Journal of Behavioral Decision Making, Vol.

I , 223-237 (1988)

An Availability Bias in Professional Judgment

LAURETTE DUBE-RIOUX AND J. EDWARD RUSSO
Cornell University, lthaca, NY, U.S.A

ABSTRACT
When branches of a fault tree are pruned, subjects d o not fully transfer the
probability of those branches to the ‘all other’ category. This underestimation of
the catch-all probability has been interpreted as an ‘out of sight, out of mind’form
of the availability bias. The present work replicates this underestimation bias with
professional managers. It then demonstrates the effectiveness of a corrective tactic,
extending the tree by generating additional causes, and also reveals that more easily
retrieved short-term causes dominate the generation process. These results d o not
differ across managers’ culture, education or experience. After evaluating such
alternative explanations as category redefinition, we conclude that availability is a
major cause, though possibly not the sole cause, of the underestimation bias.
KEY WORDS Availability Bias Debiasing Heuristic Likelihood estimation
Reason generation

In this work we investigate an availability phenomenon in the estimation of event probabilities. Before
turning to the specific task and context of the present research, we distinguish two uses of the term
availability. The original meaning denotes the ease of recall (or imagination) of instances of a specified
event (Tversky and Kahneman, 1973). When an event’s likelihood is estimated by this form of avail-
ability, the availability heuristic is said to be used. A bias occurs when such availability-based estimates
are distorted by the influence on retrieval of such factors as the concreteness, drama, familiarity,
recency, relevance, similarity or vividness of instances (Billings and Schaalman, 1980; Folkes, 1988;
Shedler and Manis, 1986; Taylor, 1982).
Availability has a second meaning, whether the event itself is available at the time subjective pro-
babilities are estimated. This all-or-none form partitions the continuum of ease of recall into the
limiting case of failure to recall versus all successful recalls regardless of difficulty. An estimation bias
occurs when those events not brought to mind are overlooked and the probabilities of retrieved events
are distorted upward. The case of retrieval failure is an availability phenomenon, but not an instance of
the availability heuristic. In what follows it is the retrieval failure form of availability that will be
investigated.
Because availability can depend on what is present in the environment as well as what can be
retrieved from memory, external decision aids come into play. One such aid is a fault tree, a hierarchical
delineation of the possible causes of some undesired terminal event (see, e.g., Von Winterfeldt and
Edwards, 1986). Exhibit 1 presents a fault tree for the failure of a restaurant.

The underestimation bias

For fault trees to be successful they must be complete in the sense of including every important branch.
0894-32571 881040223-15$07.50 Received 12 January 1988
0 1988 by John Wiley & Sons, Ltd. Revised 13 July 1988
 I Reataurant F a i l u r e Due t o Decreaaing P r o f i t . I

I
Decreaaing Revenuea

7
DecrOa8ing Number
of Cuatoura
I n c r e a a i n g Food Coata 1 I n c r e a e i ng Labor Costa 1 I n c r e a a i n g Overhead Coat

1
Incorrect Pricing
Unclear Image o f
Property
Changing Atmoaphere
Decreaaed Perceived
Value by C u a t o u r s
Incorrect Pricing
Inadequate S e r v i c e Pace
Improper Purchaai ng
and R e c e i v i n g
Menu V a r i e t y (Too
L i m i t e d o r Too
I n c r e a a i n g Overtime
Scheduling
Decreaaing P r o d u c t i v i t y
o f Employee.
7H i g h Debt S e r v i c e Coat
Poor Design o f t h e
Faci Ii t y
H i g h Occupancy Coata
Outdated Restaurant Poor Yerchandiaing Extenai ve) Poor O r g a n i z a t i o n a l Inadequate C a p i t a l
Concept Lack o f Employoe Poor Sale. F o r e c a a t i n g Climate Structure
Inadequate Promotion Yotivation Change. i n S u p p l i e r Union Rules Low Sales Volume
and A d v e r t i a i n g Changing C u a t o m r D i n i n g Yarket Improper P h y s i c a l Improper Growth Rate
Changing Eapectationa Out Budget H i g h Waate and Layout and Equipment High Adminiatrative
o f Cuatomra Changing N i x o f Food/ L e tt o v e r a Poor Employee S e l e c t i o n Coata
Poor Q u a l i t y o f Food Beverage Sa lea Inadequate Number o f and T r a i n i n g I n c o r r e c t Work Method
Changing Competition Inconeiatent Quality P r o d u c t i o n Employeea High Employee Turnover o f Personnel
Poor Q u a l i t y o f Servica o f Food High Theft Menu V a r i e t y Too E x t e n a i v Improper Buaineas
Changing Cuatomer Changing Competition Poor Superviaion Poor Supervi a i o n Hour.
Demographic. Improper P o r t i o n S i r e Incorrect Sire Inadequate Wage S t r u c t u r e Outdated F a c i l i t y and
Lack o f Yenu V a r i e t y Changing Customer o f Food P o r t i o n s I n e f f i c i e n t Employee Equipment
Changing L o c a t i o n Demographica Improper Storage Schedu I ing High I n f l a t i o n
~
Characteriatica Changing Cuatomer Taatea and I s a u i n g I n c r e a a i n g B e n e f i t Costa Poor C r e d i t R a t i n g
A l l Other A l I Other I A l I Other A l I Other A l I Other
P
%
Exhibit 1.
 L. Dubt-Rioux and J. E. Russo Availability Bias 225

