men-
tal model is activated that contains probability cues for professions, such as
hobbies and political attitudes. Once the mental model is activated, the mind
uses it as an inferential framework for similar-looking problems, that is, when
"Dick" is encountered as the second or subsequent problem. Carrying over
mental models to similar problems is analogous to perceptual judgment. We
watch the first few steps and then proceed on the hypothesis that the rest are
like the first (Gregory, 1974). This practice can sometimes make us stumble,
but this kind of uncertain and "risky" inference is what makes our perceptual
apparatus superior to any computer available today.
If the uninformative description is encountered first, however, then such a
mental model is not activated, because its probability cues would not discrim-
inate, and participants fall back on the only information available, the base
rates (which are, as I argued above, not part of the mental model of profession
guessing).
To summarize: (1) There is little justification for calling participants' judg-
ments in the engineer—lawyer problem an "error" in probabilistic reasoning,
because (aside from the frequentist argument) participants were not committed
to random sampling. (2) If one lets the participants do the random drawing,
base-rate neglect disappears. (3) That participants are sensitive to the distinc-
tion between random and selected (nonrandom) drawings shows again that the
framework of so-called "heuristics and biases" is much too narrow for under-
 HOW TO MAKE COGNITIVE ILLUSIONS DISAPPEAR 259

standing judgments under uncertainty (for similar results see Ginossar & Trope
1987; Grether, 1980; Hansen & Donoghue, 1977; Wells & Harvey, 1977; but see
Nisbett & Borgida, 1975).
Note that the critical variable here is the content of a problem. There seems
to be a class of contents for which participants know from their environment
that base rates are relevant (as do birds; see Caraco, Martindale, & Whittam,
1980) or that random sampling is common (though they need not represent
these concepts explicitly), whereas in other contents this is not the case. Pro-
fession guessing seems to belong to the latter category. In contrast, predictions
of sports results, such as those of soccer games, seem to belong to the former.
For instance, we found that participants revised information about the pre-
vious performance of soccer teams (base rates) in light of new information
(half-time results) in a way that is indistinguishable from Bayesian statistics
(Gigerenzer, Hell, & Blank, 1988). Here verbal assertion of random drawing
was sufficient—there was no need for strong measures to break apart mental
models.

Heuristics

The concept of a "heuristic" has various meanings and a long history—from

Descartes's 21 heuristic rules for the direction of the mind to Duncker's heu-
ristic methods that guide the stepwise reformulation of a problem until it is
solved (Groner, Groner, & Bischof, 1983). The cognitive revolution has rein-
troduced the concept of a heuristic into psychology, in particular in the work
of Herbert Simon (1957). Because of limited information-processing abilities,
Simon argued, humans have to construct simplified models of the world. Heu-
ristics are a product of these: They are shortcuts that can produce efficient
decisions. Simon understood heuristics such as satisficing (i.e., selecting the
first option available that meets an aspiration level) as adaptive strategies in a
complex environment, in which alternatives for action are not given but must
be sought out.
In the 1970s, Kahneman and Tversky borrowed the term "heuristic" from
artificial intelligence to explain "errors" in probabilistic reasoning: "People
rely on a limited number of heuristic principles which reduce the complex
tasks of assessing probabilities and predicting values to simpler judgmental
operations. In general, these heuristics are quite useful, but sometimes they
lead to severe and systematic errors" (Tversky & Kahneman, 1974, p. 1124).
Although they repeatedly asserted that these heuristics are useful, almost all
of their work focused on how they lead to "errors." The three heuristics pro-
posed in the early 1970s—representativeness, availability, and anchoring and
adjustment—were a first, promising step to connect the rather atheoretical
Bayesian research of the 1960s with cognitive theory. But in the 30 years of
"heuristics and biases" research since then, a lack of theoretical progress is
possibly the most striking result. The absence of a general theory or even of
specific models of underlying cognitive processes has been repeatedly criti-
260 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

cized (e.g., Jungermann, 1983; Wallsten, 1983), but to no avail. Why is this? I
believe that particular features of the use of the term "heuristic" have led to
the present conceptual dead end, and more research in a cul-de-sac will not
help. In my opinion, these features are the following.

The Function of Heuristics

In artificial intelligence research one hopes that heuristics can make computers
smart; in the "heuristics and biases" program one hopes that heuristics can
tell why humans are not smart. The fundamental problem with the latter is
that most "errors" in probabilistic reasoning that one wants to explain by heu-
ristics are in fact not errors, as I have argued above. Thus heuristics are meant
to explain what does not exist. Rather than explaining a deviation between
human judgment and allegedly "correct" probabilistic reasoning, future re-
search has to get rid of simplistic norms that evaluate human judgment instead
of explaining it.
Simon, and earlier Egon Brunswik, has emphasized that cognitive functions
are adaptations to a given environment and that we have to study the structure
of environments in order to infer the constraints they impose on reasoning.
Heuristics such as representativeness have little to say about how the mind
adapts to the structure of a given environment.

Redescription as a Substitute for Theorizing

Several of the explanations using heuristics are hardly more than redescrip-
tions of the phenomena reported. Take, for instance, the explanation of base-
rate neglect in the engineer—lawyer problem (and similar base-rate problems)
by the representativeness heuristic. Representativeness here means the per-
ceived similarity between a personality description and the participants'
stereotype of an engineer. In the vocabulary of Bayes's rule, this similarity is
a likelihood: that is, the probability of a description given that the person is
an engineer. Now we can see that Bayes's rule, in particular its concepts of
base rates (prior probabilities) and likelihoods, provides the vocabulary for
both the phenomenon and its purported explanation. The phenomenon is ne-
glect of base rates and use of likelihoods. The "explanation" is that participants
use representativeness (likelihoods) and do not use base rates. What is called
a representativeness heuristic here is nothing more than a redescription of the
phenomenon (Gigerenzer & Murray, 1987, pp. 153-155).

Heuristics Are Largely Undefined Concepts

Representativeness means similarity. Although there are numerous specific

models of similarity (including Tversky, 1977), the relationship between the
representativeness heuristic and these models has never been worked out.
Fiedler (1983), for instance, has analyzed the theoretical weakness of explain-
ing estimated frequencies of events by the availability heuristic. All three heu-
 HOW TO MAKE COGNITIVE ILLUSIONS DISAPPEAR 261

ristics, representativeness, availability, and anchoring and adjustment, are

largely undefined concepts and can post hoc be used to explain almost every-
thing. After all, what is similar to what (representativeness), what comes into
your mind (availability), and what comes first (anchoring) have long been
known to be important principles of the mind.

More Undefined Concepts, Less Theory

Instead of giving up the program of explaining deviations of human judgment

from simplistic norms by means of redescription and largely undefined heu-
ristics, the last 25 years have witnessed the effort to keep that program going
and to add further undefined concepts such as "causal base rates" and "viv-
idness" to account for contradictory results (for an analysis of the "causal base
rate" concept, see Gigerenzer & Murray, 1987, pp. 157-162). Heuristics such
as representativeness are by now riddled with exceptions, but all this tinkering
has not given us much purchase in understanding judgment under uncertainty.

Beyond Heuristics and Biases

I have argued that what have been widely accepted to be the "normative prin-
ciples of statistical prediction" (e.g., Ajzen, 1977, p. 304), against which hu-
man judgment has been evaluated as "fallacious," are a caricature of the pres-
ent state of probability theory and statistics. I have shown that several so-called
fallacies are in fact not violations of probability theory. Conceptual distinctions
routinely used by probabilists and statisticians were just as routinely ignored
in the normative claims of "fallacies." Most strikingly, in the experimental
research reviewed, "fallacies" and "cognitive illusions" tend to disappear if
we pay attention to these fundamental distinctions. I am certainly not the first
to criticize the notion of "robust fallacies." The only novelty in my research
is that the variables that bring "cognitive illusions" under experimental control
are those important from the viewpoint of probability and statistics (as op-
posed to, say, whether participants were given more or less "vivid" or "caus-
ally relevant" information).
Together, these results point to several ways to develop an understanding
of judgment under uncertainty that goes beyond the narrow notion of a "bias"
and the largely undefined notion of a "heuristic."

Use Different Statistical Models as Competing

Explanatory Models

The existence of different statistical models of inference is a rich resource for

developing theories about intuitive inference. This resource has been rarely
touched, possibly because of the misleading normative view that statistics
speaks with one voice.
262 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

For instance, despite the quantity of empirical data that has been gathered
on the cab problem, the lack of a theory of the cognitive processes involved
in solving it is possibly the most striking result. Tversky and Kahneman
claimed that the cab problem has one "correct answer" (1980, p. 62). They
attempted to explain the extent to which people's judgments deviated from
that "norm" by largely undefined terms such as "causal base rates." But sta-
tistics gives several interesting answers to the cab problem, rather than just
one "correct" answer (e.g., Birnbaum, 1983; Gigerenzer, 1998c; Levi, 1983). If
progress is to be made and people's cognitive processes are to be understood,
one should no longer try to explain the difference between people's judgments
and Tversky and Kahneman's "normative" Bayesian calculations. People's
judgments have to be explained. Statistical theories can provide highly inter-
esting models of these judgments. The only theoretically rich account of the
cognitive processes involved in solving the cab problem (or similar "eyewit-
ness testimony" problems) was in fact derived from a frequentist framework:
Birnbaum (1983) combined Neyman-Pearson theory with psychological mod-
els of judgments such as range-frequency theory.
Future research should use competing statistical theories as competing ex-
planatory models, rather than pretending that statistics speaks with one voice
(see also Cohen, 1982; Wallendael & Hastie, 1990).

Explore the Metaphor of the Mind as a Frequentist

I reported earlier the striking effect of participants' judging frequencies rather

than probabilities for single events. These results suggest that the mind distin-
guishes between frequencies and other meanings of probability, just as a stat-
istician of the frequentist school does. Because "cognitive illusions" tend to
disappear in frequency judgments, it is tempting to think of the intuitive sta-
tistics of the mind as frequentist statistics.
Processing of frequencies seems to be fairly automatic, like encoding of time
and space (e.g., Hasher & Zacks, 1979)—whereas probabilities are in evolu-
tionary terms recent tools of the mind that seem to be processed less auto-
matically. The theory of probabilistic mental models (Chapter 7) seems to be
the first frequentist theory of confidence judgments that integrates Brunswik's
frequency-learning view with the notion of mental models. The general theo-
retical point is that both single-case and frequency judgments are explained
by learned frequencies (the probability cues), albeit by frequencies that relate
to different reference classes and different networks of cues—in short, to dif-
ferent mental models.

Intuitive Statisticians Need to Check the Structure

of the Environment

Good judgment under uncertainty is more than mechanically applying a for-

mula, such as Bayes's rule, to a real-world problem. The intuitive statistician,
like his professional counterpart, must first check the structure of the environ-
 HOW TO MAKE COGNITIVE ILLUSIONS DISAPPEAR 263

ment (or of a problem) in order to decide whether to apply a statistical algo-

rithm at all, and if so, which (see Gigerenzer & Murray, 1987, pp. 162-174).
There is no good (applied) probabilistic reasoning that ignores the structure of
the environment and mechanically uses only one (usually mathematically con-
venient) algorithm. I illustrate this point with a thought experiment by Nisbett
and Ross (1980, p. 15), which I have shortened and slightly changed here (in
respects unimportant to my argument).
(i) You wish to buy a new car. Today you must choose between two
alternatives: to purchase either a Volvo or a Saab. You use only one cri-
terion for that choice, the car's life expectancy. You have information
from Consumer Reports that in a sample of several hundred cars the
Volvo has the better record. Just yesterday a neighbor told you that his
new Volvo broke down. Which car do you buy?
Nisbett and Ross comment that after the neighbor's information "the num-
ber of Volvo-owners has increased from several hundred to several hundred
and one" and that the Volvo's record "perhaps should be changed by an iota"
(p. 15). The moral of their thought experiment is that good probabilistic rea-
soning is applying an algorithm (here, updating of base rates) to the world.
There is some truth to this message of resisting the temptation of the vivid and
personal, but that is only half the story. Good intuitive statistics is more than
calm calculation; first and foremost, the structure of the environment has to
be examined. I will now vary the content of Nisbett and Ross's thought exper-
iment to make the point intuitively immediate. Here is the same problem, but
with a different content (Gigerenzer, 1990):
(ii) You live in a jungle. Today you must choose between two alterna-
tives: to let your child swim in the river, or to let it climb trees instead.
You use only one criterion for that choice, your child's life expectancy.
You have information that in the last 100 years there was only one ac-
cident in the river, in which a child was eaten by a crocodile, whereas
a dozen children have been killed by falling from trees. Just yesterday
your neighbor told you that her child was eaten by a crocodile. Where
do you send your child?
If good probabilistic reasoning means applying the same algorithm again
and again, the neighbor's testimony should make no difference. The base rates
would be updated by the testimony from one to two cases in 100 years, and
by this reasoning one would send the child into the river. The mind of a parent,
however, might use the new information to reject the updating algorithm in-
stead of inserting the new information into the algorithm. A parent may sus-
pect that the small river world has changed—crocodiles may now inhabit the
river.
Why do we have different intuitions for the Volvo and the crocodile prob-
lems? In the Volvo problem, the prospective buyer may assume that the Volvo
world is stable and that the important event (good or bad Volvo) can be con-
sidered as an independent random drawing from the same reference class. In
the crocodile problem, the parents may assume that the river world has
264 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

changed and that the important event (being eaten or not) can no longer be
considered as an independent random drawing from the same reference class.
Updating "old" base rates may be fatal for the child.
The question of whether some part of the world is stable enough to use
statistics has been posed by probabilists and statisticians since the inception
of probability theory in the mid-seventeenth century—and the answers have
varied and will vary, as is well documented by the history of insurance (Das-
ton, 1987). Like the underwriter, the layperson has to check structural as-
sumptions before entering into calculations. For instance, the following struc-
tural assumptions are all relevant for the successful application of Bayes's rule:
independence of successive drawings, random sampling, an exhaustive and
mutually exclusive set of hypotheses, and independence between prior prob-
abilities and likelihoods.
How can the intuitive statistician judge whether these assumptions hold?
One possibility is that the mind generalizes the specific content to a broader
mental model that uses implicit domain-dependent knowledge about these
structural assumptions. If so, then the content of problems is of central im-
portance for understanding judgment—it embodies implicit knowledge about
the structure of an environment.

The Surplus Structure of the Environment

Analyzing the environment (problem) using structural properties of a given

statistical model is one way to understand its structure. But natural environ-
ments often have surplus structure, that is, a structure that goes beyond prior
probabilities and likelihoods (the Bayesian structure) or entailment and con-
tradiction (the structure of binary prepositional logic). Surplus structure in-
cludes space and time (Bjorkman, 1984), cheating options, perspective, and
social contracts (Cosmides, 1989), among others. Surplus structure is the rea-
son that the notion of "structural isomorphs" has only limited value.
The idea of studying inductive reasoning using structural isomorphs (i.e.,
use a particular statistical model or formal principle and construct problems
that all have the same formal structure but different contents) is implicit in
much research on reasoning; it postulates that if two problems have different
contents, but the same formal structure (say, Bayesian probability-revision
structure), then judgments should be the same. But the structure of natural
environments is usually richer than what Bayes's rule has to offer, and two
structural isomorphs may differ on relevant surplus structure. If we understand
reasoning as an adaptation to the environment, then it should be sensitive to
surplus structure.
One way to deal with this is to devise theories that combine statistical the-
ory and psychological principles—just as the most distinguished statisticians
of this century, R. A. Fisher, J. Neyman, and E. S. Pearson, emphasized that
good statistical reasoning always consists of mathematics and personal judg-
ment. Birnbaum (1983) gave several examples of how the Neyman-Pearson
theory can be combined with psychological principles to give a theoretically
 HOW TO MAKE COGNITIVE ILLUSIONS DISAPPEAR 265

rich account of intuitive inference. Developing such integrated models is a

challenging task for future research on judgments under uncertainty.

The Social Context of Judgment and Decision

Judgment under uncertainty occurs in a social environment in which there are

other "players" who make a person's response more or less rational. Here is
an anecdote to illustrate this point.
A small town in Wales has a village idiot. He once was offered the choice
between a pound and a shilling, and he took the shilling. People came
from everywhere to witness this phenomenon. They repeatedly offered
him a choice between a pound and a shilling. He always took the shil-
ling.
Seen as a single choice (and by all monotone utility functions), this choice
would seem irrational. Seen in its social context, in which a surprising choice
increases the probability of getting to choose again, this behavior looks differ-
ent.
The following are several aspects of the social context of judgment and
decision that have been explored recently. First, human judgment seems to be
domain specific rather than guided by some general mental logic. In particular,
reasoning about social contracts seems to have its own laws. The striking
changes of judgment depending on people's perspective and cheating options
in a social contract were shown by Cosmides (1989) and Gigerenzer and Hug
(1992). Second, the role of conversational principles in social interactions,
such as that participants assume the experimenter's contribution will be co-
operative (Adler, 1984; Grice, 1975), has sometimes been acknowledged by,
but never been integrated into, the judgment under uncertainty literature.
Third, humans share knowledge and decisions, and sharing imposes con-
straints on information processing and judgment as postulated by shareability
theory (Freyd, 1983). Fourth, research on group decision making and judg-
ments negotiated by two or more people is still largely disconnected from
"individualistic" social cognition research (see Scholz, 1983).

Conclusion

A key metaphor for understanding inductive reasoning is probability theory.

Since its origins in the mid-seventeenth century and throughout the Enlight-
enment, probability theory was viewed as a mathematical codification of ra-
tionality. In Pierre Laplace's famous phrase: probability theory is "only good
sense reduced to calculus" (1814/1951, p. 196). When there was a striking
discrepancy between the judgment of reasonable people and what probability
theory dictated—as with the famous St. Petersburg paradox—then the mathe-
maticians went back to the blackboard and changed the equations (Daston,
1980).
266 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

Those good old days have gone, although the eighteenth-century link be-
tween probability and rationality is back in vogue in cognitive and social psy-
chology. If, in studies on social cognition, researchers find a discrepancy be-
tween human judgment and what probability theory seems to dictate, the
blame is now put on the human mind, not on the statistical model.
I have used classical demonstrations of overconfidence bias, conjunction
fallacy, and base-rate neglect to show that what have been called "errors" in
probabilistic reasoning are in fact not violations of probability theory. They
only look so from a narrow understanding of good probabilistic reasoning that
ignores conceptual distinctions fundamental to probability and statistics.
These so-called cognitive illusions largely disappear when one pays attention
to these conceptual distinctions. The intuitive statistician seems to be highly
sensitive to them—a result unexpected from the view that "mental illusions
should be considered the rule rather than the exception" (see Table 12.1).
Why do cognitive illusions largely disappear? The examples in this chapter
have illustrated three reasons:
1. Polysemy: not all probabilities are mathematical probabilities. Asking
a frequency as opposed to a probability question can reduce the mul-
tiple meanings (polysemy) of the English terms "probable" and
"likely." Frequency questions clarify that the question is actually
about mathematical probability and not about one of the other legit-
imate meanings (see the Oxford English Dictionary), which are often
suggested by the cover story of a problem. Reducing polysemy seems
to be the major reason the conjunction fallacy in the Linda problem
largely disappears (Hertwig & Gigerenzer, 1999).
2. A mathematical probability refers to a reference class, which may
differ depending on the task. Asking a frequency (as opposed to a
probability) question can systematicaly cue different reference classes
(and therefore, different probabilistic mental models). Changing ref-
erence classes seems to be the reason overconfidence bias appears in
probability judgments and disappears in frequency judgments (Chap-
ter 7).
3. Natural frequencies facilitate Bayesian reasoning. When information
is represented in natural frequencies rather than in conditional prob-
abilities (or relative frequencies), Bayesian computations become sim-
pler. Using natural frequencies is a powerful tool to reduce people's
mental confusion and foster Bayesian reasoning (Chaper 6).
This is not to say that frequencies always improve judgment. For instance,
the theory of probabilistic mental models specifies conditions under which
frequency judgments systematically underestimate actual frequencies, and
Chapter 6 explains why natural frequencies but not other kinds of frequencies
facilitate Bayesian reasoning. The question is not whether or not, or how often,
"cognitive illusions" disappear, but why. We need precise models of heuristics
that make surprising (and falsifiable) predictions, not vague terms that, post
hoc, explain everything and nothing. Future progress will be in understanding,
not debunking, human thinking.
 13
The Superego, the Ego, and the Id
in Statistical Reasoning

Piaget worked out his logical theory of cognitive development, Kohler the Ge-
stalt laws of perception, Pavlov the principles of classical conditioning, Skin-
ner those of operant conditioning, and Bartlett his theory of remembering and
schemata—all without rejecting null hypotheses. But by the time I took my
first course in psychology at the University of Munich in 1969, null hypothesis
tests were presented as the indispensable tool, as the sine qua non of scientific
research. Post-World War II German psychology mimicked a revolution of re-
search practice that had occurred between 1940 and 1955 in American psy-
chology.
What I learned in my courses and textbooks about the logic of scientific
inference was not without a touch of moralizing, a scientific version of the Ten
Commandments: Thou shalt not draw inferences from a nonsignificant result.
Thou shalt always specify the level of significance before the experiment; those
who specify it afterward (by rounding up obtained p values) are cheating. Thou
shalt always design thy experiments so that thou canst perform significance
testing.

The Inference Revolution

What happened between the time of Piaget, Kohler, Pavlov, Skinner, and Bart-
lett and the time I was trained? In Kendall's (1942) words, statisticians "have
already overrun every branch of science with a rapidity of conquest rivalled
only by Attila, Mohammed, and the Colorado beetle" (p. 69).
What has been termed the probabilistic revolution in science (Gigerenzer et
al., 1989) reveals how profoundly our understanding of nature changed when
concepts such as chance and probability were introduced as fundamental the-
oretical concepts. The work of Mendel in genetics, that of Maxwell and Boltz-
mann on statistical mechanics, and the quantum mechanics of Schrodinger
and Heisenberg that built indeterminism into its very model of nature are key
examples of that revolution in thought.

267
268 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

Psychology did not resist the probabilistic revolution, and psychologists in

turn actively contributed to the growth of statistics. But psychology is none-
theless a peculiar case. In psychology and in other social sciences, probability
and statistics were typically not used to revise the understanding of our subject
matter from a deterministic to some probabilistic view (as in physics, genetics,
or evolutionary biology) but rather to mechanize the experimenters' infer-
ences—in particular, their inferences from data to hypothesis. Of course, there
have been several attempts to revise our theories as well—for example, to
transform Piaget's logical determinism into a more Darwinian view, in which
variability and irregularity are seen as the motor of evolution rather than as
an annoyance (Gruber, 1977; Gruber & Voneche, 1977) or to transform Skin-
ner's theory into a probabilistic learning theory (Estes, 1959). But the real,
enduring transformation came with statistical inference, which became insti-
tutionalized and used in a dogmatic and mechanized way. This use of statis-
tical theory contrasts sharply with physics, in which statistics and probability
are indispensable in theories about nature, whereas mechanized statistical in-
ference such as null hypothesis testing is almost unknown.
So what happened with psychology? David Murray and I described the
striking change in research practice and named it the inference revolution in
psychology (Gigerenzer & Murray, 1987). It happened between approximately
1940 and 1955 in the United States and led to the institutionalization of one
brand of inferential statistics as the method of scientific inference in university
curricula, textbooks, and the editorial policies of major journals.
The ground for the inference revolution was prepared by a dramatic shift
in experimental practice. During the 1920s, 1930s, and 1940s, the established
tradition of experimenting with single participants—from Wundt to Pavlov—
was replaced in the United States by the treatment group experiment, in which
group means are compared. For instance, between 1915 and 1950, the per-
centage of empirical studies reporting only group data in the American Journal
of Psychology rose from 25% to 80%, and the reporting of only individual
data decreased from 70% to 17% (Danziger, 1990). Danziger argued that this
shift was in part due to the pressure felt by American academic psychologists
to legitimize their work through showing its practical utility. The Wundtian
type of experiment was useless to educational administrators, the largest market
for psychological products. The treatment group experiment, however, ap-
peared to fit their needs exactly, for example, by allowing them to compare
mean performance in two classrooms that were using different instruction
methods. After this change in experimental practice, null hypothesis testing
of group means appeared to be tailor-made to the new unit of research, the
group aggregate. Consistent with Danziger's argument, the institutionalization
of both the treatment group and null hypothesis testing spread from the applied
fields to the laboratories (Lovie, 1979). The contrast with Germany is inform-
ative. German academic psychologists of the early 20th century had to legit-
imize their work before a different tribunal, the values of a well-entrenched
intellectual elite (Danziger, 1990). In contrast to the United States, the German
educational system, driven by tradition rather than by experimentation, pro-
 THE SUPEREGO, THE EGO, AND THE ID IN STATISTICAL REASONING 269

vided only a limited market for psychologists. No comparable shift in exper-

imental practice happened in German psychology. It was only after World War
II that a new generation of German psychologists began to assimilate the meth-
odological imperatives imported from their colleagues in the United States.
The figures are telling. Before 1940, null hypothesis testing using analysis
of variance or t tests was practically nonexistent: Rucci and Tweney (1980)
found only 17 articles in all from 1934 through 1940. By 1955, more than 80%
of the empirical articles in four leading journals used null hypothesis testing
(Sterling, 1959). Today, the figure is close to 100%. By the early 1950s, half of
the psychology departments in leading U.S. universities had made inferential
statistics a graduate program requirement (Rucci & Tweney, 1980). Editors and
experimenters began to measure the quality of research by the level of signif-
icance obtained. For instance, in 1962, the editor of the Journal of Experimen-
tal Psychology, A. W. Melton (1962, pp. 553-554), stated his criteria for ac-
cepting articles. In brief, if the null hypothesis was rejected at the .05 level but
not at the .01 level, there was a "strong reluctance" to publish the results,
whereas findings significant at the .01 level deserved a place in the journal.
The Publication Manual of the American Psychological Association (APA,
1974) prescribed how to report the results of significance tests (but did not
mention other statistical methods) and used, as Melton did, the label negative
results synonymously with "not having rejected the null" and the label posi-
tive results with "having rejected the null."
It is likely that Piaget's, Kohler's, Bartlett's, Pavlov's, and Skinner's experi-
mental work would have been rejected for publication under such editorial
policies—these men did not set up null hypotheses and try to refute them.
Some of them were actively hostile toward institutionalized statistics. For his
part, Skinner (1972) disliked the intimate link Fisher had established between
statistics and the design of experiments: "What the statistician means by the
design of experiments is design which yields the kind of data to which his
techniques are applicable" (p. 122). And, "They have taught statistics in lieu
of scientific method" (p. 319). Skinner continued to investigate one or a few
pigeons under well-controlled conditions, rather than run 20 or more pigeons
under inevitably less well-controlled conditions to obtain a precise estimate
for the error variance. In fact, the Skinnerians were forced to found a new
journal, the Journal of the Experimental Analysis of Behavior, in order to pub-
lish their kind of experiments (Skinner, 1984, p. 138). Their focus was on ex-
perimental control, that is, on minimizing error beforehand, rather than on
large samples, that is, on measuring error after the fact.
This is not an isolated case, nor one peculiar to behaviorists. The Journal of
Mathematical Psychology is another. One of the reasons for launching this new
journal was again to escape the editors' pressure to perform institutionalized
null hypothesis testing.1 One of its founders, Luce (1988), called the institution-

1. R. Duncan Luce, personal communication, April 4, 1990. See also Luce's (1989)
autobiography, p. 270 and pp. 281-282.
270 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

alized practice a "wrongheaded view about what constituted scientific pro-

gress" and "mindless hypothesis testing in lieu of doing good research: measur-
ing effects, constructing substantive theories of some depth, and developing
probability models and statistical procedures suited to these theories" (p. 582).
Who is to blame for the present state of mindless hypothesis testing? Fisher
was blamed by Skinner, as well as by Meehl: "Sir Ronald has befuddled us, mes-
merized us, and led us down the primrose path. I believe that the almost univer-
sal reliance on merely refuting the null hypothesis . . . is ... one of the worst
things [that] ever happened in the history of psychology" (Meehl, 1978, p. 817).
I share the sentiments expressed by Luce and Meehl. But to blame Fisher,
as Meehl and Skinner did, gives us at best a spurious understanding of the
inference revolution. Fisher declared that a significance test of a null hypoth-
esis is only a "weak" argument. That is, it is applicable only in those cases in
which we have very little knowledge or none at all. For Fisher, significance
testing was the most primitive type of argument in a hierarchy of possible
statistical analyses (see Gigerenzer et al., 1989, chap. 3). In this chapter I argue
the following points:
1. What has become institutionalized as inferential statistics in psy-
chology is not Fisherian statistics. It is an incoherent mishmash of
some of Fisher's ideas on the one hand and some of the ideas of Ney-
man and E. S. Pearson on the other. I refer to this blend as the "hybrid
logic" of statistical inference. Fisher, Neyman, and Pearson would all
have rejected it, although for different reasons.
2. The institutionalized hybrid carries the message that statistics is sta-
tistics is statistics, that is, that statistics is a single integrated structure
that speaks with a single authoritative voice. This entails the claim
that the problem of inductive inference in fact has an algorithmic
answer (i.e., the hybrid logic) that works for all contents and contexts.
Both claims are wrong, and it is time to go beyond this institution-
alized illusion. We must write new textbooks and change editorial
practices. Students and researchers should be exposed to different
approaches (not one) to inductive inference and be trained to use
these in a constructive (not mechanical) way. A free market of several
good ideas is better than a state monopoly for a single confused idea.

