Journal of Economic Behavior & Organization

Vol. 35 (1998) 309±332

Strategies of search
Joep Sonnemans*
University of Amsterdam, Faculty of Economics and Econometrics, Department of General Economics/CREED,
Roetersstr. 11, 1018 WB Amsterdam, The Netherlands

Received 1 June 1996; received in revised form 7 September 1997; accepted 11 September 1997

Abstract

Two experiments are designed to examine the strategies people use in search behavior. In the first
experiment an electronic information board is used to register on which aspects of the situation
subjects focus their attention and after that subjects also submit a formal strategy. Although
efficiency is rather high, most subjects do not use the theoretical optimal strategy. Many subjects
seem to focus on the total earnings instead of the marginal return of another draw. On an average,
subjects stop searching too early. This cannot totally be explained by risk aversion. The second
experiment shows that the tendency to search too little can be (partly) explained by learning
processes. # 1998 Elsevier Science B.V.

JEL classi®cation: C91; D83

Keywords: Search; Information board; Strategy; Experiment

1. Introduction

Search problems occur frequently in economic everyday life: consumers search for the
lowest price and workers search for the highest wage. Theorists have derived optimal
search strategies for many situations (e.g. Kohn and Shavell, 1974; McMillan and
Rothschild, 1994). This study focuses on a relatively simple search situation: Offers are
independently drawn from a known distribution, the cost of one search action is constant,
the number of searches is unlimited, and recall (accepting a previously rejected offer) is
possible. The optimal strategy for a risk-neutral searcher is to continue searching as long
as the expected marginal benefit of another search exceeds its expected marginal cost. In

* Corresponding author. Tel.: +31 205254249; fax: +31 205255283; e-mail: [email protected]

0167-2681/98/$19.00 # 1998 Elsevier Science B.V. All rights reserved

PII S 0 1 6 7 - 2 6 8 1 ( 9 8 ) 0 0 0 5 1 - 1
310 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332

this case, this means that search continues until a draw is greater than the `reservation
price,' and that recall will never happen.
Whether individuals search as theoretical models predict is hard to examine in the
field. Therefore many experimental studies have been carried out (for a short review, see
Camerer, 1995). Generally, experimental studies have shown that subjects search
relatively well in the sense that their earnings are close to the optimal earnings (e.g.
Harrison and Morgan, 1990), but there is a tendency to search too little (Schotter and
Braunstein, 1981; Cox and Oaxaca, 1989, 1992; Kogut, 1990; Hey, 1987) and subjects
often recall (Schotter and Braunstein, 1981; Kogut, 1990). These results suggest that
subjects use search strategies which are highly successful, but are in some respects quite
different from the optimal strategy. Indeed, simulations by Moon and Martin (1990) show
that heuristic rules can perform very well. An interesting question is which strategies
individuals actually use. It is possible that some boundedly rational subjects reason
correctly but miscalculate; they will use a reservation price which is not necessarily
optimal. Other boundedly rational subjects may try to satisfy their aspirations in terms of
earnings; these subjects do not have a reservation price but stop searching if the resulting
earnings exceed their current aspiration level.
The main goal of this study is to answer the question: what are the strategies of search
individuals use, and how do these strategies relate to optimal strategies? Section 2
discusses several methods to study strategies of search, including electronic information
boards and the strategy method. Section 3 explains the search problem used in the
experiments and Section 4 describes the design and the results of the first experiment. An
electronic information board is used to study the process of search behavior and after
gaining experience all subjects submit a strategy. Most subjects did not use reservation
prices but used combined strategies which focus on earnings (like bounded rational
satisficers would do) and on the last (or highest) bid (like optimizers would). Like in
previous studies, efficiency is high, but there is a tendency to search too little.
If searchers are only boundedly rational their performance may be improved by
learning. Learning behavior has become a major topic in (experimental) economics over
the last few years (e.g. Selten, 1991; Milgrom and Roberts, 1991; Miller and Andreoni,
1991; Roth and Erev, 1995; Sonnemans et al., forthcoming). A common assumption of
adaptive learning models is that more successful behavior is more likely to be repeated.
The second experiment, described in Section 5, focuses on learning behavior. The
research question here is: how subjects learn and whether this learning can explain their
tendency to search to little. Finally, in Section 6 some concluding remarks are made.

2. How to study actual strategies of search

Several methods can be used to study the actual strategies of search. Hey (1982), Moon
and Martin (1990) use a thinking aloud protocol to reveal strategies. In these experiments
subjects are instructed to think aloud while deciding, and are audio-taped. In cognitive
psychology thinking aloud protocols are often used to study complex problem solving.
Although such a design can provide valuable insights in the way subjects reason, it has
the following potential pitfalls: behavior may be influenced by the instruction to think
 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332 311

aloud and by the knowledge that the experimenter will listen to the reasoning.
Furthermore, there are some practical disadvantages: you cannot run a lot of subjects at
the same time, and to analyze the audio tapes is a lot of work.
Another possible method is to confront the decisions of the subjects with different rules
of thumb, and see which rules describe the decisions best (Hey, 1987; Moon and Martin,
1990 use this method in addition to the thinking aloud protocol). The disadvantages of
this method are, first, that you need to have disposal of a complete set of possible rules of
thumb or you have to make additional assumptions; second, you only get indirect
evidence; third, many observations may be needed to determine which rule fits best; and
fourth, because strategies are likely to differ between subjects you need experiments that
last many periods.
The use of questionnaires can be problematic because subjects have no incentives to
report their strategy reliably, especially if subjects find it hard to describe their
(complicated) strategies.
Although these methods all have their limitations, they provide us with clues about
what kind of strategies subjects are likely to use. To get a more precise view, in the
present study two other methods to reveal search strategies are employed: an electronic
information board and a strategy method.
Electronic information board: Research in multi-attribute choice has often used
information boards. Traditionally, this is a cardboard with a matrix of envelopes. The
decision maker has to pull an envelope to obtain information about certain attributes of
the alternatives. In this way the information processing of the individual is monitored by
the experimenter. More recently, so-called electronic information boards are used. These
are computer programs which present hidden information on the computer screen. To
reveal the information the subject has to click with the mouse on certain regions of the
screen. The program keeps track of all the actions of the subject. For detailed information
about information boards, see Payne et al. (1993). Until now electronic information
boards have been used almost exclusively in multi-attribute choice research, but there are
good reasons to apply this method to search problems. In the experiments presented in
this paper an electronic information board presents information to subjects about the
number of bids, the last bid, the highest bid so far, the costs of the bid so far and the
earnings if (s)he stops at that point. The program registers on what information the
subject focuses his or her attention. An electronic information board enables the
researcher to study directly the process of decision making. This method may give
important indications about the kind of strategy the subject uses.
Strategy method: In experimental research on games a `strategy method' is sometimes
employed (e.g. Axelrod, 1984; Keser, 1992). In these experiments subjects have to
formulate strategies and these strategies interact in a tournament. Sometimes a series of
tournaments is organized, in which subjects receive information about outcomes after
each tournament and can adapt their strategies. Although an individual search problem
differs in many ways from games, the important conclusion here is that these studies
show that it is possible for subjects to formulate rather complicated strategies, especially
if they have previously experienced the problem. If subjects are paid according to the
performance of the strategy they submit, they have an incentive to report honestly what
they think is the best strategy. This is an important advantage of the strategy method
312 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332