If incomplete, an availability bias may distort the estimated frequency of causes. Fischhoff, Slovic and
Lichtenstein (1978) demonstrated this danger by using a fault tree with some branches pruned. One
group of subjects saw a full fault tree like Exhibit 1, and estimated the likelihoods of its major branches
plus a catch-all branch for ‘all other’ unlisted causes. A second group performed the same estimation
task on a pruned tree, where some branches had been omitted. Normative performance requires that the
relative frequencies assigned to the pruned branches be transferred to the catch-all category. Fischhoff,
et al. found this transfer to be less than a quarter of what it should have been. They interpreted this
result as an availability phenomenon, ‘what is out of sight is out of mind’ (p. 333).
We shall refer to this failure to transfer the total probability of pruned branches to the catch-all
category as an underestimation bias. This neutral term avoids labeling the result by its explanation
before that explanation has been shown to be the only one possible. Fischhoff, et al. found the underesti-
mation bias to be quite robust. It was not eliminated by (1) using professional subjects rather than
college students, (2) increasing the level of detail in each branch, or (3) instructing subjects to focus on
the ‘all other’ branch.
The goal of the present work is to investigate both the basis and remediation of the underestimation
bias. First, for a new group of professionals we test the core finding that a pruned fault tree leads to
biased probability estimates. Managers in the hospitality industry used the fault tree for the failure of a
restaurant shown in Exhibit 1. However, the main focus of the present work is three extensions of the
underestimation bias: a technique for correcting the bias, an investigation of the role of short-term
versus long-term causes, and a search for differential effects of culture and education.

Removing the bias

We propse a simple corrective technique, instructing subjects to extend the tree by recalling, listing, and
assigning probabilities to additional causes of restaurant failure. These subject-generated causes would
otherwise remain undifferentiated in the catch-all category. Several tactics for improving the estimation
of probabilities have succeeded, including imagining events (Carroll, 1978), explaining events
(Sherman, Skov, Hervitz and Stock, 1981), developing scenarios for an event (Gregory, Cialdini and
Carpenter, 1982), listing pro and con reasons (Koriat, Lichtenstein and Fischhoff, 1980; Hoch, 1984)
and considering opposite possibilities (Lord, Lepper and Preston, 1984). Like these techniques,
extending the fault tree is a specific action, in contrast to unspecific exhortations to avoid the bias which
tend to be unsuccessful. It reflects a concern with devising remedies for the long list of judgmental
shortcomings exposed in the last two decades (e.g., Fischhoff, 1982).
A technique similar to extending the fault tree was used by Mehle, Gettys, Manning, Baca and Fisher
(198 1). Informed of only a single course in which a student was enrolled, their subjects’ task was to
estimate the relative probabilities of four possible majors, three specific ones and the catch-all set of ‘all
others’ majors. One group of subjects was also instructed to list some of the other possible majors that
would fall in the catch-all category. This specification of other majors reduced the underestimation bias
by 19%.

Short-term versus long-term causes

Our second extension of the basic result is an investigation of the role of short-term versus long-term
causes. Because we expect short-term causes to be more easily retrieved, their differential frequency of
generation (when subjects extend the tree) provides direct support of an availability explanation of the
underestimation bias.
Short-term causes are ones whose natural occurrence is more apparent to restaurant managers either
because each one occurs more often or because attention is drawn to them in a bottom-up fashion. They
also tend to have shorter time spans for their origin and their correction, hence the label short-term. In
the case of a decreasing number of customers (the first branch in Exhibit l), short-term causes include
 226 Journal of Behavioral Decision Making Vol. 1, Iss. No. 4

inadequate promotion (including advertising) and a lack of menu variety. Note that promotion
decisions must be made continually, in part because salespeople from newspapers and radio stations
often contact restaurant managers. Similarly, menu variety must be reconsidered with seasonal changes
in the availability of ingredients or as customers either complain or praise certain menu items. Both
problems can arise and be rectified relatively quickly. In contrast, changing customer demographics and
unclear restaurant image are long-term causes. Both their origin and correction take more time; and
such problems are encountered by managers less frequently, and usually only when they look for them.
The ease of retrieval of short-term causes may be facilitated not only by a greater frequency of stored
instances, but also by specific retrieval strategies. For instance, recalling ‘the things customers complain
about’ is apt to produce more short- than long-term causes. In general, we expect that the short-term
causes will be more easily recalled, even though they may be neither more important nor more numerous.

Differences in education and culture

The third contribution of the present work is a test for individual differences in our sample of
professional managers. We test for differences in experience (where Fischhoff, et al. found only a small
improvement when professional auto mechanics were tested instead of college students), and for dif-
ferences in education and culture. There has been considerable speculation about the decision styles of
non-Westerners, but relatively few empirical studies. In probabilistic judgment the major exception is
the work of Wright and Phillips (1980; Phillips and Wright, 1977; Wright, Phillips, Whalley, Choo, Ng,
Tan and Wishuda, 1978) who report substantial differences between Asian and British subjects. For
instance, the latter were better calibrated, Le., more accurate in estimating the probability that they had
correctly answered a factual question (such as whether the Panama or Suez Canal is longer). Access to
an international sample of hospitality managers allows us to test for differences among Asians, Euro-
peans, and North Americans in the estimation of fault tree probabilities.

Impact on higher-level probabilities

A puzzling finding of Fischhoff, et al. (1978) was the absence of any effect of pruning branches on the
overall probability that a car would not start. It was expected that subjects seeing a pruned tree would
give lower estimates for this global probability than subjects who saw the full tree with its more
numerous listed causes of starting failure. Yet the overall probability of a car not starting showed no
difference between the two groups. The puzzle is confounded by other results that do find an impact of
the availability of lower-level causes on higher-level likelihood estimates (Hirt and Castellan, 1988). We
explore this phenomenon by assessing the impact of both pruning and generating causes on higher-level
probabilities.