The "Parents" and Their Conflicts

To understand the structure of the hybrid logic that has been taught in psy-
chology for some 50 years, I briefly sketch those ideas of Fisher on the one
hand and Neyman and Pearson on the other that are relevant to understanding
the hybrid structure of the logic of inference.
Fisher's first book, Statistical Methods for Research Workers, published in
1925, was successful in introducing biologists and agronomists to the new
techniques. However, it had the agricultural odor of issues like the weight of
pigs and the effect of manure, and, such alien topics aside, it was technically
far too difficult to be understood by most psychologists.
 THE SUPEREGO, THE EGO, AND THE ID IN STATISTICAL REASONING 271

Fisher's second statistical book, The Design of Experiments, first published

in 1935, was most influential on psychology. At the very beginning of this
book, Fisher rejected the theory of inverse probability (Bayesian theory) and
congratulated the Reverend Bayes for having been so critical of his own theory
as to withhold it from publication (Bayes's treatise was published posthu-
mously in 1763). Bayes's rule is attractive for researchers because it allows one
to calculate the probability p(H\ D) of a hypothesis H given some data D, also
known as inverse probability. A frequentist theory, such as Fisher's null hy-
pothesis testing or Neyman-Pearson theory, however, does not. It deals with
the probabilities p(D I H) of some data D given a hypothesis H, such as the
level of significance.
Fisher was not satisfied with an approach to inductive inference based on
Bayes's rule. The use of Bayes's rule presupposes that a prior probability dis-
tribution over the set of possible hypotheses is available. For a frequentist,
such as Fisher, this prior distribution must theoretically be verifiable by actual
frequencies, that is, by sampling from its reference set. These cases are rare.
But if we are ignorant and have no a priori distributional information, then
every researcher can express that ignorance in different numbers leading, for
Fisher, to an unacceptable subjectivism. As we shall see, however, Fisher
wanted to both reject the Bayesian cake and eat it, too.
Fisher proposed several alternative tools for inductive inference. In The
Design of Experiments (1935), he started with null hypothesis testing, also
known as significance testing, and he gave that tool the most space in his book.
It eventually became the backbone of institutionalized statistics in psychology.
In a test of significance, one confronts a null hypothesis with observations, to
find out whether the observations deviate far enough from the null hypothesis
to conclude that the null is implausible. The specific techniques of null hy-
pothesis testing, such as the t test (devised by Cosset, using the pseudonym
"Student," in 1908) or the Ftest (Ffor Fisher, e.g., in analysis of variance) are
so widely used that they may be the lowest common denominator of what
psychologists today do and know.
The topic of this chapter is the logic of inference rather than specific tech-
niques. Just as with Bayes's rule, the problems we encounter do not concern
the formula—the rule is a simple consequence of the definition of conditional
probability. The problems arise with its application to inductive inference in
science. To what aspect of inductive inference does a particular algorithm, or
technique, refer? What do the calculations mean? These are questions that
pertain to what I call the logic of inference.
Concerning my account of Fisher's logic of significance testing, one thing
must be said in advance: Fisher's writings and polemics had a remarkably
elusive quality, and people have read his work quite differently. During
Fisher's long and acrimonious controversy with Neyman and Pearson, which
lasted from the 1930s to his death in 1962, he changed, and sometimes even
reversed, parts of his logic of inference. Thus the following brief account of
Fisher's logic of inference represents one possible reading (for a more detailed
analysis, see Gigerenzer et al., 1989, chap. 3).
272 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

How Do We Determine the Level of Significance?

In The Design, Fisher suggested that we think of the level of significance as a

convention: "It is usual and convenient for experimenters to take 5 per cent
as a standard level of significance, in the sense that they are prepared to ignore
all results which fail to reach this standard" (1935, p. 13). Fisher's assertion
that 5% (in some cases, 1%) is a convention that is adopted by all experi-
menters and in all experiments and that nonsignificant results are to be ignored
became part of the institutionalized hybrid logic.
But Fisher had second thoughts, which he stated most clearly in the mid-
1950s. These did not become part of the hybrid logic. One of the reasons for
that revision was his controversy with Neyman and Pearson and Neyman's
(e.g., 1950) insistence that one has to specify the level of significance (which
is denoted as a in Neyman-Pearson theory) before the experiment, in order to
be able to interpret it as a long-run frequency of error. Neyman and Pearson
took the frequentist position more seriously than Fisher. They argued that the
meaning of a level of significance such as 5% is the following: If the null
hypothesis is correct and if the experiment is repeated many times, then the
experimenter will wrongly reject the null in 5% of the cases. To reject the null
if it is correct is called an error of the first kind (Type I error) in Neyman-
Pearson theory, and its probability is called alpha (a). In his last book, Statis-
tical Methods and Scientific Inference (1956), Fisher ridiculed this definition
as "absurdly academic, for in fact no scientific worker has a fixed level of
significance at which from year to year, and in all circumstances, he rejects
hypotheses; he rather gives his mind to each particular case in the light of his
evidence and his ideas" (p. 42). Fisher rejected the Neyman-Pearson logic of
repeated experiments (repeated random sampling from the same population)
and thereby rejected his earlier proposal to have a conventional standard level
of significance, such as .05 or .01. What researchers should do, according to
Fisher's second thoughts, is to publish the exact level of significance, say, p —
.03 (not p < .05), and communicate this result to their fellow research workers.
This means that the level of significance is determined after the experiment,
not, as Neyman and Pearson proposed, before the experiment.
Thus the phrase "level of significance" has three meanings: (a) the standard
level of significance, a conventional standard for all researchers (early Fisher),
(b) the exact level of significance, a communication to research fellows, deter-
mined after the experiment (late Fisher), and (c) the alpha level, the relative
frequency of Type I errors in the long run, to be decided on using cost-benefit
considerations before the experiment (Neyman & Pearson). The basic differ-
ence is this: For Fisher, the exact level of significance is a property of the data
(i.e., a relation between a body of data and a theory); for Neyman and Pearson,
alpha is a property of the test, not of the data. Level of significance and alpha
are not the same thing.
Neyman and Pearson thought their straightforward long-run frequentist in-
terpretation of the significance test—and the associated concepts of power and
of stating two statistical hypotheses (rather than only one, the null)—would
 THE SUPEREGO, THE EGO, AND THE ID IN STATISTICAL REASONING 273

be an improvement on Fisher's theory and make it more consistent. Fisher

disagreed. Whereas Neyman and Pearson thought of mathematical and con-
ceptual consistency, Fisher thought of ideological differences. He accused Ney-
man, Pearson, and their followers of confusing technology with knowledge:
Their focus on Type I and Type II errors, on cost-benefit considerations that
determine the balance between the two, and on repeated sampling from the
same population has little to do with scientific practice, but it is characteristic
for quality control and acceptance procedures in manufacturing. Fisher (1955,
p. 70) compared the Neyman—Pearsonians to the Soviets, their five-year plans,
and their ideal that "pure science can and should be geared to technological
performance." He also compared them to Americans, who confuse the process
of attaining knowledge with speeding up production or saving money. (Inci-
dentally, Neyman was born in Russia and went to Berkeley, California, after
Fisher made it difficult for him to stay on at University College in London).

What Does a Significant Result Mean?

The basic differences are these: Fisher attached an epistemic interpretation to
a significant result, which referred to a particular experiment. Neyman rejected
this view as inconsistent and attached a behavioral meaning to a significant
result that did not refer to a particular experiment but to repeated experiments.
(Pearson found himself somewhere in between.)
In The Design, Fisher talked about how to "disprove" a null hypothesis
(e.g., 1935, pp. 16—17). Whatever phraseology he used, he always held that a
significant result affects our confidence or degree of belief that the null hy-
pothesis is false. This is what I refer to as an epistemic interpretation: Signif-
icance tells us about the truth or falsehood of a particular hypothesis in a
particular experiment. Here we see very clearly Fisher's quasi-Bayesian view
that the exact level of significance somehow measures the confidence we
should have that the null hypothesis is false. But from a more consistent fre-
quentist viewpoint, as expressed by Neyman, a level of significance does not
tell us anything about the truth of a particular hypothesis; it states the relative
frequency of Type I errors in the long run.
Neyman (1957, 1977) called his frequentist interpretation behavioristic:To
accept or reject a hypothesis is a decision to take a particular action. Imagine
a typical application of Neyman—Pearson theory: quality control. Imagine you
have chosen the probability of Type I errors (false alarms) to be .10 and that
of Type II errors (misses) to be .01, because misses are much more costly to
your firm than false alarms. Every day you take a random sample from the
firm's production. Even if the production is normal, you will expect a signifi-
cant result (false alarm) on 10% of all days. Therefore, if a significant result
occurs, you will act as if the null hypothesis were false, that is, stop production
and check for a malfunction; but you will not necessarily believe that it is
false—because you expect a lot of false alarms in the long run.
Fisher rejected Neyman's arguments for "inductive behavior" as "childish"
(1955, p. 75), stemming from "mathematicians without personal contact with
274 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

the Natural Sciences" (p. 69). And he maintained his epistemic view: "From
a test of significance . . . we have a genuine measure of the confidence with
which any particular opinion may be held, in view of our particular data"
(p. 74). For all his anti-Bayesian talk, Fisher adopted a very similar-sounding
line of argument (Johnstone, 1987).

Does "Significant" Imply That There Is a Causal Effect?

Of course not. It is useful to distinguish between the statistical null hypothesis

and the substantive null hypothesis.2 Only the latter refers to the absence of a
particular cause. What is rejected in significance testing is the statistical hy-
pothesis, not the existence or absence of a cause. But in Fisher's writings we
can read both "yes" and "no" as answers to the aforementioned question.
Sometimes Fisher formulated the null hypothesis as "the treatment has no
effect, period," whereas in other places he formulated it as a statistical null
hypothesis (see Gigerenzer et al., 1989, pp. 95-97). In the famous Tea-Tasting
Experiment (Fisher, 1935), for instance, he stated clearly that we cannot con-
clude from a significant result (disproving the null) that the opposite hypoth-
esis (which is not formulated as an exact statistical hypothesis in null hypoth-
esis testing) is proven. (This experiment was designed to test a lady's claim
that she could tell whether the milk or the tea infusion was first added to a
cup.) That is, we cannot infer the existence of a causal process from a signif-
icant result—here, that the lady can discriminate between whether the milk
or the tea infusion was first added to the cup. For instance, there exist other
causal mechanisms (someone told the lady in which cups the tea infusion had
been poured first) that are consistent with rejecting the null hypothesis.

What Does a Nonsignificant Result Mean?

In The Design, Fisher (1935) proposed asymmetry: A null hypothesis can be
disproved, but "never proved or established" (p. 16), so "experimenters . . . are
prepared to ignore all [nonsignificant] results" (p. 13). This has been under-
stood by many textbook writers as saying that no conclusions can be drawn
from a nonsignificant result. And several textbook authors laid down the com-
mandment that I was taught: "Thou shalt not draw inferences from a nonsig-
nificant result." This made nonsignificance appear a negative, worthless, and
disappointing result. In Neyman-Pearson theory, in contrast, there is symme-
try, and a conclusion is drawn from nonsignificance: Act as if the null hy-
pothesis were true. The reason is that Neyman and Pearson start with a dis-
junction of two symmetric hypotheses (either H0 or Ha is true) and proceed by
induction through elimination.

2. On the distinction between statistical and substantive hypotheses, see Hager and
Westermann (1983) and Meehl (1978).
 THE SUPEREGO, THE EGO, AND THE ID IN STATISTICAL REASONING 275

Fisher (1955) again had second thoughts: "It is a fallacy . . . to conclude

from a test of significance that the null hypothesis is thereby established; at
most it may be said to be confirmed or strengthened" (p. 73). Thus, although
nonsignificant results cannot establish null hypotheses, according to his sec-
ond thoughts, we can do more than just "ignore" them: We may say that a
nonsignificant result "confirms" but does not "establish" the null hypothesis.
Now Fisher suggested that a nonsignificant result might indeed support the
null hypothesis, but he did not explain how.

Power

In null hypothesis testing, only one kind of error is defined: rejecting the null
hypothesis when it is in fact true. In their attempt to supply a logical basis for
Fisher's ideas and make them consistent, Neyman and Pearson replaced
Fisher's single null hypothesis by a set of rival hypotheses. In the simplest
case, two hypotheses, H0 and H,, are specified, and it is assumed that one of
them is true. This assumption allows us to determine the probability of both
Type I errors and Type II errors, indicated in Neyman-Pearson theory by a
and (3, respectively. If Ha is rejected although Ha is true, a Type II error has
occurred, ot is also called the size of a test, and 1 — P is called its power. The
power of a test is the long-run frequency of accepting Ht if it is true. The
concept of power makes explicit what Fisher referred to as "sensitivity."
Fisher (1935) pointed out two ways to make an experiment more sensitive:
by enlarging the number of repetitions and by qualitative methods, such as
experimental refinements that minimize the error in the measurements
(pp. 21-25). Nevertheless, he rejected the concept of Type II error and calcu-
lations of power on the grounds that they are inappropriate for scientific in-
duction. In his view, calculations of power, although they look harmless, reflect
the "mental confusion" between technology and scientific inference (Fisher,
1955, p. 73). If someone designs a test for acceptance procedures in quality
control, the goal of which is to minimize costs due to decision errors, calcu-
lations of power based on cost-benefit considerations in situations of repetitive
tests are quite appropriate. But scientific inference and discovery, in Fisher's
view, are about gaining knowledge, not saving money.
Fisher always rejected the concept of power. Neyman, for his part, pointed
out that some of Fisher's tests "are in a mathematical sense 'worse than use-
less,' " because their power is less than their size (see Hacking, 1965, p. 99).
Even in the Tea-Tasting Experiment, used by Fisher to introduce the logic of
null hypothesis testing in The Design, the power is only a little higher than
the level of significance (.05) or cannot be calculated at all, depending on the
conditions (see Neyman, 1950).

Random Sampling from Known Populations?

Acceptance procedures involve random sampling from a known population
(say, a firm's daily production). They also allow for repeated random sampling
276 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

(every day a random sample may be taken). Recall that Neyman and Pearson
based their theory on the concept of repeated random sampling, which denned
the probability of Type I and Type II errors as long-run frequencies of wrong
decisions in repeated experiments.
Fisher, in contrast, held that in scientific applications there is no known
population from which repeated sampling can be done. There are always many
populations to which a sample may belong. "The phrase 'repeated sampling
from the same population' does not enable us to determine which population
is to be used to define the probability level, for no one of them has objective
reality, all being products of the statistician's imagination" (Fisher, 1955,
p. 71). Fisher proposed to view any sample (such as the sample of participants
in a typical psychological experiment, which is not drawn randomly from a
known population) as a random sample from an unknown hypothetical infinite
population. "The postulate of randomness thus resolves into the question, 'Of
what population is this a random sample?' which must frequently be asked by
every practical statistician" (Fisher, 1922, p. 313). But how can the practical
statistician find out? The concept of an unknown hypothetical infinite popu-
lation has puzzled many: "This is, to me at all events, a most baffling concep-
tion" (Kendall, 1943, p. 17).

Mechanical Scientific Inference

One way of reading The Design suggests that null hypothesis testing is a fairly
mechanical procedure: Set up a null hypothesis, use a conventional level of
significance, calculate a test statistic, and disprove the null hypothesis, if you
can. Fisher later made clear that he did not mean it to be so. For instance, he
pointed out that the choice of the test statistic and deciding which null hy-
potheses are worth testing cannot be reduced to a mechanical process. You
need constructive imagination and much knowledge based on experience
(Fisher, 1933). Statistical inference has two components: informed judgment
and mathematical rigor.
Similarly, Neyman and Pearson always emphasized that the statistical part
has to be supplemented by a subjective part. As Pearson (1962) put it: "We left
in our mathematical model a gap for the exercise of a more intuitive process
of personal judgment in such matters—to use our terminology—as the choice
of the most likely class of admissible hypotheses, the appropriate significance
level, the magnitude of worthwhile effects and the balance of utilities"
(pp. 395-396).
In Neyman and Pearson's theory, once all judgments are made, the decision
(reject or accept) results mechanically from the mathematics. In his later writ-
ings, Fisher opposed these mechanical accept/reject decisions, which he be-
lieved to be inadequate in science, in which one looks forward to further data.
Science is concerned with the communication of information, such as exact
levels of significance. Again, Fisher saw a broader context, the freedom of the
Western world. Communication of information (but not mechanical decisions)
 THE SUPEREGO, THE EGO, AND THE ID IN STATISTICAL REASONING 277

recognizes "the right of other free minds to utilize them in making their own
decisions" (Fisher, 1955, p. 77).
But Neyman reproached Fisher with the same sin—mechanical statistical
inference. As a statistical behaviorist, Neyman (1957) looked at what Fisher
actually did in his own research in genetics, biology, and agriculture, rather
than at what he said one should do. He found Fisher using .01 as a conven-
tional level of significance, without giving any thought to the choice of a par-
ticular level dependent on the particular problem or the probability of an error
of the second kind; he accused Fisher of drawing mechanical conclusions,
depending on whether or not the result was significant. Neyman urged a
thoughtful choice of the level of significance, not using .01 for all problems
and contexts.
Both camps in the controversy accused the other party of mechanical,
thoughtless statistical inference; thus I conclude that here at least they
agreed—statistical inference should not be automatic.
These differences between what Fisher proposed as the logic of significance
testing and what Neyman and Pearson proposed as the logic of hypothesis
testing suffice for the purpose of this chapter. Both have developed further
tools for inductive inference, and so have others, resulting in a large toolbox
that contains maximum likelihood, fiducial probability, confidence interval ap-
proaches, point estimation, Bayesian statistics, sequential analysis, and ex-
ploratory data analysis, to mention only a few. But it is null hypothesis testing
and Neyman-Pearson hypothesis-testing theory that have transformed exper-
imental psychology and part of the social sciences.

The Offspring: Hybrid Logic

The conflicting views presented earlier are those of the parents of the hybrid
logic. Not everyone can tolerate unresolved conflicts easily and engage in a
free market of competing ideas. Some long for the single truth or search for a
compromise that could at least suppress the conflicts. Kendall (1949) com-
mented on the desire for peace negotiations among statisticians:
If some people asserted that the earth rotated from east to west and others
that it rotated from west to east, there would always be a few well-
meaning citizens to suggest that perhaps there was something to be said
for both sides, and maybe it did a little of one and a little of the other;
or that the truth probably lay between the extremes and perhaps it did
not rotate at all. (p. 115)
The denial of the existing conflicts and the pretense that there is only one
statistical solution to inductive inference were carried to an extreme in psy-
chology and several neighboring sciences. This one solution was the hybrid
logic of scientific inference, the offspring of the shotgun marriage between
Fisher and Neyman and Pearson. The hybrid logic became institutionalized in
278 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

experimental psychology (see Gigerenzer, 1987a), personality research (see

Schwartz & Dangleish, 1982), clinical psychology and psychiatry (see Meehl,
1978), education (see Carver, 1978), quantitative sociology (see Morrison &
Henkel, 1970), and archaeology (see Cowgill, 1977; Thomas, 1978), among oth-
ers. Nothing like this happened in physics, chemistry, or molecular biology.

The Hybrid Logic Is Born

Before World War II, psychologists drew their inferences about the validity of
hypotheses by many means—ranging from eyeballing to critical ratios. The
issue of statistical inference was not of primary importance. Note that this was
not because techniques were not yet available. On the contrary: already in
1710, John Arbuthnot proved the existence of God by a kind of significance
test, astronomers had used them during the 19th century for rejecting outliers
(Swijtink, 1987), and Fechner (1897) wrote a book on statistics including in-
ference techniques, to give just a few examples. Techniques of statistical in-
ference were known and sometimes used, but experimental method was not
yet dominated by and almost equated with statistical inference.
Through the work of the statisticians George W. Snedecor at Iowa State
College, Harold Hotelling at Columbia University, and Palmer Johnson at the
University of Minnesota, Fisher's ideas spread in the United States. Psychol-
ogists began to cleanse the Fisherian message of its agricultural odor and its
mathematical complexity and to write a new genre of textbooks featuring null
hypothesis testing. Guilford's Fundamental Statistics in Psychology and Edu-
cation, first published in 1942, was probably the most widely read textbook
in the 1940s and 1950s. In the preface, Guilford credited Fisher for the logic
of hypothesis testing taught in a chapter that was "quite new to this type of
text" (p. viii). The book does not mention Neyman, E. S. Pearson, or Bayes.
What Guilford teaches as the logic of hypothesis testing is Fisher's null hy-
pothesis testing, deeply colored by "Bayesian" thinking: Null hypothesis test-
ing is about the probability that the null hypothesis is true. "If the result comes
out one way, the hypothesis is probably correct, if it comes out another way,
the hypothesis is probably wrong" (p. 156). Null hypothesis testing is said to
give degrees of doubt such as "probable" or "very likely" a "more exact mean-
ing" (p. 156). Its logic is explained via surprising headings such as "Probability
of hypotheses estimated from the normal curve" (p. 160).
Guilford's logic is not consistently Fisherian, nor does it consistently use
"Bayesian" language of probabilities of hypotheses. It wavers back and forth
and beyond. Phrases such as "we obtained directly the probabilities that the
null hypothesis was plausible" and "the probability of extreme deviations from
chance" are used interchangeably for the same thing: the level of significance.
And when he proposed his own "somewhat new terms," his intuitive Bayesian
thinking becomes crystal clear. A p value of .015 for a hypothesis of zero
difference in the population "gives us the probability that the true difference
is a negative one, and the remainder of the area below the point, or .985, gives
us the probability that the true difference is positive. The odds are therefore
 THE SUPEREGO, THE EGO, AND THE ID IN STATISTICAL REASONING 279

.985 to .015 that the true difference is positive" (p. 166). In Guilford's hands,
p values that specify probabilities p(D I H ) of some data (or test statistic) D
given a hypothesis H turn miraculously into Bayesian posterior probabilities
p(H\D) of a hypothesis given data.
Guilford's confusion is not an exception. It marks the beginning of a genre
of statistical texts that vacillate between the researcher's "Bayesian" desire for
probabilities of hypotheses and what Fisher is willing to give them.
This first phase of teaching Fisher's logic soon ran into a serious compli-
cation. In the 1950s and 1960s, the theory of Neyman and E. S. Pearson also
became known. How were the textbook writers to cope with two logics of
scientific inference? How should the ideological differences and personal in-
sults be dealt with? Their solution to this conflict was striking. The textbook
writers did not side with Fisher. That is, they did not go on to present null
hypothesis testing as scientific inference and add a chapter on hypothesis test-
ing outside science, introducing the Neyman-Pearson theory as a logic for
quality control and related technological problems. Nor did they side with
Neyman and Pearson, teaching their logic as a consistent and improved version
of Fisher's and dispensing entirely with Fisherian null hypothesis testing.
Instead, textbook writers started to add Neyman-Pearsonian concepts on
top of the skeleton of Fisher's logic. But acting as if they feared Fisher's re-
venge, they did it without mentioning the names of Neyman and Pearson. A
hybrid logic of statistical inference was created in the 1950s and 1960s. Neither
Fisher nor Neyman and Pearson would have accepted this hybrid as a theory
of statistical inference. The hybrid logic is inconsistent from both perspectives
and burdened with conceptual confusion. Its two most striking features are (a)
it hides its hybrid origin and (b) it is presented as the monolithic logic of
scientific inference. Silence about its origin means that the respective parts of
the logic are not identified as part of two competing and partly inconsistent
theoretical frameworks. For instance, the idea of testing null hypotheses with-
out specifying alternative hypotheses is not identified as part of the Fisherian
framework, and the definition of the level of significance and the power of a
test as long-run frequencies of false and correct decisions, respectively, in re-
peated experiments is not identified as part of the Neyman-Pearson frame-
work. And, as a consequence, there is no mention of the fact that each of these
parts of the hybrid logic were rejected by the other party, and why, and what
the unresolved controversial issues are.

The Structure of Hybrid Logic

To capture the emotional tensions associated with the hybrid logic, I use a
Freudian analogy.3

3. Here I am elaborating on a metaphor suggested by Acree (1978). In a different

context, Devereux (1967) talked about the relation between anxiety and elimination of
subjectivity by method.
280 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

The Neyman-Pearson logic of hypothesis testing functions as the Superego

of the hybrid logic. It demands the specification of precise alternative hypoth-
eses, significance levels, and power in advance to calculate the sample size
necessary, and it teaches the doctrine of repeated random sampling. The fre-
quentist Superego forbids epistemic statements about particular outcomes or
intervals, and it outlaws the interpretation of levels of significance as the de-
gree of confidence that a particular hypothesis is true or false.
The Fisherian theory of significance testing functions as the Ego. The Ego
gets things done in the laboratory and gets papers published. The Ego deter-
mines the level of significance after the experiment, and it neither specifies
power nor calculates the sample size necessary. The Ego avoids precise pre-
dictions from its research hypothesis; that is, it does not specify the exact
predictions of the alternative hypothesis but claims support for it by rejecting
a null hypothesis. The Ego makes abundant epistemic statements about par-
ticular results. But it is left with feelings of guilt and shame for having violated
the rules.
Censored by both the frequentist Superego and the pragmatic Ego are state-
ments about probabilities of hypotheses given data. These form the Bayesian
Id of the hybrid logic. Some direct measure of the validity of the hypotheses
under question—quantitatively or qualitatively—is, after all, what researchers
really want.
The Freudian metaphor suggests that the resulting conceptual confusion in
the minds of researchers, editors, and textbook writers is not due to limited
intelligence. The metaphor brings the anxiety and guilt, the compulsive and
ritualistic behavior, and the dogmatic blindness associated with the hybrid
logic into the foreground. It is as if the raging personal and intellectual conflicts
between Fisher and Neyman and Pearson and between these frequentists and
the Bayesians were projected onto an "intrapsychic" conflict in the minds of
researchers. And the attempts of textbook writers to solve this conflict by de-
nying it have produced remarkable emotional, behavioral, and cognitive dis-
tortions.

Anxiety and Guilt

Editors and textbook writers alike have institutionalized the level of signifi-
cance as a measure of the quality of research. As mentioned earlier, Melton,
after 12 years editing one of the most prestigious journals in psychology, said
in print that he was reluctant to publish research with significance levels below
.05 but above .01, whereas p < .01 made him confident that the results would
be repeatable and deserved publication (1962, pp. 553-554). In Nunnally's In-
troduction to Statistics for Psychology and Education (1975), the student is
taught similar values and informed that the standard has been raised: "Up until
20 years ago, it was not uncommon to see major research reports in which
most of the differences were significant only at the 0.05 level. Now, such re-
sults are not taken very seriously, and it is more customary today to see results
reported only if they reach the 0.01 or even lower probability levels" (p. 195).
 THE SUPEREGO, THE EGO, AND THE ID IN STATISTICAL REASONING 281

Not accidentally, both Melton and Nunnally show the same weak understand-
ing of the logic of inference and share the same erroneous belief that the level
of significance specifies the probability that a result can be replicated (dis-
cussed later). The believers in the divinatory power of the level of significance
set the standards.
The researcher's Ego knows that these publish-or-perish standards exist in
the outside world and knows that the best way to adapt is to round up the
obtained p value after the experiment to the nearest conventional level, say to
round up the value p = .006 and publish p < .01. But the Superego has higher
moral standards: If you set alpha to 5% before the experiment, then you must
report the same finding (p = .006) as "significant at the 5% level." Mostly, the
Ego gets its way but is left with feelings of dishonesty and of guilt at having
violated the rules. Conscientious experimenters have experienced these feel-
ings, and statisticians have taken notice. The following comment was made in
a panel discussion among statisticians; Savage remarked on the statisticians'
reluctance to take responsibility for once having built up the Superego in the
minds of the experimenters:
I don't imagine that anyone in this room will admit ever having taught
that the way to do an experiment is first carefully to record the signifi-
cance level then do the experiment, see if the significance level is at-
tained, and if so, publish, and otherwise, perish. Yet, at one time we
must have taught that; at any rate it has been extremely well learned in
some quarters. And there is many a course outside of statistics depart-
ments today where the modern statistics of twenty or thirty years ago is
taught in that rigid way. People think that's what they're supposed to do
and are horribly embarrassed if they do something else, such as do the
experiment, see what significance level would have been attained, and
let other people know it. They do the better thing out of their good in-
stincts, but think they're sinning. (Barnard, Kiefer, LeCam, & Savage,
1968, p. 147)
Statistics has become more tolerant than its offspring, the hybrid logic.

Denial of the Parents

The hybrid logic attempts to solve the conflict between its parents by denying
its parents. It is remarkable that textbooks typically teach hybrid logic without
mentioning Neyman, E. S. Pearson, and Fisher—except in the context of tech-
nical details, such as specific tables, that are incidental to the logic. In 25 out
of 30 textbooks I have examined, Neyman and E. S. Pearson do not appear to
exist. For instance, in the introduction to his Statistical Principles of Experi-
mental Design (1971), Winer credits Fisher with inspiring the "standard work-
ing equipment" (p. 3) in this field, but a few pages later he presents the Ney-
man-Pearson terminology of Type I error, Type II error, power, two precise
statistical hypotheses, cost-benefit considerations, and rejecting and accepting
hypotheses. Yet nowhere in the book do the names of Neyman and E. S. Pear-
son appear (except in a "thank you" note to Pearson for permission to repro-
282 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

duce tables), although quite a few other names can be found in the index. No
hint is given to the reader that there are different ways to think about the logic
of inference. Even in the exceptional case of Hays's textbook (1963), in which
all parents are mentioned by their names, the relationship of their ideas is
presented (in a single sentence) as one of cumulative progress, from Fisher to
Neyman and Pearson (p. 287).4 Both Winer's and Hays's are among the best
texts, without the confusions that abound in Guilford's, Nunnally's, and a mass
of other textbooks. Nevertheless, even in these texts the parents' different ways
of thinking about statistical inference and the controversial issues are not
pointed out.

Denial of Conflicts between Parents

Thus the conflicting views are almost unknown to psychologists. Textbooks

are uniformly silent. (Some statistics teachers protest that airing these disputes
would only confuse students. I believe that pointing out the conflicting views
would make statistics much more interesting to students who enjoy thinking
rather than being told what to do next.) As a result of this silence, many a text
muddles through the conflicting issues leaving confusion and inconsistency in
its wake—at least, among the more intelligent and alert students. For instance,
Type I and Type II errors are often defined in terms of long-run frequencies of
erroneous decisions in repeated experiments, but the texts typically stop short
of Neyman's behavioral interpretation and fall back to epistemic interpreta-
tions of the two errors as levels of confidence about the validity of the hy-
potheses. In fact, the poorer texts overflow with amazing linguistic contortions
concerning what a level of significance means. For instance, within three pages
of text, Nunnally explained that "level of significance" means all of the fol-
lowing: (1) "If the probability is low, the null hypothesis is improbable"
(p. 194); (2) "the improbability of observed results being due to error" (p. 195);
(3) "the probability that an observed difference is real" (p. 195); (4) "the
statistical confidence . . . with odds of 95 out of 100 that the observed differ-
ence will hold up in investigations" (p. 195); (5) the degree to which experi-
mental results are taken "seriously" (p. 195); (6) "the danger of accepting a
statistical result as real when it is actually due only to error" (p. 195); (7) the
degree of "faith [that] can be placed in the reality of the finding" (p. 196); (8)
"the null hypothesis is rejected at the 0.05 level"; and (9) "the investigator can
have 95 percent confidence that the sample mean actually differs from the
population mean" (p. 196). And, after the last two versions, the author assured
his readers: "All of these are different ways to say the same thing" (Nunnally,
1975, p. 196).
Nunnally did not spell out the differences between the logics of Fisher,
Neyman and Pearson, and the Bayesians. He avoided the conflicting interpre-

4. In the third edition (1981), however, Hays's otherwise excellent text falls back to
common standards: J. Neyman and E. S. Pearson no longer appear in the book.
 THE SUPEREGO, THE EGO, AND THE ID IN STATISTICAL REASONING 283

tations by declaring that everything is the same. The price for this is conceptual
confusion, false assertions, and an illusory belief in the omnipotence of the
level of significance. Nunnally is a pronounced but not an atypical case.