above questionnaires.1 In the first experiment of the present study, subjects first play
many periods for a search problem and then submit a strategy according to which they
play another 20 periods.

3. The search problem

Subjects in the experiments are confronted with the following search problem (see
Section 4.2 and the Appendix A for the precise instructions to the subjects). They receive
an `article' which they have to sell to the computer. The computer starts with an initial
bid. The subject can accept this bid, or ask for another bid. The number of bids is
unlimited and at any time the subject can decide to sell for the highest bid so far. Each bid
costs 2 cents and is drawn from a discrete uniform distribution [1;100]. All bids are
stochastically independent. There is a fixed cost per period (search problem) of 50 cents.2
A risk-neutral decision maker will stop searching if the marginal cost of an additional
search action (ˆ2 cents) exceeds its expected marginal benefit:
X
100
0:01 j ÿ H† ˆ 0:005 100 ÿ H† 101 ÿ H†; with H the highest bid so far:
jˆH‡1

This means that search will continue until a bid greater than 80 cents is drawn. The
expected earnings are 30.5 cents with a standard deviation of 10.6 cents.
It is reasonable to expect that in an experiment at least some subjects are not capable of
calculating the optimal solution and can be considered `boundedly rational.' It is possible
that some of these subjects use a reservation price, not necessarily of 81 cents (e.g. they
reason correctly but miscalculate). Reservation price strategies (`Stop if last bid H')
will be labelled as `strategies of the optimal form.' Another form of boundedly rational
behavior is that someone does not maximize expected earnings, but tries to satisfy his or
her aspirations (in terms of earnings).3 A `satisficer' stops searching when the `earnings if
you stop now' exceed the aspiration level. Such an aspiration level does not necessarily
stay constant during the course of a period, but can be adapted.4 This makes it hard to
distinguish satisficers on the basis of their decisions alone, but it should be possible to
recognize satisficers by their focus on the earnings.

1
Of course, strategies are, not only interesting as descriptions of behavior, but also because in every day life
search behavior is often delegated to an agent. The instructions to the agent can take the form of a search
strategy.
2
These fixed costs (not common in other search experiments) are introduced to make the incentives for using a
good strategy more salient. For example, a strategy earning on an average 70 cents in an experiment without
fixed costs (with expected earnings of the optimal strategy of 80.5 cents) has an efficiency of 87 percent (70/
80.5), but will have in an experiment with fixed costs an efficiency of only 66 percent (20/30.5).
3
Kogut (1990) finds indications that the stopping decisions of at least some subjects are influenced by total
returns (and not only by marginal returns).
4
After some bids a target earning may become very unlikely or even impossible which may cause the subject
to lower the target.
 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332 313

4. Experiment 1

4.1. Pilot study

Before discussing the design of Experiment 1 a pilot study is briefly described. The
goals of this pilot were to test the electronic information board part of the experiment and
to elicit information about the kind of strategies subjects are likely to use (and therefore
should be possible in the strategy part of the experiment). Subjects are 5 groups of high
school students (in total 71 subjects) who visited the department of economics as part of
an educational project for potential future students. Subjects played 30 periods with the
information board and then filled in a questionnaire about the strategy he/she had used.
Not all the answers to these questions are comprehensible, which illustrates the limited
use of a questionnaire like this. Nevertheless, it is possible to get a general impression
about the strategies used. The largest group (50%) seems to focus on the highest or last
bid. However, a part of these subjects indicates that they use a combined strategy of the
form `stop if the highest bid is higher than X or the number of bids is higher than Y.' The
next largest group (approximately 25%) are satisficers who stop if their earnings are good
enough. Surprisingly there are also subjects who find the order of bids important and who
use rules like `stop if the last bid is X times in a row smaller than the highest one'
(sometimes combined with other rules). Hey (1982) finds evidence of these `bouncing
rules' in an experiment in which the subjects do not know the underlying distribution of
the bids, but only very few of his subjects seem to use these rules in another experiment
with a known distribution of bids (Hey, 1987, 1991). The results of this pilot study are
used in designing the strategy part of the first experiment.

4.2. Design Experiment 1

A total of 36 subjects participated in four sessions. Twenty-one of the 36 participants
were students of economics and 14 were known to be students of other departments; 9
were female, 26 were male and the sex of one of the subjects is unreported. The average
age was 21.75 years. Including the show-up fee (10 guilders), an average of 27 guilders
($ 16.40) was earned by subjects in about 1.5 h.
Subjects are seated behind a computer screen and cannot see the screens of the other
participants. The computer program5 is stand-alone, so that the subject has no cues about
the search behavior of the other subjects (there is no need to wait between periods). After
the instructions the subject answers some questions (see Appendix A) to check whether
the instructions are fully understood (especially the stochastic independence of the bids
and the possibility of recall). The subject starts with the practice periods only after these
questions are answered correctly.
In each period the subject is endowed with an `article' which (s)he has to sell to the
computer. The period starts with an initial bid. The subject can accept this bid, or ask for
another bid. After a bid the subject has 10 seconds to decide whether (s)he wants another
bid (a clock on the screen ticks away from 10±0 seconds). The number of bids is
unlimited and at any moment the subject can decide to sell the article for the highest bid

