L A N D

M A N A G E M E N T

H A N D B O O K

42

Statistical Methods for

Adaptive Management Studies

1998

Ministry of Forests
Research Program

Statistical Methods for

Adaptive Management Studies
Vera Sit and Brenda Taylor (Editors)

Ministry of Forests
Research Program

Citation
Sit, V. and B. Taylor (editors). 1998. Statistical Methods for Adaptive Management Studies. Res. Br.,
B.C. Min. For., Res. Br., Victoria, BC, Land Manage. Handb. No. 42.

Canadian Cataloguing in Publication Data

Main entry under title:
Statistical methods for adaptive management studies
(Land management handbook ; 42)
Includes bibliographical references: p.
ISBN 0772635129
1. Forest management Statistical methods.
2. Forest management British Columbia. I. Sit, Vera.
II. Taylor, Brenda, 1962 . III. British Columbia. Ministry
of Forests. Research Branch. IV. Series.
SD387.S73S72 1998

634.92072

C989600815

Prepared for:
B.C. Ministry of Forests
Research Branch
PO Box 9519, Stn Prov Govt
Victoria, BC V8W 9C2
Published by
B.C. Ministry of Forests
Forestry Division Services Branch
Production Resources
595 Pandora Avenue
Victoria, BC V8W 3E7
1998 Province of British Columbia
Copies of this and other Ministry of Forests
titles are available from:
Crown Publications Inc.
521 Fort Street
Victoria, BC V8W 1E7
Send comments to: Vera Sit, B.C. Ministry of Forests, Research Branch,
PO Box 9519, Stn Prov Govt, Victoria, BC V8W 9C2
Ministry of Forests Publications Internet Catalogue: www.for.gov.bc.ca/hfd

ACKNOWLEDGEMENTS

This report is the combined effort of a number of

people. We would like to thank the chapter authors
not only for their written contributions, but also for
discussions that helped to identify and refine the
contents of the report. Many thanks to our reviewers: Balvinder Biring, Ian Cameron, Phil Comeau,
Nola Daintith, Darrell Errico, Roger Green, George
Harper, Russ Horton, Val LeMay, Gordon Nigh,
Peter Ott, Rick Page, Martin Raphael, Michael Pitt,
Pasi Puttonen, Doug Steventon, Michael Stoehr, Jeff
Stone, Keith Thomas, Chris Thompson, Ian Thompson, and Rita Winkler. Comments and suggestions
from the reviewers strengthened each chapter and the

report as a whole. We would also like to extend our

appreciation to Brian Nyberg for his full support and
continued encouragement throughout the project,
and to Paul Nystedt for assistance with the reports
publication. Funding for this report was provided by
the B.C. Ministry of Forests, Forest Practices Branch.
Vera Sit
Brenda Taylor
Victoria, B.C.
March 1998

iii

LIST OF CONTRIBUTORS

Judith L. Anderson

Simon Fraser University, School of Resource and Environmental Management,

Burnaby, BC V5A 1S6

Wendy A. Bergerud

B.C. Ministry of Forests, Research Branch,

P.O. Box 9519, Station Provincial Government, Victoria, BC V8W 9C2

Bruce G. Marcot

U.S. Forest Service, Department of Agriculture, Pacific Northwest Research Station,

1221 SW Yamhill Street, Suite 200, Portland, OR 97208-3890

Amanda F. Linnell Nemec International Statistics and Research Corporation,

P.O. Box 496, Brentwood Bay, BC V8M 1R3
J. Brian Nyberg

B.C. Ministry of Forests, Forest Practices Branch,

P.O. Box 9513, Station Provincial Government, Victoria, BC V8W 9C2

Randall M. Peterman

Simon Fraser University, School of Resource and Environmental Management,

Burnaby, BC V5A 1S6

Calvin N. Peters

ESSA Technologies Ltd.,

Suite 300, 1765 West 8th Avenue, Vancouver, BC V6J 5C6

William J. Reed

University of Victoria, Department of Mathematics and Statistics,

P.O. Box 3045, Victoria, BC V8W 3P4

Richard D. Routledge

Simon Fraser University, Department of Mathematics and Statistics,

Burnaby, BC V5A 1S6

Carl J. Schwarz

Simon Fraser University, Department of Mathematics and Statistics,

Burnaby, BC V5A 1S6

Vera Sit

B.C. Ministry of Forests, Research Branch,

P.O. Box 9519, Station Provincial Government, Victoria, BC V8W 9C2

G. John Smith

Geo. John Smith Statistical Consulting Services,

2781 Point Grey Road, Vancouver, BC V6K 1A4

Brenda Taylor

B.C. Ministry of Forests, Forest Practices Branch,

P.O. Box 9513, Station Provincial Government, Victoria, BC V8W 9C2

iv

CONTENTS

Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
List of Contributors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
1 Statistics and the Practice of Adaptive Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
J. Brian Nyberg
2 Design of Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Amanda F. Linnell Nemec
3 Studies of Uncontrolled Events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Carl J. Schwarz
4 Retrospective Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
G. John Smith
5 Measurements and Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Richard D. Routledge
6 Errors of Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Judith L. Anderson
7 Bayesian Statistical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Wendy A. Bergerud and William J. Reed
8 Decision Analysis: Taking Uncertainties into Account
in Forest Resource Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Randall M. Peterman and Calvin N. Peters
9 Selecting Appropriate Statistical Procedures and Asking
the Right Questions: A Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Bruce G. Marcot
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145


3.1 Equivalencies between terms used in surveys and in experimental design . . . . . . . . . . . . . . . . 29
4.1 I/A ratio by colony 19901995 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.2 I/A ratios for 1996 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.1 Diversity measures for the abundance patterns in Figure 5.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.1 Four possible outcomes of a statistical test of a null hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . 70
7.1 Numbers of previously sampled plots from both NSR and SR cutblocks . . . . . . . . . . . . . . . . 93
7.2 Probability parameters for p(), prior distribution of (cutblock is NSR or SR),
and p(X|), conditional probability of observing X, given (cutblock is NSR or SR) . . . 94
7.3 The likelihoods (probability that X plots out of 12 are US given ) and the
posterior probability that the cutblock is NSR for all possible X values,
when the prior probability, 0 = 0.84 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7.4 Suggested cutoff values for the Bayes factor when comparing two hypotheses . . . . . . . . . . 97
7.5 Hypothetical gains for each combination of action and state of nature . . . . . . . . . . . . . . . . . 98
7.6 The Bayes posterior gain and posterior Bayes decision for all possible numbers of
understocked plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
8.1 A generalized decision table showing calculation of expected outcomes
for two potential management actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
8.2 Some possible arrangements that could be considered for a thinning experiment . . . . . . 117
8.3 Results of Sainsburys (1991) calculations of the benefits of different designs
for an active adaptive management experiment on groundfish in Australia . . . . . . . . . . . 120
9.1 Stages of an adaptive management project and sources of information
appropriate for each stage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
A1.1 Numbers of previously sampled plots (observed as US or S)
from both NSR and SR cutblocks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

vi


2.1 Relationship between the study units in a typical research experiment
and an adaptive management experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Design and analysis of an adaptive management experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.1 A classification of the methods considered in this chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 Relationship between degree of control, strength of inference, and type of study design . . . 21
3.3 Simplified outcomes in a BACI design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.4 Problems with the simple BACI design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.5 The BACI-P design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.6 The enhanced BACI-P design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.1 Comparing the development of a retrospective and prospective study . . . . . . . . . . . . . . . . . . . 46
4.2 I/A ratio vs colony size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.1 Examples of accuracy and precision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.2 Errors in x-values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.3 Estimated numbers of chinook salmon spawning in the
Upper Fraser Watershed near Tte Jaune Cache . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.4 Four dominance patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.1 The relationship between scientific and statistical hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.2 Variables influencing power to detect the difference between a
sample mean and a constant for the wood duck nest cavity example . . . . . . . . . . . . . . . . . . . . . 74
6.3 A posteriori power analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
7.1 Components of a Bayesian analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.2 Distribution of sample plots for the silviculture example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.3 Probability tree for the silviculture example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7.4 Decision tree for the silviculture example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
8.1 Changes in estimates of various physical constants as new experimental
or measurement methods were developed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
8.2 A simple example of a generalized decision tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
8.3 Decision tree for the Tahoe National Forest example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
8.4 Possible models for hypothetical data on average volume per tree
at age 100 years as a function of initial density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
8.5 Posterior probabilities for different slopes of a linear model . . . . . . . . . . . . . . . . . . . . . . . . . . 114
8.6 An example sensitivity analysis of Cohan et al.s (1984) decision analysis
on the Tahoe burning example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
8.7 Decision tree for the analysis of various management actions in
Sainsburys (1988) large-scale fishing experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
9.1 Causes and correlates: four examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

vii

1 STATISTICS AND THE PRACTICE OF ADAPTIVE MANAGEMENT

J. BRIAN NYBERG

Abstract

As adaptive management becomes more widely

recognized as a foundation element of good land
stewardship, many resource professionals are attempting to extend its theories and principles into common
practice. They wish to conduct powerful management
experiments, to monitor the outcomes effectively and
efficiently, and to use the resulting data to make reliable inferences for future decisions. Most managers,
however, have little formal training in the application
of experimental design and statistics to the problems
that they want to address through adaptive management. This chapter sets the stage for the in-depth
discussions of key aspects of statistics in adaptive management that are presented in subsequent chapters. It
includes a working definition of adaptive management, demonstrates the value of the application of
adaptive management to forestry issues, and explains
some of the differences between research studies and
adaptive management techniques.

1.1 Introduction
The concept of adaptive management (Holling [editor]1978) is steadily gaining wider acceptance in
forestry, especially in Canada and the United States
(e.g., Schmiegelow and Hannon 1993; Bormann et al.
1994; Nyberg and Taylor 1995; Covington and Wagner [technical coordinators] 1996; MacDonald et al.
1997). As a hybrid of scientific research and resource
management, adaptive management blends methods
of investigation and discovery with deliberate manipulations of managed systems. Through observation
and evaluation of the ways that human interventions
affect managed systems, new knowledge is gleaned
about system interactions and productive capacities.
This new knowledge is then applied to future decisions in a cycle of continuous improvement of
policies and field practices.
Adaptive management has somewhat different
goals from research and presents challenges that differ both in scope and nature from those posed by
typical forest research studies. Consequently, designing and analyzing adaptive management studies
involves more than simply transferring research techniques to management problems. Scientists can play

The terms manager and researcher

are used here in the following senses:
Managers (or resource managers) are responsible
for making or approving decisions about forest
resource use and conservation. Although there
are exceptions, resource managers in British Columbia usually have university or technical
institute training to the level of the undergraduate degree or diploma, and are often registered
as professional foresters, agrologists, engineers,
geoscientists, or biologists. Resource managers
are usually employed by government agencies or
private forest companies. To understand the main
ideas in this report, and to be effective in implementing adaptive management, managers should
have a basic academic background in statistics
and subsequent field experience in making or
contributing to complex resource decisions.
Researchers are applied scientists, usually from
government agencies or universities, who are responsible for conducting scientific studies of
forest ecology and management. Their goals include both furthering knowledge of forests and
explaining how human actions affect forests. In
addition to their expertise in forestry or related
disciplines, researchers usually have post-graduate
training in statistical methods and experimental
design. To benefit fully from this report, however,
they should also have considerable experience in
conducting forest research.

an important role in adaptive management (Walters

1986), but it is local resource professionals who must
become the adaptive managers if the promise of
the concept is to be realized through its application
to a large proportion of forested lands. As part of
their everyday jobs, these managers (see above for
clarification of the term) must be able to implement
or even design studies that produce reliable information about issues that concern or challenge them.
This suggests that resource managers might need
to use statistics in such studies. Few field-level
managers, however, have experience in applying
experimental designs and statistical methods, even in

situations suited to the classical statistical techniques

taught in most universities and colleges. Furthermore, the characteristics of some adaptive management studies make them unsuitable for many
familiar designs and methods, including analysis of
variance (ANOVA). Alternative approaches such as
Bayesian statistics and meta-analysis can be helpful in
some of these problematic cases, but most resource
managers are not familiar with these approaches.
To be informative and efficient, adaptive management projects must be led by people who know what
options for study designs and analyses are available,
and the relative strengths and weaknesses of each.
This is a reasonable if ambitious objective for
resource managers, whose role in adaptive management usually includes articulating questions,
selecting among alternative courses of action, and
then implementing those actions. For the researchers
and biometricians who often advise managers on the
details of study designs, sampling, and analysis, a
more comprehensive understanding of the various
statistical techniques is required.
This report has been designed as a guide to statistical methods appropriate for adaptive management
studies, with material that should interest both managers and researcher scientists. It should serve as an
introduction for some resource managers and a refresher for others on statistical methods, their
strengths and weaknesses, and their suitability for
studies of different types of management problems.
For researchers and biometricians, it should provide
a refresher on classical (familiar) methods, an introduction to less familiar methods, and a discussion of
the typical challenges that will be faced in applying
both to the unfamiliar situations of adaptive management. Although all the methods discussed here have
been previously described in other texts and reports,
that material is widely scattered in the literature and
is thus not easily available to forestry practitioners.
This report brings them together under one cover
and deals directly with their application to adaptive
management of forests.
The design of studies and analysis of datathe
themes of this reportare only two components of
the much larger topic of adaptive management. The
following section explains the procedural framework
of adaptive management. For information on other
aspects, including conceptual foundations and implementation, refer to Holling (editor, 1978), Walters
(1986), Lee (1993), Gunderson et al. (1995), and Taylor et al. (1997). In addition, Taylor et al. (1997)

include a comprehensive list of other references.

1.2 Towards a Working Definition
Adaptive management is a systematic process
for continually improving management policies
and practices by learning from the outcomes of
operational programs. Its most effective form
active adaptive managementemploys
management programs that are designed to experimentally compare selected policies or
practices, by evaluating alternative hypotheses
about the system being managed. The key characteristics of adaptive management include:
acknowledgement of uncertainty about what
policy or practice is best for the particular
management issue;
thoughtful selection of the policies or practices
to be applied;
careful implementation of a plan of action designed to reveal the critical knowledge;
monitoring of key response indicators;
analysis of the outcome in consideration of the
original objectives; and
incorporation of the results into future decisions.

Increasing use of the term adaptive management

by different agencies in different settings (e.g., Lancia
et al. 1996; Namkoong 1997) has spawned various interpretations and misinterpretations of its meaning.
Consequently it is for many little more than a fuzzy
concept. To bring the concept into sharper focus and
to encourage a shared understanding of adaptive
management among resource professionals in British
Columbia, Nyberg and Taylor (1995) proposed the
definition listed in the text above.
This definition suggests that adaptive management
must comprise an organized sequence of activities.
The sequence begins with a thorough analysis of the
problem being faced and then proceeds to the creation of a management plan that is designed to speed
learning about the system. It is not complete until the
planned management actions have been implemented, measured, and evaluated; and the resulting new
knowledge has been fed back into the decisionmaking process to aid in future planning and management. This sequence of steps can be summarized
as a six-step process: (1) problem assessment, (2) project design, (3) implementation, (4) monitoring,
(5) evaluation, and (6) adjustment of future decisions.

The sequence may need to be repeated in a continuing learning cycle if uncertainties remain unresolved
or new ones appear.
This report deals mainly with the second, fourth,
and fifth steps in the adaptive management process,
namely the design (thoughtful selection) of practices
to be studied, the measurement (monitoring) of responses, and the evaluation (analysis) of results.

rigour so as to provide reliable information in a

timely and cost-efficient manner. As part of the design process it is also critical to consider the statistical
methods that will be used to analyze the resulting
data. The following chapters describe methods that
can be used to enhance the value of data from studies
that pose some of the design problems listed above.
1.4 Need for Adaptive Management

1.3 Experiments in Adaptive Management

Adaptive management can take two different modes:
active and passive (Walters and Holling 1990). A critical feature of both modes is thorough exploration,
often through a modelling or gaming process, of
the potential effects of policies or practices that are
being considered for implementation. In passive applications only one policy or practice is explored,
whereas in active adaptive management multiple options are compared and contrasted. In both cases
subsequent management activities reveal, through
monitoring and evaluation of their results, the accuracy or completeness of the earlier predictions. These
deliberately designed activities are experiments in
the broad sense of the term; that is, deliberate tests or
trials intended to provide information about the response of the system of interest.
The notion of experimentation is central to
adaptive management. As Lee (1993, p. 9) puts it,
Adaptive management...embodies a simple imperative: policies are experiments; learn from them. In
fact, experimentation is the element that ultimately
distinguishes adaptive management and experimental research from other approaches to learning about
nature. These other approaches, including the retrospective and observational studies described in later
chapters of this report, can contribute helpful knowledge to later adaptive management work, but they
are not themselves adaptive management because
they do not include deliberately planned experimental manipulations.
Experimentation is considered at some length in
this report, but it is defined here quite broadly compared to many scientists concept of a scientific
experiment. For reasons of scale, expense, and others, adaptive management experiments will not
always include controls, replication, multiple treatments, randomization, or other features commonly
expected of traditional scientific research. Nevertheless, those designing adaptive management
experiments should strive to balance practicality with

Uncertainty drives adaptive management (Walters

1986). There would be little need to develop new
policies or methods if managers were dealing with
stable, predictable ecological and social systems. The
outcomes of management programs could be reliably
predicted, and standard practices could be taught to
each generation of young professionals. Adaptive
management and other approaches for dealing with
uncertainty would be of little value.
Resource managers, however, do not live in such a
world (Hilborn 1987). Uncertainties are pervasive in
their work. The major categories of uncertainty that
trouble managers when they consider the future are:
natural environmental variability (e.g., weather,
fire, earthquakes, avalanches, volcanoes, stream
flows, genetic composition of species, animal
movements);
human impacts on the environment through global climate change, new technology, and the growing population;
lack of knowledge about most aspects of the
ecosystems being managed; and
variations in social and political goals expressed as
varying budgets, shifting policy directions, and
changing demands for commodities, services, and
aesthetic values from forests.
Given that resource managers and policy makers
are faced with such difficult challenges, what can they
do? Scientific research is one avenue for addressing
the problem of lack of knowledge, but research programs often take years to organize, carry out, and
report results. Meanwhile, resource management decisions continue to be made and forests continue to
fall and regenerate under human hands. Money and
expertise for research in forestry and other natural
resource disciplines continue to be constrained at
levels far below those needed to address many important issues. Furthermore, scientific research is limited
in the types of questions it can answer because many
forestry practices have cumulative effects that are

only apparent at scales of time, space, or both that

are not amenable to investigation through traditional
experimental research. For example, it is impossible
to use classical experimental methods employing
controls and replicated treatments to determine the
effects of forestry on wildlife that use huge areas,
such as caribou, or that are threatened with extinction or local extirpation, such as spotted owls.
When research, education, or personal experience
fail to provide information needed for difficult decisions, managers typically turn to professional opinion followed by unstructured trial-and-error
management. This approach to learning is often inefficient and unreliable. Unless management alternatives are carefully thought out and attention is paid
to potentially confounding factors such as biases,
random errors, and unmeasured influences of weather, site, or other factors, it is often impossible to say
what really caused any observed response. This can
lead to myths being accepted widely due to the
strongly held opinions of one or a few people
opinions that are later found to be wrong.
For example, poorly conceived and unsuccessful
field trials may have been the genesis of the formerly
strong bias among foresters in central British Columbia against partial cutting in high-elevation stands of
spruce (Picea spp.) and subalpine fir (Abies
lasiocarpa). Until a few years ago, many believed that
partial cutting was unsuitable in any and all sprucefir stands. This belief was based largely on reports
that stands that had been partially cut before 1970
had all been subsequently windthrown or infested by
insects or disease. Recent partial cutting trials have
shown, however, that spruce-fir forests can be windfirm and healthy (for several years, so far) if the
harvest intensity and site and stand conditions are
appropriate.
In contrast to the basic trial-and-error approach,
adaptive management is a much more organized and
powerful approach to learning from experience.
Its greatest contribution to learning may lie in the
notion of making explicit predictions of the expected
outcomes of management actions, then comparing
actual outcomes to the predictions before adjusting
subsequent actions and the models used to make the
initial predictions. By designing management actions
as experiments stronger inferences can be drawn
from their outcomes, reducing the chance of generating false notions about forest functions and impacts.
Other potential benefits of adaptive management include more reliable answers to questions about

effects of forestry over large geographic areas and

long time frames; insight into the causes and processes of system responses; shared understanding among
scientists, policy makers, and managers; systematic
resolution of conflicts over policies and practices;
and efficient use of staff and budgets to address clear
objectives (Holling [editor] 1978; Lancia et al. 1996;
Taylor et al. 1997). All of these benefits contribute to
accelerated learning and to the ultimate goal of improved decisions and forest management in future.
As an example of a potential application of adaptive management, consider the problem that resource
managers face when they examine the question of
how to protect water quality and downstream fish
habitat in small headwater streams, while still allowing some timber harvesting to take place nearby. This
situation creates a common and difficult problem in
areas of British Columbia where small streams are
numerous, slopes are steep, and timber values are
high. The weight of expert opinion and of evidence
from larger streams suggests that some streamside
vegetation must be retained to provide shade and leaf
litter, prevent sedimentation, and prevent degradation of bank and channel structure. If large trees are
left standing in a narrow (<30 m) riparian strip after
logging of areas near small streams, however, they
are often blown down by wind, which may cause serious soil disturbance and bank damage as their root
wads pull up. The loss of potential timber revenue in
the fallen trees is exacerbated by the risk of insects
(especially bark beetles) colonizing riparian blowdown and infesting other stands. Therefore, simply
leaving narrow reserve zones of unlogged timber
along all headwater streams is often not economically
or ecologically appropriate.
Because past experience and scientific research
have not resulted in a reliable approach to managing
riparian areas adjacent to the smallest streams, what
can resource managers and scientists do? If the situation allowed, the easy answer might be to cease all
logging within, say, 50 m of such streams. But agency
direction and societal demands do not allow such a
hands off approach. Resource managers could instead postpone logging in all such areas while waiting
for more intensive research to be done; or they could
simply pick one or more methods and carry on logging while hoping for the best. Neither of these
alternatives is a suitable response in times of restricted research funding and rising public expectations
for resource stewardship.
By designing management actions as experiments,

as part of adaptive management, logging could proceed and at the same time yield important information on key indicators of stream conditions, fish and
wildlife habitat, and responses of trees and other vegetation. Through modified logging operations managers could apply several alternative treatments, such
as various widths of unharvested reserve zones or different degrees of partial cutting of riparian zones, or
both. Among other benefits, such a program might
reveal the cumulative effects on watershed dynamics
and wildlife habitat of large and widespread treatmentseffects that could not be studied in a more
traditional, small-scale research experiment.
Because adaptive management has attributes common to scientific research and sometimes draws on
research techniques such as controlled experiments,
some may assume that an ambitious program of
adaptive management would obviate the need for research. They would be wrong, for intensive research
is far more suited than adaptive management to
answering some questions. Intensive research can
produce deeper knowledge of selected system
processes, such as mechanisms of physiological response in seedlings exposed to varying temperature
and moisture regimes, than could adaptive management. It may also be the only approach suitable for
sorting out the interacting effects of a number of factors on some dependent variable. This in-depth
knowledge may be crucial for building models used
to forecast how the overall system will respond to
management.
Adaptive management is most suited to selecting
amongst alternative courses of action, such as different partial cutting treatments that could be applied to
a particular site and stand type. It can also be helpful
for testing the modelled responses of managed systems against real-world results of management,
across a much wider range of conditions than could
any practical program of intensive research.
In fact, research and adaptive management complement each other, so that the application of both
approaches to a problem will almost certainly lead to
better results than use of either alone. Adaptive management can reveal management surprises;
research can help to explain them.
Adaptive management may be valuable to anyone
facing substantial uncertainty about the best course
of management action, as long as that person has or

can obtain the authority to implement a program

that leads to learning. In cases where political or
other pressures have produced moratoriums or other
forms of management gridlock, such authority
may not be easily obtained.1 Where management interventions are going to proceed (e.g., when delays
carry unacceptable social or ecological risks), then
much can be gained by treating them as opportunities to learn. In some cases the only way to discover
how an action will affect a system is to actually try it
(Walters 1986). This is especially true for complex,
large-scale systems and effects of cumulative actions.
Although the emphasis in this report and most
other literature is the application of adaptive management to natural resources, the approach is equally
valuable for other fields. Adaptive management
shares much of its theoretical basis with the concepts
of continuous improvement in business (Deming
1986; Walton 1986), adaptive control process theory
in engineering, and operations research and management (McLain and Lee 1996). Wherever an
organized, experimental approach to difficult management questions is needed, adaptive management
could be of help.
1.5 Role of Statistics
In forest management, data and mathematical analyses are central to management decision-making, and
statistical methods play several important roles.
Modern forestry is based to a great extent on statistical descriptions of the characteristics of forests and
forest products, such as timber inventories; and on
inferences about the expected future conditions of
forests and habitat, including growth and yield relations. Statistics are also crucial in understanding how
forest resources respond to human and natural perturbations, because they allow us to distinguish
treatment effects from random and sampling
errors.
In adaptive management, statistical methods also
play a critical role. Adaptive managers will often want
to measure the initial state of the systems they administer, and they will usually need to monitor
trends over time that show the systems responses
to management policies or practices. In evaluating
outcomes, they will want to draw inferences about
the causes of any changes that are detected in the

1 It can be argued that adaptive management is the best way to resolve gridlock based on mistrust of resource managers by public groups or
other stakeholders. The same is true where the impasse arises from competing ideas and values, fear of consequences, or other concerns
rooted in lack of knowledge of the forests responses to management.

system to decide how and when to adjust actions in

the future or at comparable sites. In all of these applications statistical techniques can provide important
insights into both qualitative and quantitative measurements. Statistical analyses allow managers to
discern small but important differences in data sets,
and to distinguish patterns of correlation and interaction from background variation and sampling
errors.
Careful design of management experiments or operational trials is often the first step towards gaining
data from which reliable inferences can be drawn.
Whenever possible, adaptive management studies
should include experimental controls, unbiased sampling and allocation of treatments, and replication of
treatments. However, it is important to recognize
that the operational scale and setting of adaptive
management studies may constrain the level of statistical rigour that can be achieved. It may be
impossible, for example, to find multiple areas that
are sufficiently homogeneous to serve as replicates of
operational-scale treatments. In other cases, it may
not be feasible to meet some of the critical assumptions of the classical methods of statistical analysis,
including random allocation of treatments, homogeneity of variance, and independence of sample
variances. Perhaps even more significant is the fact
that frequentist statistical methods such as
ANOVA and regression analysis are not designed to
answer common management questions such as
What is the probability of a 50% increase in windthrow after partial cutting?
As a result, classical methods will be useful in some
adaptive management studies but not in others.
When classical methods are not appropriate, a proposed study may still be worthwhile if alternative
types of analyses can reveal important insights from
the data.
1.6 Structure of the Report
The next eight chapters provide an overview of principles and methods for a wide range of approaches to
experimentation and data analysis in operational
forestry settings. Beginning with a review of basic
concepts of experimental design and classical methods of statistical analysis in Chapter 2, the report then
covers other common approaches to studying natural
systems (Studies of Uncontrolled Events and Retrospective Studies in Chapters 3 and 4, respectively).
Chapter 5 (Measurements and Estimates) and

Chapter 6 (Errors of Inference) discuss common

mistakes in interpretation of data and statistical results, and suggest ways to avoid them. Chapters 7 and
8 give an overview of Bayesian statistics and decision
analysis, both of which are unfamiliar to many resource managers but which have great potential value
in adaptive management. Chapter 9 synthesizes the
methods discussed in this report and presents a simplified users guide to the value of different types of
information in adaptive management.
Each of the next seven chapters explains what a
statistical method can do for a project leader, when it
should be used, and what its limitations are. The authors have limited the use of formulas, mathematical
notation, and statistical jargon as much as possible,
without making the information superficial. Nevertheless, some of the concepts and methods discussed
will be unfamiliar and challenging, and some sections
may have to be read several times. Most of the technical terms are defined in the glossary at the back of
the report. Since this handbook is not comprehensive
we encourage readers wanting more detailed information to consult the references listed at the end of
each chapter. Most importantly, project leaders
should consult throughout their projects with biometricians or experienced researchers to ensure that
powerful and cost-efficient methods are used.
Finally, we recognize that there are approaches to
generating and analyzing data that this report does
not cover. For example, some readers will already be
familiar with the methods referred to as combining
information (Draper et al. 1992) and meta-analysis (Fernandez-Duque and Valeggia 1994). Because
the focus is on forestry, there is little attention to the
quantitative methods for fish and wildlife population
analysis that are treated in many adaptive management papers (e.g., Walters 1986). Readers should also
remember that this report addresses only one of the
many issues that need to be considered if adaptive
management is to succeed widely. Greater challenges
may lie in social and institutional aspects of implementing adaptive management, such as the risk
aversion of some managers (Walters and Holling
1990), inadequate institutional structures and
stakeholder participation (McLain and Lee 1996), incomplete or ineffective implementation of the study
plan, (C.J. Walters, pers. comm., 1995) uncertain or
inadequate funding for monitoring and analyses
(McLain and Lee 1996), lack of commitment to
reporting (Taylor et al. 1997), and institutional
memory loss about what has been learned

(Hilborn 1992). These problems too need careful

consideration, innovative thinking, and personal
commitment to shared learning.
With the stage now set, the remainder of this report presents ideas and methods that should be
valuable to anyone who uses or needs to use quantitative analyses in forestry. Throughout, the focus is
on designing more powerful and informative management experiments. There is meat, however, for
both statisticians and experienced researchers. We
hope that the ideas in the following chapters will contribute to more effective, efficient learning in many
different situations.
References
Bormann, B.T., P.G. Cunningham, M.H. Brookes,
V.W. Manning, and M.W. Collopy. 1994. Adaptive ecosystem management in the Pacific
Northwest. U.S. Dep. Agric. For. Serv., Gen.
Tech. Rep. PNW-GTR-341.
Covington, W. and P.K. Wagner (technical coordinators). 1996. Conference on adaptive
ecosystem restoration and management:
restoration of Cordilleran conifer landscapes of
North America. Proc. conf. June 68, 1996,
Flagstaff, Ariz. U.S. Dep. Agric. For. Serv., Gen.
Tech. Rep. RM-GTR-278.
Deming, W.E. 1986. Out of the crisis. MIT Center for
Advanced Engineering Study, Cambridge,
Mass.
Draper, D., D.P. Gaver. Jr., P.K. Goel, J.B. Greenhouse, L.V. Hedges, C.N. Morris, J.R. Tucker,
and C.M. Waternaux. 1992. Combining information: Statistical issues and opportunities for
research. National Academic Press, Washington, D.C., Contemporary Statistics No. 1.
Fernandez-Duque, E. and C. Valeggia. 1994. Metaanalysis: a valuable tool in conservation
research. Cons. Biol. 8:55561.
Gunderson, L.H., C.S. Holling, and S.S. Light (editors). 1995. Barriers and bridges to the renewal
of ecosystems and institutions. Columbia Univ.
Press, New York, N.Y.
Hilborn, R. 1987. Living with uncertainty in resource
management. N. Am. J. Fish. Manage. 7:15.

______. 1992. Can fisheries agencies learn from experience? Fisheries 17:614.
Holling, C.S. (editor). 1978. Adaptive environmental
assessment and management. J. Wiley, London,
U.K.
Lancia, R.A., C.E. Braun, M.W. Collopy, R.D.
Dueser, J.G. Kie, C.J. Martinka, J.D. Nichols,
T.D. Nudds, W.R. Porath, and N.G. Tilghmann. 1996. ARM! for the future: adaptive
resource management in the wildlife profession. Wildl. Soc.Bull. 24:43642.
Lee, K.N. 1993. Compass and gyroscope: integrating
science and politics for the environment. Island
Press, Washington, D.C.
McLain, R.J. and R.G. Lee. 1996. Adaptive management: promises and pitfalls. Environ. Manage.
20:43748.
MacDonald, G.B., R. Arnup, and R.K. Jones. 1997.
Adaptive forest management in Ontario: a literature review and strategic analysis. Ontario
Min. Nat. Resour., For. Res. Info. Pap. 139.
Namkoong, G. 1997. A gene conservation plan for
loblolly pine. Can. J. For. Res. 27:4337.
Nyberg, J.B. and B. Taylor. 1995. Applying adaptive
management in British Columbias forests. In
Proc. FAO/ECE/ILO International Forestry
Seminar, Prince George, B.C., Sept. 915, 1995,
pp. 23945. Can. For. Serv., Prince George, B.C.
Schmiegelow, F.K.A. and S.J. Hannon. 1993. Adaptive
management, adaptive science and the effects
of forest fragmentation on boreal birds in
northern Alberta. Trans. N. Am. Wildl. Nat.
Resour. Conf. 58:58497.
Taylor, B., L. Kremsater, and R. Ellis. 1997. Adaptive
management of forests in British Columbia.
B.C. Min. For., For. Practices Br., Victoria, B.C.
Walters, C.J. 1986. Adaptive management of renewable resources. McGraw-Hill, New York, N.Y.
Walters, C.J. and C.S. Holling. 1990. Large-scale
management experiments and learning by
doing. Ecology 71:20608.
Walton, M. 1986. The Deming management method.
Perigee Books, New York, N.Y.

2 DESIGN OF EXPERIMENTS
AMANDA F. LINNELL NEMEC

Abstract

Experimentation is essential for making well-informed

decisions about the management of complex forest
ecosystems. Although experiments in forest management are generally more complicated than the typical
research experiment, many issues, such as protection
against bias, repeatability of results, efficient use of
resources, and quantification of uncertainty, are the
same. Therefore, managers, as well as researchers, can
benefit from a good understanding of the principles of
sound experimental design. This chapter reviews some
fundamental concepts of experimental design and the
assumptions that provide a basis for classical statistical
inference. The importance of clear objectives and the
need for replication, randomization, and blocking are
emphasized. Practical limitations of the classical approach, as it applies to the design of adaptive
management experiments, are discussed briefly.

2.1 Introduction
Successful management of our forests is a dynamic
process in which current programs are continually
monitored and adapted as new information becomes
available and policies change. Because the outcome
of decisions is always uncertain, managers often experiment with new strategies to help determine the
best course of action. Although such tests do not
necessarily conform to the standards of a strictly controlled research experiment, many issues, such as
elimination of bias, repeatability of results, efficient
use of resources, and quantification of uncertainty,
are the same. For this reason, managers, as well as researchers, can benefit from a good understanding of
the principles of sound experimental design. This
chapter reviews the theory of classical experimental
design and the assumptions that provide a basis for
statistical inference from experimental data. Application of traditional theory to the design of adaptive
management experiments is considered.
The literature on the design of experiments is vast.
Classical books, such as The Design of Experiments by
Fisher (1935), The Design and Analysis of Experiments
by Kempthorne (1952), Experimental Designs by
Cochran and Cox (1957), and Planning of Experiments
by Cox (1958), remain valuable sources of guidance

and are recommended reading. Since the early work

of Fisher, the number of books and papers on experimental design has exploded, as indicated by the list of
more than 200 references assembled over 10 years ago
by Steinberg and Hunter (1984). During the last 10
years, improvements in computer technology have
encouraged statisticians to improve and expand their
efforts even more (e.g., refer to Atkinson 1982, 1988,
1996; Bates et al. 1996), so the trend persists. Unfortunately, much of the recent work is inaccessible to
managers because the language is overly technical
and difficult to understanda concern expressed by
Pike (1984) and several others who discuss the review
by Steinberg and Hunter (1984). Moreover, there is
usually a long delay before new methods are accepted
and incorporated into popular statistical software
packages. Therefore, readers seeking practical information on the design and analysis of experiments are
advised to consult general textbooks (e.g., John 1971;
Box et al. 1978; Steel and Torrie 1980; Mead 1988;
Montgomery 1991; Milliken and Johnson 1989, 1992)
and papers written specifically for managers or applied scientists (e.g., Stafford 1985; Penner 1989).
2.2 Definitions
An experiment is the manipulation of a population
or system (e.g., partial cutting of a forested area) as a
means of gathering information. The source of experimental material (e.g., trees, vegetation, wildlife)
is called the experimental population. In an ideal experiment, the experimental population is the same as
the target population, or forest ecosystem, to which
the results are eventually to be applied. In practice,
the two populations necessarily differ in time and
perhaps space or scale as well.
In a typical research experiment, the experimental
population might consist of: the trees in a nursery; a
collection of relatively small plots of land, which are
distributed over one or more geographic areas; or
any other set of entities. Each tree, plot, or entity is
an experimental unit that receives one of several
treatments (e.g., levels of fertilizer, partial cutting or
no cutting). These units are generally too large to
measure in their entirety and are subdivided into a
set of smaller sampling units (e.g., branches or
subplots) from which a suitable subset is selected for

measurement. To monitor trends or other effects of

time, measurements are repeated at suitable intervals
over a period of months or years.
The results of research experiments are of limited
value for making inferences about the effects of treatments or actions applied to management units. Management units (e.g., a forest stand or polygon, a
mapsheet, or a timber supply area) are areas of forest
that are convenient for administration or costeffective operations, and therefore are considerably
larger than research plots. The impact of disturbances created by large-scale operations in these units
cannot, in general, be deduced from small research
plots, where effects such as fragmentation, soil erosion, and changes in vegetation or water quality
might not be evident. Proponents of adaptive management (e.g., Walters 1986; Walters and Holling
1990; Taylor et al. 1997) argue that successful management of complex biological systems requires fullscale testing. These experiments, which are known as
adaptive management experiments, are used to test
entire management plans, with the management unit
serving as the experimental unit as illustrated in Figure 2.1. In an adaptive management experiment, one
or more systems are monitored regularly over time
and decisions about treatments or other interventions are made as the experiment progresses. Because
management units are large and complex, they must
be broken down into suitable sampling units for observation and evaluation. In this respect, adaptive
management experiments resemble research experiments, although the number and type of sampling
units might differ.

Forest ecosystem

Management unit

Experimental unit
.

10

Relationship between the study units in a

typical research experiment (left) and an
adaptive management experiment (right).

2.3 Objectives
The first requirement in any experiment is a clear
statement of the goals. Paraphrasing Box et al. (1978,
p. 15), the purpose of an experiment must be clear
and agreed upon by all interested parties; there must
be agreement on what criteria will determine whether
the objectives have been accomplished; and, finally,
in the event that the objectives change during the
course of the experiment, an arrangement must be
made by which all interested parties are informed of
the change, and agreement on the revisions can be
reached. The importance of these points cannot be
overemphasized. Without clear objectives, the outcome of an experiment is predictable: ambiguous
results and wasted time, money, and valuable (possibly irreplaceable) resources. Unnecessary waste is
always unacceptable. However, when large management units are involved, the costs can be devastating.
Defining the objectives of an experiment requires
careful consideration of the components that make
up the system under study, the forces that drive the
system, and the best means of extracting information
about both. In a small-scale research experiment, attention might reasonably be restricted to relatively
simple questionsfor instance, how does tree
growth differ under various controlled conditions?
The objectives of adaptive management experiments
typically concern more complicated issuessuch as,
how is biodiversity affected by forest practices?
In both cases, general scientific concepts (e.g., tree
growth and biodiversity) must be stated in terms of
well-defined, measurable quantities (e.g., height or
diameter increment, number of species). These
quantities provide a concrete basis for planning
experiments and for analyzing the results.
The objectives of an experiment are often posed as
hypotheses to be tested or parameters to be estimated. Special care must be taken to ensure that the
hypotheses are sensible and that the parameters are
useful for making decisions. In classical hypothesis
testing, a so-called null hypothesis is retained unless
there is convincing evidence to the contrary. Based
on the outcome of the experiment, the null hypothesis is either retained or rejected in favour of a specific
alternative hypothesis. (See Anderson, this volume,
Chap. 6, for a discussion of the associated errors of
inference.) The null and alternative hypotheses
should be defined so that both outcomes (i.e., acceptance or rejection of the null hypothesis) represent
reasonable and informative conclusions that would

justify the cost of the experiment. Implausible null

hypotheses should be rejected at the outset and replaced with something more relevant. For instance,
no sensible person would ever accept the hypothesis
that partial cutting has no impact on forest ecosystems; a more reasonable hypothesis is that partial
cutting does not reduce a key parameter (e.g., the
number of bird species) below some critical value.
The alternative hypothesis is equally, if not more, important than the null hypothesis. An experiment that
is designed to detect one type of departure from a
null hypothesis might be completely ineffective in detecting other types of deviations. For example, an
experiment to evaluate short-term impacts of partial
cutting will generally provide little information about
long-term effects.
2.4 Principles of Experimental Design
An experimental design is a detailed plan describing
all aspects of an experiment (see Bergerud 1989b).
Most experimental designs include a sampling design, which describes the nature of the sampling
units, the number of sampling units, the method of
selection, and the variables to be measured. The experimental design depends on the purpose of the
experiment and thus the entire reasoning process by
which the experimenter really hopes to link the facts
he wishes to learn from the experiment with those of
which he is already reasonably certain (Mandel 1964,
p. 2). Experimental designs are characterized by three
main components: (1) the factors and factor levels to
be investigated, (2) the amount and type of replication, and (3) the method of randomization, including
any blocking. Each of these elements should be considered carefully to assure that all data pertaining to
the objectives are collected (and can be analyzed) and
that the data be collected in the most efficient way
(i.e., optimum results for a minimum cost).
2.4.1 Experimental factors
Experimental factor means any treatment1 or variable
(e.g., harvesting method, species composition, age)
that is controlled in an experiment, either by physically applying a treatment to an experimental unit or
by deliberately selecting a unit with a particular characteristic. A covariate may be any other variable that
is measured but not influenced by the experiment.
Experiments are distinguished from observational

studies by the investigators ability to control the

experimental conditions. (Refer to Eberhardt and
Thomas 1991, and Schwarz, this volume, Chap. 3, for
a discussion of the differences among experiments,
observational studies, and various other types of
study.) This control helps justify inferences about
cause and effect. For example, to determine the best
method of partial cutting to minimize the impact on
wildlife, various levels of volume removal might be
tested. This approach would probably be more informative than simply measuring volume in existing
partial cuts and observing a correlation with number
of animals. Correlation proves little about cause and
effect because both variables might be correlated with
a third variable (e.g., elevation). An interesting discussion of inferences about cause and effect can be
found in Holland (1986).
Usually many types and levels of factors might be
investigated in an experiment. These must be selected
carefully to meet the study objectives. A simple experiment might have only one relevant factor (e.g.,
method of logging), which has relatively few predetermined values or levels (e.g., clearcutting, partial
cutting, or no cutting). An experimental design that
involves only one such factor is called a one-way design with a fixed factor (effect). A design with two
such factors is called a two-way design, a design with
three factors is a three-way design, and so on. In
some applications, the factor levels are not known
ahead of time (although the number of levels is fixed)
but are randomly selected from a well-defined population of levels. Such factors are known as random
factors (effects). For instance, a manager might want
to compare bird diversities in a random sample of 20
white pine stands from a particular site series. Here
stand is a random factor with 20 levels. Factors with a
continuum of possible values (e.g., age, diameter),
which commonly occur in regression designs, are
called continuous factors or simply variables.
Two or more factors can occur in various combinations. If an experiment includes every level of one
factor in combination with every level of the other
factor(s) then the factors are crossed and the design is
called a factorial design. The advantages of factorial
designs over designs that vary one factor at a time are
twofold: efficient replication and estimates of interactions. In some cases, the set of possible levels of one
factor depends on the level of another, in which case
the former is said to be nested in, rather than crossed

1 The term treatment will hereafter refer to a particular set of conditions, an action, or an entire management strategy.

11

with, the latter. The (relative) moisture level of a

stand might, for example, be classified differently depending on the subzone in which the stand is located.
Most well-designed experiments include a control or
standard by which the effectiveness or impact of
treatments can be judged. In a one-way design, the
control might be one level of the treatment (e.g.,
clearcut, partial cut, and an uncut control), while a
two-way factorial design might include a separate
control (e.g., clearcut with and without mechanical
site preparation, partial cutting with and without site
preparation, and an uncut control with no site preparation). Bergerud (1989a) describes the analysis of the
latter type of design.
The overall effect of a factorthat is, the average
effect for all levels of the other factorsis called its
main effect. If the effect of one factor depends on the
level of one or more other factors then there is an interaction among the factors. A certain amount of
fertilizer might, for example, increase the height of
one species but have a smaller, or even opposite, effect on another species; in this case, there is an
interaction between species and fertilizer. Factorial
designs allow an investigator to study many interactions in the same experiment. However, if some
combinations are omitted for any reason then certain
main effects or interactions may be inseparable from,
or confounded with, others. Thus the magnitude of
confounded effects cannot be estimated. In such circumstances, the experimenter must be sure that the
confounded effects are of little or no interest, or can
be assumed to be negligible.
2.4.2 Replication
Replication, a standard means of validating scientific
findings, is a cornerstone of the theory of experimental design as laid down by Fisher (1935). Experimental
conditions are replicated when the same combination of factors occurs in more than one experimental
unit. This replication provides an estimate of the experimental error, which is any variation that cannot
be explained by the experimental factors (e.g., sampling or measurement error, and natural variation
among the experimental units). In the absence of
replication, there is no way, without appealing to
nonstatistical arguments, to assess the importance
of observed differences between the experimental
units.
The type of replication should be consistent with
the objectives. The experimental units should, therefore, be as similar as possible to the elements of the

12

target population. For instance, test sites should be

selected from the same geographic area as the target
population and treatment plots should be comparable in type and size to management units. Experimentation at a single site limits the applicability of
results to a small geographic region, while failure to
identify the appropriate unit of replication results in
pseudoreplication (Hurlbert 1984; Bergerud 1988,
1991) and possibly erroneous conclusions about the
nature of an effect.
In addition to selecting the size and type of replicate, the optimum number of experimental units
(and sampling units) must be determined. Sample
sizes should be large enough that definitive conclusions about the size of an effect or the validity of a
hypothesis can be made (i.e., parameter estimates
must be sufficiently precise and tests of hypotheses
must be decisive). Simple sample size calculations
provide an estimate of the minimum number of
replicates and sampling units needed to meet this
goal (see Nemec 1991; Bergerud 1992). If the required
sample size is unrealistic, because of the high cost of
treatment or sampling, then the objectives and design
of the experiment should be re-evaluated to determine whether an observational or retrospective study
(see Schwarz, this volume, Chap. 3, and Smith, this
volume, Chap. 4) might be more cost effective.
2.4.3 Randomization and blocking
Randomization occurs when treatments are randomly assigned to the experimental units. Like replication, randomization is an essential element of
good design. It helps minimize the risk of bias by
ensuring that all unmeasured factors (e.g., soil moisture, nutrients, prevalence of root disease) are more
or less evenly distributed among the treatments, or
randomized out of the experiment. Thus each
treatment is expected to be assigned to an approximately equal number of wet and dry sites, nutrientrich and nutrient-poor sites, sites with low and high
levels of root disease, and so on.
Random assignment of treatments can be
accomplished in a variety of ways. In a completely
randomized design, treatments are assigned by
picking units at random from the experimental population. For instance, if there were two treatments
and 20 experimental units (10 replicates per treatment) then the units might be numbered from 1 to
20, with 10 numbers picked at random to determine
which units receive the first treatment and the
remainder receiving the second. Sometimes

experimental units have a variety of origins (e.g., different geographic regions) and therefore exhibit
considerable inherent variability. In such situations, a
substantial reduction in the experimental error can
often be achieved by separating the units into homogeneous groups, or blocks, according to origin, or
some other factor. Treatments are then randomly assigned within each block. This is a form of restricted
randomization because each treatment is constrained
to occur a fixed number of times in each block. In a
completely randomized design, the randomization is
unrestricted, which might result in a very uneven distribution of origins or other attributes among the
treatment groups. Many other designs with some
form of restricted randomization exist, including
split-plot designs (which impose constraints on the
assignment of treatments to two types of experimental units: main plots and subplots), Latin squares,
and lattice designs. Refer to Cochran and Cox (1957)
and Anderson and McLean (1974) for more information about these and other designs.
Random sampling is another way of avoiding bias
when factors, such as species or age, cannot be assigned. Experimental units should be randomly
chosen from the experimental population and sampling units should likewise be selected at random
from the experimental units. Random samples, unlike haphazard samples or judgement samples
(samples judged to be representative by an expert
who makes the selection), have known statistical
properties. This allows the precision of a result to be
estimated from the samplesomething that is not
possible with non-random sampling. For further discussion of the topic, see Deming (1960, pp. 3033)
and Schwarz (this volume, Chap. 3).
Randomization and random sampling are closely
related ideas leading, in many cases, to the same
mathematical models and equivalent data analyses
(Feinberg and Tanur 1987; Smith and Sugden 1988).
For instance, randomization with blocking is analogous to stratified random sampling, and split-plot
designs are comparable to cluster sampling. Despite
the parallels, randomization is traditionally discussed
in connection with experimental design, while issues
relating to sampling design (e.g., type and number of
sampling units, method of sampling) are reserved for
discussions of observational studies. For more information on the latter subject refer to Schwarz (this
volume, Chap. 3).
Randomization is one of the simplest ideals to
achieve in both small- and large-scale experiments.

The process of randomly assigning treatments to

units is the same regardless of the size or nature of
the unittree, research plot, or stand. All that is required is a list of units and a random number
generator. Despite its importance in eliminating bias
and its ease of application, randomization is sometimes resisted on grounds that environmental or
logistical constraints prohibit its use, or that it is
impractical. Although it might be tempting to assign certain treatments (or controls) to stands that
are most difficult to harvest, or to stands that are
most visible from neighbouring communities, such
practices are likely to influence the outcome and thus
invalidate any inferences about cause and effect (refer
to Section 2.5). If randomization of treatments is really not practical or possible then the whole purpose
of the experiment should be reconsidered.
Random selection of experimental units and sampling units can be more difficult to achieve than
random assignment of treatments. For instance, how
can a random sample of 25 trees be selected from a
stand without compiling a complete list of trees and
their locations? Likewise, how can a random sample
of needles be selected from a tree? Various ingenious
solutions have been proposed, such as the use of random bearings and random distances to locate sample
trees and randomized-branch sampling to sample individual trees (Gregoire et al. 1995). Thus both
randomization and random sampling (or a close approximation) are feasible in most experiments.
2.4.4 Other considerations
Replication, randomization, and blocking are usually
recognized as the most important principles of good
design. These have been expanded over the years (see
Atkinson 1982, 1988; Federer 1984; Steinberg and
Hunter 1984) to include such other features as orthogonality (which implies that all main effects and
interactions of the experimental factors can be estimated), balance (equal sample sizes), efficiency
(minimum cost), and, in recent years, optimality
(minimum variance). Efficiency and optimality,
which help ensure maximum precision for a fixed
sample size, are likely to remain important as long as
money and resources are scarce. As technology continues to improve, the emphasis on orthogonality
and other computational issues is expected to diminish. Mead (1990) has, for example, suggested that the
need for orthogonality, which excludes a large class
of potentially useful designs, primarily because of numerical complexity, should be re-evaluated in the

13

light of present computing power. As the role of experiments continues to evolve, new criteria and
principles will undoubtedly emerge (see Section 2.6).
Successful experimentation depends on more than
the choice of factors and the use of appropriate randomization, replication, and blocking. Many practical details must also be considered. All pertinent
measurements of the experimental (sampling) units
must be identified, appropriate field procedures and
data collection forms must be developed, and provision must be made for adequate supervision of the
data collection. In large-scale studies, coordination
and optimization of procedures are especially important. For a checklist of these and other aspects of the
planning and execution of a study, refer to Sit (1993).
2.5 Statistical Inference for Experiments
Statistical inference and experimental design are
closely linked. Design determines the kind of statistical inferences that are possible, while consideration
of the proposed method of analysis almost always influences design (e.g., sample-size calculations depend
on the hypothesis testing procedure or estimation
method that will be used). In fact, failure to contemplate how the data will be analyzed invariably results
in a poor design.
All statistical inferences are based on a set of assumptions that links the data to the experimental
population via the design. Thus inferences are necessarily confined to the experimental population.
Because the relationship between experimental and
target populations is unknown, extrapolation to the
latter is more or less conjecture and should be viewed
with caution, particularly when the gap between
experimental and target populations is large. The
validity of statistical inference in an experimental setting is discussed more fully by Deming (1953, 1975),
Box et al. (1978, Chap. 1), and Hahn and Meeker
(1993).
Analysis of variance (ANOVA) methods are commonly applied to experimental data, so much so that
discussions of experimental design are often more
about ANOVA than fundamental issues of design.
Both are concerned with sources of variation and the
estimation of experimental error. The basic premise
of ANOVA is that the observed variability in experimental data can be attributed to a finite number of
identifiable sources, including factors under the control of the investigator, uncontrolled experimental
errors, and various interactions. A simple, one-way,

14

fixed-effects design has a single experimental factor

(e.g., harvesting method) and a single source of
experimental error (e.g., inherent variability in the
stands that are treated). A split-plot design has two
sources of experimental error: variation among main
plots and variation among subplots. Proper data
analysis is possible only when the identification and
classification of the factors by type (e.g., fixed or random, nested or crossed) is consistent with the design.
Any restrictions on the randomization, which are imposed by design (e.g., blocking), must be duly noted
by the inclusion of appropriate error terms (a theme
emphasized by Anderson and McLean 1974).
The purpose of an ANOVA is to make inferences
about effects attributable to experimental factors
(refer to Sit 1995 for more information), while taking
into account uncertainty caused by errors. To do this
the data are assumed to be generated by a probability
model that includes all the relevant effects and has
three key assumptions: (1) selection of the experimental units and assignment of treatments are
independent of the response variables of interest,
(2) all random effects and experimental errors are
mutually independent, and (3) random effects and
experimental errors attributable to a common source
are identically distributed as (normal) random variables with a mean of zero. The first assumption is
often not stated explicitly but is crucial for making
causal inferences (refer to Holland 1986 for a mathematical explanation of this fact). If the variables of
interest somehow influence, either directly or indirectly, which experimental units receive a particular
treatment, then the potential for bias is clear (e.g., assignment of treatment on the basis on slope, aspect,
density of trees, or any other variable that might affect tree height precludes inferences about the effect
of treatment on height). Randomization and random
sampling are the only effective means of eliminating
this source of bias.
Independence among errors and random effects is
another important assumption of ANOVA. There are
two common departures from this condition: temporal and spatial autocorrelation. In forestry studies,
correlation among repeated measurements of the
same experimental unit (temporal autocorrelation)
and correlation among units in close proximity (spatial autocorrelation) are likely occurrences. The
possibility of autocorrelation can sometimes be
avoided by an appropriate choice of design (e.g., by
ensuring that the sampling units are far apart) or it
can be taken into account by adopting a suitable

spatial or time-series probability model (refer to

Nemec 1996)
Formal statistical inference based on ANOVA
methods (e.g., F-tests, computation of confidence intervals) requires knowledge of the distributional
properties of the random effects and experimental errors. The usual assumption is that all random effects
or experimental errors associated with a specific
source have the same distribution, which is generally
assumed to be normal with a mean of zero and a homogeneous variance. The impact of departures from
these assumptions varies depending on the type of
departure and on the sample size. Minor deviations
from normality have little impact on inferences
about fixed effects, especially when the sample sizes
are large. Tests for random effects tend to be more
seriously affected. Heterogeneous error variances can
distort P-values, although the degree of distortion
depends on the range of variances. Serious departures from normality or homogeneity can sometimes
be identified and corrected; some suggestions are
outlined by Snedecor and Cochran (1967, Chap. 11).
Alternatively, robust or nonparametric methods can
be used. Failure to adjust for lack of independence
(autocorrelation) tends to exaggerate effects when
the correlation is positive, or understate effects when
the correlation is negative.

Classical experimental design also provides a

framework for ANOVA, analysis of covariance,
regression methods, and other types of statistical
analysis. This framework allows investigators to estimate components of variation, and, beyond that, to
make formal statistical inferences about their significance. Hypothesis testing has been emphasized in the
past but confidence intervals, which are often more
informative because they provide an estimate of the
size of an effect, are just as easily constructed. The
usefulness of ANOVAboth the underlying models
and the components-of-variation approachcannot
be disputed. The models are very versatile. Simple
forms of temporal or spatial autocorrelation can, for
instance, be accommodated by repeated-measures or
split-plot models. However, the usual ANOVA assumptions (e.g., normality) are untenable in some
situations (e.g., if the data are discrete). This is a limitation of the probability model, not of the classical
approach to design. Although ANOVA models and
experimental design are closely linked, the two
should not be equated. Other models (e.g., log-linear
models for categorial data, nonparametric models)
and methods of analysis (e.g., regression analysis)
might be applicable when ANOVA models are not.
In addition, new models and methods continue to be
developed, many of them for classical designs.