Obsessive-Compulsive and Mechanical Behavior

As previously mentioned, statisticians have emphasized the indispensable role

of personal judgment, although with respect to different parts of their logics.
For Fisher, informed judgment was needed for the choice of the statistical
model, the test statistics, and a null hypothesis worth investigating. For Ney-
man and Pearson, personal judgment was needed for the choice of the class
of hypotheses (two hypotheses, in the simplest case), and the cost-benefit con-
siderations that lead to the choice of Type I error, power, and sample size. For
Bayesians such as de Finetti, finally, "subjectivism" and "relativism" are the
very cornerstones of 20th-century probability theory (de Finetti, 1931/1989;
Jeffrey, 1989).
The need for these kinds of informed judgments was rarely a topic in the
textbooks. Rather, a mass of researchers must have read the textbooks as de-
manding the mindless, mechanical setting up of null hypotheses and recording
of p values. Journals filled with p values, stars, double stars, and triple stars
that allegedly established replicable "facts" bear witness to this cookbook men-
tality.
Guilford's misunderstanding that to set up a null hypothesis means to pos-
tulate a zero difference or a zero correlation was perpetuated. "Null" denotes
the hypothesis to be "nullified," not that it is necessary to postulate a zero
effect. Rarely were null hypotheses formulated that postulated something other
than a zero effect (such as "the difference between the means is 3 scale
points"). Rarely were precise alternative hypotheses stated, and even if there
were two competing precise hypotheses, as in N. H. Anderson's information
integration theory, only one of them was tested as the null hypothesis, some-
times resulting in tests with a power as low as .06 (Gigerenzer & Richter, 1990).
Reasons for using a particular level of significance were almost never given,
and rarely was a judgment about the desired power made and the sample size
calculated. As a result, the power of the tests is typically quite low (below .50
for a medium effect), and pointing this out (Cohen, 1962) has not changed
practice. Two-and-a-half decades after Cohen's work, the power of the null
hypothesis tests was even slightly worse (Sedlmeier & Gigerenzer, 1989).
Rather, null hypotheses are set up and tested in an extremely mechanical way
reminiscent of compulsive hand washing. One can feel widespread anxiety
surrounding the exercise of informed personal judgment in matters of hypoth-
esis testing. The availability of statistical computer packages seems to have
reinforced this mindless, mechanical behavior. A student of mine once tested
in his thesis the difference between two means, which were numerically ex-
actly the same, by an F test. Just to say that the means are the same seemed
to him not objective enough.
284 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

The institutionalization of the hybrid logic as the sine qua non of scientific
method is the environment that encourages mechanical hypothesis testing. The
Publication Manual of the American Psychological Association (APA, 1974)
for instance, called "rejecting the null hypothesis" a "basic" assumption
(p. 19) and presupposes the hybrid logic. The researcher was explicitly told to
make mechanical decisions: "Caution: Do not infer trends from data that fail
by a small margin to meet the usual levels of significance. Such results are
best interpreted as caused by chance and are best reported as such. Treat the
result section like an income tax return. Take what's coming to you, but no
more" (p. 19; this passage was deleted in the third edition in 1983). This pre-
scription sounds like a Neyman-Pearson accept-reject logic, by which it mat-
ters for a decision only on which side of the criterion the data fall, not how
far. Fisher would have rejected such mechanical behavior (e.g., Fisher, 1955,
1956). Nevertheless, the examples in the manual that tell the experimenter
how to report results use p values that were obviously determined after the
experiment and rounded up to the next conventional level, such as p < .05,
p < .01, and p < .001 (pp. 39, 43, 48, 49, 70, 96). Neyman and Pearson would
have rejected this practice: These p values are not the probability of Type I
errors—and determining levels of significance after the experiment prevents
determining power and sample size in advance. Fisher (e.g., 1955,1956) would
have preferred that the exact level of significance, say p — .03, be reported,
not upper limits, such as p < .05, which look like probabilities of Type I errors
but aren't.

Distorted Statistical Intuitions

Mechanical null hypothesis testing seems to go hand-in-hand with distorted

statistical intuitions. I distinguish distorted statistical intuitions from the con-
fusion and inconsistency of the hybrid logic itself. The latter results from mish-
mashing Fisher and Neyman and Pearson without making the conflation ex-
plicit, as I argued earlier. The conceptual confusion of the hybrid logic
provided fertile ground for the growth of what I call distorted statistical intu-
itions. The distortions all seem to go in one direction: They exaggerate what
can be inferred from a p value.
The network of distorted intuitions makes the obsessive performance of null
hypothesis testing seem quite reasonable. Therefore, distorted intuitions serve
an indispensable function. These illusions guide the writings of several text-
book authors and editors, but they seem to be most pronounced in the users
of null hypothesis testing, researchers in psychology and neighboring fields.
Some distorted intuitions concern the frequentist part of the hybrid logic, oth-
ers the Bayesian Id. I give one example of each (there is a larger literature on
distorted statistical intuitions taught in statistical textbooks and held by ex-
perimenters; see Acree, 1978; Bakan, 1966; Brewer, 1985; Carver, 1978; Gutt-
man, 1977, 1985; Lykken, 1968; Pollard & Richardson, 1987; Rozeboom, 1960;
Tversky & Kahneman, 1971).
 THE SUPEREGO, THE EGO, AND THE ID IN STATISTICAL REASONING 285

Replication Fallacy Suppose a is set as .05 and the null hypothesis is rejected
in favor of a given alternative hypothesis. What if we replicate the experiment?
In what percentage of exact replications will the result again turn out signifi-
cant? Although this question arises from the frequentist conception of repeated
experiments, the answer is unknown. The a we choose does not tell us, nor
does the exact level of significance.
The replication fallacy is the belief that the level of significance provides
an answer to the question. Here are some examples: In an editorial in the
Journal of Experimental Psychology, the editor stated that he used the level of
significance reported in submitted papers as the measure of the "confidence
that the results of the experiment would be repeatable under the conditions
described" (Melton, 1962, p. 553). Many textbooks fail to mention that the
level of significance does not specify the probability of a replication, and some
explicitly teach the replication fallacy. For instance, "The question of statis-
tical significance refers primarily to the extent to which similar results would
be expected if an investigation were to be repeated" (Anastasi, 1958, p. 9). Or,
"If the statistical significance is at the 0.05 level . . . the investigator can be
confident with odds of 95 out of 100 that the observed difference will hold up
in future investigations" (Nunnally, 1975, p. 195). Oakes (1986, p. 80) asked
70 university lecturers, research fellows, and postgraduate students with at
least two years' research experience what a significant result (t — 2.7, df — 18,
p = .01) means. Sixty percent of these academic psychologists erroneously
believed that these figures mean that if the experiment is repeated many times,
a significant result would be obtained 99% of the time.
In Neyman and Pearson's theory the level of significance (alpha) is defined
as the relative frequency of rejections of H0 if H0 is true. In the minds of many,
1 — alpha erroneously turned into the relative frequency of rejections of HQ,
that is, into the probability that significant results could be replicated.

The Bayesian Id's Wishful Thinking I mentioned earlier that Fisher both re-
jected the Bayesian cake and wanted to eat it, too: He spoke of the level of
significance as a measure of the degree of confidence in a hypothesis. In the
minds of many researchers and textbook writers, however, the level of signif-
icance virtually turned into a Bayesian posterior probability.
What I call the Bayesian Id's wishful thinking is the belief that the level of
significance, say .01, is the probability that the null hypothesis is correct, or
that 1 — .01 is the probability that the alternative hypothesis is correct. In
various linguistic versions, this wishful thinking was taught in textbooks from
the very beginning. Early examples are Anastasi (1958, p. 11), Ferguson (1959,
p. 133), Guilford (1942, pp. 156-166), and Lindquist (1940, p. 14). But the be-
lief has persisted over decades of teaching hybrid logic, for instance in Miller
and Buckhout (1973; statistical appendix by F. L. Brown, p. 523), Nunnally
(1975, pp. 194-196), and the examples collected by Bakan (1966) and Pollard
and Richardson (1987). Oakes (1986, p. 82) reported that 96% of academic
psychologists erroneously believed that the level of significance specifies the
probability that the hypothesis under question is true or false.
286 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

The Bayesian Id has its share. Textbook writers have sometimes explicitly
taught this misinterpretation but have more often invited it by not specifying
the difference between a Bayesian posterior probability, a Neyman-Pearsonian
probability of a Type I error, and a Fisherian exact level of significance.

Dogmatism

The institutionalization of one way to do hypothesis testing had its benefits.

It made the administration of the social science research that had exploded
since World War II easier, and it facilitated editors' decisions. And there were
more benefits. It reduced the high art of hypothesis construction, of experi-
mental ingenuity, and informed judgment into a fairly mechanical schema that
could be taught, learned, and copied by almost anyone. The informed judg-
ments that remain are of a low-level kind: whether to use a one- or a two-
tailed significance test. (But even here some believed that there should be no
room for judgment, because even this simple choice seemed to threaten the
ideal of mechanical rules and invite cheating.) The final, and perhaps most
important, benefit of the hybrid logic is that it provides the satisfying illusion
of objectivity: The statistical logic of analyzing data seemed to eliminate the
subjectivity of eyeballing and wishful distortion. To obtain and maintain this
illusion of objectivity and impartiality, the hybrid logic had to deny its par-
ents—and their conflicts.
The danger of subjective distortion and selective reading of data exists, to
be sure. But it cannot be cured by replacing the distortions of particular ex-
perimenters with a collective distortion. Note that the institutionalized prac-
tice produces only selective and limited objectivity and hands other parts of
scientific practice over to rules of thumb—even parts for which the statistical
methods would be applicable. For example, during the 19th century, astron-
omers used significance tests to reject data (so-called outliers), assuming, at
least provisionally, that their hypothesis was correct (Swijtink, 1987). Social
scientists today, in contrast, use significance tests to reject hypotheses, assum-
ing that their data are correct. The mathematics do not dictate which one
the scientists should trust and which one they should try to refute. Social
scientists seem to have read the statistical textbooks as saying that statistical
inference is indispensable in selecting good from bad hypotheses but not for
selecting good from bad data. The problem of outliers is dealt with by rules of
thumb.5
The dogmatism with which the hybrid logic has been imposed on psy-
chology researchers by many textbook writers and editors and by researchers

5. So is the problem of how many replications (participants) an experiment should

use. Sedlmeier and Gigerenzer (1989) found no use of Neyman-Pearsonian calculations
of sample size in published work. Some statistical texts have explicitly encouraged this:
"Experienced researchers use a rule of thumb sample size of approximately twenty.
Smaller samples often result in low power values while larger samples often result in a
waste of time and money" (Bruning & Kintz, 1977, p. 7).
 THE SUPEREGO, THE EGO, AND THE ID IN STATISTICAL REASONING 287

themselves has lasted for half a century. This is far too long. We need a knowl-
edgeable use of statistics, not a collective compulsive obsession. It seems to
have gone almost unnoticed that this dogmatism has created a strange double
standard. Many researchers believe that their participants must use Bayes's
rule to test hypotheses, but the researchers themselves use the hybrid logic to
test their hypotheses—and thus themselves ignore base rates. There is the il-
lusion that one kind of statistics normatively defines objectivity in scientific
inference and another kind, rationality in everyday inference. The price is a
kind of "split brain," where Neyman-Pearson logic is the Superego for exper-
imenters' hypothesis testing and Bayesian statistics is the Superego for partic-
ipants' hypothesis testing.

Beyond Dogmatism: Toward a Thoughtful Use of Statistics

Here are a few first principles: Do not replace the dogmatism of the hybrid
logic of scientific inference with a new, although different one (e.g., Bayesian
dogmatism). Remember the obvious: The problem of inductive inference has
no universal mathematical solution. Use informed judgment and statistical
knowledge. Here are several specific suggestions:

1. Stop teaching hybrid logic as the sine qua non of scientific inference.
Teach researchers and students alternative theories of statistical in-
ference, give examples of typical applications, and teach the students
how to use these theories in a constructive (not mechanical) way.
Point out the confused logic of the hybrid, the emotional, behavioral,
and cognitive distortions associated with it, and insist on clarity (Co-
hen, 1990). This will lead to recognizing the second point.
2. Statistical inference (Fisherian, Neyman-Pearsonian, or Bayesian) is
rarely the most important part of data analysis. Teach researchers and
students to look at the data, not just at p values. Computer-aided
graphical methods of data display and exploratory data analysis are
means toward this end (Diaconis, 1985; Tukey, 1977). The calculation
of descriptive statistics such as effect sizes is a part of data analysis
that cannot be substituted by statistical inference (Rosnow & Rosen-
thai, 1989). A good theory predicts particular curves or effect sizes,
but not levels of significance.
3. Good data analysis is pointless without good data. The measurement
error should be controlled and minimized before and during the ex-
periment; instead one tends to control it after the experiment by in-
serting the error term in the F ratio. Teach researchers and students
that the important thing is to have a small real error in the data. With-
out that, a significant result at any level is, by itself, worthless—as
Cosset, who developed the t test in 1908, emphatically emphasized
(see Pearson, 1939). Minimizing the real error in measurements may
be achieved by an iterative method: First, obtain measurements and
look at the error variance, then try methods to minimize the error (e.g.,
stronger experimental control, investigating each participant carefully
288 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

in a single-case study rather than in a classroom), then go back and

obtain new measurements and look at the new error variance, and so
on, until improvements are no longer possible. Axiomatic measure-
ment theory that focuses on ordinal rather than numerical judgments
may help (Krantz, Luce, Suppes, & Tversky, 1971). It is all too rarely
used.
4. Good data need good hypotheses and theories to survive. We need
rich theoretical frameworks that allow for specific predictions in the
form of precise research hypotheses. The null hypothesis of zero dif-
ference (or zero correlation) is only one version of such a hypothe-
sis—perhaps only rarely appropriate. In particular, it has become a
bad habit not to specify the predictions of a research hypothesis but
to specify a different hypothesis (the null) and to try to reject it and
claim credit for the unspecified research hypothesis. Teach students
to derive competing hypotheses from competing theoretical frame-
works and to test their ordinal or quantitative predictions directly,
without using the null as a straw man.

Conclusions

Statistical reasoning is an art and so demands both mathematical knowledge

and informed judgment. When it is mechanized, as with the institutionalized
hybrid logic, it becomes ritual, not reasoning. Many colleagues have argued
that it is not going to be easy to get researchers in psychology and other so-
ciobiomedical sciences to drop this comforting crutch unless one offers an
easy-to-use substitute. But this is exactly what I want to avoid—the substitu-
tion of one mechanistic dogma for another. It is our duty to inform our students
of the many good roads to statistical inference that exist and to teach them
how to use informed judgment to decide which one to follow for a particular
problem. At the very least, this chapter can serve as a tool in arguments with
people who think they have to defend a ritualistic dogma instead of good
statistical reasoning. Making and winning such arguments is indispensable to
good science.
 14

Surrogates for Theories

I enjoy conference dinners. At such a dinner several years ago, I was crammed
in with four graduate students and four professors around a table laden with
Chinese food. The graduate students were eager to learn first-hand how to com-
plete a dissertation and begin a research career, and the professors were keen
to give advice. With authority, one colleague advised them: "Don't think big.
Just do four or five experiments, clip them together, and hand them in." The
graduate students nodded gratefully. They continued to nod when I added:
"Don't follow this advice unless you are mediocre or unimaginative. Try to think
in a deep, bold, and precise way. Take risks and be courageous." What a di-
lemma. How could these students follow these contradictory bits of advice?
Based on an analysis of articles in two major social psychology journals,
the Journal of Personality and Social Psychology and the Journal of Experi-
mental Social Psychology, Wallach and Wallach (1994, 1998) concluded that
the theoretical argument in almost half of the studies borders on tautology. If
an argument is a "near-tautology," there is no point in spending time and
money trying to experimentally confirm it. "Don't think big" seems to be a
prescription followed by many professional researchers, not merely conser-
vative advice for graduate students. Complaints about the lack of serious the-
ory in social psychology have been voiced before (e.g., Fiedler, 1991, 1996).
Atheoretical research is not specific to social psychology, however, although
some parts of psychology do better than others (Brandtstadter, 1987).
In this chapter, I address two questions: What are the surrogates for theory
in psychology? and What institutional forces perpetuate reliance on these sur-
rogates? This chapter is not intended to be exhaustive, only illustrative. The
examples I use are drawn from the best work in the areas discussed: the psy-
chology of reasoning, judgment, and decision making.

Surrogates

The problem is not that a majority of researchers would say that theory is
irrelevant; the problem is that almost anything passes as a theory. I identify

289
290 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

four species of surrogates for theory: one-word explanations, redescription,

muddy dichotomies, and data fitting. There are other species, such as Wallach
and Wallach's near-tautologies (whose banality and lack of imagination are as
striking as their circularity). What distinguishes these surrogates from genuine
theory is that they are vague, imprecise, and/or practically unfalsifiable, that
they often boil down to common sense and lack boldness and surprise.

One-Word Explanations
The first species of theory surrogate is the one-word explanation. Such a word
is a noun, broad in its meaning and chosen to relate to the phenomenon. At
the same time, it specifies no underlying mechanism or theoretical structure.
The one-word explanation is a label with the virtue of a Rorschach inkblot: A
researcher can read into it whatever he or she wishes to see.
Examples of one-word explanations are representativeness, availability, and
anchoring and adjustment, which are treated as the cognitive heuristics people
use to make judgments and decisions. These terms supposedly explain "cog-
nitive illusions" such as base-rate neglect. These "explanations" figure prom-
inently in current textbooks in cognitive psychology, social psychology, and
decision making. It is understandable that when these three terms were first
proposed as cognitive processes in the early 1970s, they were only loosely
characterized (Tversky & Kahneman, 1974). Yet 30 years and many experi-
ments later, these three "heuristics" remain vague and undefined, unspecified
both with respect to the antecedent conditions that elicit (or suppress) them
and also to the cognitive processes that underlie them (Gigerenzer, 1996a). I
fear that in another 30 years we will still be stuck with plausible yet nebulous
proposals of the same type: that judgments of probability or frequency are
sometimes influenced by what is similar (representativeness), comes easily to
mind (availability), and conies first (anchoring).
The problem with these heuristics is that, post hoc, at least one of them can
be fitted to almost any experimental result. For example, base-rate neglect is
commonly attributed to representativeness. But the opposite result, over-
weighting of base rates ("conservatism"), is just as easily "explained" by in-
voking anchoring (on the base rate) and adjustment. One-word explanations
derive their seductive power from the fact that almost every observation can
be called upon as an example.
Even better, one-word explanations can be so parsimonious that a single
one can explain both a phenomenon and its opposite (Ayton & Fisher, 1999).
For instance, Laplace (1814/1951) had described a phenomenon that is known
today as the gambler's fallacy: when in a random sequence a run is observed
(e.g., a series of red on the roulette wheel), players tend to believe that the
opposite result (black) will come up next. Tversky and Kahneman (1974) pro-
posed that this intuition is due to "representativeness," because "the occur-
rence of black will result in a more representative sequence than the occur-
rence of an additional red" (p. 1125). Gilovich, Vallone, and Tversky (1985)
have described another phenomenon known as the belief in the "hot hand":
 SURROGATES FOR THEORIES 291

when in a random sequence a run is observed (e.g., a series of hits scored by

a basketball player), players tend to believe that the same result (a hit rather
than a miss) will come up next. The hot-hand fallacy has also been attributed
to representativeness because "even short random sequences are thought to be
highly representative of their generating process" (p. 295). The word "repre-
sentativeness" can account for observing both A and non-A—here, the belief
that after a run the opposite result will come up (gambler's fallacy) and the
belief that after a run the same result will come up (the hot hand). Who could
ask for more?
One might think that researchers rely on such one-word explanations be-
cause they lack precise models. But this is not always the case. For instance,
there are several precise definitions of similarity, such as Euclidean distance,
the city-block metric, and various measures of feature overlap, including Tver-
sky's (1977) model (Shepard, 1962, 1974). However, attempts to define the
word representativeness have met with little attention (e.g., Smith & Osherson,
1989). Proponents of one-word explanations continue to defend undefined
terms. For instance, it has been argued that representativeness "can be assessed
experimentally; hence it need not be defined a priori" (Kahneman & Tversky,
1996, p. 585). The term availability is similarly vague. Sometimes it denotes
the "number" of instances that come to mind, sometimes the "ease" with
which instances come to mind, and at still other times it means something
else. Again, some researchers have attempted to pin down the meaning of the
term (e.g., Fiedler, 1983, 1991; Wanke, Schwarz, & Bless, 1995) but with little
effect on the thinking of others in the field. There is also a long tradition of
fairly precise conceptions of anchoring, such as in Kelson's adaptation level
and Parducci's range-frequency theories. But the seductive power of one-word
explanations seems to have caused collective amnesia. The strange reluctance
of many researchers of reasoning, judgment, and decision making to specify
precise and falsifiable process models and to work out the relationship be-
tween cognitive heuristics has been pointed out repeatedly (e.g., Einhorn &
Hogarth, 1981; Shanteau, 1989; Wallsten, 1983).
But one-word explanations have great advantages. As long as they are plau-
sible and remain unspecified, they are hard to falsify. And if one has three to
choose from—such as representativeness, availability, and anchoring and ad-
justment—at least one of them can "account" post hoc for almost any phe-
nomenon. The near-omnipotence of one-word explanations, however, does not
foster theory development.

Redescription
Recall Moliere's parody of the Aristotelian doctrine of substantial forms: Why
does opium make you sleepy? Because of its dormative properties. Redescrip-
tion has a long tradition in trait psychology, for instance, when an aggressive
behavior is attributed to an aggressive disposition or intelligent behavior to
high intelligence. But redescription in psychology is not limited to attributing
behaviors to traits and other essences.
292 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

Research on thinking and judgment and decision making is another field in

which redescription flourishes as a surrogate for theoretical ideas. Hallmarks
of redescription in these areas are words such as transparent, relevant, and
salient (which is not to say that every use of these terms implies redescription).
For instance, one important issue in problem solving is how the external rep-
resentation of a problem—whether it is represented in the form of text, figures,
probabilities, frequencies, and so on—influences performance. If an effect is
found, the question arises: Why does this type of problem representation elicit
better performance? Redescription creeps in when researchers propose that the
effect was obtained because "the correct answer is made transparent by the
representation," because of "a salient cue that makes the correct answer ob-
vious" (Kahneman & Tversky, 1996, p. 586), or because "the problem is now
simpler" (for more examples, see Gigerenzer, 1996a; Gigerenzer & Murray,
1987, pp. 159-162; Sahlin, 1991). That a representation makes a problem "sim-
pler" or its answer "transparent" is not an explanation but rather what needs
to be explained.

Muddy Dichotomies

Torn between being distressed over and content with the state of research on
information processing, Allen Newell (1973) entitled a commentary "You Can't
Play 20 Questions with Nature and Win." What distressed Newell was that
when behavior is explained in terms of dichotomies—nature versus nurture,
serial versus parallel, grammars versus associations, and so on—"clarity is
never achieved" and "matters simply become muddier and muddier as we go
down through time" (pp. 288—289). There is nothing wrong with making dis-
tinctions in terms of dichotomies per se; what concerned Newell were situa-
tions in which theoretical thinking gets stuck in binary oppositions beyond
which it never seems to move.
Let us consider a case in which false dichotomies have hindered precise
theorizing. Some arguments against evolutionary psychology are based on the
presumed dichotomy between biology and culture, or genes and environment
(Tooby & Cosmides, 1992). One such argument goes: Because cognition is
bound to culture, evolution must be irrelevant. But biology and culture are not
opposites. For instance, our ability to cooperate with conspecifics to whom we
are genetically unrelated—which distinguishes us humans from most other
species—is based on mechanisms of both biological and cultural origin. Sim-
ply to ask about the relative importance of each in terms of explained variance,
such as that 80% of intelligence is genetically inherited, is, however, not al-
ways an interesting question. The real theoretical question concerns the mech-
anism that combines what is termed the "biological" and the "cultural." For
biologists, the nature/nurture or biological/cultural dichotomy is a nontstarter:
Genes are influenced by their environment, which can include other genes,
and culture can change gene pools (coevolution).
Cognitive psychology is also muddied by vague dichotomies. For instance,
a popular opposition is between associations and rules. Sloman (1996) has
 SURROGATES FOR THEORIES 293

linked this dichotomy to Smolensky's distinction between an intuitive proces-

sor and a conscious rule interpreter, Hinton's distinction between intuitive and
rational processing, Schneider and Schiffrin's distinction between automatic
and controlled processing, Evans's distinction between a perceptually based
matching process and a linguistic-logical process, and Freud's distinction be-
tween primary and secondary processes. The problem is that these distinctions
are not all the same (Gigerenzer & Regier, 1996). Collecting more and more
binary oppositions—and labeling these "two-process theories"—does not nec-
essarily enhance clarity. Dichotomies can be an important first step, but they
cannot substitute for theories of cognitive processes.

Data Fitting
There are other surrogates for theories in psychology, one of which is the use
of powerful mathematical tools for data fitting in the absence of theoretical
underpinnings. Psychologists have historically embraced such new tools,
which they then propose as new theories. When factor analysis became a com-
mon tool for data processing in psychological research, humans were modeled
as a bundle of personality factors. When multidimensional scaling came along
in the 1960s and 1970s, human categorization and other mental processes were
proposed to be based on distances between points in multidimensional space.
More recently, the advent of the serial computer was followed by that of neural
networks as a model of cognitive function. There is nothing wrong with using
these mathematical tools per se. The important point with respect to surrogate
theories is whether the tool is used for modeling or for data fitting (this is itself
a false dichotomy, there being a continuum between these poles). Charles
Spearman originally designed factor analysis as a theory of intelligence, but
(in the form of principal component analysis) it ended up as a fitting tool for
all kinds of psychological phenomena. Likewise, Roger Shepard (e.g., 1962,
1974) interpreted the various Minkowski metrics that can be used in multi-
dimensional scaling as psychological theories of similarity, such as in color
perception, but multidimensional scaling ended up as a largely atheoretical
tool for fitting any similarity data, with the Euclidean metric as a conventional
routine. Similarly, neural networks can be used as constrained or structured
networks into which theoretical, domain-specific assumptions are built (e.g.,
Regier, 1996), but many applications of neural networks to modeling psycho-
logical phenomena seem to amount to data fitting with numerous free param-
eters. Neural networks with hidden units and other free parameters can be too
powerful to be meaningful—in the sense that they can fit different types of
results that were generated with different process models (Geman, Bienen-
stock, & Doursat, 1992; Massaro, 1988).
In general, mathematical structures can be used to test theories (with pa-
rameters determined by theoretical considerations, e.g., the metric in multi-
dimensional scaling) or as a fitting tool (with parameters chosen post hoc so
as to maximize the fit). Fitting per se is not objectionable. The danger is that
294 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

enthusiasm for a mathematical tool can lead one to get stuck in data fitting
and to use a good fit as a surrogate for a theory.

What Institutional Forces Support Surrogates for Theories?

There is one obvious reason why surrogates for theories come to mind more
quickly than real theories: demonstrating how a one-word explanation, a re-
description, a dichotomy, or an exercise in data fitting "explains" a phenom-
enon demands less mental strain than developing a bold and precise theory.
It takes imagination to conceive the idea that heat is caused by motion, but
only little mental effort to propose that heat is caused by specific particles that
have the propensity to be hot. In what follows, I identify two institutions that
may maintain (rather than cause) the abundant use of surrogates for theories
in some areas of psychology.

The Institutionalization of Null Hypothesis Testing

In recent years, more and more scholars have argued against the ritual of null
hypothesis testing, which was institutionalized in psychology around 1955.
Many other scholars have responded with passionate defenses of it. So far the
debate focuses on issues such as whether one should replace significance test-
ing by confidence intervals, effect sizes, or something else. In my view, how-
ever, the single most important issue is that institutionalized null hypothesis
testing allows surrogates for theories to flourish (Chapter 13). To switch to
confidence intervals will not necessarily reverse this trend.
Null hypothesis testing provides researchers with no incentive to specify
either their own research hypotheses or competing hypotheses. The ritual is
to test one's unspecified hypothesis against "chance," that is, against the null
hypothesis that postulates "no difference between the means of two popula-
tions" or "zero correlation." As Danziger (1990) has shown, the origin of this
practice is in parapsychology and education, where the interest was not in
testing positive theories but detecting effects greater than those of chance. The
problem lies not in statistical testing per se but in a specific statistical method
that became institutionalized. If psychologists had adopted statistical methods
that test two or more well-specified hypotheses against one another (such as
Neyman-Pearson or Bayesian statistics), then they would have been forced to
formulate precise hypotheses.
As long as there is an institutionalized methodology that does not encourage
researchers to specify their hypotheses, there is little incentive to think hard
and develop theories from which such hypotheses could be derived. Accord-
ingly, the voluminous instruction manual from which graduate students and
researchers learn how to write an article, the APA Publication Manual, devotes
many pages to methodology but not to theory. From this students learn not
that hypotheses and theories should be bold, surprising, and precisely stated,
but that their business is to test null hypotheses. To perform this ritual, mere
 SURROGATES FOR THEORIES 295

surrogates for thinking big are sufficient. The result has been called "null sci-
ence" (Bower, 1997). It reminds me of a mechanical maxim regarding the crit-
ical ratio (the difference between the means divided by the standard deviation
of the differences), the predecessor of the significance level: "A critical ratio
of three, or no Ph.D."