METHOD

Stimuli
The fault tree shown in Exhibit 1 contains two first-level branches, decreasing revenues and increasing
costs. Beneath these are five second-level branches, each of which heads a list of 12 third-level causes
and a thirteenth, catch-all category labelled ‘all other’. We focus primarily on these third-level causes
and treat each of the five second-level branches as if they were separate trees.
The fault tree for restaurant failure was developed after consulting practitioners, academic experts,
and publications on restaurant management. The 60 causes in the 5 branches were selected on the basis
of a pre-test in which ten experts rated a preliminary group of nearly 150 third-level causes on the
likelihood that they would play a major role in the occurrence of the second-level ones. The experts
were five professors of foodservice management and five present or former restaurant managers. Only
L. Dubt-Rioux and J. E. Russo Availability Bias 227

causes achieving a mean rating of 4 or higher on a 5-point scale (1, not at all likely, to 5, very likely)
were selected. About 15 to 18 causes in each branch met this criterion.
The same experts were also asked to classify each of the qualified causes as short-term or long-term.
The level of agreement among the experts served to qualify each reason as strongly or weakly short-
term (long-term). If at least 8 of the 10 experts agreed, the reason qualified as strong; agreements of 6
and 7 out of 10 were classified as weak. Finally, within each branch 12 causes were randomly selected
for the full tree within the following distributional constraint: 4 each of strongly short-term and strongly
long-term, and 2 each of weakly short-term and weakly long-term.

Design
In the Full condition the 12 specific causes of a branch were all displayed to subjects. To test for an
underestimation bias a Pruned condition was used in which 6 of the 12 specific causes were omitted.
The absence of any estimation bias would be manifest as the complete transfer of the six omitted
probabilities to the catch-all category. A catch-all probability less than this ideal signifies an under-
estimation bias.
The use of two Pruned conditions enabled a test for differences in bias between short-term and
long-term causes. The Pruned Long-Term condition consisted of 4 strongly short-term and 2 weakly
long-term causes, while Pruned Short-Term listed 4 strongly long-term and 2 weakly short-term causes.
Rather than using all causes of one kind, the mixed lists were designed to reduce the risk of subjects’
detecting the manipulation.
The final two conditions were the Extended Pruned Short- and Long-Term. These two differed from
the corresponding Pruned conditions only in the additional instruction to extend the list of six specific
causes before assigning probabilities.
Thus, there were five conditions in all: Full, two Pruned and two Extended. These formed a Latin
square design in which each subject in a block of five participated in the five conditions once and saw
each of the five branches once. To minimize carry over effects, the two Extended conditions were always
presented last. Except for this constraint the order of presentation was counterbalanced.

Subjects
Active professionals in the hospitality industry were recruited from registrants in a summer program at
Cornell University’s School of Hotel Administration. Eighty managers in four different courses were
asked to participate in a ‘study of why restaurants fail’. All agreed to participate and were randomly
assigned to one of the five condition-by-branch rows of the Latin square design. Eleven subjects were
eliminated because they failed to follow instructions or provided missing data. To achieve the maximum
number of complete replications of the Latin square, 9 more subjects were randomly dropped, leaving
60 subjects in total.
All subjects were currently employed in the hospitality industry: 22 in food-related management, 17
in administrative functions like personnel or accounting, 14 in general management, and 9 in other
functions like marketing and consulting. Overall they averaged more than 8 years of experience in
restaurant operations. Even for those with little experience in restaurant management, the organi-
zational structure of the hospitality industry allows us to assume that every manager had sufficient
exposure to foodservice management to complete the experimental task.
The group was comprised of 21 North Americans, 21 Europeans, 16 Asians and 2 others. All
managers were functional in English, a requirement for admission to the summer program. Their
formal education varied widely, with 26 holding an undergraduate degree in hotel or restaurant
management, 7 in business, and 13 in other fields, mainly liberal arts. Three others held technical
degrees from culinary institutes, and 11 had no formal education beyond high school.
228 Journal of Behavioral Decision Making Vol. 1, Iss. No. 4

Task and procedure

At the beginning of a regularly scheduled class subjects were given an experimental booklet containing
background information, instructions, and stimuli. The task was self-paced and took about 15 minutes.
An introductory statement reminded subjects of the complexity of the management system they were
responsible for (50% of new restaurants fail within one year). They were asked to think systematically
about potential causes of decreasing restaurant profits. They were then shown a completed example of
the task in a related context, low occupancy in a hotel. The example listed a catch-all category beneath
eight specific causes (e.g., poor quality of dining facilities, unattractiveness of the facility, and high
gasoline cost) and provided sample probabilities from a hypothetical respondent that summed to 100%.
For each of the five branches of the fault tree, subjects were asked to provide their best estimate of the
likelihood that each listed third-level cause was the major cause of the second-level one.’ Likelihood
values were expressed as percentages, and subjects were reminded that these values had to add to 100%.
The 10 (out of 300) responses that failed to sum correctly were normalized to 100%. These cases seemed
to represent minor arithmetic errors that never exceeded 10% and never occurred more than once per
subject.
In all conditions, subjects were warned to avoid exclusive reliance on the listed causes, ‘Remember to
include the likelihood of any reason not specifically listed in the “all other” category.’ For the Extended
conditions, the following instructions were added after the list of causes, ‘There may be other important
causes not listed above. Please take two minutes to consider some. Add them below and assign a
percentage to each.’
After completing all five branches, subjects were asked to estimate the percentage of restaurant
failures due to each of the five second-level causes. This task allowed us to test for an effect of pruning
and generating lower-level causes on higher-level probabilities. Recall that though a straightforward
application of availability would predict such an effect, Fischhoff, et al. failed to find it. The booklet
concluded with demographic questions.

RESULTS

The results are organized as follows. We first test for the same underestimation of the catch-all
probability reported by Fischhoff, et al. (1978). We then move to the three novel concerns of our work:
the remedial value of extending the tree, differences between short-term and long-term causes, and
whether the observed effects differ with culture or education. Finally, we consider whether pruning and
extending specific causes affects the estimates of higher-level probabilities.