2.6 Advantages of Classical Experimental Design

2.7 Experimental Design for Adaptive Management

The controlled experiment is in common wisdom

the most decisive of tests (Susser 1994, p. 831). By
controlling both the levels and combinations of factors applied to the experimental material, the
investigator can make much stronger inferences
about causal relationships and interactions than
would otherwise be possible. Replication, randomization, and blocking serve three important purposes:
elimination of systematic error (i.e., bias), quantification of uncertainty, and reduction of uncontrolled
experimental error. Because these three goals are relevant to any scientific investigation, many parallels
can be found between experimental design and the
design of a survey or an observational study (see
Schwarz, this volume, Chap. 3). Two examples mentioned previously are randomization and random
sampling to eliminate bias, and blocking and stratification to improve precision; refer to Feinberg and
Tanur (1987) and Smith and Sugden (1988) for
a discussion of these and other similarities and
differences.

An experiment is conducted to answer one or more

questions. In a research setting, the questions tend to
be relatively simple. Does a new method of storage
promote better short-term survival of seedlings than
an established method? Is there any difference in the
growth of two species of seedlings, 5 years after
planting? Are planted trees more susceptible to root
disease than naturally regenerated trees? The questions confronting managers are likely to be
considerably more challenging, involving such complex issues as maximization of sustainable yield,
avoidance of unnecessary risk, economic efficiency,
and economic stability (see Walters 1986, Chap. 2).
To limit these problems, thereby making them more
amenable to solution by experimentation, a manager
must: (1) identify key factors for analysis: what are
the main factors that distinguish strategies? which
factors can be controlled? (2) consider timing: over
what time period and how frequently should the system be monitored? (3) consider spatial scale: what are
the management units? and (4) identify quantities of

15

interest: how are quality of results, costs, and benefits

is considerably more complicated. Responses of indimeasured? For practical advice on bounding
vidual systems and the overall response of systems
problems for analysis, refer to Chapter 1 of
managed under the same plan (i.e., replicates) must
Walters (1986).
be considered (see Walters 1986, Chap. 10). This
After the objectives have been determined, the
problem has no simple solution because the link benext step is to define the target population and to
tween classical methods (e.g., ANOVA) for the analytake an inventory of the management units available
sis of replicated designs and decision analysis for
for experimentation. If the target population is a
nonreplicated management strategies is not well deunique ecosystem managed as a single unit (e.g., a
veloped. Meta-analysis (see Mann 1990 for an interparticular area of old-growth forest) then replication
esting and nontechnical discussion of meta-analysis)
is inapplicable because there can be only one reor alternative methods for integrating the results
sponse to a particular management strategy. On the
from several experiments might be useful in such sitother hand, when groups of units are sufficiently
uations, although a piece-meal analysis of large,
similar that they can be managed according to a
complex, and dynamic systems has obvious drawcommon strategy (e.g., stands with comparable ages
backs.
and species compositions), replication is desirable to
determine the range of possible responses to
that strategy. However, even when replication is theoretically possible, it might not be
Objectives
practical because of the high costs of largescale experiments, time constraints, or
limited resources.
Target
Replication is a decisive issue. Its
population
effect on the design and analysis of adaptive
management experiments is illustrated in
Figure 2.2. If replication of a treatment or
Experimental
management strategy is possible then classipopulation
cal methods (Sections 2.4 and 2.5) can be
useful. Moreover, even if treatment replication is impossible or impractical, adherence
Yes
No
Replication
to traditional principles can help ensure that
possible?
the sampling design is sound. Thus, randomization, replication, and blocking
Experimental units:
(random sampling and stratification) are efExperimental
randomization
fective means of avoiding bias and reducing
unit (1)
replication
error in both replicated and non-replicated
experiments.
Adaptive management requires a suitable
Sampling units:
Sampling units:
model for predicting transitions of a system
random sampling
random sampling
from one state to another, and a set of rules
for deciding the best action at any given
time. In the case of non-replicated experiClassical statistical
Bayesian statistics
ments (right side of Figure 2.2), various anaanalysis
Decision analysis
(e.g. ANOVA)
lytical methods have been developed (see
Walters 1986, Chap. 49 ). These methods
are based on the theory of stochastic
ACTIONS
processes, Bayesian statistics (Bergerud and
Treatment/
Reed, this volume, Chap. 7), and decision
Management Strategies
theory (Peterman and Peters, this volume,
Chap. 8). When data arise from replicated
. Design and analysis of an adaptive management experiment.
systems (left side of Figure 2.2), the problem

16

2.8 Summary
Sound experimental design is essential for adaptive
management of valuable forest resources. Adherence
to the principles of randomization, replication, and
blocking helps to ensure that an experiment meets
the basic requirements for success (Cox 1958): absence of systematic error, precision, validity for an
appropriate range of conditions, simplicity, and an
estimate of uncertainty. Failure to consider these issues leads to unnecessary waste and, in the worst
case, bad decisions resulting in serious damage to
sensitive forest ecosystems. Development and adoption of adaptive methods for carrying out large-scale
experiments has been slow, due partly to the barriers
created by excessively technical language, a lack of
analytical tools, and a limited number of successful
applications to serve as models. Overcoming these
obstacles by continuing efforts to educate and to inform (e.g., Biometrics Information series published
by Research Branch, B.C. Ministry of Forests) can
only help to improve matters in the future.

______. 1989b. What is the design? B.C. Min. For.,

Res. Br., Victoria, B.C. Biometrics Inf. Pamph.
No. 17.
______. 1991. When are blocks pseudoreplicates?
B.C. Min. For., Res. Br., Victoria, B.C. Biometrics Inf. Pamph. No. 34.
______. 1992. A general description of hypothesis
testing and power analysis. B.C. Min. For., Res.
Br., Victoria, B.C. Biometrics Inf. Pamph. No.
37.
Bergerud, W.A. and W.I. Reed. [n.d.] Bayesian statistical methods. This volume.
Box, G.E.P., W.G. Hunter, and J.S. Hunter. 1978. Statistics for experimenters: an introduction to
design, data analysis, and model building. J.
Wiley, New York, N.Y.
Cochran, W.G. and G.M. Cox. 1957. Experimental
designs. 2nd ed. J. Wiley, New York, N.Y.

References

Cox, D.R. 1958. Planning of Experiments. J. Wiley,

New York, N.Y.

Anderson, J.L. [n.d.]. Errors of inference. This

volume.

Deming, W.E. 1953. On the distinction between enumerative and analytic surveys. J. Am. Statist.
Assoc. 48:24455.

Anderson, V.L. and R.A. McLean. 1974. Design of experiments: a realistic approach. Marcel Dekker,
New York, N.Y.

______. 1960. Sample design in business research.

J. Wiley, New York, N.Y.

Atkinson, A.C. 1982. Developments in the design of

experiments. Intern. Statist. Rev. 50:16177.
. 1988. Recent developments in the methods
of optimum and related experimental designs.
Intern. Statist. Rev. 56:99115.
. 1996. Usefulness of optimum experimental
designs. J. Royal Statist. Soc., B, 58:5976.
Bates, R.A., R.J. Buck, E. Riccomgno, and H.P.
Wynn. 1996. Experimental design and observation for large systems. J. Royal Statist. Soc.,
B, 58:7794.
Bergerud, W. 1988. Understanding replication and
pseudoreplication. B.C. Min. For., Res. Br.,
Victoria, B.C. Biometrics Inf. Pamph. No. 5.
______. 1989a. ANOVA: Factorial designs with a separate control. B.C. Min. For., Res. Br., Victoria,
B.C. Biometrics Inf. Pamph. No. 14.

______. 1975. On probability as a basis for action.

Am. Statist. 29:14652.
Eberhart, L.L. and J.M. Thomas. 1991. Designing environmental field studies. Ecol. Monogr.
61:5373.
Federer, W.T. 1984. Principles of statistical design
with special reference to experiment and treatment design. In Statistics: An Appraisal. Proc.
50th Anniv. Conf. Iowa State Statist. Lab. H.A.
David and H.T. David (editors), Iowa State
Press, pp. 77105.
Feinberg, S.E and J.M. Tanur. 1987. Experimental and
sampling structures: Parallels diverging and
meeting. Intern. Statist. Rev. 55:7596.
Fisher, R.A. 1935. The design of experiments. Hafner,
New York, N.Y. Reprint 1971.

17

Gregoire, T.G., H.T. Valentine, and G.M. Furnival.

1995. Sampling methods to estimate foliage and
other characteristics of individual trees. Ecology
76:118194.
Hahn, G.J. and W.Q. Meeker. 1993. Assumptions for
statistical inference. Am. Statist. 47:111.
Holland, P.W. 1986. Statistics and causal inference.
J. Am. Statist. Assoc. 81:94560.
Hurlbert, S.H. 1984. Pseudoreplication and the design
of ecological field experiments. Ecol. Monogr.
54:187211.

Peterman, R.M. and C. Peters. [n.d.] Decision analysis: taking uncertainties into account in forest
resource management. This volume.
Pike, D.J. 1984. Discussion (of paper by Steinberg and
Hunter 1984). Technometrics 26:1059.
Schwarz, C.J. [n.d]. Studies of uncontrolled events.
This volume.
Sit, V. 1993. What do we look for in a working plan?
B.C. Min. For., Res. Br., Victoria, B.C. Biometrics Inf. Pamph. No. 44.

John, P.W.M. 1971. Statistical design and analysis of

experiments. McMillan, New York, N.Y.

______. 1995. Analyzing ANOVA designs. B.C. Min.

For., Res. Br., Victoria, B.C. Work. Pap.
07/1995. Biometrics Inf. Handb. No. 5.

Kempthorne, O. 1952. The design and analysis of experiments. Krieger, Malabar, Fla.

Smith, G.J. [n.d.] Retrospective studies. This

volume.

Mandel, J. 1964. The statistical analysis of experimental data, Dover, Mineola, N.Y.

Smith, T.M.F. and R.A. Sugden. 1988. Sampling and

assignment mechanisms in experiments, surveys, and observational studies. Int. Statist. Rev.
58:16580.

Mann, C. 1990. Meta-analysis in the breech. Science

249:47680.
Mead, R. 1988. The design of experiments, statistical
principles for practical applications. Cambridge
Univ. Press, Cambridge, U.K.
______. 1990. The non-orthogonal design of experiments. J. Royal Statist. Soc., A, 153:151201.
Milliken, G. and D.E. Johnson. 1989. Analysis of
messy data. Vol. 2: nonreplicated experiments.
Van Nostrand Reinhold, New York, N.Y.

Snedecor, G.W. and W.G. Cochran. 1967. Statistical

methods. 6th ed. Iowa State Univ. Press, Ames,
Iowa.
Stafford, S.G. 1985. A statistics primer for foresters.
J. For. 83:14857.
Steel, R.G. and J.H. Torrie. 1980. Principles and procedures of statistics: a biometrical approach.
2nd ed. McGraw-Hill, New York, N.Y.

______. 1992. Analysis of messy data. Vol. 1: designed

experiments. Wadsworth, Belmont, Calif.

Steinberg, D.M. and W.G. Hunter. 1984. Experimental design: review and comment.
Technometrics 26:71130.

Montgomery, D.C. 1991. Design and analysis of experiments. 3rd ed. J. Wiley, New York, N.Y.

Susser, M. 1994. The logic in ecologic: II. the logic of

design. Am. J. Public Health 84:8305.

Nemec, A.F.L. 1991. Power analysis handbook for the

design and analysis of forestry trials. B.C. Min.
For., Res. Br., Victoria, B.C. Biometrics Inf.
Handb. No. 2.

Taylor, B., L. Kremsater, and R. Ellis. 1997. Adaptive

management of forests in British Columbia.
B.C. Min. For., For. Practices Br., Victoria, B.C.

______. 1996. Analysis of repeated measures and time

series: an introduction with forestry examples.
B.C. Min. For., Res. Br., Victoria, B.C. Work.
Pap. 15/1996. Biometrics Inf. Handb. No. 6.
Penner, M. 1989. Optimal design with variable cost
and precision requirements. Can. J. For.
19:15917.

18

Walters, C. 1986. Adaptive management of renewable

resources. MacMillan, New York, N.Y.
Walters, C.J. and C.S. Holling. 1990. Large-scale experiments and learning by doing. Ecology
71:20608.

3 STUDIES OF UNCONTROLLED EVENTS

CARL J. SCHWARZ

Abstract

The rationale for carefully planned experiments is well

documented. A well-designed experiment will have a
high probability of detecting important, biologically
meaningful differences among the experimental
groups. Causal relationships can be inferred because
the experimental factors have been manipulated and
randomly assigned.
In many cases, controlled experiments are impractical or too expensive, and surveys of existing ecological
populations are performed, even though the resulting
inferences will be weaker than those obtained through
controlled experimentation. Consequently, nonexperimental studies, or passive adaptive management, is
primarily a tool for generating hypotheses to be tested
by careful experimentation.
Despite the weaker inferences from nonexperimental studies, the same attention must be paid to the
proper design of a survey so that the conclusions are
not tainted by inadvertent biases. This paper will review several of the standard nonexperimental studies
by presenting an overview of the study protocol, the
conclusions that can be reached, and the potential
problems that can occur.

Consequently, nonexperimental surveys, or passive

adaptive management, leads to conclusions that are
primarily a tool for generating hypotheses eventually
to be tested by careful and more efficient experimentation.
For example, observation surveys of existing lakes
showed that the more acidic lakes tended to have
fewer fish. An alternative explanation that could explain this result states that some unknown factor
causes the lake to acidify and also kills fish (i.e., the
relationship between numbers of fish and acidification is due to a common response to another factor).
However, experiments where lakes were deliberately
acidified refute this alternate explanation. No such
refutation is possible from surveys of existing populations. The primary message is that causation cannot
be inferred without active manipulation.
Despite the weaker inferences from nonexperimental surveys, the same attention must be paid to
the proper design of a survey so that inadvertent biases do not taint the conclusions. The many different
types of nonexperimental surveys are outlined in
Figure 3.1. To begin, consider the following series of
examples to illustrate the differences among these
types of surveys.

3.1 Introduction
The rationale for carefully planned experiments in
ecology is well documented (Hurlbert 1984). A welldesigned experiment will have a high probability of
detecting important, biologically meaningful differences among the experimental groups. Furthermore,
because the manager directly manipulated the experimental factors and randomly assigned the
experimental units to the particular combination of
experimental factors, the manager can infer a causal
relationship between the experimental factors and
the response variable. The manager who takes a similar approach and practices active adaptive
management can make the strongest possible inferences about the role of the experimental factors.
In many cases, controlled experiments are impractical or too expensive, and surveys of existing
ecological populations are performed, even though
the resulting inferences will be weaker than those
obtainable through controlled experimentation.

Experimental
studies

Designed
experiments

Descriptive
surveys

Observational
surveys
Non-experimental
studies
Analytical
surveys

Impact
surveys
.

A classification of the methods considered in

this chapter.

19

3.1.1 Example A: descriptive survey

A manager is interested in examining the natural regeneration in a cutblock harvested by clearcutting.
The objective is to measure the amount of regeneration. A suitable response measure will be the density
of newly grown trees. A series of sample plots is systematically located within a single cutblock and the
density is measured on each sample plot. The mean
density over all plots is computed along with a measure of precision, the standard error. The study has
only one response variable, the density on each plot,
and no explanatory variables. This is a descriptive survey as no comparisons will be made with other
cutblocks and the information pertains only to that
particular cutblock. No inferences about the density
in other cutblocks is possible.
3.1.2 Example B: observational survey
This same manager now notices that north-facing
slopes seem to have a lower insect infestation rates
than south-facing slopes. One cutblock from a northfacing slope and one cutblock from a south-facing
slope are selected. Sample plots are located on each
cutblock, and the insect infestation is measured on
each sample plot. The response variable is the
amount of infestation in each plot. The orientation of
the slope is an explanatory variable. Estimates of the
mean infestation are obtained for each block. The
sample means for each block likely differ, but with
information on the variation within each block, it is
possible to determine if the population means also
differ (i.e., to determine if the true average infestation
in the two blocks differs). This is an observational survey as two convenient cutblocks were selected and
compared. However, the results are only applicable
to the two cutblocks sampled and can neither be extrapolated to other cutblocks, nor to the effects of
north- and south-facing slopes. The reason for this
weak inference is that the observed differences between the cutblocks may be due to just natural
variation unrelated to the direction of the slope; no
information has been collected on the variability
among cutblocks with the same orientation.
3.1.3 Example C: analytical survey
The manager expands the above survey. Within the
forest management unit, blocks are randomly chosen
in pairs so that, within each pair, one cutblock is on a
north-facing slope and the other is on a south-facing
slope. Sample plots are randomly located on each
cutblock, and the insect infestation is measured on

20

each sample plot. The response variable is the

amount of infestation in each plot. The orientation is
an explanatory variable. Estimates of the mean infestation are obtained for each type of slope along with
the measures of precision. The manager then compares the two means using information on both the
within-cutblock variability and the variability among
blocks with the same orientation. It may appear that
plots on the south-facing slope have a higher infestation than plots on a north-facing slope. This is an
analytical survey, as a comparison was made over an
entire population of cutblocks in the forest management unit. This survey differs from a controlled
experiment in that the orientation of the cutblocks
cannot be controlled by the manager. An alternative
explanation for this observed result is that some
other unknown factor caused the insect infestations
to be different on the two orientations.
3.1.4 Example D: designed experiment
The manager is interested in testing the effect of two
different types of fertilizer on regeneration growth.
Experimental plots in several homogeneous cutblocks are established. Within each cutblock, plots
are randomly assigned to one of the fertilizers. The
regeneration growth of the plots treated with the two
fertilizers is then compared. The response variable is
the amount of growth; the explanatory variable is the
fertilizer type. Because plots were randomly assigned
to the fertilizers, the effects of any other, uncontrollable, lurking factor should, on average, be about
equal in the two treatment groups. Consequently,
any difference in the mean regeneration growth can
be attributed to the fertilizer. The primary differences
between this example and Example C are that the
manager controls the explanatory factor and can randomly assign experimental units to treatments. These
two differences in the protocol allow stronger inferences than in analytical surveys.
3.1.5 Example E: impact survey
The manager wishes to examine if clearcutting is
changing the water quality on nearby streams. A control site with similar soil and topography as the
experimental site, is selected, in a provincial park.
Water quality readings are taken from both streams
several times before harvesting, and several times
after harvesting. The response variable is the water
quality; the explanatory variable is the presence or
absence of nearby clearcutting.
The changes in water quality in the control and

experimental sites are compared. If the objective is

to examine if there is a difference in water quality
between these two specific sites, then the survey will
answer the question. This is similar to the strength of
inference for observational surveys (Example B). If
the objective is to extrapolate from this pair of sites
to the effects of clearcutting in general, the inference
is much more limited. First, because the control or
impact sites are not replicated it is impossible to
know if the observed differences are within the range
of natural variation. This limitation could be partly
resolved by adding multiple control sites and
assuming that the variability among control sites is
representative of that among impact sites. However,
the lack of randomization of the impact will still limit
the extent to which the results can be generalized.
But in the longer term, if there are several such pairs
of sites and all show the same type of impact, solid
grounds are established for assigning a causal relationship, even though randomization never took
place. This would be based on the idea of a superpopulation consisting of all possible pairs of sites; it is
not likely that unobservable, latent factors would be
operating in the same direction in all experiments.
This last form is the closest to a designed experiment
for an impact survey.

These five examples differ in two important dimensions:

1. The amount of control over the explanatory factor.
Descriptive surveys have the least amount of control, while designed experiments have maximal
control.
2. The degree of extrapolation to other settings.
Again, in descriptive surveys, inference is limited
to those surveyed populations, while in designed
experiments on randomly selected experimental
units, inference can be made about future effects
of the explanatory factors.
In general, the more control or manipulation present, the stronger the inferences that can be made
(Figure 3.2).
This chapter will present an overview of some of
the issues that arise in surveys lacking experimental
manipulations. It will start with an overview of the
descriptive surveys used to obtain basic information
about a population. Observational surveys will not be
explicitly addressed as their usefulness is limited and
their design and analysis are very close to analytical
surveys. Next, analytical surveys, where the goal is to
compare subsets of an existing population, will be

Degree of control

Designed
experiments
Control-Impact
surveys
Impact
surveys
Analytical
surveys
Observational
surveys
Descriptive
surveys
Strength of inference
.

Relationship between degree of control, strength of inference, and type of study design.

21

described. Impact Surveys, where one site affected by

some planned or unplanned event and a control site
where no such event occurs are compared, will then
be discussed. Finally, some general principles of nonexperimental studies will be reviewed.
3.2 Descriptive Surveys
The goal of a descriptive survey is to estimate a parameter of interest (e.g., an average, total, proportion,
or ratio) for a single population (e.g., Example A of
Section 3.1). No attempt is made to compare the parameters between two or more populations.
Many excellent references on descriptive survey
methods are available (Cochran 1977; Krebs, 1989;
Thompson 1992). Therefore, this section is limited to
a brief account of the main survey methods that
could be used in field research. Details on actual field
procedures are also available; for example, Myers and
Shelton (1980).
3.2.1 Survey methods
Simple random sampling
Simple random sampling is the basic method of selecting survey units. Each unit in the population is
selected with equal probability and all possible samples are equally likely to be chosen. This selection is
commonly done by listing all the members in the
population and then sequentially choosing units
using a random number table. Units are usually chosen without replacement (i.e., each unit in the
population can only be chosen once). In some cases,
particularly for multistage designs, there are advantages to selecting units with replacement (i.e., a unit
in the population may potentially be selected more
than once).
The analysis of a simple random sample is
straightforward. The mean of the sample is an estimate of the population mean. An estimate of the
population total is obtained by multiplying the sample mean by the number of units in the population.
The sampling fraction, the proportion of units chosen from the entire population, is typically small. If it
exceeds 20%, an adjustment (the finite population
correction) will result in better estimates of precision
(a reduction in the standard error) to account for the
fact that a substantial fraction of the population was
surveyed.
An example of a simple random sample would be a
vegetation survey in a large forest stand. The stand is

22

divided into 300 1-hectare plots, and a random sample of 20 plots was selected and analyzed using aerial
photos.
Pitfall: A simple random sample design is often hid-

den in the details of many other survey designs. For

example, many surveys of vegetation are conducted
using strip transects where the initial starting point of
the transect is randomly chosen, and then every plot
along the transect is measured. Here the strips are the
sampling unit, and are a simple random sample from
all possible strips. The individual plots are subsamples from each strip and cannot be regarded as
independent samples. For example, suppose a rectangular stand is surveyed using aerial overflights. In
many cases, random starting points along one edge
are selected, and the aircraft then surveys the entire
length of the stand starting at the chosen point. The
strips are typically analyzed section-by-section, but it
would be incorrect to treat the smaller parts as a simple random sample from the entire stand.
Solution: Note that a crucial element of simple random samples is that every sampling unit is chosen
independently of every other sampling unit. For example, in strip transects, plots along the same
transect are not chosen independently: when a particular transect is chosen, all plots along the transect
are sampled and so the selected plots are not a simple
random sample of all possible plots. Strip-transects
are actually examples of cluster-samples.
Systematic sampling
In some cases, it is logistically inconvenient to randomly select sample units from the population. An
alternative is to take a systematic sample where every
kth unit is selected (after a random starting point); k
is chosen to give the required sample size. For example, if a stream is 2 km long, and 20 samples are
required, then k=100 and samples are chosen every
100 m along the stream after a random starting point.
A common alternative when the population does not
naturally divide into discrete units is grid-sampling.
Here sampling points are located using a grid that is
randomly located in the area. All sampling points are
a fixed distance apart.
If a known trend is present in the sample, it can be
incorporated into the analysis (Cochran 1977, Chap. 8).
For example, suppose that the systematic sample follows an elevation gradient that is known to directly
influence the response variable. A regression-type
correction can be incorporated into the analysis.

However, note that this trend must be known from

external sourcesit cannot be deduced from the
survey.
Pitfall: A systematic sample is typically analyzed in

the same fashion as a simple random sample. However, the true precision of an estimator from a
systematic sample can be either worse or better than
a simple random sample of the same size, depending
if units within the systematic sample are positively or
negatively correlated among themselves. For example, if a systematic samples sampling interval
happens to match a cyclic pattern in the population,
values within the systematic sample are highly positively correlated (the sampled units may all hit the
peaks of the cyclic trend), and the true sampling
precision is worse than a simple random sample of
the same size. What is even more unfortunate is that,
because the units are positively correlated within the
sample, the sample variance will underestimate the
true variation in the population, and, if the estimated
precision is computed using the formula for a simple
random sample, a double dose of bias in the estimated precision occurs (Krebs 1989, p. 227). On the other
hand, if the systematic sample is arranged perpendicular to a known trend to try to incorporate
additional variability in the sample, the units within a
sample are now negatively correlated, the true precision is now better than an simple random sample of
the same size, but the sample variance now overestimates the population variance, and the formula for
precision from a simple random sample will overstate the sampling error.
While logistically simpler, a systematic sample is
only equivalent to a simple random sample of the
same size if the population units are in random
order to begin with (Krebs 1989, p. 227). Even worse,
no information in the systematic sample allows the
manager to check for hidden trends and cycles.
Nevertheless, systematic samples offer the following practical advantages over simple random
sampling if the bias in the estimated precision can be
corrected:
make plot relocation for long-term monitoring
easier;
allow mapping to be carried out concurrently with
the sampling effort because the ground is systematically traversed;
avoid poorly distributed sampling units, which can
occur with a simple random sample (this problem
can also be avoided by judicious stratification).

Solution: Because a strong assumption of random-

ness in the original population is necessary,

systematic samples are discouraged and statistical advice should be sought before starting such a scheme.
If no other designs are feasible, a slight variation in
the systematic sample provides some protection from
the previous problems. Instead of taking a single systematic sample every kth unit, take two or three
independent systematic samples of every 2kth or 3kth
unit, each with a different starting point. For example, rather than taking a single systematic sample
every 100 m along the stream, two independent systematic samples can be taken, each selecting units
every 200 m along the stream starting at two random
starting points. The total sample effort is still the
same, but now some measure of the large-scale spatial structure can be estimated. This technique is
known as replicated subsampling (Kish 1965, p. 127).
Cluster sampling
In some cases, units in a population occur naturally
in groups or clusters. For example, some animals
congregate in herds or family units. It is often convenient to select a random sample of herds and then
measure every animal in the herd. This is not the
same as a simple random sample of animals because
individual animals are not randomly selected; the
herds were the sampling unit. The strip-transect example in Section 3.2.1 is also a cluster sample; all plots
along a randomly selected transect are measured. The
strips are the sampling units, while plots within each
strip are subsampling units. Another example is circular plot sampling; all trees within a specified radius
of a randomly selected point are measured. The sampling unit is the circular plot while trees within the
plot are subsamples.
The reason cluster samples are used is that costs
can be reduced compared to a simple random sample
giving the same precision. Because units within a
cluster are close together, travel costs among units
are reduced. Consequently, more clusters (and more
total units) can be surveyed for the same cost as a
comparable simple random sample.
Pitfall: A cluster sample is often mistakenly analyzed

using methods for simple random surveys. Such

analysis is not valid because units within a cluster are
typically positively correlated. This erroneous analysis produces an estimate that appears to be more
precise than it really is (i.e., the estimated standard

23

error is too small and does not fully reflect the actual
imprecision in the estimate).
Solution: To be confident that the reported standard
error really reflects the uncertainty of the estimate,
the analytical methods must be appropriate for the
survey design. The proper analysis treats the clusters
as a random sample from the population of clusters.
The methods of simple random samples are applied
to the cluster summary statistics (Thompson 1992,
Chap. 12; Nemec 1993).
Multi-stage sampling
In many situations the population is naturally divided into several different sizes of units. For example,
a forest management unit consists of several stands,
each stand has several cutblocks, and each cutblock
can be divided into plots. These natural divisions can
be easily accommodated in a survey through the use
of multistage methods. Units are selected in stages.
For example, several stands could be selected from a
management area; then several cutblocks are selected
in each of the chosen stands; then several plots are selected in each of the chosen cutblocks. Note that in a
multistage design, units at any stage are selected at
random only from those larger units selected in previous stages.
The advantage of multistage designs are that costs
can be reduced compared to a simple random sample
of the same size, primarily through improved logistics. The precision of the results is less than an
equivalent simple random sample, but because costs
are less, a larger multistage survey can often be completed for the same costs as a smaller simple random
sample. This approach often results in a more precise
design for the same cost. However, due to the misuse
of data from complex designs, simple designs are
often highly preferred and end up being more costefficient when costs associated with incorrect decisions are incorporated.
Pitfall: Although random selections are made at each

stage, a common error is to analyze these types of

surveys as if they arose from a simple random sample. The plots were not independently selected; if a
particular cutblock was not chosen, then none of the
plots within that cutblock can be chosen. As in cluster samples, this erroneous analysis produces
estimated standard errors that are too small and do
not fully reflect the actual imprecision in the estimates. A manager will be more confident in the
estimate than is justified by the survey.

24

Solution: Again, it is important that the analytical

methods are suitable for the sampling design. The

proper analysis of multistage designs considers that
random samples takes place at each stage (Thompson
1992, Chap. 13). In many cases, the precision of the
estimates is determined essentially by the number of
first-stage units selected. Little is gained by extensive
sampling at lower stages.
Multiphase designs
In some surveys, multiple surveys of the same survey
units are performed. In the first phase, a sample of
units is selected (usually by a simple random sample). Every unit is measured on some variable. Then
in subsequent phases, samples are selected ONLY
from those units selected in the first phase, not from
the entire population.
Multiphase designs are commonly used in two situations. In the first case, stratifying a population in
advance is sometimes difficult because the values of
the stratification variables are not known. The first
phase is used to measure the stratification variable on
a random sample of units. The selected units are then
stratified, and each stratum is further sampled as
needed to measure a second variable. This approach
avoids having to measure the second variable on
every unit when the strata differ in importance. For
example, in the first phase, plots are selected and
measured for the amount of insect damage. The plots
are then stratified by the amount of damage, and second-phase allocation of units concentrates on plots
with low insect damage to measure total usable volume of wood. It would be wasteful to measure the
volume of wood on plots with heavy insect damage.
The second common occurrence for using a multistage design is a surrogate variable (related to the real
variable of interest) on selected units is relatively easy
to measure, and then, in the second phase, the real
variable of interest is measured on a subset of the
units. The relationship between the surrogate and desired variable in the smaller sample is used to adjust
the estimate based on the surrogate variable in the
larger sample. For example, managers need to estimate the volume of wood removed from a harvesting
area. A large sample of logging trucks is weighed
(which is easy to do), and weight will serve as a surrogate variable for volume. A smaller sample of
trucks (selected from those weighed) is scaled for volume and the relationship between volume and
weight from the second-phase sample is used to predict volume based on weight only for the first-phase

sample. Another example is the count plot method of

estimating volume of timber in a stand. A selection of
plots is chosen and the basal area determined. Then a
sub-selection of plots is rechosen in the second
phase, and volumes are measured on the secondphase plots. The relationship between volume and
area in the second phase is used to predict volume
from area measurements seen in the first phase.
Repeated sampling
One common objective of long-term surveys is to
investigate changes over time of a particular population. This investigation, which involves repeated sampling from the population, has three common designs.
First, separate independent surveys can be conducted at each time point. This is the simplest design
to analyze because all observations are independent
over time. For example, independent surveys can be
conducted at 5-year intervals to assess regeneration of
cutblocks. However, precision of the estimated
change may be poor because of the additional variability introduced by having new units sampled at
each time point.
At the other extreme, units are selected in the first
survey and the same units are remeasured over time.
For example, permanent plots that are remeasured
for regeneration over time can be established. The
advantage of permanent plots is that comparisons
over time are free of additional variability introduced
by new units being measured at every time point.
One possible problem is that survey units have become damaged over time, and the sample size will
tend to decline over time. An analysis of these types
of designs is more complex because of the need to account for (1) the correlation over time of
measurements on the same sample plot and (2) possible missing values when units become damaged
and are dropped from the survey.
Intermediate to the previous two designs are partial replacement designs where a portion of the survey
units is replaced with new units at each time point.
For example, 20% of the units could be replaced by
new units at each time point; units would normally
stay in the survey for a maximum of five time periods. These types of designs require complex analysis.
Pitfall: The most common error made in analyzing

repeated sampling designs is to treat observations as

being independent. This typically leads to estimated
precisions that appear too precise (i.e., the real precision is much poorer).

Solution: The analysis of repeated samples is quite

complexit is important to consult with an expert

in this field.
Designs for wildlife sampling
Two common survey designs for measuring wildlife
abundance are capture-recapture surveys and distance surveys.
In capture-recapture surveys (Otis et al. 1978;
Pollock et al. 1990), animals are captured, tagged, and
released on each of a number of time points. The pattern of recaptures of the observed animals is used to
estimate survival rates and abundance. Skalski and
Robson (1992) discuss the design of surveys using
capture-recapture methods.
In distance surveys (Buckland et al. 1993), an observer follows a transect and notes the angle and distance of animals from the transect line. A detection
function is constructed that relates the probability of
spotting an animal as a function of the distance from
the transect line and this is used to estimate abundance.
This section is deliberately brief as many complex
planning problems are associated with using these
methods and expert assistance is strongly recommended.
3.2.2 Refinements that affect precision
Sampling with unequal probability
All of the designs discussed in previous sections have
assumed that each sample unit was selected with
equal probability. In some cases, it is advantageous to
select units with unequal probabilities, particularly if
they differ in their contribution to the overall total.
This technique can be used with any of the sampling
designs discussed earlier. An unequal probability
sampling design can lead to smaller standard errors
(i.e., greater precision) for the same total effort compared to an equal probability design. For example,
forest stands may be selected with probability proportional to the area of the stand (i.e., a stand of
200 ha will be selected with twice the probability than
a stand of 100 ha) because large stands contribute
more to the overall population and it would be
wasteful to spend much sampling effort on smaller
stands.
The variable used to assign the probabilities of selection to individual survey units does not need to
have an exact relationship with an individual contribution to the total. For example, in probability
proportional to prediction (3P) sampling, all trees in

25

a small area are visited. A simple, cheap characteristic, which is used to predict the value of the tree, is
measured. A subsample of the trees is then selected
with probability proportional to the predicted value,
remeasured using a more expensive measuring device. The relationship between the cheap and
expensive measurement in the second phase is used
with the simple measurement from the first phase to
obtain a more precise estimate for the entire area.
This example illustrates two-phase sampling with unequal probability of selection.
Stratification
All survey methods can potentially benefit from stratification (also known as blocking in the experimentaldesign literature). Stratification groups survey units
into homogeneous groups before conducting the survey, and then conducts independent surveys in each
stratum. At the end of the survey, the stratum results
are combined and weighted appropriately. For example, a watershed might be stratified by elevation into
three strata, and separate surveys are conducted within each elevation stratum. The separate results would
be weighted proportionally to the size of the elevation strata. Stratification will be beneficial whenever
variability among the sampling units can be anticipated and strata can be formed that are more
homogeneous than the original population.
A major question with stratified surveys is the
allocation of sampling units among the strata. Depending upon the goals of the survey, an optimal
allocation of sampling units can be one that is equal
in all strata, that is proportional to the stratum size,
or that is related to the cost of sampling in each stratum (Thompson 1992, Chap. 11). Equal allocation
(where all strata have the same sample size) is preferred when equally precise estimates are required for
each stratum as well as for the overall population.
Proportional allocation (where the sample size in
each stratum is proportional to the population size)
is preferred when more precise estimates are required
in larger strata. If the costs of sampling vary among
the strata, then an optimal allocation that accounts
for costs would try to obtain the best overall precision at the lowest cost by allocating units among the
strata accounting for the costs of sampling in each
stratum.
Stratification can be carried out prior to the survey
(pre-stratification) or after the survey (post-stratification). Pre-stratification is used if the stratum

26

variable is known in advance for every plot (e.g., elevation of a plot). Post-stratification is used if the
stratum variable can only be ascertained after measuring the plot (e.g., soil quality or soil pH). The
advantages of pre-stratification are that samples can
be allocated to the various strata in advance to optimize the survey and the analysis is relatively
straightforward. With post-stratification, there is no
control over sample size in each of the strata, and the
analysis is more complicated (the samples sizes in
each stratum are now random). Post-stratification
can result in significant gains in precision but does
not allow for finer control of the sample sizes as
found in pre-stratification.
Auxiliary variables
An association between the measured variable of interest and a second variable of interest can be
exploited to obtain more precise estimates. For example, suppose that growth in a sample plot is
related to soil nitrogen content. A simple random
sample of plots is selected and the height of trees in
the sample plot is measured along with the soil nitrogen content in the plot. A regression model is fit
(Thompson 1992, Chap. 7 and 8) between the two
variables to account for some of the variation in tree
height as a function of soil nitrogen content. This approach can be used to make precise predictions of the
mean height in stands if the soil nitrogen content can
be easily measured. This method will be successful if
a direct relationship exists between the two variables.
The stronger the relationship, the more effective this
method will be. This technique is often called ratioestimation or regression-estimation.
Notice that multiphase designs often use an auxiliary variable but this second variable is only
measured on a subset of the sample units.
Unit size
A typical concern with any of the survey methods occurs when the population does not have natural
discrete sampling units. For example, a large section
of land may be arbitrarily divided into 1 m2 plots, or
10 m2 plots. A natural questionis what is the best
size of unit?has no simple answer and depends
upon several factors, which must be addressed for
each survey:
Cost: All else being equal, sampling many small
plots may be more expensive than sampling fewer

larger plots. The primary difference in cost is the

overhead in travel and setup to measure the unit.
Size of unit: An intuitive feeling is that more smaller plots are better than few large plots because the
sample size is larger. This will be true if the characteristic of interest is patchy, but surprisingly,
makes no difference if the characteristic is randomly scattered throughout the area (Krebs 1989,
p. 64). Indeed if the characteristic shows avoidance, then larger plots are better. For example,
competition among trees implies they are spread
out more than expected if they were randomly located. Logistical considerations often influence the
plot size. For example, if trampling the soil affects
the response, then sample plots must be small
enough to measure without trampling the soil.
Edge effects: Because the population does not have
natural boundaries, decisions must often be made
about objects that lie on the edge of the sample
plot. In general, larger square or circular plots are
better because of smaller edge-to-area ratio. (A
long narrow rectangular plot can have more edge
than a similar-area square plot.)
Size of object being measured: Clearly, a 1 m2 plot
is not appropriate when counting mature Douglasfir, but may be appropriate for a lichen survey.
A pilot survey should be carried out prior to a
large-scale survey to investigate factors that influence
the choice of sampling unit size.
Sample size determination
An important question in survey design is the choice
of sample size, which is the primary determinant of
the costs of the survey and of precision. The sample
size should be chosen so that the final estimates have
a precision that is adequate for the management
question. Paradoxically, to determine the proper
sample size, some estimate of the population values
needs to be known before the survey is conducted!
Historical data can sometimes be used. In some cases,
pilot surveys will be needed to obtain preliminary
estimates of the population values to plan the main
survey. (Pilot surveys are also useful to test the protocol. Refer to Section 3.5).
Unfortunately, sometimes even pilot surveys cannot be done because of the difficulty in sampling or
because the phenomenon is a one-time event. If a
study has multiple objectives, reconciling the sample

size requirements for each objective may also be difficult. In these and many other cases, sample sizes are
determined solely by the budget for the survey.
3.3 Analytical Surveys
In descriptive surveys, the objective was to simply
obtain information about one large group. In
observational surveys, two deliberately selected subpopulations are chosen and surveyed, but the results
are not generalized to the whole population. In
analytical surveys, subpopulations are selected and
sampled to generalize the observed differences
among the subpopulation to this and other similar
populations.
As such, analytical and observational surveys and
experimental design are similar. However, the primary difference is that, in experiments, the manager
controls the assignment of the explanatory variables
while measuring the response variables, whereas in
analytical and observational surveys, neither set of
variables is under the control of the manager. (Refer
to Section 3.1, Examples B, C, and D). The analysis of
complex surveys for analytical purposes can be very
difficult (Sedransk 1965a, 1965b, 1966; Rao 1973; Kish
1984, 1987).
The first step in analytical surveys is to identify potential explanatory variables (similar to factors in
experiments). At this point, analytical surveys can be
usually further subdivided into three categories depending on the type of stratification:
the population is pre-stratified by the explanatory
variables and surveys are conducted in each stratum to measure the outcome variables;
the population is surveyed in its entirety, and poststratified by the explanatory variables; and
the explanatory variables can be used as auxiliary
variables in ratio or regression methods.
In very complex surveys, all three types of stratification may take place.
The choice between the categories is usually made
by the ease with which the population can be prestratified and the strength of the relationship between
the response and explanatory variables. For example,
sample plots can be easily pre-stratified by elevation
or by exposure to the sun, but it would be difficult to
pre-stratify by soil pH.
Pre-stratification has the advantage that the manager controls the number of sample points collected
in each stratum. However, the numbers are not

27

controllable in post-stratification and may lead to

very small sample sizes in certain strata just because
the strata form only a small fraction of the population.
For example, a manager may wish to investigate
the difference in regeneration (as measured by the
density of new growth) as a function of elevation.
Several cutblocks will be surveyed. In each cutblock,
the sample plots will be pre-stratified into three elevation classes, and a simple random sample will be
taken in each elevation class. The allocation of effort
in each stratum (i.e., the number of sample plots)
will be equal. The density of new growth will be measured on each selected sample plot. On the other
hand, suppose that the regeneration is a function of
soil pH. This cannot be determined in advance, and
so the manager must take a simple random sample
over the entire stand, measure the density of new
growth and the soil pH at each sampling unit, and
then post-stratify the data based on measured pH.
The number of sampling units in each pH class is not
controllableindeed it may turn out that certain pH
classes have no observations.
If explanatory variables are treated as a auxiliary
variables, then a strong relationship must exist between the response and explanatory variables and the
auxiliary variable must be able to be measured precisely for each unit. Then, methods like multiple
regression can also be used to investigate the relationship between the response and the explanatory
variable. For example, rather than classifying elevation into three broad elevation classes or soil pH into
broad pH classes, the actual elevation or soil pH must
be measured precisely to serve as an auxiliary variable
in a regression of regeneration density versus elevation or soil pH.
If the units have been selected using a simple random sample, then the analysis of the analytical
surveys proceeds along similar lines as the analysis of
designed experiments (Kish 1987; Nemec, this volume, Chap. 2). In most analyses of analytical surveys,
the observed results are postulated to have been
taken from a hypothetical super-population of which
the current conditions are just one realization. In the
above example, cutblocks would be treated as a
random blocking factor, elevation class as an explanatory factor, and sample plots as samples within
each block and elevation class. Hypothesis testing
about the effect of elevation on mean density of regeneration occurs as if this were a planned
experiment.

28

Pitfall: Any one of the sampling methods described

in Section 3.2 for descriptive surveys can be used for
analytical surveys. Many managers incorrectly use the
results from a complex survey as if the data were
collected using a simple random sample. As Kish
(1987) and others have shown, this mistake can lead
to substantial underestimates of the true standard
error (i.e., the precision is thought to be far greater
than is justified based on the survey results). Consequently, the manager may erroneously detect
differences more often than expected (i.e., make a
Type I error) and make decisions based on erroneous
conclusions.
Solution: As in experimental design, it is important
to match the analysis of the data with the survey design used to collect it. The major difficulties in
analyzing analytical surveys are:

1. Recognizing and incorporating the sampling

method used to collect the data in the analysis. The
survey design used to obtain the sampling units
must be taken into account in much the same way
as the analysis of the collected data is influenced by
actual experimental design. Equivalencies between terms in a sample survey and terms in
experimental design are provided in Table 3.1. No
quick and easy method is available for the analysis
of complex surveys (Kish 1987). The super-population approach seems to work well if the selection
probabilities of each unit are known (these are
used to weight each observation appropriately)
and if random effects corresponding to the various
strata or stages are employed. The major difficulty
caused by complex survey designs is that the observations are not independent of each other. This
nonindependence, if properly incorporated into
the analysis, can improve precision. If not accounted for, nonindependence will lead to
seriously biased estimates of precision.
2. Unbalanced designs (e.g., unequal numbers of
sample points in each combination of explanatory
factors). This difficulty typically occurs if poststratification is used to classify units by the explanatory variables but can also occur in pre-stratification if the manager decides not to allocate equal
effort in each stratum. The analysis of unbalanced
data is described by Milliken and Johnson (1984).
3. Missing cells (i.e., certain combinations of explanatory variables may not occur in the survey).
The analysis of such surveys is complex, but refer
to Milliken and Johnson (1984).

4. If the range of the explanatory variable is naturally

limited in the population, then extrapolation
outside of the observed range is not recommended.
More sophisticated techniques can also be used in
analytical surveys. For example, correspondence
analysis, ordination methods, factor analysis, multidimensional scaling, and cluster analysis all search
for associations among measured variables that may
give rise to hypotheses for further investigation. Unfortunately, most of these methods assume that units
have been selected independently of each other using
a simple random sample; extensions where units
have been selected via a complex sampling design
have not yet developed. Simpler designs are often
highly preferred to avoid erroneous conclusions
based on inappropriate analysis of data from complex designs.
Pitfall: While the analysis of analytical surveys and

designed experiments are similar, the strength of the

conclusions is not. In general, causation cannot be
inferred without manipulation. An observed relationship in an analytical survey may be the result of
a common response to a third, unobserved variable.
For example, consider the two following experiments. In the first experiment, the explanatory
variable is elevation (high or low). Ten stands are
randomly selected at each elevation. The amount of
growth is measured and it appears that stands at

higher elevations have less growth. In the second experiment, the explanatory variable is the amount of
fertilizer applied. Ten stands are randomly assigned
to each of two doses of fertilizer. The amount of
growth is measured and it appears that stands that
receive a higher dose of fertilizer have greater growth.
In the first experiment, the manager cannot say
whether the differences in growth are due to differences in elevation or amount of sun exposure or soil
quality as all three may be highly related. In the second experiment, all uncontrolled factors are present
in both groups and their effects will, on average, be
equal. Consequently, the assignment of cause to the
fertilizer dose is justified because it is the only factor
that differs (on average) among the groups.
As noted by Eberhardt and Thomas (1991), rigorous application of the techniques for survey sampling
is needed when conducting analytical surveys, otherwise these surveys are likely to be subject to biases.
Experience and judgement are very important in
evaluating the prospects for bias, and attempting to
find ways to control and account for these biases. The
most common source of bias is the selection of survey units; the most common pitfall is to select units
based on convenience rather than on a probabilistic
sampling design. The potential problems that this
can lead to are analogous to those that occur when it
is assumed that callers to a radio phone-in show are
representative of the entire population.

Equivalencies between terms used in surveys and in experimental design

Survey term

Experimental design term

Simple random sample

Completely randomized design

Cluster sampling

(a) Clusters are random effects; units within a cluster treated as subsamples; or
(b) Clusters treated as main plots; units within a cluster treated as subplots in a
split-plot analysis

Multi-stage sampling

(a) Nested designs with units at each stage nested in units in higher stages. Effects
of units at each stage treated as random effects; or
(b) Split-plot designs with factors operating at higher stages treated as main plot
factors and factors operating at lower stages treated as subplot factors

Stratification

Fixed factor or random block depending on the reasons for stratification

Sampling unit

Experimental unit or treatment unit

Subsample

Subsample

29

3.4 Impact Surveys

Probably the most important and controversial use
of surveys is to investigate the effects of large-scale,
potentially unreplicated events. These impact surveys
investigate the impact of an event or process. In many
cases, this survey must be done without having the
ability or resources to conduct a planned experiment.
Consider three examples: the impact of a hydroelectric dam on water quality of the dammed stream;
the impact of clearcuts on water quality of nearby
streams; and the effect of different riparian zone
widths along streams near clearcuts. First, randomization and replication are not possible in the first
example. Only one dam will be built on one stream.
In the other two examples, it is possible to randomize
and replicate the experiment and so the principles of
experimental design may be useful. Second, the impact of the first two examples can be compared to a
control or non-treated site while in the last example
impacts are compared: the two different riparian
zone widths.
Regardless of the control over randomization and
replication, the goal of impact surveys is typically to
measure ecological characteristics (usually over time)
to look for evidence of a difference (impact) between
the two sites. Presumably, this impact will be attributed to the event, but, as shown later, the lack of
replication and randomization may limit the generalization of the findings. Then, based on the findings,
remediation or changes in future events will be
planned. In all cases, the timing of the event must be
known in advance so that baseline information can
be collected.
A unifying example for this section will be an investigation of the potential effects of clearcuts on
water quality of nearby streams. Several, successively
more complex impact designs will be considered.
3.4.1 Designs
Before-after contrasts at a single site
This is the simplest impact design. A single survey is
taken before and after a potential disturbance. This
design is widely used in response to obvious accidental incidences of potential impact (e.g., oil spills,
forest fires), where, fortuitously, some prior information is available. From this survey, the manager
obtains a single measurement of water quality before
and after the event. If the second survey reveals a
change, this difference is attributed to the event.

30

Pitfalls: The observed event and the changes in the

response variable may not be relatedthe change

may be entirely coincidental. Even worse, no information is collected on the natural variability of the
water quality over time and the observed differences
may simply be due to natural fluctuations over time.
Decisions based on this design are extremely hard to
justify. This design cannot be used if the event cannot
be planned and no prior data are available. In these
cases, little can be said about the impact of the event.
Repeated before-after sampling at a single site
An embellishment on the previous sampling scheme
is to perform multiple surveys of the stream at multiple time points before and after the event. In this
design, information is collected on the mean water
quality before and after the impact. As well, information is collected on the natural variability over time.
This design is better than the previous design in that
observed changes due solely to natural fluctuations
over time can be ruled out and consequently any observed change in the mean level is presumably real.
The choice between regular intervals and random
intervals depends upon the objectives of the survey.
If the objective is to detect changes in trend, regularly
spaced intervals are preferred because the analysis is
easier. On the other hand, if the objective is to assess
differences before and after impact, then samples at
random time points are advantageous, because no
cyclic differences unforeseen by the sampler will influence the size of the difference. For example,
surveys taken every summer for several years before
and after the clearcutting may show little difference
in water quality but potentially significant differences
in the winter may go undetected.
Pitfall: Despite repeated surveys, this design suffers

from the same flaw as the previous design. The repeated surveys are pseudoreplications in time rather
than real replicates (Hurlbert 1984). The observed
change may have occurred regardless of the clearcut
because of long-term trends over time. Again, decisions based on this design are difficult to justify.
BACI: Before-after-control-impact surveys
As Green (1979) pointed out, an optimal impact survey has several features:
the type of impact, time of impact, and place of
occurrence should be known in advance;
the impact should not have occurred yet; and
control areas should be available.

The first feature allows the surveys to be efficiently

planned to account for the probable change in the
environment. The second feature allows a baseline
survey to be established and to be extended as needed. The last feature allows the surveyor to distinguish
between temporal effects unrelated to the impact and
changes related to the impact.
The simplest BACI design will have two times of
sampling (before and after impact) in areas (treatment and a control) with biological and environmental variables being measured in all combinations of
time and area. In this example, two streams would
be sampled. One stream would be adjacent to the
clearcut (the treatment stream); the second stream
would be adjacent to a control site that is not clearcut
and should have similar characteristics to the treatment stream and be exposed to similar climate and
weather. Both streams are sampled at the same time
points before the clearcut occurs and at the same
time point after the clearcut takes place. Technically,
this is known as an area-by-time factorial design, and
evidence of an impact is found by comparing the

before and after samples for the control site with the
before and after samples for the treatment sites. This
contrast is known as the area-by-time interaction
(see Figure 3.3).
This design allows for both natural stream-tostream variation and coincidental time effects. If the
clearcut has no effect, then change in water quality
between the two time points should be the same (i.e.,
parallel lines in Figures 3.3a and b). On the other
hand, if the clearcut has an impact, the time trends
will not be parallel (Figures 3.3c, d, and e).
Pitfalls: Hurlbert (1984), Stewart-Oaten et al. (1986),

and Underwood (1991) discuss the simple BACI design and point out concerns with its application.
First, because impact to the sites was not randomly
assigned, any observed difference between control
and impact sites may be related solely to some other
factor that differs between the two sites. One could
argue that it is unfair to ascribe the effect to the impact. However, as Stewart-Oaten et al. (1986) point
out, the survey is concerned about a particular impact

Water quality

(b)

Water quality

(a)

Before
After
Time

Before
After
Time
.

(e)
Water quality

(d)
Water quality

Water quality

(c)

Before
After
Time

Before
After
Time

Before
After
Time

Simplified outcomes in a BACI design.

The change in a measured variable from two sampling occasions (dots at before and after the impact) in the
control (solid line) or impact (shaded line) sites. In (a) and (b) the lines are parallel and there is no evidence of
an impact. The difference in (b) between control and impact sites reflects area differences, but both sites
experience the same temporal trend. In (c), (d), and (e), the change over time differs between the control and
impact sites. This change is evidence of a time-treatment interaction, or that the impact has had an effect.

31

in a particular place, not in the average of the impact

when replicated in many different locations. Consequently, detecting a difference between these two
specific sites may be possible; however, without randomization of replicate treatments at many different
sites, the findings from this survey cannot be generalized to other events on different streams.
This concern can be reduced by monitoring several control sites (Underwood 1991). However, two
assumptions must be made: (1) the variation in the
(AfterBefore) measurements of the multiple control sites is the same as the variation among potential
impact sites, and (2) the variability over time between
the control sites is not correlated. Then the plausability of the difference observed in the impact site can
be estimated given the observed variability in the
changes in the control sites. In our example, several
control streams could be monitored at the same time

points as the single-impact stream. Then if the observed difference in the impact stream is much
different than could be expected based on the multiple-control streams, the event is said to have caused
an impact. When several control sites are monitored,
the lack of randomization is less of a concern because
the replicated control sites provide some information
about potential effects of other factors.
The second and more serious concern with the
simple BACI design with a single sampling point before and after the impact is that it fails to recognize
that natural fluctuations in the characteristic of interest that are unrelated to any impact may occur
(Hurlbert 1984; Stewart-Oaten et al. 1986). For example, consider Figure 3.4. If there were no natural
fluctuations over time, the single samples before and
after the impact would be sufficient to detect the
effects of the impact. However, if the population also

Water quality

(a)

Before

After
Time
(c)

Water quality

Water quality

(b)

Before

After
Time

32

Before

After
Time

Problems with the simple BACI design.

The change in a measured variable from two sampling occasions (dots at before and after the impact) in the
control (solid line) or impact (shaded line) sites. In (a), there is little natural variation in the response over time
and so the measured values indicate a change in the mean level. In (b) and (c), natural variation is present,
but, because only one point was sampled before and after impact, it is impossible to distinguish between no
impact (b) and impact (c) on the mean level.

(a)

(b)

Control

Impact

Water quality

Water quality

Control

Impact

Difference
Difference
Before

After
Time

Before

After
Time

The BACI-P design.