Disciplinary Isolation
Over the course of the 20th century, academic psychology has become more
and more compartmentalized into subdisciplines such as social psychology,
cognitive psychology, developmental psychology, and so on. Each subdisci-
pline has its own journals, reviewers, and grant programs, and one can have
a career in one of them without ever reading the journals of neighboring
subdisciplines. In addition, job searches are often organized according to
these categories. This territorial organization of psychology discourages re-
searchers from engaging with psychological knowledge and colleagues out-
side of their territory, not to mention with other disciplines. As Jerry Fodor
(1995) put it:
Unfortunately, cognitive psychology as people are trained to practice it,
at least in this country, has been traditionally committed to methodolog-
ical empiricism and to disciplinary isolationism, in which it was, for
example, perfectly possible to study language without knowing anything
about linguistics, (pp. 85-86)
This isolationism is by no means restricted to the study of language. For
instance, the experimental study of logical thinking in arguably the most re-
searched problem, the Wason selection task, has been carried out with little
reference to modern logic, and the study of statistical reasoning has been con-
ducted with little attention to the relevant issues in statistics (see Gigerenzer,
1994a; Oaksford & Chater, 1994).
Intellectual inbreeding can block the flow of positive metaphors from one
discipline to another. Neither disciplines nor subdisciplines are natural cate-
gories. Interdisciplinary exchange has fueled the development of some of the
most influential new metaphors and theories in the sciences, such as when
Ludwig Boltzmann and James Clerk Maxwell developed statistical mechanics
by borrowing from sociology. Bolzmann and Maxwell modeled the behavior
of gas molecules on the behavior of humans as Adolphe Quetelet had portrayed
it: erratic and unpredictable at the individual level but exhibiting orderly sta-
tistical laws at the level of collectives (Gigerenzer et al., 1989, ch. 2). Territorial
science, in contrast, blocks the flow of metaphors and the development of new
theories. Distrust and disinterest in anything outside one's subdiscipline sup-
ports surrogates for theory.
296 COGNITIVE ILLUSIONS AND STATISTICAL RITUALS

Data without Theory

In this chapter, I have specified four surrogates for theory and two possible
institutional reasons why some of these surrogates flourish like weeds. These
two reasons certainly cannot explain the whole story.
Several years ago, I spent a day and a night in a library reading through
issues of the Journal of Experimental Psychology from the 1920s and 1930s.
This was professionally a most depressing experience, but not because these
articles were methodologically mediocre. On the contrary, many of them make
today's research pale in comparison with their diversity of methods and sta-
tistics, their detailed reporting of single-case data rather than mere averages,
and their careful selection of trained participants. And many topics—such as
the influence of the gender of the experimenter on the performance of the
participants—were of interest then as now. What depressed me was that almost
all of this work is forgotten; it does not seem to have left a trace in the collec-
tive memory of our profession. It struck me that most of it involved collecting
data without substantive theory. Data without theory are like babies without
parents: Their life expectancy is low.
 REFERENCES

Aaronson, D., Grupsmith, E., & Aaronson, M. (1976). The impact of computers
on cognitive psychology. Behavioral Research Methods and Instrumenta-
tion, 8, 129-138.
Acree, M. C. (1978). Theories of statistical inference in psychological research:
A historicocritical study. Dissertation Abstracts International, 39 (10),
5037B. (University Microfilms No. H790 H7000.)
Adler, J. E. (1984). Abstraction is uncooperative. Journal for the Theory of So-
cial Behavior, 14, 165-181.
Adler, J. E. (1991). An optimist's pessimism: Conversation and conjunction. In
E. Eells & T. Maruszewski (Eds.), Probability and rationality: Studies on
L. Jonathan Cohen's philosophy of science (pp. 251-282). Amsterdam:
North-Holland Rodopi.
Ajzen, I. (1977). Intuitive theories of events and the effects of base-rate infor-
mation on predictions. Journal of Personality and Social Psychology, 35,
303-314.
Ajzen, I., & Fishbein, M. (1975). A Bayesian analysis of attribution processes.
Psychological Bulletin, 82, 261-277.
Alba, J. W., & Marmorstein, H. (1987). The effects of frequency knowledge on
consumer decision making. Journal of Consumer Research, 14, 14—26.
Alibi des Schornsteinfegers: Unwahrscheinliche Wahrscheinlichkeitsrech-
nungen in einem Mordprozess [Alibi of a chimney sweep: Improbable
probability calculations in a murder trial]. Rheinischer Merkur, No. 39.
Airport, D. A. (1975). The state of cognitive psychology. Quarterly Journal of
Experimental Psychology, 27, 141-152.
Allwood, C. M., & Montgomery, H. (1987). Response selection strategies and
realism of confidence judgments. Organizational Behavior and Human De-
cision Processes, 39, 365-383.
American Psychological Association (1974). Publication Manual of the Amer-
ican Psychological Association (2nd ed.). Baltimore: Garamond/Pridemark
Press.
American Psychological Association (1983). Publication Manual of the Amer-
ican Psychological Association (3rd ed.). Menasha, WI: Banta.
Anastasi, A. (1958). Differential psychology (3rd ed.) New York: Macmillan.

297
298 REFERENCES

Anderson, J. R. (1980). Cognitive psychology and its implications. San Fran-

cisco, CA: Freeman.
Anderson, J. R. (1990). The adaptive character of thought. Hillsdale, NJ: Erl-
baum.
Anderson, J. R., & Milson, R. (1989). Human memory: An adaptive perspective.
Psychological Review, 96, 703-719.
Anderson, N. H. (1986). A cognitive theory of judgment and decision. In B.
Brehmer, H. Jungermann, P. Lourens, & G. Sevon (Eds.), New directions in
research on decision making (pp. 63-108). Amsterdam: North-Holland.
Arbib, M. A. (1993). Allen Newell, unified theories of cognition. Artificial In-
telligence, 59, 265-283.
Arbuthnot, J. (1710). An argument for Divine Providence, taken from the con-
stant regularity observ'd in the births of both sexes. Philosophical Trans-
actions of the Royal Society, 27, 186-190.
Arkes, H. R. (1981). Impediments to accurate clinical judgment and possible
ways to minimize their impact. Journal of Consulting and Clinical Psy-
chology, 49, 323-330.
Arkes, H. R., Christensen, C., Lai, C., & Blumer, C. (1987). Two methods of
reducing overconfidence. Organizational Behavior and Human Decision
Processes, 39, 133-144.
Arkes, H. R., & Hammond, K. R. (Eds.). (1986). Judgment and decision making:
An interdisciplinary reader. Cambridge, England: Cambridge University
Press.
Armelius, B., & Armelius, K. (1974). The use of redundancy in multiple-cue
judgments: Data from a suppressor-variable-task. American fournal of Psy-
chology, 87, 385-392.
Armelius, K. (1979). Task predictability and performance as determinants of
confidence in multiple-cue judgments. Scandinavian fournal of Psychol-
ogy, 20, 19-25.
Ash, M. G. (1987). Psychology and politics in interwar Vienna: The Vienna
Psychological Institute, 1922-1942. In M. G. Ash & W. R. Woodward
(Eds.), Psychology in twentieth-century thought and society (pp. 143-164).
Cambridge: Cambridge University Press.
Aspray, W. (1990). John von Neumann and the origins of modern computing.
Cambridge, MA: MIT Press.
Atran, S. (1990). Cognitive foundations of natural history: Towards an anthro-
pology of science. Cambridge: Cambridge University Press.
Axelrod, R. (1984). The evolution of cooperation. New York: Basic Books.
Ayton, P., & Fisher, I. (1999). The gambler's fallacy and the hot-hand-fallacy:
Two faces of subjective randomness. Manuscript submitted for publica-
tion.
Baars, B. J. (1986). The cognitive revolution in psychology. New York: Guilford
Press.
Babbage, C. (1812-1994). Passages from the life of a philosopher (M. Campbell-
Kelley, Ed.). Piscataway, NJ: IEEE Press. (Original work published 1812)
Bakan, D. (1966). The test of significance in psychological research. Psycho-
logical Bulletin, 66, 423-437.
Bar-Hillel, M. (1980). The base-rate fallacy in probability judgments. Acta Psy-
chologica, 44, 211-233.
Barnard, G. A. (1979). Discussion of the paper by Professors Lindley and Tver-
 REFERENCES 299

sky and Dr. Brown. Journal of the Royal Statistical Society, Series A, 142,
171-172.
Barnard, G. A., Kiefer, J. C., LeCam, L. M., & Savage, L. J. (1968). Statistical
inference. In D. G. Watts (Ed.), The future of statistics (p. 147). New York:
Academic Press.
Baron-Cohen, S. (1995). Mindblindness. Cambridge, MA: MIT Press.
Barsalou, L. W., & Ross, B. H. (1986). The roles of automatic and strategic pro-
cessing in sensitivity to superordinate and property frequency. Journal of
Experimental Psychology: Learning, Memory, and Cognition, 12, 116-134.
Bayes, T. (1763). A essay towards solving a problem in the doctrine of chances.
Philosophical Transactions of the Royal Society of London, 53, 370-418.
Bazerman, M. H. (1990). Judgment in managerial decision making. New York:
Wiley.
Bazerman, M. H., & Neale, M. A. (1986). Heuristics in negotiation: Limitations
to effective dispute resolution. In H. R. Arkes & R. R. Hammond (Eds.),
Judgment and decision making: An interdisciplinary reader (pp. 311-321).
Cambridge, England: Cambridge University Press.
Becker, G. (1976). The economic approach to human behavior. Chicago: Uni-
versity of Chicago Press.
Beretty, P. M., Todd, P. M., & Martignon, L. (1999). Categorization by elimina-
tion: Using few cues to choose. In G. Gigerenzer, P. M. Todd, & the ABC
Group, Simple heuristics that make us smart (pp. 235-254). New York:
Oxford University Press.
Binet, A., & Simon, T. (1914). Mentally defective children. London: Edward
Arnold.
Birnbaum, M. H. (1983). Base rates in Bayesian inference: Signal detection
analysis of the cab problem. American Journal of Psychology, 96, 85-
94.
Bjorkman, M. (1984). Decision making, risk taking and psychological time:
Review of empirical findings and psychological theory. Scandinavian
Journal of Psychology, 25, 31-49.
Bjorkman, M. (1987). A note on cue probability learning: What conditioning
data reveal about cue contrast. Scandinavian Journal of Psychology, 28,
226-232.
Blackwell, R. J. (1983). Scientific discovery: The search for new categories.
New Ideas in Psychology, I , 111-115.
Boesch, C., & Boesch, H. (1984). Mental map in wild chimpanzees: An analysis
of hammer transports for nut cracking. Primates, 25, 160-170.
Boole, G. (1958). An investigation of the laws of thought on which are founded
the mathematical theories of logic and probabilities. New York: Dover.
(Original work published 1854)
Borges, B., Goldstein, D. G., Ortmann, A., & Gigerenzer, G. (1999). Can igno-
rance beat the stock market? In G. Gigerenzer, P. M. Todd, & the ABC
Group, Simple heuristics that make us smart (pp. 59-72). New York: Ox-
ford University Press.
Borgida, E., & Brekke, N. (1981). The base rate fallacy in attribution and pre-
diction. In J. H. Harvey, W. J. Ickes, & R. F. Kidd (Eds.), New directions in
attribution research (Vol. 3, pp. 63-95). Hillsdale, NJ: Erlbaum.
Boring, E. G. (1942). Sensation and perception in the history of experimental
psychology. New York: Appleton-Century-Crofts.
300 REFERENCES

Boring, E. G. (1957). A history of experimental psychology. New York:

Appleton-Century-Crofts.
Boring, E. G. (1963). History, psychology, and science: Selected papers. New
York: Wiley.
Bower, B. (1997). Null science. Science News, 151, 356-357.
Brandtstadter, J. (1987). On certainty and universality in human development: De-
velopmental psychology between apriorism and empiricism. In M. Chapman
& R. A. Dixon (Eds.), Meaningand the growth of understanding: Wittgenstein's
significance for developmental psychology (pp. 69-84). Berlin: Springer.
Erase, G. L., Cosmides, L., & Tooby, J. (1998). Individuation, counting, and
statistical inference: The role of frequency and whole object representa-
tions in judgment under uncertainty. Journal of Experimental Psychology:
General, 127, 3-21.
Brehmer, A., & Brehmer, B. (1988). What has been learned about human judg-
ment from thirty years of policy capturing? In B. Brehmer & C. R. B. Joyce
(Eds.), Human judgment: The SJT view (pp. 75-144). Amsterdam: Elsevier.
Brehmer, B. (1988). The development of social judgment theory. In B. Brehmer
& C. R. B. Joyce (Eds.), Human judgment: The SJT view (pp. 13-40). Am-
sterdam: Elsevier.
Brehmer, B. (1994). The psychology of linear judgment models. Acta Psychol-
ogica, 87, 137-154.
Brehmer, B., & Joyce, C. R. B. (Eds.). (1988). Human judgment: The SJT view.
Amsterdam: Elsevier.
Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1993). Classification
and regression trees. New York: Chapman & Hall.
Brewer, J. K. (1985). Behavioral statistics textbooks: Source of myths and mis-
conceptions? Journal of Educational Statistics, 10, 252-268.
Broder, A. (in press). Assessing the empirical validity of the "Take The Best"
heuristic as a model of human probabilistic inference. Journal of Experi-
mental Psychology: Learning, Memory and Cognition.
Brown, N. R., & Siegler, R. S. (1993). Metrics and mappings: A framework for
understanding real-world quantitive estimation. Psychological Review,
100, 511-534.
Brown, W., & Thomson, G. H. (1921). The essentials of mental measurement.
Cambridge, England: Cambridge University Press.
Bruning, J. L., & Kintz, B. L. (1977). Computational handbook of statistics (2nd
ed.). Glenview, IL: Scott, Foresman.
Brunswik, E. (1934). Wahrnehmung und Gegenstandswelt: Grundlegung einer
Psychologie vom Gegenstand her. Leipzig, Germany: Deutike.
Brunswik, E. (1937). Psychology as a science of objective relations. Philosophy
of Science, 4, 227-260.
Brunswik, E. (1939). Probability as a determiner of rat behavior. Journal of
Experimental Psychology, 25, 175-197.
Brunswik, E. (1943). Organismic achievement and environmental probability.
Psychological Review, 50, 255-272.
Brunswik, E. (1955). Representative design and probabilistic theory in a func-
tional psychology. Psychological Review, 62, 193-217.
Brunswik, E. (1956). Perception and the representative design of psychological
experiments. Los Angeles: University of California Press.
Brunswik, E. (1957). Scope and aspects of the cognitive problem. In H. Gruber,
 REFERENCES 301

R. Jessor, & K. R. Hammond (Eds.), Cognition: The Colorado Symposium

(pp. 5-40). Cambridge, MA: Harvard University Press.
Bmnswik, E. (1964). Scope and aspects of the cognitive problem. In J. S. Bru-
ner, E. Brunswik, L. Festinger, F. Heider, K. F. Muenzinger, C. E. Osgood,
& D. Rapaport (Eds.), Contemporary approaches to cognition (pp. 4-31).
Cambridge, MA: Harvard University Press.
Bundeszentrale fur gesimdheitliche Aufklarung. (1988—1993). Wissenswertes
tiber den HIV-Test, Issues 1-10. Cologne: Bundeszentrale fur gesundheit-
liche Aufklarung.
Burke, D. S., Brundage, J. F., Redfield, R. R., Damato, J. J., Schnabel, C. A., Put-
man, P., Visitine, R., & Kim, H. J. (1988), Measurement of the false positive
rate in a screening program for human immunodeficiency virus infection.
New England Journal of Medicine, 319, 961-964.
Busch, M. P. (1994). HIV testing in blood banks. In G. Schochetman & J. R.
George (Eds.), AIDS testing: A comprehensive guide to technical, medical,
social, legal, and management issues (pp. 224-236). New York: Springer.
Buss, D. M. (1987). Sex differences in human mate selection criteria: An evo-
lutionary perspective. In C. Crawford, M. Smith, & D. Krebs (Eds.), Socio-
biology and psychology: Ideas, issues and applications (pp. 335-351).
Hillsdale, NJ: Erlbaum.
Byrne, R., & Whiten, A. (Eds.). (1988). Machiavellian intelligence: Social ex-
pertise and the evolution of intellect in monkeys, apes and humans. Ox-
ford, England: Clarendon Press.
Caraco, T., Martindale, S., & Whittam, T. S. (1980). An empirical demonstration
of risk-sensitive foraging preferences. Animal Behavior, 28, 820-830.
Carnap, R. (1969). The logical structure of the world (Trans. R. A. George),
Berkeley: University of California Press. (Original work published 1928)
Carpenter, R. G., Gardner, A., McWeeny, P. M., & Emery, J. L. (1977). Multistage
scoring system for identifying infants at risk of unexpected death. Archives
of Disease in Childhood, 53, 606-612.
Carroll, J. B. (1982). The measurement of intelligence. In R. J. Sternberg (Ed.),
Handbook of human intelligence (pp. 29—120). Cambridge, England: Cam-
bridge University Press.
Carver, R. P. (1978). The case against statistical significance testing. Harvard
Educational Review, 48, 378-399.
Casscells, W., Schoenberger, A., & Grayboys, T. (1978). Interpretation by phy-
sicians of clinical laboratory results. New England Journal of Medicine,
299, 999-1000.
Castellan, N. J. (1981). On-line computers in psychology: The last 10 years, the
next 10 years—the challenge and the promise. Behavioral Research Meth-
ods and Instrumentation, 13, 91-96.
Castellan, N. J. (1991). Computers and computing in psychology: Twenty years
of progress and still a bright future. Behavior Research Methods, Instru-
ments, and Computers, 23, 106-108.
Catalan, J., & Pugh, K. (1995). Suicidal behavior and HIV infection—is there a
link? AIDS Care, 7 (Suppl. 2), 117-121.
Cheng, P. W., & Holyoak, K. J. (1985). Pragmatic reasoning schemas. Cognitive
Psychology, 17, 391-416.
Cheng, P. W., & Holyoak, K. J. (1989). On the natural selection of reasoning
theories. Cognition, 33, 285-313.
302 REFERENCES

Christensen-Szalanski, J. J. J., & Beach, L. R. (1982). Experience and the base-

rate fallacy. Organizational Behavior and Human Performance, 29, 270-
278.
Cohen, J. (1962). The statistical power of abnormal-social psychological re-
search: A review. Journal of Abnormal and Social Psychology, 65, 145-
153.
Cohen, J. (1990). Things I have learned (so far). American Psychologist, 45,
1304-1312.
Cohen, L. J. (1981). Can human irrationality be experimentally demonstrated?
Behavioral and Brain Sciences, 4, 317-370.
Cohen, L. J. (1982). Are people programmed to commit fallacies? Further
thoughts about the interpretation of experimental data on probability. Jour-
nal for the Theory of Social Behaviour, 12, 251-274.
Cohen, L. J. (1986). The dialogue of reason. Oxford, England: Clarendon Press.
Cohen, L. J. (1989). The philosophy of induction and probability. Oxford, En-
gland: Clarendon Press.
Colebrooke, H. (1825). Address on presenting the Gold Medal of the Astro-
nomical Society to Charles Babbage. Memoirs of the Astronomical Society,
1, 509-512.
Cooper, W. S. (1989). How evolutionary biology challenges the classical theory
of rational choice. Biology and Philosophy, 4, 457-481.
Cooper, W. S., & Kaplan, R. (1982). Adaptive "coin-flipping": A decision-
theoretic examination of natural selection for random individual variation.
Journal of Theoretical Biology, 94, 135-151.
Cosmides, L. (1985). Deduction or Darwinian algorithms? An explanation of
the "elusive" content effect on the Wason selection task. Dissertation Ab-
stracts International, 47 (1), 358B. (University Microfilms No. 86-02206.)
Cosmides, L. (1989). The logic of social exchange: Has natural selection shaped
how humans reason? Studies with the Wason selection task. Cognition,
31, 187-276.
Cosmides, L., & Tooby, J. (1992). Cognitive adaptations for social exchange. In
J. Barkow, L. Cosmides, & J. Tooby (Eds.), The adapted mind Evolutionary
psychology and the generation of culture (pp. 163-228). New York: Oxford
University Press.
Cosmides, L., & Tooby, J. (1994a). Beyond intuition and instinct blindness:
Towards an evolutionary rigorous cognitive science. Cognition, 50, 41-77.
Cosmides, L., & Tooby, J. (1994b). Origins of domain specificity: The evolution
of functional organization. In L. A. Hirschfeld & S. A. Gelman (Eds.), Map-
ping the mind: Domain specificity in cognition and culture (pp. 85-116).
Cambridge, England: Cambridge University Press.
Cosmides, L., & Tooby, J. (1996). Are humans good intuitive statisticians after
all? Rethinking some conclusions from the literature on judgment under
uncertainty. Cognition, 58, 1-73.
Cowgill, G. L. (1977). The trouble with significance tests and what we can do
about it. American Antiquity, 42, 350-368.
Cronbach, L. J. (1957). The two disciplines of scientific psychology. American
Psychologist, 12, 671-684.
Cummins, D. D. (1998). Social norms and other minds: The evolutionary roots
of higher cognition. In D. D. Cummins & C. Allen (Eds.), The evolution of
mind (pp. 30—50). New York: Oxford University Press.
 REFERENCES 303

Curd, M. (1980). The logic of discovery: An analysis of three approaches. In

T. Nickles (Ed.), Scientific discovery, logic, and rationality (pp. 201-219).
Dordrecht, Holland: Reidel.
Daly, M., & Wilson, M. (1988). Homicide. New York: Aldine de Gruyter.
Dannenberg, A. L., McNeil, J. G., Brundage, J. F., & Brockmeyer, R. (1996). Sui-
cide and HIV-infection. Journal of the American Medical Association, 276,
1743-1746.
Danziger, K. (1985). The methodological imperative in psychology. Philosophy
of the Social Sciences, 15, 1-13.
Danziger, K. (1987). Statistical method and the historical development of re-
search practice in American psychology. In L. Kriiger, G. Gigerenzer, &
M. S. Morgan (Eds.), The probabilistic revolution: Vol. 2. Ideas in the sci-
ences (pp. 35-47). Cambridge, MA: MIT Press.
Danziger, K. (1990). Constructing the subject: Historical origins of psycholog-
ical research. Cambridge, England: Cambridge University Press.
Darwin, C. (1965). The expressions of the emotions in man and animal. Chi-
cago: University of Chicago Press. (Original work published 1872)
Dashiell, J. F. (1939). Some rapprochements in contemporary psychology. Psy-
chological Bulletin, 36, 1-24.
Daston, L. (1980). Probabilistic expectation and rationality in classical proba-
bility theory. Historia Mathematica, 7, 234-260.
Daston, L. (1981). Mathematics and the moral sciences: The rise and fall of the
probability of judgments, 1785-1840. In H. N. Jahnke & M. Otte (Eds.),
Epistemological and social problems of the sciences in the early nine-
teenth century (pp. 287-309). Dordrecht, Holland: Reidel.
Daston, L. (1987). The domestication of risk: Mathematical probability and
insurance 1650-1830. In L. Kriiger, L. Daston, & M. Heidelberger (Eds.),
The probabilistic revolution: Vol. 1. Ideas in history (pp. 237-260). Cam-
bridge, MA: MIT Press.
Daston, L. (1988). Classical probability in the Enlightenment. Princeton, NJ:
Princeton University Press.
Daston, L. (1992). The naturalized female intellect. Science in Context, 5, 209-
235.
Daston, L. (1994). Enlightenment calculations. Critical Inquiry, 21, 182-202.
Dawes, R. M. (1979). The robust beauty of improper linear models. American
Psychologist, 34, 571-582.
Dawes, R. M. (1980). Confidence in intellectual judgments vs. confidence in
perceptual judgments. In E. D. Lantermann & H. Feger (Eds.), Similarity
and choice: Papers in honor of Clyde Coombs (pp. 327—345). Bern, Swit-
zerland: Huber.
Dawkins, R. (1976). The selfish gene. Oxford, England: Oxford University
Press.
Dehaene, S. (1997). The number sense: How the mind creates mathematics.
New York: Oxford University Press.
Dershowitz, A. (1996). Reasonable doubts: The criminal justice system and the
O. J. Simpson case. New York: Simon & Schuster.
Deutscher Bundestag. (1988). AIDS: Fakten und Konsequenzen. Progress report
of the Enquete Committee of the llth German Bundestag, 3/88. Bonn,
Germany: Bonner Universitats Buchdruckerei.
Deutscher Bundestag. (1990). AIDS: Fakten und Konsequenzen. Final report of
304 REFERENCES

the Enquete Committee of the llth German Bundestag, 13/90. Bonn, Ger-
many: Bonner Universitats Buchdruckerei.
Devereux, G. (1967). From anxiety to method in the behavioral sciences. Paris:
Mouton.
de Waal, F. B. M., & Luttrell, L. M. (1988). Mechanisms of social reciprocity in
three primate species: Symmetrical relationship characteristics or cogni-
tion? Ethology and Sociobiology, 9, 101-118.
Diaconis, P. (1985). Theories of data analysis: From magical thinking through
classical statistics. In D. C. Hoaglin, F. Mosteller, & J. W. Tukey (Eds.), Ex-
ploring data tables, trends and shapes (pp. 1-36). New York: Wiley.
Dietz, K., Seydel, J., & Schwartlander, B. (1994). Back-projection of German
AIDS data using information on dates of tests. Statistics in Medicine, 13,
1991-2008.
DiFonzo, N. (1994). Piggybacked syllogisms for investor behavior: Probabilistic
mental modeling in rumor-based stock market trading. Ph.D. diss., Temple
University, Philadelphia.
Doherty, M. E., & Kurz, E. M. (1996). Social judgment theory. Thinking and
Reasoning, 2, 109-140.
Doll, L. S., & Kennedy, M. B. (1994). HIV counseling and testing: What is it
and how well does it work? In G. Schochetman & J. R. George (Eds.), AIDS
testing: A comprehensive guide to technical, medical, social, legal, and
management issues (pp. 302-319). New York: Springer.
Dugatkin, L. A. (1996). Interface between culturally based preferences and ge-
netic preferences: Female mate choice in Poecilia Reticulata. Proceedings
of the National Academy of Sciences, 93, 2770-2773.
Duncker, K. (1945). On problem solving (Trans. T. L. S. Lees). Psychological
Monographs, 58 (5, Whole No. 270). (Original work published 1935.)
Barman, J. (1992). Bayes or bust? A critical examination of Bayesian confir-
mation theory. Cambridge, MA: MIT Press.
Eberle, J. F., Deinhardt, K. O., & Habermehl, M. A. (1988). Die Zuverlassigkeit
des HIV-Antikorpertests. Deutsches Arzteblatt, 85, 1512-1514.
Eddy, D. M. (1982). Probabilistic reasoning in clinical medicine: Problems and
opportunities. In D. Kahneman, P. Slovic, & A. Tversky (Eds.), Judgment
under uncertainty: Heuristics and biases (pp. 249-267). Cambridge, En-
gland: Cambridge University Press.
Edgington, E. S. (1974). A new tabulation of statistical procedures used in APA
journals. American Psychologist, 29, 25-26.
Edwards, W. (1954). The theory of decision making. Psychological Bulletin, 51,
380-417.
Edwards, W. (1962). Dynamic decision theory and probabilistic information
processing. Human Factors, 4, 59-73.
Edwards, W. (1966). Nonconservative information processing systems. Rep.
No. 5893-22-F. Ann Arbor: University of Michigan, Institute of Science
and Technology.
Edwards, W. (1968). Conservatism in human information processing. In B.
Kleinmuntz (Ed.), Formal representation of human judgment (pp. 17-52).
New York: Wiley.
Edwards, W., Lindman, H., & Savage, L. J. (1963). Bayesian statistical inference
for psychological research. Psychological Review, 70, 193-242.
Edwards, W., & von Winterfeldt, D. (1986). On cognitive illusions and their
 REFERENCES 305

implications. In H. R. Arkes & K. R. Hammond (Eds.), Judgment and de-

cision making: An interdisciplinary reader (pp. 642-679). Cambridge, En-
gland: Cambridge University Press.
Einhorn, H. J. (1970). The use of nonlinear, noncompensatory models in
decision-making. Psychological Bulletin, 73, 221-230.
Einhorn, H. J., & Hogarth, R. M. (1975). Unit weighting schemes for decision
making. Organizational Behavior and Human Performance, 13, 171-192.
Einhorn, H. J., & Hogarth, R. M. (1981). Behavioral decision theory: Processes
of judgment and choice. Annual Review of Psychology, 32, 53-88.
Elstein, A. S. (1988). Cognitive processes in clinical inference and decision
making. In D. C. Turk & P. Salovey (Eds.), Reasoning, inference and judg-
ment in clinical psychology (pp. 17-50). New York: Free Press.
Elster, J. (1979). Ulysses and the sirens: Studies in rationality and irrationality.
Cambridge, England: Cambridge University Press.
Elster, J. (1990). When rationality fails. In K. S. Cook & M. Levi (Eds.), The
limits of rationality (pp. 19-51). Chicago: University of Chicago Press.
Epstein, J. S. (1994). Regulation of HIV-related tests and procedures. In G.
Schochetman & J. R. George (Eds.), AIDS testing: A comprehensive guide
to technical, medical, social, legal, and management issues (pp. 52-61).
New York: Springer.
Ericsson, K. A., & Simon, H. A. (1984). Protocol analysis: Verbal reports as
data. Cambridge, MA: MIT Press.
Estes, W. K. (1959). The statistical approach to learning theory. In S. Koch (Ed.),Psy-
chology:A study of science (Vol. 2, pp. 380-491). New York: McGraw-Hill.
Estes, W. K. (1976). The cognitive side of probability learning. Psychological
Review, 83, 37-64.
Evans, R. B. (1990). Robert Sessions Woodworth and the "Columbia Bible":
How the psychological experiment was redefined. American Journal of
Psychology, 103, 391-401.
A false HIV test caused 18 months of hell. (1993, March 5). Chicago Tribune.
Fechner, G. T. (1897). Kollektivmasslehre [The measurement of collectives].
Leipzig, Germany: Engelmann.
Ferguson, L. (1959). Statistical analysis in psychology and education. New
York: McGraw-Hill.
Feynman, R. (1967). The character of physical law. Cambridge, MA: MIT Press.
Fiedler, K. (1983). On the testability of the availability heuristic. In R. W.
Scholz (Ed.), Decision making under uncertainty (pp. 109-119). Amster-
dam: North-Holland.
Fiedler, K. (1988). The dependence of the conjunction fallacy on subtle lin-
guistic factors. Psychological Research, 50, 123-129.
Fiedler, K. (1991). Heuristics and biases in theory formation. Theory and Psy-
chology, 1, 407-430.
Fiedler, K. (1996). Explaining and simulating judgment biases as an aggregation
phenomenon in probabilistic, multiple-cue environments. Psychological
Review, 103, 193-214.
Fillenbaum, S. (1977). Mind your p's and q's: The role of content and context
in some uses of and, or, and if. Psychology of Learning and Motivation,
11, 41-100.
Finetti, B. de (1989). Probabilism. Erkenntnis, 31, 169-223. (Original work
published 1931)
306 REFERENCES