5
The program is developed in VISPRO/REXX for an OS/2 environment and is available from the author.
314 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332

the computer has made so far in that period. Each bid costs 2 cents and is drawn
independently from a discrete uniform distribution [1;100]. The earnings in a period
equal the highest bid at the moment of stopping, minus the fixed costs (50 cents) and the
costs of all bids in that period.
At any moment the subject can get information about the number of bids so far in that
period, about the last bid received, about the highest bid so far, about the costs of the bids
so far and about what the earnings would be if (s)he would stop at that moment. In the
four practice periods the subject cannot earn any money, and the information is always
presented on the screen. In part 1 (15 periods) the subject has unlimited access to this
information, but has to click with the mouse or use the keyboard to reveal information. In
part 2 (15 periods) the subject can reveal only one piece of information after each bid. In
part 3 (15 periods) a period starts in the situation that a number of bids is already made.
Before the start of the period the computer has made 0, 1, 2, 3 or 4 bids (all these
situations are equally likely). The costs of the bids still equal the number of bids times 2
cents; so that the bids the computer has made before the start of the period also cost
money. After each bid the subject can reveal only one piece of information.
After part 3 the subjects fill out a questionnaire about their strategies and after that they
receive instructions for part 4 on paper (refer to Appendix A). They are then asked to
formulate a strategy. The computer plays that strategy for 20 periods and the results are
paid to the subject. The strategy is a stopping criterion which consists of one, or a
combination of more, stopping conditions. The stopping conditions are derived from the
answers to the questionnaire of the pilot experiment 6 and have the form:

I want to stop:

A. If the last bid is greater than or equal to: ......

B. If the highest bid is greater than or equal to: ......
C. If the number of bids is greater than or equal to: ......
D. If my earnings are greater than or equal to: ......
E. If ...... bids in a row are smaller than the highest bid so far.

The subject fills in numbers for the stopping condition(s) he or she wants to use, and
formulates the stopping criterion `X' if only stopping condition X is to be used, or a
logical combination of stopping conditions (e.g. `X or Y', or `X and Y', or `(X or Y) and
Z).' After the subject has formulated a strategy, the experimenter reformulates the
stopping condition in words to check whether the strategy is really what the subject wants
it to be (e.g. `So you want to stop if either the last bid is greater or equal to 85, or the
number of bids is greater or equal to 6, is that right?'). The experimenter also warns
subjects who formulate a strategy which may never stop (e.g. the strategy `stop if my
earnings are greater or equal to 40' stops after one of the first 5 bids or never).7 If the

6
All strategies formulated by the subjects of the pilot experiment can be expressed as a combination of one of
more of these stopping conditions.
7
In two cases (participants 10 and 11) the experimenter forgot to warn the subject. The strategies of these
subjects are excluded from the analysis (refer to Section 4.3).
 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332 315

subject confirms the strategy, the experimenter types the strategy into the computer, and
20 periods are played with this strategy. The subject sees the sequence of bids, after
which bid the strategy stops searching and the resulting earnings, but is not allowed to
change the strategy. In a concluding questionnaire the questions are asked: ``To what
extent is your strategy in part 4 similar to your behavior in the first 2 parts of the
experiment? (7-point scale very different ± exactly the same)'' and the open question
``Can you indicate the difference between your behavior in parts 1 and 2 and your
strategy in part 4?'' The questionnaire also contains a few questions about personal
characteristics like age and gender.
All bids are drawn beforehand and in each period all the subjects receive the same bids
in the same order. This is done to assure that possible individual differences are not due to
different experiences in the experiment.

4.3. Results
Search behavior: Table 1 shows that subjects stopped optimally in more than 60
percent of the cases. There is a tendency to search too little. Although efficiency levels
are fairly high, 32 out of the 36 subjects earned significantly (p<0.05) less than the
optimal searcher (a 2-sided Wilcoxon rank test was calculated for each individual with
the earnings in one period as one observation).
Subjects who use a reservation price never recall (never accept a previous bid).
However, in 13.9 percent (part 1) and 16.7 percent (part 2) of the periods subjects
recalled. There are large differences between subjects: 5 subjects never recalled, 14
almost never (0%, 10%), 15 sometimes recalled (10%, 50%), 2 subjects recalled most of
the time (50%, 100%), and no subjects recalled always. This suggests that there are
individual differences in the strategies subjects use.
Information: The kind of search strategies subjects use is indicated by the information
they reveal. Subjects who employ the optimal strategy (stop if the last bid is 81 or higher)
or a strategy of the same form (stop if the last bid is X or higher) will focus on the `last
bid' or `highest bid so far.' Subjects who are not maximizers but are satisficers will focus
on the `earnings if you stop now.' Table 2 shows that almost 70 percent of the subjects
focus on the information an optimal strategist would focus on, and 30 percent focus on
the earnings and therefore may be satisficers.
The fact that someone always reveals the `highest bid so far' in the first two parts, does
not necessarily mean that he or she uses a strategy of the optimal form. For example, it is
possible that someone counts and remembers the number of bids and/or calculates the
`earnings if you stop now.' In part 3 a subject using a strategy of the optimal form should

Table 1
Search behavior in experiment 1, 36 subjects, 15 periods per part. The earnings are in cent per period

Mean number of bids Stopping Mean earnings

Optimal Actual Early Optimal Late Optimal Actual Efficiency

Part 1 4.73 4.07 40% 54% 6% 29.4 24.7 84%
Part 2 5.53 4.62 17% 74% 9% 27.3 25.8 95%
Part 3 5.67 4.55 30% 61% 9% 29.9 26.8 90%
316 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332

Table 2
The first four rows show the percentages of subjects who revealed that information most often. Total number of
subjects is 36. The last row shows the percentages of subjects who revealed that information first at the start of
periods in part 3 most often. Nobody focused on the information about the costs of the bids