The change in a measured variable from multiple randomly chosen sampling occasions (dots at before and
after the impact) in the control (solid line) or impact (shaded line) sites. In (a), there is no impact and the
mean level of the difference (bottommost line) is constant over time. In (b), there is an impact, and the mean
level of the difference (bottommost line) changes over time.

has natural fluctuations over and above the longterm average, then distinguishing between cases
where there is no effect from those where there was
impact is impossible. In terms of our example, differences in the water quality may be artifacts of the
sampling dates and natural fluctuations may obscure
differences or lead to the conclusion that differences
are present when they are not.
BACI-P: Before-after-control-impact
paired designs
Stewart-Oaten et al. (1986) extended the simple BACI
design by pairing surveys at several selected time
points before and after the impact. Both sites are
measured at the same time points. An analysis of how
the difference between the control and impact sites
changes over time would reveal if an impact has occurred (Figure 3.5). The rationale behind the design is
that repeated sampling before the development indicates the pattern of differences over several periods of
potential change between the two sites. This survey
design provides information both on the mean difference in the water quality before and after impact, and
on the natural variability of the water quality measurements. If the changes in the mean difference are
large relative to natural variability, the manager has
detected an effect.
The decision between random and regularly

spaced intervals has been discussed in an earlier

sectionthe same considerations apply here.
Pitfall: As with all surveys, numerous assumptions

need to be made during the analysis (Stewart-Oaten

et al. 1992; Smith et al. 1993). The primary assumption is that the responses over time are independent
of each other. A lack of independence over time
tends to produce false-positives (Type I errors) where
the manager may declare that an impact has occurred
when in fact, none has. In these cases formal time series methods may be necessary (Rasmussen et al.
1993). (The analysis of time series is easiest with regularly spaced sampling points).
Two other assumptions are made: that the difference in mean level between control and impact sites
is constant over time in the absence of an impact effect and that the effect of the impact is to change the
arithmetic difference. In our example, the difference
in the mean water quality between the two sites
would be assumed to be constant over time. The
mean water quality measurements may fluctuate over
time, but both sites are assumed to fluctuate in lockstep with each other maintaining the same average
arithmetic difference. One common way this assumption is violated is if the response variable at the
control site is a constant multiple of the response
variable at the impact site. Then arithmetic differences

33

will depend upon the actual levels. For example, suppose that the readings of water quality at two sites at
the first time point were 200 versus 100, which has an
arithmetic difference of 100; at the second time point,
the readings were 20 versus 10, which has an arithmetic difference of 10; but both pairs are in a 2:1 ratio
at both time points. The remedy is simple: a logarithmic transform of the raw data converts a multiplicative difference into a constant arithmetic
difference on the logarithmic scale. This problem is
commonly found when water quality measurements
are concentrations (e.g., pH).
Underwood (1991) also considered two variations
on the BACI-P design. First, it may not be possible to
sample both sites simultaneously for technical or logistical reasons. Underwood (1991) discussed a
modification where sampling is done at different
times in each site before and after impact (i.e., sampling times are no longer paired), but notes that this
modification cannot detect changes in the two sites
that occurred before the impact. For example, differences in water quality may show a gradual change
over time in the paired design prior to impact. Without paired sampling, it would be difficult to detect
this change. Second, sampling only a single control
site still has the problems identified earlier of not
knowing if observed differences in the impact and
the control sites are site-specific. Again, Underwood
(1991) suggests that multiple control sites should be
monitored. In our example, more than one control

site would be measured at each time point. The variability in the difference between each control site and
the impact site provides information on generalization to other sites.
Enhanced BACI-P: Designs to detect acute versus
chronic effects or to detect changes in variation as
well as changes in the mean
As Underwood (1991) pointed out, the previous designs are suitable for detecting long-term (chronic)
effects in the mean level of some variable. In some
cases, the impact may have an acute effect (i.e., effects only last for a short while) or may change the
variability in response (e.g., seasonal changes become
more pronounced). Underwoods solution is to
modify the sampling schedule so that it occurs on
two temporal scales (Figure 3.6). For example, groups
of surveys could be conducted every 6 months with
three surveys 1 week apart randomly located within
each group. The analysis of such a design is presented
in Underwood (1991). Again, several control sites
should be used to confound the argument about detected differences being site-specific.
This design is also useful when there are different
objectives. For example, the objective for one variable
may be to detect a change in trend. The pairing of
sample points on the long time scale leads to efficient
detection of trend changes. The objectives for another variable may be to detect differences in the mean
level. The short time scale surveys randomly located

Water quality

Control

Impact

Period 1
Period 2
Before impact

Period 1

Period 2
After impact

Time
.

34

The enhanced BACI-P design.

The change in a measured variable from multiple randomly chosen sampling occasions in two periods (dots at
before and after the impact) in the control (black line) or impact (shaded line) sites. The two temporal scales
(sampling periods vs sampling occasions) allows the detection of a change in mean and in a change in
variability after impact.

in time and space are efficient for detecting differences in the mean level.
3.4.2 Issues in impact surveys
Time dependence
Many of the analyses proposed for the above surveys
(e.g., regression or ANOVA) have methodological
problems that need to be resolved before interpreting
the results.
In regression of the characteristics versus time, the
estimated slope is often used as evidence of a longterm change. However, data collected over time
violate the assumption of independence required for
ordinary regression. The estimate of the slope remains unbiased, but typically the estimated standard
error of the slope is too small. The results appear to
be statistically significant when, in fact, there is no
evidence of a change (Neter et al. 1990, Chap. 13) and
a Type I error would have been made.
Comparing the means before and after impact
using ANOVA methods also suffers from the same
problem of correlation among the measurements.
Again, a typical result is that the estimated standard
error of the difference is too small, and results are declared statistically significant when in fact they are
not, and a Type I error would have been made.
An alternative analysis is to use time-series methods that incorporate temporal correlation. The
analysis of time series is quite complex (Nelson 1973)
particularly if the time points are unequally spaced. If
the data points are taken before and after the impact,
the time series analysis can be extended using intervention analysis to test if an impact changed the level
of the series (Rasmussen et al. 1993).
Temporary or permanent monitoring sites
A common question in monitoring surveys is the use
of temporary or permanent monitoring sites. For example, should permanent water quality sampling
sites that are remeasured over time, or temporary
sampling sites that are re-randomized at each time be
used? Many of the concerns are similar to those for
repeated sampling designs discussed earlier. Permanent plots give better estimates of change over time
because the extra plot-to-plot variability caused by
bringing in new plots each year is removed. However,
the costs of establishing permanent plots are higher
than for temporary sites, and the first randomization
may lead to a selection of plots that have some

strange characteristics. Of course, if the measurement

process alters the sampling unit, new plots will have
to be selected for each survey. A compromise solution is a rotating panel survey, where only a part of
the sample is changed at each time point. In large,
complex, long-term designs with multiple objectives,
permanent plots are often the preferred solution
since no survey design is optimal for all objectives
and the objectives change over time.
3.4.3 Impact surveys summary
As noted by Smith et al. (1993), the BACI-P design
and its extensions are one of the best models for impact assessment. These designs can show that
observed differences in ecological variables between
the control and impact sites are neither artifacts of
sampling nor due to temporal trends unrelated to the
impact. The strength of the inference is directly related to the design issues directly under the control of
the managers such as the frequency of sampling and
number of control sites. Because of the potentially
large amounts of data collected, quality assurance
methods need to be employed throughout the length
of the survey so that problems in data management,
data handling, or changes in personnel do not compromise the survey.
3.5 Conclusion
Green (1979) gave 10 principles applicable to any
sampling design; these principles have been paraphrased, reordered, and extended below. Underwood
(1994) also gives some advice on areas of common
misunderstanding between environmental biologists
and statisticians.
1. Formulate a clear, concise hypothesis
The success or failure of a sampling program often
hinges on clear, explicit hypotheses. Woolly thinking
at this stage frequently leads to massive amounts of
data collected without enough planning as to how, to
what end, and at what cost the information can be
subsequently handled. Hypotheses should be stated
in terms of direct, measurable variables (e.g., action
X will cause a decrease in Y). The hypotheses to be
tested have implications for what and how data are to
be collected.
2. Ensure that controls will be present
Most surveys are concerned with changes over time,
typically before and after some impact. Effects of an

35

impact cannot be demonstrated without the presence

of controls serving as a baseline so that changes over
time, unrelated to the impact, can be observed. Without controls, no empirical data are available to refute
the argument that observed changes might have occurred regardless of impact.
3. Stratify in time and space to reduce heterogeneity
If the area to be sampled is large and heterogeneous
(highly variable), then sampling from the entire area,
ignoring the known heterogeneity, reduces the precision of the estimate. Extra variation may be
introduced to the measured variable solely by differences within the survey area unrelated to the
treatment. By stratifying the survey area in advance
(also known as blocking in the experimental design
literature), this extra variability can be accounted for.
The judicious choice of auxiliary variables can also be
used to increase precision of the estimates.
4. Take replicate samples within each combination
of time, space, or any other controlled variable
Differences among treatments can only be demonstrated by comparing the observed differences among
treatments with differences within each treatment.
Lack of replication often restricts the interpretation
of many experiments and surveys to the sampled
units rather than to the entire population of interest.
It is imperative that the replicates be true replicates
and not pseudoreplicates (Hurlbert 1984), where the
same experimental unit is often measured many times.
5. Determine the size of a biologically meaningful,
substantive difference that is of interest
A sufficiently large survey (i.e., with large sample
sizes) can detect minute differences that may not be
of biological interest. It is important to quantify the
size of a difference that is biologically meaningful before a survey begins so that resources are not wasted
either by performing a survey with an excessive sample size or by performing a survey that has lower
power to detecting this important difference.
6. Estimate the required sample sizes to obtain adequate power to detect substantive differences
or to ensure sufficient precision of the estimates
In this era of fiscal restraint, it is unwise to spend significant sums of money on surveys or experiments
that have only a slight chance of detecting the effect
of interest or give estimates that are so imprecise as
to be useless. Such designs are a waste of time and
money.

36

If the goal of the survey is to detect a difference

among populations, the required sample sizes will
depend upon the magnitude of the suspected difference, and the amount of natural variation present.
Estimates of these qualities can often be obtained
from experience, literature reviews of similar surveys,
or pilot surveys. Simulation studies can play an important role in assessing the efficiency of a design.
If the goal is descriptive, then the required sample
sizes will depend only upon the natural variation present. As mentioned, estimates of the variability can
be obtained from experience, literature reviews, or a
pilot survey.
As noted earlier, it may be infeasible to conduct a
pilot survey, historical data may not exist, or it may
be difficult to reconcile sample sizes required for different objectives. Some compromise will be needed
(Cochran 1977, pp. 8182).
One common misconception is that sample size is
linked to the size of the population. To the contrary,
the sample sizes required to estimate a parameter in a
small population with a specified precision are the
same as in a large population. This non-intuitive result has a direct analogue in testing a pot of soup for
saltthe cook tastes only a spoonful regardless of
pot size.
7. Allocate replicate samples using probabilistic
methods in time and space
There is a tendency to allocate samples into representative or typical locations. Even worse are
convenience samples where the data are collected at
sampling points that are easily accessible or close-athand. The key to ensuring representativeness is
randomization. Randomization ensures that the effects of all other uncontrollable variables are equal,
on average, in the various treatment groups or that
they appear in the sample, on average, in the same
proportions as in the population. Unless the manager
is omniscient, it is difficult to ensure that representative or typical sites are not affected by other,
unforeseen, uncontrollable factors.
Notice that a large sample size does not imply representativeness. Randomization controls
representativeness; sample size controls statistical
power.
8. Pretest the sampling design and sampling
methods
It is difficult to spend effort on a pilot survey
knowing that the data collected may not contribute
to the final results and may be thrown away. Howev-

er, this approach is the only way to check if serious

problems exist in the survey, if the size of the survey
unit is appropriate, if the data collection forms are
adequate, and if the actual level of variability is present in the field, etc.
After a pilot survey has been conducted, its results
can be used to modify the proposed design and finetune such aspects as the required sample size. In
many cases, a pilot survey shows that the objectives
of the proposed survey are unobtainable for the projected cost and effort and the survey must be
substantially modified or abandoned.
9. Maintain quality assurance throughout the
survey
Despite best efforts, plans will deviate during the
course of the survey, particularly if the survey extends
over many years and personnel changes. Many of the
principles of statistical process control can be applied
here (Montgomery 1991). For example, ensure that
instruments are recalibrated at regular intervals, sampling protocols are followed consistently among
different team members, and data are being keyed
correctly.
10. Check the assumptions of any statistical analysis
Any statistical procedure makes explicit and implicit
assumptions about the data collected. Match the
analysis with the survey design. In many cases, a statistically significant result can be obtained
erroneously if assumptions necessary for the analysis
were violated.
11. Use the Inter-Ocular Trauma Test
Presentation of final results is just as important as design, execution, and analysis. A survey will be of
limited usefulness if it sits on a shelf because other
readers are unable to interpret the findings. Good
graphical methods (figures, plots, charts, etc.) or presentations will pass the Inter-Ocular Trauma Test
(i.e., the results will hit you between the eyes!)
Despite their limitations, uncontrolled events can
play a useful role in adaptive management. The study
of uncontrolled events and designed experiments differ in two important dimensions:
1. The amount of control. As the name implies, the
study of uncontrolled events does not give the
manager the ability to manipulate the explanatory
variables.
2. The degree of extrapolation to other settings. The
lack of randomization implies that the manager

must be careful in extrapolating to new situations

because of the possible presence of latent, lurking
factors.
These differences imply that inferences are not as
strong as those made after carefully controlled experiments, but the results often lead to new hypotheses
being tested in future research. Despite the weaker
inferences from studying uncontrolled events, the
same attention must be paid to the proper design of a
survey so that inadvertent biases do not taint the
conclusions.
References
Box, G.E.P. and G.C. Tiao. 1975. Intervention analysis with applications to economic and
environmental problems. J. Am. Statist. Assoc.
70:709.
Buckland, S.T., D.R. Anderson, K.P. Burnham, and
J.L. Laake. 1993. Distance sampling: estimating
abundances of biological populations. Chapman and Hall, London, U.K.
The standard monograph on the statistical analysis of distance sampling experiments.
Cochran, W.G. 1977. Sampling techniques. J. Wiley,
New York, N.Y.
One of the standard references for survey sampling. Very technical.
Eberhardt, L.L. and J.M. Thomas 1991. Designing
environmental field studies. Ecol. Monogr.
61:5373.
An overview of the eight different field situations
as shown in Figure 1.
Fletcher, D.J. and B.F.J. Manly (editors). 1994. Statistics in ecology and environmental monitoring.
Univ. Otago Press, Dunedin, N.Z.
A collection of papers at a moderate to advanced
level on a wide range of topics.
Green, R.H. 1979. Sampling design and statistical
methods for environmental biologists. J. Wiley,
New York, N.Y.
One of the first comprehensive unified treatments
of sampling issues for environmental biologists.
Very readable.
______. 1993. Application of repeated measures designs in environmental impact and monitoring
studies. Austr. J. Ecol. 18:8198.

37

Hurlbert, S.H. 1984. Pseudoreplication and the design

of ecological field experiments. Ecol. Monogr.
52:187211.
A critique of many common problems encountered in ecological field experiments.

Nelson, C.R. 1973. Applied time series analysis for

managerial forecasting. Holden-Day, San Francisco, Calif.
A primer on the basic time series analysis
methods.

Keith, L.H. (editor). 1988. Principles of environmental sampling. Am. Chem. Soc., New York, N.Y
A series of papers on sampling mainly for environmental contaminants in ground and surface
water, soils, and air. A detailed discussion on
sampling for pattern.

Nemec, A.F.L. 1993. Standard error formulae for cluster sampling (unequal cluster sizes). B.C. Min.
For., Res. Br., Victoria, Biometric Inf. Pamph.
No. 43.

Kish, L. 1965. Survey sampling. J. Wiley, New York,

N.Y.
An extensive discussion of descriptive surveys
mostly from a social science perspective.

Neter, J.N., W. Wasserman, and M.H. Kutner. 1990.

Applied linear statistical models: regression,
analysis of variance, and experimental designs,
3rd ed. Irwin, Boston, Mass.
A standard treatment of regression and experimental design suitable after a first course in
statistics.

______. 1984. On analytical statistics from complex

samples. Sur. Methodol. 10:17.
An overview of the problems in using complex
surveys.
______. 1987. Statistical designs for research. J. Wiley,
New York, N.Y.
One of the more extensive discussions of the use of
complex surveys in analytical surveys. Very technical.
Krebs, C.J. 1989. Ecological methodology. Harper and
Row, New York, N.Y.
A methods books for common techniques used in
ecology.
Milliken, G.A. and D.E. Johnson 1984. The analysis of
messy data: Vol. 1: designed experiments. Van
Nostrand Reinhold, New York, N.Y.
A complete treatise on the analysis of unbalanced
data in designed experiments. Requires a background in the use of ANOVA methodology.
Montgomery, D.C. 1991. Introduction to statistical
quality control. J. Wiley, New York, N.Y.
A standard introduction to the principles of
process control.
Myers, W.L. and R.L. Shelton 1980. Survey methods
for ecosystem management. J. Wiley, New
York, N.Y.
Good primer on how to measure common ecological data using direct survey methods, aerial
photography, etc. Includes a discussion of common survey designs for vegetation, hydrology,
soils, geology, and human influences.

38

______. [n.d.]. Design of experiments. This volume.

Otis, D.L., K.P. Burnham, G.C. White, and D.R.

Anderson. 1978. Statistical inference from capturedata on closed animal populations. Wildl.
Monogr. 62.
The standard monograph on the statistical analysis of mark-recapture experiments in closed
populations.
Pollock, K.H., J.D. Nichols, C. Brownie, and J.E.
Hines 1990. Statistical inference from
capturerecapture experiments. Wildl. Monogr. 107.
The standard monograph on the statistical analysis of mark-recapture experiments in open
populations.
Rao, J.N.K. 1973. On double sampling for stratification and analytical surveys. Biometrika
60:12533.
Rasmussen, P.W., D.M. Heisey, E.V. Nordheim, and
T.M. Frost 1993. Time series intervention analysis: unreplicated large-scale experiments. In
Design and analysis of ecological experiments.
S.M. Scheiner and J. Gurevitch (editors). pp.
13858. Chapman and Hall, New York, N.Y.
Scheiner, S.M. and J. Gurevitch 1993 (editors). Design
and analysis of ecological experiments. Chapman and Hall, New York, N.Y.
Sedransk, J. 1965a. A double sampling scheme for analytical surveys. J. Am. Statist. Assoc.
60:9851004.

______. 1965b. Analytical surveys with cluster sampling. J. Royal Statist. Soc., B, 27:26478.
______. 1966. An application of sequential sampling
to analytical surveys. Biometrika 53:8597.
Skalski, J.R. and D.S. Robson 1992. Techniques for
wildlife investigations: design and analysis of
capture data. Academic Press, New York, N.Y.
Presents methods for conducting experimental inference and mark-recapture statistical studies for
fish and wildlife investigations.
Smith, E.P., D.R. Oruos, and J. Cairns Jr. 1993.
Impact assessment using the before-aftercontrol-impact (BACI) model: concerns and
comments. Can. J. Fisheries Aquatic Sci.
50:62737
Stewart-Oaten, A., J.R. Bence, and C.W. Osenberg.
1992. Assessing effects of unreplicated perturbationsno simple solutions. Ecology
73:1396404.
Stewart-Oaten, A., W.M. Murdoch, and K. Parker.
1986. Environmental impact assessment:
pseudoreplication in time? Ecology
67:92940.
One of the first extensions of the BACI design discussed in Green (1979).
Thompson, S.K. 1992. Sampling. J. Wiley, New York,
N.Y.
A good companion to Cochran (1977). Has many
examples of using sampling for biological populations. Also has chapters on mark-recapture,
line-transect methods, spatial methods, and
adaptive sampling.
Underwood, A.J. 1991. Beyond BACI: experimental
designs for detecting human environmental
impacts on temporal variations in natural populations. Austr. Marine and Freshwater Res.
42:56987.
A discussion of current BACI designs, and an enhanced BACI design to detect changes in
variability as well as in the mean response.
______. 1994. Things environmental scientists (and
statisticians) need to know to receive (and give)
better statistical advice. In Statistics in ecology
and environmental monitoring. D.J. Fletcher
and B.F. Manly (editors). Univ. Otago Press,
Dunedin, N.Z.

39

4 RETROSPECTIVE STUDIES
G. JOHN SMITH

Statistics are extremely important to resource management. The rigour of gathering and analyzing data for
a proper statistical analysis often conflicts with the
need to obtain the required information within a short
time frame and within a limited budget. Retrospective
studies are one alternative to a fully controlled, or
prospective, study. These studies offer a compromise,
which uses existing data or circumstances. This approach greatly shortens the time between the inception of the study and the presentation of the results,
as well as reduces the cost. A considerable degree of
methodological correctness can be maintained by
careful design, analytical techniques, and presentation
of results.
As with any compromise, retrospective studies must
be used carefully. In retrospective analyses, often the
results are preliminary, and sometimes do not allow
for quantitative model building, hypothesis testing, or
point estimation. However, by carefully presenting results and designing the study, and being aware of the
pitfalls inherent in individual analyses, a great deal of
useful information can be obtained. Even if the results
are interim, such efforts can be beneficial to the gathering of future information, and to decision-making
processes. Managers need all the tools available to
properly manage forest resources and adapt to changing conditions and priorities.
In this chapter, a definition and many examples are
presented to demonstrate the differences between
prospective and retrospective studies. Each example is
reviewed with an emphasis on contrasting retrospective and prospective studies, and pointing out the
strengths and weaknesses of the retrospective approach. Finally, some suggestions are given regarding
the design of retrospective studies and the analysis of
retrospective data.

that all comparisons and estimates are statistically

valid and free of bias and confounding factors.
The values of expedience and rigour usually conflict. However, both are important. In dynamic
natural systems such as forests, many effects of events
(such as particular logging practices or forest fires)
take many years to manifest themselves. Studies have
limited time and resources. In the real world, statistics involves art as well as science. To provide valid
and accurate results, technical considerations such as
experimental or survey design, sample size and allocation, and the desired accuracy or precision of the
results are necessary. However, with limited resources, these requirements must be balanced with
constraints such as the ability to execute the field
procedures, weather, training and management of
participants and field personnel, financial resources,
and time available to gather and analyze the data.
This chapter discusses some compromises that
attempt to satisfy both values. As with any compromise, sometimes it provides excellent results, at other
times it is the best of a difficult situation, and occasionally it is unworkable.
Nemec (this volume, Chap. 2) discusses designed
experiments, those where the experimenter assigns
treatments and can manipulate the experimental factors at will. Schwarz (this volume, Chap. 3) discusses
the study of uncontrolled events, those where the experimenter has a very limited ability to manipulate
the experimental factors, and methods for improving
the information that can be gained from them. However, what happens when the results of the studies
will not be available for a long time, or where it is unrealistic or unacceptable to implement past practices
that have been condemned? An example might be
a study of the effect of large-scale clearcutting. How
can we use data that already exist and what are the
advantages and pitfalls?

4.1 Introduction

4.2 Definitions

In any study involving data, two values help determine the methodology to apply. The first is
expedienceto complete the work as quickly and efficiently as possible to meet deadlines and minimize
cost. The second is rigourto scrupulously apply statistical methods and experimental controls to ensure

All studies that involve the gathering and synthesis of

data can be placed in one of two categories, depending on the nature of the design, the data, and the
analysis. The first category is called a prospective
study. This term, which has been used extensively in
biostatistical literature relating to medical science

Abstract

41

(Bailar and Mosteller 1986; Rosner 1986), indicates

that the data are collected and analyzed without reference to past data or circumstances.
A prospective study may be either designed as described in Chap. 2 or uncontrolled (see Chap. 3). The
second category is called a retrospective study. In this
type of study, data that have already been collected
for other purposes, or that are of useful historical
circumstances, are used directly.
To clarify this point, let us refer to the following
definitions from Rosner (1986, p. 326). Although
these definitions relate specifically to medical science,
the underlying ideas are directly applicable to
forestry.
Definition 10.4: A prospective study is a study in
which we identify a group of disease-free individuals at one point in time and follow them over a
period of time until some of them develop the
disease. We then try to relate the development of
disease over time to other variables measured at
baseline.
Definition 10.5: A retrospective study is a study in
which we initially identify two groups of individuals: (1) a group that has the disease under study
(the cases) and (2) a group that does not have the
disease under study (the controls). We then try to
relate their prior health habits to their current
disease status.
Although not from forestry, these definitions emphasize conceptual differences between the two
approaches for designing a study. In the prospective
case, all of the data collection and design of the study
are based on events that happen after the inception
of the study. In the previous definitions, we start out
with similar subjects and then observe what happens.
Subjects will be divided into categories based on
factors that are suspected to be related to the development of the disease. The researcher will then try to
relate the development of the disease to various characteristics and behaviours of the individuals.
In the retrospective case, data that are already
available, or are of circumstances or events that have
already happened before the study is initiated, are
used extensively. In the previous definitions, we start
out with two different groups of subjectsone with
the disease, the other withoutand then work backwards to determine relationships between the
acquisition of the disease and the characteristics and

42

behaviours of individuals. The References section

contains many references to the statistical treatment
of data in particular instances. Most of the statistical
literature relates to the medical field (e.g., Bailar and
Mosteller 1986; Greenberg 1988; Hoogstraten and
Koele 1988; Koele and Hoogstraten 1988; Sikkel 1990;
Weinberg et al. 1993).
The distinction between prospective and retrospective is significant, as it represents a great difference in the measure of control over the studys design
and execution. Note that this method of categorizing
studies does not refer to particular statistical or survey methodologies.
If all things were equal, the obvious choice would
be a prospective approach because it provides more
control over the design and implementation of the
study. A prospective study offers the option to determine exactly what data to obtain, to determine the
survey techniques or the design of the experiment,
and, in some cases, to exercise control over the types
of treatments and their assignment to sampling units.
However, all things are not equal. The use of prior
knowledge or taking advantage of existing results is
critical in forest management for many reasons. First,
enormous savings of time, labour, and financial resources can be realized if data or results from prior
studies or surveys are used. Second, in some instances it may be impossible to re-create the exact
circumstance that occurred at some previous time or
in another location. Third, social and political pressures may prevent the execution of treatments that
would likely have significant negative impacts on an
ecosystem or community. Retrospective analysis is
especially useful in planning and designing future
studies. For example, prior studies can help in estimating sample sizes, determining statistical power
(Anderson, this volume, Chap. 6) or estimating prior
probabilities in a Bayesian analysis (Bergerud and
Reed, this volume, Chap. 7). Also, prior data can be
used to develop or refine hypotheses or provide information on the required time span of a prospective
study.
Another important consideration in the design of
a study is the fact that resource managers must make
decisions regularly. Because making decisions based
on some rather than no knowledge is better, retrospective analysis can be effectively used. A carefully
designed retrospective study can be accomplished
within a much shorter time frame than a prospective
one. The researcher may have less freedom to control
the design, but this shortcoming is offset by the

requirement for fewer resources and the availability

of other sources of knowledge and data. Thus, answers can be obtained quickly from a retrospective
analysis, whereas a prospective study might not be
completed before a decision must be made.
4.3 Examples
To further illustrate the concepts mentioned in the
previous sections, let us consider the following
examples.
4.3.1 Example A: effect of landslides on water
quality
To assess the impact of major natural disturbances
such as landslides on water quality, it would be socially unacceptable to create such events. An alternative is to consider areas where these events have
already occurred. Furthermore, significant time, effort, and money are saved by using data from previous natural disasters to compare or observe trends
rather than waiting for future events to occur. This
analysis would be termed retrospective since the data
collected will be gathered from events that have already transpired.
This application fits the definition for a retrospective study given in Section 4.2. Let us compare this
example with the definition point by point. In the example, we identify two types of areas (corresponding
to groups of individuals in the medical definition): in
one area landslides have already occurred (cases), in
the other they have not (controls). We then relate the
water quality (prior habits) to the current status of
the area. The experimental design will not be as rigorous as for a prospective study because the sampling
will be opportunistic rather than being based on a statistically rigorous design. To compare the water
quality in slide and non-slide areas, it would be ideal
to compare areas that were identical in all respects
except for the fact that a slide occurred in one and
not the other. If this comparison were achievable,
then any difference in water quality would be due
solely to the slide. However, this comparison will
generally not be possible, and hence the analysis and
interpretation of the data will need to take this into
account.
4.3.2 Example B: forest bird populations
Retrospective analysis is extremely useful in the study
of long-term trends. Consider, for example, a study
of the long-term effects of clearcutting on forest bird

populations. A prospective approach would consist

of clearcutting several areas according to a specific
design in which natural factors, which affect the
regeneration and stand development, can be controlled. The next step would involve collecting data
on these areas for the next, say 80 years. Since managers would likely have to make a decision before the
results from this study are available, the disadvantage
is obvious. Retrospective analysis can help the manager by providing timely information.
An alternative to the prospective approach is to
consider areas that have been clearcut in the past.
Bird populations can be assessed at various stages of
regeneration after clearcutting. This method is a retrospective study because advantage is being taken of
circumstances that serve as proxies (alternatives) for
treatments (in this case, areas that have been clearcut previously). This approach greatly decreases the
duration of the study. As in the previous example, a
compromise must be made in the experimental design because the sampling will be opportunistic and
natural factors such as those previously mentioned
would not be controllable to the same degree as in a
prospective study.
4.3.3 Example C: economic effect of forest fires
Often starting a study from scratch is difficult. For
example, past studies have determined the extent to
which long-run timber supply and the flow of economic benefits from a forest can be reduced by fire.
To assess the costs and benefits of a fire protection
program, the probability of destructive fires must be
estimated (Reed 1995) in one of two ways. First, retrospective analysis of historical fire data could be
undertaken. The second option, a prospective analysis involving the selection of several areas based on a
suitable sampling plan, would involve, according to
plan, setting fires in some of the areas, and leaving
others undisturbed. The resulting economic effects in
the burned areas could be compared to those in areas
that were not set ablaze. While this second option is
an interesting design, and from a purely statistical
point of view has many advantages, it is socially unacceptable. Furthermore, this study would clearly be
very long and highly impractical.
4.3.4 Example D: effect of herbicides
The provision of information to direct future studies
is an important contribution of retrospective analysis. In 1991, a study investigated the impacts of
herbicides on grizzly bear forage production in the

43

Coastal Western Hemlock zone (Hamilton et al.

1991). Retrospective analysis of past herbicides treatments provided information on the relationship
between stand structure, density, and forage availability in glyphosate-treated stands. This information
was used to guide the range of treatments to use for
further investigation and testing.
4.3.5 Example E: preservation of biodiversity
The idea that green tree retention mimics the stand
structure remaining after natural disturbances such
as fires is an important assumption that currently
guides much of the management for biodiversity
(B.C. Ministry of Forests and B.C. Environment
1995). An example of this principle is found in Traut
(1995). He studied the preservation of biodiversity in
green tree retentiona logging practice where some
large trees are left uncut in each cutting unitby examining areas that had been ravaged by fire. The
effects of fire were assumed to be analogous to the
effect of green tree retention, because several trees
survive the fire as they would survive the logging.
Although fire and logging using green tree retention
are different, if they have similar effects on biodiversity, effects can be studied that would not be possible
otherwise, except by waiting several decades.
Acker et al. (1995) did a similar study in the
Willamette National Forest, as did Zenner (1995) on
Douglas-fir and western hemlock in the Western
Central Oregon Cascades. Peck and McClune (1995)
did a similar study in western Oregon, but targeted
canopy lichen communities.
4.3.6 Example F1: site index versus tree density
Goudie (1996) presented a paper discussing the relationship between the site index and the tree density
in lodgepole pine stands. He postulates that increasing density represses growth, and that to measure the
site index on dense stands underestimates the index.
It is often assumed that these two factorsdensity
and growthare independent, thus challenging the
usefulness of thinning as a measure to promote
growth. The investigation of this phenomenon with
a prospective study would require a design in which
stands of various densities would be subjected to a
variety of thinning regimes, and the effect on the site
indices observed.
Goudie extensively used stands that had been
thinned in the past, sometimes by natural causes
such as fire. He takes great effort to ensure that
stands representing various tree densities are similar

44

in all respects except for site density so that a difference in site index can be directly attributable to the
density. He correctly points out the potential biases
in the selection of plots, and although the design has
some shortcomings, he finds compelling evidence to
suggest a density-dependent repression on site index.
The dependence was most noticeable in very dense
stands and hardly noticeable in stands of lower density. Thus growth and yield models that do not take
repression into account would not be applicable to
very dense stands
Here is an example where existing data have been
used considerably. The results were presented along
with a discussion of the potential flaws and biases,
and a great deal of information and knowledge was
gained. Researchers would have had to wait years for
the results from a prospective study.
4.3.7 Example F2: site index versus tree density
Thrower (1992) also conducted a study at Larsons
Bench, east of Chilliwack, B.C., on the relationship
between density and height-and-diameter growth in
coastal Douglas-fir stands. This study, too, tried to
mimic an experimental design by comparing a natural (unspaced) and a previously logged (spaced) area,
each with several similar ecological units. This study
used an existing thinning, which was not designed for
a research study. Some matching of units was possible. However, since the two areas did not have the
same conditions at the time of thinning, it was acknowledged that the comparison of the growth rates
in the two areas may be a combination of growth rate
and initial conditions.
4.4 Contrasting Data Collected from Prospective
and Retrospective Studies
Data collected from a prospective study can be used
and analyzed directly, and the interpretation can be
based on sound statistical design and analysis. By
contrast, data from a retrospective study have fewer
statistical controls, and often some components cannot be combined with other components of the data
unless additional assumptions are made about their
comparability.
Consider Example B (forest bird populations) in
the previous section. For a prospective study, the same
treatments would be used throughout 80 years of the
study. The analysis of trends for various species is relatively straightforward. In the retrospective study
there may be a variety of stands where clearcutting

was done at various times in the past. The choice of

stands will be limited and comparisons may have to
be made among stands that differ not only in the
time since being clearcut, but also in other variables
such as weather, initial stand structure, slope, aspect,
and elevation (which cannot be controlled).
In a retrospective analysis, stands must be matched
as closely as possible (i.e., choose stands for comparison that are as similar as possible in all ways except
for the time since the clearcut). If we can achieve this,
then the difference in the populations will be due to
clearcutting and not to other factors. Lack of matching requires making additional assumptions about
the comparability of the stands included in the analysis. For example, you might assume that for different
areas, the weather conditions since the clearcutting
will have affected all the clearcuts in the same way.
Before making this assumption, a study of each area
would be advisable to assess the assumptions plausibility. Thus the lower cost and shorter time of the
study are somewhat offset by a design with fewer
controls. Because of the greater use of untested assumptions, the interpretation of the results will
demand greater caution than in a prospective study.
For example, in many instances, results cannot be
generalized, nor hypotheses tested. However, hypotheses can usually be generated for future studies.
Retrospective analysis offers other benefits over
long-term experiments. In Example B (forest bird
populations), a prospective experiment has the risk
that some of the treatment areas may eventually be
used for purposes other than forests (e.g., become
agricultural land or urban areas) during the course of
the experiment. Hence the treatment areas are lost to
the final analysis. This risk is present even in studies
that last for considerably shorter periods of time.
4.5 Comparing the Development of a Retrospective
and Prospective Study
A comparison of the development of prospective and
retrospective studies, especially where they differ, will
help us to appreciate how to use retrospective analysis effectively. Figure 4.1 outlines the basic steps involved in the development of the two types of studies.
The left side of Figure 4.1 follows the steps in a retrospective study, while the right side displays the
steps in a prospective study. In a retrospective study,
several steps replace the experimental design stage in
a prospective study. Comparable data sources must
be found that replace rigorous statistical and field

procedures. Any difficulty in finding truly comparable data sources inhibits a rigorous statistical design.
The lack of control in the retrospective study means
that additional assumptions will be required to
perform the analysis, implying that more care is
needed in the interpretation of the results. For example, if in Example B (forest bird populations) we
were unable to match clearcut and natural stands, we
may find it necessary to compare a clearcut stand
with a natural one that has a different aspect and tree
species mix. We would assume that these factors in
the two stands do not make a difference to the
species present.
In Example F2 (site index versus tree density) the
comparability of thinned and unthinned stands was
somewhat compromised because of differences in the
conditions of the stands at the time of clearcutting.
Consequently, the results may be less widely applicable. The advantages, however, are the potential to
greatly shorten the studys time frame, and to reduce
the effort and resources required.
The previous comments do not imply that retrospective analyses are inadequate, but do indicate the
importance of additional diligence during their design, interpretation, and analysis. Furthermore, the
study of alternative data sources greatly enhances the
ability to design an efficient study using knowledge
already gained about the subject under investigation.
This study may lead to rejecting some scenarios or
considering others that might not otherwise be
obvious. Also, retrospective data provide advance
warnings of difficulties one might expect, which
could lead to the failure of an experimental design. A
retrospective study can often be used as a pilot study
to obtain qualitative information.
4.6. Studies with Significant Retrospective and
Prospective Components
Sometimes it is not clear whether to classify a study
as prospective or retrospective. The study may appear to be a prospective study, but after scrutiny may
be found to be more like a retrospective one. Thinking about which category the study belongs to will
help us understand how the assumptions might affect
the interpretation of results. The following examples
illustrate this point.
4.6.1 Example G: change in the measuring instrument
Consider the case where a questionnaire survey of

45

Retrospective Study

Common Elements

Prospective Study

Dynamic
natural system

Set goals

Search for
retrospective data
or circumstances

Plan project,
study or analysis
Experimental
design

Determine
applicability
of retrospective
data to study

Determine
methods for
integrating data

Field procedures
and data collection

Analysis

Evaluation,
approximate
interest

Inference
Interpretation
of results

Dissemination
of results,
recommendation

Action

46

Comparing the development of a retrospective and prospective study.

hunters has been done for many years, yielding information about hunter activity, number of animals or
birds killed, etc. With advancing technology and
more knowledge of the resource, the survey is
redesigned, new computer technology is used, the
questions are clearer, and perhaps one or two are
added or deleted. The same series of statistics is generated before and after the redesign.
Now suppose that later a trend analysis of the
number of days hunted and the number of birds of
various species killed is required over a time period
that spans both the old and the new methodology.
Because the same series of statistics has been generated throughout the period covered by the study, the
researcher might infer that it would be valid simply
to use the data without further consideration. However, changes in the way a question is asked or in the
way the questions are edited may influence a persons
response and therefore an apparent trend may be due
to the question rather than the activity. Any analysis
would require assumptions about the comparability
of the data. The interpretation of the results would
need to acknowledge these assumptions and examine
their potential implication. Cooch et al. (1978) describe some of the effects of survey changes on results

in the Canadian Wildlife Services National Waterfowl Hunter Surveys.

4.6.2 Example H: changing statistical methodology
Consider the following simplified example. Suppose
that for a number of years the ratio of immature to
adult Ancient Murrelets (Synthliboramphus antiquus), called the I/A ratio, has been collected for a
specific population consisting of five colonies. This
ratio is of interest because the higher it is, the greater
the number of young per nest. This is one measure of
the health of the population. These birds are colonial
(i.e., nest close together in small areas or colonies).
Each year a sample of approximately 20 nests is observed from each of five colonies under study, and
the number of adults and young are counted. The
overall I/A ratio was computed by simply averaging
the ratios from the five colonies (Table 4.1).
In 1996, the size (i.e., the total number of nests) of
each colony, in addition to the number of immatures
and adults in the observed nests, is recorded. The
results are given in Table 4.2.
As in previous years the comparable I/A ratio can
be computed as the simple average of the I/A ratios
for the five colonies. Its value is 1.07.

. I/A ratio by colony 19901995

I/A ratio by colony

Year

Average
I/A ratio

1990
1991
1992
1993
1994
1995

0.83
0.90
0.95
0.92
0.85
0.96

1.10
1.02
0.97
1.03
0.93
1.05

0.90
0.84
0.88
0.81
0.78
0.91

1.12
1.25
1.12
1.17
1.08
1.20

1.35
1.41
1.22
1.27
1.19
1.36

1.06
1.08
1.03
1.04
0.97
1.10

Total

. I/A ratios for 1996

Colony

Nests observed
Immatures
Adults
I/A ratio
Colony size

20
29
40
0.73
58

20
39
40
0.98
75

C
18
34
36
0.94
105

D
17
39
34
1.15
178

20
63
40
1.58
285

95
148
190
701

47

I/A ratio

1.5

1.0

0.5
0

100

200

300

Colony size
. I/A ratio vs colony size.

In 1996, with the information about colony size

being added, the I/A ratio can be weighted by the size
of the colony. The I/A ratio calculated in this way will
be 1.24. In this estimation procedure, the counts are
weighted within each colony, which tends to reduce
the bias. Furthermore, the I/A ratios are examined as
a function of colony size (Figure 4.2).
Subject to confirmation by a statistical test,
Figure 4.2 seems to indicate that larger colonies are
more productive. The difference between the two estimates of the overall I/A ratio arises because the
larger colonies tend to have higher I/A ratios and, in
the second estimate, the colonies influence the estimate in proportion to their size.
For management purposes, the health of the
population in the five colonies is needed. Should the
lower figure (I/A=1.07) be presented because it provides the best comparison with information from
previous years? Or should the higher figure (I/A=1.24),
which will be less biased and a better estimate, be used?
Alternatively, the analysis from previous years
could be redone by weighting the results prior to
1996 by the 1996 colony sizes. If colony sizes do not
vary a great deal from one year to the next, this
method may be a good way to compare as well as update the results from previous years. The downside to
this procedure is that if colony sizes fluctuate considerably over time, poorer estimates may result.
Furthermore, the agency may present an image of incompetence by not presenting a coherent
methodology.
Several solutions may be correct. The choice will
depend on the priorities of the agency and the proposed use of the information obtained. If the data are
used as input to a mathematical model, using the best
information available would be the highest priority;
it may be essential to adjust previous years information if there is reason to believe this is better. On the

48

other hand, if trends over time are desired, then

consistency is important, even at the expense of a systematic bias in each years results.
The previous two examples represent cases in a
continuing study. The researcher has introduced the
change in the circumstance through a change in
methodology. Often, these cases might not be considered retrospective but simply studies with
statistical bias. However, the same factors that exist
in Examples A through F, exist here. We are using
previously collected data that are not entirely compatible with data we currently collect. The same
caveats exist here as in the previous examples. In this
sense, Example H is similar to a retrospective study.
The following example demonstrates that a study
designed as a prospective one often has elements of a
retrospective nature. The existence of such elements
should not necessarily result in a study being classified as retrospective. Many studies, of necessity, contain some retrospective and some prospective
elements. This happens because we often have limited control over the data we collect. Hence many
studies are hybrids and it is an oversimplification to
classify every study as either purely prospective or
purely retrospective. Consider the following example
from wildlife conservation.
4.6.3 Example I: Pacific Brant
Each year, Pacific Brant (Branta bernicla) migrate
north in the spring and early summer and south in
the autumn. During the northward migration, one of
the major stopping areas is the Parksville-Qualicum
area on the eastern shore of Vancouver Island. The
birds stay there for anywhere from several hours to
over 10 days before moving on. Proper conservation
and management of the Pacific Brant requires
knowledge of how many brant use this area, and for
how long.
A solution is to count the number of brant in the
area, which seems easy because the vast majority feed
on the beaches along the seashore. For counts, where,
how often, and what data to collect must be determined.
If counts are done at any regular intervals (e.g.,
daily) then those birds that stay several days may be
counted more than once. Furthermore, brant that
stay for a shorter time than this interval may be
missed altogether. Certainly both situations cannot
be accommodated unless other data are used.
Multiple counts can be accounted for through the
observation of banded birds since the unique band

number provides a method for identifying individual

birds. This method can work quite well, but simplifying assumptions are necessary. First, assumptions
would be required for any prospective banding study,
such as the randomness of the banded birds among
the population as a whole, and the likelihood of seeing a banded bird (e.g., the band may be hidden if the
bird is swimming, or its identification number may
not be readable by the observer).
Banding will be expensive unless brant that have
already been banded in previous studies can be used.
The incorporation of bands from elsewhere introduces a retrospective component into the study and
the requirement for additional assumptions. We
must consider where the brant were banded. Were
brant from some wintering areas subjected to more
intensive banding efforts than those from others?
Even if we can answer this question, we need to know
how many brant from various wintering areas pass
through Parksville-Qualicum. Some components of
the population may have a high proportion of banded birds, and others may have none. It may be necessary to make some simplifying assumptions about
this source of data as well as an assessment of its
validity.
The decisions relating to these issues will result in
a compromise between optimum statistical methods,
which may be impossible to implement, and allowing
less desirable retrospective components. The additional assumptions under which the analysis was
done should be stated clearly. More information
about the Pacific Brant is given in Campbell et al.
(1990a), and estimation procedures in Routledge et
al. (1998).
4.7 Guidelines for Designing Retrospective Studies
Several examples of retrospective studies, along with
some weaknesses and strengths, have been discussed.
As practitioners and decision-makers, we need guidance concerning the factors that distinguish a good
retrospective study from a mediocre one. Although
retrospective and prospective designs and analyses
have basic differences, we should endeavour to apply
sound statistical principles to both. The main difference is the rigour with which these principles can be
applied to each, and to what extent compromises
must be made. The following list provides some principles for good design. Although not exhaustive, this
list focuses on several points where retrospective and
prospective studies tend to differ.

Probability sampling
Awareness and clear statement of assumptions
Design considerationscontrolling variation
Use of direct measurements rather than proxies for
the measurements

Let us look at these four principles and consider

how we might adhere to these in both retrospective
and prospective analyses.
4.7.1 Probability sampling
Probability sampling introduces an element of randomness into the process of selecting the sampling
units. As pointed out by Schwarz (this volume,
Chap. 3) randomness is essential to virtually all statistical methods. It serves to remove many inadvertent
biases and is central to the strict application of statistical theory to experimental design (Nemec, this
volume, Chap. 2) and to the proper assessment of inference (Anderson, this volume, Chap. 6).
Probability sampling is often difficult in a
prospective study, and is even more arduous in a retrospective one. Consider Example A (effect of
landslides on water quality). If we use past data, we
have limited choices for our sampling units. For example, to estimate the water quality (chemical
composition and concentration of various impurities
in the water) 10 years after a landslide, we are limited
to the areas where slides have actually occurred and
are of the appropriate age. These may not necessarily
be representative of the region in which we are interested. For example, if most of the slides were at lower
elevations, or in a particular valley, then the data
from individual slides may have to be weighted to
compensate for their geographical distribution, or a
non-random sample be chosen that will have a more
representative geographic distribution. This last procedure violates the concept of probability sampling
but may represent its best approximation using limited past data.
4.7.2 Awareness and clear statement of assumptions
Often assumptions must be made that are clearly not
true, but the consequences of which are hopefully
minimal. Consider Example I (Pacific Brant). A retrospective component to this study is the use of
bands that were affixed to birds elsewhere.
To simplify the discussion, assume that the brant
came from two wintering areas, area X in which a
large proportion of brant had been banded, and area

49

Y in which only a few had been banded. The average

length of stay of the brant in Parksville-Qualicum
(the first step for estimating the total population
passing through the area) can be obtained from repeated band sightings. However, if brant from
wintering area X tend to spend more time in the area,
our estimate of length of stay will be too high since a
disproportionate number of birds seen with bands
are from area X. Unless we have detailed banding information and know the proportion of birds from
each wintering area, we can do little about this source
of error except to assume that the length of stay for
the X and Y birds is the same, acknowledge the potential bias, and assess its possible effects on the
estimates.
4.7.3 Design considerationscontrolling variation
In any study, it is essential to reduce unwanted variation as much as possible. The more we succeed in
doing this, the better we can succeed in establishing
relationships among variables, testing hypotheses, or
obtaining precise point estimates.
In Example F1 (site index versus tree density),
Goudie (1996) recognized that stands on which site
index were compared must be as similar as possible,
leaving tree density as the only variable. His success
in determining a valid relationship between site index
and tree density depended upon his success in finding nearly identical sites except for the density.
Because he did a retrospective study, the job is more
difficultthere are fewer stands to choose from. In
a prospective study he would have a much larger
choice of sampling unitsa comprehensive sampling
universe from which to sample.
4.7.4 Use of direct measurements rather than proxies for the measurements
Sometimes there are insufficient sampling units with
the properties we want to study. In Example E
(preservation of biodiversity), this was the case. An
assumption was made that a burnt stand with trees
left after a fire is equivalent to a logged stand with
green tree retention. This assumption allowed a
greater choice of stands and thus the potential for
better control in the design. However, in doing this,
the burnt areas are a proxy for green tree retention.
The cost of greater control is the assumption of similarity between burnt and green tree retention stands.

50

4.8 Roles of Retrospective Analysis in Adaptive

Management
Adaptive management is a systematic approach to
improving managerial techniques through learning
from the outcomes of interventions by those involved with the administration of the resource. This
definition implies designing interventions and monitoring programs to provide reliable feedback
concerning the outcomes and their causes. Managers
need all the tools available to properly manage forest
resources, and must be flexible and able to adapt to
quickly changing conditions and priorities.
Retrospective analysis can be used to provide input
in five important ways:
1. Assessing long-term management actions without waiting until the effect of the action is
realized. In Example B (forest bird populations),
assessing long-term trends in bird populations
after clearcutting would require many years. Observing the effect on already clearcut stands can
provide useful and expedient information.
2. Assessing impacts of natural phenomena that
cannot be created for the purpose of a study.
Consider Example A (effects of landslides). Landslides are rarely produced on purpose. In cases
such as in road construction in mountainous terrain where some slides may be created through
blasting, landslides would likely not provide useful
information for a study of water quality. A more
representative sample could be obtained from natural slides.
3. Studying historical patterns of events such as disturbance by fire, fluctuations in weather, or
outbreaks of parasites such as the gypsy moth.
The study in Example C (economic effect of forest
fires) looks at previous patterns of fire and uses
these data to assess the economic impacts of the
fires. This study also would fit in category 2, above,
since fires would not be set simply for estimating
impacts, economic or otherwise. Controlled burns
for preventing larger fires in the future would not
suffice for this type of analysis because they would
not furnish a suitable sample for assessing the impacts of large devastating fires.
4. Collecting background information to aid in the
design of a related study. Example D (effect of
herbicides) demonstrates a retrospective study that

was performed to take advantage of existing data

to aid in the planning of a more suitable study that
will have more focused objectives.
5. Providing interim information for making decisions when results from long-term studies will
not be available until after the intended deadline
has passed. The studies in Example E (preservation of biodiversity) yield considerable information on the relationship between biodiversity and
certain logging practices. This information would
take many years to collect and would be difficult to
implement as a prospective study. Example F1 (site
index versus tree density) falls into this category
too. A prospective study would take years to complete as the database would build very slowly.
In discussing the role of retrospective studies in
adaptive management, their relationship to prospective studies must be considered. In the following
discussion, the two methods will be compared and
contrasted.
Each of the five ways of using retrospective analysis
listed previously can be broken into two components:
1. Feedback for management. The opportunity to
obtain information for improving current management practices often exists. Regular feedback
should not be restricted to retrospective studies.
As prospective studies progress, they offer greater
potential for feedback, but results typically take a
longer time. Wherever possible, prospective
studies should report interim results. If this information is prepared with the proper interpretation,
the manager will be aware of current ideas, and realize that they are subject to change as more
information becomes available. We should not be
reluctant to assess the most current knowledge and
use it to modify management practices.
2. Feedback for understanding. This is just as important as feedback for management. In addition to
knowing that a particular course of action works,
we need to understand why it works. What are the
underlying natural processes and relationships that
cause the results we observe? It is only when we
answer these types of questions that we can generalize our ideas and progress. This type of feedback
will be assimilated into the knowledge base for the
resource, and while it may not be useful for current decisions, it will become useful in the future
when combined with other information.

Feedback for understanding has two components.

The first is direct input to the knowledge base.
Prospective studies are the best way to achieve this as
they are carefully controlled, and statistical accuracy
is more important than short execution time. The
second component is the provision of direction for
future studies. This can often be accomplished by a
retrospective study since only general directions are
required, and time need not be spent obtaining rigorous results.
4.9 Conclusions
Retrospective analysis is an essential tool for research
and management. It allows the researcher to augment
and design studies by relying on previous data or circumstances. Its advantages include great savings of
time and resources since much work has already been
done. In most instances, forest managers cannot afford to wait a long time for results from ideal studies,
and often cannot afford to do ideal studies at all. The
strength of retrospective analysis lies in its ability to
provide an alternative to a purely prospective approach by combining historical data with current
information for the production of interim results.
Retrospective analysis is an important predictor
for the future. Even though quantitative probabilities
(P-values) cannot always be attached to hypotheses, a
qualitative understanding of processes or estimation
of parameters can be obtained. This knowledge can
provide valuable background information and will,
at the very least, provide information for future research.
Finally, we should not depend entirely on retrospective studies. Because of their inherent
weaknesses, their long-term importance lies in providing information concerning future management
and research directions, and pointing out where
more detailed prospective studies are necessary. At
some point we need a quantitative verification of hypotheses with statistically sound results. Where
possible, we must ultimately supplement retrospective studies with prospective ones.
Difficult management decisions of today based on
sparse information should become routine decisions
tomorrow based on solid information. This transition is ongoing, because with each puzzle we solve, a
new one seems to be waiting. Retrospective analysis
can be applied to timely but tentative results, followed by primarily prospective studies, and then to
thorough investigation of a phenomenon. Our activi-

51

ties must be divided between fighting fires and

planning for the future in areas not yet under threat.

Cochran, W.G. 1977. Sampling techniques. 3rd ed. J.

Wiley, New York, N.Y.

References

Cooch, F.G., S. Wendt, G.E.J. Smith, and G. Butler.

1978. The Canada migratory game bird hunting
permit and associated surveys. Can. Wildl. Serv.
Rep. Ser. 43:839.

Acker, S.A., P.S. Muir, G.A. Bradshaw, E. Cazares,

R.P. Griffiths, G.W. Lienkaemper, B. Marks, B.
McCune, A.R. Moldenke, R. Molina, J-L.E.
Peck, J. Smith, B.H. Traut, and E.K. Zenner.
1995. Retrospective studies of the effects of
green tree retention of conifer production and
biodiversity on the Willamette National Forest,
Dep. For. Sci., Oreg. State Univ. Suppl.
Agreem. PNW 920289.
Anderson, J.L. [n.d.]. Errors in inference. This
volume.
Bailar, J.C. and F. Mosteller. 1986. Medical uses of
statistics, NEJM Books, Waltham, Mass.
Bergerud, W.A. and N.J. Reed. [n.d.]. Bayesian statistical methods. This volume.
British Columbia Ministry of Forests and Environment 1995. Biodiversity guidebook. Victoria,
B.C. Forest Practices Code guidebook.
Bunnell, F.L. 1995. Forest-dwelling vertebrate faunas
and natural fire regimes in British Columbia:
patterns and implications for conservation.
Cons. Biol. 9:63644.
Campbell, R.W., N.K. Dawe, I. McTaggart-Cowan,
J.M. Cooper, G.W. Kaiser, and M.C.E. McNall.
1990a. The birds of British Columbia, Vol. I.
Royal B.C. Museum, Victoria, B.C.
______. 1990b. The birds of British Columbia,
Vol. II. Royal B.C. Museum. Victoria, B.C.
Campbell, R.W., N.K. Dawe, I. McTaggart-Cowan,
J.M. Cooper, G.W. Kaiser, M.C.E. McNall, and
G.E.J. Smith. 1997. The birds of British Columbia, Vol. III. Univ. B.C. Press, Vancouver, B.C.
Ciampi, A., S. Franceschi, G. Franchini, J. Ghiffault,
D. Crivellari, S. Tumolo, and G. Bassignano.
1990. Missing information and confounding in
retrospective studies: Statistical methods and an
example of data analysis. Statistica Applicata
2:3751.

52

Goudie, J.W. 1996. The effect of stocking on estimated site index in the Morice, Lakes and
Vanderhoof timber supply areas in central
British Columbia. Presented at N. Int. Veg.
Manage. Ann. Meet. Jan. 2425, 1996,
Smithers, B.C.
Greenberg, R.S. 1988. Retrospective studies (including case-control). Encycl. Statist. Sci. 8:1204.
Gyug, L.W. and S.P. Bennett. 1995. Bird use of
wildlife tree patches 25 years after clearcutting.
Rep. prep. for B.C. Min. Environ. Penticton,
B.C.
Hahn, G.J. 1980. Retrospective studies versus planned
experimentation. Chem. Techn. 10:3723.
Hamilton, A.N., C.A. Bryden, and C.J. Clement. 1991.
Impacts of glyphosate application on grizzly
bear forage production in the coastal western
hemlock zone. For. Can. and B.C. Min. For.
FRDA Rep. No. 165.
Hartman, G.F. and M. Miles. 1995. Evaluation of fish
habitat improvement projects in B.C. and recommendations on the development of
guidelines for future work. B.C. Min. Environ.,
Lands and Parks, Fish. Br., Victoria, B.C.
Hoogstraten, J. and P. Koele. 1988. A method for analyzing retrospective pre-test / post-test designs:
II. Application. Bull. Psychometric Soc.
26:1245.
Hunter, M.L. 1990. Wildlife, forests, and forestry
principles of managing forests for biological diversity. Prentice-Hall, Englewood Cliffs, N.J.
Kelsall, J.P., E.S. Telfer, and T.D. Wright. 1977. The
effects of fires on the ecology of the boreal forest with particular reference to the Canadian
north. Can. Wildl. Serv., Occas. Pap. No. 32.