Fischer, D. (1990). AIDS-Modellprogramm der Bundesregierung—eine Zwi-

schenbilanz. Offentliches Gesundheitswesen, 52, 425-431.
Fischhoff, B. (1977). Perceived informativeness of facts. Journal of Experimen-
tal Psychology: Human Perception and Performance, 3, 349-358.
Fischhoff, B. (1982). Debiasing. In D. Kahneman, P. Slovic, & A. Tversky (Eds.),
Judgment under uncertainty: Heuristics and biases (pp. 422-444). Cam-
bridge, England: Cambridge University Press.
Fischhoff, B. (1988). Judgment and decision making. In R. J. Sternberg & E. E.
Smith (Eds.), The psychology of human thought (pp. 153-187). Cambridge,
England: Cambridge University Press.
Fischhoff, B., & MacGregor, D. (1982). Subjective confidence in forecasts. Jour-
nal of Forecasting, 1, 155-172.
Fischhoff, B., Slovic, P., & Lichtenstein, S. (1977). Knowing with certainty: The
appropriateness of extreme confidence. Journal of Experimental Psychol-
ogy, 3, 552-564.
Fishburn, P. C. (1988). Nonlinear preference and utility theory. Baltimore:
Johns Hopkins University Press.
Fishburn, P. C. (1991). Nontransitive preferences in decision theory. Journal of
Risk and Uncertainty, 4, 113-134.
Fisher, R. A. (1922). On the mathematical foundations of theoretical statistics.
Philosophical Transactions of the Royal Society of London, A, 222, 309-
368.
Fisher, R. A. (1925). Statistical methods for research workers. Edinburgh, Scot-
land: Oliver & Boyd.
Fisher, R. A. (1933). The contributions of Rothamsted to the development
of the science of statistics. Annual Report of the Rothamsted Station, 43-
50.
Fisher, R. A. (1935). The design of experiments. Edinburgh, Scotland: Oliver
& Boyd.
Fisher, R. A. (1955). Statistical methods and scientific induction. Journal of the
Royal Statistical Society, Series B, 17, 69-78.
Fisher, R. A. (1956). Statistical methods and scientific inference. Edinburgh,
Scotland: Oliver & Boyd.
Fodor, J. A. (1983). The modularity of mind. Cambridge, MA: MIT Press.
Fodor, J. A. (1995). The folly of simulation. In P. Baumgartner & S. Payr (Eds.),
Speaking minds: Interviews with twenty eminent cognitive scientists
(pp. 85-100). Princeton, NJ: Princeton University Press.
Fraisse, P. (1964). The psychology of time. London: Eyre & Spottiswoode.
Fretwell, S. D. (1972). Populations in seasonal environments. Princeton, NJ:
Princeton University Press.
Frey, B. S. (1990). Entscheidungsanomalien: Die Sicht der Okonomie. Psy-
chologische Rundschau, 41, 67-83.
Freyd, J. J. (1983). Shareability: The social psychology of epistemology. Cog-
nitive Science, 7, 191-210.
Galison, P. (1987). How experiments end. Chicago: University of Chicago Press.
Gallistel, C. R. (1990). The organization of learning. Cambridge, MA: MIT Press.
Gallistel, C. R., Brown, A. L., Carey, S., Gelman, R., & Keil, F. C. (1991). Lessons
from animal learning for the study of cognitive development. In S. Carey
& R. Gelman (Eds.), The epigenesis of mind: Essays on biology and cog-
nition (pp. 3-36). Hillsdale, NJ: Erlbaum.
 REFERENCES 307

Gallistel, C. R., & Gelman, R. (1992). Preverbal and verbal counting and com-
putation. Cognition, 44, 43-74.
Garcia, J., & Koelling, R. A. (1966). The relation of cue to consequence in avoid-
ance learning. Psychonomic Science, 4, 123-124.
Garcia y Robertson, R., & Garcia, J. (1985). X-rays and learned taste aversions:
Historical and psychological ramifications. In T. G. Burish, S. M. Levy, &
B. E. Meyerowitz (Eds.), Cancer, nutrition and eating behavior: A biobe-
havioral perspective (pp. 11-41). Hillsdale, NJ: Erlbaum.
Gardner, H. (1988). Creative lives and creative works: A synthetic scientific
approach. In R. J. Sternberg (Ed.), The nature of creativity (pp. 298-321).
Cambridge, England: Cambridge University Press.
Gavanski, I., & Hui, C. (1992). Natural sample spaces and uncertain belief.
Journal of Personality and Social Psychology, 63 (5), 766-780.
Gavin, E. A. (1972). The causal issue in empirical psychology from Hume to
the present with emphasis upon the work of Michotte. Journal of the His-
tory of the Behavorial Sciences, 8, 302-320.
Geman, S., Bienenstock, E., & Doursat, R. (1992). Neural networks and the bias/
variance dilemma. Neural Computation, 4, 1-58.
George, J. R., & Schochetman, G. (1994). Detection of HIV infection using se-
rologic techniques. In G. Schochetman & J. R. George (Eds.), AIDS testing:
A comprehensive guide to technical, medical, social, legal, and manage-
ment issues (pp. 62-102). New York: Springer.
Gigerenzer, G. (1981). Messung und Modellbildung in der Psychologie [Mea-
surement and modeling in psychology]. Munich: Reinhardt.
Gigerenzer, G. (1984). External validity of laboratory experiments: The
frequency-validity relationship. American Journal of Psychology, 97 (2),
185-195.
Gigerenzer, G. (1987a). Probabilistic thinking and the fight against subjectivity.
In L. Kriiger, G. Gigerenzer, & M. S. Morgan (Eds.), The probabilistic revo-
lution: Vol. 2. Ideas in the sciences (pp. 11-33). Cambridge, MA: MIT Press.
Gigerenzer, G. (1987b). Survival of the fittest probabilist: Brunswik, Thurstone,
and the two disciplines of psychology. In L. Kriiger, G. Gigerenzer, & M.
Morgan (Eds.), The probabilistic revolution: Vol. 2. Ideas in the sciences
(pp. 49-72). Cambridge, MA: MIT Press.
Gigerenzer, G. (1990). Strong AI and the problem of "second-order" algorithms.
Behavioral and Brain Sciences, 13 (4), 663-664.
Gigerenzer, G. (199la). From tools to theories: A heuristic of discovery in cog-
nitive psychology. Psychological Review, 98, 254-267.
Gigerenzer, G. (1991b). How to make cognitive illusions disappear: Beyond
"heuristics and biases." In W. Stroebe & M. Hewstone (Eds.), European
Review of Social Psychology, 2, 83-115.
Gigerenzer, G. (1991c). Does the environment have the same structure as Bayes'
theorem? Behavioral and Brain Sciences, 14, 495.
Gigerenzer, G. (199ld). On cognitive illusions and rationality. In E. Eells & T.
Manuszewski (Eds.), Reasoning and rationality: Essays in honour of L. J.
Cohen (pp. 225-249). Amsterdam: Radop.
Gigerenzer, G. (1993a). The superego, the ego, and the id in statistical reason-
ing. In G. Keren & G. Lewis (Eds.), A handbook for data analysis in the
behavioral sciences: Methodological issues (pp. 311-339). Hillsdale, NJ:
Erlbaum.
308 REFERENCES

Gigerenzer, G. (1993b). The bounded rationality of probabilistic mental mod-

els. In K. I. Manktelow & D. E. Over (Eds.), Rationality: Psychological and
philosophical perspectives (pp. 284-313). London: Routledge.
Gigerenzer, G. (1994a). Why the distinction between single-event probabilities
and frequencies is relevant for psychology (and vice versa). In G. Wright
& P. Ayton (Eds.), Subjective probability (pp. 129-162). New York: Wiley.
Gigerenzer, G. (1994b). Where do new ideas come from? In M. A. Boden (Ed.),
Dimensions of creativity (pp. 53-74). Cambridge, MA: MIT Press.
Gigerenzer, G. (1995). The taming of content: Some thoughts about domains
and modules. Thinking and Reasoning, 1, 324-333.
Gigerenzer, G. (1996a). On narrow norms and vague heuristics: A reply to
Kahneman and Tversky. Psychological Review, 103, 592-596.
Gigerenzer, G. (1996b). The psychology of good judgment: Frequency formats
and simple algorithms. Journal of Medical Decision Making, 16, 273-280.
Gigerenzer, G. (1996c). Rationality: Why social context matters. In P. B. Baltes,
& U. M. Staudinger (Eds.), Interactive minds: Life-span perspectives on the
social foundation of cognition (pp. 319-346). Cambridge, England: Cam-
bridge University Press.
Gigerenzer, G. (1997). The modularity of social intelligence. In A. Whiten &
R. W. Byrne (Eds.), Machiavellian intelligence: II (pp. 264-288). Cam-
bridge, England: Cambridge University Press.
Gigerenzer, G. (1998a). Ecological intelligence: An adaptation for frequencies.
In D. D. Cummins & C. Allen (Eds.), The evolution of mind (pp. 9-29).
New York: Oxford University Press.
Gigerenzer, G. (1998b). Surrogates for theories. Theory and Psychology, 8,195—
204.
Gigerenzer, G. (1998c). Psychological challenges for normative models. In
D. M. Gabbay & P. Smets (Eds.), Handbook of defeasible reasoning and
uncertainty management systems: Vol. 1. Quantified representation of un-
certainty and imprecision (pp. 441—467). Dordrecht: Kluwer.
Gigerenzer, G. (in press). Ideas in exile: The struggles of an upright man. In
K. R. Hammond & T. R. Stewart (Eds.), The essential Brunswik: Begin-
nings, explications, applications. New York: Oxford University Press.
Gigerenzer, G., & Goldstein, D. G. (1996a). Reasoning the fast and frugal way:
Models of bounded rationality. Psychological Review, 103, 650-669.
Gigerenzer, G., & Goldstein, D. G. (1996b). Mind as computer: Birth of a meta-
phor. Creativity Research Journal, 9, 131-144.
Gigerenzer, G., Hell, W., & Blank, H. (1988). Presentation and content: The use
of base rates as a continuous variable. Journal of Experimental Psychology:
Human Perception and Performance, 14, 513-525.
Gigerenzer, G., & Hoffrage, U. (1995). How to improve Bayesian reasoning with-
out instruction: Frequency formats. Psychological Review, 102, 684-704.
Gigerenzer, G., & Hoffrage, U. (1999). Overcoming difficulties in Bayesian rea-
soning: A reply to Lewis & Keren and Mellers & McGraw. Psychological
Review, 106, 425-430.
Gigerenzer, G., Hoffrage, U., & Ebert, A. (1998). AIDS counseling for low-risk
clients. AIDS Care, 10, 197-211.
Gigerenzer, G., Hoffrage, U., & Kleinbolting, H. (1991). Probabilistic mental
models: A Brunswikian theory of confidence. Psychological Review, 98,
506-528.
 REFERENCES 309

Gigerenzer, G., & Hug, K. (1992). Domain-specific reasoning: Social contracts,

cheating, and perspective change. Cognition, 43, 127-171.
Gigerenzer, G., & Kurz, E. M. (in press). Vicarious functioning reconsidered: A
fast and frugal lens model. In K. R. Hammond & T. R. Stewart (Eds.), The
essential Brunswik: Beginnings, explications, applications. New York: Ox-
ford University Press.
Gigerenzer, G., & Murray, D. J. (1987). Cognition as intuitive statistics. Hills-
dale, NJ: Erlbaum.
Gigerenzer, G., & Regier, T. (1996). How do we tell an association from a rule?
Comment on Sloman (1996). Psychological Bulletin, 119, 23-26.
Gigerenzer, G., & Richter, H. R. (1990). Context effects and their interaction
with development: Area judgments. Cognitive Development, 5, 235—264.
Gigerenzer, G., & Selten, R. (Eds.). (2000). Bounded rationality: The adaptive
toolbox. Cambridge, MA: MIT Press.
Gigerenzer, G., Swijtink, Z., Porter, T., Daston, L., Beatty, J., & Kriiger, L. (1989).
The empire of chance: How probability changed science and everyday life.
Cambridge, England: Cambridge University Press.
Gigerenzer, G., & Todd, P. M. (1999). Fast and frugal heuristics: The adaptive tool-
box. In G. Gigerenzer, P. M. Todd, & the ABC Research Group, Simple heuris-
tics that make us smart (pp. 3-34). New York: Oxford University Press.
Gigerenzer, G., Todd, P. M., & the ABC Research Group (1999). Simple heuris-
tics that make us smart. New York: Oxford University Press.
Gillespie, J. H. (1977). Natural selection for variances in offspring numbers: A
new evolutionary principle. American Naturalist, 111, 1010-1014.
Gilovich, T., Vallone, R., & Tversky, A. (1985). The hot hand in basketball: On
the misconception of random sequences. Cognitive Psychology, 17, 295—
314.
Gilpin, M. E. (1975). Limit cycles in competition communities. American Nat-
uralist, 109, 51-60.
Ginossar, Z., & Trope, Y. (1980). The effects of base rates and individuating
information on judgments about another person. Journal of Experimental
Social Psychology, 16, 228-242.
Ginossar, Z., & Trope, Y. (1987). Problem solving in judgment under uncer-
tainty. Journal of Personality and Social Psychology, 52, 464-474.
Gleick, J. (1992). Genius: The life and science of Richard Feynman. New York:
Pantheos.
Gluck, D., Vornwald, A., Grossau, E., & Kubanek, B. (1990). HIV prevalence in
blood donors in urban and rural areas of the Federal Republic of Germany.
Blut, 60, 304-307.
Goldstein, D. G. (1994). The less-is-more effect in inference. M.A. thesis, Uni-
versity of Chicago.
Goldstein, D. G., & Gigerenzer, G. (1999). The recognition heuristic: How ig-
norance makes us smart. In G. Gigerenzer, P. M. Todd, & the ABC Research
Group, Simple heuristics that make us smart (pp. 37-58). New York: Ox-
ford University Press.
Good, I. J. (1971). 46656 varieties of Bayesians. American Statistician, 25, 62-
63.
Good, I. J. (1995). When batterer turns murderer. Nature, 375, 541.
Gould, S. J. (1992). Bully for brontosaurus: Further reflections in natural his-
tory. New York: Penguin Books.
310 REFERENCES

Gould, S. J., & Vrba, E. S. (1982). Exaptation—a missing term in the science of
form. Paleobiology, 8, 4-15.
Gregory, R. L. (1974). Concepts and mechanisms of perception. New York:
Scribner.
Grether, D. M. (1980). Bayes rule as a descriptive model: The representative-
ness heuristic. Quarterly Journal of Economics, 95, 537-557.
Grice, H. P. (1975). Logic and conversation. In P. Cole & J. L. Morgan (Eds.),
Syntax and semantics: Vol. 3. Speech acts (pp. 41—58). New York: Aca-
demic Press.
Griffin, D., & Tversky, A. (1992). The weighing of evidence and the determi-
nants of confidence. Cognitive Psychology, 24, 411-435.
Griggs, R. A., & Cox, J. R. (1982). The elusive thematic-materials effect in Wa-
son's selection task. British Journal of Psychology, 73, 407-420.
Groner, M., Groner, R., & Bischof, W. F. (1983). Approaches to heuristics: A
historical review. In R. Groner, M. Groner, & W. F. Bischof (Eds.), Methods
of heuristics (pp. 1-18). Hillsdale, NJ: Erlbaum.
Gruber, H. E. (1977). The fortunes of a basic Darwinian idea: Chance. In R. W.
Rieber & K. Salzinger (Eds.), The roots of American psychology: Historical
influences and implications for the future (pp. 233-245). New York: New
York Academy of Sciences.
Gruber, H. E. (1981). Darwin on man: A psychological study of scientific cre-
ativity (2nd ed.). Chicago: University of Chicago Press.
Gruber, H. E., & Voneche, J. J. (Eds.). (1977). The essential Piaget. New York:
Basic Books.
Guilford, J. P. (1942). Fundamental statistics in psychology and education.
New York: McGraw-Hill.
Guilford, J. P. (1954). Psychometric methods (2nd ed.). New York: McGraw-Hill.
Guilford, J. P., & Hoepfner, R. (1971). The analysis of intelligence. New York:
McGraw-Hill.
Guttman, L. (1977). What is not what in statistics. Statistician, 26, 81-107.
Guttman, L. (1985). The illogic of statistical inference for cumulative science.
Applied Stochastic Models and Data Analysis, 1, 3-10.
Hacking, I. (1965). Logic of statistical inference. Cambridge, England: Cam-
bridge University Press.
Hacking, I. (1975). The emergence of probability. Cambridge, England: Cam-
bridge University Press.
Hacking, I. (1983). Representing and intervening. Cambridge, England: Cam-
bridge University Press.
Hackmann, W. D. (1979). The relationship between concept and instrument
design in eighteenth-century experimental science. Annals of Science, 36,
205-224.
Hager, W., & Westermann, R., (1983). Zur Wahl und Pruning statistischer Hy-
pothesen in psychologischen Untersuchungen. Zeitschrift fur experimen-
ted und angewandte Psychologie, 30, 67-94.
Hamilton, W. D. (1964). The genetic evolution of social behavior: Parts 1 and
2. Journal of Theoretical Biology, 7, 1-52.
Hammerton, M. (1973). A case of radical probability estimation. Journal of
Experimental Psychology, 101, 252-254.
Hammond, K. R. (1966). The psychology of Egon Brunswik. New York: Holt,
Rinehart & Winston.
 REFERENCES 311

Hammond, K. R. (1980). Introduction to Brunswikian theory and methods. New

Directions for Methodology of Social and Behavioral Science, 3,1-11.
Hammond, K. R. (1990). Functionalism and illusionism: Can integration be
usefully achieved? In R. M. Hogarth (Ed.), Insights in decision making
(pp. 227-261). Chicago: University of Chicago Press.
Hammond, K. R., Hursch, C. J., & Todd, F. J. (1964). Analyzing the components
of clinical inference. Psychological Review, 71, 438-456.
Hammond, K. R., Stewart, T. R., Brehmer, B., & Steinmann, D. O. (1975). Social
judgment theory. In M. F. Kaplan & S. Schwartz (Eds.), Human judgment
and decision processes (pp. 271-312). New York: Academic Press.
Hammond, K. R., & Summers, D. A. (1965). Cognitive dependence on linear
and nonlinear cues. Psychological Review, 72, 215-244.
Hammond, K. R., & Wascoe, N. E. (Eds.). (1980). Realizations of Brunswik's
representative design. New Directions for Methodology of Social and Be-
havioral Science, 3, 271-312.
Hansen, R. D., & Donoghue, J. M. (1977). The power of consensus: Information
derived from one's own and others' behavior. Journal of Personality and
Social Psychology, 35, 294-302.
Hanson, N. R. (1958). Patterns of discovery. Cambridge, England: Cambridge
University Press.
Harcourt, A. H., & de Waal, F. B. M. (Eds.) (1992). Coalitions and alliances in
humans and other animals. Oxford, England: Oxford University Press.
Harvard University Center for Cognitive Studies. (1963). Third annual report.
Cambridge, MA: Harvard University, Center for Cognitive Studies.
Harvard University Center for Cognitive Studies. (1964). Fourth annual report.
Cambridge, MA: Harvard University, Center for Cognitive Studies.
Harvard University Center for Cognitive Studies. (1966). Sixth annual report.
Cambridge, MA: Harvard University, Center for Cognitive Studies.
Harvard University Center for Cognitive Studies. (1968). Eighth annual report.
Cambridge, MA: Harvard University, Center for Cognitive Studies.
Harvard University Center for Cognitive Studies. (1969). Ninth annual report.
Cambridge, MA: Harvard University, Center for Cognitive Studies.
Hasher, L., Goldstein, D., & Toppino, T. (1977). Frequency and the conference of
referential validity. Journal of Verbal Learning and Behavior, 16, 107—112.
Hasher, L., & Zacks, R. T. (1979). Automatic and effortful processes in memory.
Journal of Experimental Psychology: General, 180, 356-388.
Hasher, L., & Zacks, R. T. (1984). Automatic processing of fundamental infor-
mation. American Psychologist, 39, 1327-1388.
Hays, W. L. (1963). Statistics for psychologists (2nd ed.). New York: Holt, Rine-
hart & Winston.
Heider, F. (1958). The psychology of interpersonal relations. New York: Wiley.
Heider, F., & Simmel, M. (1944). An experimental study of apparent behavior.
American Journal of Psychology, 57, 243-259.
Herbart, J. F. (1834). A textbook in psychology (Trans. M. K. Smith). New York:
Appleton-Century-Crofts.
Hertwig, R., & Gigerenzer, G. (1999). The "conjunction fallacy" revisited: How
intelligent inferences look like reasoning errors. Journal of Behavioral De-
cision Making, 12, 275-305.
Hertwig, R., Gigerenzer, G., & Hoffrage, U. (1997). The reiteration effect in
hindsight bias. Psychological Review, 104, 194-202.
312 REFERENCES

Hertwig, R., Hoffrage, U., & Martignon, L. (1999). Quick estimation: Letting the
environment do the work. In G. Gigerenzer, P. M. Todd, & the ABC Group,
Simple heuristics that make us smart (pp. 209-234). New York: Oxford
University Press.
Hilgard, E. R. (1955). Discussion of probabilistic functionalism. Psychological
Review, 62, 226-228.
Hilton, D. J. (1995). The social context of reasoning: Conversational inference
and rational judgment. Psychological Bulletin, 118, 248-271.
Hintzman, D. L., & Block, R. A. (1972). Repetition and memory: Evidence for
multiple trace hypothesis. Journal of Experimental Psychology, 88, 297-
306.
Hintzman, D. L., Nozawa, G., & Irmscher, M. (1982). Frequency as a nonpro-
positional attribute of memory. Journal of Verbal Learning and Verbal Be-
havior, 21, 127-141.
Hirschfeld, L. A., & Gelman, S. A. (Eds.). (1994a). Mapping the mind: Domain
specificity in cognition and culture. Cambridge, England: Cambridge Uni-
versity Press.
Hirschfeld, L. A., & Gelman, S. A. (1994b). Toward a topography of mind: An
introduction to domain specificity. In L. A. Hirschfeld & S. A. Gelman
(Eds.), Mapping the mind: Domain specificity in cognition and culture
(pp. 3-35). Cambridge, England: Cambridge University Press.
Hoffman-Valentin, F. (1991). AIDS: Gefahren, Schutz, Vorsorge, Behandlungs-
moglichkeiten. Landsberg, Germany: Ecomed.
Hoffrage, U. (1994). Zur Angemessenheit subjektiver Sicherheits-Urteile: Eine
Exploration der Theorie der probabilistischen mentalen Modelle [On the
validity of confidence judgments: A study of the theory of probabilistic
mental models]. Ph.D. diss., University of Salzburg.
Hoffrage, U., & Gigerenzer, G. (1996). The impact of information representation
on Bayesian reasoning. In G. Cottrell (Ed.), Proceedings of the Eighteenth
Annual Conference of the Cognitive Science Society (pp. 126-130). Mah-
wah, NJ: Erlbaum.
Hoffrage, U., & Gigerenzer, G. (1998). Using natural frequencies to improve
diagnostic inferences. Academic Medicine, 73, 538—540.
Hoffrage, U., & Hertwig, R. (1999). Hindsight bias: A price worth paying for
fast and frugal memory. In G. Gigerenzer, P. M. Todd, & the ABC Group.
Simple heuristics that make us smart (pp. 191-208). New York: Oxford
University Press.
Hofstatter, P. R. (1939). Uber die Schatzung von Gruppeneigenschaften. Zeit-
schrift fur Psychologie, 145, 1-44.
Holland, J. H., Holyoak, K. J., Nisbett, R. E., & Thagard, P. R. (1986). Induction:
Processes of inference, learning and discovery. Cambridge, MA: MIT Press.
Holton, G. (1988). Thematic origins of scientific thought (2nd ed.). Cambridge,
MA: Harvard University Press.
Howell, W. C., & Burnett, S. (1978). Uncertainty measurement: A cognitive
taxonomy. Organizational Behavior and Human Performance, 22, 45—68.
Huber, O. (1989). Information-processing operators in decision making. In H.
Montgomery & O. Svenson (Eds.), Process and structure in human deci-
sion making (pp. 3-21). New York: Wiley.
Hull, C. L. (1943). The uniformity point of view. Psychological Review, 50,
203-216.
 REFERENCES 313

Humphrey, G. (1951). Thinking. New York: Wiley.

Humphrey, N. K. (1988). The social function of intellect. In R. Byrne & A.
Whiten (Eds.), Machiavellian intelligence (pp. 13-26). Oxford, England:
Clarendon Press. (Original work published 1976)
Hutchins, E. (1995). Cognition in the wild. Cambridge, MA: MIT Press.
Huttenlocher, J., Hedges, L., & Prohaska, V. (1988). Hierarchical organization
in ordered domains: Estimating the dates of events. Psychological Review,
95, 471-484.
Inhelder, B., & Piaget, J. (1969). The early growth of logic in the child. New
York: Norton Library.
Jeffrey, R. (1989). Reading Probabilismo. Erkenntnis, 31, 225-237.
John Paul II (1997). The pope's message on evolution and four commentaries.
Quarterly Review of Biology, 72, 381-406.
Johnson-Laird, P. N. (1983). Mental models. Cambridge, England: Cambridge
University Press.
Johnstone, D. J. (1987). Tests of significance following R. A. Fisher. British Jour-
nal of the Philosophy of Science, 38, 481-499.
Jones, E. E., & McGillis, D. (1976). Correspondent inferences and the attribution
cube: A comparative reappraisal. In J. H. Harvey, W. J. Ickes, & R. F. Kidd
(Eds.), New directions in attribution research, Vol. 1 (pp. 389-420). Hills-
dale, NJ:Erlbaum.
Jonides, J., & Jones, C. M. (1992). Direct coding for frequency of occurrence.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 18,
368-378.
Jungermann, H. (1983). The two camps on rationality. In R. W. Scholz
(Ed.), Decision making under uncertainty (pp. 63-86). Amsterdam:
Elsevier.
Juslin, P. (1993). An explanation of the hard-easy effect in studies of realism
of confidence in one's general knowledge. European Journal of Cognitive
Psychology, 5, 55-71.
Juslin, P. (1994). The overconfidence phenomenon as a consequence of infor-
mal experimenter-guided selection of almanac items. Organizational Be-
havior and Human Decision Processes, 57, 226—246.
Juslin, P., Olsson, H., & Winman, A. (1998). The calibration issue: Theoretical
comments on Suantak, Bolger, and Ferrell (1996). Organizational Behavior
and Human Decision Processes, 73, 3-26.
Juslin, P., Winman, A., & Persson, T. (1995). Can overconfidence be used as an
indicator of reconstructive rather than retrieval processes? Cognition, 54,
99-130.
Juslin, P., Winman, A., & Olsson, H. (2000). Naive empiricism and dogmatism
in confidence research: A critical examination of the hard-easy effect. Psy-
chological Review, 107, 384-396.
Kadane, J. B., & Lichtenstein, S. (1982). A subjectivist view of calibration. De-
cision Research Rep. No. 82-6. Eugene, OR: Decision Research.
Kahneman, D., & Tversky, A. (1972). Subjective probability: A judgment of
representativeness. Cognitive Psychology, 3, 430-454.
Kahneman, D., & Tversky, A. (1973). On the psychology of prediction. Psycho-
logical Review, 80, 237-251.
Kahneman, D., & Tversky, A. (1982). On the study of statistical intuitions. In
D. Kahneman, P. Slovic, & A. Tversky (Eds.), Judgment under uncertainty:
314 REFERENCES

Heuristics and biases (pp. 493-508). Cambridge, England: Cambridge Uni-

versity Press.
Kahneman, D., & Tversky, A. (1996). On the reality of cognitive illusions. Psy-
chological Review, 103, 582-591.
Keeney, R. L., & Raiffa, H. (1993). Decisions with multiple objectives. Cam-
bridge, England: Cambridge University Press.
Kelley, H. H. (1967). Attribution theory in social psychology. In D. Levine (Ed.),
Nebraska symposium on motivation (Vol. 15, pp. 192-238). Lincoln: Uni-
versity of Nebraska Press.
Kelley, H. H., & Michaela, I. L. (1980). Attribution theory and research. Annual
Review of Psychology, 31, 457-501.
Kendall, M. G. (1942). On the future of statistics. Journal of the Royal Statis-
tical Society, 105, 69-80.
Kendall, M. G. (1943). The advanced theory of statistics (Vol. 1). New York:
Lippincott.
Kendall, M. G. (1949). On the reconciliation of theories of probability. Bio-
metrika, 36, 101-116.
Keren, G. (1987). Facing uncertainty in the game of bridge: A calibration study.
Organizational Behavior and Human Decision Processes, 39, 98-114.
Keren, G. (1988). On the ability of monitoring non-veridical preceptions and
uncertain knowledge: Some calibration studies. Ada Psychologica, 67,
95-119.
Kita, S. (1993). Language and thought interface: A study of spontaneous ges-
tures and Japanese mimetics. Ph.D. diss., University of Chicago.
Klayman, J., & Ha, Y. (1987). Confirmation, discontinuation, and information
in hypothesis testing. Psychological Review, 94, 211-228.
Klayman, J., Soil, J., Gonzalez-Vallejo, C., & Barlas, S. (1999). Overconfidence?
It depends on how, what, and whom you ask. Organizational Behavior
and Human Decision Processes, 79, 216-247.
Kleiter, G. D. (1994). Natural sampling: Rationality without base rates. In G. H.
Fischer & D. Laming (Eds.), Contributions to mathematical psychology,
psychometrics, and methodology (pp. 375-388). New York: Springer.
Klix, F. (1993). Erwachendes Denken: Geistige Leistung aus evolutionspsy-
chologischer Sicht [Thought coming to life: Cognitive performance from
an evolutionary psychological perspective]. Heidelberg, Germany: Spek-
trum.
Koehler, J. J. (1992). Probabilities in the courtroom: An evaluation of the ob-
jections and policies. In D. K. Kagehiro & W. S. Laufer (Eds.), Handbook
of psychology and law (pp. 167-184). New York: Springer.
Koehler, J. J. (1996). The base rate fallacy reconsidered: Descriptive, normative
and methodological challenges. Behavioral and Brain Sciences, 19, 1-53.
Koriat, A., Lichtenstein, S., & Fischhoff, B. (1980). Reasons for confidence.
Journal of Experimental Psychology: Human Learning and Memory, 6,
107-118.
Kosslyn, S. M., & Pomerantz, J. R. (1977). Imagery, propositions, and the form
of internal representations. Cognitive Psychology, 9, 52-76.
Krantz, D. H., Luce, R. D., Suppes, P., & Tversky, A. (1971). Foundations of
measurement (Vol. 1). New York: Academic Press.
Krauss, S., Martignon, L., & Hoffrage, U. (1999). Simplifying Bayesian infer-
ence: The general case. In L. Magnani, N. Nersessian, & N. Thagard (Eds.),
 REFERENCES 315