Information

Number of bids Last bid Highest bid so far Earnings

Part 1 0 61.1% 16.7% 22.2%
Part 2 0 58.3% 11.1% 30.6%
Part 3 2.8% 27.8% 30.6% 38.9%
Total 0.9% 49.1% 19.4% 30.6%
Information revealed first in part 3 0 17.1% 31.4% 51.4%

always first look at the `highest bid so far,' while a satisficer would first look at the
`earnings if you stop now.' The last row of Table 2 shows that more than 50 percent
behave like satisficers. These are for the most part the same subjects who focus on the
earnings in parts 1 and 2.
Strategies: All strategies had to be of the format mentioned in Section 5.1. Although all
strategies described by the subjects of Experiment 1 fit into this format, it is possible that
someone wants to use another kind of strategy. In the instructions in part 4 the subjects
are asked to ask the experimenter if they think that their strategy cannot be described
within the presented format. Only 2 of the 36 subjects wanted to submit a strategy not
fitting the format.8 This suggests that the enforced format did not restrict the subjects very
much.
Four subjects indicated in the questionnaire that the strategy submitted in part 4 was
fundamentally different from their behavior in the first two parts. Two of these subjects
formulated a strategy which not always stopped and were mistakenly not warned for this
by the experimenter. The two others used an `and' statement where they meant an `or'
statement. These 4 strategies are not included in further analysis. Some subjects indicated
that they used essentially the same strategy in parts 1 and 2 of the experiment, but that in
the first parts they sometimes deviated from these rules of thumb and stopped earlier or
later based upon a `feeling,' or they sometimes had the `temptation' to search longer. For
these subjects the strategy seems to work as a commitment to act in the way they think is
best.
Table 3 displays the 32 remaining strategies. None of the strategies is exactly the
optimal strategy, but the strategies of subjects 6, 8, and 24 come very close. Seven
strategies (22%) are of the optimal form,9 the remaining strategies almost always admit
the earnings and/or the number of bids in the stopping criterion. The strategies and the
actual behavior of the subjects can be compared by applying the strategies upon the same
sequence of bids as the subjects experienced in the first 2 parts of the experiment (see

8
These two subjects wanted to use complicated strategies of the form `stop if the earnings (casu quo. highest
bid) so far is greater than X(t)', in which X(t) depended upon the number of bids so far. The experimenter told
these subjects to formulate another strategy within the restrictions.
9
The strategy of subject 34 is not formulated as a strategy of the optimal form (and not scored as such), but is
in practice of the optimal form because if the number of bids is greater or equal to 5 and the earnings are greater
or equal to 10 there has been at least one bid greater than 70.
 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332 317

Table 3
The strategies of the subjects in part 4 of experiment 2. The characters in the second column have the following
meaning: Nˆnumber of bids; Lˆlast bid; Hˆhighest bid so far; Eˆearnings if you stop now; Oˆnumber of bids
in a row smaller than the highest bid so far. The earnings and the percentage same as optimal are based upon a
simulation, Nˆ160 000

Subject Strategy Earnings Stopping decision

Mean SD % Same % Same % Same

as actual as actual as optimal
part 1 part 2

2 L80 or H80 OR N10 OR E30 29.7 11.8 53 a 67 86

4 L90 OR (H80 AND N10) OR E30 30.0 11.5 47 a 73 76
5 (L65 AND E6) OR H70 OR N15 OR O5 26.6 12.1 80 87 55
6 H80 30.5 10.4 47 a 87 95
7 L80 OR (H66 AND N8) OR E20 OR O5 28.5 12.1 47 67 69
8 H80 30.5 10.4 80 87 95
9 L75 29.8 10.0 67 100 77
12 L80 OR E20 29.8 9.9 53 a 93 81
13 L100 OR H70 OR N10 28.2 11.3 47 53 63
14 H84 OR E25 30.3 10.9 80 80 88
15 L80 OR (N10 AND E0) 30.1 10.8 80 100 88
16 H70 OR E10 26.9 11.1 80 67 a 56
17 L80 OR E24 OR O8 30.0 10.9 73 87 86
18 L72 OR H75 OR N15 OR E20 29.1 10.4 47 a 73 68
19 (H80 AND E20) OR (N10 AND O8) 29.1 13.8 60 73 84
20 L70 OR H70 28.6 10.4 67 a 87 64
21 H76 OR E18 29.3 9.9 80 93 72
22 N21 OR E20 29.1 11.7 53 a 67 73
23 H70 28.6 10.4 47 53 64
24 L82 30.5 10.9 80 80 95
25 H75 OR N10 OR E15 28.1 11.2 67 87 63
26 (H85 AND E18) OR N10 29.5 12.5 40 a 67 72
27 H70 OR N15 OR E10 26.9 11.2 47 53 55
28 (H75 OR N6) AND E15 29.8 10.2 47 a 73 76
29 (N6 AND H80) OR E15 28.7 10.1 67 53 68
30 H85 OR N8 OR E20 28.5 12.2 60 73 69
31 L75 OR N10 OR (H50 AND O5) 28.7 11.9 60 73 70
32 L82 OR N9 29.4 12.4 67 73 82
33 N7 OR E15 27.1 12.7 67 a 67 59
34 L70 OR (N5 AND E10) 28.6 10.4 60 a 80 64
35 N10 OR E10 26.6 11.7 80 73 55
36 H68 28.0 10.7 73 87 60
a
The strategy differs significantly from the actual behavior, Wilcoxon 2-sided test p<0.05. This is tested by
comparing the actual number of bids with the number of bids of the strategy, applied to the same sequence of
bids.

Table 3). The strategies are very close to the behavior in part 2 and somewhat less close
to the decisions in part 1.
To compare the strategies with the optimal strategy, a number of simulations were run.
The results of these simulations for all the individual strategies are displayed in Table 3.
On an average 73 percent of the stopping was optimal, and when the strategies did not
318 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332

stop optimally, they stopped mostly too early. The average number of bids is 3.93, one bid
less than the optimal strategy. The average earnings are 28.8 cents, an efficiency of 94
percent. These results are very comparable to the results in the first two parts (Table 2).
In one respect the strategies seem to be more in line with the optimal strategy than the
actual behavior is: recall occurs less in the strategies (4.1%) than in the actual behavior
(13.9% in part 1 and 16.7% in part 2). Fifteen strategies (47%) recall in less than 1
percent of the cases and only 6 strategies (19%) recall in more than 10 percent of the
cases.