Koele, P. and J. Hoogstraten. 1988. A method for analyzing retrospective pre-test/post-test designs: I
Theory. Bull. Psychometric Soc. 26:514.

Thrower, J.S. 1992. Height and diameter growth of

Douglas-fir site trees and Larsons Bench. Rep.
to Silv. Br., B.C. Min. For.

Lehmkuhl, J.F., L.F. Ruggiero, and P.A. Hall. 1991.

Landscape-scale patterns of forest fragmentation and wildlife richness and abundance in the
southern Washington Cascade Range. U.S.
Dep. Agric., For. Serv., Pac. NW Res. Sta., Gen.
Tech. Rep. PNW-GTR-285.

Traut, B.H. 1995. Effects of variation in ecosystem

carryover on biodiversity and community
structure of forest floor bryophytes and understory vascular plants: a retrospective approach.
M.Sc. thesis. Dep. For. Sc., Oreg. State Univ.,
Corvallis, Oreg.

McAllister, M.K. and R.M. Peterman. 1992. Experimental design in the management of fisheries: a
review. N. Am. J. Fish. Manage. 12:118.

Wang, M.C. 1992. The analysis of retrospectively ascertained data in the presence of reporting
delays. J. Am. Statist. Assoc. 87:397406.

Marcot, B.G. 1989. Putting data, experience and professional judgment to work in making land
management decisions. In Proc. B.C.U.S. Dep.
Agric. For. Serv. Workshop, Oct. 1620, 1989,
pp. 14061.

Weinberg, C.R., D.D. Baird, and A.S. Rowland. 1993.

Pitfalls inherent in retrospective time-to-event
studies: The example of time to pregnancy. Statistics in Medicine 12:86779.

Maymin, Z. and S. Gutmann. 1992. Testing retrospective hypotheses. Can. J. Statistics 20:33545.
Nemec, A.F.L. [n.d.]. Design of experiments. This
volume.

Zenner, E.K. 1995. Effects of residual trees on growth

of young to mature Douglas-fir and western
hemlock in the western central Oregon Cascades. M.Sc. thesis. Dep. For. Sci., Oreg. State
Univ., Corvallis, Oreg.

Peck, J.E. and B. McClune. 1995. Remnant trees in relation to canopy lichen communities in western
Oregon: a retrospective approach. Dep. Bot.
Plant Path., Oreg. State Univ., Corvallis, Oreg.
Tech. Rep.
Prentice, R. 1976. Use of the logistic model in retrospective studies. Biometrics 32:599606.
Reed, W.J. 1995. Estimating the historic probability of
stand-replacement fire using the age-class distribution of undisturbed forest. For. Sci.
40:10419.
Rosner, B.A. 1986. Fundamentals of biostatistics.
PWS Publishers, Boston, Mass.
Routledge, R., G.E.J. Smith, L. Sun, N. Dawe, E. Nygren, and J. Sedinger. Biometrics [1998].
Estimating the size of a transient population.
Biometrics (in press).
Schwarz, C.J. [n.d.]. Studies of uncontrolled events.
This volume.
Sikkel, D. 1990. Retrospective questions and group
differences. J. Official Statist. 6:16577.

53

5 MEASUREMENTS AND ESTIMATES

RICHARD D. ROUTLEDGE

Abstract

Measurements and estimates are never perfect. The

potential for error is always present, especially in field
studies on large ecosystems such as British Columbias
forests. In the past, inattention to measurement errors
has led to serious management failures, most notably
in the related field of fisheries management. This
chapter describes fundamental measurement concepts, ways to assess and reduce the potential size of
measurement errors, and ways to adjust subsequent
analyses and associated management actions to account for these errors.

5.1 Introduction
Do any of these statements describe your views?
It is a waste of time to worry about measurement
errors. I have enough practice in my field to have
reduced measurement errors to a negligible size.
If I know that my measurements are not perfect,
then I should take several, and average them,
maybe throwing out the odd one that is far from
the others.
I have the resources only to make a subjective guess
at the abundance of some minor species. Surely this
will be adequate. After all, I am only looking for
trends. If the measurement errors are large, and are
consistently present, cant we ignore them when we
are looking for trends?
I dont have to worry about measurement errors. I
always take repeated observations and use standard
statistical techniques to deal with them. If my measurements do contain large errors, then cant I just
take repeated measurements, do a routine statistical
analysis, and quote a P-value to silence the pesky
biometricians?
I have an important job to do. I dont have the time
or luxury of worrying about statistical niceties like
academics and scientists. I need to get on with
managing for forest production.
If you agree with any of these opinions, then you
may find this chapter unsettling.

Adaptive management is a methodology for producing information on the consequences of different

management practices. These consequences must
eventually be measured, and hence measurement is a
key part of any adaptive management project. In general forest management, measurements are taken to
evaluate outcomes of management actions, assess
trends, and evaluate current states of forest ecosystems. This chapter focuses on measurements of
natural resources, such as timber volumes, fish
population parameters, or species diversity of communities affected by forest management. The types of
measurements that are considered range from counting, to direct physical measurements, to educated
guesses.
Careful attention to measurement errors is an essential component of a successful forest management
project. In field work, we often deal with quantities
that are difficult to measure, and the errors in these
measurements are often large. It is tempting to ignore these errors, and manage as though the
estimates reflect the true state of the resource. This
temptation must be resisted as it could lead to ecological, social, and economic disasters. The collapse
of our Atlantic cod (Gadus morhua) stocks can be attributed in part to unwarranted confidence in official
abundance estimates. Here in British Columbia, inadequate attention to measurement and estimation
errors put the Adams River sockeye (Oncorhynchus
nerka) at risk in 1994.
Even when measurement errors are small, as is
often the case in the physical sciences such as astronomy, we eventually push our measurement systems
to their limit when we address increasingly difficult
questions. If we are not vigilant, management expectations can increase beyond the capacity of the
measurement system.
Key components of major projects are often measured by subjective guesses or eyeball estimates.
Although this approach is often the only feasible way
to proceed, such estimates are notoriously vulnerable
to large errors that violate the basic requirements of
statistical analysis. Without realizing it, we may gradually grow to rely on these measurements for more
than they can deliver. Furthermore, statistical analyses typically require specific assumptions about
measurement errors that are often not satisfied by

55

visual estimates. Close attention to these assumptions

will ensure that reliable information necessary for
making management decisions is extracted from the
data.
It is also easy to be fooled into a false sense of the
reliability of ones own measurements. Subtle biases
can infiltrate the best measurement protocols. Any
measurement processes that provide data for key
management decisions need to be checked on an ongoing basis through proper quality control
procedures.
This chapter will discuss two types of measurement errorsrandom and systematicand how
they can be identified and addressed. The assumptions about errors required for standard statistical
analyses and the consequences of violating these assumptions will be examined. Specific issues
associated with special types of measurements will
also be discussed. This chapter focuses on measurement errors themselves. Sampling, the procedure
that defines how, when, and where measurements are
taken, is discussed in Schwarz (this volume, Chap. 3).
5.2. Measurement Errors
5.2.1 Types of errors
Almost all measurements contain errors. The existence of errors does not imply that the worker has
made a mistake. Although careful attention to equipment and protocol may reduce their average size,
errors can never be eliminated entirely. Errors come
in two types: random and systematic.
For example, repeat measurements of the length of
a sawlog will typically differ by small amounts. The
differences will be caused by a myriad of small factors, such as placement of the ends of the measuring
tape, displacement of the middle of the tape by irregularities in the log, or minor errors in reading the
scale. These are random (or chance) errors. If the
tape were nylon and had been stretched substantially
by extended use, then all the measurements would
underestimate the true length of the log. This systematic error or bias often goes undetected.
The following example illustrates the distinction
between chance and systematic errors. It also illustrates the need for care even in simple measures such
as counting.
Aerial surveys have become a popular way of assessing the size of populations of big game animals.
It can be tempting to believe that, from an aircraft,
every individual can be seen and counted. Goddard

56

(1967) conducted one of the early assessments of this

belief. From extensive field work, he ascertained that
69 black rhinoceros (Diceros bicornis) occupied Olduvai Gorge in Tanzania. However, in each of 18
overflights, he never spotted more than 50% of this
population, even under ideal conditions. Under
poorer conditions, he spotted as few as 35% of the
animals. Caughley (1974) reported sighting rates for
other African populations ranging from 23% for
some African mammals to 89% for cattle. Subsequent studies have confirmed these early findings for
a variety of species and habitats.
Goddards survey results were unpredictable.
That is, they were subject to chance errors, attributable to many factors, including type of aircraft,
altitude, speed, visibility conditions, observer alertness, and happenstance. For example, if a rhinoceros
happened to move just as the observer was scanning
that location, it would be more likely to be detected.
Goddards results also consistently underestimated
the true population size. That is, they were subject to
systematic errors. In each survey, some animals were
not seen. Some animals will always go undetected.
They may be hidden under trees, tucked behind
cliffs, camouflaged by underbrush, or simply missed
because the observer was concentrating on another
location, or was fatigued.
The presence of chance errors is relatively easy to
detectrepeat the measurement process several
times under similar conditions. Different results,
with no apparent pattern to the variation, show
chance errors at work. In contrast, systematic errors
cannot be detected by repeat measurements of an unknown value. Had Goddard just taken repeat counts
of a population of unknown size, he could not have
detected the bias. Repeated measurements of a
known quantity are needed to assess the bias.
5.2.2 Gauging the size of the systematic and chance
errors
The bias, or systematic error, in a measurement
process is the difference between the mean of an indefinitely long list of measurements and the true
value of the quantity being measured. It can be
estimated by taking the average of repeated measurements of a known quantity.
Once the bias has been determined, then as long
as it does not change, it will be predictable and can
therefore be eliminated from the results. In contrast,
chance errors are never predictable. At best, we can
make some statement about their distribution by

inspecting a list of measurements of the same quantity. Deviations from the average of this list will
estimate the chance errors. The average size of these
deviations (calculated formally through a root mean
square) is called the standard deviation.
For a list of measurements, x1,x2,x3,xn, the average or mean is given by
n

= x1 + x2 +x3 ++xn = i =1 i ,
x

n
n
and the standard deviation by
SD(x) =

(x )2.
x

n 1

i =1

The standard deviation, SD(x), gauges the average

size of the chance errors (i.e., the average amount by
which an observation deviates from the average.)
The standard deviation tells us absolutely nothing
about the bias. In repeated measurements of a known
quantity,

x (the known quantity being measured)

gauges the bias.
5.2.3 The average of repeated measurements and its
standard error
By taking repeated measurements the size of the
chance errors can be assessed. Averaging these
repeated measurements also allows the researcher/
observer to reduce their impact. By the very nature of
chance errors, some measurements will be positive,
and others negative. In the averaging process, some
of the positive and negative errors will cancel each
other. The remaining chance error in the average will
tend to be somewhat smaller than the errors in individual measurements. The standard deviation of an
average of n independent measurements taken under
similar conditions will be decreased through averaging by the factor . Formally,
n
SD(X)

SD(X ) = ,

where X represents the average of n independently generated values of a variable labelled X.

The standard deviation of an average is often referred to as a standard error. The standard error of a
sample average gauges the average size of the means
fluctuations from sample to sample. The sample

standard deviation (Section 5.2.2) gauges the average

size of the fluctuations of the values within a sample.
In many applications, the sample standard deviation
is used solely to estimate the standard error in the
sample average.
These two quantitiesstandard deviation and
standard errorprovide markedly different information. For example, an adaptive management
experiment of the effects of logging on invertebrate
stream fauna, might investigate the effects on the
abundance of mayfly larvae. The stream bottom will
likely have been sampled. The standard deviation of
the numbers of mayfly larvae per sampling unit describes the inherent variation in mayfly abundance.
Although any effects of forest management practices
on this inherent variation may be important, the primary concern is usually about changes in the mean
or overall abundance. To this end, the sample average
is calculated and its standard error is used to gauge
the average size of the chance errors in this estimate.
Does it always make sense to average all the repeated measurements? No. The positive and negative
chance errors will tend to cancel each other out unless a single, very large error dominates all the others.
Such observations should be singled out for special
attention as described in the following section.
5.2.4 Discarding aberrant measurements
From time to time, a data set will contain one or
more measurements that look aberrant. They stand
apart from the others, and appear as if some gross
error was made in taking them. (Statisticians call
such values outliers in that they lie outside the range
of the rest of the values.) It can be very tempting to
discard these measurements. Not only do they seem
unreasonable, but they also draw attention to possible deficiencies in the measurement process.
Resist the temptation to discard outliers. Instead,
take a thorough look at how they were generated.
Here are the most likely possibilities.
1. They might just result from an obvious error, such
as a misplaced decimal point or inadvertently including counts of other animal species in with the
rhinoceroses in the previous example. Such errors
should be corrected.
2. Outliers might point to a fascinating and important discovery (e.g., that an anomalously large
count of a resident bird population is attributable
to a previously unnoticed transient population).
Such novel insight will usually be worth exploiting.

57

3. They might be due to an inherent part of the natural variability that should not be ignored.
If aberrant values are routinely thrown out, the
resulting data set will give a false impression of the
structure of the system being measured. For example,
in a wildlife habitat study, denning habitat might be
predicted in an area based on the distribution of the
diameter at breast height for a stand. For a stand
dominated by small trees, the diameter measurements associated with the few mature trees could be
considered as outliers. Discarding these outliers
could lead to the conclusion that the stand contains
only small-diameter trees incapable of providing
valuable denning sites. Management errors could
then arise from this erroneous impression.
Furthermore, opportunities for identifying and
controlling large sources of error can be lost, and
clues to new discoveries may go undetected. For example, a few extraordinarily large-diameter trees in a
replanted stand may lead to valuable genetic insight;
a pocket of unusually small ones may betray the arrival of a new insect pest.
Outliers must be treated carefully for another reason: they can invalidate standard statistical inference
procedures (see section 5.3.1, last paragraph).
5.2.5 Accuracy and precision
A measurement process is said to have high accuracy
if all errors are typically small. This requires that both

(a)

the bias and chance errors be small. A measurement

process is said to have high precision if the chance errors are typically small. Measurements that are
precise, but severely biased, are still inaccurate.
Figure 5.1 provides a graphical illustration.
Figure 5.1a depicts a high accuracy situation, with
small bias and small chance errors. This situation is
clearly desirable.
The poor precision in Figure 5.1b can often be improved through careful attention to the sources of
chance errors. Goddards aerial counts, for example,
were conducted on a low budget. He essentially
hitched a ride on an aircraft flying over the area every
chance that came along. At greater cost, he could
have arranged for all flights to be conducted with the
same aircraft, over a standard flight path, at a constant speed, and under similar visibility conditions.
Similar improvements can be achieved in an experimental design context through blocking (see Nemec,
this volume, Chap. 2), and in sampling through stratification (see Schwarz, this volume, Chap. 3).
The situation depicted in Figure 5.1c is insidious.
The high precision creates a false sense of reliability if
the practitioner is unaware of the high bias. Standard
statistical analysis techniques do not confront this
difficulty. They are designed to deal with chance errors only, not bias. For further technical discussion on
accuracy and precision, see Cochran (1977, pp. 1516).
Another measurement problem arises from modern technology. Digital displays are featured in many

(b)

(c)

. Examples of (a) high accuracy (both small bias and small chance errors), (b) low accuracy caused by poor
precision (small bias but large chance errors), and (c) low accuracy but high precision (large bias but small
chance errors).

58

measurement instruments. It is remarkably easy to

forget the inevitable inaccuracy of a measurement, and
to accept on faith the accuracy of every digit in an electronic display. When much of the measurement process
has been automated, we lose reminders of potential inaccuracy. We must treat a digital readout with as much
skepticism as a manual reading of an old-fashioned
vernier scale.
5.3 Standard Statistical Techniques and Measurement Errors
5.3.1 The standard model for measurement errors
In applying standard statistical analyses, measurement errors are assumed to:
1. contain no systematic component;
2. be independent;
3. have a constant standard deviation; and
4. feature a distribution that follows a normal curve.
If any of these model assumptions is violated, then
standard statistical analyses may not work properly.
Furthermore, as discussed in the next section, a standard regression analysis requires that no appreciable
measurement errors be made in the x-variable.
Assumption 1 is critical. Consider, for example, a
confidence interval designed to capture the true volume of timber in an experimentally managed stand
19 times out of 20. By adding and subtracting roughly
two standard errors for the estimate, the formula
compensates for chance errors but not for any systematic component. A large systematic error (e.g.,
generated from a faulty formula for converting diameter and height measurements to volume estimates)
would invalidate the confidence interval.
Systematic errors can be insidious. They can grow
imperceptibly as equipment ages or observers making subjective judgements change their perspectives.
To deal thoroughly with systematic errors, we need
to directly assess the bias. If the bias has been determined to a high degree of accuracy, then its value can
be subtracted from the measurements. But if the bias
estimate is subject to substantial uncertainty, then
this factor must be explicitly taken into account in
any subsequent analyses. It may be possible to put
reasonable bounds on the bias and then manage the
resource conservatively.
Bayesian statistics are also becoming popular
(Taylor et al. 1997; Bergerud and Reed, this volume,
Chap. 7). One might be able to incorporate subjective
assessments of the size of the systematic error into a

formal Bayesian analysis. Nonetheless, the credibility

of this approach will hinge critically on the knowledge base supporting these subjective assessments.
Assumption 2 on independence is also extremely
important. This assumption means that in a sequence of independent errors, the value of the next
error does not depend on the previous ones. In practice, the error in the next measurement can often be
related to the previous one. If, for example, a
measurement of nitrate concentration in soil samples
were to depend noticeably on the kit purchased to
make the measurement, and the next measurement
were to be taken with the same kit as the previous
one, then the errors would be dependent. If the error
in the previous measurement were to be positive,
then the next one would likely be positive as well.
In this example, a set of measurements all taken
with this same kit would show less variability than a
set of measurements each taken with a different kit.
Hence, dependent errors can often lead to a false
sense of precision. They can often be reduced
through randomization (e.g., randomly selecting the
measurement kit for each measurement). However,
such complete randomization is rarely practical.
When other, less complete randomization procedures are followed, the resulting dependent errors
can be handled through more advanced analysis
techniques such as time-series analysis and analysis
of variance for nested or hierarchical designs.
Assumption 3, although often not as critical as the
others, can play a crucial role in prediction intervals
from regression studies. Prediction intervals in a region of unusually small variability will be wider than
necessary, while those in a region of unusually large
variability will fail to contain the predicted value with
the stated probability.
For example, tree height measurements typically
have a nonconstant standard deviation. The standard
deviation increases with height. Furthermore, height
itself will tend to increase with diameter. In a regression of tree height against diameter, prediction
intervals for height would be too wide for smalldiameter trees and too narrow for large-diameter
trees. (Huang et al. 1992 discuss a common solution
to this problem in the context of height-diameter
relationships.)
Assumption 4, that the errors have an approximately normal distribution, is not too critical with
one important exceptionoutliers. Mild departures
from the normal curve may not seriously alter the behaviour of standard inference procedures, especially

59

if many measurements are being averaged. But watch

out for extreme outliers; they can invalidate even the
most basic confidence interval or t-test. This is another important reason to investigate thoroughly any
aberrant-looking measurements (see Section 5.2.4).

relationship. Nigh (1995) proposes using the geometric mean regression line in this context. However,
this technique, promoted by Ricker (1973) in fisheries
analysis, is highly controversial and may itself
provide inaccurate slope estimates. The most appropriate technique will depend upon the specific
application and on the relative sizes of the variation
in the x- and y-directions. See Fuller (1987) for a
thorough discussion of the handling of estimation
errors in the x-variable.
Outliers are also particularly troublesome in regression analyses. Points that are far from the
regression line have considerable influence. Rumours
abound of practitioners routinely discarding such
points. See Section 5.2.4 on discarding aberrant measurements.

5.3.2 Measurement errors in regression analysis

The standard model for regression analysis assumes
errors only in the y-variable. Errors in the x-variable
will tend to disperse the scatterplot in the horizontal
direction. Consequently, not only does the scatter
away from the line increase, but the slope of the regression line also decreases (Figure 5.2). For example,
in the management of mixed-species stands such as
lodgepole pine and white spruce, a researcher might
be interested in establishing a relationship between
the site index of white spruce and the site index of
lodgepole pine. Site index is an estimated value based
on height and age, and is therefore subject to error.
The error could be large enough to substantially
affect the apparent strength of the regression

5.4 Types of Measurements and Associated Issues

Following is a discussion of different types of measurements and typical problems associated with each.

(a) No errors in x
25

20

15
7

10

11

12

13

11

12

13

X
(b) Errors in x
25

20

15
7

10

X
. Errors in the x-values typically not only increase the scatter in the picture, but also spread out the points in the
x-direction. This result reduces the slope of the regression line. In this example, the slope is reduced from 1.95 to
1.10.

60

5.4.1 Counting
Counting is a basic method for assessing the size of a
population. As discussed in the example on rhinoceros surveys, it is easy to overestimate the ability to
count animals. Be prepared for substantial undercounts. Counts of spawning salmon, for example, can
be out by a factor of 5 to 10.
Both over- and undercounting can have serious
management implications, and of course these implications extend to other forms of estimates. In a
silvicultural management experiment, for example,
the numbers of trees in sampled stands may be
counted to estimate stand density. If the stand is too
dense, then by law it must be thinned; if the stand is
understocked, then a remedial measure such as
planting must be taken. In this case, over- and underestimation would incur unnecessary cost to the forest
manager. The uncertainty in an estimate must always
be acknowledged, and quantified where possible
through estimates of bias and standard error, along
with confidence limits if appropriate.
5.4.2 Direct physical measurements
Measurements of simple physical attributes such as
length and mass are usually highly accurate. However, some situations may demand extraordinarily high
accuracy. For example, complex models will require
extremely accurate measurements if measurement
errors are compounded in these models.
5.4.3 Indirect physical measurements
Many measurements of physical quantities are indirect. A liquid thermometer, for example, displays the
height of a column of liquid, and measures temperature only insofar as this height is related to
temperature. The manufacturer must calibrate the
instrument by testing it out at known temperatures.
We cannot safely ignore the manufacturers methodology if we push the instrument close to its limits.
These limits can be exceeded by:
1. demanding more accuracy than can be expected of
the instrument (e.g., trying to measure altitude to
the nearest metre when repeat readings at the
same elevation show chance variation over a range
of 20 m);
2. taking measurements outside of the range of values for which the instrument was designed
(e.g., using an altimeter calibrated for use up to
3000 m in a flight over Mount Waddington, whose
summit is at 4019 m); or
3. using the instrument under unusual conditions

not tested in the calibration process (e.g., using an

altimeter at extremely cold temperatures, at which
it was not tested).
Calibration is not an easy task. The scale on a thermometer has evenly spaced markings, which reflect
the approximate straight-line relationship between
temperature and volume (and, hence, height of the
column of liquid). Many other relationships are
curved. Hence, the calibration process involves fitting
a nonlinear relationship between actual values and
measured responses. Other variables may also influence the relationship. Furthermore, the actual
values will in fact be measurements themselves.
Fuller (1987, pp. 177179) provides a brief introduction to some of the statistical issues associated with
calibration. Also, see the warning in Section 5.3.2
about measurement errors in an x-variable used in
a regression analysis.
5.4.4 Subjective estimates and trend indicators
Informal management schemes often rely on subjective estimates. Many of British Columbias smaller
salmon populations, for example, have traditionally
been subjected to annual, visual surveys from which
subjective estimates were made.
Many people regard annual surveys, such as those
for British Columbia salmon populations, as useful
primarily for monitoring trends. In many instances,
reliable indicators of trend are needed instead of unbiased measures of population size. A downward
trend in an abundance indicator signals the need for
increased protection, perhaps regardless of the actual
size of the population.
However, anyone relying on a trend indicator for
management decisions must be very careful. Coho
salmon (Oncorhynchus kisutch) estimation on British
Columbias Central Coast illustrates two of the problems. Coho are notoriously hard to spot, and
abundance estimates of the smaller stocks have been
very subjective. An observers subjective guess one
year may be strongly influenced by reported estimates from previous years. Hence, a substantial
overestimate in one of the early years might easily
have become propagated through subsequent years.
Such an error would leave an inflated impression of
the historic abundance levels. Attempts to rebuild
stocks to perceived historic levels would be misplaced.
Furthermore, incomplete records over time can
frustrate attempts to determine trendsdetecting a

61

10 000
9 000

Spawning estimates

8 000
7 000
6 000
5 000
4 000
3 000
2 000
1 000
0
1950

1960

1970

1980

1990

2000

Year
. Estimated numbers of chinook salmon spawning in the Upper Fraser Watershed near Tte Jaune Cache.

trend is impossible if no estimates are being collected.

Figure 5.3 illustrates a further problem in using
subjective estimates for assessing trends. This graph
shows the abundance estimates produced by the Department of Fisheries and Oceans for the chinook
salmon (Oncorhynchus tschawytscha) that spawn in
the Upper Fraser Watershed near Tte Jaune Cache.
It looks as if the population has recovered from a
near-calamitous decline in the 1960s. One might also
argue that the recovery gained strength in the late
1970s, and that abundance reached a short-term
plateau around 1980. But the department hired a new
local observer in the late 1970s. He proved to be an
extraordinarily dedicated employee, who developed a
detailed knowledge of the local spawning populations
over his early tenure. We shall never know how much
of the increase in estimated abundance is attributable
to the enhanced knowledge of the resource that this
dedicated observer developed.
A rigid adherence to an unambiguous protocol
would have eliminated this problem. But we need to
foster opportunities for improvement, not stifle them
with an inflexible protocol. Furthermore, salmon
spawning patterns change over time, sometimes
abruptly. The measurement protocol must be free
to adapt to these changes. A suggestion for how this
transition might be achieved follows.

62

Often, the conflicting needs for continuity and

change can be addressed by developing a new protocol while retaining the old one. Continuity can be
maintained by running both protocols in parallel
during a test and phase-in period. This continuity
also provides an opportunity for fine-tuning the new
protocol. Important oversights can be corrected
before the new protocol is used for making management decisions.
Problems with observer bias can also be handled
through a similar strategy. When observers are
changed, the measurement series often changes
abruptly. Using multiple observers in a phase-in
period can provide a series of parallel observations
that can in turn be used to assess the change in observer bias, and provide a training period for the new
observer before his/her results are used for making
key management decisions. If thorough attention is
paid to the measurement protocol, then a regression
or time-series analysis can provide more definitive
evidence of a real trend.
5.4.5 Turning indices into unbiased estimates
In managing fish or waterfowl populations, we can
sometimes reduce reporting biases to obtain reasonably accurate estimates of the numbers killed by
humans. However, we often want to know the

fraction killed. To do this, we need an unbiased estimate of abundance. By contrast, it is often feasible
only to obtain an index of abundance for much of
the population. Indices can be turned into unbiased
estimates through a sort of calibration process.
In resource management work, this goal is often
achieved through double sampling and ratio estimation. The Canadian Wildlife Service, for example,
conducts an annual survey of breeding waterfowl.
Aerial surveys are used to obtain rough abundance
estimates. These figures are then supplemented by
more thorough ground surveys. The results of the
ground surveys are used to adjust the less accurate,
but more extensive, aerial surveys for bias. In a sense,
the ground surveys are used to calibrate the aerial
surveys.
This technique of double sampling is described in
more detail by Cochran (1977, Chapter 12) and
Thompson (1992, Chapter 14). It is a valuable tool in
a wide variety of contexts. For example, in a management experiment involving the monitoring of grass
biomass over time, definitive estimates of biomass
could be obtained only through destructive sampling.
By contrast, extensive but imprecise information
could be obtained from subjective, visual estimates.
By sampling a small number of quadrats destructively, the results of a more extensive set of visual
estimates can be adjusted for bias.
5.4.6 Quantitative measures of imprecise concepts
Forest management is closely linked to ecology,
which in turn is full of vague concepts such as niche
width, niche overlap, similarity, importance, competition, and diversity (Krebs 1994). Developing precise,
quantitative measures of these concepts is one of the
enduring challenges of the subject. The following discussion illustrates both the need for precise,
quantitative measures and common problems encountered in their construction and use. These are
illustrated in the context of diversity measures.
Biodiversity has recently received increasing attention in resource management. Yet it is not easy to
define and measure. The Biodiversity Guidebook in
the Forest Practices Code (B.C. Ministry of Forests
and B.C. Environment 1995) defines the concept as
follows:
Biological diversity (or biodiversity) is the diversity of plants, animals and other living organisms
in all their forms and levels of organization, and
includes the diversity of genes, species and

ecosystems, as well as the evolutionary and functional processes that link them.
This definition explicitly mentions the organisms
and levels of organization to be considered, but
leaves the word diversity undefined. The guidebook gives directions on how to manage forests to
maintain biodiversity. Some of these directions are
based on assumptions on how forest ecosystems
function, including:
The more that managed forests resemble the
forests that were established from natural
disturbances, the greater the probability that
all native species and ecological processes will be
maintained.
Although this assumption has considerable
intuitive and practical appeal, how can we be sure
that it is valid and that the strategy is working? We
need to have some way of quantifying biodiversity so
that we can monitor it directly. This in turn requires
a quantitative definition that pins down this vague
concept.
No single definition will be universally applicable.
A wildlife biologist will be interested in maintaining
wildlife diversity by maintaining the structural diversity in a forest. Thus, two concepts of diversity are
invoked: the species diversity of the wildlife and the
structural diversity of the forest. A fisheries biologist
will focus on maintaining the diversity of individual
fish stocks (not species), which in turn depends on
maintaining a diverse set of healthy fish habitats.
Indices are useful measures of abstract concepts.
However, a single measure may not capture the concept fully, and several different types of indices may
be needed to measure an imprecise concept. Consider an analogy of blind people trying to describe an
elephant. Each person examines a different part of
the elephant. No one person will obtain an accurate
overall impression of the elephant. Obviously, the
more of the elephant that you can include in the operational definition the better, but there will always
be limitations.
Now consider species diversity: its simplest definition is the number of species. But counting or
estimating the number of species in a community is
very difficult. An indefinite number of rare or cryptic
species may go undetected in any survey.
Furthermore, diversity depends not only on
the number of species, but also on the lack of

63

predominance of any one species. Few people would

view a cornfield with a few scattered weeds from
seven other species as being as diverse as an alpine
meadow containing a more even mixture of eight
species. Alfred Russel Wallaces (1895) description of
the diversity in the Amazon rainforest captures the
essence of the concept:
If the traveller notices a particular species and
wishes to find more like it, he may turn his eyes
in vain in any direction. Trees of varied forms,
dimensions, and colours are around him, but he
rarely sees any one of them repeated. Time after
time he goes towards a tree which looks like the
one he seeks, but a closer examination proves it
to be distinct.
Numerous researchers have attempted to develop
a quantitative measure of diversity that would embody such an impressiona measure that both
captures the essence of the concept and is easy to implement. To date, most attention has been focused
on three measures of diversity (Routledge 1979). All
but the first require data on relative abundances of
the species. These measures are defined as follows
(with p1, p2, p3, ..., ps representing the proportion of
the total abundance contributed by species 1, 2, 3, ,s):
1. the species richness,
N0 = s ;
2. an index related to the Shannon-Wiener index
from information theory,

N1 = exp

p 1n(p )
i

i =1

; and

3. an index related to an early proposal by Simpson,

s

N2 =

p2

[ ]

i =1

In each instance, if all species are equally abundant, the diversity reduces to the species count, N0.
It is customary to view N1 and N2 as describing the
number of equally common species that would
produce an equivalent impression of diversity.
Which index should be used? A choice of index
should in general depend on the properties of the
index in relation to the goals of a study. The example
portrayed in Figure 5.4 and Table 5.1 illustrates some
of the differences amongst these three indices.
Figure 5.4 displays the abundance patterns; Table 5.1,
the values of the diversity indices.

64

Note how the species richness is unaffected by the

relative abundances. Of the other two indices, Simpsons index is less sensitive to the rarer species.
Compare, for example, abundance patterns (c) and
(d). As the rarer species diminish in abundance, the
Shannon-Wiener index declines from 2.19 to 1.34
(i.e., by 0.85). By contrast, Simpsons index declines
by only 0.42. If, from (d), the three rarest species
were to go extinct, and the abundances of the other
five were to remain unchanged, then the ShannonWiener index would decline by a further 7% versus
only 2% for Simpsons index. Thus, the ShannonWiener index is more sensitive to the abundance of
the rarer species.
If it is important to focus on the rarer species, then
Simpsons index may be inappropriate. But the sensitivity of the Shannon-Wiener index to the rarer
species comes at a cost. It is usually difficult to assess
the abundance of rare species, and the sensitivity of
the Shannon-Wiener index to these quantities makes
the index very hard to estimate (Routledge 1980). To
date, no generally reliable method has been found to
estimate this index.
When choosing the suite of quantities to be measured, we need to remember the specific goals of the
project. Simpsons index, for example, may play a
useful role in assessing trends in dominance of the
more abundant species, but is inadequate in management projects where the preservation of rarer species
is a priority. If focusing on the rarer species is important, the Shannon-Wiener index would be more
useful if only its bias and standard error could be
more readily predicted. A sensible choice would be to
estimate Simpsons index along with the abundances
of important rare species.
When using an index measure in an analysis, we
must be aware of the range and scale of the index so
as to make proper interpretations. If, for example, a
diversity index were to drop from 2 to 1, should we be
concerned? If the diversity were measured by either
N1 or N2, this could be interpreted as follows: What
was once comparable to a community of two equally
abundant species has dropped essentially to a monoculture. Such a decrease would be clearly noticeable,
and worthy of attention. Had diversity been measured by ln(N1), as is sometimes suggested, then a decline from 2 to 1 would have been hard to interpret.
Choosing an index to quantify an imprecise
concept is tricky. One must pick an aspect of the
elephant that is relatively easy to measure and relevant to the purpose of the study. Overall weight may

 . Diversity measures for the abundance patterns in Figure 5.4

Community
a
b
c
d

Species richness (N0)

8
8
8
8

be hard to measure; however, a simple measure of

height may serve as a useful proxy. Following is a
set of guidelines for selecting and using indices.
5.4.7 Guidelines for selecting and using indices
1. Determine the goals of the experiment.
2. Define, as clearly as possible, the abstract concept
that is to be assessed.
3. Keeping in mind the goals of the experiment, decide what aspect(s) of the underlying concept are
important in terms of the management or scientific objectives.

Shannon-Wiener (N1)
8
4.81
2.19
1.34

Simpsons (N2)
8
4.38
1.53
1.11

4. Identify indices that have been developed to quantify this abstract concept.
5. Assess the sensitivity of the indices to the important aspects identified in step 3.
6. Examine the range and scale of the indices. Is it
easy to tell whether an observed value represents a
desirable level, or whether an observed change in
the index represents an important change in the
achievement of management objectives?
7. Ensure that the bias and standard error of all selected indices are well understood and predictable.
Each index shoul not be overly sensitive to small

(a) Even distribution

(b) Four dominant species

(c) One dominant species

(d) A virtual monoculture

. Four dominance patterns. In each instance, eight species are present, but the communities are increasingly
dominated by fewer species.

65

measurement errors. Large, unquantifiable biases

are particularly troublesome and can easily arise if
the index depends on complex mathematical functions of the measured values, but can also arise in
simple indices, such as the species richness.
5.5 Reducing Measurement Errors
Although some measurement errors are inevitable,
they can often be reduced substantially. The resulting
benefits to the experiment can be considerable. Following are some guidelines for achieving this goal.
5.5.1 Guidelines for reducing measurement errors
1. Counting
Ensure that new personnel are trained by experienced workers.
Where feasible, mark or discard items previously counted to reduce double-counting.
Anticipate undercounting. Try to assess its extent by taking counts of populations of known
size.
Try to reduce errors by taking counts only in
favourable conditions and by implementing a
rigorous protocol.
2. Physical measurements
Instruments should be calibrated before first
use, and periodically thereafter.
Personnel should be trained in the use of all
measuring devices.
Experienced personnel, as part of an overall
quality control program, should spotcheck measurements, particularly those taken by new
personnel.
Incorporate new equipment where appropriate
(e.g., lasers and ultrasound, for distance measurements).
3. Remeasurement
Watch for the transfer of errors from previous
measurements (e.g., a mistaken birth from an
item erroneously marked as dead).
Reduce errors in relocating the site of previous
measurements through more careful marking,
use of modern electronic GPS technology, etc.
Ensure that bias is not propagated through the
use of previous measurements as guides to subsequent ones. (This issue is particularly
troublesome in subjective estimates.)

66

4. Visual estimates
Ensure that all visual estimates are conducted
according to rigorous protocols by well-trained
observers.
Pay particular attention to observer bias. When
bringing a new observer into the program, ensure that his/her results are backed up by an
experienced observers.
If sites or times are to be selected as part of the
collection of visual estimates, eliminate selection bias by providing a protocol for site- or
time-selection. Do not, for example, let vegetation samplers pick modal sites.
5. Data handling
Record data directly into electronic form where
possible.
Back up all data frequently.
Use electronic data screening programs to
search for aberrant measurements that might be
due to a data handling error.
Design any manual data-recording forms and
electronic data-entry interfaces to minimize
data-entry errors. In the forms, include a field
for comments, encourage its use, and ensure
that the comments are not lost or ignored.
5.6 Summary
A century ago, British Columbias renewable resources seemed so limitless that we asked very little of
our measurements of the resources and their support
systems. With new requirements imposed (e.g., by
the Forest Practices Code) and with increased harvesting capacity, we are escalating our demands on
the measurement systems. In recent years, our systems for estimating fish populations have let us
down. The recent controversy over the management
of Fraser River sockeye (Fraser et al. 1995), has not
been so much about a breakdown in the quality of
the measurement procedures, as about the fact that
our management expectations have increased beyond
the capacity of the measurement system.
Assess continually the adequacy of a measurement
system to improve it where possible and to point out
when its limitations may be exceeded. The assessment should include:
the choice of quantities to be measured;
the procedures and equipment for taking the measurements;
any associated sampling;

 the processing, storage, and analysis of the data;

and
the demands and expectations of the resource
managers or others who use the results.
Good quality control procedures are an essential
component of any measurement process. This applies
even if the measurements are to be used solely for indicating trends. A well-designed protocol must be
followed, and the reliability of the data must be commensurate with management needs.
Just as expectations can rise without notice, so can
a gradual deterioration in quality go undetected.
Consider, for example, the monitoring of spawning
habitat in Devoe Creek far up the North Arm of
Quesnel Lake. The creek enters the lake through an
old western redcedar (Thuja plicata) grove, containing many downed trees and rampant with devils
club (Oplopanax horridus). Workers faced with a rising wind on the long stretch of water back to the
landing would be tempted to cut corners if they believed that no one valued their work or would ever
check on their accuracy. For field measurements,
often taken in isolated conditions, quality checks will
inevitably be infrequent. The quality of the data will
depend critically on the reliability and commitment of
the field staff; sound, quality management practices will
foster the required spirit.
The myths presented at the beginning of the chapter should now have been dispelled.
It is a waste of time to worry about measurement
errors. I have enough practice in my field to have
reduced measurement errors to a negligible size.
Measurement errors in many field studies are
large, and inadequate attention to them has led to
major management disasters.
If I know that my measurements are not perfect,
then I should take several, and average them,
maybe throwing out the odd one that is far from
the others.
Taking repeated measurements allows the researcher to assess the average size of the chance errors. Averaging these measurements will usually
reduce the impact of the chance errors. However,
aberrant measurements should be singled out for
special attention, not casually or routinely discarded.
They could provide valuable insight, and are an important part of the information collected. In addition, averaging will not reduce any systematic bias.

I have the resources only to make a subjective guess

at the abundance of some minor species. Surely this
will be adequate. After all, I am only looking for
trends. If the measurement errors are large, and are
consistently present, cant we ignore them when we
are looking for trends?
We need to know enough about the errors to be
able to distinguish between a trend in the quantity
being measured and in the measurement errors. Furthermore, a false estimate of the historical state of the
forests could lead to inappropriate management actions.
Trend indicators are often set up when it seems
too difficult or costly to implement the rigorous procedures required to produce unbiased abundance
estimates. Trend indicators demand almost as much
rigour, and measurement procedures must be rigorous enough to rule out any cause for a trend other
than a change in abundance.
I dont have to worry about measurement errors.
I always take repeated observations and use standard statistical techniques to deal with them. If my
measurements do contain large chance errors, then
cant I just take repeated measurements, do a routine statistical analysis, and quote a p-value to
silence the pesky biometricians?
The standard statistical analysis procedures require
specific assumptions about the measurement errors.
Violated assumptions lead to questionable analyses
and management decisions.
I have an important job to do. I dont have the
time or luxury of worrying about statistical niceties
like academics and scientists. I need to get on with
managing for forest production.
Adaptive management of British Columbias
forests is an important task. Thorough attention to
measurement errors and other statistical niceties will
help, not hinder, the ongoing development of improved management strategies.
5.6.1 Guidelines for developing a measurement
protocol
Following are recommended guidelines for improving the quality and value of measurements. As with
any general guidelines, these will need to be adapted
to specific applications, and are intended more as an
initial checklist of important items to consider than
as an inflexible set of rules.

67

1. Determine what parameter needs to be measured. In

so doing, pay attention both to the relevance of the
quantity being measured and the practical difficulties
in obtaining reliable measurements or estimates.
Consider both the concept and its measurement.
2. Determine the demands that need to be placed on the
quality of the resulting measurements or estimates.
3. Devise a measurement system that will meet these
demands. If this task is impossible, reassess the management objectives and strategies and revisit
guidelines 1 and 2.
4. Assess the accuracy of the proposed measurement system by taking repeated measurements of known
quantities under a variety of conditions.
5. Establish an unambiguous protocol for taking measurements, and ensure its proper implementation.
6. Implement a system of periodic checks on the continuing performance of the measurement system in light
of internal changes and external demands. Watch for
subtle increases in the demands placed on the system.

Huang, S., S.J. Titus, and D.P. Wiens. 1992. Comparison of nonlinear height-diameter functions for
major Alberta tree species. Can. J. For. Res.
22:1297304.
Krebs, C.J. 1994. Ecology: The experimental analysis
of distribution and abundance. 4th ed. Harper
and Collins, New York, N.Y.
Nemec, A.F.L. [n.d.]. Design of experiments. This
volume.
Nigh, G.D. 1995. The geometric mean regression line:
a method for developing site index conversion
equations for species in mixed stands. For. Sci.
41:8498.
Ricker, W.E. 1973. Linear regressions in fisheries research. J. Fish. Res. Board Can. 30:40934.
Routledge, R.D. 1979. Diversity indices: Which ones
are admissible? J. Theoret. Biol. 76:50315.
______. 1980. Bias in estimating the diversity of large,
uncensused communities. Ecology 61:27681.

7. Look for ways to develop incremental improvements

while maintaining the integrity of any long-range
data series. When implementing changes, phase them
in, running new and old methods in parallel during
a transition period.

Schwarz, C.J. [n.d.]. Studies of uncontrolled events.

This volume.

References

Thompson, S.K. 1992. Sampling. Wiley-Interscience,

New York, N.Y.

Bergerud, W.A. and W.J. Reed. [n.d.]. Bayesian statistical methods. This volume.
British Columbia Ministry of Forests and B.C. Environment. 1995. Biodiversity guidebook.
Victoria, B.C. Forest Practices Code guidebook.
Caughley, G. 1974. Bias in aerial survey. J. Wildl.
Manage. 36:13540.
Cochran, W.G. 1977. Sampling techniques. 3rd ed.
J. Wiley, New York, N.Y.
Fuller, W.A. 1987. Measurement error models.
J. Wiley, New York, N.Y.
Goddard, J. 1967. The validity of censusing black rhinoceros populations from the air. East Afr.
Wildl. J. 5:1823.

68

Taylor, B., L. Kremsater, and R. Ellis. 1997. Adaptive

management of forests in British Columbia.
B.C. Min. For., For. Practices Br., Victoria, B.C.

Wallace, A.R. 1895. A narrative of travels on the Amazon and Rio Negro: with an account on the
native tribes, and observations on the climate,
geology, and natural history of the Amazon
Valley. Greenwood Press, New York, N.Y.
Reprinted 1969.

6 ERRORS OF INFERENCE
JUDITH L. ANDERSON
Abstract

Occasional erroneous conclusions (errors of inference)

are unavoidable in the analysis of results from management experiments and monitoring programs.
However, their probability of occurrence in a given experiment can be controlled. In an experiment comparing two treatments, conclusions can be incorrect in
two ways: (1) concluding that a difference between
the treatments is real when in fact it is not (a Type I
error), or (2) concluding that there is no difference between treatments when in fact a difference exists (a
Type II error). Both types of error can be costly in
typical adaptive management experiments, where
treatments involve the effects of commercial activities,
such as harvesting, on ecosystems. A Type I error may
lead to unnecessary limitations on commercial activities, while a Type II error may result in the continuation of activities damaging to the ecosystem. Type I
error is limited by the conventional significance level
of statistical tests to a frequency of less than five errors
per 100 tests performed. The method for estimating
and limiting Type II error rate (statistical power analysis) is less well known but just as important. This
chapter discusses conceptual and practical aspects of
statistical power analysis (including references and
software that aid in performing power analysis) and its
role in the design of large-scale experiments in forest
management.

factors that influence the probability of a Type II

error;
strategies by which those factors can be manipulated in experimental design to control error rates in
ecological studies; and
software and literature dealing with theoretical and
practical aspects of Type II error and its management.
6.2 What Are Type I and Type II Errors, and How
Do They Fit into Statistical Inference?
6.2.1 Statistical inference is an important part of
the process of evaluating a scientific hypothesis
Statistical inference is a form of reasoning that leads
to rational conclusions about states of nature when
the available information comes from a sample of the
population or system under study (Kirk 1982).
Figure 6.1 illustrates the position of statistical inference in the iterative process of evaluating a scientific
hypothesis. Usually the first step is the statement of
the hypothesis as a testable proposition that is tentatively adopted as an explanation for observed facts
and as a guide for investigation (Kirk 1982). For example, an ecologist studying the effect of logging
practices on Coeur dAlene salamanders near S4

Scientific
hypothesis

Deductive
logic

Statistical
hypotheses:
H0 , HA

6.1 Introduction: Type I and Type II errors

@@
?h@@
@@
?h@@
@@
?h@@
@@
?h@@
@@
?h@@
@@
@@
@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@?h@@
?e@@@@@@
@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@?e@@@@@@@@
@@
@@
@@
@@
@@
@@
@@
@@

@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@

@@
@@
@@
@@
@@
@@
@@
@@

@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@

Development of
measurement
and sampling
procedures

@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@

@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@

Inductive
logic

@@
@@
@@
@@
@@
@@
@@
@@

@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@
@@

@@@@@@@@? @@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@? @@@@@@@@

@@@@@@@@? @@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@?e@@@@@@@@e?@@@@@@@@? @@@@@@@@
@@?
g@@
@@?
g@@
@@
@@?
g@@
@@?
g@@
@@?
g@@
@@?
g

A major goal in adaptive and experimental management is to improve our understanding of managed
biological systems by making reliable conclusions
(inferences) from experiments and monitoring programs. However, any experimental inference has a
chance of being incorrect, and these errors can result
in large economic and ecological costs. Therefore, experimenters must understand how errors of
inference occur and how to control them. This chapter discusses the following topics:
the relationship between the two types of errors of
inference in statistical tests, with a focus on the
category of error most often ignoredType II
error, failure to reject a null hypothesis when it is
in fact false;

Statistical
inference:
Reject H0?

Data collection
and parameter
estimation

. The relationship between scientific and statistical

hypotheses (adapted from Kirk 1982). HO
denotes the null hypothesis, HA denotes the
alternative hypothesis.

69

streams in the Kootenay Valley might begin with

the scientific hypothesis: The 10 m wide riparian
reserve zone recommended by the Forest Practices
Code (B.C. Ministry of Forests and B.C. Environment 1995) is sufficient to maintain salamander populations as a component of vertebrate biodiversity.
Next, deductive logic is used to generate a statistical hypothesis that follows from the scientific
hypothesis. This statistical hypothesis is a quantitative statement about some variable whose sampling
distribution can be described statistically. A statistical
hypothesis that proceeds logically from the salamander example might then be as follows: Population
density of salamanders is not reduced in experimental 10 m wide riparian reserves on S4 streams,
compared with the density of control populations
in comparable habitat in unlogged watersheds.
After identifying the experimental units, the experimenter decides on the statistical parameter to be
estimated and develops appropriate measurement
and sampling procedures. In the salamander example, the experimental units would be watersheds and
the estimated parameter would be mean population
density across sampled watersheds in two treatment
groups: an experimental group of logged watersheds
with riparian reserves and a control group of unlogged watersheds. Further refinement would
transform the statistical hypothesis into mutually exclusive null and alternative hypotheses, denoted HO
and HA, respectively. Examples might be HO: mean
population density in the experimental riparian reserves (E) mean control population density (C),
and HA: E < C.
Statistical inference provides a logical structure for
drawing conclusions about whether the null hypothesis can be rejected. The basis for this decision is the
statistical test, which uses a statistical model to assess
the probability that the observed data could have

come from a population of experimental units for

which the null hypothesis is true. If the probability is
sufficiently low, we reject the null hypothesis HO and,
by implication, provisionally accept the alternative
hypothesis HA. This decision inductively influences
our degree of belief in the scientific hypothesis under
consideration. Typically, in basic science, several iterations of this process are required before consensus
is reached about the plausibility of the scientific hypothesis.
In the salamander example, if the mean of population density measurements in watersheds with
riparian reserves were significantly lower than that of
controls, statistical inference would reject the null
hypothesis. This experiment alone would provide
one line of evidence about the scientific hypothesis,
suggesting that the recommended riparian reserves
may not be sufficient to preserve the salamander
population as an element of biodiversity. If making a
decision about the adequacy of this management recommendation was not urgent, further studies would
be desirable to provide a more thorough test of the
scientific hypothesis.
6.2.2 Statistical inference can lead to any one of
four outcomes, two of which are incorrect
The four possible outcomes of a statistical hypothesis
test are shown in Table 6.1. If the null hypothesis is
really true, then two outcomes are possible: Not rejecting HO is a correct inference, while rejecting it
constitutes a Type I error. Similarly, if HO is really
false, the correct inference is to reject it, and failing to
do so constitutes a Type II error. The probability of
committing a Type I error is , while the probability
of a Type II error is . The power of the test (1) is
the probability of correctly rejecting HO when it is really false.
In the salamander example, a Type I error would

. Four possible outcomes of a statistical test of a null hypothesis. The probability of each outcome is given in
parentheses. Management decisions that might proceed from the inference are indicated in parentheses. Adapted
from Toft and Shea (1983).

Inference
State of nature

Do not reject HO
(Manage as though HO were true)

Reject HO
(Manage as though HA were true)

HO true

Correct (1):
Correctly infer that no
treatment effect exists

Type I error ():

Infer that treatment effect exists
when in fact there is none

HO false

Type II error ():

Infer that no treatment effect
exists when in fact there is one

Correct (power = 1):

Correctly infer that
treatment effect exists

70

be committed if the experimenter concluded that the

density of experimental populations had decreased
when in fact they were really doing at least as well as
the controls. If management decisions followed from
this erroneous inference, the forest industry might
unnecessarily be required to leave a wider reserve
along S4 streams.
On the other hand, a Type II error would be committed if the experimental population had really
declined relative to controls, but the experiment
failed to detect the decline. In this case, managers
would conclude that 10-m riparian reserve zones
were sufficient to maintain the species. They might
implement this level of riparian protection on a large
scale, without realizing that the practice, in fact,
failed to meet the objective of maintaining the salamander population in the community.
6.2.3 For a complete analysis of possible errors of
inference, the alternative hypothesis must indicate
a treatment effect worth detecting
In statistical tests, the data actually observed in an experiment are compared with the data that could be
observed if the population under study had a particular assumed value for one of its statistical parameters.
To represent that assumed value, a null hypothesis
must always include an equality relationship (often
called a point null hypothesis). It may include an
inequality as well if treatment effects are expected to
occur in one direction. Thus, HO: 1 2 and HO: 1= 2
are valid null hypotheses, but HO: 1< 2 is not.
In the salamander example, only the equality relationship of the null hypothesis (HO: E C) is
actually used in the statistical test. The experimenter
calculates the observed value of the t-statistic

xC
xE
(xE )/s, where is the mean of population
density estimates in experimental (logged with 10-m

riparian reserve zones) watersheds; xC is the mean of

population density estimates in control (unlogged)
watersheds; and s is the standard error of the difference between the sample means. This observed value
of t is compared with the distribution of estimates of
t that could arise if the salamander population densities really were equal in both the experimental and
control watersheds. HO is rejected if the calculated t
value is extreme, such that the probability that it
could come from that t distribution is less than the
critical (the maximum acceptable rate of Type I
error, set by convention to 0.05). In other words, if
HO were really true and the experiment were repeated
many times, the null hypothesis would be rejected

(incorrectly) in at most 5% of the replicate experiments.

Because traditional practice has focused only on
rejection of the null hypothesis, scientists have usually been content to use alternative hypotheses with no
equality relationship, such as HA: 1 < 2. However,
to analyze the probability of Type II error, a statistical
parameter based on the data must also be compared
to a distribution that could occur if the alternative
hypothesis were true; this comparison requires that
HA includes an equality. The difference between the
equality relationships in HO and HA expresses the
minimum change the experimenter is interested in
detecting reliably (i.e., with a probability of 1). In
the salamander example, if the experimenter decides
that it is important to detect a 10% reduction in population density, she would state the alternative
hypothesis as HA: E 0.9 C.
The requirement that HA be bounded by an equality forces us to recognize that statistical significance is
not the same thing as biological significance. Given
a large enough sample and enough decimal places,
even the most trivial differences between means can
be declared statistically significant. To avoid this outcome, the experimenter in the salamander example
has clearly chosen what differences are important to
detectdeclines in population density of 10% or
more. This difference, the distance between the
equality relationships of the two hypotheses, is called
the effect size (see Section 6.5). For differences less
than the biologically significant effect size, the experimenter is willing to provisionally accept HO when it
is actually false. Having made this decision, she can
design the experiment without wasting resources on
the detection of biologically trivial differences.
6.3 Why Are Type I and Type II Errors Important
in Adaptive Management?
6.3.1 In applied science and management, statistical inference often leads directly to decisions
Progress in basic science has traditionally depended
upon a conservative approach to the acceptance of
new hypotheses. This approach is used because
waiting for evidence to accumulate has little disadvantage; a true scientific hypothesis will surely be
discerned by the scientific community eventually
after repeated iterations of the process in Figure 6.1.
On the other hand, erroneously adopting a
hypothesis that later proves to be false can waste time
and effort. Therefore, scientists doing basic research

71

try to minimize the probability of Type I error by

keeping small, without much explicit concern for
the magnitude of . In basic science, failure to reject
the statistical hypothesis HO does not imply the immediate acceptance of the scientific hypothesis from
which it is derived, but rather a further suspension
of judgement about it. A consensus to accept a
scientific hypothesis is achieved only after repeatedly
placing it at risk by testing its predictions (Platt 1964).
In contrast, applied scientists and resource managers must be concerned with both types of error.
Decision-makers pressured by time constraints often
must make judgements before adequate evidence is
accumulated; sometimes they have only a single statistical test of the scientific hypothesis in question
(Peterman 1990a). Thus, failure to reject a statistical
null hypothesis may result in the de facto acceptance
of the null scientific hypothesis. Management actions
proceed as though the null hypothesis were indeed
true, despite the finite probability that this conclusion might be a Type II error (Table 6.1).
6.3.2 Actions based only on either type of error
can be costly
As in other resource management contexts, experiments or monitoring programs in managed forest
ecosystems frequently involve tests of hypotheses
about the effect of human activity on the ecosystem.
The statistical tests used to evaluate these hypotheses
are usually structured on null hypotheses of no effect
on ecosystem variables; for example, HO: Water
quality in streams within 100 m of a logging road is
at least as good as that in watersheds without roads.
With this null hypothesis, a Type I error occurs if
managers conclude that water quality is reduced
when in fact it is not. The costs of Type I errors usually are borne by those interested in exploiting the
resource. For example, logging companies suffer financially if their roadbuilding activities are curtailed
due to an incorrect rejection of HO.
However, the costs of a Type II error may be even
greater, and they often affect the ecosystem, public
health, or the public purse. With a Type II error, managers would incorrectly conclude that roadbuilding
does not affect water quality. The consequences
would include both the effects of reduced water quality on humans and biota, and the costs of correcting
the damage. Thus, in contrast to the relaxed approach to Type II error common in basic science,
applied experimentation and management require
attention to both types of error (Parkhurst 1990; Peterman 1990a, 1990b).