Model-based reasoning in scientific discovery (Vol. 1, pp. 165-179). New

York: Academic Press.
Krebs, J. R., & Davies, N. B. (1987). An introduction to behavioral ecology (2nd
ed.). Oxford, England: Blackwell.
Krech, D. (1955). Discussion: Theory and reductionism. Psychological Review,
62, 229-231.
Kroznick, J. A., Li, F., & Lehman, D. R. (1990). Conversational conventions,
order of information acquisition, and the effect of base rates and individ-
uating information on judgments. Journal of Personality and Social Psy-
chology, 59, 1140-1152.
Kriiger, L., Daston, L., & Heidelberger, M. (Eds.) (1987). The probabilistic rev-
olution: Vol. 1. Ideas in history. Cambridge, MA: MIT Press.
Kriiger, L., Gigerenzer, G., & Morgan, M. S. (Eds.). (1987). The probabilistic
revolution: Vol. 2. Ideas in the sciences. Cambridge, MA.: MIT Press.
Kuhn, T. (1970). The structure of scientific revolutions (2nd ed.). Chicago: Uni-
versity of Chicago Press.
Kummer, H. (1988). Tripartite relations in hamadryas baboons. In R. Byrne &
A. Whiten (Eds.), Machiavellian intelligence: Social expertise and the ev-
olution of intellect in monkeys, apes and humans (pp. 113-121). Oxford,
England: Clarendon Press.
Kummer, H., Daston, L., Gigerenzer, G., & Silk, J. (1997). The social intelligence
hypothesis. In P. Weingart, P. Richerson, S. D. Mitchell, & S. Maasen (Eds.),
Human by nature: Between biology and the social sciences (pp. 157-179).
Hillsdale, NJ: Erlbaum.
Kyburg, H. E. (1983). Rational belief. Behavioral and Brain Sciences, 6, 231-
273.
Landman, J., & Manis, M. (1983). Social cognition: Some historical and theo-
retical perspectives. Advances in Experimental Social Psychology, 16, 49-
123.
Langley, P., Simon, H. A., Bradshaw, G. L., & Zytkow, J. M. (1987). Scientific
discovery. Cambridge, MA: MIT Press.
Laplace, P.-S. (1951). A philosophical essay on probabilities (Trans. F. W. Trus-
cott & F. L. Emory). New York: Dover. (Original work published 1814)
Larkin, J., & Simon, H. (1987). Why a diagram is (sometimes) worth ten thou-
sand words. Cognitive Science, 11, 65-99.
Leary, D. E. (1987). From act psychology to probabilistic functionalism: The
place of Egon Brunswik in the history of psychology. In M. G. Ash & W. R.
Woodward (Eds.), Psychology in twentieth-century thought and society
(pp. 115—142). Cambridge, England: Cambridge University Press.
Legendre, G., Raymond, W., & Smolensky, P. (1993). Analytic typology of case
marking and grammatical voice. Proceedings of the Berkeley Linguistics
Society, 19, 464-478.
Legrenzi, P., & Murino, M. (1974). Falsification at the pre-operational level.
Italian Journal of Psychology, 1, 363-368.
Leibniz, G. W. (1951). The horizon of human doctrine. In P. P. Wiener (Ed.), Se-
lections (pp. 73-77). New York: Scribner's. (Original work published 1690)
Leibniz, G. W. (1951). Toward a universal characteristic. In P. P. Wiener (Ed.),
Selections (pp. 17-25). New York: Scribner's. (Original work published
1677)
Lenoir, T. (1986). Models and instruments in the development of electrophy-
316 REFERENCES

siology, 1845-1912. Historical Studies in the Physical and Biological Sci-

ences, 17, 1-54.
Lenoir, T. (1988). Practice, reason, context: The dialogue between theory and
experiment. Science in Context, 2, 3-22.
Leslie, A. (1994). ToMM, ToBY, and agency: Core architecture and domain
specificity. In L. A. Hirschfeld & S. A. Gelman (Eds.), Mapping the mind:
Domain specificity in cognition and culture (pp. 119-148). Cambridge, En-
gland: Cambridge University Press.
Levelt, W. J. M. (1989). Speaking: From intention to articulation. Cambridge,
MA: MIT Press.
Levi, I. (1983). Who commits the base rate fallacy? Behavioral and Brain Sci-
ences, 6, 502-506.
Levinson, S. (1992). How to think in order to speak Tzeltal. Unpublished man-
uscript, Max Planck Institute for Psycholinguistics, Cognitive Anthropol-
ogy Group, Nijmegen, Holland.
Lewontin, R. C. (1968). Evolution of complex genetic systems. In M. Gersten-
haber (Ed.), Some mathematical questions in biology (pp. 62-87). Provi-
dence, RI: American Mathematical Society.
Lichtenstein, S., & Fischhoff, B. (1977). Do those who know more also know
more about how much they know? The calibration of probability judg-
ments. Organizational Behavior and Human Performance, 20, 159-183.
Lichtenstein, S., & Fischhoff, B. (1981). The effects of gender and instruction
on calibration. Tech. Rep. No. PTR-1092-81-7. Eugene, OR: Decision Re-
search.
Lichtenstein, S., Fischhoff, B., & Phillips, L. D. (1982). Calibration of proba-
bilities: The state of the art to 1980. In D. Kahneman, P. Slovic, & A. Tver-
sky (Eds.), Judgment under uncertainty: Heuristics and biases (pp. 306-
334). Cambridge, England: Cambridge University Press.
Light, P., Girotto, V., & Legrenzi, P. (1990). Children's reasoning on conditional
promises and permissions. Cognitive Development, 5, 369-383.
Lindquist, E. F. (1940). Statistical analysis in educational research. Boston:
Houghton Mifflin.
Little, I. M. D. (1949). A reformulation of the theory of consumers' behavior.
Oxford Economic Papers, 1, 90-99.
Locke, J. (1959). An essay concerning human understanding (A. C. Fraser, Ed.).
New York: Dover. (Original work published 1690)
Loftus, G. R. (1991). On the tyranny of hypothesis testing in the social sciences.
Contemporary Psychology, 36, 102-104.
Lopes, L. L. (1981). Decision making in the short run. Journal of Experimental
Psychology: Human Learning and Memory, 7, 377-385.
Lopes, L. L. (1982). Doing the impossible: A note on induction and the expe-
rience of randomness. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 8, 626-636.
Lopes, L. L. (1991). The rhetoric of irrationality. Theory and Psychology, 1,65-82.
Lopes, L. L. (1992). Three misleading assumptions in the customary rethoric
of the bias literature. Theory and Psychology, 2, 231-236.
Lopes, L. L. (1994). Psychology and economics: Perspectives on risk, cooper-
ation, and the marketplace. Annual Review of Psychology, 45, 197-227.
Lopes, L. L. (1995). Algebra and process in the modeling of risky choice. In J. R.
Busemeyer, R. Hastie, & D. Medin (Eds.), Decision making from theperspec-
 REFERENCES 317

tive of cognitive psychology (pp. 177-220). New York: Academic Press.

Lovie, A. D. (1979). The analysis of variance in experimental psychology:
1934-1945. British Journal of Mathematical and Statistical Psychology, 32,
151-178.
Lovie, A. D. (1983). Attention and behaviorism—fact and fiction. British Jour-
nal of Psychology, 74, 301-310.
Lovie, A. D., & Lovie, P. (1986). The flat maximum effect and linear scoring
models for prediction. Journal of Forecasting, 5, 159—168.
Lubek, I., & Apfelbaum, E. (1987). Neo-behaviorism and the Garcia effect: A
social psychology of science approach to the history of a paradigm clash.
In M. Ash & W. Woodward (Eds.), Psychology in twentieth-century thought
and society (pp. 59-92). Cambridge, England: Cambridge University Press.
Luce, R. D. (1977). Thurstone's discriminal processes fifty years later. Psychom-
etrika, 42, 461-489.
Luce, R. D. (1988). The tools-to-theory hypothesis. Review of the book Cogni-
tion as intuitive statistics. Contemporary Psychology, 33, 582-583.
Luce, R. D. (1989). Autobiography. In G. Lindzey (Ed.), Psychology in autobi-
ography (Vol. 8, pp. 245-289). Stanford, CA: Stanford University Press.
Luchins, A. S., & Luchins, E. H. (1994). The water jar experiments and Ein-
stellung effects: I. Early history and surveys of textbook citations. Gestalt
Theory, 16, 101-121.
Lykken, D. T. (1968). Statistical significance in psychological research. Psycho-
logical Bulletin, 70, 151-159.
Lyon, D., & Slovic, P. (1976). Dominance of accuracy information and neglect
of base rates in probability estimation. Ada Psychologica, 40, 287-298.
MacGregor, D., & Slovic, P. (1986). Perceived acceptability of risk analysis as
a decision-making approach. Risk Analysis, 6, 245-256.
Manktelow, K. I., & Evans, J. St. B. T. (1979). Facilitation of reasoning by re-
alism: Effect or non-effect? British Journal of Psychology, 70, 477-488.
Mansson, S. A. (1990). Psycho-social aspects of HIV testing—the Swedish case.
AIDS Care, 2, 5-16.
Marr, D. (1982). Vision: A computational investigation into the human repre-
sentation and processing of visual information. San Francisco: Freeman.
Martignon, L., & Hoffrage, U. (1999). Why does one-reason decision making
work? A case study in ecological rationality. In G. Gigerenzer, P. M. Todd,
& the ABC Research Group, Simple heuristics that make us smart (pp.
119-140). New York: Oxford University Press.
Martignon, L., & Laskey, K. B. (1999). Bayesian benchmarks for fast and frugal
heuristics. In G. Gigerenzer, P. M. Todd, & the ABC Research Group, Sim-
ple heuristics that make us smart (pp. 169-188). New York: Oxford Uni-
versity Press.
Marzuk, P. M., & Perry, S. W. (1993). Suicide and HIV: Researchers and clini-
cians beware. AIDS Care, 5, 387-390.
Massaro, D. W. (1987). Speech perception by ear and eye. Hillsdale, NJ: Erl-
baum.
Massaro, D. W. (1988). Some criticisms of connectionist models of human per-
formance. Journal of Memory and Language, 27, 213-234.
Massaro, D. W. (1998). Perceiving talking faces. Boston: MIT Press.
Maurer, C., Kiehl, W., & Altmann, D. (1993). Zur Pravalenz und HIV-Inzidenz
bei Blutspendern in Baden-Wiirttemberg. In V. Kretschmer, W. Stangel, &
318 REFERENCES

R. Eckstein (Eds.), Transfusionsmedizin 1992/93. Beitrag der Infusions-

therapie, 31 (pp. 5-9). Freiburg, Germany: Karger.
May, R. S. (1986). Overconfidence as a result of incomplete and wrong knowl-
edge. In R. W. Scholz (Ed.), Current issues in West German decision re-
search (pp. 13-30). Frankfurt am Main, Germany: Lang.
May, R. S. (1987). Realismus von subjektiven Wahrscheinlichkeiten: Eine
kognitionspsychologische Analyse inferentieller Prozesse beim Over-
confidence-Phdnomen [Calibration of subjective probabilities: A cognitive
analysis of inference processes in overconfidencej. Frankfurt am Main,
Germany: Lang.
Mayseless, O., & Kruglanski, A. W. (1987). What makes you so sure? Effects of
epistemic motivations on judgmental confidence. Organizational Behavior
and Human Decision Processes, 39, 162-183.
McArthur, L. A. (1972). The how and what of why: Some determinants and
consequences of causal attribution. Journal of Personality and Social Psy-
chology, 22, 171-193.
McClelland, A. G. R., & Bolger, F. (1994). The calibration of subjective proba-
bilities: Theories and models 1980-1994. In G. Wright & P. Ayton (Eds.),
Subjective probability (pp. 453-482). Chichester, England: Wiley.
McClennen, E. F. (1990). Rationality and dynamic choice. Cambridge, England:
Cambridge University Press.
McCloskey, D. N. (1985). The rhetoric of economics. Madison: University of
Wisconsin Press.
McCorduck, P. (1979). Machines who think. San Francisco: Freeman.
McCulloch, W. S. (1965). Embodiments of mind. Cambridge, MA: MIT Press.
McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent
in nervous activity. Bulletin of Mathematical Biophysics, 5, 115-133.
McKenzie, C. R. (1994). The accuracy of intuitive judgment strategies: Covar-
iation assessment and Bayesian inference. Cognitive Psychology, 26, 209—
239.
Medin, D. L., Wattenmaker, W. D., & Michalski, R. S. (1987). Constraints and
preferences in inductive learning: An experimental study of human and
machine performance. Cognitive Science, 11, 299-339.
Meehl, P. E. (1978). Theoretical risks and tabular asterisks: Sir Karl, Sir Ronald,
and the slow progress of soft psychology. Journal of Consulting and Clin-
ical Psychology, 46, 806-834.
Melton, A. W. (1962). [Editorial]. Journal of Experimental Psychology, 64, 553-
557.
Michotte, A. (1963). The perception of causality. London: Methuen. (Original
work published 1946)
Miller, G. A., & Buckhout, R. (1973). Psychology: The science of mental life.
New York: Harper & Row.
Miller, G. A., Galanter, E., & Pribram, K. H. (1960). Plans and the structure of
behavior. New York: Holt, Rinehart & Winston.
Miller, G. F., & Todd, P. M. (1995). The role of mate choice in biocomputation:
Sexual selection as a process of search, optimization, and diversification.
In W. Banzhaf & F. H. Eeckman (Eds.), Evolution and biocomputation:
Computational models of evolution (pp. 169-204). Berlin: Springer.
Miller, G. F., & Todd, P. M. (1998). Mate choice turns cognitive. Trends in Cog-
nitive Sciences, 2, 190-198.
 REFERENCES 319

Millikan, R. (1984). Language, thought, and other biological categories. Cam-

bridge, MA: MIT Press.
Mineka, S., & Cook, M. (1988). Social learning and the acquisition of snake
fear in monkeys. In T. R. Zentall & B. G. Galef (Eds.), Social learning: Psy-
chological and biological perspectives (pp. 51-73). Hillsdale, NJ: Erlbaum.
Mises, R. von (1957). Probability, statistics, and truth. London: Allen & Unwin.
(Original work published 1928)
Mitchell, S. D., Daston, L., Gigerenzer, G., Sesardic, N., & Sloep, P. (1997). In
the service of pluralism and interdisciplinarity. In P. Weingart, P. Richer-
son, S. D. Mitchell, & S. Maasen (Eds.), Human by nature: Between biology
and the social sciences (pp. 103-150). Hillsdale, NJ: Erlbaum.
Morrison, D. E., & Henkel, R. E. (Eds.). (1970). The significance test contro-
versy. Chicago: Aldine.
Mueser, P. R., Cowan, N., & Mueser, K. T. (1999). A generalized signal detection
model to predict rational variation in base rate use. Cognition, 69, 267-
312.
Murdock, B. B., Jr. (1982). A theory for the storage and retrieval of item and
associative information. Psychological Review, 89, 609-626.
Nesse, R. M., & Williams, G. C. (1995). Why we get sick: The new science of
Darwinian medicine. New York: Vintage Books.
Neumann, J. von (1958). The computer and the brain. New Haven, CT: Yale
University Press.
Newcombe, R. D., & Rutter, D. R. (1982). Ten reasons why ANOVA theory and
research fail to explain attribution processes: I. Conceptual problems. Cur-
rent Psychological Reviews, 2, 95-108.
Newell, A. (1973). You can't play 20 questions with nature and win: Projective
comments on the papers of this symposium. In W. G. Chase (Ed.), Visual
information processing (pp. 283-308). New York: Academic Press.
Newell, A., Shaw, J. C., & Simon, H. A. (1958). Elements of a theory of human
problem solving. Psychological Review, 65, 151-166.
Newell, A., & Simon, H. A. (1972). Human problem solving. Englewood Cliffs,
NJ: Prentice-Hall.
Neyman, J. (1937). Outline of a theory of statistical estimation based on the
classical theory of probability. Philosophical Transactions ot the Royal
Society of London, Series A, 236, 333-380.
Neyman, J. (1950). First course in probability and statistics. New York: Holt.
Neyman, J. (1957). Inductive behavior as a basic concept of philosophy of
science. International Statistical Review, 25, 7-22.
Neyman, J. (1977). Frequentist probability and frequentist statistics. Synthese,
36, 97-131.
Neyman, J., & Pearson, E. S. (1928). On the use and interpretation of certain
test criteria for purposes of statistical inference: Part 1. Riometrika, 20A,
175-240.
Nickles, T. (1980). Introductory essay: Scientific discovery and the future of
philosophy of science. In T. Nickles (Ed.), Scientific discovery, logic, and
rationality (pp. 1-59). Dordrecht, Holland: Reidel.
Nisbett, R. E., & Borgida, E. (1975). Attribution and the psychology of predic-
tion. Journal of Personality and Social Psychology, 32, 932-943.
Nisbett, R. E., & Ross, L. (1980). Human inference: Strategies and shortcomings
of social judgment. Englewood Cliffs, NJ: Prentice Hall.
320 REFERENCES

Nisbett, R. E., & Wilson, T. D. (1977). Telling more than we can know: Verbal
reports on mental processes. Psychological Review, 84, 231-259.
Nunnally, J. C. (1975). Introduction to statistics for psychology and education.
New York: McGraw-Hill.
Oakes, M. (1986). Statistical inference: A commentary for the social and the
behavioral sciences. Chichester, England: Wiley.
Oaksford, M., & Chater, N. (1994). A rational analysis of the selection task as
optimal data selection. Psychological Review, 101, 608-631.
Over, D. E., & Manktelow, K. I. (1993). Rationality, utility and deontic reason-
ing. In K. I. Manktelow & D. E. Over (Eds.), Rationality: Psychological and
philosophical perspectives (pp. 231-259). London: Routledge.
Packer, C. (1977). Reciprocal altruism in Papio annubis. Nature, 265, 441-443.
Parducci, A. (1965). Category judgment: A range-frequency model. Psycholog-
ical Review, 72, 407-418.
Paulos, J. A. (1988). Innumeracy: Mathematical illiteracy and its consequences.
New York: Vintage Books.
Payne, J. W., Bettman, J. R., & Johnson, E. J. (1993). The adaptive decision
maker. New York: Cambridge University Press.
Pearson, E. S. (1939). "Student" as statistician. Biometrika, 30, 210-250.
Pearson, E. S. (1962). Some thoughts on statistical inference. Annals of Math-
ematical Statistics, 33, 394-403.
Peichl-Hoffman, G. (1991). Spezifitatsprobleme bei der Testung auf Anti-HIVl
bzw. Anti-HIV2 in der Routineuntersuchung von Blutspendern. Klinisches
Labor, 37, 320-328.
Peirce, C. S., & Jastrow, J. (1884). On small differences of sensation. Memoirs
of the National Academy of Sciences, 3, 75-83.
Peterson, C. R., DuCharme, W. M., & Edwards, W. (1968). Sampling distribu-
tions and probability revision. Journal of Experimental Psychology, 76,
236-243.
Phelps, R. H., & Shanteau, J. (1978). Livestock judges: How much information
can an expert use? Organizational Behavior and Human Performance, 21,
209-219.
Phillips, L. D., & Edwards, W. (1966). Conservatism in a simple probability
model inference task. Journal of Experimental Psychology, 72, 346-354.
Piaget, J. (1930). The child's conception of causality. London: Kegan Paul,
Trench, & Trubner.
Piaget, J., & Inhelder, B. (1975). The origin of the idea of chance in children.
New York: Norton. (Original work published 1951)
Piattelli-Palmarini, M. (1991, March/April). Probability blindness: Neither ra-
tional nor capricious. Bostonia, 28—35.
Piattelli-Palmarini, M. (1994). Inevitable illusions: How mistakes of reason rule
our minds. New York: Wiley.
Pinker, S. (1994). The language instinct. London: Penguin Press.
Platt, R., & Griggs, R. (1993). Darwinian algorithms and the Wason selection
task: A factorial analysis of social contract selection task problems. Cog-
nition, 48, 163-192.
Politzer, G., & Nguyen-Xuan, A. (1992). Reasoning about promises and warn-
ings: Darwinian algorithms, mental models, relevance judgments or prag-
matic schemas? Quarterly Journal of Experimental Psychology, 44A, 402-
421.
 REFERENCES 321

Pollard, P. (1982). Human reasoning: Some possible effects of availability. Cog-

nition, 12, 65-96.
Pollard, P., & Evans, }. St. B. T. (1983). The role of "representativeness" in
statistical inference: A critical appraisal. In J. St. B. T. Evans (Ed.), Think-
ing and reasoning: Psychological approaches (pp. 107-134). London:
Routledge & Kegan Paul.
Pollard, P., & Richardson, J. T. E. (1987). On the probability of making Type I
errors. Psychological Bulletin, 102, 159-163.
Popper, K. R. (1959). The logic of scientific discovery. New York: Basic Books.
(Original work published 1935)
Porter, T. (1986). The rise of statistical thinking, 1820-1900. Princeton, NJ:
Princeton University Press.
Premack, D. (1990). The infant's theory of self-propelled objects. Cognition, 36,
1-16.
Premack, D., & Premack, A. J. (1994). Moral belief: Form versus content. In
L. A. Hirschfeld & S. A. Gelman (Eds.), Mapping the mind: Domain spe-
cificity in cognition and culture (pp. 149-168). Cambridge, England: Cam-
bridge University Press.
Prince, A., & Smolensky, P. (1991). Notes on connectionism and harmony the-
ory in linguistics. Tech. Rep. No. CU-CS-533-91. Boulder: University of
Colorado, Department of Computer Science.
Pugh, K., O'Donnell, I., & Catalan, J. (1993). Suicide and HIV disease. AIDS
Care, 5, 391-400.
Pylyshyn, Z. W. (1973). What the mind's eye tells the mind's brain: A critique
of mental imagery. Psychological Bulletin, 80, 1-24.
Quetelet, L.AJ. (1969). A treatise on man and the development of his faculties
(Trans. R. Knox). Gainesville, FL: Scholars' Facsimiles and Reprints. (Orig-
inal work published 1842)
Regier, T. (1996). The human semantic potential: Spatial language and con-
strained connectionism. Cambridge, MA: MIT Press.
Reichenbach, H. (1938). Experience and prediction. Chicago: University of Chi-
cago Press.
Rieskamp, J., & Hoffrage, U. (1999). When do people use simple heuristics and
how can we tell? In G. Gigerenzer, P. M. Todd, & the ABC Research Group,
Simple heuristics that make us smart (pp. 141-167). New York: Oxford
University Press.
Ronis, D. L., & Yates, J. F. (1987). Components of probability judgment accu-
racy: Individual consistency and effects of subject matter and assessment
method. Organizational Behavior and Human Decision Processes, 40,
193-218.
Rosch, E. (1978). Principles of categorization. In E. Rosch & B. B. Lloyd (Eds.),
Cognition and categorization (pp. 27-48). Hillsdale, NJ: Erlbaum.
Rosnow, R. L., & Rosenthal, R. (1989). Statistical procedures and the justifica-
tion of knowledge in psychological science. American Psychologist, 44,
1276-1284.
Rouanet, H. (1961). Etudes de decisions experimentales et calcul de probabil-
ites [Studies of experimental decision making and the probability calcu-
lus]. In Colloques Internationaux du Centre National de la Recherche
Scientifique (pp. 33-43). Paris: Editions du Centre National de la Recher-
che Scientifique.
322 REFERENCES

Rozeboom, W. W. (1960). The fallacy of the null hypothesis significance test.

Psychological Bulletin, 57, 416-428.
Rucci, A. J., & Tweney, R. D. (1980). Analysis of variance and the "second
discipline" of scientific psychology: A historical account. Psychological
Bulletin, 87, 166-184.
Sahlin, N. (1991). Baconian inductivism in research on human decision-
making. Theory and Psychology, 1, 431-450.
Sahlin, N. (1993). On higher order beliefs. In J. Dubucs (Ed.), Philosophy of
probability (pp. 13-34). Dordrecht, Holland: Kluwer.
Saks, M., & Kidd, R. F. (1980). Human information processing and adjudica-
tion: Trial by heuristics. Law and Society Review, 15, 123-160.
Samuels, R., Stich, S., & Bishop, M. (in press). Ending the rationality wars:
How to make disputes about human rationality disappear. In R. Elio (Ed.),
Vancouver studies in cognitive science: Vol. 11. Common sense, reasoning
and rationality. New York: Oxford University Press.
Samuelson, P. A. (1938). A note on the pure theory of consumers' behavior.
Economica, 5, 61-71.
Savage, L. J. (1954). The foundations of statistics. New York: Wiley.
Schaefer, R. E. (1976). The evaluation of individual and aggregated subjective
probability distributions. Organizational Behavior and Human Perfor-
mance, 17, 199-210.
Schaffer, S. (1992). Disappearing acts: On Gigerenzer's "Where do new ideas
come from?" Unpublished manuscript.
Schaffer, S. (1994). Babbage's intelligence: Calculating engines and the factory
system. Critical Inquiry, 21, 203-227.
Schering AG (1992). AIDS Information. Berlin: Schering AG.
Scholz, R. W. (Ed.) (1983). Decision making under uncertainty. Amsterdam:
North-Holland.
Scholz, R. W. (1987). Cognitive strategies in stochastic thinking. Dordrecht,
Holland: Reidel.
Schonemann, P. H. (1983). Some theory and results for metrics for bounded
response scales, /ournal of Mathematical Psychology, 72, 407-418.
Schrage, G. (n. d.). Schwierigkeiten mit der stochastischen Modellbildung—
zwei Beispiele aus der Praxis. [Problems with stochastic models—two ex-
amples from real life]. Unpublished manuscript.
Schum, D. A. (1994). The evidential foundations of probabilistic reasoning.
New York: Wiley.
Schwartz, S., & Dangleish, L. (1982). Statistical inference in personality re-
search. Journal of Research in Personality, 16, 290-302.
Schwartz, S., Kinosian, B. P., Pieskalla, W. P., & Lee, H. (1990). Strategies for
screening blood for human immunodeficiency virus antibody: Use of a
decision support system. Journal of the American Medical Association,
264, 1704-1710.
Schwarz, N., Strack, F., Hilton, D., & Naderer, G. (1991). Base rates, represen-
tativeness and the logic of conversation: The contextual relevance of "ir-
relevant" information. Social Cognition, 9, 67-84.
Searle, J. (1984). Minds, brains and science. Cambridge, MA: Harvard Univer-
sity Press.
Sedlmeier, P. (1997). BasicBayes: A tutor system for simple Bayesian inference.
Behavior Research Methods, Instruments, and Computers, 29, 328-336.
 REFERENCES 323

Sedlmeier, P., & Gigerenzer, G. (1989). Do studies of statistical power have an

effect on the power of studies? Psychological Bulletin, 105, 309-316.
Sedlmeier, P., & Gigerenzer, G. (in press). Teaching Bayesian reasoning in less
than two hours. Journal of Experimental Psychology: General.
Sedlmeier, P., Hertwig, R., & Gigerenzer, G. (1998). Are judgments of the po-
sitional frequencies of letters systematically biased due to availability?
Journal of Experimental Psychology: Learning, Memory, and Cognition, 24
(3), 754-770.
Seligman, M. E. P., & Hager, J. L. (Eds.). (1972). Biological boundaries of learn-
ing. New York: Appleton-Century-Crofts.
Sen, A. (1993). Internal consistency of choice. Econometrica, 61, 495-521.
Shafer, G. (1976). A mathematical theory of evidence. Princeton, NJ: Princeton
University Press.
Shafer, G. (1978). Non-additive probabilities in the work of Bernoulli and Lam-
bert. Archive for the History of Exact Sciences, 19, 309-370.
Shannon, C. (1940). Symbolic analysis of relay and switching circuits. M.S.
thesis, MIT.
Shanteau, J. (1989). Cognitive heuristics and biases in behavioral auditing: Re-
view, comments and observations. Accounting Organizations and Society,
14, 165-177.
Shaughnessy, J. M. (1992). Research on probability and statistics: Reflections
and directions. In D. A. Grouws (Ed.), Handbook of research on mathe-
matical teaching and learning (pp. 465—494). New York: Macmillan.
Shepard, R. N. (1962). The analysis of proximities: Part 2. Psychometrika, 27,
219-246.
Shepard, R. N. (1967). On subjectively optimum selections among multi-
attribute alternatives. In W. Edwards & A. Tversky (Eds.), Decision making
(pp. 257-283). Baltimore: Penguin Books.
Shepard, R. N. (1974). Representation of structure in similarity data: Problems
and prospects. Psychometrika, 39, 373-421.
Shepard, R. N. (1990). Mind sights. New York: Freeman.
Shepard, R. N. (1992). The perceptual organization of colors: An adaptation to
regularities of the terrestrial world? In J. H. Barkow, L. Cosmides, & J.
Tooby (Eds.), The adapted mind: Evolutionary psychology and the gen-
eration of culture (pp. 495-532). New York: Oxford University Press.
Sherman, S. J., Judd, C. M., & Park, B. (1989). Social cognition. Annual Review
of Psychology, 40, 281-326.
Simon, H. A. (1945). Administrative behavior: A study of decision-making pro-
cesses in administrative organization. New York: Free Press.
Simon, H. A. (1956). Rational choice and the structure of the environment.
Psychological Review, 63, 129-138.
Simon, H. A. (1957). Models of man: Social and rational. New York: Wiley.
Simon, H. A. (1969). The sciences of the artificial. Cambridge, MA: MIT Press.
Simon, H. A. (1973). Does scientific discovery have a logic? Philosophy of Sci-
ence, 40, 471-480.
Simon, H. A. (1979). Information processing models of cognition. Annual Re-
view of Psychology, 30, 363-396.
Simon, H. A. (1982). Models of bounded rationality. Cambridge, MA: MIT
Press.
Simon, H. A. (1986). Rationality in psychology and economics. In R. Hogarth
324 REFERENCES