4.4. Discussion
The results of this experiment are generally in line with the findings in other studies
(Schotter and Braunstein, 1981; Hey, 1987; Cox and Oaxaca, 1989, 1992; Kogut, 1990):
search is highly efficient (in terms of earnings), but fewer bids are asked than optimal and
there is some tendency to recall. This suggests that the different way of presenting
information to the subjects (compared with other studies) did not influence the search
behavior in an important manner.
There are individual differences in the use of information, the frequency of recall and
the kind of strategy described in the questionnaire. The data of the questionnaires fit the
data of the information board closely. This means that the use of an electronic
information board can help to understand the decision strategies of subjects. We also
conclude that the formal strategies in part 4 come very close to the actual behavior in the
first parts of the experiment.
The information board results show that approximately one-third of the subjects focus
mainly on the earnings. This suggests that these subjects behave more like satisficers than
as optimizers. The strategy part of the experiment shows that most subjects use combined
strategies containing both `earnings' and `last or highest bid.' Almost 70 percent of the
strategies submitted in the second part of the experiment include an earning condition and
only 22 percent of the strategies are of the optimal form. Nevertheless, efficiencies are
generally very high. In this study, as in most previous studies, subjects tend to ask a few
bids. Schotter and Braunstein (1981), Cox and Oaxaca (1989) suggest that early stopping
can be explained by risk aversion. Risk aversion can be interpreted as the willingness to
accept lower average earnings in exchange for a lower variability of the earnings.
Fig. 1 displays strategies, with the mean earnings on the vertical axis and the standard
deviation of the earnings on the horizontal axis. The strategies of the subjects are depicted
as black squares. The line shows the strategies of the optimal form; strategies on the
broken line are dominated in the sense that another strategy of the optimal form has
greater mean earnings with the same standard deviation (e.g. subjects 20, 23 and 34
would have earned on an average 2 cents more with the same standard deviation if they
did not stop at a last bid of 70 but at a last bid of 80). Only 2 strategies of a non-optimal
form (subjects 12 and 21) and 3 of the optimal form (subjects 6, 8 and 9) can be
theoretically explained by risk aversion, 22 strategies (69%) have a greater standard
deviation and a smaller mean than the optimal strategy. These strategies can neither be
explained by risk-liking; for all but one strategy (subject 24) there is a strategy of the
optimal form with a greater mean and the same standard deviation.
 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332 319

Fig. 1. Mean and standard deviation of the earnings of the strategies of the optimal form (solid and dotted curve)
and the strategies of the subjects (black squares, see Table 3). The risk-neutral optimal strategy is indicated by
the O (SDˆ10.6, Meanˆ30.5). The strategies on the dotted curve are of the optimal form with reservation prices
lower than 75. The strategies under the solid curve cannot be explained by risk aversion because there exist
strategies of the optimal form with higher mean earnings and the same standard deviation.

Learning: The results clearly show that almost none of the subjects can be considered
fully rational. While fully rational subjects would start with the optimal strategy and
would stay with that strategy during the whole experiment, boundedly rational subjects
are likely to adjust their strategy based upon their experiences. A possible solution to the
early-stopping enigma may be found in this learning behavior of the subjects. Two kinds
of learning can be distinguished: naive and sophisticated learning. A sophisticated learner
realizes that she does not know the right strategy, and will be prepared to experiment a
little in order to learn. A naive learner does not deliberately try out different strategies,
but tends to (not) repeat behavior that worked (not) well in the past.
The role of the feedback the individuals get about the effectiveness of their behavior is
essential in naive learning. In a search experiment subjects who stop early have less
opportunity to learn than subjects who stop late. This can be demonstrated with an
example. Suppose subjects A and B have tendencies to stop respectively too early and too
late and there is a series of bids `23, 75, 82, 43, 18, 69, 52, 86, etc.' Subject B will stop too
late, say after the eighth bid (86). She has observed the whole series of bids and may
realize that if she stopped after the third bid (82), she would have earned more
(82ÿ50ÿ6ˆ26 cents instead of 86ÿ50ÿ16ˆ20 cents), so she may learn to stop earlier.
Subject A stops after the second bid and does not observe the other bids in the series and
will never know that he would have earned more if he would have asked another bid. The
difference in the amount of information gathered by the subjects causes a difference in
320 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332

learning opportunity and subjects who stop late will eventually learn to stop earlier while
subjects who stop early may have difficulty in learning at all.
If at least part of the learning is naive, one would expect a asymmetry of learning
possibilities for subjects who have a tendency to stop late or too early. There are some
indications that this kind of naive learning does indeed take place in the experiment. On
average subjects search 0.66, 0.91, 1.12 and 1.07 bids less than optimal in respectively
parts 1, 2, 3 and their strategy. So if subjects learn, that learning seems to go in the
direction of fewer bids. For each subject and period a variable R is calculated as the actual
earnings minus the highest earnings one could have earned if one had stopped earlier. In
the above example the R of Subject B is 20ÿ26ˆÿ6 and of Subject A is 21‡29ˆ50. A
negative R means regret (one could have earned more) and negative reinforcement while
a positive R means that one is happy about your decision (positive reinforcement). The
data show that when R is negative, subjects stop at the next period on an average 3.6 bids
earlier and when R is positive they stop 0.637 bids later. However, these data are not easy
to interpret. Negative regret is more likely to occur in a period with a relatively
unfavourable sequence of bids, the sequence of the next period will then on an average be
more favourable and the subject will stop earlier even if his or her strategy did not
change. Statistical tests are further complicated by the fact that all subjects experienced in
each period the same sequence of bids. Therefore Experiment 2 was designed to
investigate learning behavior.
Sophisticated learners who first (systematically) test different strategies and after that
follow the most successful one, will not necessarily learn to stop too early on an average.
However, sophisticated learning can also cause systematical biases. If we assume that a
subject uses a strategy of a specific form (e.g. the optimal form) and varies a parameter in
this strategy (e.g. the reservation price) the rate of learning will be influenced by the
steepness of the parameter ± average earnings function. Where this function is relatively
flat, experimenting with different values in the neighborhood will give similar earnings
and little learning will occur. A relatively steep function will facilitate learning. In the
first experiment subjects could, not only vary parameters, but also test different forms.
This makes it difficult to study the effect of sophisticated learning in Experiment 1. In
Experiment 2, subjects have to use a reservation price and the possible effect of
sophisticated learning can be observed.

5. Experiment 2

The second experiment focuses on three research questions. One, will forcing subjects
to use a reservation price enhance performance? Second, do subjects learn? Third, can the
tendency to stop searching too early, found in Experiment 1 and several other studies, be
explained by naive or sophisticated learning?