72

6.3.3. Decision-makers need quantitative information about probabilities of errors of inference

Management of forest ecosystems is inherently uncertain, and the costs associated with both types of
errors may be high. Decision-makers must quantify
uncertainties and their associated costs, including
them explicitly in their plans (Morgan and Henrion
1990). Decision analysis (Raiffa 1968), an effective
method for analyzing alternative management actions in the face of uncertainty, requires estimates of
the probability that both null and alternative hypotheses are actually true (i.e., probabilities for the
states of nature in Table 6.1). These probabilities can
be calculated using Bayesian methods of statistical inference (Bergerud and Reed, this volume, Chap. 7;
Peterman and Peters, this volume, Chap. 8). While it
is not possible to calculate the probability that a hypothesis is true using classical statistical tests, the
probability of incurring either a Type I or a Type II
error can be controlled to acceptable levels through
careful experimental design and statistical power
analysis.
6.4 What Factors Limit Rates of Type I and
Type II Error?
6.4.1 Experimenters can control Type I error rate
by choosing the critical level of
The traditional moment of truth in statistical procedures is the significance test. The P-value calculated
in most familiar statistical tests indicates the probability of obtaining a test statistic at least as extreme as
the one calculated from the data, if HO were true. The
significance level is a critical value of the maximum probability of Type I error (rejecting HO when
it is true) that the scientist is willing to tolerate. Thus,
when a P-value is less than 0.05 (the usual critical
value of ), the experimenter rejects the null hypothesis with the guarantee that the chance is less than 1 in
20 that a true null hypothesis has been rejected.
This guarantee about the probability of making a
Type I error implicit in significance testing is valid
only if the assumptions of the test are met. Otherwise, it can be difficult to determine the actual
probability of Type I error. The assumptions of various significance tests are the subject matter of most
practical reference books on biometrics and will not
be discussed here.
The significance test analyzes only the case in
which HO is really true (the top row of Table 6.1), and
thus tells only half the story. If in fact HO is not true,

Type II error becomes the concern (bottom row of

Table 6.1). The significance test alone provides no information about its probability.
6.4.2 Experimenters can control the Type II error
rate by planning for a suitable level of statistical
power
While the significance test limits the rate of Type I
error, it is equally important to control the probability of Type II error. To achieve this, power analysis is
used to estimate , the probability of Type II error,
and its complement, statistical power (1), the
probability of detecting a specified treatment effect
when it is present (Cohen 1992).
Statistical power is a function of several variables:
sample size, N;
variance of the observed quantities, s 2;
effect size (the treatment effect the experimenter
wants to be able to detect); and
(the maximum rate of Type I error tolerated).
In general, given any four of these five variables
(power, sample size, variance, effect size, and ),
the fifth can be solved for.
To understand how each of these variables influences power, consider a study designed to address the
question, Is this area good breeding habitat for
wood ducks? The experimenter knows from other
studies that the ducks exhibit the strongest preference for nest cavities with a diameter of 25 cm and
that the standard deviation for diameter of cavities is
10.0 cm in the mixture of tree species present. He decides to sample 20 cavities and to analyze the sample
mean using a z-test, with the statistical hypotheses
HO: = 25 and HA: < 25.
This statement of the statistical hypotheses is not
adequate, because statistical power analysis requires
that HA must be bounded by an equality relationship,
as discussed in Section 6.2.3. In this case, the experimenter suspects that the ducks will not use cavities
with a diameter less than 22 cm, so he would like to
detect reliably a difference in mean diameter of 3
cmthe biologically significant effect size. Thus, for
power analysis, the hypotheses are HO: = 25 and
HA: 22, with the implicit provision that the experimenter is not concerned about rejecting HO if the
mean cavity diameter lies between 22 and 25 cm.
Figure 6.2a shows the expected sampling distribution, under HO and HA, for estimates of mean
diameter of cavity based on samples of 20 cavities.
The light shaded area under the left tail of the O

distribution is 5% of the total area under the curve

(below 21.4 cm). If the sample mean falls in this region, the null hypothesis will be rejected. If HO is true
(i.e., the true mean is indeed 25 cm), yet the sample
mean falls in this tail, a Type I error will occur.
The A distribution represents the sampling distribution of estimates of the mean diameter if the
equality relationship of HA is true (i.e., = 22). In
this case, observing a sample mean under 21.4 cm will
lead to the correct inference (that HO is false). Statistical power, the probability of correctly rejecting the
null hypothesis, is represented by the unshaded area
under the A curve (1). The dark area under the
right portion of the A curve represents , the probability of committing Type II error. If the true mean
cavity size is in fact 22 cm, yet the observed sample
mean is greater than 21.4 cm, the null hypothesis
( = 25) will not be rejected and a Type II error will
occur. In this example, , the probability of Type I
error, is constrained to 0.05, while , the probability
of Type II error, is 0.60 and statistical power (1) is
0.40. Hence, the experimenter has only a 40% chance
of reliably identifying inadequate breeding habitat for
wood ducks.
What can the experimenter do to increase the
power of this test? One possibility is to increase sample size. Figure 6.2b shows the case in which sample
size has been doubled to 40. Because increasing N reduces the standard error of the mean (s/ the
N),
distributions under HA and HO become narrower.
With a narrower distribution, the 5% cutoff point for
hypothesis testing increases to 22.5 cm. Therefore, the
dark area under the A curve representing is now
reduced from 0.60 to 0.38 of the total area. In this example, therefore, doubling the sample size has
increased statistical power from 0.40 to 0.62.
Increasing the effect size of interest also increases
power. Suppose the experimenter believes that wood
ducks will use cavities with diameters as small as
19 cm. He would like to reject HO reliably if the mean
is 19 cm, but is content to accept HO incorrectly if
the mean is between 19 and 25 cm . Figure 6.2c shows
the sampling distributions for an effect size of 6 cm,
twice as large as that in Figure 6.2a. In this case, the
overlap between the two distributions of estimates of
the mean is reduced because O and A are farther
apart. The dark area representing is now 0.14 of the
total area, so statistical power is 0.86. The experimenter has an 86% chance of correctly identifying
unsuitable wood duck habitat, more than double the
power of the design in the basic scenario.

73

0.4

(a)

A
(b)

(d)

Probability

0.4

0.4

(c)

(e)

A
1

0
15

20

25

30

15

20

25

30

Nest cavity diameter (cm)

. Variables influencing power to detect the difference between a sample mean and a constant for the wood duck
nest cavity example. For the basic scenario (a), HO: =25, HA: 22; desired effect size for detection = 3, = 10,
= 0.05, N = 20. Figures (b)(e) are calculated from the basic scenario with one change in each. In (b), N = 40;
in (c), effect size = 6; in (d), = 5; in (e), = 0.1. The light shaded area under the HO curve represents , the
probability of a Type I error. The dark shaded area under the HA curve represents , the probability of a Type II
error. The unshaded area under the HA curve represents statistical power (1). Each of the variations in (b)(e)
increases power relative to the basic scenario (a).

74

Decreasing the variance of the observed variable

(by improving measurement precision or by various
techniques of experimental design, such as blocking)
also increases power by reducing the standard error
of the estimated means. Figure 6.2d represents the
wood duck observations as in the basic scenario, with
the exception that the experimenter has decided to
measure only those cavities in the species of tree most
favoured by the ducks. The SD is known to be 5 cm
in that species, half that in the basic scenario. Here
again, the dark area under the A curve representing
the probability of Type II error is reduced to 0.14, so
statistical power is 0.86.
Finally, increasing the probability of making a
Type I error will increase power. In Figure 6.2e, the
critical level (the grey tail under the O distribution) is doubled to 0.10. Consequently the range of
estimated means for which HO is correctly rejected
increases, resulting in a modest increase of power to
0.52. This situation requires a trade-off: the experimenter is willing to run a 10% chance of incorrectly
labelling good habitats as unsuitable for wood ducks
to produce a 52% chance of correctly identifying unfavourable habitats.
6.5 The Concept of Effect Size is Complex
6.5.1 How are biologically significant, detectable,
and standardized effect sizes applied?
In Section 6.2.3, effect size was introduced as the difference between the equality components of the null
and alternative hypotheses, usually chosen to represent a biologically significant difference. In the wood
duck example, the biologically significant effect size of
interest was easily understood as the difference between the minimum acceptable nest cavity diameter
of 22 cm and the preferred size of 25 cm. This concept
is most useful when the experimenter uses the relationships involving power and sample size to address
a question such as, To have an 80% chance of detecting a treatment effect at least as large as the
biologically significant one, what sample size should I
use?
On the other hand, sometimes the sampling program is already set, and the experimenter would like
to know how large an effect size could be detected
with, say, 80% reliability. In that case, detectable effect
size is the unknown to be calculated, rather than an
input determined by biological processes or the experimenters choice. See Section 6.6.2 for more
discussion of detectable effect size.

These concepts of effect size require that all the

treatment effects of the experiment be summarized in
a single parameter. This approach is straightforward
for some designs. In the wood duck example, the biologically significant effect size was the difference
between a population mean and a known constant
(the preferred cavity size). Similarly, when comparing two populations means or two correlation
coefficients, the estimate of effect size is simply the
difference between the two values. However, formulas for effect size become more complex in designs
that involve many relationships among statistical parameters, such as multiple regression or analysis of
variance.
In addition, to facilitate calculation of power and
comparison between experiments, formulas for effect
size are usually presented in a standardized form, including measures of variance as well as summaries of
the magnitude of treatment effects. For example, the
difference between two means is expressed as a standardized effect size by dividing by the standard
deviation: (12)/, where 1, and 2 indicate the
true population means and indicates the (common) standard deviation of the populations (Cohen
1988). Standardized effect size has several advantages.
It combines into a single parameter two of the four
variables that influence powereffect size and variance. Because it has no units, standardized effect size
allows comparisons to be made among different experiments.
6.5.2. Arbitrary effect sizes can also be useful
Ideally the effect size to be detected should be biologically significant, but in many cases this value
cannot be expressed quantitatively due to lack of information. To assist in determining sample size in
these cases, Cohen (1988) identifies representative
small, medium, and large standardized effect
sizes, based on the range observed in the behavioural
sciences literature. Because they convey an intuitive
sense about the data irrespective of the subject area,
these levels can be useful as standard, if arbitrary,
substitutes for a biologically significant effect size in
ecological power analysis as well. In Cohens (1988)
classification system, effect sizes are typically categorized as small when they are subtle. Small effect
sizes are often associated with newly detected phenomena or very noisy systems. Medium effect sizes
are large enough to be perceived in the course of normal experience, while large effect sizes are easily
perceived at a glance.

75

To get a feeling for small, medium, and large standardized effect sizes for the wood duck example,
suppose the investigator really had no idea what deviation from the preferred mean of 25 cm would
constitute an unsuitable habitat. He might decide
to detect a medium effect size, reasoning that a difference perceptible to a human observer should be
evident to the ducks. Cohen (1988) defines a
medium standardized effect size for a t-test as
(1 2)/ = 0.5. Using the standard deviation
among cavity diameters (10 cm) in the formula, the
experimenter would need to set HA: 2 20 cm. If the
experimenter felt that even a subtle difference in cavity size could be important to the ducks, he would
choose Cohens small standardized effect size (0.2)
and set HA: 2 23 cm. Finally, if he felt that only an
obvious difference was worth detecting reliably, he
could choose a large standardized effect size (0.8)
and set HA: 2 17.
Because of their importance and subtlety, effect
size concepts have received considerable attention.
Further discussion can be found in: Toft and Shea
(1983); Rotenberry and Wiens (1985); Tanke and
Bonham (1985); Stewart-Oaten, et al. (1986); Kraemer
and Thiemann (1987); Millard (1987a); Cohen (1988,
1992); Forbes (1990); Parkhurst (1990); Peterman
(1990a); Fairweather (1991); Faith et al. (1991); Matloff (1991); McGraw and Wong (1992); Nicholson and
Fryer (1992); Schrader-Frechette and McCoy (1992);
Stewart-Oaten et al. (1992); Scheiner (1993); Osenberg et al. (1994); and Mapstone (1995).
6.6 How Should Power Analysis be Used in Experimental Adaptive Management?
6.6.1 Power considerations are an intrinsic part of
experimental design
The selection of experimental design is largely a question of managing the factors that influence Type II
error rate: sample size, variance, effect size, and .
These variables affect statistical power in different
ways. In the wood duck example, halving the sample
variance and doubling the biologically significant effect size improved power more than did doubling
sample size or doubling the critical level of . Thus,
while is completely under the experimenters control and there are good reasons to choose critical
-levels other than 0.05 (see Section 6.7), changing
critical -level is not the most effective way to gain
power (Lipsey 1990). Instead, as we will discuss,
efficient design usually requires specific knowledge

76

about the experimental system, which suggests that

pilot studies may sometimes be essential.
Choice of sample size and distribution of sampling
effort in time and space are largely under the control
of the experimenter and have been the subject of
much theoretical development (Pentico 1981; OBrien
1982; Hinds 1984; Hurlbert 1984; Andrew and Mapstone 1987; Ferraro et al. 1989; Green 1989; Krebs
1989; Kupper and Hafner 1989; Lipsey 1990; Ferraro
et al. 1994; Magnussen and Boyle 1995; Mapstone
1995). For many frequently used statistical tests, the
sample size necessary to achieve a given level of
power is easy to look up in such references as Cohen
(1988) or to calculate using software packages
(Thomas 1997b; Thomas and Krebs 1997). However,
in large management experiments where information
is costly to gather, decisions about sample size must
consider sampling costs as well.
While sample size is relatively straightforward to
adjust, experimenters should not overlook other aspects of experimental design that can be equally
effective at reducing rates of Type II error. First,
treatments can be chosen to increase effect size detectable in an experiment. For example, in a
management experiment involving effects of fertilization on green-up, it is important to know when the
fertilizer should be applied to maximize its effect.
Second, choosing dependent variables that are sensitive to treatments or impacts can also increase effect
size (Lipsey 1990). Third, the power of an experimental design can be strongly dependent upon the shapes
of response curves (Lipsey 1990; Nicholson and Fryer
1992). For example, if the relationship between the
independent and dependent variable is known to be
relatively steep in slope over some range of the independent variable, treatments targeted to that range
would produce the greatest effect size.
Taylor and Gerodette (1993) discussed an interesting example of the choice among dependent
variables, in which the power to detect a decline in a
Northern Spotted Owl population was influenced by
the variable monitored. Based on power analysis of a
number of simulated monitoring programs, the
authors made specific recommendations about the
choice of dependent variable. At low population
densities, estimates of birth and death rates provided
higher power to detect population declines than did
estimates of population size, whereas the reverse was
true at high population densities.
In another study designed to identify the most
useful variables for assessing impacts on water

quality, Osenberg et al. (1994) compared the sensitivity of population-based biological variables,
individual-based biological variables, and chemicalphysical variables. They found that standardized
effect sizes (and hence power) were greatest for individual-based biological variables because they
responded most sensitively to impacts and exhibited
relatively low error variance.
Ecological variables often exhibit large variance, so
strategies for reducing variance are especially important for achieving high statistical power. The total
variance is the sum of error variance and variance
that can be accounted for with additional information. Therefore, it is useful to account for as much of
the variance as possible in the experimental design.
Examples include blocking (Hurlbert 1984; Krebs
1989), covariate analysis (Wiens and Parker 1995),
and controlling for spatial heterogeneity (Dutilleul
1993). Error variance can also be reduced by improving measurement precision and reliability (Williams
and Zimmerman 1989; Lipsey 1990).
How could these strategies be applied to the
salamander example from Section 6.2.2? The experimenter might suspect that population density is
affected by aspect, irrespective of the width of the riparian reserve. The sample watersheds could be
divided into three blocks (north-facing, south-facing,
and east- or west-facing), to control for the variance
associated with aspect. Similarly, the experimenter
could include, as covariates, information about the
quantity and decay state of coarse woody debris in
each sample plot to account for some of the variance.
The error variance in estimates of population density
might also be reduced by sampling more intensively.
Finally, the experimenter might decide to estimate
recruitment and death rates of salamanders in addition to population density in an effort to measure
more responsive variables.
Finally, the most sophisticated design is not always
the most powerful. In more complex experimental
designs, statistical power is really a function of degrees of freedom, rather than straightforward sample
size. Because degrees of freedom are influenced by
both the extent of replication and the number of parameters to be estimated, increasing the complexity
of the design can be counterproductive with respect
to power. For example, in analysis of variance
(ANOVA), increasing the number of factors increases the number of parameters (treatment means) to be
estimated and this decreases the effective number of
replicates per cell, reducing power (Cohen 1988).

Therefore, to maintain comparable power to detect a

given main effect within a factorial ANOVA design,
more replicates per cell are required than if that same
main effect were tested within a one-way design. In
addition, interactions are detected with lower power
than main effects. When total sampling effort is restricted, it may be impossible to increase replication
adequately to justify adopting a complex design
(OBrien 1982; Cohen 1988; Lipsey 1990).
6.6.2 A posteriori power analysis can help interpret
nonsignificant results
When interpreting a statistically nonsignificant result
for a completed experiment, an experimenter should
calculate power a posteriori rather than uncritically
accepting the null hypothesis (Peterman 1990b;
Thomas 1997a). A posteriori power analysis answers
the question, What was the probability that this experiment could have detected a specific, biologically
important effect size? Calculating power a posteriori
is of course the only option if the Type II error rate
was not considered in the design of the experiment,
but it may also be necessary in well-designed experiments when sample variance turns out to be higher,
or samples smaller, than expected. Examples of
a posteriori power analysis are becoming increasingly
common in ecological literature (Thompson and
Neill 1991, 1993; Greenwood 1993; DeGraaf 1995; Reed
and Blaustein 1995; Lertzman et al. 1996; Pattanavibool and Edge 1996).
Effect size can be problematic in a posteriori
analyses. It is tempting to use the actual observed difference as a measure of effect size, but the analysis is
meaningful only if based on an effect size chosen
independently of the data, such as a biologically significant effect size (Thomas 1997a). If there is no
obvious biologically significant effect size, Cohens
(1988) representative small, medium, or large standardized effect sizes can be used.
Alternatively, in the absence of a biologically significant effect size,non-significant results can be
evaluated a posteriori with reverse power analysis,
which addresses the question, What effect size could
have been detected with acceptable power? This
technique involves solving for a different variable,
detectable effect size rather than power (Underwood
1981). Lertzman (1992) and Schieck et al. (1995)
demonstrate examples of this analysis.
Frequently, it is hard to decide what constitutes
acceptable power in a reverse power analysis. In
this case, it is more informative to present the results

77

not as a point calculation on some arbitrarily selected

level of power, but rather as a graph of power versus
detectable effect size for the experiment under consideration. Figure 6.3 demonstrates this relationship
for the nonsignificant correlation coefficients between
density of trees growing in gaps and gap area (Lertzman 1992), based on sample sizes of 3638. Lertzman
(1992) reported that he could have detected correlations of 0.45 with 80% probability; the graph presents
a wider picture (e.g., the analysis could have had a
50% chance of detecting correlations of 0.32 and a
95% chance of detecting correlations of r = 0.54 or
larger).

from passive (nonexperimental) monitoring, if principles of experimentation can be applied to decrease

the probability of erroneous conclusions. Therefore,
information-gathering protocols should be designed
so that they include clearly contrasting treatments,
controls, and attention to measurement error even
where the situation permits only passive monitoring.
The ongoing development of Before-After-ControlImpact paired (BACI) designs in environmental
impact research is an instructive example of this approach. These sampling designs have been proposed
to bring some benefits of experimentation to the
monitoring of unreplicated environmental impacts,
especially those impacts that are accidental (Bernstein and Zalinski 1983; Faith et al. 1991; Underwood
1991, 1994; Stewart-Oaten et al. 1992; Schroeter et al.
1993; Osenberg et al. 1994). See Schwarz (this volume,
Chap. 3) for a discussion on sampling designs for impact studies.

6.6.3 Principles of sound experimentation increase

power and improve inferences about causation in
nonexperimental situations
When the management problem does not demand or
allow direct experimentation, managers can still learn

0.8

Power

0.6

0.4

0.2

0
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Correlation coefficient (r)

. A posteriori power analysis: relationship between Pearson correlation coefficients and the power with which
they could be detected, given N = 38.

78

6.7 How do Experimenters Decide What Probabilities of Error are Acceptable?

6.7.1 Error rates have traditionally been set by convention
Errors of inference are inherent in any system that
involves learning from incomplete information, but
managers and experimenters can control the frequency of these errors. While the Type I error rate
has conventionally been limited to 0.05, no universally accepted limit for Type II error rates has emerged.
Cohen (1988) suggested that experiments should be
designed to have a power of 0.80. Others have proposed that and should be made equal (Lipsey
1990). These conventions are arbitrary and do not reflect any measure of the relative importance of the
two types of error or the gravity of their consequences.
6.7.2 In applied science, costs of errors should be
taken into account
In resource management experiments, where errors
of inference may lead to substantial costs, those costs
must be considered. Certainly the status quo, in
which is ignored, is undesirable because the probability of incurring large and possibly irreversible costs
associated with Type II error can be very high and
unknown.
One possibility is to set the expected costs of the
two types of errors equal, where expected cost is the
cost of the error times its probability (Toft and Shea
1983; Peterman 1990a). Thus, if C1 and C2 are the respective costs of Type I and Type II errors, we set
C1 = C2 . An acceptable value for should then be
(C1/ C2) . Mapstone (1995) suggests that both and
may need to vary when costs of sampling or other
circumstances do not allow to become as small as
prescribed by this formula. While issues of cost and
error rates have been discussed by many
authors (Pentico 1981; Hinds 1984; Mapstone and
Andrew 1987; Millard 1987a, 1987b; Ferraro et al. 1989;
Loftis et al. 1989; Davis and Hardy 1990; Holland and
Ordoukhani 1990; Parkhurst 1990; Fairweather 1991;
Schroeter et al. 1993; Power at al. 1995; Wiens and
Parker 1995), few examples of these considerations
are being practiced.
6.7.3 Other factors should also influence the choice
of and
The purpose of the research should be considered
when setting and . The efficiency of exploratory

research or data snooping will often improve if is

set relatively high and relatively low, because it is
important to detect previously unknown relationships. Careful follow-up studies can discover any
cases of Type I errorincorrect identification of a
relationship where in reality none exists.
Another issue in setting error rates is the prior
probability that each hypothesis is true. A hypothesis
that seems likely to be true, based on related research,
should be treated more cautiously with respect to erroneous rejection than a hypothesis that seems less
credible (Lipsey 1990). For example, suppose that we
can reasonably expect the community of soil microorganisms to be less disturbed in a retention
silvicultural system than in large clearcuts, but that
the difference may be subtle. An experiment comparing these silvicultural systems should be designed
with high power to detect the difference (i.e.,
should be set low).
6.8 Recommendations for Adaptive Management
6.8.1 Explicit attention to rates of Type I and Type
II error should become standard practice
Several authors have surveyed experimental literature
and found few examples that address Type II error
(Sedlmeier and Gigerenzer 1989; Peterman 1990a,
1990b; Fairweather 1991; Searcy-Bernal 1994). Although there has been some discussion about
reconsidering the arbitrary limit on Type I error
( < 0.05), that limit is rarely reviewed in discussions
of either power analysis or significance testing. However, the number of journal articles reporting new
theoretical developments in the application of power
analysis to ecological problems and the increasing
availability of software for the purpose suggest that
errors of inference will be easier to estimate and
interpret in the future. The following recommendations apply to all ecological research, but especially to
large-scale management experiments:
Experimenters and decision-makers should embrace the concept that some errors of inference are
unavoidable, but their frequency can be controlled
by astute design of experiments and monitoring
systems.
A priori power analysis, with an explicit statement
of desirable levels of and , should be included
in the design process for all experiments and monitoring programs.

79

 All reports of nonsignificant results should mention the effect size and power of the experiment.
Where appropriate, a posteriori power analysis may
be used.
Where potential costs of the errors of inference to
various stakeholders can be quantified, these costs
should be included in decisions about acceptable
levels of and .
Where currently available experimental designs
lack power, research should be directed toward developing new, powerful methodologies, such as
Before-After-Control-Impact paired designs (Underwood 1994).
Resources should be allocated to pilot studies that
will help to improve the power of large experiments.
A priori power analyses are often difficult because
not enough is known about potential response variables, biologically significant effect sizes, and spatial
and temporal variability. It would be useful to carry
out large-scale, long-term monitoring of these variables in forest ecosystems, with the express purpose of estimating them for use in future power
analyses and choices about experimental design
(Osenberg et al. 1994). Standard response variables
such as those proposed by Keddy and Drummond
(1996) for eastern deciduous forests could become
a starting point for this sort of database.
6.9 Relevant Literature and Software
6.9.1 A few key references guide experimenters
through power analysis for the most frequently
used statistical tests
The classic reference to statistical power is Cohen
(1988). Cohen provides clearly written instructions
for calculating standardized effect size and other
input parameters to power and sample size tables. He
provides these tables for t-tests, tests involving correlation coefficients, tests involving proportions, the
sign test, chi-square tests for goodness of fit and contingency, analysis of variance and covariance,
multiple regression and correlation, and set correlation and multivariate methods (e.g., canonical
correlation, MANOVA, and MANCOVA).
Zar (1996) presents a graph of power and sample
size for analysis of variance, as well as formulas for
calculating power and required sample size for a variety of other tests. While he does not include tabled

80

values, the formulas are discussed with the details of

the tests themselves, including biological examples.
In addition, Zar discusses examples of a posteriori
power analyses.
Nemec (1991) introduces power analysis using examples from forest research. Included are example
routines for the SAS statistical software package that
compare the power of completely randomized and
randomized block analysis of variance designs, and
calculate power for one- and two-sample t-tests and
one- and two-way ANOVA. Several pamphlets addressing various aspects of power analysis are also
available from the B.C. Ministry of Forests (Bergerud
1992, 1995a, 1995b, 1995c, 1995d; Sit 1992).
Lipsey (1990) discusses and compares many factors
that affect statistical power, including adjustment of
critical and strategies for optimizing standardized
effect size. This book presents power and sample size
relationships as graphs, at some cost to precision in
reading off the numbers.
6.9.2 Power analysis is available in many software
packages
Over the last few years, the variety of software packages that perform power analysis, sample size
determination, and effect size operations has greatly
increased. Thomas (1997b) maintains an annotated
list of software packages on the World Wide Web.
Many of these packages are reviewed in Thomas and
Krebs (1997). For other discussions of software, see
Goldstein (1989), Borenstein et al. (1990, 1992),
Rothstein et al. (1990), Steiger and Fouladi (1992),
and Meyer (1995). In addition, on-line power calculations are available for ANOVA (Friendly 1996) and
for correlation coefficients and tests of parameters
for normal, Poisson, and exponential distributions
(Bond 1996).
6.9.3 Power analysis procedures have been developed for many specialized applications relevant to
ecological experimentation
Numerous examples are available of power analyses
for complex designs, including factorial and repeated
measuresanalysis of variance (OBrien 1982;
Bittman and Carniello 1990; Muller et al. 1992), moderated multiple regression (Stone-Romero et al.
1994), and multivariate general linear hypotheses
(Muller and Peterson 1984). These papers focus on
practical methods for addressing questions related to
power, effect size, and sample size.
Long-term ecological experiments and monitoring

programs, like long-term medical trials, may not

need to be carried to completion if early data are sufficiently decisive (Allison et al. 1997). These studies
should therefore be subject to interim analysis to address the question, Given the results so far, what are
the chances that the conclusions will change if the
experiment or monitoring is continued to completion? As an example, Davis and Hardy (1990)
described one decision process for interim analysis of
long-term medical trials. This processstochastic
curtailmentinvolves calculating, given the current
data, the maximum possible levels of and that
would occur if the experiment were carried to its
conclusion. If, at some interim point, those maximum error levels are suitably low, the experiment or
monitoring program could be terminated.
Deviations from normal distributions are often of
concern to ecologists. Gingerich (1995) compared the
power of three tests designed to distinguish normal
from lognormal distributions. Based on Monte Carlo
simulations, he concluded that the Anderson-Darling
test is the most powerful, but even it requires very
large sample sizes to achieve high power. The power
of all the tests was sensitive to the coefficient of varia
tion (V = s/x ) in the sample; power increases as V
increases. Sawilowsky and Blair (1992) discussed the
Type I and Type II error properties of the t-test applied to measures with frequency distributions that
differ grossly from normal. They concluded that the
t-test is subject to increased Type I error under deviations from normality when the test is one-tailed or
when sample sizes are small or unbalanced. Type II
error does not increase dramatically, but Sawilowsky
and Blair suggested that nonparametric substitutes
often have higher power than the t-test when its assumptions of normality are seriously violated.
Measures of biodiversity are often important response variables in adaptive management
experiments. To estimate statistical power for studies
involving these measures, their sampling variance
and other distributional properties must be known.
Methods to address this issue are currently under
development. For example, Magnussen and Boyle
(1995) provide guidelines for power and sample size
for tests involving the Shannon-Weaver and Simpson
diversity indices. Mark-recapture methods are being
adapted to improve the estimation of species richness
and its distributional properties (E. Cooch, Simon
Fraser University, pers. comm., 1996).
Other types of statistical tests, specifically designed
for ecological contexts, have been subjected to power

analysis. Some examples that may be useful in adaptive management experiments include:
density dependencepower and sample size for
tests designed to detect whether population parameters vary as a function of population density
(Solow and Steele 1990; Dennis and Taper 1994);
trend detectionpower and sample size for tests
designed to detect whether a variable is changing
with time (Hinds 1984; Tanke and Bonham 1985;
Harris 1986; Gerodette 1987, 1991; Whysong and
Brady 1987; Kendall et al. 1992; Loftis et al. 1989);
detection of rare speciessample sizes necessary
to detect rare species, based on the Poisson distribution (Green and Young 1993);
resource selectionpatterns of Type I and Type II
errors for four tests of habitat/resource selection
(Alldredge and Ratti 1986);
home range independencepower of Schoener
statistic for independence of animal home ranges
(Swihart and Slade 1986);
environmental monitoringpower, sample size,
and cost considerations for programs of environmental impact monitoring (Skalski and McKenzie
1982; Millard 1987b; Ferraro and Cole 1990;
Ferraro et al. 1989; Parkhurst 1990; Smith and
McBride 1990; Ferraro et al. 1994; Wiens and Parker 1995);
analysis of covariance in environmental monitoringanalysis of Type I and Type II error for
ANCOVA with examples from water quality monitoring (Green 1986);
environmental impact detection, unique cases, and
before-after-control-impact design issues (Bernstein and Zalinski 1983; Faith et al. 1991;
Underwood 1991, 1994; Stewart-Oaten et al. 1992;
Schroeter et al. 1993; Osenberg et al. 1994; Allison
et al. 1997; Gorman and Allison 1997); and
spatial patterns and heterogeneitypower analysis
for experimental designs that take spatial patterns
and heterogeneity into account (Andrew and Mapstone 1987; Downing and Downing 1992; Scharf
and Alley 1993).
6.9.4 When analytic methods are not appropriate,
Monte Carlo simulation can be used to estimate
power
Many ecological analyses involve specialized statistics
or experimental designs for which no analytic formulas exist for calculating power. In such cases, Monte
Carlo simulation can be used to produce many simulated data sets generated from distributions with

81

known parameter values corresponding to given null

and alternative hypotheses. The experimenter can
then estimate statistical power by tallying the frequency with which HO is correctly rejected by the
simulated data.
Taylor and Gerodette (1993) used this method to
determine the power to detect population declines in
the Northern Spotted Owl when the monitored variables were estimates of survival and reproductive
rates. They simulated hypotheses about trends in
population density: the null hypothesis represented a
stable population ( = 1, where is the geometric
rate of population change), while a set of alternative
hypotheses represented populations declining at
specified rates (e.g., = 0.96). Survival, reproduction, capture, and recapture of individual birds in
each declining or stable population were modelled
stochastically using known ranges of variation over a
number of years; the mark-recapture estimates of demographic parameters were analyzed to generate an
estimate of . Each population and its associated estimates of were simulated several thousand times.
For stable populations (i.e., ones for which HO was
true), the critical value of estimates of was determined by finding the fifth percentile of the frequency
distribution of simulated estimates. Below that value,
HO would be rejected at = 0.05. Then, for simulated
populations in which HA was true, the power to detect the decline could be estimated by tallying the
proportion of parameter estimates for which HO was
rejected at the critical value.
Other examples of Monte Carlo methods for
power analysis include Loftis et al. (1989), who used it
to identify the most effective tests for detecting trends
in long-term water quality monitoring, and Alldredge and Ratti (1986), who simulated the statistical
behaviour of a number of models of habitat selection.
Using Monte Carlo simulations of fish populations,
Peterman and Routledge (1983) estimated the power
of proposed management experiments, and Peterman and Bradford (1987) demonstrated the difficulty
of reliably detecting time trends in recruitment.
6.10 Conclusions
Scientists and resource managers are benefiting
from a new level of sophistication in the statistical interpretation of management experiments and
monitoring programs. This involves an explicit and
balanced focus on errors of inference, including Type
II error (failure to reject a false null hypothesis) as

82

well as Type I error (rejecting a true null hypothesis).

As this new standard of practice evolves, decisionmakers and experimenters will find it helpful to bear
in mind the following points:
Experimenters can control statistical power a priori
by making wise decisions regarding pilot studies,
sampling methodology, and experimental or monitoring design. The advice of statistical consultants
is increasingly important in the planning process
as standards of experimentation improve.
If the outcome of an experiment or monitoring
program is likely to influence a management decision, statistical power should be reported when
results are non-significant.
In many situations, it is useful to free Type I and
Type II error rates from traditional arbitrary values. Desirable levels of both error rates can be
selected rationally as a function of the goal of the
study, current understanding of the system under
investigation, and costs of each type of error.
Software and written explanations, designed to
make statistical power analysis accessible to a wide
range of users, are widely available.
Acknowledgements
Thanks to all those who helped with this chapter.
Randall Peterman, Vera Sit, and Brenda Taylor provided much help and encouragement on the early
drafts. Doug Steventon, Rita Winkler, Wendy
Bergerud, Jeff Stone, Keith Thomas, Randall Peterman, and Michael Stoehr thoroughly reviewed the
manuscript and provided valuable suggestions for
improving the final version.

References
Alldredge, J.R. and J.T. Ratti. 1986. Comparison of
some statistical techniques for analysis of resource selection. J. Wildl. Manage. 50:15765.
Allison, D.B., J.M. Silverstein, and B.S. Gorman. 1997.
Power, sample size estimation, and early stopping rules. In Design and analysis of single case
research. R.D. Franklin, D.B. Allison, and B.S.
Gorman (editors). Erlbaum, Mahwah, N.J., pp.
33572.
Andrew, N.L. and B.D. Mapstone. 1987. Sampling
and the description of spatial pattern in marine
ecology. Oceanography and Marine Biology
Annual Reviews 25:3990.
Bergerud, W. 1992. A general description of hypothesis testing and power analysis. B.C. Min. For.,
Res. Br., Victoria, B.C. Biometrics Inf. Pamph.
No. 37.
______. 1995a. Power analysis and sample sizes for
completely randomized designs with subsampling. B.C. Min. For., Res. Br., Victoria, B.C.
Biometrics Inf. Pamph. No. 49.
______. 1995b. Power analysis and sample sizes for
randomized block designs with subsampling.
B.C. Min. For., Res. Br., Victoria, B.C. Biometrics Inf. Pamph. No. 50.
______. 1995c. Programs for power analysis/sample
size calculations for CR and RB designs. B.C.
Min. For., Res. Br., Victoria, B.C. Biometrics
Inf. Pamph. No. 51.
______. 1995d. Post-hoc power analysis for ANOVA
F-tests. B.C. Min. For., Res. Br., Victoria, B.C.
Biometrics Inf. Pamph. No. 52.
Bergerud, W.A. and W.J. Reed. [n.d.] Bayesian statistical methods. This volume.
Bernstein, B.B. and J. Zalinski. 1983. An optimum
sampling design and power tests for environmental biologists. J. Environ. Manage. 16:3543.
Bittman, R.M. and M.L. Carniello. 1990. The design
of an experiment using statistical power with a
startle chamber study as an example. J. Appl.
Toxicology 10:1258.

Bond, J. 1996. Online power calculator. Stat. Cons.

Cen., UCLA, Los Angeles, Calif. Available online:</www.stat.ucla.edu/calculators/
powercalc.>
Borenstein, M., J. Cohen, H.R. Rothstein, S. Pollack,
and J.M. Kane. 1990. Statistical power analysis
for one-way analysis of variance: a computer
program. Beh. Res. Methods, Instr. Comp.
22:27182.
______. 1992. A visual approach to statistical power
analysis on the microcomputer. Beh. Res.
Methods, Instr. Comp. 24:56572.
British Columbia Ministry of Forests and B.C. Environment. 1995. Riparian area management
guidebook. Victoria, B.C. Forest Practices Code
guidebook.
Cohen, J. 1988. Statistical power analysis for the behavioral sciences. Lawrence-Erlbaum, Hillsdale,
N.J.
______. 1992. A power primer. Psychol. Bull.
112:1559.
Davis, B.R. and R.J. Hardy. 1990. Upper bounds for
Type I and Type II error rates in conditional
power calculations. Commun. Statist.: Theory
Methodol. 19:357184.
DeGraaf, R.M. 1995. Nest predation rates in managed
and reserved extensive northern hardwood
forests. For. Ecol. Manage. 79:22734.
Dennis, B. and M.L. Taper. 1994. Density dependence
in time series observations of natural populations: Estimation and testing. Ecol. Monogr.
64:20524.
Downing, J.A. and W.L. Downing. 1992. Spatial aggregation, precision, and power in surveys of
freshwater mussel populations. Can. J. Fish.
Aquat. Sci. 49:98591.
Dutilleul, P. 1993. Spatial heterogeneity and the design of ecological field experiments. Ecology 74
(6):164658.
Fairweather, P.B. 1991. Statistical power and design
requirements for environmental monitoring.
Austr. J. Marine Freshwater Res. 42:55567.

83

Faith, D.P., C.L. Humphrey, and P.L. Dostine. 1991.

Statistical power and BACI designs in biological
monitoring: Comparative evaluation of measures of community dissimilarity based on
benthic macroinvertebrate communities in
Rockhole Mine Creek, Northern Territory,
Australia. Austr. J. Marine Freshwater Res.
42:589602.

Green, R.H. 1986. Some applications of linear models

for analysis of contaminants in aquatic biota. In
Statistical aspects of water quality monitoring,
A.H. El-Shaarawi and R.E. Kwiatkowski (editors). Elsevier, New York, N.Y.
______. 1989. Power analysis and practical strategies
for environmental monitoring. Environ. Res.
50:195205.

Ferraro, S.P. and F.A. Cole. 1990. Taxonomic level

and sample size sufficient for assessing pollution impacts on the Southern California Bight
benthos. Marine Ecol. Prog. Ser. 67:25162.

Green, R.H. and R.C. Young. 1993. Sampling to detect rare species. Ecol. Applic. 3:3516.

Ferraro, S.P., F.A. Cole, W. DeBen, and R.C. Swartz.

1989. Power-cost efficiency of eight macrobenthic sampling schemes in Puget Sound,
Washington, U.S.A. Can. J. Fish. Aquat. Sci.
46:215765.

Harris, R.B. 1986. Reliability of trend lines obtained

from variable counts. J. Wildl. Manage.
50:16571.

Ferraro, S.P., R.C. Swartz, F.A. Cole, and W. DeBen.

1994. Optimum macrobenthic sampling protocol for detecting pollution impacts in the
Southern California Bight. Environ. Monitoring Assessm. 29:12753.
Forbes, L.S. 1990. A note on statistical power. The
Auk 107:43853.

Greenwood, J.J.D. 1993. Statistical power. Animal Behavior 46:1011.

Hinds, W.T. 1984. Towards monitoring of long-term

trends in terrestrial ecosystems. Environ. Cons.
11:118.
Holland, B. and N.K. Ordoukhani. 1990. Balancing
Type I and Type II error probabilities: Further
comments on proof of safety vs. proof of hazard. Communications in statistics: theory and
methodology 19:355770.

Friendly, M. 1996. Power analysis for ANOVA designs. York Univ., Toronto, Ont. Available on
line: <www.math.yorku.ca/SCS/Demos/
power/>[January 1996].

Hurlbert, S.H. 1984. Pseudoreplication and the design

of ecological field experiments. Ecol. Monogr.
54:187211.

Gerodette, T. 1987. A power analysis for detecting

trends. Ecology 68:136472.

Keddy, P.A. and C.G. Drummond. 1996. Ecological

properties for the evaluation, management, and
restoration of temperate deciduous forest
ecosystems. Ecol. Applic. 6:74862.

______. 1991. Models for power of detecting trends: a

reply to Link and Hatfield. Ecology 72:88992.
Gingerich, P.D. 1995. Statistical power of EDF tests of
normality and the sample size required to distinguish geometric-normal (lognormal) from
arithmetic-normal distributions of low variability. J. Theoretical Biol. 173:12536.
Goldstein, R. 1989. Power and sample size via MSPCDOS computers. Am. Statist. 43:25360.
Gorman, B.S. and D.B. Allison. 1997. Statistical alternatives for single-case research. In Design and
analysis of single case research. R.D. Franklin,
D.B. Allison, and B.S. Gorman (editors). Erlbaum, Mahwah, N.J. pp. 159214.

84

Kendall, K.C., L.H. Metzgar, D.A. Patterson, and

B.M. Steele. 1992. Power of sign surveys to
monitor population trends. Ecol. Applic.
2:42230.
Kirk, R.E. 1982. Experimental design: procedures for
the behavioral sciences. Brooks/Cole, Monterey, Calif.
Kraemer, H.C. and S. Thiemann. 1987. How many
subjects? Statistical power analysis in research.
Sage, London, U.K.
Krebs, C.J. 1989. Ecological methodology. Harper
Collins, New York, N.Y.

Kupper, L.L. and K.B. Hafner. 1989. How appropriate

are popular sample size formulas? Am. Statist.
43:1015.
Lertzman, K.P. 1992. Patterns of gap-phase replacement in a subalpine forest. Ecology 73:65769.
Lertzman, K.P., G.D. Sutherland, A. Inselberg, and
S.C. Saunders. 1996. Canopy gaps and the landscape mosaic in a coastal temperate rainforest.
Ecology 77:1254.
Lipsey, M.W. 1990. Design sensitivity: statistical
power for experimental research. Sage, London,
U.K.
Loftis, J.C., R.C. Ward, R.D. Phillips, and C.H. Taylor. 1989. An evaluation of trend detection
techniques for use in water quality monitoring
programs. Environ. Res. Lab., Off. Res. Dev.,
U.S. Environ. Prot. Ag., Corvallis, Oreg.
McGraw, K.O. and S.P. Wong. 1992. A common language effect size statistic. Psychol. Bull.
111:3615.
Magnussen, S. and T.J.B. Boyle. 1995. Estimating
sample size for inference about the ShannonWeaver and the Simpson indices of species
diversity. For. Ecol. Manage. 78:7184.
Mapstone, B.D. 1995. Scalable decision rules for environmental impact studies: effect size, Type I,
and Type II errors. Ecol. Applic. 5:40110.
Matloff, N.S. 1991. Statistical hypothesis testing:
Problems and alternatives. Environ. Entom.
20:124650.

Muller, K.E., L.M. LaVange, S.L. Ramey, and C.T.

Ramey. 1992. Power calculations for general linear multivariate models including repeated
measures applications. J. Am. Statist. Assoc.
87:120926.
Muller, K.E. and B.L. Peterson. 1984. Practical methods for computing power in testing the
multivariate general linear hypothesis. Comput.
Statist. Data Anal. 2:14358.
Nemec, A.F.L. 1991. Power analysis handbook for the
design and analysis of forestry trials. B.C. Min.
For., For. Sc. Res. Br., Victoria, B.C. Biometrics
Inf. Handb. No. 2.
Nicholson, M.D. and R.J. Fryer. 1992. The statistical
power of monitoring programmes. Marine Pollut. Bull. 24:1469.
OBrien, R.G. 1982. Performing power sensitivity
analyses on general linear model hypotheses.
Proc. Statistical Computing Section, Am. Statist. Assoc.
Osenberg, C.W., R.J. Schmitt, S.J. Holbrook, K.E.
Abu-Saba, and A.R. Flegal. 1994. Detection of
environmental impacts: natural variability, effect size, and power analysis. Ecol. Applic.
4:1630.
Parkhurst, D.F. 1990. Statistical hypothesis tests and
statistical power in pure and applied science. In
Acting under uncertainty: multidisciplinary
conceptions. G.M. von Furstenberg (editor).
Kluwer Acad. Publ, Boston, Mass.

Meyer, G.E. 1995. Power & Effect: A statistical utility

for Macintosh and Windows systems. Beh. Res.
Methods, Instr. Comp. 27:1348.

Pattanavibool, A. and W.D. Edge. 1996. Single-tree

selection silviculture affects cavity resources in
mixed deciduous forests in Thailand. J. Wildl.
Manage. 60:6773.

Millard, S.P. 1987a. Environmental monitoring, statistics, and the law: Room for improvement.
Am. Statist. 41:24953.

Pentico, D.W. 1981. On the determination and use of

optimal sample sizes for estimating the difference in means. Am. Statist. 35:402.

______. 1987b. Proof of safety vs. proof of hazard.

Biometrics 43:71925.

Peterman, R.M. 1990a. Statistical power analysis can

improve fisheries research and management.
Can. J. Fish. Aquat. Sc. 47:215.

Morgan, M.G. and M. Henrion. 1990. Uncertainty: A

guide to dealing with uncertainty in quantitative risk and policy analysis. Cambridge Univ.
Press, Cambridge, Mass.

______. 1990b. The importance of reporting statistical power: The forest decline and acidic
deposition example. Ecology 71:30247.

85

Peterman, R.M. and M. Bradford. 1987. Statistical

power of trends in fish abundance. Can. J. Fish.
Aquat. Sci. 44:187989.
Peterman, R.M. and R. Routledge. 1983. Experimental management of Oregon coho salmon
(Oncorhynchus kisutch): designing for yield of
information. Can. J. Fish. Aquat. Sci.
40:121223.
Peterman, R.M. and C. Peters. [n.d.] Decision analysis: taking uncertainties into account in forest
resource management. This volume.
Platt, J.R. 1964. Strong inference. Science 146:34752.
Power, M., G. Power, and D.G. Dixon. 1995. Detection and decision-making in environmental
effects monitoring. Environ. Manage.
19:62939.
Raiffa, H. 1968. Decision analysis: introductory lectures on choices under uncertainty.
Addison-Wesley, Reading, Mass.
Reed, J.M. and A.R. Blaustein. 1995. Assessment of
nondeclining amphibian populations using
power analysis. Cons. Biol. 9:12930.
Rotenberry, J.T. and J.A. Wiens. 1985. Statistical
power analysis and community-wide patterns.
Am. Nat. 125:1648.
Rothstein, H.R., M. Borenstein, J. Cohen, and S. Pollack. 1990. Statistical power analysis for
multiple regression/correlation: a computer
program. Educ. Psychol. Measure. 50:81930.
Sawilowsky, S.S. and R.C. Blair. 1992. A more realistic
look at the robustness and Type II error properties of the t-test to departures from
population normality. Psychol. Bull. 111:35260.
Scharf, P.C. and M.M. Alley. 1993. Accounting for
spatial yield variability in field experiments increases statistical power. Agron. J. 85:12546.
Scheiner, S.M. 1993. Introduction: theories, hypotheses, and statistics. In Design and analysis of
ecological experiments. S.M. Scheiner and
J. Gurevitch (editors). Chapman and Hall,
New York, N.Y.
Schieck, J., K.P. Lertzman, B. Nyberg, and R. Page.
1995. Effects of patch size on birds in oldgrowth montane forests. Cons. Biol. 9:107284.

86

Schrader-Frechette, K.S. and E.D. McCoy. 1992. Statistics, costs, and rationality in ecological
inference. Trends in Ecology and Evolution
7:969.
Schroeter, S.C., J.D. Dixon, J. Kastendiek, R.O.
Smith, and J.R. Bence. 1993. Detecting the ecological effects of environmental impacts: a case
study of kelp forest invertebrates. Ecol. Applic.
3:33150.
Schwarz, C.J. [n.d.]. Studies of uncontrolled events.
This volume.
Searcy-Bernal, R. 1994. Statistical power and aquacultural research. Aquaculture 127:137368.
Sedlmeier, P. and G. Gigerenzer. 1989. Do studies of
statistical power have an effect on the power of
studies? Psychol. Bull. 105:30916.
Sit, V. 1992. Power analysis and sample size determination for contingency table tests. B.C. Min.
For., Res. Br., Victoria, B.C. Biometrics Inf.
Pamph. No. 41.
Skalski, J.R. and D.H. McKenzie. 1982. A design for
aquatic monitoring programs. J. Environ. Manage. 14:23751.
Smith, D.G. and G.B. McBride. 1990. New Zealands
national water quality monitoring network
design and first years operation. Water Res.
Bull. 26:76775.
Solow, A.R. and J.H. Steele. 1990. On sample size, statistical power, and the detection of density
dependence. J. Animal Ecol. 59:10736.
Steiger, J.H. and R.T. Fouladi. 1992. R2: A computer
program for interval estimation, power calculations, sample size estimation, and hypothesis
testing in multiple regression. Beh. Res. Methods, Instr. Comp. 24:5812.
Stewart-Oaten, A., J.R. Bence, and C.W. Osenberg.
1992. Assessing effects of unreplicated perturbations: no simple solutions. Ecology
73:13961404.
Stewart-Oaten, A., W.W. Murdoch, and K.R. Parker.
1986. Environmental impact assessment:
pseudoreplication in time? Ecology
67:92940.

Stone-Romero, E.F., G.M. Alliger, and H. Aguinis.

1994. Type II error problems in the use of moderated multiple regression for the detection of
moderating effects of dichotomous variables.
J. Management. 20:16778.
Swihart, R.K. and N.A. Slade. 1986. The importance
of statistical power when testing for independence in animal movements. Ecology 67:2558.
Tanke, W.C. and C.D. Bonham. 1985. Use of power
curves to monitor range trend. J. Range Manage. 38:42831.
Taylor, B.L. and T. Gerodette. 1993. The uses of statistical power in conservation biology: The
vaquita and northern spotted owl. Cons. Biol.
7:489500.

______. 1994. On beyond BACI: sampling designs

that might reliably detect environmental disturbances. Ecol. Applic. 4:315.
Whysong, G.L. and W.W. Brady. 1987. Frequency
sampling and Type II errors. J. Range Manage.
40:4724.
Wiens, J.A. and K.R. Parker. 1995. Analyzing the effects of accidental environmental impacts:
approaches and assumptions. Ecol. Applic.
5:106983.
Williams, R.H. and D.W. Zimmerman. 1989. Statistical power analysis and reliability of
measurement. J. Gen. Psychol. 116:35969.
Zar, J.H. 1996. Biostatistical analysis (3rd ed.) Prentice Hall, Upper Saddle River, N.J.

Thomas, L. 1997a. Retrospective power analysis.

Cons. Biol. 11:27680.
______. 1997b. A comprehensive list of power analysis software for PCs. Math. and Comp. Sci.,
Univ. St. Andrews, Scotland. Available online:
<www.interchg.ubc.ca/cacb/power/>
[Oct. 17, 1997].
Thomas, L. and C.J. Krebs. [1997]. A review of statistical power analysis software. Bull. Ecol. Soc.
Am. In press.
Thompson, C.F. and A.J. Neill. 1991. House wrens do
not prefer clean nestboxes. Animal Behavior
42:10224.
______. 1993. Statistical power and accepting the null
hypotheses. Animal Behavior 46:1012.
Toft, C.A. and P.J. Shea. 1983. Detecting communitywide patterns: estimating power strengthens
statistical inference. Am. Nat. 122:61825.
Underwood, A.J. 1981. Techniques of analysis of variance in experimental marine biology and
ecology. Oceanogr. and Marine Biol. Ann. Rev.
19:513605.
______. 1991. Beyond BACI: Experimental designs
for detecting human environmental impacts on
temporal variations in natural populations.
Australian J. Marine and Freshwater Res.
42:56987.

87

7 BAYESIAN STATISTICAL METHODS

WENDY A. BERGERUD AND WILLIAM J. REED
Abstract

This chapter provides a brief introduction and

overview of the Bayesian approach to statistical inference. The Bayesian approach is particularly suitable for
adaptive management since it provides a way of incorporating prior knowledge with newly acquired
knowledge and, based on a model, can be used to assign probabilities to possible states of nature. These
probabilities are directly useful in decision theory applications. In contrast, the familiar and well-known
frequentist methods treat each experiment or study
on its own, formally ignoring other information, and
thus syntheses of information must be done in an ad
hoc fashion.
Frequentist and Bayesian statistical approaches are
briefly contrasted by way of a forestry example. This
example is then simplified and used to demonstrate
Bayesian methods. Bayesian decision theory and
model building are also described.

7.1 Introduction
In most introductory statistics courses, students encounter basic concepts of the familiar frequentist
paradigm, such as sampling distributions, significance tests, P-values, and confidence intervals. These
concepts, which are further developed in specialized
courses on regression, design of experiments, and
sampling methods, form the basis of the statistical
toolkit that most graduates carry with them into the
world of science and management. When faced with
problems involving experiments, data, and decisions,
most foresters, biologists, and forest managers will
naturally reach into this toolkit. However, the familiar statistical toolkit often proves inadequate for dealing with management problems. Managers make decisions in an environment of uncertainty, where the
best choice among several alternatives is unknown.
They often want to know the probability that a hypothesis is true, or the degree to which it is trueinformation that frequentist statistics does not directly
provide. Despite its limitations for management, the
frequentist framework is seldom questioned by practitioners, partly because the limitations of frequentist

methods are rarely discussed in introductory courses,

but perhaps more importantly, because practitioners
know of no alternative.
Bayesian statistics, on the other hand, is an alternative theory, which is particularly suitable for
adaptive management since it provides a way of incorporating prior knowledge with newly acquired
data and of assigning probabilities to possible states
of nature.
The essential difference between the Bayesian and
the frequentist approaches hinges on the way in
which uncertainty is modelled, and on the way in
which randomness is conceived. To the Bayesian, uncertainty can be represented by assigning
probabilities to various possibilities. These probabilities may be obtained from previous data or from the
subjective beliefs of those knowledgeable of the subject matter. In contrast, a frequentist considers that
probabilities can only be meaningfully assigned to
events in repeatable experiments.1 It is claimed that
the road to scientific progress lies through objectively
falsifying or confirming scientific hypotheses, rather
than through adjusting ones beliefs concerning their
plausibility. However, adaptive management involves exactly the second kind of process, as
information on the consequences of management
actions becomes available as part of an ongoing
process.
This chapter presents a brief overview of the
Bayesian approach to statistical inference and decision-making. Bayesian ideas are also contrasted with
those from the dominant frequentist paradigm. Peterman and Peters (this volume, Chap. 8) discuss in
more depth the application of Bayesian statistics to
management decisions.
7.2 Frequentist and Bayesian Statistics
The dominant paradigm for statistical inference is
based on a frequentist notion of probability, which
we will call frequentist statistics. Data generated from
an experiment or study are fitted by a statistical
model. This model usually includes one or more unknown quantities called parameters, which are estimated during the fitting process. The fitted model

1 An accessible discussion of some of these points can be found in Swindel (1972), Dennis (1996), and Edwards (1996), and in the Teachers
Corner of The American Statistician, Vol. 51, (3), pp. 24174 (several articles).