& M. Reder (Eds.), Rational choice: The contrast between economics and
psychology (pp. 25-40). Chicago: University of Chicago Press.
Simon, H. A. (1987). Bounded rationality. In J. Eatwell, M. Milgate, & P. New-
man (Eds.), The New Palgrave: A dictionary of economics (pp. 266-268).
London: Macmillan.
Simon, H. A. (1990). Invariants of human behavior. Annual Review of Psy-
chology, 41, 1-19.
Simon, H. A. (1991). Models of my life. New York: Basic Books.
Simon, H. A. (1992a). What is an "explanation" of behavior? Psychological
Science, 3, 150-161.
Simon, H. A. (1992b). Economics, bounded rationality, and the cognitive rev-
olution. Aldershot Hants, England: Elgar.
Simon, H. A., & Kulkarni, D. (1988). The processes of scientific discovery: The
strategy of experimentation. Cognitive Science, 12, 139-175.
Simon, H. A., & Newell, A. (1986). Information processing language V on the
IBM 650. Annals of the History of Computing, 8, 47-49.
Sivyer, M., & Finlay, D. (1982). Perceived duration of auditory sequences. Jour-
nal of General Psychology, 107, 209-217.
Skinner, B. F. (1972). Cumulative record. New York: Appleton-Century-Crofts.
Skinner, B. F. (1984). A matter of consequences. New York: New York Univer-
sity Press.
Sloman, S. A. (1996). The empirical case for two systems of reasoning. Psy-
chological Bulletin, 119, 3-22.
Slovic, P., Fischhoff, B., & Lichtenstein, S. (1976). Cognitive processes and
societal risk taking. In J. S. Carroll & J. W. Payne (Eds.), Cognition and
social behavior (pp. 165-184). Hillsdale, NJ: Erlbaum.
Smith, E. E., & Osherson, D. N. (1989). Similarity and decision making. In S.
Vosniadou & A. Ortony (Eds.), Similarity and analogical reasoning
(pp. 60-75). Cambridge, England: Cambridge University Press.
Smith, L. D. (1986). Behaviorism and logical positivism. Stanford, CA: Stanford
University Press.
Snell, J. J. S., Supran, E. M., & Tamashiro, H. (1992). WHO international qual-
ity assessment scheme for HIV antibody testing: Results from the second
distribution of sera. Bulletin of the World Health Organization, 70, 605-
613.
Sniezek, J. A., & Buckley, T. (1993). Becoming more or less uncertain. In N. J.
Castellan (Ed.), Individual and group decision making (pp. 87-108). Hills-
dale, NJ: Erlbaum.
Sperber, D. (1994). The modularity of thought and the epidemiology of rep-
resentations. In L. A. Hirschfeld & S. A. Gelman (Eds.), Mapping the mind:
Domain specificity in cognition and culture (pp. 39-67). Cambridge, En-
gland: Cambridge University Press.
Sperber, D., Cara, F., & Girotto, V. (1995). Relevance theory explains the selec-
tion task. Cognition, 57, 31-95.
Sperber, D., & Wilson, D. (1986). Relevance: Communication and cognition.
Oxford, England: Blackwell.
Spettell, C. M., & Liebert, R. M. (1986). Training for safety in automated person-
machine systems. American Psychologist, 41, 545-550.
Spielberg, F., Kabeya, C. M., Ryder, R. W., Kifuani, N. K., Harris, J., Bender,
T. R., Heyward, W. L., & Quinn, T. C. (1989). Field testing and comparative
 REFERENCES 325

evaluation of rapid visually read screening essays for antibody to human

immunodeficiency virus. Lancet, 1, 580-584.
Statistisches Bundesamt. (1986). Statistisches Jahrbuch 1986 fur die Bundes-
republik Deutschland [Statistical yearbook 1986 for the Federal Republic
of Germany]. Stuttgart, Germany: Kohlhammer.
Statistisches Bundesamt (1994). Statistisches Jahrbuch 1994 fur die Bundes-
republik Deutschland. Wiesbaden, Germany: Statistisches Bundesamt.
Stephens, D. W., & Krebs, J. R. (1986). Foraging theory. Princeton, NJ: Princeton
University Press.
Sterling, T. D. (1959). Publication decisions and their possible effects on infer-
ences drawn from tests of significance—or vice versa. Journal of the Amer-
ican Statistical Association, 54, 30-34.
Stich, S. P. (1990). The fragmentation of reason. Cambridge, MA: MIT Press.
Stigler, J. W. (1984). The effect of abacus training on Chinese children's mental
calculation. Cognitive Psychology, 16, 145-176.
Stigler, S. M. (1983). Who discovered Bayes's theorem? American Statistican,
37, 290-196.
Stine, G. J. (1996). Acquired immune deficiency syndrome: Biological, medical,
social, and legal issues (2nd ed.). Englewood Cliffs, NJ: Prentice Hall.
Stone, G. O. (1986). An analysis of the delta rule and the learning of statistical
associations. In D. Rumelhart, J. McClelland, & the PDF Research Group
(Eds.), Parallel distributed processing: Explorations in the microstructure
of cognition (pp. 444-459). Cambridge, MA: MIT Press.
Strack, F. (1988). Social cognition: Sozialpsychologie innerhalb des Paradigmas
der Informationsverarbeitung. Psychologische Rundschau, 39, 72-82.
"Student" [W. S. Cosset]. (1908). The probable error of a mean. Biometrika, 6,
1-25.
Swets, J. A., Tanner, W. P., Jr., & Birdsall, T. G. (1964). Decision processes in
perception. In J. A. Swets (Ed.), Signal detection and recognition in human
observers (pp. 3-57). New York: Wiley.
Swijtink, Z. G. (1987). The objectification of observation: Measurement and
statistical methods in the nineteenth century. In L. Kriiger, L. Daston, &
M. Heidelberger (Eds.), The probabalistic revolution: Vol. 1. Ideas in his-
tory (pp. 261-285). Cambridge, MA: MIT Press.
Tanner, W. P., Jr. (1965). Statistical decision processes in detection and recog-
nition (Technical Report). Ann Arbor: University of Michigan, Sensory In-
telligence Laboratory, Department of Psychology.
Tanner, W. P., Jr., & Swets, J. A. (1954). A decision-making theory of visual
detection. Psychological Review, 61, 401-409.
Teigen, K. H. (1974). Overestimation of subjective probabilities. Scandinavian
Journal of Psychology, 15, 56-62.
Teigen, K. H. (1983). Studies in subjective probability: IV. Probabilities, confi-
dence, and luck. Scandinavian Journal of Psychology, 24, 175-191.
Terman, L. M. (1916). The measurement of intelligence. Boston: HoughtonMif-
flin.
Terman, L. M., & Merrill, M. A. (1937). Measuring intelligence. Boston: Hough-
ton Mifflin.
Tetlock, P. E. (1992). The impact of accountability on judgment and choice: To-
ward a social contingency model. In M. Zanna (Ed.}, Advances in experimen-
tal social psychology (Vol. 25, pp. 331-376). New York: Academic Press.
326 REFERENCES

Thaler, R. H. (1991). Quasi rational economics. New York: Russell Sage Foun-
dation.
Thines, G., Costall, A., & Butterworth, G. (Eds.). (1991). Michotte's experimen-
tal phenomenology of perception. Hillsdale, NJ: Erlbaum.
Thomas, D. H. (1978). The awful truth about statistics in archaeology. Ameri-
can Antiquity, 43, 231-244.
Thorndike, R. L. (1954). The psychological value system of psychologists.
American Psychologist, 9, 787-789.
Thurstone, L. L. (1927). A law of comparative judgment. Psychological Review,
34, 273-286.
Titchener, E. B. (1896). An outline of psychology. New York: Macmillan.
Todd, P. M. (2000a). Fast and frugal heuristics for environmentally bounded
minds. In G. Gigerenzer & R. Selten (Eds.), Bounded rationality: The adap-
tive toolbox (pp. 51-70). Cambridge, MA: MIT Press.
Todd, P. M. (2000b). The ecological rationality of mechanisms evolved to make
up minds. American Behavioral Scientist, 43, 940-956.
Tooby, ]., & Cosmides, L. (1992). The psychological foundations of culture. In
J. Barkow, L. Cosmides, & J. Tooby (Eds.), The adapted mind: Evolutionary
psychology and the generation of culture (pp. 19-136). New York: Oxford
University Press.
Tooby, J., & DeVore, I. (1987). The reconstruction of hominid behavioral evo-
lution through strategic modeling. In W. G. Kinzey (Ed.), The evolution of
human behavior: Primate models (pp. 183-237). Albany, NY: State Uni-
versity of New York Press.
Toulmin, S., & Leary, D. E. (1985). The cult of empiricism in psychology, and
beyond. In S. Koch & D. E. Leary (Eds.), A century of psychology as science
(pp. 594-617). New York: McGraw-Hill.
Tu, X. T., Litvak, E., & Pagano, M. (1992). Issues in human immunodeficiency
virus (HIV) screening programs. American Journal of Epidemiology, 136,
244-255.
Tukey, J. W. (1977). Exploratory data analysis. Reading, MA: Addison-Wesley.
Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59, 433-
460.
Turing, A. M. (1969). Intelligent machinery. In B. Meltzer & D. Michie (Eds.),
Machine intelligence (Vol. 5, pp. 3-23). Edinburgh, Scotland: Edinburgh
University Press. (Original work published 1947)
Tversky, A. (1977). Features of similarity. Psychological Review, 84, 327-352.
Tversky, A., & Kahneman, D. (1971). Belief in the law of small numbers. Psy-
chological Bulletin, 76, 105-110.
Tversky, A., & Kahneman, D. (1973). Availability: A heuristic for judging fre-
quency and probability. Cognitive Psychology, 5, 207-232.
Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty: Heuristics
and biases. Science, 185, 1124-1131.
Tversky, A., & Kahneman, D. (1980). Causal schemata in judgments under un-
certainty. In M. Fishbein (Ed.), Progress in social psychology (Vol. 1,
pp. 49-72). Hillsdale, NJ: Erlbaum.
Tversky, A., & Kahneman, D. (1982a). Judgments of and by representativeness.
In D. Kahneman, P. Slovic, & A. Tversky (Eds.), Judgment under uncer-
tainty: Heuristics and biases (pp. 84-98). Cambridge, England: Cambridge
University Press.
 REFERENCES 327

Tversky, A., & Kahneman, D. (1982b). Evidential impact of base rates. In D.

Kahneman, P. Slovic, & A. Tversky (Eds.), Judgment under uncertainty:
Heuristics and biases (pp. 153-160). Cambridge, England: Cambridge Uni-
versity Press.
Tversky, A., & Kahneman, D. (1983). Extensional versus intuitive reasoning:
The conjunction fallacy in probability judgment. Psychological Review,
90, 293-315.
Tweney, R. D., Doherty, M. E., & Mynatt, C. R. (Eds.). (1981). On scientific
thinking. New York: Columbia University Press.
Tweney, R. D., & Walker, B. J. (1990). Science education and the cognitive psy-
chology of science. In B. F. Jones & L. Idol (Eds.), Dimensions of thinking
and cognitive instruction (pp. 291-310). Hillsdale, NJ: Erlbaum.
Wanke, M., Schwarz, N., & Bless, H. (1995). The availability heuristic revisited:
Experienced ease of retrieval in mundane frequency estimates. Acta Psy-
chologica, 89, 83-90.
Wallach, L., & Wallach, M. A. (1994). Gergen versus the mainstream: Are hy-
potheses in social psychology subject to empirical test? Journal of Person-
ality and Social Psychology. 67, 233-242.
Wallach, M. A., & Wallach, L. (1998). When experiments serve little purpose:
Misguided research in mainstream psychology. Theory and Psychology, 8,
183-194.
Wallendael, L. R. van, & Hastie, R. (1990). Tracing the footsteps of Sherlock
Holmes: Cognitive representations of hypothesis testing. Memory and Cog-
nition, 18, 240-250.
Wallsten, T. S. (1983). The theoretical status of judgmental heuristics. In R. W.
Scholz (Ed.), Decision making under uncertainty (pp. 21-37). Amsterdam:
Elsevier.
Ward, J. W. (1994). Testing for human retrovirus infections: Medical indica-
tions and ethical considerations. In G. Schochetman & J. R. George (Eds.),
AIDS testing: A comprehensive guide to technical, medical, social, legal,
and management issues (pp. 1-14). New York: Springer.
Wason, P. C. (1966). Reasoning. In B. M. Foss (Ed.), New horizons in psychol-
ogy (pp. 135—151). Harmondsworth, England: Penguin.
Wason, P. C. (1968). Reasoning about a rule. Quarterly Journal of Experimental
Psychology, 20, 273-281.
Wason, P. C. (1981). The importance of cognitive illusions. Behavioral and
Brain Sciences, 4, 356.
Wason, P. C., & Johnson-Laird, P. N. (1970). A conflict between selecting and
evaluating information in an inferential task. British Journal of Psychology,
61 (4), 509-515.
Weber, U., Bockenholt, U., Hilton, D. J., & Wallace, B. (1993). Determinants of
diagnostic hypothesis generation: Effects of information, base rates, and
experience. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 19, 1151-1164.
Wells, G. L., & Harvey, J. H. (1977). Do people use consensus information in
making causal attributions? Journal of Personality and Social Psychology,
35, 279-293.
Wells, G. L., & Harvey, J. H. (1978). Naive attributors' attributions and predic-
tions: What is informative and when is an effect an effect? Journal of Per-
sonality and Social Psychology, 36, 483-490.
328 REFERENCES

Wendt, D. (1966). Versuche zur Erfassung eines subjektiven Verlasslichkeits-

niveaus. Zeitschrift fur Psychologic, 172, 40-81.
Whitehead, A. N., & Russell, B. (1935). Principle mathematica (2nd ed., Vol.
1) Cambridge, England: Cambridge University Press.
Whiten, A., & Byrne, R. W. (1988). The Machiavellian intelligence hypotheses
[Editorial]. In R. Byrne & A. Whiten (Eds.), Machiavellian intelligence
(pp. 1-9). Oxford, England: Clarendon Press.
Wickelgren, W. A., & Norman, D. A. (1966). Strength models and serial posi-
tion in short-term recognition memory. Journal of Mathematical Psychol-
ogy, 3, 316-347.
Wilber, J. C. (1991). New development in diagnosing infections. In P. Volbering &
M. A. Jacobson (Eds.), AIDS clinical review (pp. 1-15). New York: Dekker.
Wilkinson, G. S. (1990, February). Food sharing in vampire bats. Scientific
American, 76-82.
Williams, G. C. (1966). Adaption and natural selection: A critique of some
current evolutionary thought. Princeton, NJ: Princeton University Press.
Wimsatt, W. C. (1976). Reductionism, levels of organization, and the mind-
body problem. In G. G. Globus, G. Maxwell, & I. Savodnik (Eds.), Con-
sciousness and the brain: A scientific and philosophical inquiry (pp. 199-
267). New York: Plenum.
Wimsatt, W. C. (1986). Developmental constraints, generative entrenchment,
and the innate-acquired distinction. In P. W. Bechtel (Ed.), Integrating sci-
entific disciplines (pp. 185-208). Dordrecht, Holland: Martinus-Nijhoff.
Windeler, J., & Kobberling, J. (1986). Empirische Untersuchung zur Einschat-
zung diagnostischer Verfahren am Beispiel des Haemoccult-Tests [An
empirical study of the judgments about diagnostic procedures using the
example of the hemoccult test]. Klinische Wochenschrift, 64, 1106-1112.
Winer, B. J. (1971). Statistical principles in experimental design. New York:
McGraw-Hill.
Winterfeldt, D. von, & Edwards, W. (1982). Costs and payoffs in perceptual
research. Psychological Bulletin, 91, 609-622.
Winterfeldt, D. von, & Edwards, W. (1986). Decision analysis and behavioral
research. Cambridge, England: Cambridge University Press.
Wise, M. N. (1988). Mediating machines. Science in Context, 2, 77-113.
Wittkowski, K. (1989). Wann ist ein HIV Test indiziert? Schlusswort. Deutsches
Arzteblatt, 86, B-138-140.
Woodworth, R. S. (1938). Experimental psychology. New York: Holt, Rinehart,
& Winston.
Woodworth. R. S., & Schlosberg, H. (1954). Experimental psychology. New
York: Holt, Rinehart, & Winston.
World Health Organization. (1996). Global AIDS statistics. AIDS Care, 8, 501.
Zacks, R. T., Hasher, L., & Sanft, H. (1982). Automatic encoding of event fre-
quency: Further findings. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 8, 106-116.
Zimmer, A. C. (1983). Verbal vs. numerical processing by subjective probabil-
ities. In R. W. Scholz (Ed.), Decision making under uncertainty (pp. 159-
182). Amsterdam: North-Holland.
Zimmer, A. C. (1986). What uncertainty judgments can tell about the under-
lying subjective probabilities. Uncertainty in Artificial Intelligence, 249-
252.
 NAME INDEX

Aaronson, D., 39 Axelrod, R., 213-214 Blackwell, R. J., 24

Aaronson, M., 39 Ayton, P., 290 Blank, H., 20, 116, 256-
ABC Research Group, 257, 259
58, 125, 197 Baars, B. J., 34 Bless, H., 291
Abelson, R., 37 Babbage, C., 1, 27, 28- Block, R. A., 64
Acree, M. C., 279n, 284 29, 32, 42 Blumer, C., 150
Adler, J. E., 224, 265 Bacon, R, 4 Bockenholt, U., 188
Ajzen, I., 9, 93, 110, Bakan, D., 284-285 Boesch, C., 226
244, 251, 254, 261 Bar-Hillel, M., 93, 98, Boesch, H., 226
Alba, J. W., 181 243 Bolger, R, 174
Allais, M., 194 Barnard, G. A., 19, 250, Boltzmann, L., 23, 267,
Allport, D. A., 163 281 295
Allwood, C. M., 132, Baron-Cohen, S., 226 Boole, G., 35n, 92, 166
157 Barsalou, L. W., 64 Borges, B., 174
Anastasi, A., 285 Bartlett, R, 34, 40, 267, Borgida, E., 93, 242,
Andersen, H.-C., 238 269 251, 259
Anderson, J. R., 10, 14, Bayes, T., 60, 98, 100, Boring, E., G., 46-47
24, 166, 233 103, 122, 271, 278 Bower, B., 295
Anderson, N. H., 151, Bazerman, M. H., 242- Bradshaw, G. L., 22, 35
283 243, 246 Braithwaite, R. B., 3
Apfelbaum, E., 212 Beach, L. R., 52, 121 Brandstadter, J., 289
Arbib, M. A., 35 Bernoulli, J., 35n, 201 Erase, G. L., 64
Arbuthnot, J., 10, 278 Berretty, P. M., 197 Brehmer, A., 196
Aristotle, 8, 227 Bettman, J. R., 181 Brehmer, B., 12, 138,
Arkes, H. R., 138, 150, Bienenstock, E., 293 164, 166, 196
246 Binet, A., 235n Breiman, L., 191, 197
Armelius, B., 164, 194 Birdsall, T. G., 7 Brekke, N., 93, 242,
Armelius, K., 138, 164, Birnbaum, M. H., 8, 16, 251
194 262, 264 Brewer, J. K., 284
Ash, M., 45 Bischof, W. R, 259 Broder, A., 174
Aspray, W., 30 Bishop, M., 237 Brown, A. L., 192
Atran, S., 230 Bjorkman, M., 164, 264 Brown, R L., 285

329
330 NAME INDEX

Brown, N. R., 169 Condorcet, M. J. A. N., DiPonzo, N., 174

Brown, W., 7 9, 27, 92 Doherty, M. E., 14, 54
Bruning, J. L., 286n Cook, M., 228 Doll, L. S., 82
Brunswik.E., 11-13,44- Cooper, W. S., 207-208 Donoghue, J. M., 259
55, 57-58, 63, 133, Copernicus, N., 23 Doursat, R., 293
135, 137, 163, 170, Cosmides, L., 53, 59, Dreyfus, A., 72
182, 196-197, 205, 64, 122, 199, 213- DuBois-Reymond, E.,
260 218, 220, 228-229, 23
Buckhout, R., 285 234, 252-253, 264- DuCharme, W. M., 74
Buckley, T., 169 265, 292 Dugatkin, L. A., 174
Biihler, K., 44, 45, 50, Costall, A., 224 Dunker, K., 33-34,
55 Cowan, N., 20, 255 164
Bundeszentrale fur Cowgill, G. L., 278
gesundheitliche Cox, J. R., 223 Barman, J., 100, 122
Aufklarung (Federal Cronbach, L., 12, 48 Ebbinghaus, H., 169
Center for Health Cummins, D. D., 216 Eberle, J. R, 80-81
Education), 90 Curd, M., 3 Ebert, A., 77
Burke, D. S., 81 Eckert, J. P., 30
Burnett, S., 137 d'Alembert, J., 27 Eddy, D., 60-61, 65, 88,
Busch, M. P., 80 Daly, M., 71 98
Buss, D., 53 Dangleish, L., 278 Edwards, W., 14, 24, 92-
Butterworth, G., 224 Dannenberg, A. L., 78n 93, 130, 157, 159,
Byrne, R., 224, 226, Danziger, K., 5, 23, 25, 193, 242, 247
234 37, 48, 268, 294 Einhorn, H. J., 21, 164,
Darwin, C., 23, 53, 166, 181-182, 191, 193,
Cara, R, 191 213, 227, 230 195-196, 291
Caraco, T., 259 Dashiell, J. R, 12 Einstein, A., 24
Carey, S., 192 Daston, L., 9, 17, 27- Elstein, A. S., 88
Carnap, R., 24 28, 29n, 60, 62, 92, Elster, J., 191, 209
Carnot, L., 28 166, 201, 227, 235n, Emery, J. L., 193
Carpenter, R. G., 193 241, 246, 264-265 Epstein, J. S., 80n
Carroll, J. B., 227 Davies, N. B., 166 Ericsson, K. A., 110
Carver, R. P., 278, 284 Dawes, R., 161-162, Estes, W. K., 64, 205,
Cascells, W., 65, 98, 180-181, 184, 186, 268
122, 242, 251, 253 193 Evans, R. B., 48
Castellan, N. J., 39 Dawkins, R., 230 Evans, J. St. B. T., 217,
Catalan, J., 78n de Mere, A. G., 204- 252, 293
Chater, N., 295 205
Cheng, P. W., 216, 220- de Waal, R B. M., 213, Paraday, M., 4
221, 224 229 Pechner, G. T., 4, 7, 10,
Chiles, L., 77 DePinetti, B., 18, 246, 36, 40, 45, 278
Christensen, C., 150 283 Ferguson, L., 285
Christensen-Szalanski, Dehaene, S., 96 Fermat, P., 204
J. J. J., 52, 121 Dershowitz, A., 68-70 Feynman, R., 32n, 94,
Clinton, W. J., 78 Deutscher Bundestag, 121
Cohen, L. J., 19, 21, 81-82, 90 Fiedler, K., 19, 249,
61n, 155, 245n, 262, Devereux, G., 279 260, 289, 291
283 Devore, L, 213 Fillenbaum, S., 223
Colebrooke, H., 27 Diaconis, P., 287 Finlay, D., 162
Condillac, E. B., 27 Dietz, K., 79n Fischer, D., 82
 NAME INDEX 331

Fischhoff, B., 17, 51, 169, 174-175, 180, Hager, J. L., 212, 274n
93, 130, 135, 141, 192, 194, 196-197, Hamilton, W. D., 213
144, 145, 150-151, 217-222, 224-225, Hammerton, M., 98
161, 169, 174, 243, 228, 233-234, 237- Hammond, K. R., 12,
245, 279 238, 241, 245n, 247, 46-47, 50, 54, 93,
Fishbein, M., 9 249, 250, 256-257, 138, 164, 166, 169,
Fishburn, P. C., 191, 259, 260-263, 265- 182, 193, 196
194 268, 270, 274, 278, Hansen, R. D., 259
Fisher, I., 290 283, 286n, 290, 292- Hanson, N. R., 4
Fisher, R. A., 8, 10-11, 293, 295 Harcourt, A. H., 229
14-15, 36, 47, 49-50, Gilbert, W., 26n Harvey, J. H., 257-259
102, 115, 167, 250, Gillespie, J. H., 208n Hasher, L., 64, 137, 262
264, 266, 269, 270- Gilovich, T., 290 Hastie, R., 262
277, 280-285 Gilpin, M. E., 195 Hays, W. L., 281-282
Fodor, J. A., 41, 212, Ginossar, Z., 21, 257, Hedges, L., 169
228-229, 295 259 Heidelberger, M., 241
Fraisse, P., 162 Girotto, V., 191, 216 Heider, F, 12, 224, 232
Frenkel-Bmnswik, E., Gleik, J., 32n Heisenberg, W., 267
196 Gliick, D., 79n Hell, W., 20, 116, 208n,
Fretwell, S. D., 206 Godel, K., 44 256-257, 259
Freud, S., 237, 250, 293 Goldstein, D., 137 Helmholtz, H. von, 9
Frey, B. S., 242 Goldstein, D. G., 26, Helson, H., 291
Freyd, J. J., 265 126-127, 166, 174- Hempel, C. G., 24
Friedman, J. H., 191 175, 180, 192, 233 Henkel, R. E., 278
Good, I. J., 16, 68-70 Herbart, J. F, 7
Galanter, E., 34 Cosset, W. S., 10, 271 Hertwig, R., 51, 137,
Galison, P., 25 Gould, S. J., 61, 229, 169, 174, 197, 225,
Gallistel, C. R., 64, 192, 249 249-250, 266
205, 206, 234 Grayboys, T., 65, 242, Hilgard, E., 12, 47, 49
Gallon, F., 4, 49, 227- 251, 253 Hilton, D. J., 188, 225
228 Green, B., 37 Hinton, G. E., 293
Garcia, J., 211-212, 228 Gregory, R. L., 9, 10, Hintzman, D. L,, 64,
Garcia y Robertson, R., 258 137
212 Grether, D. M., 259 Hirschfeld, L. A., 212,
Gardner, H., 13 Grice, H. P., 224, 265 225-226
Gardner, A., 193 Griggs, R. A., 217, 223 Hoepmer, R., 227
Gauss, K. F., 27 Groner, M., 259 Hoffman-Valentin, F.,
Gavanski, I., 115, 121 Groner, R., 259 79n
Gavin, E. A., 8 Gruber, H. E., 14, 211, Hoffrage, U., 18, 51-52,
Gelman, R., 64, 192 268 65, 68, 77, 91-92,
Gelman, S. A.,212,225- Gmpsmith, E., 39 122, 126 ,129, 169,
226 Guilford, J. P., 7, 227, 174, 197, 247
Geman, S., 293 278-279, 282-283, Hoffstatter, P. R., 242
George, J. R., 80 285 Hogarth, R. M., 21, 164,
Gigerenzer, G., 4, 6, 7, Guttman, L., 284 181-182, 195-196,
10, 14, 18, 20, 24, 36, 291
45, 47-48, 51, 54, 58, Ha, Y., 186 Holland, J. H., 256-257
61n, 65, 68, 72, 75, Hacking, L, 17, 25, 62, Holton, G., 24
91, 116, 125, 127, 245n, 275 Holyoak,K.J.,216, 220-
136-137, 162, 164, Hackmann, W. D., 23 221, 224
332 NAME INDEX

Retelling, H., 278 Kaplan, R., 207-208 Lai, C., 150

Howell, W. C., 137 Keeney, R. L., 168, 182, Landers, A., 71-72
Huber, O., 181 191 Landman, J., 251-252
Hug, K., 217-222, 224, Keil, F. C., 192 Langley, P., 22, 35
265 Kekule, F. A., 4 Laplace, P., 9, 35n, 92,
Hui, C., 115, 121 Kelley, H. H., 6, 8-9, 166, 265, 290
Hull, C. L., 23, 47, 54, 12, 40, 50, 241 Larkin, J., 94
196 Kendall, M. G., 11, 267- Laskey, K. B., 234
Humphrey, G., 24 277 Leary, D. E., 6, 11, 25,
Humphrey, N. K., 234- Kennedy, M. B., 82 163
235 Kepler, J., 22 LeCam, L. M., 281
Hursch, C. J., 54, 169 Keren, G., 18, 143, 159, Legendre, A., 28
Hutchins, E., 42 162-163 Legendre, G., 191
Huttenlocher, J., 169 Kidd, R. E, 242, 246 Legrenzi, P., 215-
Kiefer, J. C., 281 216
Inhelder, B., 92, 103, Kintz, B. L., 286n Lehman, D. R., 117
201, 250 Kita, S., 41 Leibniz, G. W., 201-
Irmscher, M., 137 Klayman, J., 127, 186 202, 225
Kleinbolting, H., 18, Lenoir, T., 4, 23, 25
James, W., 34 126, 129, 247 Leslie, A., 226
Jastrow, J., 242 Kleiter, G. D., 63, 64n, Levelt, W. J. M., 41
Jeffrey, R., 283 96, 100 Levi, L, 16, 262
John Paul II, 53 Kobberling, J., 67-68 Levinson, S., 41
Johnson, E. J., 181 Koehler, J. J., 61, 72, Lewin, K., 47
Johnson, P., 278 93 Lewontin, R. C., 195
Johnson-Laird, P., 42, Koelling, R. A., 211 Li, F., 117
132, 170, 212, 215, Kohler, W., 33-34, 40, Lichtenstein, S., 17-18,
256n 267, 269 51, 93, 135, 141, 145,
Johnstone, D. J., 274 Kolmogoroff, A. N., 150, 151, 174, 243,
Jones, E. E., 9 167, 266 245
Jones, C. M., 64 Koriat, A., 135, 153, Liebert, R. M., 246
Jonides, J., 64 157, 159, 160-161, Light, P., 216
Joyce, C. R. B., 138, 174, 245 Lindman, H. 14
164 Kosslyn, S. M., 94 Lindquist, E. F., 285
Judd, C. M., 242 Krantz, D. H., 288 Little, I. M. D., 203
Jung, C. G., 250 Krauss, S., 122 Locke, J., 55, 201
Jungermann, H., 61, Krebs, J. R., 102, 122, Loftus, G. R., 238
260 166, 233 Lopes, L. L., 14, 19,
Juslin, P., 51, 127, 169, Krech, D., 47 61n, 93, 168, 191,
174 Kroznick, J. A., 117 193, 208n
Kriiger, L., 241 Lovie, A. D., 6, 180,
Kadane, J. B., 18 Kruglanski, A. W., 157 191, 193, 268
Kahneman, D., 14-16, Kiihberger, A., 126 Lovie, P., 180, 193
19-20, 51, 61, 93, Kuhn, T. S., 25 Lubek, L, 212
102, 110, 151, 158, Kulkarni, D., 22 Luce, R. D., 9, 13, 40,
167, 237, 242-244, Kiilpe, O. 37 191, 269-270, 288
248-249, 251, 254- Kummer, H., 224, 226, Luchins, A. S., 188
259, 262, 284, 290- 231 Luchins, E. H., 188
292 Kurz, E. M., 54, 196 Luttrell, L. M., 213
Kant, L, 55 Kyburg, H. E., 155 Lyon, D., 98
 NAME INDEX 333