5.1. Design
A total of 19 subjects participated in two sessions. Twelve of the 19 participants were
students of economics and 7 were known to be students of other departments; 6 were
female, 13 were male. The average age was 23.0 years. Including the show-up fee (10
 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332 321

guilders), an average of 28.50 guilders ($ 17.30) was earned by subjects in about 1 hour.
None of these subjects participated in Experiment 1.
The search problem is the same as in Experiment 1. Subjects receive instructions
and three questions to check whether the instructions are fully understood. They
start with the first period only after these questions are answered correctly. There are
no practice periods. For the first period they fill in a reservation price. The
computer draws bids until a bid is equal to or exceeds the reservation price. The
computer screen displays the sequence of bids and the earnings of that period (refer to
Appendix B). In the next period the subject can change the reservation price, or play
another period with the same reservation price. Subjects play 45 periods. The screen
shows the total earnings and bids and the results of the last 15 periods (results of the
periods before that are available by scrolling the window). The bids are drawn by the
computer and differ between subjects. After the last experiment a small questionnaire is
administered.10

5.2. Results
Reservation prices: The mean reservation price over all subjects and periods is 76.3
cents (last period 77.8 cents). Only once was a reservation price of exactly 81 reported.
Table 4 shows that there are large differences between subjects: average reservation
prices range between 60 and 90 cents. Most of the time (66%) subjects have a reservation
price lower than optimal and 52 percent of the reservation prices cannot be explained by a
risk attitude.
The mean number of bids is 5.18, a little more than the theoretical optimal of 5, but this
is caused by a few outliers. Without 6 periods with more than 50 bids (the most prominent
outlier has 299 bids!), the mean number of bids is 4.47. This is a little closer to the
optimal (5) then in Experiment 1. Overall efficiency is 90.0 percent.

5.2.1. Learning
The data (Table 5) clearly show evidence of learning. Subjects changed the reservation
price in 18.5 percent of the periods. Efficiency increases from 85 percent in the first 15
periods to approximately 95 percent in the last periods (differences between periods 1±
15, 16±30 and 31±45 are statistically significant at p<0.05, Wilcoxon test with subjects as
observations). Fig. 2 shows the proportion of changes of reservation prices upwards and
downwards for different categories of reservation prices. In all categories change in the
right direction is observed more often than change in the wrong direction. The average
reservation price after which upwards (downwards) adjustment occurs, is 69.5 (81.7)
cents.
Naive learning: Defining regret R as the actual earnings minus the highest earnings one
could have earned if one had stopped earlier, we find that subjects after experiencing

10
After 45 periods with a reservation price the subjects could play 20 periods with a strategy of the more
general form of part 2 of Experiment 1. Like in Experiment 1, they could not change their strategy between
periods. Subjects in Experiment 2 did not have the experience of the electronic information board and most of
them found it very difficult to formulate a strategy. The results of this second part of the experiment will not be
discussed in this paper.
322 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332

Table 4
Reservation prices (RP) per subject. Reservation prices between and including 76 and 81 cents can
(theoretically) be explained by risk aversion and RP>81 can be explained by risk-liking. A RP 75 cannot be
explained by risk attitude because other RP>75 can have higher mean earnings with the same standard deviation
(refer to Fig. 1)

Subject RP75 76RP81 RP>81 Mean RP

14 43 1 1 60.2
19 41 1 2 60.7
9 45 0 0 66.1
8 43 2 0 67.6
16 45 0 0 71.6
7 45 0 0 73.9
3 41 0 4 74.8
5 24 21 0 74.9
18 41 3 1 75.3
13 20 10 15 76.0
4 24 10 11 76.2
12 15 9 21 79.7
17 7 24 14 80.0
2 1 24 20 80.9
6 7 9 29 82.8
1 1 4 40 83.0
10 0 0 45 85.9
11 0 6 39 89.6
15 0 0 45 89.9
Total 51.9% 14.5% 33.6% 76.3

Table 5
Reservation prices and changes in reservation prices

Period

1±15 16±30 31±45 Total

Reservation price Mean 74.6 76.5 77.7 76.3

Standard deviation 13.1 10.8 9.1 11.2
Changes in RP Downward 8.6% 8.8% 6.7% 8.0%
Upward 16.9% 8.1% 7.0% 10.5%
Total 25.6% 16.8% 13.7% 18.5%
Efficiency 84.7% 90.8% 94.5% 90.0%

negative regret (Nˆ72) change their reservation price in 31 percent of the cases, and if
they change they lower their reservation price on an average by 9.97 cents. In contrast,
only 18 percent of the subjects who experienced positive regret (Nˆ562) change their
reservation price with a change of on an average ‡0.32 cents.11 Regret is not defined in

11
In only 7 cases the regret is exactly zero (in one case the reservation price was adapted upwards, in 6 cases
there was no change).
 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332 323

Fig. 2. Change of the reservation price. Black means a change downwards, grey means a change upwards and
white means no change. The number of cases are between brackets.

the periods in which the first bid is accepted (Nˆ195). In these cases 16 percent of
the subjects change their reservation price, the change here is on an average ‡11.64
cents.12
Sophisticated learning: There are several indications that sophisticated learning occurs.
First, there are many changes of reservation price when there is no negative regret (Many
changes after periods in which the first bid is accepted are remarkable). Second, relatively
many changes in reservation price (in both directions) are made after experiencing a
positive regret. Third, several subjects remarked in the questionnaire that they started
with experimenting with some reservation prices, and after that used the reservation price
which had the best results, or tried to fine-tune with small adjustments. This can be
interpreted as evidence that our subjects are not naive learners, but realize that they do not
know the right strategy and have to experiment. Fig. 3 shows the relation between
reservation price and efficiency. The graph is much steeper for reservation prices higher
than optimal than for reservation prices lower than optimal. This means that, for example,
a sophisticated learner who tries some numbers around 71 is less likely to conclude from

12
This number is partly caused by two outliers: subjects who in period 1 had reservation prices of 1 and 7 and
changed the next period to 78 and 77 cents. Without these two outliers the average change is ‡7.73 cents.
324 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332

Fig. 3. Efficiency of different reservation prices (expected earnings divided by the expected earnings of the
optimal reservation price of 81).

this experimenting that he should move upwards than a subject who does the same around
91 that he should move downwards.