89

summarizes the data and can be used to answer questions or to calculate confidence limits for the parameters. Questions are posed as hypotheses, the most
frequent being the well-known null hypothesis (that
a particular parameters unknown value is zero). By
considering many hypothetical replications of the experiment, statisticians can determine the behaviour
of fitted parameter values or of some function of the
parameters. This behaviour is described by a frequency or sampling distribution (such as the normal, t-,
and F-distributions) and is used to calculate the familiar P-values and confidence intervals.
One could argue that frequentist statistics is only
really applicable to the analysis of data that arise from
a procedure that is intrinsically repeatable. Such a restriction would severely limit its use. For example, it
would rule out almost completely its use in time
series analysis, time being universally recognized as
non-repeatable. This restriction would also rule out
applying frequentist statistics in many areas of
forestry, because trees generally take a long time to
grow and growing conditions could change over the
course of an experiment. For example, an experiment
involving growing various tree species would not really be repeatable, given the dependence of growth
upon weather over several decades and the possibilities of site changes. However, this objection is usually
overcome by observing that replication is regarded in
a hypothetical sense for the purpose of interpreting
results, and that even in truly repeatable experiments,
the experiment is seldom actually repeated. Rather,
in both cases one contemplates a universe of possible
replications for the purpose of comparing the actual
observed results.
In forest management, managers would like simple
answers to practical questions from sampling procedures and studies, whether experimental, observational, or a combination of the two. For example, an
assessment of the probability that one or more hypotheses are true, or the probability that an estimate
of a parameter is reasonably close to its unknown
true value, would be useful information when formulating decisions. Insofar as frequentist statistical
methods do not directly provide this information (although the results of frequentist statistical analysis
are often loosely misinterpreted in this way) they
may be of limited use to the forest manager. Bayesian
methods on the other hand can provide precisely this
sort of information.

Bayesian inference allows direct computations of

the probability of a hypothesis being true, or of the
probability distribution of a parameter in a statistical
model. An essential difference from the frequentist
approach is that probabilities are assigned to possible
parameter values before the study is conducted.
These probabilities can incorporate knowledge
gained by the investigators from their own or others
data, their experience, and even, if so desired, their
intuition. However, because it is impossible to interpret such a probability as a long-run relative
frequency, a wider definition of probability known as
subjective probability must be adopted. While this definition includes the frequentist notion of
probability, it is now extended or broadened and
makes meaningful such questions as, What is the
probability that the Vancouver Canucks will win the
Stanley Cup in 2000? or What is the probability
that it will rain tomorrow? These questions have no
meaning in the frequentist sense (because the Stanley
Cup competition is played only once in 2000, and tomorrow is similarly unique), but certainly make
sense to the person in the street, whether sports fan,
gambler, or simply somebody who listens to a weather forecast.
In the Bayesian paradigm, this prior information is
described by a prior probability distribution for the
model parameters and reflects personal knowledge or
beliefs before the study is conducted. The data modify this distribution to produce a posterior probability
distribution, which describes the accumulated information. This relationship, known as Bayes theorem,2
can be written in equation form as:
posterior = (prior) (likelihood of data
given prior) (constant)

Although a Bayesian statistician can meaningfully

make statements such as: With posterior probability
0.95 the total volume of wood in the stand lies between 46 000 and 52 000 m3, or that the posterior
probability of a certain hypothesis being true is 0.08,
making such appealing conclusions bears a price
prior probabilities must be specified and the results
depend upon this specification.
Specifying the prior distribution can be challenging. The practitioner who may shudder at the
thought of determining a prior distribution, should
be reassured that the relative importance of the newly
collected data to that of the prior distribution in

2 An informal proof of Bayes theorem using the example in Section 7.3 is presented in Appendix 1.

90

(1)

determining the posterior distribution increases with

the sample size of the newly collected data. Thus,
when sufficiently large amounts of data are collected,
the prior distribution becomes relatively unimportant and the posterior distribution depends almost
exclusively on the data via the likelihood function.
Often a convenient prior distribution that reflects ignorance is the reference prior, which roughly speaking
assumes that all possible values of the parameter (reflecting all possible hypotheses) are equally probable.
Readable discussions of the fundamentals of
Bayesian statistics can be found in, for example,
Wonnacott and Wonnacott (1977, Chap. 19),
Kennedy (1985, Chap. 12), Berry (1996), and Ellison
(1996). More complete and mathematical discussions
of Bayesian methods (in rough order of increasing
mathematical difficulty) can be found in, for example, Cox and Hinkley (1982), Gelman et al. (1995),
and Box and Tiao (1973).
The following example will help to show the
differences between the frequentist and Bayesian
paradigms and how the parts of the Bayesian
approach work together.
7.2.1 A silviculture example
For planning purposes, suppose that a district silviculturist is interested in knowing if a recently logged
cutblock will naturally regenerate within the next
2 years. If it does regenerate then treeplanting would
not be necessary and resources could be directed
elsewhere. In 2 years time, the cutblock will be sampled according to a standard protocol and the results
used in a decision rule to determine if planting is
necessary.
Suppose quite a bit of information is already available on other, older cutblocks, which are similar in
all relevant respects and thus belong to the same stratum.3 The silviculturist might apply this information
to plan the use of available resources. This process
could be done informally, by using professional
judgement, or formally, by developing a statistical
model to predict how likely it is that planting would
be necessary. The Bayesian would develop the statistical model and call it the prior distribution for the
cutblocks density in 2 years time. Regardless of its
name, the model could be used to calculate the probability that planting will be necessary at that time.
The standard frequentist approach would, at least,
appear to ignore this prior information when the

cutblock is sampled in 2 years time. For instance, the

protocol might be to calculate the 90% confidence
limits around the density estimate of the number of
stems per hectare (stems/ha) in the cutblock. If the
lower confidence limit was below a minimum stocking standard (MSS) then planting would proceed.
Formally, this decision rule does not use the prior
distribution or any other ancillary information. The
silviculturists final decision may, of course, be based
on many factors in addition to the numerical decision rule results.
In contrast, the Bayesian would explicitly include
the information contained in the prior distribution
by combining it with the collected data via Bayes
theorem to produce a posterior distribution for the
estimate of the cutblocks density. This result could
then be used to calculate a posterior probability interval (also called a credibility interval). While the
fundamental concepts underlying this interval are
quite different from those underlying the confidence
interval, this credibility interval can be used in a similar fashion. Thus a similar decision rule could be
used, namely, if the lower 90% posterior probability
interval is above the MSS then conclude that the cutblock is satisfactorily restocked and does not need
planting.
Both statistical approaches will sample the cutblock in 2 years, and use the data to decide if planting
is necessary. Formally, this sample data will stand on
its own for the frequentist while the Bayesian will use
all the datanot only the newly collected data but
also that collected earlier.
For this example, the Bayesian posterior probability interval may be shorter than the frequentists
confidence interval because of the inclusion of prior
information (and effective increase in sample size).
The mean of the posterior distribution (called the
posterior mean) is a weighted average of the prior and
observed means. If these values are not too different,
then the posterior mean will be similar to that of the
frequentist. On the other hand, if the observed mean
is quite different from that of the prior distribution,
then the posterior mean is shifted from the observed
mean towards the prior mean. The amount of shift
depends upon the ratio of the variance of the observed mean to that of the prior distribution. If the
observed mean is well known (as indicated by a small
standard error) then the shift will be small. On the
other hand, if the standard error is relatively large,

3 This sort of stratum is known as a working group within the B.C. Ministry of Forests.

91

then the shift could be substantial because the data

have not added much information to the posterior
distribution.
7.2.2 Interpreting confidence limits and P-values
This section discusses the correct interpretation of
confidence intervals and P-values. Rarely do these
tools directly answer the research questions under
consideration, but they are often incorrectly interpreted as if they do.
To illustrate, let us consider a simple estimation
problem in which a forester wishes to estimate the
total volume of wood in a cutblock. From the sample
data obtained, a 95% confidence interval is calculated
for the total volume. Let us suppose that this confidence interval ranges from 16 000 to 24 000 m3. The
correct interpretation of this interval is that if the
forester were to sample again, and again and again,
each time calculating a 95% confidence interval, 95
out of every 100 calculated intervals would contain
the true total volume in the long run. Or, put another
way, in the long run only 5% of all 95% confidence
intervals (assuming everything else is correct) would
miss the markthat is, not contain the parameter
(here, the total volume) being estimated.
To the manager reading the foresters report which
has a 95% confidence interval of 16 00024 000 m3
for the total volume, it seems much the same as saying that with probability 0.95 the total volume lies
between 16 000 and 24 000 m3. Certainly most
users (including some statisticians) would treat the
result in such a fashion. However, under the frequentist paradigm, probability is the long-run relative
frequency of an event in many trials. In this example,
the parameter, total volume of wood, is a fixed but
unknown quantity, not a random variable. The parameter is either captured or not captured in the
confidence interval. Because a relative frequency
could never relate to the parameter, a researcher cannot correctly talk about probabilities concerning the
parameter. Nonetheless, confidence intervals will inevitably be wrongly interpreted in this way. Most
non-specialists would be hard pressed to distinguish
between a statement with 95% confidence and one
with 95% probability of being correct.
The same is true of P-values. For example, a computed P-value of 0.03 will often be misinterpreted as
the probability that the null hypothesis is true. In
fact, this P-value should be interpreted as the probability (in the frequentist sense) of obtaining
observations as discrepant or more discrepant with

92

the null hypothesis, than those actually observed, if

in fact the null hypothesis is true. In other words, if
the null hypothesis were true and the experiment
were repeated many times, the results of only 3% of
the experiments would be as great or greater than the
discrepancy observed in the actual experiment. A
mental shorthand for this is to think of the P-value
as a measure of evidence against the null hypothesis
(with small P corresponding to strong evidence).
While these common tools are useful, they only indirectly answer most research questions. In contrast,
Bayesian methods often directly answer such questions.
7.3 The Fundamentals of Bayesian Statistics
This section will briefly describe Bayesian statistical
methodology, and then use a simple numerical example to demonstrate the steps involved.
7.3.1 Components of a Bayesian analysis
To make statistical comparisons or estimations, a
probabilistic model must be developed to describe
how the observations are generated. As in frequentist
statistics, this model will involve unknowns called
parameters. While frequentist methods treat these
unknowns as fixed, the Bayesian approaches these
unknowns as random variables. Thus, in addition to
the probability model, Bayesian inferential methods
require probability distributions for the unknown
parameters in that model. A Bayesian analysis requires three groups of probability distributions:
1. The first is the prior probability distribution for
the unknown parameters in the data probability
model. These parameters are developed before any
data are collected and describe what the researcher
might observe when the data are collected. This
prior distribution may be developed: (1) empirically, by using previously collected data (noncontroversial but technically more demanding and
known as empirical Bayes); or (2) subjectively, by
carefully thinking about the situation and the science involved to develop a model that captures
what the researcher might reasonably expect will
be observed. When very little is known beforehand, then noninformative or reference prior distributions are used. These are convenient and give
all plausible values of the unknown parameters a
reasonable weighting so that the chosen prior distribution has little influence on the final outcome.

2. The second group of distributions form the

probability model for the data collected from the
research study. Along with the data, these distributions are used to develop the likelihood. This
information is then fed into Bayes theorem along
with the prior distribution to generate the third
component.

. Numbers of previously sampled plots (observed

as US or S) from both NSR and SR cutblocks. An
average of 10 plots was placed in each cutblock.

Cutblock status

US

Total

3. The posterior probability distribution is the output

or outcome of a Bayesian analysis and summarizes
what is known. It can be used to develop credibility intervals and probabilities for interesting values
of the model parameters.

NSR
SR
Total

672
72
744

168
88
256

840
160
1000

The relationship between these components of a

Bayesian analysis is shown in Figure 7.1.

Prior
distribution
Bayes
theorem

Stocking status of plots

100 cutblocks to build this probability model and so

eventually develop the posterior distribution.
Suppose that an average of 10 plots was placed in
each cutblock and that each plot was classified as understocked (denoted by US) or stocked (denoted by
S), with the results shown in Table 7.1. These data are
also shown pictorially in Figure 7.2. For this simple

Posterior
distribution
672 plots
observed to be
understocked

Data

. Components of a Bayesian analysis.

7.3.2 Example with single sample plot

For illustration purposes, let us rework the previous
silviculture example. Recall that the silviculturist is
interested in determining whether the cutblock will
be satisfactorily restocked (SR) or not satisfactorily
restocked (NSR). This measurement will determine if
planting will be necessary. Suppose that this particular stratum has 100 previously studied cutblocks; of
those, 16 were SR (16%) while 84 were NSR (84%).4
This information provides the prior probability distribution,5 namely the probability that the new
cutblock is NSR is 0.84. Thus, without sampling the
new cutblock, the odds are 84 to 16 or, about 5 to 1,
that it will need planting.
Suppose that we put just one plot into this cutblock, and determine that it is understocked (i.e., the
observed number of well-spaced trees in that plot is
less than the MSS). These data are used with the probability model to develop the likelihood. We can use
more of the previously collected data on the similar

840 plots
from NSR
cutblocks

168 plots
observed to
be stocked
1000
plots
72 plots
observed to be
understocked
160 plots
from SR
cutblocks
88 plots
observed to
be stocked

. Distribution of sample plots for the silviculture

example.

4 Assume that this information is known without error.

5 Assuming, of course, that this new cutblock is correctly identified as belonging to this stratum.

93

example, we can actually skip the development of the

likelihood and directly obtain the posterior distribution from Table 7.1. Notice that 744 of all the
observed plots were found to be US. Of those, 672
were sampled from NSR cutblocks. Therefore, the
posterior probability that the current cutblock is NSR
given that the one observed plot is US is

state of nature and is denoted by p(X |) (the vertical

bar | means given). This symbol is read as the
probability of observing the data, X, given (a specific
value of) the parameter, . In this case, two such
models are usedone for each state of nature, . For
a sample of one plot, both models follow a Bernoulli
distribution. From Table 7.1, we see that
672
p(X = US| = NSR) = = 0.80 ,
840

672
= 0.903.
744
With relevant prior information and the sampling
results of just one plot, the Bayesian approach allows
the determination of the posterior probability that
the cutblock is NSR. On the other hand, the frequentist approach formally ignores the prior information
and so could do little with just one plot. In any case,
regardless of approach, it is unwise to decide the
management of the cutblock on the basis of one plot.
In Section 7.3.1, we will extend the methodology to
samples of several plots. For the rest of this section,
we will fill in the steps just skipped by developing the
necessary statistical notation, the probability model
for the data, and the posterior probability distribution.
The two possible true states of the cutblock can
be denoted by the random variable , which can have
one of two values, either NSR or SR. Based on
Table 7.1, the prior probability that the cutblock is
NSR is denoted by p( = NSR) = 0 = 0.84 while the
prior probability that the cutblock is SR is denoted by
p( = SR) = 10 = 0.16. This prior distribution has a
Bernoulli distribution with parameter 0 (a special
case of the binomial distribution when the sample
size is one).
The data, denoted by X, can have one of two values: either X = US or X = S. The probability model
provides the probability of the data given the true

while
72
p(X =US| = SR) = = 0.45.
160
These probabilities are parameters of the probability
models for the two states of nature and will be denoted by 1 and 2 , respectively. They are also
conditional probabilities because they are the probability that X = US given (or conditional on) the true
state of nature. Furthermore, these probabilities are
the likelihood6 functions, because they provide a measure of the likelihood of observing the data, X, given
specific values for the state of nature, . The model
parameters are summarized in Table 7.2 and in the
tree diagram in Figure 7.3.
A derivation of Bayes theorem using this example
is presented in Appendix 1. The conclusion from that
appendix is that the posterior probability of a particular state of nature, , given the data X is given by:
p(|X) = p() p(X|).
p(X)
For this example, the probability that the new cutblock is NSR given that the observed plot is US is:

Table 7.2 The probability parameters for p(), the prior distribution of (cutblock is NSR or SR), and
p(X|), the conditional probability of observing X, given (cutblock is NSR or SR). All values
were calculated from those in Table 7.1.

Conditional probabililty or likelihood, p(X|) for

Probability parameters
Cutblock status

Prior probability: p()

X = US

X=S

0 = 0.84
10 = 0.16

1 = 0.80
2 = 0.45

11 = 0.20
12 = 0.55

= NSR
= SR

6 Not all conditional probabilities are likelihoods.

94

States of nature, ,
for the cutblock

Data, X, of the one

observed plot

1 = 0.80

X = US

= NSR
0 = 0.84

11 = 0.20

2 = 0.45

10 = 0.16

X=S

X = US

= SR
12 = 0.55

X=S

. Probability tree for the silviculture example.

. The likelihoods (probability that X plots out of 12 are US given ) and the posterior probability that the cutblock
is NSR for all possible X values, when the prior probability, 0 = 0.84

X=
number of
US plots
observed

Likelihood that
plots are US
when = NSR
p(X | = NSR)

Likelihood that
plots are US
when = SR
p(X | = SR)

Posterior
probability
for = NSR
p( = NSR| X)

Cutblock is
most likely

Management
decision

0.000
0.000
0.001
0.003
0.016
0.073
0.277
0.652
0.902
0.978
0.995
0.999
1.000

SR
SR
SR
SR
SR
unclear
unclear
unclear
unclear
NSR
NSR
NSR
NSR

not plant
not plant
not plant
not plant
not plant
unclear
unclear
unclear
unclear
plant
plant
plant
plant

0
1
2
3
4
5
6
7
8
9
10
11
12

0.000
0.000
0.000
0.000
0.001
0.003
0.016
0.053
0.133
0.236
0.283
0.206
0.069

0.001
0.008
0.034
0.092
0.170
0.222
0.212
0.149
0.076
0.028
0.007
0.001
0.000

Total

1.000

1.000

95

posterior probability = p( = NSR|X = US)

0 1
=
0 1 +(10) 2
= 0.903

calculation determines the ratio of the posterior

probability that the cutblock is NSR (denoted by
p( = NSR|X)) to the posterior probability that the
cutblock is SR (denoted by p( = SR|X)), namely:
p( = NSR|X) p( = NSR) p(X| = NSR)
=
p( = SR|X)
p( = SR) p(X| = SR)

as earlier found directly from Table 7.1.

7.3.3 Example with several sampling plots
Making management plans by sampling one plot is
unrealistic. Therefore, we extend the Bayesian
methods for one plot to that of several plots so that
the cutblock can be properly sampled. The mathematical development for this case is given in
Appendix 2. Suppose a sample of n plots is placed in
the cutblock with the same prior probabilities as
shown in Table 7.1. We will now use X to represent
the number of plots observed to be understocked
(US). From Appendix 2, the posterior probability
that the cutblock is NSR given that X plots are observed to be US is:
p( = NSR| X) =

p( = NSR) p(X | = NSR)

.
p(X)

(2)

Table 7.3 gives the likelihoods and posterior probabilities that the cutblock is NSR for n = 12 sample
plots (using a protocol of one plot per hectare of cutblock to determine that 12 plots are required) and for
all possible X values (X = 0 to X = 12). The posterior
probabilities are calculated using equation (2), with
the prior probability being p( = NSR) = 0 = 0.84.
If four or fewer plots out of 12 are observed to be
US, then deciding that the cutblock is SR seems clear
because the posterior probability is less than 0.05.
Also, if eight or more plots are observed to be US
then the NSR decision is clear because the posterior
probability is greater than 0.95. But when 5, 6, or 7
out of 12 plots are observed to be US then the management decision is not clear. These results depend
on the model probability values of 1 and 2. If their
values had been more widely separated then the unclear decision would have occurred for fewer
X-values; if their values had been more similar, then
the undecided decision would have occurred for
more X-values. Odds ratios provide another way to
express these results.
7.3.4 Odds ratios
Another way to look at these results is to calculate the
posterior odds that the cutblock is NSR ( = NSR)
given X plots observed to be understocked. This

96

=
=

(3)

p( = NSR) p(X| = NSR)

p(X| = SR)
p( = SR)

0
p(X| = NSR)
.

10
p(X| = SR)

The posterior odds is composed of two parts:

1. the ratio of the prior probabilities, or the prior odds
is
0
0.84
= = 5.25, and
10 0.16
2. the ratio of the two conditional probabilities
known as the Bayes Factor,
p(X| = NSR)
p(X| = SR)
For n = 1 and X = US, the Bayes factor is:
Bayes factor = BF =

p(X = US| = NSR)

1
=
2
p(X = US| = SR)

( )

0.80
= = 1.78
0.45
and so the posterior odds are:

1
0
p( = NSR|X)
=
2
(10)
p( = SR|X)

( )( )
( )( )

0.84 0.80 0.672

= = = 9.3.
0.16
0.45 0.072
This value means that, having observed that one
plot is US, the odds are about 9 to 1 that the cutblock
is NSR. If the observed plot had been S (stocked), the
odds would have been

(0.84) (0.20) = (5.25 0.364) = 1.91,

0.16
0.55
about 2 to 1 that the cutblock is NSR, or about 1 to 2
that the cutblock is SR. Notice how the data have
modified the prior odds of about 5 to 1.
For multiple sample plots, n = 12, with seven plots
(X = 7) found to be NSR, the prior odds remain at

5.25 while the Bayes factor will be (values from

Table 7.3) (0.053/0.149) = 0.36. Thus the posterior
odds is now (5.25 0.36) = 1.78, meaning that the
odds are about 9 to 5 that the cutblock is NSR.
A value of the Bayes factor greater than 1 indicates
that the data contain evidence to favour the hypothesis that the cutblock is NSR, while a value less than 1
indicates the opposite (evidence favours the hypothesis that the cutblock is SR).
A common reference prior (used, for instance, if no
data had been available to develop the prior probabilities) would give both possible states, NSR or SR,
equal probabilities so that the prior odds would be 1.
In this case, the Bayes factor would directly provide
the posterior odds. For further discussion, see Kass
and Raftery (1995).
7.3.5 Bayes factor for significance testing
The Bayes factor is a measure of the evidence for two
competing hypotheses, and can be directly used to
help choose one hypothesis over the other. Example
cutoff values for both are shown in Table 7.4 and
were taken from Ellison (1996), who quotes Kass and
Raftery (1995, Sec. 3.2).
. Suggested cutoff values for the Bayes factor (BF)
when comparing two hypotheses

BF

Evidence against H: Cutblock

is SR as opposed to NSR

01
13
3 20
20 150
> 150

Nothing to mention
Not worth more than a bare mention
Positive
Strong
Very strong

For the multiple sampling plot example in the last

section, the Bayes factor was 0.36, which suggests little evidence against the hypothesis that the cutblock
is SR. Alternatively, the Bayes factor for the hypothesis that the cutblock is NSR is 1/0.36 = 2.78, which,
while still low, suggests that the evidence is more
supportive of the hypothesis that the cutblock is SR
than NSR. The Bayes factor can be used similarly to
the P-values in hypothesis testing. An advantage of
Bayes factor is that it is not sensitive to sample sizes,

whereas frequentist P-values can be dramatically affected by unusually large or small sample sizes (Cox
and Hinkley 1974, Table 10.2; Ellison 1996).
7.4 Bayesian Decision Theory
Both the inferential problems of estimation and hypothesis testing can be viewed as problems in
decision theory, for which a complete Bayesian theory has been developed. However, Bayesian decision
theory can also be used in applied problems of decision-making when information is obtained through
experience and experimentation. For instance, the
natural regeneration example previously discussed
could be formulated as a Bayesian decision theory
problem, as could many other questions relating to
forest management.
The basic framework of decision theory assumes
a set of possible, but unknown, states of nature,
= {1, 2, }, and a set of possible actions
A = {a1, a2, } available to the decision-maker.7 If the
decision-maker chooses an action, a1, when the state
of nature is 1 then an incurred loss can be calculated
by a function denoted by L(1, a1). This loss could
also be written as a gain G(1, a1) = L(1, a1). For the
natural regeneration example, the set has two states
of nature: = {1= NSR, 2= SR}. The two possible
actions under consideration are A = {a1= plant, a2 =
not plant}. For illustration purposes,8 some arbitrary
numbers will be used for the gain function and are
presented in Table 7.5 and Figure 7.4. This figure
shows a simple decision-tree diagram often used in
decision analysis. This example will be used to illustrate the basic concepts in decision analysis, which
are developed in more detail by Peterman and Peters
(this volume, Chap. 8).
The decision-maker wants to keep losses as small
as possible or, alternately, the gains as high as possible. The difficulty lies in the fact that there is usually
not a unique action, a*, for which the gain is maximized for all states of nature, . For some states of
nature one action maximizes the gain, while for others a different action will provide a maximum. In
such cases, since the state of nature is unknown, an
unambiguously best action cannot be chosen. For
example, planting a site when it is sufficiently regenerated is a waste of resources and may require further
resources later, if, for instance, the stand is too dense

7 and A are names used to represent sets of things, which consist of the possible states of nature: 1, 2,..., and the possible actions a1, a2 ...,
respectively.
8 Although we have used the gain function when writing this section because of its more positive point of view, the literature mostly uses the
loss function.

97

Gain

State of nature

Management action

1= cutblock is NSR

a1

2 = cutblock is SR

1 0

a2

G (1,a1) = $200/ha

G (2,a1) = $1200/ha

1 = cutblock is NSR

G (1,a2) = $1800/ha

2 = cutblock is SR

1 0

G (2,a2) = $500/ha

. Decision tree for the silviculture example.

and thinning is required. On the other hand, not

planting the site when needed may mean that, at rotation (harvest), the stand produces much less
volume than it could have, resulting in a significant
loss in revenue.
The best action, called the Bayes action, minimizes
the expected loss, or Bayes loss, over all possible actions, a, with respect to the prior distribution. This
action is equivalent to maximizing the expected gain.
Table 7.5 shows some hypothetical gains for each action under each state of nature. Using the prior
probability, p(), to model the probability of a particular state of nature, we can calculate the expected
gains (or losses) for each combination of and a.
The Bayes gain (BG) for the first action (a1 = plant)
can be calculated by:
BG(a1) = 0G(1,a1) + (10)G(2, a1)
= 0.84 $200 + 0.16 (-$1200) = -$24/ha,
and for the second action (a2 = not plant):

BG(a2) = 0G(1,a2) + (10)G(2, a2)

= 0.84 (-$1800) + 0.16 $500 = -$1432/ha.
Since the Bayes gain is greatest (-$24/ha) for the action a1 (plant), then the recommended Bayes action
is to plant the cutblock.
So far, the decision has been based on the prior
distribution. We can use data to update the Bayes
gain to obtain a Bayes posterior gain. We can maximize this gain to choose the action, a*(X), from all
the possible actions A calculated for every possible
value of the data. Thus we would have an optimal decision rule (or policy) prescribing the optimal action
for any observed data value. This policy is known as
the Bayes decision rule, and can be shown to minimize what is known as the Bayes risk over all decision
rules that assign an action to every possible value of
the data, X.
Continuing the example, the Bayes posterior gain
can be calculated using the posterior probability,
p( = NSR|X), whose values are presented in Table 7.6.

. Hypothetical gains for each combination of action and state of nature

Possible action
State of nature

1 = NSR
2 = SR

98

a1 = plant
G(1,a1) = $200/ha
G(2,a1) = -$1200/ha

a2 = not plant
G(1,a2) = -$1800/ha
G(2,a2) = $500/ha

 . The Bayes posterior gain and posterior Bayes decision for all possible numbers of understocked plots

Number of
understocked
plots
observed

Posterior
probability
for = NSR
(p = p( = NSR|X))

Posterior gain
for action:
a1 = plant

Bayes
posterior gain
for action:
a2 = not plant

0.000
0.000
0.001
0.003
0.016
0.073
0.277
0.652
0.902
0.978
0.995
0.999
1.000

-1200
-1200
-1199
-1195
-1178
-1098
-812
-287
62
169
194
199
200

500
500
498
492
464
333
-137
-1000
-1574
-1750
-1790
-1798
-1800

0
1
2
3
4
5
6
7
8
9
10
11
12

Because this notation is cumbersome, we will use

p = p( = NSR|X) for the rest of the section. For the
observed data of 7 US plots out of 12, the Bayes posterior gain for the first action (a1 = plant) can be
calculated by
BG(a1) = pG(1, a1) + (1p)G(2, a1)
= 0.652 x $200 + 0.348 x (-$1200) = -$287/ha,
and for the second action (a2 = not plant):
BG(a2) = pG(1, a2) + (1p)G(2, a2)
= 0.652 (-$1800) + 0.348 $500
= -$1000/ha.

not plant
not plant
not plant
not plant
not plant
not plant
not plant
plant
plant
plant
plant
plant
plant

Prior
Bayes
decision
plant
plant
plant
plant
plant
plant
plant
plant
plant
plant
plant
plant
plant

The left-hand side is the posterior odds (see equation (2)). If it is greater than the ratio of gain
differences on the right-hand side, then planting will
be the Bayes decision. If this odds is less, then not
planting would be the Bayes decision. Thus the condition (equation (4)) can be expressed as: plant if
and only if the evidence for an NSR cutblock is sufficiently high. How high it has to be depends on the
prior odds, and on the anticipated gains under all
scenarios (via the right-hand side of equation(4)).
For our example, the ratio of gains is:
G(2, a2) G(2, a1) 500 (-1200) 1700
=
= = 0.85.
G(1, a1) G(1, a2) 200 (-1900) 2000

Given this data, the best action would be to plant

the cutblock if 7 out of 12 plots were observed to be
US. Bayes posterior gains have been calculated for
each value of X and the resulting Bayes decisions
presented in Table 7.6.
The action with the higher Bayes posterior gain
would be optimal, that is, it would be optimal to
plant (a1) if

p G(1,a1) + (1p ) G(2,a1) > p G(1,a2)

+ (1p ) G(2,a2),
or after rearranging:
P(X| 1) G(2, a2) G(2,a1)
p
0
= >

1 p 1 0 P(X |2) G(1, a1) G(1, a2).

Posterior
Bayes
decision

(4)

When the posterior odds are greater than 0.85 then

the decision is to plant. For the example, this occurs
for all X greater than 6. If the posterior odds is less
than 0.85 then the decision is to not plant. Note that
the posterior decision depends on the data while the
prior decision does not (because the prior odds was
5.25 > 0.85, the prior decision was to plant). Recall
that the prior Bayes decision was calculated previous
to any data collection and thus is constant for all possible data values.
This basic framework can be extended in many
ways. For example, in a sequential decision problem,
the decision-maker can decide at each step, either to:
(1) collect more data and defer choosing an action a
from A, or (2) stop data collection and choose an

99

action a from A. Sequentially at each stage in the

sampling, the decision-maker can make a choice
based on current data, or decide to defer choice and
collect more data. Details of this and other problems
in Bayesian decision theory can be found in Berger
(1985). More discussions on decision analysis are presented in Peterman and Peters (this volume, Chap. 8).
7.5 Bayesian Model Building and Prediction
Modelling is a common tool for simulating the underlying processes of a system. In forestry, for
example, models are developed to simulate tree
growth or timber and then predict tree volume in the
forests and timber supply. These predictions could be
one of the factors considered in setting forest management policy, so the reliability of these models is
very important. The development of models involves
statistical analysis to decide which factors are important, to choose how these factors should be
represented, and to validate the output of a model
against observed behaviour.
In classical statistics, given a certain set of data
(e.g., for each experimental unit there is a response y
and a set of regressor variables x1, x2, , xp) the first
step is usually to identify a plausible model, and then
use that model to answer the questions of interest to
the experimenter. For example, in a forestry study y
might be the volume of timber at a certain age, with
the x variables corresponding to species type, spacing
treatment, fertilization treatment, site index, site altitude, and site aspect for various test plots. First, some
variable selection technique would be used to decide
which regressor variables, with what transformations
and what interaction terms, should be included in
the model. After a model had been satisfactorily
identified, the analyst would address such questions
as the efficacy of spacing and fertilization treatments.
However, a possible weakness with this approach
is that the final inferences are all contingent on the
model selected. Several different models may all have
a similar degree of plausibility, which could yield
somewhat different predicted outcomes and lead to
different decisions about treatments. Which model
should you choose? This situation can be handled
quite easily in the Bayesian framework. Essentially,
prior probabilities are assigned to possible models,
and via Bayes theorem the data are used to obtain
posterior probabilities. Then many possible models
are usually eliminated by restricting attention to a
few with relatively high posterior probabilities.

100

Subsequent analysis can be carried out by averaging

(with posterior probability as weights) over the set of
plausible models. Thus the estimated effect of a certain silvicultural treatment would be obtained as a
weighted average of the estimated effect for a number
of different plausible models. Similarly, a predicted
volume for a certain combination of treatments and
site characteristics could be obtained using the posterior distributions of regression coefficients in each
plausible model, and averaging with the posterior
weights for the various models. That is, instead of
using a single model, prediction would be based on a
number of highly probable models. With adaptive
management, managers can design management actions as experiments to distinguish between plausible
models, thus improving future predictions, management decisions, and outcomes.
When the number of regressor variables is large,
numerous subset models may be generated, possibly
too many to handle even with modern computing
power. A number of methods have been proposed to
reject implausible models (described in Raftery 1994).
7.6 Conclusion
Bayesian methods provide an attractive alternative to
the frequentist methods of classical statistics. The
Bayesian approach systematically includes prior information in the analysis, thus better matching the
manner in which we learn. Another attraction is that
it permits direct computation of the probability of an
hypothesis being true, and the probability that an estimate of a parameter is reasonably close to the
unknown true value, hence aiding managers in decision-making. Bayesian methods also allow a
common-sense interpretation of statistical conclusions, instead of the rather convoluted frequentist
explanations of confidence levels and P-values. In recent years in applied statistics, interval estimation has
increasingly been emphasized over the use of hypothesis tests. This shift provides a strong impetus for
using Bayesian methods, because it seems highly unlikely that most users give confidence intervals
anything other than a common-sense Bayesian interpretation. Furthermore, where learning and
experimentation take place sequentiallyas occurs
in adaptive managementthe Bayesian approach
seems the natural way to update knowledge.

The basic steps in a Bayesian analysis are:

1. Setting up a full probability modela distribution
of all observable quantities conditional on the
parameters (unobservable quantities). The extra
specification that the Bayesian requires over the
frequentist is a prior distribution for the parameters of the data probability model. The frequentist
regards these parameters as simply unknown
quantities, whereas the Bayesian regards them as
random variables, and uses a probability distribution to reflect the current state of knowledge
concerning their value.
2. Obtaining a posterior distribution of the parameters by conditioning on the observed data (via
Bayes theorem). In other words, obtaining the
conditional probability distribution of the unobserved parameters, given the observed data.
3. While not discussed in this chapter, the fit of the
model can be evaluated by answering questions
such as: Does the model fit the data? and how do
the conclusions depend on the modelling assumptions in step 1? If necessary the model can be
revised, and the three steps repeated.
One of the strongest objections to Bayesian statistics is the requirement for a prior distribution.
However, with a sufficiently large amount of data,
the prior distribution becomes unimportant and the
posterior probability depends almost entirely on the
data. When data are scarce, all results, whether obtained by the frequentist or Bayesian methods,
should be interpreted with caution.
One of the central features of the Bayesian approach is that it permits a direct quantification of
uncertainty. This means that there are no impediments to fitting models with many parameters and
complicated probability specifications, except for the
practical ones of computing complicated multidimensional posterior distributions. However, recent
advances in computing power have greatly expanded
the possibilities in this area, leading to a remarkable
renaissance in Bayesian statistics. The recent book by
Gelman et al. (1995) provides an up-to-date exposition of the theoretical and practical aspects of
modern Bayesian methodology.
Forest managers must make sound management
decisions based on their knowledge of the system
being managed (the system may include the forest
ecosystem as well as economic and social elements)

and existing data. Bayesian methods provide a way of

explicitly integrating a managers accumulated
knowledge with experimental data in a statistical
analysis or decision-making process.
Acknowledgements
We would like to thank Brian Nyberg, Vera Sit, and
Brenda Taylor for involving us in this project. Furthermore, we benefitted from review comments from
Susan Chen, Darrell Errico, George Harper, Peter
Ott, Paul Reschler, William G. Roland, Vera Sit, Dr.
Tim Swartz, and Brenda Taylor. Figure 7.2 is borrowed from a diagram used by Judith Anderson in a
talk about probability. The figures were prepared by
Courtney Walls.
References
Berger, J.O. 1985. Statistical decision theory and
Bayesian analysis. 2nd ed. Springer-Verlag, New
York, N.Y.
Berger, J.O. and T. Selke. 1987. Testing a point hypothesis: the irreconcilability of P-values and
evidence. J. Am. Statist. Assoc. 82:11239.
Berry, D.A. 1996. Statistics: A Bayesian perspective.
Duxbury Press, Belmont, Calif.
Box, G.E.P. and G.C. Tiao. 1973. Bayesian inference
in statistical analysis. J. Wiley, New York, N.Y.
Cox, D.R. and D.V. Hinkley. 1982. Theoretical statistics. Chapman and Hall, New York, N.Y.
Reprint.
Dennis, B. 1996. Discussion: Should ecologists become Bayesians? Ecol. Applic. 6:1095103.
Dixon, P. and A.M. Ellison. 1996. Bayesian inference
- introduction: ecological applications of
Bayesian inference. Ecol. Applic. 6:10345.
Edwards, D. 1996. Comment: The first data analysis
should be journalistic. Ecol. Applic. 6:10904.
Ellison, A.M. 1996. An introduction to Bayesian inference for ecological research and
environmental decision-making. Ecol. Applic.
6:103646.
Gelman, A., J.B. Carlin, H.S. Stern, and D.B. Rubin.
1995. Bayesian data analysis. Chapman and
Hall, New York, N.Y.

101

Kass, R.E. and A.E. Raftery. 1995. Bayes factors. J.

Am. Statist. Assoc. 90:77395.
Kennedy, P. 1985. A guide to econometrics. 2nd ed.
MIT Press, Cambridge, Mass.
Ludwig, D. 1996. Uncertainty and the assessment of
extinction probabilities. Ecol. Applic.
6:106776.
Peterman, R.M. and C. Peters. [n.d.]. Decision analysis: taking uncertainties into account in forest
resource management. This volume.
Raftery, A.E. 1994. Bayesian model selection in social
research (with discussion). In Sociological
Methodology. P.V. Marsden (editor). Blackwell, Cambridge, Mass. pp. 11195. first ed.
Swindel, B.F. 1972. The Bayesian controversy. U.S.
Dep. Agric. For. Serv. Res. Pap. SE-95. Southeastern For. Exper. Sta., Asheville, N.C.
Wonnacott, T.H. and R.J. Wonnacott. 1977. Introductory statistics. 3rd ed. J. Wiley, New York,
N.Y.
References for Some Example Applications
Burk, T.E. and A.R. Ek. 1987. Prior information reduces the error in timber sale appraisals. N. J.
Appl. For. 4:3.
Ek, A.R. and J.N. Issos. 1978. Bayesian theory and
multi-resource inventory. In Proc. Integrated
inventories of renewable natural resources
workshop, Biometrics, U.S. Dep. Agric. For.
Serv., Gen. Tech. Rep. RM-55, pp. 2918.
Gertner, G. 1987. A procedure to evaluate sampling
schemes for improving already calibrated models. For. Sci. 33:63243.
Green, E.J., M. Kohl, and W.E. Strawderman. 1992.
Constrained empirical Bayes estimates for cell
values in a two-way table. Can. J. For. Res.
22:19837.
Green, E.J. and W.E. Strawderman. 1992. A comparison of hierarchical Bayes and empirical Bayes
methods with a forestry application. For. Sci.
38:35066.

102

Johnson, D.H. 1989. An empirical Bayes approach to

analyzing recurring animal surveys. Ecology
70:94552.
Taylor, B.L., P.R. Wade, R.A. Stehn, and J.F.
Cochrane. 1996. A Bayesian approach to classification criteria for spectacled eiders. Ecol.
Applic. 107789.
Ver Hoef, J.M. 1996. Parametric empirical Bayes
methods for ecological applications. Ecol. Applic. 104755.
Wolfson, L.J., J.B. Kadane, and M.J. Small. 1996.
Bayesian environmental policy decisions: two
case studies. Ecol. Applic. 105666.

Appendix 1: Demonstration of Bayes Theorem

using the silviculture example
This appendix uses the example described in Section
7.3 to demonstrate the validity of Bayes Theorem.
The relevant numbers for that example are summarized in Table A1.1
The data in Table A1.1 were previously used to calculate interesting probabilities such as the prior
probability that the cutblock is NSR:
840
p(=NSR) = 0 = = 0.84 ;
1000
the probability of observing an understocked plot
when the cutblock is NSR:
672
p(X=US|=NSR) = 1 = = 0.80 ;
840
and the probability of observing an understocked
plot when the cutblock is SR:
72
p(X=US|=SR) = 2 = = 0.45.
160
Note that there are other interesting probabilities
to calculate from Table A1.1 For instance, the probability that both X = US and = NSR occur is known
as the joint probability and is denoted by
672

p(=NSR, X=US) = 1000 = 0.672.

A general rule is that the joint probability is the product of the prior probability (a marginal probability
because it is calculated from the margins of the table)
and the conditional probability of the data, X, given
the true state of nature, denoted, mathematically
by:
(A1.1)
p( =NSR, X=US) = p( =NSR) p(X=US| =NSR).

From Table A1.1, p( = NSR) = 0 = 0.84 and

p(X=US| =NSR) = 1 = 0.80 so that their product is:
p( =NSR, X=US) = 0.840 0.80 = 0 1 = 0.672.
Because the order does not matter, equation (A1.1)
can be rewritten as:
(A1.2)
p( =NSR, X =US) = p(X = US) p( =NSR |X =US),
where p(X = US) is the probability that the one plot
will be found to be US, and p( =NSR|X=US) is the
probability that the cutblock is really NSR if the plot
is observed to be US. Notice that this last probability
is known as the posterior probability and is what we
want to determine from the sampling. It is the probability for a state of nature given our observed data.
Relation (A1.2) can be confirmed from Table A1.1 by
noting that p(X= US) = 744/1000 and that
p( =NSR | X=US) = 672/744 so that:
p( =NSR, X=US) = 0.744 0.80 = 0.672.
Equations (A1.1) and (A1.2) can be set equal to
each other:
p( =NSR) p(X=US | =NSR)
= p(X= US) p( =NSR | X=US).
This equation can be rearranged to obtain a relationship for the posterior probability:
(A1.3)
p( = NSR) p(X = US| = NSR)
p( =NSR | X=US) =
.
p(X = US)
In more general (and more readable) terms this
equation can be written as:
p() p(X|)
.
(A1.4)
p(|X) =
p(X)

. Numbers of previously sampled plots (observed as US or S) from both NSR and SR cutblocks. Parameters for the
prior probability distribution and the two probability models are also shown.

Probability parameters

Cutblock is

= NSR
= SR
Total

Joint distribution
(probability model parameters)
Prior probability:
p()

X = US

X=S

840 plots
(0 = 0.84)
160 plots
(10 = 0.16)

672 plots
(1 = 0.80)
72 plots
(2 = 0.45)

168 plots
(11 = 0.20)
88 plots
(12 = 0.55)

1000

744

256
103

This relationship, known as Bayes theorem, forms

the core of the Bayesian statistics methodology. We
can confirm that this relationship is true for the example by using the values from Table A1.1 to
calculate:
p( =NSR|X =US)
840/1000 672/840 672
=
= = 0.903.
744
744/1000
Note that this result agrees with that obtained directly from the first column of data in Table 1 and
discussed thereafter.
Equation (A1.3) can be written in words as:
(The posterior probability of a true state of nature
given the data) = (the prior probability of that true
state of nature) times (the likelihood of the observed
data given that true state of nature) divided by (the
probability of observing the data).
Notice that this definition is a more detailed version of equation (1) in Section 7.2. The relationship
between the components of Bayesian statistics was
presented pictorially in Figure 7.1.
In general, the denominator in equation (A1.3), the
marginal probability p(X= US), can be calculated by
summing all the possible values for the numerator of
equation (A1.3). For the example, this calculation is:
p(X = US) = p( = NSR) p(X=US| =NSR)
+ p(=SR) p(X=US|=SR)

Appendix 2: Calculations using several sampling

plots for the cutblock
In this appendix we will calculate the posterior
probability that the cutblock is NSR (not satisfactorily restocked) given that it has been sampled with
several plots. While the probabilities remain the same
(0 =0.84, 1 = 0.80, and 2 = 0.45), the probability
models for the data [p(X = US| = NSR) and
p(X = US| = SR)] are now more complicated.
The number of plots observed to be US will be denoted by X, with n representing the number of plots
sampled. The probability of observing X out of n
plots given a specific state of nature, , is given by the
binomial9 distribution:
n
p(X| = NSR) = X 1x (11) (nX) , and

()
n
p(X| = SR) = (X )

(12) (nX).

(A2.1)

Suppose that 12 plots were placed in the cutblock

and that 7 of them were found to be US. Then the
conditional probability for the observed data are:
12
p(X = 7| = NSR) = 7 0.807 (1 0.80)5 = 0.053
and
12
p(X = 7| = SR) = 7 0.457 (1 0.45)5= 0.149
We can use equation A1.4 (in Appendix 1) to calculate the posterior probabilities. The denominator
now becomes

( )

( )

p(X) = 0 p(X= 7| =NSR) + (10) p(X =7|=SR),

or

which is calculated by:

p(X = US) = 0 1 + (10) 2,

p(X) = 0.84 0.053 + 0.16 0.149 = 0.0685.

so that numerically,
p(X = US) = 0.84 0.80 + 0.16 0.45 = 0.744.
Thus, for the example, equation (A1.3) can be written as:
posterior probability = p( = NSR|X = US)
=

0 1
0 1 + (10) 2.

Thus the posterior probability that the cutblock is

NSR ( = NSR) given that 7 of the 12 plots were US is:
p(X = 7|)
(A2.2)
p( = NSR|X = 7) = 0
p(X)
0.84 0.053
=
= 0.652
0.0685
The posterior probability that the cutblock is SR is:
p( = SR|X = 7) = 1 0.652 = 0.348.
The posterior probabilities, p( = NSR|X) for all possible values of X, are shown in Table 7.3.

n!
n
n
9 This distribution is described in most standard introductory statistical textbooks. X is known as the binomial coefficient and X = .
X!(n X)!
If X = 7 and n = 12 then this is equal to 792. When n = 1, the binomial distribution becomes the Bernouilli.

()

104

()

8 DECISION ANALYSIS: TAKING UNCERTAINTIES INTO ACCOUNT IN FOREST RESOURCE MANAGEMENT

RANDALL M. PETERMAN AND CALVIN N. PETERS
Abstract

For forest resource managers, uncertainties are unavoidable because of natural ecological variability and
our imperfect knowledge of ecosystems. Nevertheless,
management decisions must be made and actions
must be taken. Decision analysis, a quantitative
method of evaluating management options, can
greatly assist that decision-making process because it
explicitly uses information on uncertainties. Although
widely used in business, decision analysis is particularly
useful for forest management because it accounts for
uncertainty about states of nature (e.g., current timber volume per hectare, the slope of the relationship
between survival rate of a rare bird species and size of
patches of mature stands of trees). Decision analysis,
in conjunction with Bayesian statistical methods, permits calculation of the potential outcomes of
management actions, considering each hypothesized
state of nature weighted by its probability of occurrence. Given a clear objective, managers can then
rank their management options. A sensitivity analysis
can determine how sensitive this ranked order of management options is to different assumptions or
parameter values. Sensitivity analysis can also identify
research priorities and help resolve conflicts between
interest groups about objectives or beliefs about how
a forest ecosystem works. Decision analysis is particularly appropriate for the planning stage of an active
adaptive management initiative because it can compare the expected performance of different proposed
experimental plans, taking into account various uncertainties. This procedure can help identify the best
experimental design for an adaptive management
plan, as well as its associated monitoring program.

8.1 Introduction
As noted in Nyberg (this volume, Chap. 1) uncertainties are pervasive in natural resource management.
Our knowledge of ecosystems is incomplete and imperfect, which creates imprecision and bias in data
used to quantitatively describe the dynamics of these
systems. Despite the presence of these uncertainties,
decisions must be made and regulations must be developed. One purpose of this chapter is to discuss
why it is important for decision-makers to explicitly

consider uncertainties when evaluating possible

management actions, including different designs of
adaptive management plans or monitoring programs. Another purpose is to describe decision
analysis, a formal, quantitative method that helps decision-makers take uncertainties into account in
analyses of options by breaking down the decision
problem into tractable components. Several examples
will illustrate the benefits and limitations of decision
analysis.
8.2 Sources of Uncertainty
Several sources of uncertainties exist in management
of forest ecosystems. First, natural variability over
space and time is inherent in ecological processes.
For example, growth rates of trees and animals may
differ among sites, years, and individuals. Such natural variability makes it difficult to forecast responses
of ecological systems to different management actions with accuracy or precision. Variability in
human behaviour also makes it difficult to forecast
how human harvesters and industry will respond to
management regulations. Second, further uncertainty
exists in data because sampling techniques imperfectly estimate quantities such as density of a certain bird
species in a forest, volume of merchantable timber
present per hectare, or natural mortality and reproductive rates of mammals. These methods thus create
further imprecision and bias in estimates of quantities that vary naturally. Therefore, managers will
forecast imperfectly, making it more difficult to
achieve a given management objective. Third, management objectives are frequently uncertain, either
because they are not well defined or because they
change over time. These uncertainties create complications for managers who try to choose the best
management option. The challenge for resource
managers is how to fully account for the type, direction, and magnitude of uncertainties when making
management decisions. One purpose of this chapter
is to address this challenge.
Forest managers must recognize that they are not
alone in dealing with uncertain systems; uncertainties
are present in all natural systems, not just biological
ones. For example, we now take for granted the
values of several fundamental physical constants such

105

as the speed of light and the charge of an electron,

but their estimated values have changed dramatically
over time as new experimental techniques emerged
(Figure 8.1). It is unsettling to note that several estimates of these physical constants even changed to
values well outside the confidence intervals of the
previous estimate! Uncertainties due to such measurement biases and errors are likely to be even more
pronounced in ecological systems that are relatively
variable and complex. Thus, scientists and managers
should expect to estimate with error quantities such
as the volume of merchantable timber per hectare,
abundance of a particular species of cavity-nesting
bird, proportion of seedlings surviving per hectare
per year, or offspring produced per female mule deer
per year. Even if scientists and managers recognize
and admit that uncertainties exist, they should not be
overconfident about the accuracy or precision of estimated quantities. Because of uncertainties, they
cannot expect to know the right answer, but should
be prepared to use the best estimates along with explicit measures of their uncertainty.
Ecological uncertainties create the potential for
making incorrect decisions because they prevent
managers from exactly predicting the outcome of a
particular management action. When incorrect decisions are made, losses result. In decision theory, a
loss is defined as an undesirable outcome of a decision. Losses can be direct losses, such as the
elimination of some rare or important species of bird
or mammal. Incorrect decisions can also result in
opportunity losses when the outcome of the decision
is worse than what could have been obtained if the
correct decision had been made. For example, an opportunity loss is incurred when a particular thinning
regime results in lower net timber revenues than
those that could have been generated if a different
thinning regime had been implemented. The probability of incurring such losses depends on the degree
and type of uncertainty arising from the sources discussed above. Decision theorists define the term
risk as expected loss, which is the weighted average loss. This quantity is calculated by multiplying
each possible magnitude of loss by a weighting term,
which is its probability of occurrence (Morgan and
Henrion 1990). To minimize such risks for users as
well as management agencies, both scientists and decision-makers should systematically and
comprehensively take uncertainties into account.
However, this approach is not often taken, as we
discuss next.

106

8.3 Approaches to Making Decisions in the Presence of Uncertainty

Management agencies have historically used several,
often ineffective, approaches to making decisions in
the presence of uncertainties.
8.3.1 Best estimate approach
One common approach to managing wildlife, forests,
and fisheries is to ignore the uncertainties and base
management decisions only on the best estimates of
all parameters and other quantities. For example,
into the 1970s, allowable annual cut (AAC) in British
Columbia was calculated with a formula using only
the best estimates of parameters, without taking uncertainties into account (Pearse 1976). The problem
with this approach is that incorrect estimates can lead
managers to make incorrect or suboptimal decisions.
Nevertheless, this focus on using the best point estimates is very common, especially where admitting
uncertainty would provide user groups with leverage
to argue for increased harvests or decreased protection of non-timber values such as wildlife. To avoid
such debates, managers sometimes request that scientists only provide them with their best point
estimates, even though everyone is aware that uncertainties exist.
8.3.2 Qualitative approach
A second approach to making decisions takes uncertainties into account, but only qualitatively or
crudely, rather than rigorously. This approach is
manifested in four ways:
1. First are cases where managers use ecological uncertainties to justify maintaining the status quo.
For instance, in 1991 the Forest Resources Commission recommended that the Allowable Annual
Cut of Timber Supply Areas or Tree Farm Licenses
not be raised or lowered until and unless new timber inventory data and subsequent yield analysis
clearly justify an adjustment, except in those obvious cases where current information strongly
support a change (Peel 1991, p. 84, recommendation #87). In other words, the default is to
maintain the status quo until uncertainties are
clarified to the point where a change in AAC is
clearly indicated.
2. Some people have used a qualitative approach to
justify extreme pessimism about the response to a
management action. For example, the public

Deviation from 1973 estimate (ppm)

(a) Electron charge

50
0
-50
-100
-150
-200
-250
1950

1955

1960

1965

1970

1975

1970

1975

1970

1975

Year of estimate

Deviation from 1973 estimate (ppm)

(b) Plancks constant

50
0
-50
-100
-150
-200
-250
1950

1955

1960

1965

Deviation from 1973 estimate (ppm)

Year of estimate

(c) Inverse ne structure constant

25
20
15
10
5
0
-5
1950

1955

1960

1965

Year of estimate

. Changes in estimates of various physical constants as new experimental or measurement methods were
developed. Data points are mean estimates of physical constants and vertical bars represent standard errors of
the mean estimates. All values are in units of deviations from their 1973 estimates, in parts per million (ppm).
(Adapted from Henrion and Fischhoff 1986.)

107

opposition to plans for spraying for gypsy moth in

New Westminster, B.C., arose partly because the
effects of spraying on public health were uncertain.
3. Uncertainties can also be used qualitatively to justify a moderately pessimistic outlook and to
implement a conservative approach to management. For example, when choosing the density of
lodgepole pine seedlings to replant, the density is
often increased by some arbitrary amount to allow
for uncertainty in tree mortality due to attack by
pests (Errico 1989).
4. Finally, resource users or managers may use uncertainties qualitatively to justify an optimistic view of
how systems will respond to management. Many
cases in forestry and in fisheries indicate that industry has used uncertainties in this way to
promote increased harvests or reduced protection
of the resource. For example, harvest rates of
forests have been increased in some areas in part
based on optimistic predictions about the future
effects of thinning and other enhanced silvicultural
activities on timber supply.
These four qualitative approaches to considering
uncertainty in decision-making may result in either
unnecessarily restrictive or excessively lenient policies
because the effects of uncertainties on the outcomes
of management actions are not considered quantitatively and explicitly.
8.3.3 Quantitative approach
A third and better approach to making decisions is to
take uncertainties into account quantitatively by considering a range of possible responses of an ecological
system to each management action. In doing so,
managers can select the option that minimizes the
risk. For example, several possible growth responses
of trees to specific amounts of thinning could be
explicitly considered to reflect uncertainty in that
response, rather than choosing just one of those
responses as the sole basis for the decision.
Several methods are available for taking uncertainties into account quantitatively, including decision
analysis, Monte Carlo simulation, and formal optimization techniques. Decision analysis was developed
in the 1960s in business (Raiffa 1968) to help decision-making in the presence of economic
uncertainties, which is directly analogous to making
decisions in the presence of ecological uncertainties.