MacGregor, D., 138, Mineka, S., 228 Packer, C., 213

161 Mises, R. von, 18, 141, Papert, S., 30
Manis, M., 251-252 244, 246 Parducci, A., 152, 291
Manktelow, K. I., 216- Mitchell, S. D., 230 Park, B., 242
217, 224 Moliere, J. B. P., 291 Pascal, B., 204
Mansson, S. A., 78 Montgomery, H. 132, Paulos, J. A., 74
Marmorstein, H., 181 157 Pavlov, I. P., 40, 267,
Marr, D., 94 Morgan, M. S., 241 268, 269
Martignon, L., 52, 122, Morgenbesser, S., 230 Payne, J. W., 181, 182
197, 234 Morrison, D. E., 278 Pearson, E. S., 11, 15,
Martindale, S., 259 Mueser, P. R., 20, 255 17, 115, 264, 270,
Marzuk, P. M., 78n Mueser, K. T., 20, 255 272-273, 275-285,
Massaro, D. W., 9, 14, Murdock, B. B., Jr., 8 287
122, 293 Murino, M., 215 Pearson, K., 49-50
Mauchly, J. 30 Murray, D. J., 2, 4, 7, Peichl-Hoffman, G., 81
Maurer, C., 79n 16, 45, 61n, 164, 169, Peirce, C. S., 242
Maxwell, J. C., 23, 267, 196, 260-261, 263, Perry, S. W., 78n
295 268, 292 Persson, T., 174
May, R. S., 143, 151- Mynatt, C. R., 14 Peterson, C. R., 74
152, 157, 248 Phillips, L. D., 17, 51,
Mayseless, O., 157 Neale, M.A., 243 92-93, 145, 245
Me Arthur, L. A., 8 Nesse, R. M., 59 Piaget, J., 8, 40, 92,
McClelland, A. G. R., Neumann, J. von, 21, 103, 166, 201-202,
174 29-31 211-212, 214-215,
McClennen, E. E, 173 Neurath, O., 44 225, 232, 250, 267-
McCloskey, D. N., 166 Newcombe, R. D., 241 269, 276
McCorduck, P., 30-31, Newell, A., 31-39, 41,- Piatelli-Palmarini, M.,
33, 35n 42, 163, 292 61, 65, 237, 243
McCulloch, W. S., 29- Newton, L, 4 Pinker, S., 226
30 Neyman, J., 11, 14-15, Pitts, W., 29
McGillis, D., 9 17, 102, 115, 167, Platt, R., 217
McKenzie, C. R., 115 244, 250, 264, 266, Poincare, H., 4
McWeeny, P. M., 193 270, 272-273, 275- Poisson, S. D., 92
Medin, D. L., 115 285 Politzer, G., 221
Meehl, P. E., 270, 274n, Nguyen-Xuan, A., 221 Pollard, P., 217, 252,
278 Nickles, T., 3, 4 184, 285
Melton, A. W., 11, 269, Nisbett, R. E., 110, 242, Pomerantz, J. R., 94
280-281, 285 259, 263 Popper, K., 3-4, 24, 28,
Mendel, G., 267 Norman, D. A., 8 35, 44
Merrill, M. A., 235n Nozawa, G., 137 Porter, T., 23, 246
Metropoli, N., 32n Nozick, R., 115 Premack, D., 224, 232
Michaela, I. L., 8, 241 Nunnally, J. C., 280- Premack, A. J., 232
Michalski, R. S., 115 283, 285 Pribram, K. H., 34
Michelson, A. A., 24 Prince, A., 191
Michotte, A., 8, 40, 224 Oakes, M., 285 Prohaska, V., 169
Miller, G. A., 34, 37- Oaksford, M., 295 Prony, G. R. de, 28, 32
38, 285 Olshen, R. A., 191 Pugh, K., 78n
Miller, G. R, 229 Olsson, H., 51, 127 Pylyshyn, Z. W., 94
Millikan, R. A., 24, 226 Osherson, D. N., 291
Milson, R., 10, 14 Over, D. E., 216, 224 Quetelet, A., 23, 295
334 NAME INDEX

Raiffa, H., 168, 182, Searle, J., 21 Stevens, S. S., 45

191 Sedlmeier, P., 51, 75, Stewart, T. R., 12, 164
Ramsey, F. P., 194 137, 283, 286n Stich, S., 237, 242
Raymond, W., 191 Seligman, M. E. P., 212 Stigler, S. M., 60n, 94
Regier, T., 293 Selten, R., 125 Stine, G. J., 77, 80n, 82,
Reichenbach, H., 3, 4, Sen, A., 202-203 88
11, 22, 24, 44, 52, Shafer, G., 155, 245n Stone, G. O., 182
170 Shannon, C., 35n Stone, C. J., 191
Richardson, J. T. E., Shanteau, J., 242, 291 Strack, R, 242
284, 285 Shaughnessy, J. M., 74 "Student," 10, 271
Richter, H. R., 162, 283 Shaw, J. C., 33 Summers, D. A., 193
Rieskamp, J., 51, 174 Shepard, R. N., 37, 57, Suppes, P., 288
Rittner, C., 73 59, 19, 196, 291, 293 Swets, J. A., 6-8, 24,
Rohm, E., 73 Sherman, S. J., 242 34, 36, 162
Ronis, D. L., 152, 154- Siegler, R. S., 169 Swijtink, Z., 9, 278,
155, 158-159 Simmel, M., 232 286
Rosch, E., 133, 224 Simon, H. A., 5, 22, 31-
Rosenthal, R., 287 39, 41-42, 52-54, 94, Taine, H., 227
Rosnow, R. L., 287 102, 110, 125, 163, Tanner, W. P., Jr., 6-8,
Ross, L., 64, 242, 263 167-168, 170, 177, 24, 34, 36, 162
Rouanet, H., 92 193, 209, 259 Teigen, K. H., 19, 158
Rozeboom, W. W., 284 Simon, T., 235n Terman, L., 235n
Rucci, A. J., 11, 268 Simpson, O. J., 68, 70 Tetlock, P. E., 197, 225
Russell, B., 31, 33, 35n Sivyer, M., 162 Thaler, R. H., 243
Rutter, D. R., 241 Skinner, B. R, 40, 49, Thines, G., 224
52, 211, 228, 230, Thomas, D. H., 278
Sahlin, N., 100, 292 267-270 Thomson, G. H., 7
Saks, M., 242, 246 Sloman, S. A., 292 Thomson, W., 23
Samuels, R., 237 Slovic, P., 98, 138, 150, Thorndike, R. L., 12, 48
Samuelson, P. A., 203 243 Thurstone, L. L., 13,
Savage, L. J., 14, 122, Smith, L. D., 23 36, 227, 228
194, 281 Smith, A., 28 Todd, R J., 54, 169
Schaefer, R. E., 74 Smith, E. E., 291 Todd, P. M., 58, 125,
Schaffer, S., 26n, 28 Smolensky, P., 191, 293 197, 228-229, 233
Schering AG, 79n Snedecor, G. W., 278 Tolman, E. C., 23, 34,
Schiffrin, R. M., 293 Snell, J. J. S., 80 163
Schlick, M., 44-45, 53 Sniezak, J. A., 169 Tooby, J., 53, 59, 64,
Schlosberg, H., 48 Spearman, C., 227, 293 122, 213-214, 216,
Schneider, W. 293 Spencer, H., 227 228-229, 234, 252-
Schochetman, G., 80 Sperber, D., 191, 224, 253, 292
Schoenberger, A., 65, 229 Toppino, T., 137
242, 251, 253 Spettel, C. M., 246 Toulmin, S., 6, 25
Scholz, R. W., 93, 265 Spielberg, R, 80 Trope, Y., 21, 257, 259
Schonemann, P. H., 152 Statistisches Tu, X. T., 80-81
Schrage, G., 73-74 Bundesamt, 79, 143 Tukey, J. W., 22, 287
Schrodinger, E., 267 Steinmann, D. O. 164 Turing, A., 29-31
Schum, D. A., 115 Stephens, D. W., 102, Tversky, A., 14-16, 19-
Schwartz, S., 80, 278 122, 166, 233 20, 51, 61, 93, 102,
Schwarz, N., 225, 291 Sterling, T. D., 11, 268 110, 151, 158, 167,
 NAME INDEX 335

237, 242-244, 248- Wason, P., 214-216 Wimsatt, W., 167, 231
249, 251, 254-260, Wattenmaker, W. D., Windeler, J., 67
262, 284, 288, 290- 115 Winer, B. J., 281-282
292 Weber, E. H., 7 Winman, A., 51, 127,
Tweney, R. D., 11, 14, Weber, U., 188 174
186, 268 Wells, G. L., 257-259 Winterfeldt, D. von, 93,
Wendt, D., 242 130, 157, 159, 193,
Vallone, R., 290 Wertheimer, M., 34 247
Voneche, J. J., 211, 268 Westermann, R., 274n Wise, M. N., 23
Vrba, E. S., 230 White, S., 199 Wittgenstein, L., 44
Whitehead, A. N., 31, Wittkowski, K., 80-81
Walker, B. J., 186 33, 35n Woodworth, R. S., 48-
Wallace, B., 188 Whiten, A., 224, 226, 49
Wallach, L., 289, 290 234 World Health
Wallach, M. A., 289, Whittam, T. S., 259 Organization, 79-80
290 Wickelgren, W. A., 8 Wundt, W., 37, 40, 268
Wallendael, L. R., van Wiegand, W., 73
262 Wilkinson, G. S., 213 Yates, J. R, 152, 154-
Wallsten, T. S., 260, Williams, G.C., 59, 212- 155, 158-159
291 213
Wanke, M., 291 Wilson, M., 71 Zacks, R. T., 64, 262
Ward, J. W., 82 Wilson, T. D., 110 Zimmer, A. C., 155
Wascoe, N. E., 164, 193 Wilson, D., 224 Zytkow, J. M., 35
This page intentionally left blank
 SUBJECT INDEX

acceptance of new theories, 5, 10, artificial intelligence, 31, 227, 244,

26, 37-39, 42 260
accountability, 225 aspiration level, 167, 193, 259
accuracy, 53, 177, 183, 188, 233 astronomers, 23, 27, 230, 278,
achievement, 46, 51, 53, 163 286
adaptation, 10, 11, 53, 59, 61, 138, attribution theory, 8, 9, 40
141, 164, 168, 212, 230, 259, availability, 51, 216, 217, 220, 242,
260 259-261, 290, 291
adaptive
coin-flipping, 207, 208 BACON, 5, 22
problem, 200, 211, 214, 229 base rate, 16, 20, 67, 72, 93, 99, 100,
toolbox, 228 106, 151, 255-260, 263
unpredictability, 233 neglect (fallacy), 16, 20, 51, 52, 57,
advertisement, 192 61, 65, 93, 115, 212, 242, 243,
AIDS counseling, 58, 71, 72, 77-91 251-254, 266, 290
altruism, 213, 216, 217. See also only, 114, 116, 118, 120
cooperation battering, 58, 68-71
ambiguity of cues, 44, 46 Bayes's rule, 10, 16, 17, 51, 60-62,
American Psychological Association, 66-69, 75, 81, 92, 93, 99-110,
12, 212, 284 117, 121, 122, 166, 201, 232,
Publication Manual, 269, 294 233, 255, 260, 262, 264, 271,
analysis of variance (ANOVA), 6, 8, 287
12, 40, 49, 50, 55, 193, 233, 241, physical analog of, 103
269 Bayesian
Analytical Engine, 28, 30 reasoning, 52, 59-61, 63, 64, 70,
anchoring and adjustment, 259, 261, 74, 75, 92-123, 167
290 statistics, 9, 10, 14, 201, 241, 259,
anthropocentrism, 230 277
anthropomorphism, 230 Bechterev's disease, 67
anxiety, 280, 281 behaviorism, 6, 31, 37, 212, 228,
Archimedian axiom, 174 230, 269, 273, 277

337
338 SUBJECT INDEX

blood donors, 77-79 conjunction fallacy, 19, 212, 242,

bounded rationality, viii, 8, 53, 54, 248-250, 252, 266. See also
58, 123, 125, 167, 168, 170, 190, Linda problem
192, 197 conservatism, 92, 290
breast cancer, 58, 60, 61, 65, 66, 95- consistency, 18, 151, 197, 201-204,
98, 101, 102, 108, 110, 245 209, 232, 233, 246, 250
brochures, 89, 90 content effect, 217
Brunswik ratio, 11 context, 44-46
bureaux de calculs, 27, 32, 42 of discovery, 3, 4, 13, 24, 25
of justification, 3, 13, 20, 21, 23,
cab problem, 16, 98, 102, 111, 169, 42
244, 262 cooperation, 213, 214, 216, 218, 226.
calibration, 138, 145, 146, 148, 149, See also altruism
151, 153-155, 161, 248 copyediting, 1
causal counseling, 77-91
attribution theory, 8, 9, 40 criminal law, 58, 68-74
base rates, 261 cue, 170-176
reasoning, 6-9, 12, 40, 50, 241 ambiguity, 44, 46
cheating detection, 199, 211, 213, integration, 164
216, 217, 220-222, 224, 264 redundancy, 196
classification and regression trees, substitution, 131-135, 164, 172,
197 173
cognitive utilization, 46, 133, 137
illusion, 51, 57, 157, 237, 242- validity, 115, 133-135, 137, 164,
244, 246-248, 252, 261, 262, 188, 190
266, 290
maps, 23
revolution, 1, 4, 6, 29-31, 53, 228, Darwinian psychology, 53
241, 259 data fitting, 290, 293, 294
coherence, 250. See also consistency decision
color constancy, 59, 61, 63 criterion, 7, 13
colorectal cancer, 67, 68 rule, 173
Columbia Bible, 47, 48, 50 decomposition, 35
commensurability, 191, 195 degree of belief, 18, 246, 273
compulsive hand washing, 55 determinism, 47, 91, 268
computational dichotomies, 290-293
complexity, 99, 121, 227 disciplinary isolation, 295
simplicity, 63, 64, 99, 101-103, disconfirmation strategies,
107, 108, 116 153
computer, 1, 21, 26, 27, 29-33, 35, discovery, 2, 4, 5, 8, 10, 13, 23-26,
38, 42, 43, 293 32, 34, 35, 40, 42, 60, 126, 127,
confidence, 18, 19, 129, 135, 136, 275
149, 150, 153-155, 169, 246, discrimination, 7, 8, 34, 36
250, 274, 282 rate, 176
confidence-frequency effect, 140, division of labor, 28, 32, 43, 200
147, 148, 151, 152, 155-158, domain
165, 174 actual, 229, 230
confirmation bias, 159-161, 174, generality, 228
186, 216, 245, 246, 248 proper, 229, 230
 SUBJECT INDEX 339

domain specificity, 199, 200, 211- false

213, 216, 222, 224, 225, 228, negative, 80, 91. See also miss
229, 265, 293 positive, 71, 80, 82, 85-87, 89, 90,
double standards, 17, 51, 287 91
positive rate, 60, 61, 67, 79, 82, 96
false alarm, 7, 107, 273. See also
ecological
false positive
rationality, viii, 8, 57, 58, 105,
frequencies, 100
125, 199
rate, 96, 100, 106, 107
validity, 11, 45, 46, 52, 53, 133,
fast and frugal
137, 141, 158, 175, 176, 181,
heuristic, 123, 126, 168, 171, 175,
182
177, 184, 186, 198, 232, 233,
editors, 11, 49, 212, 238, 269, 280,
234
285, 286
lens model, 54, 196
Ego, 280, 281
Fisherian statistics, 9, 17, 39, 114,
Einstellung effect, 187
115, 118-120
electromagnetism, 26
flat maximum effect, 51, 193
ELISA (enzyme-linked
food, 192, 213, 214, 234
immunoabsorbent assay) test,
foraging, 102, 206
71, 72, 77, 80, 81, 87, 89, 107
frame problem, 227
emotion, 53, 199, 232, 234
free will, 30
engineer-lawyer problem, 20, 51,
French Revolution, 29, 62, 92
244, 254-258, 260
frequency, 57, 91, 253
enlightenment, 27, 28, 34, 35, 55, 92,
format, 66, 99, 100, 101, 107, 108,
93, 122, 166, 167, 198, 201, 232,
116, 121, 122, 253
265
judgments, 18, 19, 247-251
view of intelligence, 31, 42
learning, 64, 262
environment, 58, 61, 62, 64, 131,
long run, 18, 248
158, 168, 175, 260
frugality, 177, 183, 188
of evolutionary adaptiveness, 59
functional validity, 11, 46
laboratory, 41, 57
structure of, 52-54, 58, 105, 125,
Galton-Pearson program, 12
168, 182, 190, 262-264
Gambler's fallacy, 290, 291
surplus structure, 264
game theory, 213
evolution, 53, 58, 62, 64, 76, 212,
general-purpose
227, 268
intelligence, 227, 228
of cooperation, 217, 223
mechanism, 211, 212
evolutionarily stable strategy, 206,
generative entrenchment, 231
213, 214
genius, 27, 28
evolutionary theory, 199, 200, 210,
German measles, 103-110, 117, 120
213, 216, 292
God, 10
exaptation, 230
Gricean conversational principle,
expert witness, 72-74
225
guilt, 280, 281
factor analysis, 12, 55, 227, 293 Guthrie test, 67
faculty, 200, 227, 228
horizontal, 228 hard-easy effect, 126, 127, 129, 141-
psychology, 227 143, 157, 159, 162, 164, 174
vertical, 228 reversal, 142, 159
340 SUBJECT INDEX

hemoccult test, 67 menu, 94-122

heuristics, 3, 4, 52, 125, 167, 173, processing, 31-39
185, 187, 194, 259-261 representation, 52, 58, 62, 63, 76,
and biases program, 15-21, 92, 88, 94, 100
166, 167, 237, 242-266 innumeracy, 58, 74, 95
of discovery, 22 insight, 1, 4, 6, 33, 58, 67, 95, 126
fast and frugal, 123, 126, 168, 171, institutionalization of inferential
175, 177, 184, 186, 198, 232, statistics, 6, 8, 11-13, 267-288,
233, 234 294
ignorance-based decision making, intellectual integrity, 46, 49, 50, 55,
192 238
one-reason decision making, 173, intelligence, 26, 27, 29, 32, 33, 41,
190, 193, 197 42, 48, 227, 228, 292
satisficing, 167, 170, 192, 193, 259 artificial, 31, 227, 244, 260
See also anchoring and as a combinatorial calculus, 27,
adjustment, availability, 34, 35
Minimalist heuristic, recognition distributed, 43
heuristic, representativeness, Machiavellian, 222, 224, 226, 235
Take The Best heuristic, Take machine, 28
The Last heuristic, tallying, tools- primate, 226
to-theories heuristic test, 235
hierarchical organization, 231, 232 intransitivity, 173, 194, 195, 197
hit rate, 60, 61, 67, 72, 96, 100, 106. intuitive statistics, 6, 11, 12, 14, 17,
See also sensitivity, power 20, 21, 31, 54, 241, 242, 263,
HIV (human immunodeficiency 264, 266
virus), 71, 72, 77-91 irrationality, 103, 216, 237, 249, 265
counseling, 71, 72, 77-91
test, 58, 71, 77, 78, 107, 122 joint occurrence, 114, 115, 118-120
HL-antigen-B27 test, 67 Journal of the Experimental Analysis
hologram, 21 of Behavior, 49, 269
homme moyen, 23 Journal of Mathematical Psychology,
hommes eclaires, 27, 92, 122 269
homunculus, 7, 9, 13, 50, 233, 249
hot hand fallacy, 290, 291 knowledge, limited, 170, 171, 177,
179, 183, 184, 186, 187, 195
Id, 280, 284-286
idiot savant, 27 Laplacean demon (superintelligence),
ignorance-based decision making, 12, 167, 168, 170, 183, 194, 233,
192 234
illusion of validity, 17 law of comparative judgement, 13
imitation, learning through, 228 Leibniz's dream, 201, 202, 211
inclusive fitness, 213 lens model, 54, 196
incommensurability, 191, 195 less-is-more effect, 126, 127, 174,
inference 179, 184
from givens, 169 level of significance, 15, 267, 269,
from memory, 169, 194 271, 272, 277, 278, 280-284,
revolution, 267, 268 287
information exact, 115, 272, 285, 286
format, 93-122 lexicographic strategy, 168, 191
 SUBJECT INDEX 341

life insurance, 78 motivation, 202, 203, 209, 210

likelihood, 260, 264 multidimensional
ratio, 7, 13, 69, 70 psychophysics, 45, 50
subtraction, 114, 115, 118, 119 scaling, 55, 293
Linda problem, 19, 169, 197, 248- multiple regression, 4, 54, 182, 183,
250. See also conjunction 185, 186, 189, 190, 193, 194,
fallacy 196
local mental model (LMM), 130, 132, murder, 68-70, 73, 74
133, 136
logic, 132, 166, 197, 199, 201, 202, name recognition, 192, 230
211, 215, 216, 221, 222, 225, natural
265 environment, 57, 132, 133, 139,
of discovery, 3, 4, 22, 35 140, 143, 163, 233
Logic Theorist (LT), 32, 33, 36, 37 frequencies, 58, 62-68, 70, 72, 75,
logical 76, 81-83, 88-91, 95, 96, 98,
empiricists, 25 122
positivism, 44 sampling, 52, 57, 58, 62, 63, 95,
96, 99, 100, 122
mammography, 58, 60, 62, 65-68, selection, 208, 212, 227
75, 97, 98, 101-112, 122 neural networks, 182, 190, 193, 293
Manhattan Project, 32 Neyman-Pearson theory, 7, 8, 16, 34,
manufacturing, 1, 28, 32, 42 36, 115, 233, 262, 264, 271-273,
maximizing, 204-209 275, 277, 279, 280, 284, 285,
expected utility, 205 287
expected value, 206 normal distribution, 13, 23
medical diagnosis, 58, 60, 65-68 normative, 167, 202, 242-244, 246,
mental 251, 255, 261, 262, 266
arithmetic, 27 null hypothesis, 6, 267-288
logic, 223 testing, 115, 238, 239, 267-288,
model, 131, 212, 256, 258, 259 294. See also significance testing
metaphor, 1, 5, 13, 14, 21, 28, 33, null science, 295
38, 39, 41, 42, 54, 58, 125, 247,
262, 265, 280, 295 odds, 69, 70, 89
methodological dogma, 51 one-reason decision making, 173,
mind 190, 193, 197, 196
as a computer, 33-41 one-word explanations, 290, 291
as an intuitive statistician, 6, 7, optimality theory, 191, 192
10-14, 31, 50, 54, 247, 264, overconfidence bias, 17, 18, 51, 129-
266 165, 174, 242-248, 252, 266
Minimalist heuristic, 184, 185, 188, Oxford English Dictionary, 250, 266
189, 194, 195, 197
miss, 7, 36, 273. See also false p value, 49, 238, 239, 267, 281-287
negative perceptual
modularity, 226-235 achievement, 44
strong, 229 compromises, 45
weak, 229 constancies, 11, 44
module, 227, 229-232, 234, 235 illusion, 161, 162, 237
moral, 29, 230, 235 perspective change, 220-222
sentiment, 27 phenylketonuria, 67
342 SUBJECT INDEX

physicians, 64, 66-68, 81, 83, 85, 88, rape, 73

98, 188 rationality, 21, 92, 167, 193, 200,
politeness, 203, 204, 233 201, 206, 209, 210, 243, 246
positive predictive value (PPV), 79, classical, 167, 168, 173, 175, 186,
81-83, 85-91 191, 194
posterior war, 237
distribution, 100 recognition, 8, 9, 177, 184
odds, 69 heuristic, 126, 127, 172, 178, 186,
probability, 20, 60, 62, 98, 100, 187, 191, 192, 234
101, 106, 107, 279, 285 validity, 178, 180
power, 15, 275, 279, 281, 283. See redescription, 260, 261, 290-292
also sensitivity reference class, 133, 136, 142, 143,
prevalence, 79, 81, 83-88, 90, 91. 151, 163, 164, 170, 171, 176,
See also base rate 256, 263, 264
principle of isolation, 41 relative frequency, 17, 19, 102, 116,
prior 117, 246
odds, 69 format, 119, 120
probability, 16, 60, 255, 260, 264, relevance theory, 191
271 replication fallacy, 285
probabilistic representation, 52, 58, 62, 63, 76, 88,
functionalism, 11, 45, 46, 48 94, 100. See also information
mental model (PMM), 19, 54, format, information menu
126, 129-165, 170, 172, 174, training, 75
190, 191, 194, 197, 248, 256, representative
262 design, 45, 46, 48, 49, 51, 54
revolution, 238, 241, 267, 268 sampling, 51, 63, 137, 138, 141,
probability, 57, 58, 62, 63, 70, 81, 144, 150, 151, 158, 162
83, 91, 120, 166, 212 representativeness, 19, 20, 45, 93,
cues, 133, 136 242, 248, 250, 255, 256, 258-
epistemic interpretation, 15, 18, 261, 290, 291
246, 273 (see also degrees of ritual, 39, 49, 238, 239, 280, 294
belief) rule training, 75
format, 66-68, 100, 105, 107, 108,
112, 119, 121
frequentist interpretation, 15, 17- sampling
19, 246, 247, 280 random, 20, 254-257, 259, 263,
matching, 205 264, 275, 276, 280
nonadditive, 245, 266 systematic, 95, 96
second-order, 64, 100 representative, 51, 63, 137, 138,
of single events, 17-19, 65, 69, 141, 144, 150, 151, 158, 162
70, 96, 99, 102, 103, 117, 121, satisficing, 167, 170, 192, 193, 259
136, 141, 158, 245-248, 252, scientific community, 5, 12, 13, 26
253 screening, 58, 66, 67, 96, 102
theory, 72, 92, 141, 199, 201, search, 171, 173, 180, 230
202, 204, 232, 246, 249, 250, exhaustive, 193
261 limited, 193
process analysis, 108-111 rule, 172, 187, 188
Property Alpha, 202-204, 209 selection task, 197, 199, 214, 215,
psychoanalysis, 19, 54, 196, 237 217, 218, 295
psychophysics, 7, 8, 45, 50 self-identity, 229
 SUBJECT INDEX 343

sensitivity, 60, 61, 67, 72, 79-81, Neyman-Pearson theory, 7, 8, 16,

83-91, 275. See also hit rate, 34, 36, 115, 233, 262, 264, 271-
power 273, 275, 277, 279, 280, 284,
shortcut, 58, 103, 105-107, 109 285, 287
signal-detection theory. See theory of stimulus
signal detectability equipotentiality, 211, 212
significance testing, 9-11, 14, 40, detection, 6, 7, 13, 34, 36
115, 270, 271, 274, 275, 278, discrimination, 7
280, 286. See also null- stopping rule, 172, 174, 175, 191,
hypothesis testing 193-195
similarity, 19, 260, 261, 291 strategies
simulation 36, 38, 39, 42, 183, 187, non-Bayesian, 65, 114-116, 118-
192 120
social noncompensatory, 173, 174, 190,
computer, 42, 43 191, 196
context, 265 nonlinear, 173, 174, 193, 194
contract, 211, 213, 216-224, 229- See also Bayes's rule, heuristics,
231, 234, 264 unit-weight linear model
environment, 1, 199 structural isomorphs, 264
exchange, 199, 211 structure of environments, 52, 53,
intelligence, 226, 227, 229-231, 58, 106, 125, 168, 182, 190, 262-
233, 235 264
interaction, 206, 207 subroutine, 32, 35, 41
objective, 202-204, 209, 210, 213 suicide, 77, 78
organization of work, 1, 26-28 Superego, 280, 281, 287
origin of computer metaphor, 1, 26 surrogates for theory, 289, 290, 292-
rationality, viii, 8, 199, 224 295
tool, 226
worker, 83, 85, 86 Take The Best, 125, 168, 171-197,
special theory of relativity, 24 233, 234
specificity, 79, 80, 81, 83, 88, 89. See Take The Last, 184, 185, 187-190,
also power, false positive rate 195
St. Petersburg paradox, 208, 265 tallying, 180, 181, 183, 185-187
statistical teaching, 30, 91
inference, 14, 15, 20, 276, 278, statistical reasoning, 74, 75
282, 287 (see also Bayes's rule, testimony
Bayesian statistics, statistics, eye-witness, 8, 16, 262
null-hypothesis testing) expert witness, 72—74
mechanics, 23 textbooks, 15, 67, 75, 99, 229, 238,
thinking, 49, 58, 75, 94, 239 278, 279, 282-286
statistics theory of signal detectability (TSD),
Fisherian, 9, 17, 39, 114, 115, 118- 6, 7, 13, 34, 36
120 think-aloud protocol, 36, 37
hybrid logic of, 277, 279-281, 284, thinking as calculation, 27-29, 32,
287, 288 41, 42
inferential, 6, 11 threat, 231
institutionalization of inferential, threshold, 7, 36
6, 8, 11-13, 267-288, 294 Tit For Tat, 214
intuitive, 6, 11, 12, 14, 17, 20, 21, tools, 1, 5
31, 54, 241, 242, 263, 264, 266 of justification, 5
344 SUBJECT INDEX

tools-to-theories heuristic, 2-26, 31, utility, 6, 167, 203, 209, 276

36, 38, 40, 42, 57, 193 maximization, 201
transitivity. See intransitivity
transparency, 193 vicarious functioning, 46, 48, 53, 54,
triggering algorithm, 231, 232, 235 134, 164, 196
true positives, 82. See also Vienna Circle, 4, 5, 11, 24, 44, 53,
sensitivity, power 55
two-process theories, 293 visual illusion, 242
Type I error, 7, 36, 246, 272, 273,
Wason selection task. See selection
275, 276, 281-284, 286
task
Type II error, 7, 15, 36, 246, 273,
weighted
275, 276, 281-283
linear model, 182, 183, 185-187
tallying, 181, 183, 185, 193
unconscious inferences, 9 Western blot test, 71, 72, 77, 80, 81,
unit-weight linear model, 181, 183, 87, 89
185-187, 193 window period, 84, 86, 87, 88, 91
Universal Characteristic, 201, 211, write-aloud protocols, 109, 111,
225 115