Calculation of bias
Any underestimation bias was measured as the difference between the ideal and observed probabilities
assigned to the catch-all category in a Pruned condition. The ideal probability was computed from the
Full condition by cumulating the probabilities assigned to the six pruned causes and the catch-all
category. In the Extended conditions, the ideal was compared to an extended catch-all value that
included the probabilities assigned to all subject-generated causes as well as to the ‘all other’ category.
Observed effects were tested for significance by an analysis of variance (ANOVA) with the factors of
condition (treatment), branch, square and subjects within square. Because square differences consisted
only of subject differences, the latter two factors were combined into a single subjects factor. To stabilize
the variance of the data, which were probability estimates bounded by 0 and 1, an arcsin transform was
performed prior to the ANOVA (Neter and Wasserman, 1974). The specific research hypotheses were
tested by planned comparisons, for example, between a Pruned condition and the Full tree.
 L. Dub&-Riouxand J. E. Russo Availability Bias 229

Ideal Observed Extended

Experimental Catch-all Catch-all Catch-all Underestimation
Condition Probability Probability Probability Bias

Full Tree ___ .036 ___ ___

Pruned
Long-Term 334 .074 ___ S10"
Short-Term .497 .065 ___ .432"
Combined .541 .070 ___ .471"

Extended Pruned
Long-Term 3 4 .024 .I86 .39gab
Short-Term .497 .027 .244 .254'b
Com bined .541 .025 .215 .326ab

"Significantly greater than zero (p < .05).

bSignificantlyless than the corresponding Pruned condition above (p < .05).

Exhibit 2. Mean observed probabilities and underestimation bias

Is there an underestimation bias?

The main results are shown in Exhibit 2. The observed probability assigned to the catch-all category in
both Pruned conditions, .074 for Long-Term and .065 for Short-Term, is significantly less than the ideal;
F(1,114) 5 4 7 . 8 , <
~ .001 and F(1, 114) = 8 6 9 . 0 , <
~ .001, respectively. (Neither the branch nor subjects
factor in the ANOVA was significant at the .05 level.) The corresponding underestimations of the
Pruned catch-all probability, .51 and .43, are substantial, respresenting 93% and 94% of the maximum
possible. In other words, subjects showed little sensitivity to the incompleteness of the pruned trees,
adjusting the probability assigned to 'all other' causes only 6% or 7% as much as the normative
benchmark. Furthermore, in only 1 of 120 cases did the Pruned catch-all probability reach the
normative level. This replicates the basic finding of Fischhoff, et al. for our sample of professional
managers.

Does extending the tree reduce the estimation bias?

The results shown in Exhibit 2 confirm the effectiveness of the extension tactic in reducing the under-
estimation bias. Compared to the corresponding Pruned conditions the observed bias is significantly
lower for both Extended Pruned Long-Term, F(1, 114) = 11.26,p < .01, and Extended Pruned Short-
Term, F(1,l 14) = 2 9 . 0 4 , ~< .001. Combining over both Short- and Long-Term conditions, the catch-all
category is three times greater in Extended than Pruned. This corresponds to a residual bias that is 65%
of the maximum, substantially less than the 93%observed under the Pruned condition, but still far from
ideal.
Differences among subjects in the number of causes generated suggest a stricter test of the extension
tactic: Does the remedial value of extending the tree depend directly on the amount of extension?
Exhibit 3 displays the underestimation bias as a function of the number of extended causes. There is a
nearly monotone relation between the number of generated reasons and the size of the bias. Note also
that when a subject listed no additional causes, the mean estimation bias, .52, closely matched the .47
observed for the Pruned conditions.
To test the reliability of this decrease as well as to estimate the asymptote when the number of
extended causes becomes very large, we fit to these data an exponential function of the form
230 Journal of Behavioral Decision Making Vol. I , Iss. No. 4

where Y is the underestimation bias and X the number of generated causes. The overall fit of the model
was highly significant, F(3, 117) = 180.6, p < .001, and accounted for 81.8% of the variance. The
estimate of the asymptote a,which represents the predicted bias when a very large number of causes is
generated, was -. 114 with a 95 percent confidence interval of (-.368, .139). This estimate is conditioned
on the substantial assumptions that the relation between causes and bias is exponential (among many
other functions with a finite asymptote) and can be extrapolated to much larger numbers of causes than
those observed.
The data shown in Exhibit 3 strongly suggest that if enough explicit causes are added to the Pruned
list the underestimation bias may be completely eliminated. However, zero bias relative to the Full
condition is probably not accurate with respect to the natural ecology of restaurant failures. If the Full
condition itself suffers from the unavailability of causes beyond the 12 listed, as seems plausible given
the results of the Pruned treatment, then the true value of the Full catch-all probability is higher than
the .036 measured. This means that the true underestimation bias is greater than our estimates, and the
measured disappearance suggested by the trend in Exhibit 3 probably does not represent a full remedia-
tion of the problem. The preceding argument also implies that the negative estimate of the asymptote
(-. 1 14) may be entirely reasonable. In other words, generating a very large number of additional causes
might correct not only the bias caused by the Full-Pruned difference but also some of that inherent in
using only 12 explicit causes in the Full condition.