108

Decision analysis is becoming a popular tool in resource management (e.g., Lord 1976; Walters 1981,
1986; Cohan et al. 1984; Parkhurst 1984; Bergh and
Butterworth 1987; Holbert and Johnson 1989; Parma
and Deriso 1990; McAllister and Peterman 1992a; McDaniels 1992; Thompson 1992; Hilborn et al. 1994;
Maguire and Boiney 1994; Reckhow 1994; Adkison
and Peterman 1996). This popularity is due to several
reasons.
First, most problems in resource management are
too complex (with lags, nonlinearities, threshold
phenomena, and cumulative effects) to permit the
use of formal optimization techniques (see Clark
1990 for some exceptions). Second, decision analysis
can help managers rank proposed management actions based on quantitative assessments of
probabilities of uncertain events and the desirability
of possible outcomes (Keeney 1982; Howard 1988;
Clemen 1996). Decision analysis can be thought of as
one type of risk assessment in that it considers the
uncertainties that create risks. Although decision
analysis cannot guarantee that a correct decision will
be made each time, it will improve the quality of several similar decisions over the long term because it
explicitly takes uncertainties into account quantitatively (Von Winterfeldt and Edwards 1986). Similarly,
taking the optimal action identified by a decision
analysis does not guarantee a certain desired outcome, but it increases the probability of a desirable
outcome occurring. Finally, decision analysis can
combine Bayesian statistical analysis and stochastic
models (Monte Carlo simulations) into a structured,
systematic approach to making decisions. Complex
decision problems are broken down into smaller and
more manageable components; these components
are then recombined to determine the optimal action. This process makes decision analysis a useful
tool for decisions involving complex ecological and
human responses to management actions, which certainly characterize forest management.
8.4 Eight Components of Decision Analysis
To make a complex decision problem in forestry
more tractable, decision analysis breaks the problem
down into eight components:
1. management objectives;
2. management options;
3. uncertain states of nature;
4. probabilities on the uncertain states of nature;

 . A generalized decision table showing calculation of expected outcomes for two potential management
actions, given two possible states of nature (Hypothesis 1 and 2) with their associated probabilities (P1 and
P2). Compare with Figure 8.2.

Hypotheses or uncertain
states of nature

Probabilities

Potential
management action #1

Potential management
action #2

Hypothesis 1

Probability that
Hypothesis 1 is
correct (P1)

Consequence of action 1
if Hypothesis 1 is
correct (C11)

Consequence of action 2
if Hypothesis 1 is
correct (C21)

Hypothesis 2

Probability that
Hypothesis 2 is
correct (P2)

Consequence of action 1
if Hypothesis 2 is
correct (C12)

Consequence of action 2
if Hypothesis 2 is
correct (C22)

Expected consequence
of action 1 =
(P1 C11)+(P2 C12)

Expected consequence
of action 2 =
(P1 C21)+(P2 C22)

5. model to calculate the outcome of each management action for each state of nature;
6. decision tree or decision table;
7. ranking of management actions; and
8. sensitivity analyses.
A generalized decision table (e.g., Table 8.1) can be
used to structure the decision analysis of simple
problems. In this table, two alternative management
actions are listed across columns and alternative hypotheses or uncertain states of nature, with their
associated probabilities (P1 and P2), are placed in
rows. For each combination of action and hypothesis,

Management actions

Probabilities of
states of nature

the consequences or outcomes (C11, C12, etc.) are calculated using a model. The expected value of the
consequence for a particular management action
(last row) is then calculated from the weighted average of all possible consequences for that action,
where the weighting is the probability of the hypothesis that gives rise to each consequence.
For more complex problems, a decision tree can be
used to structure the analysis (Render and Stair 1988;
Clemen 1996). The generalized decision tree in Figure
8.2 corresponds to the decision table in Table 8.1. Alternative management actions in Figure 8.2 are
represented by branches emerging from a square
States of nature
or hypotheses
Hypothesis 1

P1

Outcomes or
consequences
C11

Action 1
P2

P1

Hypothesis 2

Hypothesis 1

C12

C21

Action 2
P2

Hypothesis 2

C22

. A simple example of a generalized decision tree showing two different management actions and two possible
states of nature (Hypothesis 1 and 2) with their associated probabilities (P1 and P2). The square at the left is the
decision node and the circles are chance nodes. The consequences associated with each combination of
management action, i, and state of nature, j, are designated Cij. This decision tree is the graphical equivalent of
the general decision table shown in Table 8.1.

109

decision node, and uncertain states of nature or hypotheses are represented as branches coming from
the circular chance nodes. The probability of each
uncertain state of nature is shown explicitly for each
state-of-nature branch. Outcomes or consequences
of each management action, given each state of nature, are shown on the right. Decision trees can
accommodate much more complexity than a decision table by including numerous branches and
uncertainty nodes.
We will use an application of decision analysis to
forest management in Tahoe National Forest, California (Cohan et al. 1984) to illustrate the eight
components of this method. The purpose of Cohan
et al.s particular decision analysis (referred to as the
Tahoe example) was to determine what treatment
should be applied before a prescribed burn on a
recently harvested forest site. Figure 8.3 shows the
decision tree for this problem; its components are
explained below.
8.4.1 Management objectives
Decision analysis requires a clearly defined management objective or goal so that the different
management actions can be ranked by how well they
are expected to attain the objective. The objective is
usually stated explicitly in terms of maximizing (or
minimizing) one or more quantitative measures of
performance (such as expected value of future timber
harvests). However, decision analysis can also accommodate situations in which the objective is to choose
an action that produces a performance measure, such
as abundance of some rare bird species, that is within
an acceptable range of values. In this case, actions
that do not lead to outcomes within this range can be
discarded, and some secondary criterion (such as
minimizing cost) can be used to choose from the remaining actions. As emphasized by Keeney (1992),
identifying objectives requires carefully applying various procedures to ensure, for instance, that
fundamental objectives are not confused with the
means needed to attain them.
In the Tahoe example (Figure 8.3), the management objective was to maximize the expected net
resource value of the forest following the prescribed
burn. That value took into account the value of the
timber harvested, as well as the cost of carrying out
the pre-burn treatment (if any), the cost of the prescribed broadcast burn, and the cost incurred from
an escaped fire (if one escaped).
In the case of British Columbias forests, manage-

110

ment objectives can involve timber value, recreational use, wildlife habitat, and quality of viewscapes in
various combinations and with various relative importances. For example, a primary management
objective in Clayoquot Sound is to maintain longterm productivity and natural diversity of the area.
Subgoals include maintaining watershed integrity, biological diversity, and cultural, scenic, recreational,
and tourism values (Scientific Panel for Sustainable
Forest Practices in Clayoquot Sound 1995).
8.4.2 Management options
Managers need to define a list of alternative actions
from which to choose the best option. Considerable
thought should be put into developing innovative
options, as well as into identifying feasible ones
(Keeney 1982).
The Tahoe prescribed burn problem has two alternative management actions. These alternatives are
shown in Figure 8.3 as two branches emerging from
the square decision node. One choice was to conduct the prescribed broadcast burn without any
pre-burn treatment of the site (burn only). The
other alternative was to pile up timber slash from the
clearcut harvest before the broadcast burn (YUM
and burn). This latter treatment, referred to as yarding unmerchantable material (YUM), incurs
additional costs but reduces the probability of fire escaping and increases the chances of a successful burn.
Cohan et al.s (1984) question was, Is YUM worth
the additional cost?
8.4.3 Uncertain states of nature
Uncertain states of nature are parameters or quantitative hypotheses that are treated explicitly as
uncertainties in an analysis, usually by considering a
range of values for one or more parameters in a
model (see Section 8.4.5). Such uncertain parameters
lead to a corresponding range of forecasts of outcomes of management actions. For instance, it may
be difficult to estimate the effect of different sizes of
leave patches in a retention harvesting strategy on
abundance of a bird population because of uncertainty about how the probability of blowdown is
affected by patch size (i.e., whether that probability is
a steeply rising function of patch size or a relatively
flat one). There is also considerable uncertainty
about the benefits of some requirements in the
British Columbia Forest Practices Code for meeting
objectives related to biodiversity or recreational use.
For example, it is unclear whether the survival rate of

juvenile coho salmon is greatly or only slightly affected by the width of riparian forest that the Code
requires to be left along stream banks.
Two major uncertainties in the Tahoe example
(Figure 8.3) involved the fire behaviour and the magnitude of costs associated with what Cohan et al.
(1984) referred to generally as problem fires.
Uncertainty in fire behaviour was represented by
defining three types of fires: a successful burn, a
problem burn, and an escaped fire. The second uncertainty was the cost of a problem burn (high,
intermediate, or low cost). These uncertain states of
nature are shown as branches emerging from circular
chance nodes in Figure 8.3.

Management actions

Treatments

8.4.4 Probabilities on the uncertain states of nature

In forest management, considerable uncertainty usually exists about states of nature such as those listed
in Section 8.4.3 because of short data series, natural
variability, and measurement error and bias. However, scientists and decision-makers need to state a
relative degree of belief, or probability, for these
different states of nature so that they can forecast the
expected outcome of each possible management action and determine its ranking. For example, wide
confidence limits on the slope of a relationship between survival rate of trees to a given age and initial
stocking (density of seedlings) can produce a range of
forecasts about future harvests from stands that are

States of nature and their associated probabilities

Fire behaviour

Cost of problems

Successful burn
0.899

Outcomes
Resource Treatment Problem/
value
cost
escape cost

Net
value

6 620

4 858

1 762

6 620

4 858

3 010

-1 248

6 620

4 858

1 400

362

6 620

4 858

700

1 062

Escaped re
0.001

6 620

4 858

40 000

-38 238

Successful burn
0.8485

6 620

4 550

2 070

6 620

4 550

3 360

-1 290

6 620

4 550

1 750

320

6 620

4 550

1 050

1 020

6 620

4 550

40 000

-37 930

High cost
YUM and
burn
$1 559

0.25
Problems
0.100

Int. cost
0.50
Low cost
0.25

High cost
0.25

$1 713

Burn only

Problems
0.150

Int. cost
0.50
Low cost
0.25

Escaped re
0.0015

. Decision tree for the example described in the text for the Tahoe National Forest. The management options
(treatments) are to burn only or YUM and burn; the latter refers to yarding unmerchantable material,
where the slash material from the logging operation is piled up before burning. Outcomes are costs in dollars for
a 14-acre site. The resulting expected net resource values for each management option are indicated next to the
option. See text for details. (Adapted from Cohan et al. 1984.)

111

in favour of the alternative hypothesis, HA, or it is

not, based on a comparison of the computed P-value
with the pre-determined . For two reasons, this dichotomous approach to describing the state of nature
(HO versus HA) is inappropriate to describe ecological
uncertainty in decision analyses. First, managers need
to consider several different HA estimates of the slope
as possible states of nature, not just H0 and a single
HA, because different slopes may have very different
implications for the selection of an initial density of
seedlings to replant (e.g., Figure 8.4). Second, the Pvalue resulting from a standard hypothesis test refers
to the probability of obtaining the test-statistic by
chance alone if the HO were true and does not state
the probability that HO or any other possible hypothesis is correct. Therefore, P-values do not provide the
decision analyst with the required probability for
even one state of nature (Berger and Berry 1988), let
alone several. Thus, the classical statistical approach
to hypothesis testing is not a useful framework for
characterizing ecological uncertainties as input to decision analyses as described here.

Loge volume per tree at age 100

replanted at a specific density. In this case, managers

need a probability for various slopes of that relationship to estimate the expected harvest levels for
different stocking densities.
Unfortunately, classical statistics do not provide
such probabilities. Most ecologists describe uncertainty in estimates of states of nature (e.g., slopes or
other parameters) with standard errors, confidence
limits, and coefficients of variation. They also routinely apply classical statistical inference methods to
test point null hypotheses. However, such procedures
are inadequate for decision-making for the following
reasons.
First, hypothesis tests are too restrictive for making
decisions in the presence of uncertainties because
they only provide information relevant to two states
of nature: the null hypothesis and a specified alternative. In hypothesis testing, a point null hypothesis,
HO, (e.g., the slope of the relationship between average volume per tree at a given age and initial density
= 0) is tested with data by some classical method
such as a t-test. The null hypothesis is either rejected

H0

best
estimate

?
Initial density
. Possible models for hypothetical data on average volume per tree at age 100 years as a function of initial
density. The solid line is the best-fit regression line; dashed lines represent other possible, but less likely,
hypotheses about the true underlying relationship, including the null hypothesis, HO , of no relationship.

112

For similar reasons, standard errors of parameter

estimates and 95% confidence limits are also not useful characterizations of uncertainties for making
decisions. Specifically, they do not indicate the
probability to be placed on different possible states of
nature, even though scientists commonly make this
misinterpretation about confidence intervals (e.g.,
see Sokal and Rohlf 1969, p. 142; Bergerud and Reed,
this volume, Chap. 7). Some readers might consider
using statistical power analysis for this purpose,
which is the probability of correctly rejecting HO
(using a particular method such as a t-test) when a
specific alternative hypothesis is true. While power
analysis is indeed useful for designing experiments
and monitoring programs that will reduce uncertainty (e.g., Peterman 1990a, 1990b; Osenberg et al. 1994;
Mapstone 1995), statistical power does not indicate
the probability that a given HA might be true in nature. Thus, statistical power analysis does not provide
sufficient information to decision-makers about the
relative degree of belief in alternative states of nature.
Instead, quantifying uncertainties in states of nature requires an assessment of the relative merits of
rival hypotheses in the light of observational or experimental data that bear upon them... (Edwards 1992,
p. 1; emphasis ours). Three techniques are available
to do this. First, long and detailed historical data sets
can provide information about the relative frequency
of events. For example, historical data on forest fires
can provide probabilities for different intensities and
sizes of forest fires in a specific region. Similarly, data
from stream gauges can quantify the probability of
different heights of streams. Unfortunately, such
lengthy continuous records are not common. Second, where data are inadequate, estimates of
probabilities for different states of nature can be
elicited from experts using various techniques, based
on their knowledge of the system under study (Morgan and Henrion 1990). Third, Bayesian statistical
analysis is appropriate where some but not enough
data are available to generate frequency distributions
as discussed in the first technique. Bayesian statistical
methods use these data to generate probabilities representing degrees of belief for different values of
parameters (see Bergerud and Reed, this volume,
Chap. 7). This approach uses Bayes theorem to calculate the posterior probability that a particular
hypothesis is correct, given the data and some prior
probability distribution of the hypotheses based either on other, independent data (when available) or
on expert opinion (Box and Tiao 1973; Press 1989).

For example, rather than merely considering the two

possibilities that the slope of the relationship between
volume per tree and initial density in Figure 8.4 is either significantly different from zero or not, a
Bayesian statistical analysis would use the observed
data to calculate probabilities for a wide range of
different slopes (Figure 8.5), each of which has different implications for the management decision of how
many seedlings to replant. Thus, the output of
Bayesian analyses provides exactly the information
required for a decision analysis: the probabilities associated with each of several uncertain states of
nature. Ellison (1996) presents an easy-to-read introduction to use of Bayesian statistics in applied
ecology; the Crome et al. (1996) paper in the same
issue applies Bayesian statistics to the problem of detecting effects of logging on forest-dwelling birds.
In the Tahoe example in Figure 8.3, forestry staff
provided estimates of probabilities of the three types
of burns and the three levels of costs associated with
problem fires. These probabilities appear on the
branches corresponding to the appropriate uncertain
state of nature (Figure 8.3). The probabilities placed
on the three types of burns were different for the two
management options (YUM and burn, or burn
only, because treating the site before burning increased the probability of a successful burn and
reduced the probability of either a problem burn or
an escaped fire. Managers estimated that the probabilities of high, intermediate, and low costs of a
problem burn (if one occurred) would be the same
for both management actions.
8.4.5 Model to calculate outcomes
Another key element of decision analysis is the model
used to calculate the consequences of each combination of a particular management action and each
possible state of nature. The bounds, assumptions,
and complexity of the model will depend on the
availability of data. Whatever type of model is used,
it must produce quantitative indicators of the consequences of each alternative management action, such
as revenue from timber or an index of bird diversity.
Those indicators must relate directly to the management objective stated in the first component of
decision analysis.
In the Tahoe example, the relative costs and timber revenues resulting from different treatments were
estimated using models of fire behaviour, fire effects,
and economics developed by timber management
staff in Tahoe National Forest. The forecast outcomes

113

of these models for each alternative management action and for each uncertain state of nature are shown
on the right side of Figure 8.3. The net resource value
is the resource value minus the treatment cost and
the cost of a problem fire or escaped fire. For example, the simulated net resource value of the YUM
and burn option, if a successful burn resulted, was
$1762 on this 14-acre site. For the same action, but assuming that a problem fire recurred that had high
costs, their simulated net resource value was -$1248.

principle unknowable, given the uncertainty.

The decision tree for the Tahoe problem
(Figure 8.3) illustrates this structure. For each management action (type of pre-burn treatment), each
possible behaviour of fire, and each possible level of
cost of problem fires, there is a resulting value of the
timber resource, a treatment cost, and a cost resulting from problem or escaped fires. The expected
value of each alternative action is the sum of the net
resource value for each state of nature, multiplied by
the probability of that state occurring. Thus the
YUM and burn alternative has an expected value

8.4.6 Decision tree or decision table

A decision tree or decision table provides a logical
framework for ranking the different management actions by combining the states of nature, their
probabilities of occurrence, and their outcomes.
These rankings are based on expected values of
outcomes, or weighted average outcomes, for each
action. That is, each outcome is weighted by the
probability assigned to the associated state of nature
(parameter value or hypothesis). Summing these
weighted outcomes for each management action
gives the expected value of that action. Thus, the expected value as defined by decision theorists
represents the weighted average quantity, not necessarily the specific value that you would expect to see
in the short term (Lindley 1985). The latter is in

EV = (0.899 1762) + (0.1 0.25 (-1248))

+ (0.1 0.5 362)
+ (0.1 0.25 1062)
+ (0.001 (-38 238))
= $1559.
By similar calculation, the expected value of the
option without a pre-burn treatment is $1713.
Although the probability of an escaped fire is very
low (0.001 or 0.0015), its cost could contribute significantly to the total expected net resource value of each
management option. This example shows how even
low-probability events may affect which action is optimal, if the costs of such events are large enough.

0.16

Posterior probability

0.12

0.08

0.04

0
-0.3

-0.22

-0.14

-0.06

0.02

Slope of loge (average volume per tree) vs initial density

. Posterior probabilities for different slopes of a linear model for the hypothetical data shown in Figure 8.4.
Posterior probabilities were calculated using Bayesian statistics. The best-fit line shown in Figure 8.4 has the
highest posterior probability, but other lines with different slopes also have reasonably high probabilities. These
probabilities can be used in a decision analysis to represent the relative degree of belief in the different slopes.

114

8.4.7 Ranking of management options

The management actions are ranked by applying the
management objectives identified in the first component of decision analysis. For instance, if the
objective is to maximize the expected value of outcomes, each management action can be ranked using
the calculations from the decision tree. In the Tahoe
example in Figure 8.3, the optimal action is to burn
only without treating the site beforehand; its expected value of $1713 was greater than the $1559 for the
other option. This burn only option maximizes the
expected net resource value, even though the probabilities of a problem fire or escaped fire are higher
with this alternative than with the YUM and burn.
By ranking management options in this way, decision analysis explicitly considers uncertainties by
taking into account the probability that different
states of nature may exist, as well as their effect on
the expected outcomes of each management action.
The optimal decision identified when uncertainties
are used in this manner is referred to as the Bayes
decision (Morgan and Henrion 1990).
8.4.8 Sensitivity analyses
Decision analysis provides only one possible answer
to a decision problem because the optimal decision
may depend on the assumptions made, the value of
various parameters in the model, the structural form
of relationships in the model, or the probabilities
placed on the states of nature. Therefore, managers
must also be given results of sensitivity analyses,
which directly show how the rank order of management actions (i.e., the best decision) is affected by
these assumptions. If such analyses show that a given
action is still optimal over a wide range of assumptions, then such assumptions can be deemed relatively unimportant and managers will be confident that
the recommended action is indeed the best one.
However, if the rank order of management actions is
sensitive to different assumptions, then more data
must be collected for that particular parameter or assumption. In this manner, a sensitivity analysis of a
decision analysis can identify future research priorities.
Although Cohan et al. (1984) did not conduct a
formal sensitivity analysis of their Tahoe example
shown in Figure 8.3, several parameters could affect
the optimal decision, including the additional costs
of performing the YUM treatment, the costs associated with an escaped fire, and the probability of an
escaped fire if the YUM treatment is not used.

To demonstrate sensitivity analysis, we calculated the

effect of the last parameter on the optimal decision
by repeating the decision analysis using several possible values of the probability of the fire escaping for
the burn only option. The parameter values investigated ranged from 0.001 to 0.009 (Figure 8.6).
Results show that the burn only option remained
the best option (i.e., generated the largest expected
dollar value of the resource) as long as the probability
of having an escaped fire was less than 0.0055. However, if that probability was actually 0.0055 or greater,
then the YUM and burn option became the one
with the largest expected dollar value of the resource.
Thus, over a certain range of this parameter value,
the decision that was optimal in the original baseline
case (burn only) remained optimal, but, outside of
that range, the optimal decision switched to YUM
and burn. Such results should be presented to experts to determine whether a value greater than or
equal to 0.0055 for the probability of an escaped fire
without YUM is within the realm of possibility. If
this range is not plausible, decision-makers can be
confident that uncertainty in this parameter does not
affect their decision. However, if a value in this range
is plausible, the value of this parameter in nature becomes important for decision-making, and high
priority should be placed on obtaining a better estimate of this probability.
Sensitivity analyses can also be used to show how
different management objectives may or may not affect the choice of the optimal decision. This is
particularly important when objectives include more
than just maximizing the expected value of timber
harvested. Diverse objectives of various stakeholder
groups are currently commonplace in forest management in British Columbia. For example, objectives in
the Kamloops Land and Resource Management Plan
also include protection of habitat, maintenance of
diverse recreational fishing opportunities, and conservation of Pacific salmon (Westland Resource
Group 1995). In this type of situation, a quantitative
sensitivity analysis using the method of decision
analysis can show how similar to one another objectives would have to be lead to the same management
option being chosen (e.g., Maguire and Boiney 1994).
In some cases, relatively little change in the objectives
of one or more interest groups may lead them to recommend the same action, thereby resolving a
conflict.

115

1750

Expected net value of the resource ($)

1700

YUM and burn

1650

Burn only

1600

1550

1500

1450

1400
0

0.001

0.002

0.003

0.004

0.005

0.006

0.007

0.008

0.009

Probability of an escaped re without YUM (yarding unmerchantable material)

. An example sensitivity analysis of Cohan et al.s (1984) decision analysis on the Tahoe burning example (Figure
8.3). Lines show the expected net dollar value of the resource for different probabilities of an escaped fire under
the burn only option (i.e., without YUM). The solid line represents the YUM and burn option; the dashed
line is for the burn only option. The best estimate provided by forestry staff of the probability of having a fire
escape under the burn only option was 0.0015 (i.e., 1.5 chances out of 1000), but there is uncertainty in this
estimate. The sensitivity analysis shows that the burn only option had the highest expected dollar value as
long as this probability was less than 0.0055. Above that value, the expected value of the YUM and burn
option was greater than that of the burn only option.

8.5 Application of Decision Analysis to Adaptive

Management
Because management of forests is an uncertain science, Walters (1986) argued that resource managers
should manage in an active adaptive manner. In
other words, they should carefully design management actions as experiments, just as laboratory
experiments or monitoring programs would be

116

designed before their implementation (Hairston [editor] 1989). Well-designed experiments generate
rigorous new information about the relative effectiveness of each action or about the different
hypotheses about biological processes. Acting adaptively will tend to reduce future uncertainties and
thereby improve future management (Peterman and
McAllister 1993).
If decision-makers take this approach, they must

be able to evaluate alternative management actions,

including different adaptive management plans,
based on the plans abilities to generate timely and
cost-effective information. The key question is,
which experimental design is the most appropriate?
Dozens of possible experimental designs could be
implemented, not all of which are going to be equally
informative, let alone feasible. If a suboptimal design
is chosen, the information may be too costly or may
not reduce uncertainties. Therefore, decision-makers
need some way to compare different experimental
designs and identify the one that is expected to maximize the benefits of adaptive management. Decision
analysis is an appropriate method to do this.
For instance, suppose that we are planning an experiment such as the ones currently investigating
silviculture techniques in British Columbia. Assume
that an objective is to maximize timber value. Different randomly selected plots can be thinned to
different densities 20 years after replanting to stimulate growth of remaining trees. However, many
possible arrangements of treatments exist (Table 8.2
shows only a few examples). The question is, which
of these arrangements should be used? Decision
analysis can help answer these questions by comparing the expected outcomes of different options,
taking uncertainties into account about the amount
of release to be experienced by trees in different densities at age 20. In this sense, decision analysis can
integrate several of the methods described in previous chapters (e.g., power analysis, sampling,
experimental design) and provides a structured way
to choose among many possible arrangements of experiments or adaptive management plans.

8.6 Other Examples of Decision Analysis

Many examples are available from fields within resource management where decision analysis or
similar methods of accounting for uncertainty have
been used to help structure a resource management
problem and to identify an optimal action. For instance, in addition to the Tahoe example of pre-burn
treatments previously described, Cohan et al. (1984)
presented several other cases that applied decision
analysis to fire management in U.S. National Forests.
In all cases, useful insights into prescribed burning
resulted from taking uncertainties into account and
breaking the complex problems into understandable
components. The authors also noted that using decision analysis to document the rationale for decisions
improved the rate of learning by management agencies. Managers involved with silviculture experiments
have also used decision analysis to compare the expected performance of different planting, thinning,
and fertilization strategies.
Stahl et al. (1994) evaluated different but very important questions for forest managers: what is the
optimal method for conducting forest inventories,
given that more precise methods cost more, and how
often should an inventory be done on a stand?
While the researchers used formal optimization
methods, their approach was structured much like a
decision analysis; they identified uncertainties about
the state of nature (current timber volume) when
comparing the effects of different inventories on the
expected value of net timber revenue (value of timber minus costs of harvesting and conducting
inventories). This analysis considered such uncertainties explicitly by assuming that each of three

. Some possible arrangements that could be considered for a thinning experiment. Each arrangement
consists of a different number of replicates at various densities of trees, which might be necessary because
of logistical constraints.

Number of replicate plots at each density

Density (stems/ha)

Option 1

Option 2

Option 3

250
500
750
Control (unthinned)

3
2
3
2

3
3
2
2

4
4
0
2

117

inventory methods would produce a probability distribution of estimates of timber volumes at any given
time. The inventory methods differed in cost (high,
medium, or low) and precision (high, medium, or
low).
Stahl et al. (1994) found that, in general, several inexpensive and less precise inventories taken only a
few times during the life of a stand resulted in a higher expected net income than a single, expensive but
very precise inventory. In addition, the authors concluded that precise inventory information was more
valuable when the potential losses in income due to
incorrect decisions were large. This conclusion is perhaps intuitive, but Stahl et al. were able to
quantitatively estimate the relative value of different
methods of doing forest inventories by explicitly considering uncertainties in information.
In wildlife management, Maguire (1986) used
decision analysis to recommend an appropriate conservation strategy for Whooping Crane populations
to minimize their probability of extinction. Maguire
evaluated whether it is better from a biodiversity
standpoint to create a single large population or
several small ones, given that random catastrophic
events can occur (a common debate in conservation
biology; see Simberloff and Abele 1976). In the
Whooping Crane situation, when Maguire (1986)
took the uncertainties associated with severe storms
into account, the optimal action was to move some
of the Whooping Cranes and create two separate
populations. This approach was better than keeping
them as a single population that had a higher probability of extinction if a rare severe storm occurred in
that one location.
Decision analysis has also been applied to complex
land use issues, such as the decision whether to preserve and/or mine in the Tatshenshini area of
wilderness in northwestern British Columbia (McDaniels 1992). There, the major uncertainties
included the environmental values associated with
preserving the area, the tax revenue to be generated
by mining, the question of whether mining would
actually go ahead given the regulatory process, and
other uncertainties. The analysis suggested that
preservation of the entire area as a park would have
the greatest expected value, taking into account the
nonmarket value of the wilderness.
Within the field of natural resources, decision
analysis has been used most widely in fisheries management. For instance, several authors have used

118

decision analysis to identify optimal management actions for Pacific salmon (e.g., Lord 1976; Walters 1981,
1986). Decision analysis was also able to identify the
optimal precautionary safety margin to apply to harvest rates of other marine fish species, given
uncertainties in stock abundance and productivity
(Frederick and Peterman 1995).
A final fisheries example from the northwestern
shelf of Australia (Sainsbury 1988, 1991; Sainsbury et
al. 1997) demonstrates particularly well how decision
analysis can be used in the design phase of an experimental, or active adaptive management program.
Foresters can learn considerably from this case study
because it is one of the few large-scale active adaptive
management experiments ever implemented, as well
as one of the few to use formal decision analysis in
the planning stage (also see Walters 1986; McAllister
and Peterman 1992b). This case study is therefore
worth discussing in detail.
The objectives of this experiment were to determine why the abundances of two economically
valuable groups of groundfish species were declining
relative to less valuable species and to take appropriate management action (Sainsbury 1988). In 1985,
Sainsbury proposed four different hypotheses, or
states of nature, that could potentially explain the
historical decrease in abundance of the valuable
species relative to the less valuable ones. These hypotheses were an intraspecific mechanism that
inhibited the valuable species, two different interspecific interactions between the valuable and
less-valuable species that kept the former at low
abundances, and a mechanism in which the existing
trawl fishery disrupted the preferred ocean floor
habitat of the valuable species. Sainsbury proposed
five experimental, or active adaptive, management
regimes to distinguish among these hypotheses (see
WA to WE in Figure 8.7, which shows the major elements of Sainsburys decision analysis). These
experimental management strategies ranged from
continuing the existing trawl fishery, to stopping the
trawl fishery for some period and using a trap fishery
only, to several activities in various spatial areas (including no fishing, trap fishing only, trawl fishing
only, or both). Sainsburys decision analysis forecasted the expected economic value of the fish catch for
each of these management strategies for each of the
four possible states of nature. These states of nature were weighted by their probability of occurrence
(P1 to P4), as derived from historical data and

Management actions

States of nature and associated probabilities

Hypotheses about
sh community

Time period (yr)

Management time

Hypothesis 1
P1

P2
P3

WA

t=5
WB

WC

Hypothesis 2
Hypothesis 3

P4

t = 10

Hypothesis 4

Value of catch
(millions $)

C1
C2
C3
C4

t = 20

Hypothesis 1
WD

Outcomes

t=5

P1

t = 10

P2
P3

WE
t = 20

Hypothesis 2
Hypothesis 3

P4
Hypothesis 4

Hypothesis 1
t=5

P1

t = 10
t = 20

P2
P3

Hypothesis 2
Hypothesis 3

P4
Hypothesis 4

C5
C6
C7
C8

C9
C10
C11
C12

. Decision tree for the analysis of various management actions in Sainsburys (1988) large-scale fishing
experiment. Management strategies (WA to WE ), time periods, hypotheses, and outcomes are described in the
text and Table 8.3. Only a subset of the complex branches is shown.

119

 . Results of Sainsburys (1991) calculations of the benefits of different designs for an active adaptive management
experiment on groundfish in Australia. Management strategies WA to WE are defined in the text; they differed in
how much fishing occurred and when, what type of gear was used, and whether the strategies were based only
on existing information as of 1985 (WA and WB) or on information resulting from the active adaptive experiment
(WC , WD , WE ). WB ,1 to WB ,4 refer to four different long-term harvesting strategies; time period, t, is the duration
of the experiment in years. Expected values of catch are in millions of Australian dollars. See text for details.
(Adapted from Sainsbury 1991.)

Strategy

Expected Value of catch (millions $)

WA
WB,1
WB,2
WB,3
WB,4
WC,t = 5
WC,t = 10
WC,t = 20
WD,t = 5
WD,t = 10
WD,t = 20
WE,t = 5
WE,t = 10
WE,t = 20

Bayesian statistical analysis. Several management

strategies, implemented from 5 to 20 years, were simulated under various scenarios, starting in 1985. After
this simulated learning period, the model then determined which of the four hypotheses was the most
likely, which in turn would suggest a long-term management plan. Thus, Sainsburys expected value of
the catch included the value during both the learning
period and the subsequent period of implementing
the derived optimal action.
Sainsbury (1991) found that the expected value of
the catch was maximized at AUS $40.6 million by applying strategy WE for 5 years. This experimental
strategy had some replicate areas open to trawling
and others closed. The other adaptive management
regimes had lower expected values (Table 8.3), which
illustrates the benefits of applying decision analysis to
compare different designs of experimental management plans. Without this type of rigorous analysis,
a suboptimal experimental design might have been
chosen. The value of collecting information was included in the calculated economic value of the catch
because a management strategy that produced highquality data during the learning period led to

120

9.96
27.2
35.4
31.8
9.96
35.6
29.7
21.2
37.4
37.2
36.3
40.6
40.5
38.6

improved understanding of which of the four hypotheses was responsible for the decline in
abundance of the valuable species. This approach allowed a more accurate decision to be made about
which long-term harvesting strategy was most likely
to reverse the problem and increase the value of the
catch. (Incidentally, Sainsbury et al. 1997 reported
that the experimental management strategy WE generated data by 1991 that strongly supported the fourth
hypothesisthat trawling detrimentally affected the
habitat of the more valuable groundfish species.
Trawling was subsequently reduced.)
8.7 Value of Information
By taking uncertainty into account quantitatively in
decision analyses, analysts can quantify the effects of
considering or reducing uncertainties when making
decisions. Several types of analyses are possible: expected value of including uncertainty (EVIU),
expected value of sample information (EVSI), expected value of perfect information (EVPI), and
expected value of experimental or adaptive management. (See Morgan and Henrion 1990 for more
details.)

The expected value of including uncertainty

(EVIU) provides a measure of how much better the
expected value of some decision will be if analysts
consider uncertainty through a decision analysis, as
opposed to the common approach of using only the
best point estimates of parameters to make decisions.
EVIU is calculated as the difference between the expected outcome of a decision based on a probabilistic
decision analysis (the Bayes decision) and a decision
based only on using the best point estimates of uncertain parameters (the deterministic decision).
Therefore, EVIU represents the increase in expected
benefits or reduction in expected losses that results
from using decision analysis and can be used to determine whether to spend the additional time
necessary to gather the data and complete a decision
analysis. Note that EVIU is always 0 because the
Bayes decision accounts for the potentially useful information (contained in the uncertainties) that is lost
if these uncertainties are ignored.
Another measure, the expected value of sample information (EVSI), estimates the benefits of reducing
uncertainties by collecting additional data through a
monitoring program or adaptive management experiment. EVSI requires calculating the expected value
of a decision made with improved information, compared to the current level of uncertainty. In practice,
this value is estimated by adjusting the probabilities
placed on the states of nature to reflect more certainty (i.e., making the probability distribution more
precise or more accurate) and then repeating the decision analysis using this adjusted distribution. EVSI
is then the difference between this value and the expected value of the Bayes decision. The effect of
additional information on the probabilities for the
states of nature can sometimes be estimated using
sampling theory. If the cost of carrying out the sampling program can be estimated, the ratio of the
benefits (EVSI) to the costs provides one way to evaluate the effectiveness and efficiency of planned
sampling programs and to allocate research budgets
among different sampling programs.
The expected value of perfect information (EVPI)
is a measure of the increase in expected benefits or
decrease in expected losses if we could forecast the
outcomes of management actions with complete certainty. Consequently, this value is the maximum
amount that we should be willing to pay for research
that will generate information and reduce uncertainties. Although this value is hypothetical because
uncertainties are always present, it is often instructive

to compute it because this value provides an upper

bound on EVIU and EVSI.
These general concepts of value of information are
directly relevant to adaptive or experimental management because the expected value of an experiment
can be calculated explicitly using decision analysis.
For instance, Table 8.3 shows Sainsburys (1991) estimates of the expected value of various management
plans as calculated before the experiment began in
1985. For example, the expected value of the catch
from allowing the existing trawl fishery to continue
(Strategy WB,4) was $9.96 million, given the
uncertainty that existed in 1985. Immediate implementation of long-term harvesting strategy WB,2
(moderate-intensity trap fishery) in 1985, without
collecting any additional information, would have increased the expected value of the catch to $35.4
million. This maximum expected value of the catch
represents what could have been realized given the
level of uncertainty in 1985. However, several of the
proposed experimental management strategies (WC,
WD, and WE) produced even larger expected values of
catch (Table 8.3) because of the value of the information about the uncertain biological hypotheses that
were generated by the experiment. For example, as
noted previously, the experimental strategy WE with
a learning period of 5 years maximized the expected
value of the catch at $40.6 million. This amount was
$5.2 million more than the next best strategy that
could have been implemented in 1985 without doing
experimental management (WB,2).
8.8 Quantifying Management Objectives
For a decision analysis to rank the management options, one or more management objectives must be
identified. However, disagreement about what the
management objective should be is common, as in
the land use issue in the Tatshenshini described previously (McDaniels 1992). Disagreement often occurs
when management objectives are based on consultation with a wide range of managers and stakeholders
(Bell et al. [editors] 1977). These disagreements can
be resolved by repeating the decision analysis using
several different management objectives, each representing a different viewpoint. The key question in
such analyses is how much the optimal decision
changes when different management objectives are
used. As noted previously in the sensitivity analysis
section, addressing this issue can help resolve conflicts by identifying which assumptions or elements

121

of the objectives lead to different recommended

management actions. In some cases, participants may
not disagree once the quantitative decision analysis is
done.
Conflicting management objectives can be treated
more formally using multi-attribute utility theory
(Keeney and Raiffa 1976). Utility is a unitless measure
of satisfaction obtained from different quantitative
outcomes of decisions. Utility analysis converts into
common units (utilities) different kinds of outcomes,
or attributes, such as dollar value of timber and an
index of biodiversity. Utility functions permit this
conversion and the shapes of these functions reflect
the degree of risk aversion of the stakeholder. Once
converted to utilities, these attributes can be combined into a weighted average utility, where
weightings placed on different attributes reflect their
relative importance to different interest groups.
Multi-attribute utility analysis thus provides a
quantitative method for incorporating multiple and
conflicting management objectives into the decisionmaking process. The disadvantage of combining
these objectives into a single weighted average utility
is that the trade-offs implicit in combining multiple
objectives are hidden from the decision-maker. For
that reason, it is often preferable to show the separate
attributes as functions of the management decision
along with the results of the multi-attribute utility
analysis. This explicitly shows decision-makers the
trade-offs that are inherent in particular decisions.
8.9 Communicating Uncertainty and Results of
Decision Analyses
To establish confidence in the analysis, all users of
the results of a decision analysis must be informed
not only of the optimal decision, but also of the assumptions for which that action is optimal. This
approach is necessary because, as noted under sensitivity analysis, different parameter values, model
structures, or management objectives can sometimes
lead to a different optimal decision. One of the
advantages of decision analysis is that these assumptions are made explicit, and consequently the effects
of these assumptions on the optimal decision can be
explored quantitatively. However, these advantages
are lost unless the decision analyst communicates
these results clearly and effectively to decision-makers. Several steps can be taken to ensure good
communication.

122

First, the decision-making process used to identify

the optimal decision must be adequately documented. This information includes full documentation of
the data, key assumptions, and methods; a list of
which factors were considered uncertain and why;
the results and implications of sensitivity analyses for
the decision-maker; and limitations of the decision
analysis. Such documentation shows the decisionmaker exactly how an optimal decision was derived
and provides an indication of how robust that decision is to the assumptions. The documentation also
allows decisions to be made consistently when the
same decision must be made by different people at a
different time or place. As well, good documentation
allows decisions to be subjected to an iterative
process of external peer review and evaluation.
In doing so, analysts and decision-makers can learn
from successes and failures, leading to progressive
improvements in decision-making.
The second step toward good communication is to
choose appropriate methods for presenting results of
decision analyses and sensitivity analyses. What
seems most appropriate for technical analysts may be
confusing for non-specialists. For instance, Ibrekk
and Morgan (1987) showed that the conventional way
for scientists to express uncertainty in some quantity,
a probability distribution, is not the most effective
way to convey understanding of that uncertainty to
others (i.e., they found a cumulative probability
distribution to be best). In some cases, a computerbased hierarchical information system has been used
to present and explain results of analyses (e.g., Gobas
1993). Such information systems allow the user to select the level of detail, ranging from summary graphs
to complete, detailed numerical tables.
A final measure to ensure good communication
between various individuals is to involve decisionmakers and other interested parties in the analysis
from the beginning. This step can be done through
workshops and by allowing users to conduct whatif runs with simulation models. Early participation
by decision-makers and stakeholders avoids
misunderstandings and misinterpretations of key
assumptions, data, methods, and results. It also increases the chance that these important groups will
accept the results and will support the analysis by
providing necessary information.

8.10 Benefits and Limitations of Decision Analysis

8.10.1 Benefits
To summarize, several key benefits of decision analysis have made it an increasingly popular method for
quantitatively considering uncertainties when making decisions in resource management.
1. By systematically accounting for complexities and
uncertainties, decision analysis can improve the
quality of decisions made, resulting in more
favourable outcomes over the long term.
2. Using a decision tree to structure a problem helps
identify what specific data and assumptions are
needed to perform the analysis.
3. Going through a formal decision analysis requires
explicit statements of the assumptions, parameter
values, and management objectives, including
views about risk aversion.
4. Decision analysis allows explicit comparison and
ranking of alternative management actions.
5. Sensitivity analyses help to set priorities for future
research and establish confidence in the analysis by
identifying the robustness of a recommended action.
6. Explicitly taking uncertainties into account permits calculation of the benefits of considering
uncertainty compared to only using the best point
estimates (EVIU), and the value of reducing these
uncertainties through the implementation of a research sampling program (EVSI).
7. A systematic approach documents the method by
which decisions were reached and thus indicates
which methods of analysis and management actions work and why.
8. Decision analysis can be used for conflict resolution between interest groups.
8.10.2 Limitations
As with any method that assists decision-making,
decision analysis also has its limitations. A major limitation is that the amount of data required to conduct
a decision analysis of a complex problem can be
large. States of nature, probabilities on those states,
and outcomes of management actions must all be
quantified to apply decision analysis. In many cases,
these data may not be available in sufficient quantity
or quality to allow formal decision analysis. Another

limitation of decision analysis is that quantifying

management objectives is sometimes difficult. This
situation is especially problematic when diverse user
groups or stakeholders are part of the decision-making body or are involved in consultations with
managers. Under these circumstances, identifying
quantitative indicators of management objectives can
be difficult even if multiattribute utility analysis is
applied.
These limitations can, in some cases, make
application of decision analysis impossible or unwarranted. When this happens, management actions
should be taken cautiously, given the inevitable presence of uncertainties.
Another limitation of decision analysis stems not
so much from the method itself as from the way in
which results are used. As described previously, decision theorists define risk as expected loss. Thus,
when decision analysis compares the expected values
of outcomes for various possible management actions, it essentially calculates the risk associated with
each action. However, when managers or scientists
present such results to stakeholders, they may interpret them quite differently. Substantial research
shows that such people often perceive risks quite
differently from the amount of risk estimated from
quantitative analyses (Slovic 1987). The magnitude of
this difference depends on factors such as the amount
of control over the risk, the level of trust in the experts, and the immediacy of the effect (Slovic 1987).
8.10.3 Evaluation of quality of decisions
The benefits and limitations of decision analysis lead
to two important points about how decisions should
be evaluated. First, the quality of decisions should be
evaluated based on the process used to make them, not
on their short-term outcome. This view is based on the
observation that, because of the complexity of forest
ecosystems and the number of factors influencing
them, favourable outcomes might arise as much from
fortuitous events as from good decisions. Thus, for
instance, if chance events happened to lead to a good
outcome in some situation, in spite of an incorrect
decision, managers might unjustifiably conclude that
the decision they made was correct and they might
repeat it in similar circumstances. However, the
chance events might not occur again in the managers
favour. Similarly, a correct decision might lead to
some detrimental effect because of an unfavourable
chance event that coincided with the decision. For
this reason, conclusions about the quality of a

123

decision should not be based on short-term outcomes;

they should be based on how systematically, rigorously, and comprehensively the decision options
were evaluated before making the decision. Studies
show that decisions based on rigorous analyses that
quantitatively account for uncertainties will, in the
long term, produce better results than decisions made
using other approaches (Von Winterfeldt and Edwards 1986). Thus, decisions that are based on a
rigorous approach to analyzing options and uncertainties should be labelled good decisions, whereas
others should be described as unacceptable, regardless
of the short-term outcomes of any particular decision.
The second point is that the decision-making
process should be judged not on an absolute scale,
but relative to other methods available. Decision
analysis has some potentially serious limitations, but
few alternative methods have been demonstrated to
provide a better approach to using information on
the uncertainties and complexities inherent in resource management decisions. Because of this, decision analysis is being increasingly applied to a wide
range of problems in toxicology, water quality,
forestry, fisheries, and wildlife management.
However, methods for quantitative policy analysis
are continually being improved, and analysts should
be aware of developments in the field so that they use
the best methods available to make decisions.
8.11 Final Recommendations for Decision Analysts
and Decision-makers
Do not push scientists to state their best estimate
despite the uncertainties because this effectively
ignores uncertainties and will often lead to management actions with undesirable results. Instead,
for choosing among options, use a systematic
method such as decision analysis, which takes uncertainties into account explicitly.
Do not forget the caveats and limitations of the
various components of a decision analysis. For example, recognize the trade-offs between the
complexity of models and their reliability. Acknowledge the assumptions behind formulation
and parameterization of models, and use sensitivity analyses to explore how these factors affect the
optimal decision.

124

When doing a decision analysis, adhere to the

following guidelines to ensure that the decisionmaking process is the best available:
1. Clearly identify the main goal of the decision
analysis.
2. Ensure that interaction occurs early and
periodically among scientists, analysts, decision-makers, the public, user groups, and other
stakeholders.
3. Document all steps in the analysis.
4. Do not assume that everyone will agree with
your methods (e.g., Bayesian statistics), estimates of parameters, or interpretation of data.
5. State the assumptions and data used, carefully
qualify the conclusions, and clearly define the
limits of the analysis.
6. Present extensive sensitivity analyses that focus
on:
a) how the rank order of management options
changes with different assumptions; and
b) research prioritiesthe most important
areas for getting more data.
7. Be cautious: not only could the analysis be incomplete but it will almost certainly be missing
components. These factors may affect your results.
8. When communicating information about risks
or uncertainties, think about what it is like not
to know the material.
9. The entire process of decision analysis and
communication should be iterative, continually
moving toward improving decisions.
10. Insist on objective science and rigorous external
peer review of analyses.
11. When decision analysis is used to evaluate different proposed adaptive management actions,
implement a monitoring program in conjunction with the chosen action to ensure that the
maximum amount of information is obtained.
12. Recognize that decision analysis is only one part
of the whole process for making decisionsit is
NOT the entire process. However, if decision
analysis is one component, it can help improve
environmental decision-making.

 Not all circumstances warrant a full, formal quantitative decision analysisjustifiable usage of
decision analysis is case-specific. For example, decision analysis is more feasible if at least some
reliable data are available and clear management
objectives are stated. Furthermore, decision analysis is more appropriate when costs of incorrect
decisions are potentially large. First-time users of
this approach are encouraged to use the references
here and to discuss the approach with experts who
have previously used decision analysis. Regardless
of the specific situation, it is always worth at least
thinking about a decision-making problem in
terms of the components of decision analysis as
described, even if final calculations are never carried out due to limitations in data or other
problems. The mere process of describing each
component helps to clarify and organize the decision-making process and to identify research
needs.
Acknowledgements
We are grateful to Vera Sit, Brenda Taylor, Darrell
Errico, and Milo Adkison for many useful suggestions, comments, and discussions. Milo Adkison also
suggested the format for Table 8.1. Several reviewers
also provided helpful comments: Russ Horton,
Wendy Bergerud, Michael Stoehr, Jeff Stone, and
Peter Ott.
References
Adkison, M.D. and R.M. Peterman. 1996. Results of
Bayesian methods depend on details of implementation: an example of estimating salmon
escapement goals. Fish. Res. 25:15570.
Bell, D.E., R.L. Keeney, and H. Raiffa (editors). 1977.
Conflicting objectives in decisions. J. Wiley,
New York, N.Y.
Berger, J.O. and D.A. Berry. 1988. Statistical analysis
and the illusion of objectivity. Am. Sci.
76:15965.
Bergerud, W.A. and W.J. Reed. [n.d.]. Bayesian statistical methods. This volume.

Bergh, M.O. and D.S. Butterworth. 1987. Towards rational harvesting of the South African anchovy
considering survey imprecision and recruitment variability. S. African J. Marine Sci.
5:93751.
Box, G.E.P. and G.C. Tiao. 1973. Bayesian inference
in statistical analysis. Addison-Wesley, Reading,
Mass.
Clark, C.W. 1990. Mathematical bioeconomics: the
optimal management of renewable resources. J.
Wiley, New York, N.Y.
Clemen, R.T. 1996. Making hard decisions: an introduction to decision analysis. 2nd ed. Duxbury
Press, Wadsworth Publ. Co., Belmont, Calif.
Cohan, D., S.M. Haas, D.L. Radloff, and R.F. Yancik.
1984. Using fire in forest management: decision
making under uncertainty. Interfaces 14:819.
Crome, F.H.J., M.R. Thomas, and L.A. Moore. 1996.
A novel Bayesian approach to assessing impacts
of rain forest logging. Ecol. Applic. 6:110423.
Edwards, A.W.F. 1992. Likelihood: expanded edition.
Johns Hopkins Univ. Press, Baltimore, Md.
Ellison, A.M. 1996. An introduction to Bayesian inference for ecological research and
environmental decision-making. Ecol. Applic.
6:103646.
Errico, D. 1989. Choosing stand density when spacing
lodgepole pine in the presence of risk of pest attack. B.C. Min. For., Victoria, B.C. Res. Rep.
89004-HQ.
Frederick, S.W. and R.M. Peterman. 1995. Choosing
fisheries harvest policies: when does uncertainty matter? Can. J. Fish. Aquat. Sci. 52:291306.
Gobas, F.A.P.C. 1993. A model for predicting the
bioaccumulation of hydrophobic organic
chemicals in aquatic food-webs: application to
Lake Ontario. Ecol. Modeling 69:117.
Hairston, N.G. (editor). 1989. Ecological experiments: purpose, design, and execution.
Cambridge Univ. Press, Cambridge, U.K.

125

Henrion, M. and B. Fischhoff. 1986. Assessing uncertainty in physical constants. Am. J. Physics
54:7917.

Maguire, L.A. 1986. Using decision analysis to manage endangered species populations. J. Environ.
Manage. 22:34560.

Hilborn, R., E.K. Pikitch and M.K. McAllister. 1994.

A Bayesian estimation and decision analysis for
an age-structured model using biomass survey
data. Fish. Res. 19:1730.

Maguire, L.A. and L.G. Boiney. 1994. Resolving environmental disputes: a framework
incorporating decision analysis and dispute resolution techniques. J. Environ. Manage.
42:318.

Holbert, D. and J.C. Johnson. 1989. Using prior information in fisheries management: A
comparison of classical and Bayesian methods
for estimating population parameters. Coastal
Manage. 17:33347.
Howard, R.A. 1988. Decision analysis: practice and
promise. Manage. Sci. 34:67995.
Ibrekk, H. and M.G. Morgan. 1987. Graphical communication of uncertain quantities to
non-technical people. Risk Analysis 7:51929.
Keeney, R.L. 1982. Decision analysis: an overview.
Operations Res. 30:80338.
______. 1992. Value-focused thinking. Harvard Univ.
Press, Cambridge, Mass.

Mapstone, B.D. 1995. Scalable decision rules for environmental impact studies: effect size, Type I,
and Type II errors. Ecol. Applic. 5:40110.
Morgan, M.G. and M. Henrion. 1990. Uncertainty:
a guide to dealing with uncertainty in quantitative risk and policy analysis. Cambridge Univ.
Press, Cambridge, U.K.
Nyberg, J.B. [n.d.]. Statistics and the practice of
adaptive management. This volume.
Osenberg, C.W., R.J. Schmitt, S.J. Holbrook, K.E.
Abu-Saba, and A.R. Flegal. 1994. Detection of
environmental impacts: natural variability, effect size, and power analysis. Ecol. Applic.
4:1630.

Keeney, R.L. and H. Raiffa. 1976. Decisions with multiple objectives: preferences and value
trade-offs. J. Wiley, New York, N.Y.

Parkhurst, D.F. 1984. Decision analysis for toxic

waste releases. J. Environ. Manage. 18:10530.

Lindley, D.V. 1985. Making decisions. Wiley Interscience, New York, N.Y.

Parma, A.M. and R.B. Deriso. 1990. Experimental

harvesting of cyclic stocks in the face of alternative recruitment hypotheses. Can. J. Fish.
Aquat. Sci. 47:595610.

Lord, G.E. 1976. Decision theory applied to the simulated data acquisition and management of a
salmon fishery. Fish. Bull. (U.S.) 74:83746.
McAllister, M.K. and R.M Peterman. 1992a. Decision
analysis of a large-scale fishing experiment
designed to test for a genetic effect of sizeselective fishing on British Columbia pink
salmon (Oncorhynchus gorbuscha). Can. J. Fish.
and Aquat. Sci. 49:130514.
______. 1992b. Experimental design in the management of fisheries: a review. N. Am. J. Fish.
Manage. 12:118.
McDaniels, T. 1992. A multiple objective decision
analysis of land use for the Tatshenshini-Alsek
area. Appendix to report for B.C. Commission
on Resources and Environment, Victoria, B.C.

126

Pearse, P.H. 1976. Timber rights and forest policy in

British Columbia. Rep. Royal Comm. For. Res.,
Victoria, B.C.
Peel, A.L. 1991. Forest resources commission: the future of our forests. B.C. Forest Resources
Commission, B.C. Min. of For., Victoria, B.C.
Peterman, R.M. 1990a. Statistical power analysis can
improve fisheries research and management.
Can. J. Fish. Aquat. Sci. 47:215.
______. 1990b. The importance of reporting statistical power: the forest decline and acidic
deposition example. Ecology 71:20247.

Peterman, R.M. and M.K. McAllister. 1993. A brief

overview of the experimental approach to reducing uncertainty in fisheries management
an extended abstract. In Risk evaluation and biological reference points for fisheries
management. S.J. Smith, J.J. Hunt, and D. Rivard (editors). Can. Spec. Pub. Fish. Aquat. Sci.
120:41922.

Simberloff, D.S. and L.G. Abele. 1976. Island biogeographic theory and conservation practice.
Science 191:2856.

Press, S. J. 1989. Bayesian statistics: principles, models, and applications. J. Wiley, New York, N.Y.

Stahl, G., D. Carlsson, and L. Bondesson. 1994. A

method to determine optimal stand data acquisition policies. For. Sci. 40:63049.

Raiffa, H. 1968. Decision analysis: introductory lectures on choices under uncertainty.

Addison-Wesley, Don Mills, Ont.
Reckhow, K.H. 1994. Importance of scientific uncertainty in decision making. Environ. Manage.
18:1616.
Render, B. and R.M. Stair. 1988. Quantitative analysis
for management. Allyn and Bacon, Boston,
Mass.
Sainsbury, K.J. 1988. The ecological basis of multispecies fisheries, and management of a
demersal fishery in tropical Australia. In Fish
Population Dynamics, J.A. Gulland (editor),
2nd ed. J. Wiley, New York, N.Y. pp. 34982.
______. 1991. Application of an experimental approach to management of a tropical
multispecies fishery with highly uncertain dynamics. ICES (International Council for the
Exploration of the Sea) Marine Sci. Symp.
193:30120.

Slovic, P. 1987. Perception of risk. Science 236:2805.

Sokal, R. and F.J. Rohlf. 1969. Biometry: the principles and practice of statistics in biological
research. W.H. Freeman, San Francisco, Calif.

Thompson, G.G. 1992. A Bayesian approach to management advice when stock-recruitment

parameters are uncertain. Fish. Bull. (U.S.).
90:56173.
Von Winterfeldt, D. and W. Edwards. 1986. Decision
analysis and behavioral research. Cambridge
Univ. Press, Cambridge, U.K.
Walters, C.J. 1981. Optimum escapements in the face
of alternative recruitment hypotheses. Can. J.
Fish. Aquat. Sci. 38:67889.
______. 1986. Adaptive management of renewable resources. MacMillan, New York, N.Y.
Westland Resource Group. 1995. Environmental indicators for land and resource management
planning. Prep. for B.C. Min. of Environ.,
Lands and Parks, Victoria, B.C.

Sainsbury, K.J., R.A. Campbell, R. Lindholm, and

A.W. Whitelaw. 1997. Experimental management of an Australian multispecies fishery:
examining the possibility of trawl induced
habitat modification. In Global trends: fisheries
management. E.L. Pikitch, D.D. Huppert, and
M.P. Sissenwine (editors). American Fisheries
Society Symposium, Vol. 20, Bethesda, Md,
pp. 10712.
Scientific Panel for Sustainable Forest Practices in
Clayoquot Sound. 1995. Sustainable ecosystem
management in Clayoquot Sound: planning
and practices. Rep. 5, Victoria, B.C.