5.3. Discussion
Forcing subjects to use a reservation price seems not to move behavior in the direction
of the theoretical optimum: the mean number of bids was somewhat closer to the
theoretically expected 5 bids, but the efficiency was only a little lower. Like in
Experiment 1, reservation prices tend to be too low and hence most subjects
search too little. Subjects seem to learn (on an average in the direction of the optimal
reservation price). Subjects tend to lower their reservation price after experiencing
negative regret (they could have earned more if they had stopped earlier). Although this
last result is in line with the naive learning hypothesis, early stopping cannot be explained
completely by this hypothesis. Over all periods reservation prices are more often adjusted
upwards than downwards, against the hypothesis. Furthermore, negative regret is
relatively rare (9% of the cases). Upwards adjustment of the reservation price especially
takes place if the first bid is accepted. Some subjects seem to reason: ``If I immediately
get what I want, maybe I should ask for more,'' a heuristic that in everyday life may be
rather successful.
Besides naive learning, sophisticated learning can explain part of the early stopping
tendency. In both experiments the bids were uniformly distributed. It is easy to see
that this effect of sophisticated learning may be even larger with other kinds of
distributions like a normal or triangular distribution (e.g. Hey, 1982; Schotter and
Braunstein, 1981).
 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332 325

6. Conclusion

The main goal of this study was to answer the question `what are the strategies of
search individuals use?' Results from previous studies are replicated: search is rather
efficient in terms of earnings but most subjects search too little and they often recall
previous bids. Remarkable individual differences are observed. Approximately 30 percent
of the subjects seem to focus their attention more on the `earnings if you stop
now' than on the highest bid so far (or the last bid). These subjects may be `satisficers', in
contrast with the `maximizers' supposed by economic theory. However, the strategy
method part of the first experiment shows that most subjects use combined strategies
containing both `earnings' and `last or highest bid' (e.g. subject 21 who submitted
the strategy, `stop if the earnings are greater then 18 or the highest bid is higher then 76'
is a satisficer during the first 4 bids and after that has a reservation price like a
maximizer).
That non-optimal strategies can be rather efficient in terms of earnings, can be
explained by the fact that the choices of satisficers who adapt their aspiration
level during a period can be very similar to the choices of the optimal strategy.
This can be illustrated by the following anecdote. A subject asked for help while
formulating his strategy and explained to me that in the first parts he always focused on
the earnings, but that he was satisfied with lower earnings if he had already got many
bids. He wanted to formulate a stopping criterion in which he stopped if the earnings are
greater than 35 minus the cost of the bids so far. I pointed out to him that because the
earnings equal the highest bid minus 50 cents minus the costs of the bids, his criterion is
in fact ``stop if the highest bid is higher than 85,'' which fits the imposed restrictions. This
illustrates that a satisficer who adapts his/her aspiration level after every bid is in fact
using a strategy of the optimal form. When, for example, the aspiration level is adapted
only after every 3 bids, the stopping behavior will somewhat resemble a strategy with a
reservation price, with the important difference that sometimes previous bids will be
recalled.
Subjects cannot be considered as being fully rational. Bounded rational subjects are
likely to change their behavior based upon their experiences. Two kinds of learning are
considered: naive and sophisticated learning. A sophisticated learner realizes that she
does not know the right strategy, and will be prepared to experiment a little in order to
learn. A naive learner does not deliberately try out different strategies, but tends to (not)
repeat behavior that worked (not) well in the past. In a search experiment naive learners
who stop early have less opportunity to learn than naive learners who stop late. This may
explain the early stopping tendency found in the present and previous studies. The results
of Experiment 2 show that naive learning can partly explain early stopping. The effects of
sophisticated learning are unclear if subjects can use different kinds of strategies
(Experiment 1), but if they are forced to use a reservation price (Experiment 2), there will
be a bias to use too low reservation prices. At least some subjects in Experiment 2 seem to
be sophisticated learners.
The methodological question, whether the electronic information board and
the strategy method can be useful tools in search behavior research, can be answered
positively. Search behavior seems to be comparable to previous studies, so the use
326 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332

of an electronic information board appears not to interfere with actual search

behavior, while it provides a lot of information about the cognitive processes of
the subjects. However, an information board gives only indications about the
kind of strategy a subject employs, it does not reveal the exact strategy. The strategy
method asks subjects for their strategies, and that appears to work very well in this
study.
Further research, using these research tools, may address the additional questions
about learning and the role of feedback in learning to search. Another interesting
topic is the way in which strategies may change if recall of previous bids is not
allowed.

Acknowledgements

Financial support by the Netherlands' Organization for Scientific Research (NWO) is

gratefully acknowledged. I am very thankful to Theo Offerman, JoÈrgen Wit, Arthur
Schram, Claudia Keser, Tanga McDaniels and two anonymous referees for their help and
comments.

Appendix A

These are the instructions of Experiment 1.

Instruction (on computer screen and on paper)

You are participating in an experiment about decision making. The experiment consists
of several periods. At the start of each period you will get an article, you will sell this
article to the computer. The computer will make as many bids as you like, and during any
moment you can decide to sell the article for the highest bid the computer has made so far
that period. Three things are very important to know:

1. In each period you have fixed costs of 50 cents (you can interpret this as the cost of the
article). In addition every bid costs 2 cents. So your COSTS in a period equals 50‡
(number of bids in that period)*2 cents.
2. Your EARNINGS in a period equals the HIGHEST BID at the moment you stop,
MINUS YOUR COSTS in that period.
3. All bids are integers between and including 1 and 100. Each of these integers is
equally likely, and independent of previous bids.

After each bid a clock on your screen will tick away from 10±0 seconds. After a
bid you have 10 seconds to decide whether you want another bid. You decide by
clicking the mouse on one of the red buttons `Another bid' or `Stop.' If you are not used
to a mouse, you can also use the keyboard and press `a' for another bid or `s' to stop.
 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332 327

During the 10 seconds you can change your mind and your decision by pushing the other
button.
In this experiment you can get information about the number of bids so far in this
period, about the last bid you received, about the highest bid so far, about the costs of the
bids so far and about what your earnings would be if you stopped at that moment. We will
give an example:

Example
So far somebody received the bids 43, 12, 78 and 63. In that case the available
information is:

4 Number of bids
63 Last bid
78 Highest bid
8 Cost of bids if you stop now (ˆnumber of bids times 2 cents)
20 Earnings if you stop now (ˆhighest bid minus the costs: 78ÿ8ÿ50ˆ20)

It is possible to have negative earnings in a period (when on the moment of stopping

the costs are higher than the highest bid). If this happens, you will lose money that period
and that loss will be deducted from the earnings in other periods.
Your total earnings will be paid to you immediately after the experiment has ended,
personally, confidentially and in cash. Your earnings are your own business, you do not
have to talk to anybody about it.
Before we start with some practice periods, we want to check whether you understand
the instructions. On your table you find 3 exercises. After you have finished these
exercises, the experimenter will come to your table to check your answers. Then the
practice periods will start. If you have questions, please raise your hand. The
experimenter will come to your table. You cannot earn or lose money in the four
practice periods.