60
44
I

GI Number of Observations
40 -

20 -

0-

0 2 4 6 8
Number of Extended Causes

Exhibit 3. Impact of the number of extended causes on the underestimation bias

 L. Dub&-Riouxand J. E. Russo Availability Bias 23 1
Is there a difference between pruning short-term and long-term causes?
We expected short-term causes to be easier to recall than long-term ones. However, before any greater
frequency of short-term causes is attributed to ease of retrieval two potential confounds, importance
and base rate, must be dealt with. Concerning importance, any confound between it and short-term
should have been removed by our selection procedure which assured that all listed causes, whether
short- or long-term, were judged as approximately equal in importance. Although the original ratings
by experts were unavailable, the gist of these data was that there was no apparent difference between the
importance ratings of short- and long-term causes. The second possible confound, namely different base
rates for the two types of causes, was also ruled out by these data. Roughly equal numbers of both kinds
of causes from the initial list qualified as important. It was not the case that more long-term (or
short-term) causes exceeded the criterion for importance. Thus, differences in the occurrence of short-
term and long-term causes should not be attributable to differences in either importance or base rates.
If short-term causes were easier to recall we would expect the two Short-Term conditions (in which
more short-term causes were pruned) to exhibit a smaller underestimation bias than the corresponding
Long-Term treatments. This hypothesis is supported by the direction of the results displayed in Exhibit
2. For the Extended conditions, the Short-Term bias, .25, is substantially less than the Long-Term, .40.
This difference is marginally significant under a one-tailed alternative hypothesis; 61,232) 2.63, or
equivalently t(232) 1.62, p = .051. For the Pruned conditions, however, the effect is only half as large
and not statistically significant.
A second test of the differential recall of short- and long-term causes can be based directly on the
generated causes themselves. Were more of these short-term, or was there an equal number of each?*All
generated causes were classified as short-term or long-term using the same criteria as the ten experts
who originally performed this task. Of the 107 causes, 86 were short-term, a proportion (.80)
significantly above the (approximate) base rate of S O (p < .OOI).’
More telling than the overall frequency of short-term causes is whether the Extended Pruned Short-
Term condition’s greater reduction of the underestimation bias was due specifically to the generation of
more short-term causes. This would establish a direct link between the easier availability of short-term
causes and the effectiveness of extending the tree. In the Extended Pruned Short-Term condition the
proportion of short-term generated causes was .93 (50 short-term out of 54 total), while in the corres-
ponding Long-Term condition this proportion was only .68 (36 of 53). The difference is statistically
reliable; z 4.48, p < .001. Thus, the smaller underestimation bias in the Extended Pruned Short-Term
condition (.25 versus .40 for Long-Term) seems directly attributable to the relative ease of generating
short-term causes. We note that this phenomenon is not attributed to short-term causes being either
more important or more prevalent; they are just easier to retrieve.

Do individual differences explain the underestimation bias?

Three subject characteristics were tested as possible predictors of the observed results. The first, years of
experience, was assessed in four ways: total managerial restaurant experience, total non-managerial
(i.e., operational) restaurant experience, total restaurant experience of any kind, and total managerial
experience of any kind. When added as covariates in the ANOVA, none of these variables significantly
predicted either the underestimation bias or its reduction via the extension tactic. For example, the
mean underestimation bias (from the combined Pruned conditions) was .472 for managers with direct
experience in restaurant management and .470 for those without it. Similarly, none of the four
experience measures reliably predicted the number of extended causes. In fact, the two subjects who
generated the most causes were the two with the least experience.
Education showed similar null results. Subjects were partitioned into five groups: college graduates
with a hotel/ restaurant management degree, business degree or other degree, graduates of a culinary
institute, and graduates of high school only. The mean underestimation bias ranged from a low of .433
232 Journal of Behavioral Decision Making Vol. I , Zss. No. 4

(for culinary institute graduates) to a high of .494 (for business school graduates). These differences with
education were not statistically reliable. There was a similar absence of reliable effects in the reduction
of bias and the number of extended causes.
Finally, we searched for differences among the Asian, European and North American managers. The
mean underestimation bias for these three groups, in order, was .464, .488 and .459. These differences
were not statistically reliable, as was the case for the reduction in bias and the number of extended
causes. However, the absence of significant cultural differences should not be taken as conclusive.
Because the managers in our sample had all been selected to participate in the same summer program at
an American university, they may have been more homogeneous and less representative of their
cultures. We view the question of cultural differences in probabilistic judgment to be important and
largely unaddressed (except for the work of Wright and Phillips, e.g., 1980). We can say only that we
observed no reliable differences in our international sample of foodservice professionals.

Does lower-level availability affect higher-level probabilities?

Fischhoff, et al. (1978) unexpectedly found that subjects who saw a pruned tree did not estimate a
smaller absolute probability of a randomly selected car not starting than subjects who saw the full tree.
That is, manipulating the availability of level-three causes had no effect on the likelihood estimates of
the level-one event. Although a few of Fischhoff, et al.’s other manipulations did yield significant effects
(e.g., calling attention to the tree’s ‘completeness’), the substantial lack of impact of lower-level changes
on a higher-level probability is both noteworthy and puzzling.
Because we did not use a between subjects design in which separate groups saw the Full, Pruned and
Extended trees, we could not search for a cross-level effect on the highest-level probability that a
restaurant would fail. Instead, we obtained the relative probabilities that each of the five second-level
branches was the main cause of restaurant failure. The availability hypothesis would predict that
whenever a second-level cause headed a Full tree, its likelihood would be increased above the base rate
of .20 by the relatively long list of 12 specific third-level causes. Similarly, a second-level cause heading a
Pruned tree should exhibit likelihoods below .20. The mean estimated probabilities were: Full, .196,
Pruned (combined), .201; and Extended (combined), .201. These differences are trivial and not
statistically reliable; chi-square (4) = 1.88, p > 30.
Given the success of the tactic of extending the tree, we also tested whether generating more causes in
the Extended Pruned conditions might produce higher estimates of the probability of the corresponding
branch. An availability hypothesis would predict a monotonically increasing relation between the
number of additional causes generated and the probability estimated for the head of that branch. A
one-way ANOVA yielded no effect; F(7,106) = .31, p > S O . Thus, we find a complete absence of
higher-level effects from either pruning or generating lower-level causes. We address this anomaly in the
discussion as it bears directly on whether availability genuinely plays a role in estimating likelihoods in
fault trees.