127

9 SELECTING APPROPRIATE STATISTICAL PROCEDURES AND ASKING THE RIGHT QUESTIONS: A SYNTHESIS
BRUCE G. MARCOT
Abstract

In this chapter, I synthesize statistical approaches and

considerations for adaptive management studies. I review approaches to learning from management
actions and address questions of space and time. I also
present a set of guidelines for asking the right questions about statistical reliability, for selecting the
appropriate adaptive management study, and for
guiding how different types of information can contribute at different stages in adaptive management.
These guidelines are presented in a table, which can
be used as a decision tree to determine the best kinds
of studies for each step in the adaptive management
process, and the most appropriate use of exisiting
information.

9.1 Introduction
How should managers and researchers select an approach for designing an adaptive management study
and analyzing the results? The chapters in this report
provide some guidance; for example, Nemec (this
volume, Chap. 2), summarizes principles of experimental design, and Schwarz (this volume, Chap. 3)
lists types of nonexperimental and experimental designs. Other publications (e.g., Green 1979), while not
specific to adaptive management as defined in this
volume, also provide guidance on designing ecological studies. This chapter reviews issues to consider in
designing adaptive management studies, synthesizes
the methods discussed in preceding chapters of this
report, and summarizes the roles different types of
information can play in adaptive management.
Statistical approaches and study designs can be selected only when the management question is first
well articulated. In the first section of this chapter,
I review three types of monitoring, differentiated by
the types of question they each address, and then address how the spatial and temporal elements of a
management question can influence study design.
In the second section, I review the characteristics of
powerful studies and the principles of experimental
design. The third section summarizes various types of
information (including existing data, retrospective
studies, and nonexperimental studies) and experimental studies, and how they can contribute to

adaptive management. In the final section, I discuss

some points to consider in interpreting and communicating the results from adaptive management
studies, and in particular the difficulty in unravelling the causal web. Throughout this chapter, I use
the oversimplistic labels researcher and manager,
fully realizing that in the real world many resource
professionals don both hats.
9.2 Types of Questions Addressed in Adaptive
Management
A little experience often upsets a lot of theory.
Cadman
The B.C. Ministry of Forests defines adaptive management as a formal process entailing problem
assessment, study design, implementation, monitoring, evaluation, and feedback (B.C. Ministry of
Forests 1996). In this approach, management activities are crafted as experiments to fill critical gaps in
knowledge. The key questions are: (1) To what extent
did the managment action lead to the measured outcome? and (2) Are our assumptions valid about how
the system works?
Other institutions use the term adaptive management differently. For example, the USDA Forest
Service incorporates the general concepts of adaptive
management into its planning, but not as a formal
process. Regardless of the definition of adaptive management and how it is institutionalized, monitoring
activities and evaluation of data are key steps in
adaptive management. The statistical approaches
discussed in this report can help in both the design of
monitoring activities and in the interpretation of
data.
9.2.1 Types of monitoring
There are three types of monitoring: implementation
monitoring, effectiveness monitoring, and validation
monitoring. Each type of monitoring serves a unique
function in an adaptive management study.
Implementation monitoring
Implementation monitoring (or compliance monitoring) essentially asks: Have the management guidelines been implemented correctly (Collopy et al. 1993)?

129

Correct implementation can be determined by a

complete census of all activities or by sampling activities stratified by administrative unit or location. Obviously, asking more detailed questions of the effects
and validity of particular management activities
should proceed only when they have been correctly
implemented. Implementation monitoring, however,
does not teach us about effects of management actions. Thus, the focus of adaptive management is effectiveness and validation monitoring.
Effectiveness monitoring
Effectiveness monitoring asks: Are the management
guidelines and activities producing the desired effects? Do the management activities really alter the
biophysical conditions as expected? Many questions
can be asked of the effects of management guidelines.
Highest priority should be directed to potential effects that have the most serious economic, biological,
or ecological ramifications, and those carrying the
greatest uncertainty.
Validation monitoring
Validation monitoring, technically the most difficult
of the three kinds of monitoring, asks: Are the ultimate expectations for the guidelines being met? Are
the basic assumptions about how the biophysical system operates really correct, or does it operate in a
very different way that would invalidate the selected
management approach? If so, how?
Validation monitoring may be used to validate
ecosystem models (Gentiol and Blake 1981), which is
vital to ensuring the models successful and appropriate use. In adaptive management, validation monitoring should focus on the ecosystem elements that
have the greatest implications on the decision about
the best course of action. Problem assessmentidentifying which relationships to validateis the first
step of adaptive management.
9.2.2 Issues of space and time
Issues of space and time will in part determine the
type of study design that is possible. For example,
studies of large geographic areas may preclude replication, suggesting before-after-control-impact paired
(BACI-P) study (Schwarz, this volume, Chap. 3).
Similarly, long response times may suggest retrospective analysis of past actions to provide a preliminary
assessment of the impact of a proposed action.

130

Issues of space
The five kinds of spatial effects to consider can influence the design of a study as well as the interpretation
of its results.
1. What is the influence of on-site management
activities on off-site conditions? That is, local
management may influence remote conditions,
both directly and indirectly (Loehle 1990). An example is the downstream effect of stream
temperature or sedimentation on fish populations
due to local reduction, removal, or restoration of
riparian vegetation cover.
2. What is the relative influence of off-site management activities on on-site (desired) conditions?
On-site conditions can be influenced by other offsite activities. For example, despite protection of
old-growth forest groves, some arboreal lichens
might nonetheless decline because of degraded air
quality from industrial pollutants originating elsewhere in the airshed. The potential influence of
downstream dams and fish harvesting on the
abundance of local fish populations is another
example.
3. To what degree do local management activities influence the on-site (desired) conditions? That is,
to what extent do background noise and other environmental factors affect on-site conditions?
Local management may influence only a portion
of the total variation in local conditions. For example, providing local breeding habitat only
partially succeeds in conserving populations of
neotropical migratory birds, whose numbers may
still decline due to pesticide loads or habitat loss
encountered during wintering in the neotropics.
4. What is the relative influence of conditions and activities from different spatial scales, particularly the
effects on local stand-level conditions from broader landscape-level factors? That is, desired
conditions and management actions are best addressed at appropriate scales of geography. As
examples, effects of forest management on abundance of coarse woody debris are best assessed at
the stand level; effects of forest management on
vegetation conditions that affect visual quality or
goshawk (Accipiter gentilis) habitat are best assessed at the landscape level; and effects of overall
management policy and ownership patterns on
grizzly bear (Ursus arctos) populations are best
assessed at subregional or regional levels.

5. What are the cumulative effects of stand-level

treatments as they spread across the landscape?
For example, wind fetch and thus wind speed may
increase as clearcuts become wider with sequential,
adjacent cuts. Thus, the windthrow hazard in one
cutblock may increase as adjacent areas are cut,
and the windthrow hazard in those cutblocks
cannot simply be extrapolated from the hazard
measured in a single cutblock surrounded by trees.
For each of these five kinds of spatial effects,
adaptive management monitoring studies would be
designed and implemented differently. Where this is
not possible, spatial influences should at least be acknowledged as potential sources of variation and
included in the analysis.
Issues of time
Answering questions about time effects can help distinguish true cause from non-causal correlation, and
treatment effects from natural variation. Three typical time scale issues follow.
1. What are the response times of variables? For
some variables, response may be apparent in a relatively short period of time; others may respond
more slowly. Examples are the relatively short and
quick response time of seedling survival compared
with the long and slow response times associated
with many biodiversity indices (e.g., changes in
grizzly bear populations).
2. What are the lag times of variables? Some variables may not immediately respond to a treatment
or may depend greatly on site history. For example, because acorn woodpeckers (Melanerpes
formicivorous) show high fidelity to particular sites,
a lag will exist before they respond to the removal
of granary trees (Ligon and Stacey 1996). This lack
of short-term response should not lead one to conclude that management actionsin this example,
the reduction or removal of granary treeshave
no effect. Sometimes these lags in response result
when conditions from prior time periods overwhelm or influence responses from current
actions. For example, the intensity of a fire will be
influenced by site history, in addition to current
management actions. Thus short-term changes in
a response variable may reflect both the management action and past site history. Some time-lag
effects can be quite variable and manifest as nonmonotonic (up and down) trends over the long

term. For example, annual non-monotonic variations in bird populationsboth increases and
decreasesmay belie truer long-term declines in
some population counts (Thomas and Martin
1996).
3. What are the cumulative effects of a variable over
time? Some variables do not make a mark except
over time or until a particular threshold has been
exceeded. An example is the adverse effect of certain pesticides on wildlife reproduction. The
detrimental effect may not be apparent until the
pesticide concentrations reach a particular level of
toxicity (Tiebout and Brugger 1995).
The design of adaptive management studies and
selection of analysis methods are guided in part by
these considerations of space and time. For example,
replication is one major consideration in designing
studies. Given a large geographic area, as tends to be
the focus in ecosystem management, or a rare condition, such as a threatened species population, are
spatial replicates possible? That is, can landscapes or
threatened populations be replicated at all, or in
adequate numbers? If the conditions cannot be replicated, then pseudoreplication (e.g., dividing a single
area into smaller blocks) may be the only recourse
(Hurlbert 1984). Alternatively, other kinds of studies
(e.g., analytical surveys, expert testimony) might help
in assessing the impact of the treatments, although
they do not allow strong inference about cause.
Similarly, long response times and time lags make
temporal replication difficult. Retrospective studies
(see Smith, this volume, Chap. 4) provide one alternative for gaining insight into the long-term effects
of management actions. In cases where either spatial
or temporal replication is severely limited, a higher
probability of Type I and II errors might need to be
tolerated (see Anderson, this volume, Chap. 6).
In some cases, a powerful adaptive management
study may be possible but managers, decision-makers, industries, or other interested bodies may not be
willing to bear the cost, duration, and tight controls
on management activities. The consequences of not
using an optimum study must be explicitly considered and understood by all.

131

9.3 Considerations in Designing an Adaptive Management Study

9.3.1 Characteristics of a powerful adaptive management study
To help in evaluating management actions and validating functional and causal relationships, an
adaptive management study should be consistent
(i.e., should represent the system of interest), accurate, precise, and unbiased (see Routledge, this
volume, Chap. 5). Managers and researchers should
work together in designing an adaptive management
study that represents the real system and provides information within acceptable limits of Type I and
Type II errors (Anderson, this volume, Chap. 6).
They may also want to consider the trade-offs inherent in relaxing any of the conditions, such as
accepting a lower but still acceptable level of precision in exchange for lower cost or more rapid results.
The study design should also be independently reviewed to assess its capability to meet the desired
(and often conflicting) criteria of high consistency,
high accuracy, high precision, and low bias.
9.3.2 What managers need to ask of reliability
Managers should ask four general questions regarding the reliability of adaptive management studies
and their results.
1. What confidence can I have in the results of this
adaptive management study, particularly for
avoiding false positives? Statistically, this question
can be answered by calculating the probability of a
Type I error (Anderson, this volume, Chap. 6).
2. What power do the results provide for avoiding
false negatives (Anderson, this volume, Chap. 6)?
Statistically, this can be answered by calculating
the probability of a Type II error (although
Bayesian approaches differ significantly in not
dealing with questions of confidence and power).
Type I and Type II errors hold different implications for managers (Marcot 1986; Anderson, this
volume, Chap. 6). For example, if the adaptive
management study is aimed at determining adverse effects of some management activity on a
wildlife species that is threatened, then managers
may be more tolerant of a Type I error than of a
Type II error. However, if the species is not threatened and the activity results in important
commodity production and economic return,
then they may be more tolerant of a Type II error.

132

3. What is the relevance of the results? How representative is the study of other sites or conditions?
Some studies may reveal only local conditions and
the chance effects of unique site histories, rather
than overall effects, or they may pertain to only
one vegetation type or climatic condition. The
manager should know the contexts under which
results apply. For example, results of a forest thinning operation may apply to only a particular
initial stand density or forest type.
4. Were the effects truly a result of the management
activity? This question cuts to the heart of separating cause from noise, and determining what really
influenced the outcome. The experimental studies
that are central to adaptive management are designed to determine causality. Researchers and
managers should not assume that demonstration
of pattern and correlation constitutes valid evidence of causation.
9.3.3 Principles of experimental design
To help ensure success in evaluating management
actions, researchers should review adaptive management studies for the four main principles of
experimentation: randomization, replication, blocking, and representation (see Nemec, this volume,
Chap. 2). Randomization reduces bias. Replication
allows an estimation of variance, which is vital for
confirming observed differences. Blocking increases
precision and reduces cost and sample size. Representation helps to ensure study of the correct universe of
interest.
In the real world, these four principles cannot always be met and compromises are necessary. It is
often impossible to fully randomly allocate treatments, such as forest clearcuts or fire locations. In
such cases, study sites may be randomly selected
from existing clearcuts or fire locations, resulting in
nonexperimental studies (e.g., observational studies,
analytical surveys, retrospective studies, or impact
studies; see Schwarz, this volume, Chap. 3). When interpreting study results, researchers should account
for the site-specific characteristics leading to the initial nonrandom assignment of the treatment.
Furthermore, the researcher should recognize that
the altered study can no longer provide reliable
knowledge of cause, but only generates hypotheses
for validation when future management actions are
implemented.
When replication is not possible, suspected causal
effects can be masked by confounding hidden causes

or by spurious correlations. Researchers may be

tempted to resolve the problem by taking multiple
samples as pseudoreplications. The drawback of this
solution is that study results apply to study areas
only and cannot be generalized to the entire system
of interest.
When blocking is not feasible, precision suffers.
Larger sample sizes, hence increased cost, are necessary to achieve desired levels of confidence and
power. Finally, when a study considers only a portion
of the system of interest (due to lack of randomization, replication, or funding), generalization of the
results to the entire system could be inappropriate
and misleading. In this case, researchers and managers together must re-evaluate the study objectives
and scope.
Even though researchers are responsible for
designing studies, managers and decision-makers
should be aware of these issues and possible limitations. Other useful aspects of measurement errors are
reviewed by Routledge (this volume, Chap. 5), who
presents a useful set of criteria for selecting indices.
9.4 Types of Information and Study Designs
Study the past if you would divine the future.
Confucius
Information from sources other than management
experiments can play important roles in adaptive
management. For example, expert judgement, anecdotes, existing data, and literature can help in
building simulation models used to explore alternative scenarios and identify key uncertainties.
Information from these sources can also provide supporting evidence, which becomes important when
real world limitations prevent the design of ideal
management experiments. Each source of information provides different levels of reliability.
9.4.1 Learning from existing data, expertise, and
expert testimony
Using existing data and literature
In the initial stages of adaptive management, existing
data and literature can be used to evaluate scenarios,
project effects, or devise guidelines. However, the
ability to determine treatment effects from existing
data is often limited because such data may not cover

the range of environments or treatments proposed,

or may be knitted from disparate databases. In addition, the spatial, temporal, or ecological scope and
the degree of reliability of such data may be poorly
documented. Perhaps a good reminder of the potential weaknesses of using existing information is to
remember the acronym for best available data.
When existing data are used, how well they can address the critical management question should be
assessed honestly and accurately.
Gathering expertise and expert testimony
Another source of information is expert judgement,
review, and testimony. Broad-scale assessments of
wildlife population viability conducted recently by
federal resource management agencies of the western
United States have relied on panels of experts and
contracted technical reports to fill in many gaps left
by existing databases and publications (e.g., Schuster
et al. 1985). In my own work using expert-panel
approaches, I have modified1 the Delphi technique
(Zuboy 1981; Richey, Horner, and Mar 1985) to collect expert knowledge and judgement (Marcot et al.
1997). However, expert judgement cannot replace
statistically sound experiments.
9.4.2 Learning from management actions
Probably the most reliable means of gathering information for assessing the impact of management
actions is to conduct field studies. But, like publications and expert opinion, empirical evidence comes
in many forms and levels of usefulness. A few key
sources of evidence for the manager to know about
listed here in increasing order of reliabilityinclude
anecdotes and expert judgement, retrospective studies, nonexperimental (observational) studies, and
experimental manipulation.
Anecdotes and expert judgement
The results of management actions are often evaluated informally by simple observations with no measurements. Such opportunistic observations are
a two-edged foil: while the collective expertise from
field experts can constitute a valuable and irreplaceable pool of wisdom, individual anecdotes can prove
strikingly misleading. As a whole, anecdotal information should be used with a great deal of cautionor
at least with rigorous peer reviewto help avoid
problems such as motivational bias (Marcot et al. 1997).

1 Modifications addressed the need to adhere to the U.S. Federal Advisory Committee Act, by polling individual experts for basic ecological information and not reaching group consensus on specific management actions.

133

Anecdotes and expert judgement alone are not recommended for evaluating management actions because of their low reliability and unknown bias.
In the BC Forest Service, use of this source of information alone to evaluate management actions is not
considered adaptive management.
Retrospective studies
Sometimes the results of management actions are
provided by measuring the outcomes of future actions taken in the past. Retrospective studies (evaluating the outcomes of actions taken in the past) are
valuable for helping to predict the outcomes of future
actions. These studies can provide some insights to
support or refute proposed hypotheses, and are particularly valuable for problems where some indicators take a long time to respond. However, because
the treatments might not have been randomly assigned, and the initial conditions and the details of
the treatments are often unknown, teasing out causal
factors may be challenging at best and misleading at
worst.
Nonexperimental (observational) studies
Nonexperimental studies (called observational
studies by some authors) are the most common kind
of field studies reported in wildlife journals. Like retrospective studies, nonexperimental studies are not
based on experimental manipulations. Although it
may be debatable whether nonexperimental studies
should entail hypothesis testing, they should
nonetheless meet statistical assumptions, including
adequacy of samples sizes and selection of study sites,
to ensure reliable results. Much can be learned from
taking advantage of existing conditions and unplanned disturbances (Carpenter 1990; Schwarz, this
volume, Chap. 3).
Nonexperimental studies usually entail analysis of
correlations among environmental and organism parameters, such as studying the correlations between
clearcutting and wildlife response. Causes are
inferred and corroborated through repeated observations under different conditions. Because results may
be confounded by uncontrolled (and unknown) factors, nonexperimental studies are best interpreted as
providing only insights to cause. These insights can
be valuable in predicting outcomes of actions, but
again, the veracity of such predictions and the effects
of management actions are best evaluated through
controlled experiments (McKinlay 1975, 1985). Of

134

nonexperimental studies, BACI-P designs allow the

strongest inferences about causes (Schwarz, this volume, Chap. 3).
Inventories and surveys are not the same as nonexperimental studies; they display patterns but do not
reveal correlates. Nevertheless, inventories and surveys can be useful in adaptive management. They
provide information from which to select random
samples, or a baseline of conditions from which to
monitor changes over time. Inventories and surveys
should still adhere to strict sampling protocols and
can use more advanced statistical methods to streamline efficiencies (Schwarz, this volume, Chap. 3). For
example, Max et al. (1990) presented an inventory
method of random sampling of Northern Spotted
Owl (Strix occidentalis caurina) territories with partial, annual replacement of samples to increase
accuracy and reduce bias in estimates of site occupancy.
One particularly terse version of inventories is
rapid assessment procedure (RAP) or rapid survey,
used by some conservation groups running ahead of
the bulldozers to survey biota of tropical forests
(Oliver and Beattie 1993, 1996). Rapid surveys may
prove useful in some temperate ecosystems as well,
but should be used only to provide quick, initial,
mostly qualitative or categorical information from
which to design more formal adaptive management
studies.
Experimental manipulation
Management actions can best be evaluated
through experimentation (Nemec, this volume,
Chap. 2). Experimental manipulations can be used
to quantify the contributions from each suspected
causal factor, and ultimately to develop, refine, and
validate prediction models. The kind of experimentation referred to here involves deliberate, planned
alterations of one or more sites, one of which may be
an unaltered control.
Finally, demonstrations are not adaptive management per se, but often appear in the adaptive
management literature (e.g., Yaffee et al. 1996).
Demonstrations are designed to showcase the execution of specific management activities such as
silvicultural techniques but they do not provide the
evidence that controlled, replicated experiments do.
When faced with a proposal for a demonstration
study, the manager might first ask if they need evidence of cause and effect, and, if so, if a management

experiment with controls and replicated treatments

would better provide evidence as well as the opportunity to demonstrate the activities.
9.4.3 Information for improving study designs
Study designs can be improved by using prior knowledge of the system of interest gained through
retrospective analysis of past events, existing literature, and expert testimony. This information can aid
in blocking samples to increase study efficiency, and
in ensuring correct spatial and temporal representation of samples.
Study design can also benefit from initial field
sampling. This sampling can provide preliminary
estimates of variance of parameters that can be used
to calculate sample size necessary to meet desired
levels of precision. Initial field sampling also gives
information on stratification or blocking strategy
and helps to reveal conditions not originally considered in a study.
The relative merit of alternative study designs can
be assessed using the tools of quantitative decision
analysis, including Bayesian statistics (see Bergerud
and Reed, this volume, Chap. 7; Peterman and Peters,
this volume, Chap. 8). Such analysis may suggest, for
example, the sampling period, sampling frequency,
and sample size necessary for providing reliable
information in a suitable time frame and at an acceptable cost.
The past several sections have discussed characteristics of AM study designs and use of information
sources. I turn next to the topic of integrating study
results into statements of risk. The topic of risk is also
addressed by Peterman and Peters (this volume,
Chap. 8).
9.5 Risk Analysis and Risk Management
Lots of folks confuse bad management with destiny.
Kin Hubbard
9.5.1 Risk: speaking the same language between
analysis and management
The concept of risk has pervaded much of the adaptive management literature and much of land
management planning. However, researchers and
managers often use the term risk in vastly different
ways. This use can lead to, at best, confusion in interpreting results, or, at worst, misrepresentation of
study results. For adaptive management, risk is

defined as the expected value of adverse outcomes of

a management action.
It is useful in adaptive management to differentiate
risk analysis from risk management. In risk analysis,
the researcher lists possible outcomes, estimates their
likelihoods under one or more alternative future scenarios, and calculates their individual utilities by
weighting outcome likelihoods by outcome values.
These values are usually derived by managers and
may pertain to social, economic, or political interests,
as well as to legal regulations and objectives for resource management. Weighting outcome values with
outcome likelihoods helps the manager to determine
the overall risk of a management action. Then, in risk
management, the manager defines and applies their
risk attitude (their degree of risk-seeking, risk-neutral, or risk-avoidance behaviour) and then decides
on the best course of action. In separating risk analysis from risk management, the onus of articulating
outcome values, describing personal attitudes to risk,
and defining personal decision criteria is correctly
placed on the manager, not the researcher.
Formal decision analysis (Peterman and Peters,
this volume, Chap. 8) is a method for assessing the
risk of alternative outcomes of actions, taking uncertainty into account. Most managers do weigh the
relative values or outcomes, their likelihoods, and a
host of other factors that limit the decision space,
such as political acceptability, effects on career, and
effects on potential future decisions. However, decision analysis in your head is a poor substitute for
quantitative decision analysis. At a minimum, managers should explicitly reveal their own outcome
values, risk attidudes, and decision criteria.
9.5.2 Expressing uncertainties and unknowns
Uncertainty is a hallmark of science. However, managersas well as politicians, the media, the public,
and courtstypically view the issue of uncertainty
differently than do researchers. To the researcher,
uncertainty in adaptive management may represent
error of measurement, confounding effects of natural
variation, or other unstudied causes; such uncertainty is to be expected and results are to be treated with
due care (Kodrick-Brown and Brown 1993). In some
sense, the researcher may be certain of a particular
level of variance, and may still view adaptive management study results as strong evidence of some effect
of a management activity within some range of outcome. To the manager and others, however, such

135

variance may be seen as lack of evidence of effects, or

even as strong evidence of little or no effect, if the researcher cannot be certain of the outcome.
Scientific unknowns should be treated as a qualitatively different beast than scientific uncertainty. For
the researcher, uncertain outcomes can be quantified
as some measure of variation (such as variance or
confidence interval), but unknowns cannot be quantified at all. The influence of unknowns may be
deterministic, stochastic, strong, weak, or nonexistent; the researcher often simply cannot say. Again,
however, the manager might erroneously view unknowns as lack of evidence of effect and thus as
justification to proceed unless some contrary proof
is provided.
Managers also need to understand how to interpret results of adaptive management studies,
particularly in the context of a risk analysis. If adaptive management studies are designed as good
statistical investigations, then results can serve to either falsify, or fail to falsify, the null hypothesis being
tested; results can never prove a hypothesis.2 Failing to falsify the null hypothesis of no effect lends
only incremental credence to the management hypothesis. One of the ways to lend greater credence is
through replicate findings that would further corroborate results.
Therefore, researchers and managers (as well as
courts, media, and the public) must come to a common understanding of the concepts and implications
of scientific uncertainty, unknowns, risk and associated concepts of proof, errors, and statistical
falsification. Otherwise, results of adaptive management studies can be severely misrepresented,
misunderstood, and misapplied.
9.5.3 Unravelling the causal web: when is it our
fault and what can be done?
One of the main reasons for conducting adaptive
management studies of resource use or ecosystem elements is to determine not just patterns and trends
but also their causes. The manager should ask: What
is the true cause? Do our management activities directly affect the outcome, or merely set the stage for
other, more direct factors? To what degree do our
management activities influence the outcome?
Untangling the causal web in field situations can
be a great challenge. Seldom are causal factors affecting ecosystems single, simple, or easily quantified.
Most often, factors interact in complex ways, such as

with indirect and secondary effects, and through

feedback relations (Figure 9.1). Even in the simplest
model (Figure 9.1a), the relative contributions of
known and unknown causes must be estimated. In
simple models, the contribution from linear associationswhich may or may not be causalis indicated
by the value of the coefficient of determination R2 (or
adjusted R2), with the contribution from unknown
associations being 1R2. In more complex models,
(Figures 9.1b, c, d), estimating relative contributions
can be more involved. In real-world cases, it is not always evident which factors act as proximate causes,
which act as less direct causes, which are mere correlates with no causal relation, and which participate in
obligate feedback relations.
Of course, some relations are obvious, such as removal of forest canopy causing the local elimination
of obligate canopy-dwelling organisms. But less obvious effects or gradations, though difficult to unravel,
may be of great interest to the manager. For example,
what degree of effect does partial removal of the forest canopy have on local plant or animal populations
that are only facultatively associated with canopy environments? Might there be compounding, cumulative effects that exacerbate or ameliorate such effects,
such as wider regional loss of forest canopies, or
restoration of canopy conditions in adjacent stands?
To determine the relative influence of specific
management activities, the researcher may turn to
statistical techniques using estimation of partial
correlations. These methods help determine the
contribution of one factor, such as a management activity, given the effects of all other factors (e.g., other
activities, natural changes in environments, unknown
causes). Traditional analyses such as step-wise multiple regression help identify such partial influences.
Other, less well-known techniques such as regression
trees and path regression analysis (e.g., Schemske and
Horvitz 1988) can also be used.
Determining the relative influence of management
actions is vital for setting realistic expectations for
management results. For example, determining that
fragmentation of local forests affects breeding habitat
for migrating songbirds (Wilcove 1985) is only part of
the puzzle; loss of habitat on neotropical wintering
grounds is also a significant cause of declines in songbird populations. Therefore, changing local
management to reduce fragmentation should be expected to have only a partial impact on songbird
populations.

2 Some authors suggest that Bayesian analyses also can be interpreted as the testing of null hypotheses, that is, the prior probabilities.

136

(a)

(d)

E1

E1
E2

S1
S2

Ei

(b)

Sj

E2
S2

(c)

E1

?
S1

E2
. Causes and correlates: four examples. In all figures, S = wildlife species response; ? = unexplained variation due
to measurement error, experimental error, or effects of other environmental or species factors; solid arrows =
causal relations; dotted arrows = correlational relations that may or may not be causal. (a) In this simplest case,
some wildlife species response S, such as population presence or abundance, is assumed to be explained and
caused by some environmental factor E. (b) In a more complex case, we may be measuring one environmental
factor E1 when the real cause is another environmental factor E2. (c) Getting closer to the real world, a second
species response S2 may be part of the cause. (d) Most like the real world, with feedback relations among the
dependent (response) variables S. (Adapted from Morrison et al. 1998, Fig. 10.2.)

9.5.4 A dilemma for managers: when samples are

few and crises are many
One bane of adaptive management is that, in many
cases, the unique field conditions make it difficult to
correctly design statistical studies to identify and
quantify causes. Especially when studying landscapes,
ecosystems, rare or threatened species, and infrequent events, the major problems in the design of
such studies are small sample size and inability to
replicate conditions. In such circumstances, what can
the researcher do, and how should the manager interpret results? The answer may be found in several

courses of action: selecting correct indicators, merging disparate lines of evidence, and using statistical
procedures that take advantage of prior knowledge or
that function adequately with small sample sizes.
Selecting correct indicators
Indicators that are objective, repeatable measurements, whose quality is documented quantitatively
should be selected. For adaptive management studies,
an indicator should (1) respond rapidly to changes,
(2) signal changes in other variables of interest, (3) be
monitored efficiently, and (4) be causally linked to

137

changes in stressors. Most ecological indicators

purported to fit these criteria usually fail (Block et al.
1987; Patton 1987; Landres et al. 1988). For example,
the Northern Spotted Owl, often selected by USDA
Forest Service as an old-growth indicator, may
serve criterion (4), but fails with the other three criteria: spotted owls have low reproductive rates and
long life spans, so they respond slowly to changes;
changes in their populations may not necessarily correlate well with other desired facets of old-growth
forests (e.g., habitat for anadromous fish); and their
population trends are terribly costly to monitor.
Indicators that do meet these criteria include soil
arthropods as indicators of soil productivity (McIver
et al. 1990; Moldenke and Lattin 1990; Pankhurst et
al. [editors] 1997); butterfly diversity as an indicator
of overall ecosystem diversity (Kremen 1994); and
some arboreal lichens as indicators of air quality
(Stolte et al. 1993; Geiser et al. 1994) or persistence of
old forests (Tibell 1992). See Murtaugh (1996) for a
review of the statistical basis of ecological indicators.

tions, the data from several studies might be combined into an overall regression. This regression
might suggest a significant correlation between
clearcutting and grizzly bear populations. However,
grizzly bears within individual study areas might respond differently to clearcutting because they come
from different geographic areas, latitudes, or forest
types. Thus the correlation may reflect these differences between populations, rather than any
treatment effect. The incorrect conclusion of correlation would arise because such an analysis violates an
assumption underlying regression: that the data
come from the same statistical population with the
same causal mechanisms. On the other hand, a formal meta-analysis approach would analyze results
from each study with differences among studies as an
explanatory factor. CI has great utility, especially
where powerful experimental studies are difficult.
However, managers and researchers must be careful
in its use, ensuring that studies are truly from the
same causal web.

Merging disparate lines of evidence

Merging different study results is a second tactic that
can help in identifying causal relations when good
experimental design is impossible or impractical. In
statistics, this process is called combining information (CI). Draper et al. (1992) provide a useful
overview of various CI techniques, including methods of meta-analysis (Hedges and Olkin 1985) that
can be useful in conservation research (FernandezDuque and Valeggia 1994). For example,
meta-analysis was used by Burnham et al. (1996) to
determine overall trends of Northern Spotted Owls
by combining results from individual population
demography studies.
CI is not a panacea, as it can be fraught with
difficulties such as matching consistency and representativeness among studies designed for different
initial objectives. Still, the researcher may wish to use
CI methods to merge lines of evidence taken from
available information. This available information
could include anecdotes and local experience, retrospective studies, observational studies, experimental
manipulations, and demonstrations. The reliability
of each source for inferring causes should be judged
very carefully.
In contrast to formal meta-analysis, simply pooling data from different studies could lead to spurious
and misleading conclusions. For example, to assess
the impact of clearcutting on grizzly bear popula-

Using statistical procedures that take advantage of

prior knowledge
Bayesian statistics were developed specifically for
using prior knowledge and incrementally gathered
field data (Ver Hoef 1996). Bayesian statistical techniques include empirical Bayes and sequential Bayes
procedures, in which initial estimates of the likelihood of conditions become incrementally adjusted
and refined over time as new evidence is gathered
(e.g., Gazey and Staley 1986; Link and Hahn 1996).
Expert opinion, existing literature and data, retrospective studies, and non-experimental studies can all
be used to establish preliminary values of prior probabilities in a Bayesian analysis. Bayesian methods
were reviewed by Bergerud and Reed (this volume,
Chap. 7), who advocate their use to incorporate accumulated knowledge of experts.

138

9.6 Conclusions and Recommendations

Knowledge is the small part of ignorance that we
arrange and classify.
Ambrose Bierce
9.6.1 A decision tree for managers
The six stages of adaptive management and sources
of information appropriate for each stage are presented in Table 9.1. This table can be used by
managers as a decision tree to guide the choice of

139

Literature
review

"

"

[ ]

[ ]

(pilot)

(pilot)

[ ]b

[ ]b

Experimental
study

"

"

Nonexperimental
study

"
"
"

Retrospective
study

"
"
"

Anecdote

"
"
"

Demonstration

Monitoring design is determined at the Design project stage.

Expert
judgement

a Experimental and nonexperimental studies can provide information on patterns, correlates, etc., but typically these studies will not be done by taking advantage of management
actions, but rather as part of applied research.
b All else being equal, if the cost of conducting a nonexperimental study is significantly less than that of an experimental study, choose the former.
c In the Evaluation stage, existing information based on literature, expert judgement, and retrospective analysis is updated using data collected from the management experiment to assess the effect or outcome of an action. It can also be used to determine the relative plausibility of suspected causes, and to estimate the prior probabilities in a Bayesian
analysis.

6. Adjust management action

5. Evaluate
Interpretation

4. Monitor

3. Implement

2. Design project
Determine treatments to
implement, sample size,
effect size, power, etc.

1. Assess problem
Identify potential impacts of
management actions and the
potential reasons for them.
Identify patterns and trends
Identify correlates
Identify potential causes
of suspected impact a

AM stages

. Stages of an adaptive management (AM) project and sources of information appropriate for each stage. This table can be used by managers as a
decision tree to guide (1) the choice of study for each AM stage (reading across rows), and (2) the use of existing information (reading down
columns). = recommended sources; " = not recommended; [ ] = most recommended for a given project stage; = does not apply;
(pilot) = use for pilot study only. See Chapter 1 for full descriptions of AM stages.

study for each stage of adaptive management, as well

as to guide the use of existing information.
At the problem assessment stage, existing information is valuable for identifying potential impacts of
management actions. At the project design stage,
pilot studies (experimental or nonexperimental) are
recommended for fine tuning the study methodology. Pilots can be used to estimate variability in the
response variables; these estimates can then be used
to determine sample size, effect size, and power for
the study. Controlled experiments allow the strongest
inference about the actual impacts of management
actions. Once a study has been implemented, relevant data are collected through a monitoring process.
The data are then analyzed using appropriate statistical methods to answer questions set out at the
beginning of the adaptive management project. In
the evaluation stage, existing knowledge (based on
literature, expert judgement, and retrospective analysis) is updated using data collected from the
management experiment to assess the effect or outcome of an action. Using Bayesian analysis, existing
knowledge together with collected data can also be
used to determine the relative plausibility of suspected causes. Management actions are then adjusted
based on this updated knowledge. During the course
of the management experiment, new questions may
arise that then lead to further problem assessment,
project design, implementation, and so on, in a continuous cycle of learning.

making management decisions and defending such

decisions legally, politically, and scientifically.
Short of this ideal, both researchers and managers
have their work cut out for them. They should maximize the use of available information, but not draw
undue conclusions about causes. It may be useful to
explicitly array the various available lines of evidence
and to articulate the confidence in identifying causes
from each. Managers and researchers must look for
similarities and disparities among lines of evidence
and investigate reasons for the differences. Moreover,
they should seek peer review to ensure appropriate
and rigorous use of various sources of information.
Repeatability of findings and suspected causes is the
basis for true scientific understanding and predictability.
Real-world adaptive management problems are
often complicated by time exigencies or finite funding so that powerful experiments are not possible.
If the study design must be compromised, then the
ramifications of drawing incorrect conclusions
should be thought out. When turning to field experts
for their professional opinions, managers should be
aware of potential problems such as motivational
and personal bias.
Managers should weigh the benefits and cost of a
more or less rigorous approach. For, in the end, reliable knowledge is a hard-won but essential
commodity for ensuring successful conservation
practices for future generations.

9.6.2 Is there a best statistical approach to adaptive management?

The answer to this question is an unqualified yes.
The best approach for answering the questions Did
this action have the desired effect? and Are the
basic assumptions underlying management decisions
correct? is to use controlled, randomized experiments with sufficient sample sizes and duration. This
approach provides the best understanding of causal
relations and the best basis for validating predictive
modelsassuming that the models can provide
testable predictions.3 Nevertheless, the informativeness of a statistical approach must also be weighed
against its costs (ecological, social, and economic).
Ultimately, designing management actions as controlled, randomized experiments will provide the best
evidence for managers who face the difficult task of

Acknowledgements
My thanks to Brian Nyberg and Brenda Taylor for
suggesting topics to cover in this chapter and for
their technical reviews of the manuscript. Roger
Green, Rick Page, Martin Raphael, and Ian Thompson also provided technical reviews. Thanks to Vera
Sit and Brenda Taylor for their fine edits. My gratitude to Tim Max for his technical review and for
thoughtful discussions on the role of statistics in
adaptive management.

3 As expressed by one statistician, if predictive models are so complex that they become essentially untestable, then they are nothing more
than belief structures and their relation to science is questionable at best (T. Max, pers. comm., 1997).

140

References
Anderson, J.L. [n.d.]. Errors of inference. This volume.
Block, W.M., L.A. Brennan, and R.J. Gutierrez. 1987.
Evaluation of guild-indicator species for use in
resource management. Environ. Manage.
11:2659.

Gentiol, S. and G. Blake. 1981. Validation of complex

ecosystem models. Ecol. Modelling 14:2138.
Green, R.H. 1979. Sampling design and statistical
methods for environmental biologists. J. Wiley,
New York, N.Y.
Hedges, L.V. and I. Olkin. 1985. Statistical methods
for meta-analysis. Academic Press, New York,
N.Y.

British Columbia Ministry of Forests. 1996. Adaptive

management: learning from our forests.
Brochure. Integrated Resources Policy Branch,
Victoria, B.C.

Hurlbert, S.H. 1984. Pseudoreplication and the design

of ecological field experiments. Ecol. Monogr.
54:187211.

Burnham, K.P., D.R. Anderson, and G.C. White.

1996. Meta-analysis of vital rates of the northern spotted owl. Stud. Avian Biol. 17:92101.

Kodric-Brown, A. and J. H. Brown. 1993. Incomplete

data sets in community ecology and biogeography: a cautionary tale. Ecol. Applic. 43:73642.

Carpenter, S.R. 1990. Large-scale perturbations: opportunities for innovation. Ecology 71:203843.

Kremen, C. 1994. Biological inventory using target

taxa: a case study of the butterflies of Madagascar. Ecol. Applic. 4:40722.

Collopy, M., J. Cannell, S. Daniels, D. Dippon, E.

Gaar, G. Grant, B. Mulder, C. Philpot, J. Steffenson, and F. Swanson. 1993. Implementation
and adaptive management. In Forest ecosystem
management: an ecological, economic, and social assessment. Forest Ecosystem Management
Assessment Team. U.S. Dep. Agric. For. Serv.,
Washington, D.C.

Landres, P.B., J. Verner, and J.W. Thomas. 1988. Ecological uses of vertebrate indicator species: a
critique. Cons. Biol. 2:31628.
Ligon, J.D. and P.B. Stacey. 1996. Land use, lag times
and the detection of demographic change: the
case of the acorn woodpecker. Cons. Biol.
10:8406.

Draper, D., D.P. Gaver, Jr., P.K. Goel, J.B. Greenhouse, L.V. Hedges, C.N. Morris, J.R. Tucker,
and C.M. Waternaux. 1992. Combining information: statistical issues and opportunities for
research. National Academic Press, Washington, D.C. Contemporary statistics No. 1.

Link, W.A. and D.C. Hahn. 1996. Empirical Bayes estimation of proportions with application to
cowbird parasitism rates. Ecology 77:252837.

Fernandez-Duque, E. and C. Valeggia. 1994. Metaanalysis: a valuable tool in conservation

research. Cons. Biol. 8:55561.

McIver, J.D., A.R. Moldenke, and G.L. Parsons. 1990.

Litter spiders as bio-indicators of recovery after
clearcutting in a western coniferous forest. NW
Environ. J. 6:4102.

Gazey, W.J. and M.J. Staley. 1986. Population estimation from mark-recapture experiments using a
sequential Bayes algorithm. Ecology 67:94151.
Geiser, L.H., C.C. Derr, and K.L. Dillman. 1994. Air
quality monitoring on the Tongass National
Forest: methods and baselines using lichens.
U.S. Dep. Agric. For. Serv., Alaska Region. Petersburg, Alaska. R10-TB-46.

Loehle, C. 1990. Indirect effects: a critique and alternate methods. Ecology 71:23826.

McKinlay, S.M. 1975. Comprehensive review on design and analysis of observational studies. J.
Am. Statist. Assoc. 70:50320.
______. 1985. Observational studies. In Encyclopedia
of statistical sciences. Vol. 6. J. Wiley, New
York, N.Y. pp. 397401.

141

Marcot, B.G. 1986. Summary: biometric approaches

to modeling: the managers viewpoint. In
Wildlife 2000: modeling habitat relationships of
terrestrial vertebrates. J. Verner, M.L. Morrison, and C.J. Ralph (editors). Univ. Wis. Press,
Madison, Wis. pp. 2034.
Marcot, B.G., M.A. Castellano, J.A. Christy, L.K.
Croft, J.F. Lehmkuhl, R.H. Naney, R.E. Rosentreter, R. E. Sandquist, and E. Zieroth. 1997.
Terrestrial ecology assessment. In An assessment of ecosystem components in the interior
Columbia Basin and portions of the Klamath
and Great Basins. T.M. Quigley and S.J. Arbelbide (editors). Pac. N. Res. Sta., Portland, Oreg.
U.S. Dep. Agric., For. Serv. Gen. Tech. Rep.
PNW-GTR-405. pp. 14971713.
Max, T.A., R.A. Souter, and K.A. OHalloran. 1990.
Statistical estimators for monitoring spotted
owls in Oregon and Washington in 1987. U.S.
Dep. Agric. For. Serv., Portland Oreg. Res. Pap.
PNW-RP-420.
Moldenke, A.R. and J.D. Lattin. 1990. Density and diversity of soil arthropods as biological probes
of complex soil phenomena. NW Environ. J.
6:40910.
Morrison, M.L., B.G. Marcot, and R.W. Mannan
[1998]. Wildlife-habitat relationships: concepts
and applications. 2nd ed. Univ. Wis. Press,
Madison, Wis. In press.
Murtaugh, P.A. 1996. The statistical evaluation of
ecological indicators. Ecol. Applic. 6:1329.
Nemec, A.F.L. [n.d.]. Design of experiments. This
volume.
Oliver, I. and A.J. Beattie. 1993. A possible method for
the rapid assessment of biodiversity. Cons. Biol.
7:5628.
______. 1996. Designing a cost-effective invertebrate
survey: a test of methods for rapid assessment
of biodiversity. Ecol. Applic. 6:594607.
Pankhurst, C., B. Doube, and V. Gupta (editors).
1997. Biological indicators of soil health. Oxford Univ. Press, New York, N.Y.
Patton, D.R. 1987. Is the use of management indicator species feasible? W. J. For. 2:334.

142

Peterman, R.M. and C. Peters. [n.d.]. Decision analysis: taking uncertainties into account in forest
resource management. This volume.
Richey, J.S., R.R. Horner, and B.W. Mar. 1985. The
Delphi technique in environmental assessment.
II. Consensus on critical issues in environmental monitoring program design. J. Environ.
Manage. 21:14759.
Richey, J.S., B.W. Mar, and R.R. Horner. 1985. The
Delphi technique in environmental assessment.
I. Implementation and effectiveness. J. Environ.
Manage. 21:13546.
Routledge, R.D. [n.d.]. Measurements and estimates.
This volume.
Schemske, D.W. and C.C. Horvitz. 1988. Plant-animal interactions and fruit production in a
neotropical herb: a path analysis. Ecology
69:112838.
Schuster, E.G., S.S. Frissell, E.E. Baker, and R.S. Loveless. 1985. The Delphi method: application to
elk habitat quality. U.S. Dep. Agric. For. Serv.
Res. Pap. INT-353.
Schwarz, C.J. [n.d.]. Studies of uncontrolled events.
This volume.
Smith, G.J. [n.d.]. Retrospective studies. This volume.
Stolte, K., D. Mangis, R. Doty, and K. Tonnessen.
1993. Lichens as bioindicators of air quality.
Rocky Mtn. For. Range Exp. Sta., Fort Collins
Colo. U.S. Dep. Agric. For. Serv. Gen. Tech.
Rep. RM-224.
Thomas, L. and K. Martin. 1996. The importance of
analysis method for breeding bird survey population trend estimates. Cons. Biol. 10:47990.
Tibell, L. 1992. Crustose lichens as indicators of forest
continuity in boreal coniferous forests. Nor. J.
Bot. 12:42750.
Tiebout, H.M. III and K.E. Brugger. 1995. Ecological
risk assessment of pesticides for terrestrial vertebrates: evaluation and application of the U.S.
Environmental Protection Agencys quotient
model. Cons. Biol. 9:160518.

Ver Hoef, J.M. 1996. Parametric empirical Bayes

methods for ecological applications. Ecol. Applic. 6:104755.
Wilcove, D.S. 1985. Forest fragmentation and the decline of migratory songbirds. Ph.D. thesis,
Princeton Univ., Princeton, N.J.
Yaffee, S.L., A.F. Phillips, I.C. Frentz, P.W. Hardy,
S.M. Maleki, and B.E. Thorpe. 1996. Ecosystem
management in the United States. Island Press,
Covelo, Calif.
Zuboy, J.R. 1981. A new tool for fishery managers: the
Delphi technique. N. Am. J. Fish. Manage.
1:559.

143

144

GLOSSARY

A posteriori: Referred to after the data have been collected and examined.
A priori: Referred to before the data are collected and
examined.
Accuracy: The nearness of a measurement to the
actual value of the variable being measured.
Active adaptive management: Management is designed as an experiment to compare alternative
actions (treatments) or discriminate among alternative hypotheses about how the system responds
to actions. Active adaptive management can involve deliberate probing of the system to
identify thresholds in response and clarify the
shape of the functional relationship between actions and response variables.
Alternative hypothesis: A claim or research hypothesis that is compared with another (usually null)
hypothesis.
Analysis of variance (ANOVA): A group of statistical procedures for analyzing continuous data
sampled from two or more populations, or from
experiments in which two or more treatments are
used. ANOVA procedures partition the variation
observable in a response variable into two basic
components: (1) variation due to assignable causes
and (2) uncontrolled or random variation. Assignable causes refer to known or suspected sources of
variation from variates that are controlled (experimental factors) or measured (covariates) during
an experiment. Random variation includes the effects of all other sources not controlled or
measured during the experiment.
Analytical survey: A type of nonexperimental study
where groups sampled from a population of units
are compared.
Autocorrelation: Occurrence when consecutive measurements in a series are not independent of one
another. Also called serial correlation.
Bayes decision: The optimal decision identified
when uncertainties are considered using a formal
decision analysis.
Bias: The deviation of a statistical estimate from the
quantity it estimates. Bias can be a systematic
error introduced into sampling or testing. Positive
bias will overestimate the parameter; negative bias
will underestimate it.

Blocking: A design technique where experimental

units are grouped into homogeneous blocks, according to some identifiable characteristic(s).
Successful blocking reduces the experimental error
that results from variation among heterogeneous
units.
Conditional probability, P(A|B): The probability of
the event A given that a related event B has taken
place.
Confidence limits: Confidence limits indicate the
precision of a parameter estimate. If samples of
size n were repeatedly obtained from the population and constructed (1)% confidence limits
for each, the expected result would be that
100(1) out of 100 confidence limits would
contain the true parameter.
Confounding: Confounding occurs when one or
more effects cannot be unambiguously attributed
to a single factor or interaction.
Control: A treatment level included in an experiment
to show what would have happened if no treatments had been applied to the experimental
material.
Controlled experiment: An experiment in which the
experimenter controls the treatments to be compared, and can randomly assign experimental
units to the treatments. Also called a designed experiment.
Correlation: A measure of the strength of the linear
relationship between two random variables. A
strong correlation between two random variables
does not necessary signify a causal relationship between them.
Covariate: A variable that influences the response
but is unaffected by any other experimental factors. Including a covariate in the analysis may
increase the power of the analysis to detect treatment effects.
Decision analysis: A structured, formalized method
for ranking management actions that are being
considered. It quantitatively takes into account
uncertainties.
Delphi technique: A procedure for interviewing experts and capturing their expertise by striving to
reach consensus in an expert panel or group setting.

145

Effect size: The treatment effect the experimenter

wants to be able to detect. Effect size influences
the statistical power of an analysis: larger effect
size yields greater statistical power.
Expected value of an outcome: The weighted average
outcome, where each outcome is weighted by the
probability assigned to its branch on the decision
tree. Summing these weighted outcomes for each
management action gives the expected value of
that action.
Experimental design: A plan for assigning treatments to experimental units and the statistical
analysis associated with the plan. It includes formulation of statistical hypotheses, choice of
experimental conditions, specification of the
number of experimental units required and the
population from which they are to be sampled, assignment of treatments to experimental units,
determination of the dependent variables to be
measured, and the statistical analysis to be performed.
Experimental error: Any variation, including sampling (measurement) error and natural variation
error, that cannot be explained by the experimental factors.

planned event is compared with a control site not

affected by the event. Impact surveys are typically
used to investigate the effects of large-scale, unreplicated events. Types of impact surveys include
BACI (Before-After-Control-Impact) where variables in both an impact (treatment) and control
site are compared before and after some event,
and BACI-P (Before-After-Control-Impactpaired)an extension of BACI where control and
impact sites are sampled at the same points in
time, both before and after the event.
Null hypothesis: A statistical hypothesis that states
that there is no difference between the true
value of a parameter and the hypothesized value,
or that there is no effect of a treatment.
Observational survey: A type of nonexperimental
study where results from two units or sites are
compared. Results are applicable only to the units
sampled, and cannot be extrapolated to other
units or sites.
Parameter: A numerical characteristic of a population. It is often estimated by a sample statistic.

Experimental factor: Any treatment or variable that

is controlled in an experiment, either by physically
applying a treatment to an experimental unit or by
deliberately selecting a unit with a particular characteristic.

Passive adaptive management: Managers implement

what they assume, based on existing information,
is the best action (i.e., the action most likely to
produce the desired outcome). Adjustments are
made when actual outcomes deviate from predictions. The limitation of passive adaptive
management is that it can be difficult to determine
why actual outcomes deviate from predictions.

Experimental unit: The entity to which one treatment (level of one or more factors) is applied.
Also called a treatment unit.

Power (1): The probability of correctly rejecting a

null hypothesis when it is actually false. Also called
statistical power, or the power of a test.

Explanatory variable: A variable that is thought to

provide information on the value of the response
variable.

Precision: The closeness to each other of repeated

measurements of the same quantity. Precision
should not be confused with accuracy. Imagine a
dart board: accuracy refers to the distance of the
dart from the bulls-eye; precision refers to how
tightly grouped repeated dart throws are.

Homogeneous: Experimental units are homogeneous when they do not differ from one another
in any systematic fashion and are as alike as possible on all characteristics that might affect the
response.
Hypothesis: A tentative assumption, adopted to account for certain facts that can be tested.
Hypothesis testing: A type of statistical inference for
assessing the validity of a hypothesis by determining whether it is consistent with the sample data.
Impact survey: A type of nonexperimental study
where one site affected by some planned or un-

146

Prospective study: A study where actions (treatments) have not yet been applied, and data have
not yet been collected. Prospective studies may be
either experimental or nonexperimental. Contrast
with retrospective study.
Pseudoreplication: Refers to various violations of the
assumption that replicated treatments are independent. A common form of pseudoreplication
occurs when multiple subsamples from one treat-

ment unit, rather than samples from multiple

(replicated) treatment units, are used to calculate
the statistical probability of a treatment effect.
P-value: Probability of obtaining a value for a test
statistic that is as extreme as or more extreme than
the observed value, assuming the null hypothesis
is true. In classical hypothesis testing, the null hypothesis is rejected when the P-value is less than
the chosen significance level ().
R2: A statistic that assesses how well a regression
model describes the relationship between the dependent and independent variables. For
comparisons of several models using the same
data, it is more appropriate to use the adjusted
R2R2 adjusted by the model degrees of freedom.
Randomization: Treatments are randomly assigned
to the experimental units, so that each unit has a
known and independent chance of being allocated
a particular treatment. Randomization protects
against possible bias (systematic error) by ensuring that all unmeasured factors are more or less
evenly distributed among treatments.
Random sampling: A scheme for choosing subjects
from a population, so that each member of the
population has a known (often equal) and independent chance of being selected. Random
sampling allows you to generalize the results of the
experiment to the population from which the
sample was drawn.
Regression: A relationship where the magnitude of
one variable (the dependent variable) is determined in part by the magnitude of another
variable (the independent variable).
Replication: Replication involves applying the same
combination of factors to more than one experimental unit. Replication is a means of assessing
the variability that is not attributable to the treatment.
Response variable: A variable measured to assess the
outcome of an experiment. In regression, the response variable is referred to as the dependent
variable.
Retrospective study: A study that uses data already
collected for other purposes or actions (treatments) that have already been implemented. A
retrospective study is a type of uncontrolled (nonexperimental) study. Contrast with prospective
study.

Risk: The expected loss associated with an outcome

or decision. Risk is the product of the possible
magnitude of a loss and the probability of it occurring.
Sample: A subset of measurements or observations
taken from a population. Conclusions about the
characteristics of the population can be drawn
from the characteristics of the sample.
Sampling design: A plan that describes the nature of
the sampling units, the number of sampling units,
the method of selection, and the variables to be
measured.
Sampling unit: A basic unit selected for sampling.
Scientific (research) hypothesis: A testable proposition that is tentatively adopted to account for
observed facts and to guide investigation.
Sensitivity analysis: A procedure for assessing the
degree to which predicted outcomes vary with
changes in assumptions about parameter values.
Significance level (): The probability of making a
Type I error (i.e., rejecting a true null hypothesis).
In hypothesis testing, indicates the maximum
amount of Type I error the experimenter is willing
to tolerate.
Standard deviation: A measure of the dispersion
(variability) of the data. The deviations of individual observations from the sample mean are
squared, the squares are averaged, and the square
root of the result is calculated.
Standard error: The standard deviation of a sample
statistic. Standard deviation is a measure of the
dispersion of the individual observations from
their mean; standard error is a measure of the dispersion of repeated sample statistics from their
mean.
Statistic: A numerical characteristic that is computed
from a sample of observations and that estimates a
population parameter.
Statistical hypothesis: It states the scientific hypothesis in precise, quantitative terms, often as a
variable whose sampling distribution can be described statistically.
Statistical independence: Observations are statistically independent if the value of one of the
observations does not influence the value of any
other observations. Simple random sampling produces independent observations.

147

Statistical inference: The act of drawing a conclusion

about the characteristics of a population by analyzing the characteristics of a sample (i.e.,
generalizing about the whole, based on information from a part of the whole).
Stochastic process: A process that is not completely
predictable.
Stratification: Survey units are grouped into homogeneous groups, according to some identifiable
characteristic(s). Each stratum is then surveyed.
Stratification in sampling is analogous to blocking
in experimental design.
Systematic sampling: A sampling scheme where
every kth unit in a population is sampled (with the
result that sampling points are a fixed distance
apart).
Type I error: The error of rejecting a null hypothesis
that is true.
Type II error: The error of not rejecting a null hypothesis that is false.
Uncontrolled experiment: An experiment in which
the investigator does not control the selection of
treatments or the assignment of treatments to the
experimental units.
Variance: A measure of the dispersion (variability) of
the data. The variance is the square of the standard
deviation.

148