DISCUSSION

Although our empirical results seem clear, their interpretation is not so transparent. The absence of any
impact of lower-level availability on higher-level probabilities is puzzling. And even whether the under-
estimation bias can be attributed to availability has been questioned in recent work by Hirt and
Castellan (1988). We attempt to resolve both issues via a consideration of the processes underlying the
estimation of event likelihoods. We begin by considering the estimation process when the events whose
likelihoods are to be estimated have already been identified. The process of generating the specific
events, which of course must precede likelihood estimation, is then considered.
L. DubO-Rioux and J. E. Russo Availability Bias 233

Strategies for likelihood estimation

There are two general approaches to estimating the probabilities of prespecified causes, the availability
heuristic and inference from related knowledge. The availability heuristic means basing the probability
estimate on the ease or frequency of retrieval of instances of the specified event. In the present context
the likelihood of a second-level cause, like increased labor costs, could be estimated either from the
recall of known instances or from the specific third-level causes (of increased labor costs) that had been
presented earlier. Assuming that the latter basis is used by at least some subjects, the availability
heuristic should produce an effect of lower-level manipulations on higher-level probabilities - an effect
that was not observed. Alternatively, the likelihood of restaurant failure due to increasing labor costs
might be inferredfrom related knowledge, such as low wage inflation or an influx of new workers that
has depressed wage increases. We distinguish such causal inference from the cumulating of instances or
lower-level probabilities that is the essence of the availability heuristic. An inference strategy would not
require a lower-level impact on higher-level likelihood estimates, and thus is compatible with the
observed absence of such effect^.^
Although we know of no data that can conclusively discriminate between these two processes for
estimating event likelihoods, discussions with our foodservice experts and informal debriefing of
subjects yielded many examples of inferring higher-level probabilities from relevant knowledge and few,
if any, clear instances of an availability strategy based on lower-level causes. Thus, we find it plausible
that the absence of an impact of lower-level manipulations on higher-level probability estimates in
Fischhoff, et a1.k and in our own work might be accounted for by the use of an inferential strategy.
A recent study by Hirt and Castellan (1988) reports a significant effect of a lower-level manipulation
on higher-level probabilities. They used Fischhoff, et a1.k automobile fault tree with student subjects,
half of whom were instructed (prior to seeing any fault tree) to generate ‘as many problems as they
could think of that could delay a car’s starting for at least one minute’ (p. 123). This generation
manipulation had systematic effects on higher-level probabilities. The estimated probability of a car not
starting was increased from a median of .050 for non-generation subjects to .089 for those who initially
generated their own list of possible causes. More closely related to our data, the relative probability of a
second-level cause was weakly but positively related to the number of third-level causes generated in the
corresponding branch.
How might we account for the total absence of cross-level effects in our experiment and their
significant presence in Hirt and Castellan’s work? Although any attempt to reconcile these different
results is necessarily post hoc, we note the use of different subjects, professionals versus students.
Whereas our professional subjects probably had sufficient knowledge to support an inferential strategy,
Hirt and Castellan’s student subjects may not have. As a consequence, they may have had to rely on the
availability heuristic, which itself is influenced by pruning and extending a lower-level of the tree.
Student subjects may even have felt somewhat obligated to connect the earlier generation task to the
later estimation of second-level probabilities in a way that professionals would not. Thus, the different
results might plausibly be explained by the differential use of availability versus inference strategies in
the two subject populations. We reiterate the post hoc status of such an account. It may be safest to
conclude only that the discrepant results are not irreconcilable, and that the actual explanation awaits
subsequent experimentation.

Biases in likelihood estimation

Having discussed the process of probability estimation, let us turn to possible sources of bias. We
consider three. The first was demonstrated by Mehle, et al. (198 I) who had subjects assess the likelihood
of possible college majors for a student identified only by a single currently taken course. While one
group used an experimenter-provided list of possible majors, a second group generated their own.
Although both groups overestimated the probabilities of the specific majors, the error was nearly three
 234 Journal of Behavioral Decision Making Vol. 1, Iss. No. 4

times greater for the subject-generated hypotheses. This suggests that in extending fault trees, the
likelihood of self-generated causes may be overestimated relative to externally provided causes. This
may partly compensate for retrieval failure and thereby reduce the underestimation bias - or even
reverse it if enough causes are generated. Conclusive estimates of such overestimation (and any bias
reduction) require veridical probabilities. Such values could be obtained by Mehle, et al. in a closed
system like course enrollments, but not by us for a less well structured domain like restaurant failure.
A second source of systematic error in estimating the probabilities of prespecified events follows from
a new explanation for the underestimation bias, category redefinition. Hirt and Castellan (1988) claim
that ‘all pruned causes may be fully available when probabilities are estimated, but because of uncert-
ainty about which category each one belongs in, not all are assigned to the catch-all category as
normatively predicted’ (p. 127). That is, the problem may not be the unavailability of unlisted causes,
but rather confusion over whether they belong in the catch-all or some other category.
Both availability and category redefinition can fully account for the failure to completely transfer the
probabilities of pruned causes to the catch-all category. In addition, some thought will reveal that the
redefinition hypothesis can also account for the bias-reducing effect of generating additional causes. We
believe, however, that the observed differences in the generation of short-term and long-term causes
support the availability hypothesis, and cannot easily be accounted for by category redefinition. Short-
term causes are easier to retrieve, which is the foundation of an availability explanation; yet they do not
seem to be either more or less susceptible to category redefinition, which is necessary if the latter
hypothesis is to apply. Thus, we believe that the short- versus long-term results support an availability
phenomenon and disconfirm the redefinition hypothesis. Nonetheless, category redefinition may still
explain some of the observed underestimation bias. Only the use in future research of a fault tree with
unambiguous category boundaries5 will remove the possibility of category redefinition.
One last possible contributor to the underestimation bias deserves mention. Subjects may presume
that the catch-allprobability is always low in any good fault tree. It doesn’t make much sense for fault
tree builders to leave causes in the ‘all other’ category whose probability equals that of some listed
causes. If subjects presume a low catch-all probability, the likelihood of pruned causes might not be
fully transferred to the catch-all category creating, at least in part, the underestimation bias.
In sum, we have speculated about strategies and biases in estimating likelihoods and how their
differential use might account for the underestimation bias as well as the conflicting results observed in
our work, Hirt and Castellan’s (1988) and Fischhoff, et al.’s (1978). Post hoc it is impossible to identify
the processes actually used. Only future work can choose among the competing explanations for the
observed effects.

Bias in the generation of events

We turn now to the second basic process, the retrieval (or generation) of the specific causes whose
likelihoods are to be estimated. Because identifying possible causes necessarily precedes their
estimation, the former may be more likely to limit the usefulness of a fault tree. If an unspecified cause is
implicitly assigned a probability of zero, the resulting error may well exceed any bias in the likelihood
estimation process.
Hoch (1984) surveys retrieval difficulties in the generation of candidate causes. One of those
difficulties, the interference of initially retrieved items with subsequent recall is relevant to our work.
The phenomenon of part-list interference, sometimes known as the ‘part-set cuing effect’, is robust and
easily demonstrated (Nickerson, 1984; Peynircioglu, 1987). In the prototypical procedure, all subjects
study a list of items to be recalled later. One group is then shown part of the list and asked to recall the
remainder. A control group is given no cues and instructed to recall all items on the original list. The
part-list cued group recalls fewer of the ‘remainder’ items than does the uncued group. Various inter-
L. Dubk-Rioux and J. E. Russo Availability Bias 235

ference processes have been hypothesized to account for this phenomenon. For our purpose it is
sufficient to note that such interference might inhibit the retrieval of causes and thereby provide one
specific foundation for an availability explanation of the observed underestimation bias. Part-list
interference, however, is only one retrieval phenomenon relevant to cause generation. A thorough
survey, though highly valuable, is beyond our scope.
What might be done to reduce the biases created by retrieval problems? Hoch (1984) has shown that
inserting a delay disables the interference from causes already generated. This suggests that a ‘spaced’
strategy for cause generation may prove superior to a single ‘massed’ effort, even holding constant the
total time on task. Another different but potentially useful technique involves shifting from a forward to
a backward perspective when anticipating future causes of trouble. This shift is accomplished by
phrasing a future event in the past, e.g., by assuming that it is two years in the future and a restaurant
has, in fact, failed. Mitchell, Russo and Pennington (1988) show that this shift of temporal perspective
generates more causes of an event (though not necessarily causes of higher quality which was not
assessed in their work).
In general, many of the techniques for aiding the retrieval of natural categories may prove useful for
improving how we extend fault trees. Similarly, known biases in likelihood estimation might inform the
fault tree task, and point the way toward improving its performance. Thus, we envision a more positive
prospect for studies of likelihood estimation than just the exposure of human shortcomings. What we
hope will eventually emerge is a program for aiding the performance of professionals in the use of real
world fault trees.

NOTES

I. This instruction was designed to avoid the problem of multiple causes that are neither necessary nor sufficient
(Downing, Sternberg and Ross, 1985; Einhorn and Hogarth, 1986). Instructions permitting subjects to interpret the
likelihood of a given cause as the portion of causality attributed to that cause are greatly complicated by non-
mutually exclusive causes. For instance, there is no guidance on how joint causality should be partitioned between
two overlapping causes. Estimating the likelihood that a given cause is the single major cause avoids these
problems. Note that the automobile fault tree of Fischhoff, et al. (1978) does not encounter such difficulties because
the specific causes listed there are mutually exclusive and each is sufficient by itself.
2. In answering this question, we acknowledge that not all third-level causes are mutually exclusive. For instance,
changing customer demographics appears in both of the first two branches, and poor quality of service (in branch
I) is related to lack of employee motivation (in branch 2). Since the two Extended conditions were always presented
last, some of the generated causes may have been primed by a preceding listing. Because such priming does not
reflect naturally occurring retrieval processes, the causes generated by subjects were purged of any that were
identical or very similar to those on an earlier Full or Pruned list. This removed 47 of the 154 total causes generated
in the two Extended conditions. Although the analyses that follow are based on the 107 purged data only, they were
also performed on the complete set of extended causes. The results with all causes included were diluted, but not
otherwise changed; all significant differences were significant with either data set.
3. Because the exact proportion of causes classified by experts as short-term was unavailable, we also computed the
highest base rate such that an observed proportion of .80 (in a sample of 107) would still fall reliably above it
@<.05). That value, .74, is comfortably above the approximate base rate of SO.
4. There is at least one other strategy, the recall of a previously computed estimate of the required probability. We
presume that a recalled estimate could have originally been derived either from the availability heuristic or an
inference process. Thus, though such direct retrieval is possible, it does not expand the two main approaches and
will not be considered separately.
5. Whereas laboratory studies of availability can avoid unambiguous categories, professionally realistic categories
may often entail some ambiguity. Even here, however, we suspect that confusion about category definitions
decreases with professional experience. Thus, although category redefinition may routinely occur and legitimately
account for at least some underestimation bias, we suspect that it may be less prevalent for professionals.
236 Journal of Behavioral Decision Making Vol. 1, Iss. No. 4

ACKNOWLEDGEMENTS

T h e authors thank Leo Renaghan for this assistance in planning and conducting the experiment and
Stephen J. Hoch, Deborah J. Mitchell, Bernd H.Schmitt and Paul J. H. Schoemaker for comments on
an earlier version of this paper.

REFERENCES

Billings, R. S. and Schaalman, M. L. ‘Administrators’ estimation of the probability of outcomes of school

desegregation: A field test of the availability heuristic’, Organizational Behavior and Human Performance, 26,
1980,97-114.
Carroll, J. S. ‘The effect of imagining an event on expectations for the event: An interpretation in terms of the
availability heuristic’, Journal of Experimental Social Psychology, 14, 1978, 88-96.
Downing, C. J., Sternberg, R. J. and Ross, B. H. ‘Multicausal inference: evaluation of evidence in causally complex
situations’, Journal of Experimental Psychology: General, 1 14, 1985, 230-263.
Einhorn, H. J. and Hogarth, R. M. ‘Judging probable cause’, Psychological Bulletin, 99, 1986, 3-19.
Fischhoff, B. ‘Debiasing’, in D. Kahneman, P. Slovic and A. Tversky (Eds.), Judgement Under Uncertainty:
Heuristics and Biases. Cambridge, England: Cambridge University Press, 1982.
Fischhoff, B., Slovic, P. and Lichtenstein, S. ‘Fault trees: Sensitivity of estimated failure probabilities to problem
representations’, Journal of Experimental Psychology: Human Perception and Performance, 4, 1978,330-344.
Folkes, V. S. ‘The availability heuristic and perceived risk’, Journal of Consumer Research, 15, 1988, 13-23.
Gregory, W. L., Cialdini, R. B. and Carpenter, K. M. ‘Self-relevant scenarios as mediators of likelihood estimates
and compliance: Does imagining make it so?’, Journal of Personality and Social Psychology, 43, 1982,89-99.
Hirt, E. R. and Castellan, N. J., Jr. ‘Probability and category redefinition in the fault tree paradigm’, Journal of
Experimental Psychology: Human Perception and Performance, 14, 1988, 122-131.
Hoch, S. J. ‘Availability and inference in predictive judgment’, Journal of Experimental Psychology: Learning,
Memory, and Cognition, 10, 1984,649-662.
Koriat, A., Lichtenstein, S. and Fischhoff, B. ‘Reasons for confidence’, Journal of Experimental Psychology:
Human Learning and Memory, 6, 1980, 107-1 18.
Lord, C. G., Lepper, M. R. and Preston, E. ‘Considering the opposite: A corrective strategy for social judgment’,
Journal of Personality and Social Psychology, 47, 1984, 1231-1243.
Mehle, T., Gettys, C. F., Manning, C., Baca, S. and Fisher, S. ‘The availability explanation of excessive plausibility
assessment’, Acta Psychologica, 49, 1981, 127-140.
Mitchell, D. J., Russo, J. E. and Pennington, N. ‘Back to the future: Temporal perspective in the explanation of
events’, Journal of Behavioral Decision Making, forthcoming, 1988.
Neter, J. and Wasserman, W. Applied Linear Statistical Models, Homewood, Illinois: Richard D. Irwin, 1974.
Nickerson, R. S. ‘Retrieval inhibition from part-set cuing: A persisting enigma in memory research’, Memory and
Cognition, 12, 1984, 53 1-552.
Peynircioglu, Z. F. ‘On the generality of the part-set cuing effect: Evidence from nonmemory tasks’, Journal of
Experimental Psychology: Learning, Memory and Cognition, 13, 1987,43742.
Phillips, L. D. and Wright, G. N. ‘Cultural differences in viewing uncertainty and assessing probabilities’, in H.
Jungermann and G. de Zeeuw (Eds.), Decision Making and Chance in Human Affairs, Amsterdam: D. Reidel,
1977.
Shedler, J. and Manis, M. ‘Can the availability heuristic explain vividness effects?’, Journal of Personality and
Social Psychology, 5 I, 1986, 26-36.
Sherman, S. J., Skov, R. B., Hervitz, E. F. and Stock, C. B. ‘The effects of explaining hypothetical future events:
From possibility to probability to actuality and beyond’, Journal of Experimental Social Psychology, 17, 1981,
142-158.
Taylor, S. E. ‘The availability bias in social perception and interaction’, in D. Kahneman and A. Tversky
(Eds.), Judgment Under Uncertainty: Heuristics and Biases, Cambridge, England: Cambridge University Press,
1982.
Tversky, A. and Kahneman, D. ‘Availability: A heuristic for judging frequency and probability’, Cognitive
Psychology, 5 , 1973,207-232.
 L. Dubk-Rioux and J. E. Russo Availability Bias 237

Von Winterfeldt, D. and Edwards, W. Decision Analysis and Behavioral Research, Cambridge, England:
Cambridge University Press, 1986.
Wright, G. N. and Phillips, L. D. ‘Cultural variations in probabilistic thinking: Alternative ways of dealing with
uncertainty’, International Journal of Psychology, 15, 1980, 239-257.
Wright, G. N., Phillips, L. D., Whalley, P. C., Choo, G. T. G., Ng, K. O., Tan, I. and Wishuda, A. ‘Cultural
differences in probabilistic thinking’, Journal of Crosscultural Psychology, 9, 1978, 285-299.

Authors ’ biographies:
Laurette DubC-Rioux teaches foodservice management at the University of Montreal. She holds an MBA from
L’Ecole des Hautes Etudes Commerciales and is currently completing doctoral studies at Cornell University. She
has also worked at different levels of management in foodservice operations, and is the co-author of a textbook on
cost and quality control in foodservices.

J. Edward Russo is Associate Professor of Marketing and Behavioral Science at the Johnson Graduate School of
Management of Cornell University. He earned a Ph.D. in psychology from the University of Michigan in 1971. His
current research focus is decision aiding, especially that of consumers.