Exercises to check understanding of the instructions (on paper)

Exercise 1
If the last 4 bids were all under 50 cents, what then is the probability that the next bid
will be greater than 50?

a. less than 50%

b. exactly 50%
c. more than 50%

Exercise 2
Someone has received successively the bids 23, 42, 3, 64 and 50. Fill out the schema
below for this case.
328 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332

Exercise 3
Someone has received successively the bids 77, 34 and 68. Fill out the schema below
for this case.

Instructions part 1, on computer screen

You just finished the practice periods. In a moment the first part of the experiment will
start. There is an IMPORTANT DIFFERENCE between the practice periods and the real
experiment: you will NOT AUTOMATICALLY RECEIVE INFORMATION about the
number of bids, the last bid, the highest bid so far, the costs so far or your earnings if you
stop at that moment anymore.
You still can get the information, but you have to click the mouse on the buttons.
If you are not used to a mouse, you can also use the keyboard and type the first
character of the text in the button. The first part of the experiment consists of 15
periods.
 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332 329

Instructions part 2, on computer screen

You just finished the first part of the experiment. In a moment the second part of the
experiment will start. The situation will be quite similar to the situation in the first part.
The only difference is that from now on AFTER EACH BID YOU CAN CLICK ONLY
ONE BUTTON FOR INFORMATION.
After each bid you can get information about the number of bids, OR the last bid, OR
the highest bid so far, OR the costs so far, OR the earnings if you would stop on that
moment. The second part of the experiment consists of 15 periods.

Instructions part 3, on computer screen

You just finished the second part of the experiment. In a moment the third part will
start. The situation is similar to the situation in the second part, with one important
difference. From now on a period starts in the situation that a number of bids are already
made. Before the start of the period the computer has made 0, 1, 2, 3 or 4 bids (all these
situations are equally likely).
330 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332

The costs of the bids still equal the number of bids times 2 cents; so the bids the
computer has made before the start of the period also cost money! The fixed costs are still
50 cents each period. After each bid you can push only one button for information. The
third part of the experiment consists of 15 periods.

Instructions part 4, on paper

In part 4 of the experiment we ask you to formulate a strategy. The computer will play
20 periods with this strategy; the results of these 20 periods will be paid to you. If you
have decided on a strategy, it will be fixed for all 20 periods; you will not get an
opportunity to change it anymore.
The situation is the same as in parts 1 and 2 of the experiment:
 The costs in a period equal 50 cents plus the number of bids multiplied with 2
cents;
 The earnings equal the highest bid on the moment of stopping, minus the
costs;
 All bids are integers between and including 1 and 100. Each of these integers is equally
likely, and independent of previous bids.

In a previous experiment participants filled out the same questionnaire that you have
finished a moment ago. These questionnaires gave us an idea of the kind of strategies
participants use.

If the your stopping criterion depends on one of the stopping conditions A, B, C, D or

E, you fill in a value after that condition and you put that letter after `Stopping criterion.'
It is also possible to make a stopping criterion which is a combination of the stopping
conditions. In that case you fill in the values for the conditions you use, and you use the
words and or or to make the stopping criterion.
For example, if you want to stop if both conditions A and B are fulfilled, you make the
criterion `A and B.' At the other hand, if you want to stop if one of the conditions A or B
is fulfilled, you make the criterion `A or B.' It is also possible to make a criterion which
combines more than two conditions. In that case you may need to use brackets, for
example, `(A or B) and C' or `A or (B and C)' (note that these are different criteria).
 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332 331

Examples given are completely arbitrarily and do not indicate what kind of strategy
will result in a large profit. If you want to use a strategy which cannot be formulated with
the above 5 conditions, please ask the experimenter. If something is not totally clear,
please raise your hand.

Appendix B

Computer screen Experiment 2

References

Axelrod, R., 1984. The Evolution of Cooperation, Basic Books New York.
Camerer, C., 1995. Individual decision making. In: Kagel, J., Roth, A. (Eds.), Handbook of Experimental
Economics. Princeton University Press, Princeton, NJ.
Cox, J.C., Oaxaca, R.L., 1989. Laboratory experiments with a finite-horizon job-search model. Journal of Risk
and Uncertainty 2, 301±330.
Cox, J.C., Oaxaca, R.L., 1992. Direct tests of the reservation wage property. The Economic Journal 102, 1423±
1432.
Harrison, G.W., Morgan, P., 1990. Search intensity in experiments. The Economic Journal 100, 478±486.
Hey, J.D., 1982. Search for rules for search. Journal of Economic Behavior and Organization 3, 65±81.
Hey, J.D., 1987. Still searching, Journal of Economic Behavior and Organization 8, 137±144.
Hey, J.D., 1991. Experiments in Economics. Blackwell, Oxford.
Keser, C., 1992. Experimental Duopoly Markets with Demand Inertia, Game-Playing Experiments and the
Strategy Method, Springer, Heidelberg.
Kogut, C.A., 1990. Consumer search behavior and sunk costs. Journal of Economic Behavior and Organization
14, 381±392.
332 J. Sonnemans / J. of Economic Behavior & Org. 35 (1998) 309±332

Kohn, M.G., Shavell, S., 1974. The theory of search. Journal of Economic Theory 9, 93±123.
McMillan, J., Rothschild, M., 1994. Search. In: Aumann R., Hart S. (Eds.) Handbook of Game Theory, vol. 2,
pp. 905±927.
Milgrom, P., Roberts, J., 1991. Adaptive and sophisticated learning in normal form games. Games and Economic
Behavior 3, 82±100.
Miller, J.H., Andreoni, J., 1991. Can evolutionary dynamics explain free riding in experiments?. Economic
Letters 36, 9±15.
Moon, P., Martin, A., 1990. Better heuristics for economic search ± experimental and simulation evidence.
Journal of Behavioral Decision Making 3, 175±193.
Payne, J.W., Bettman, J.R., Johnson, E.J., 1993. The Adaptive Decision Maker, Cambridge University Press,
Cambridge.
Roth, A.E., Erev, I., 1995. Learning in extensive-form games: Experimental data and simple dynamic models in
the intermediate term. Games and Economic Behavior 8, 164±212.
Schotter, A., Braunstein, Y.M., 1981. Economic search: An experimental study. Economic Inquiry 19, 1±25.
Selten, R., 1991. Evolution, learning and economic behavior. Games and Economic Behavior 3, 3±24.
Sonnemans, J., Schram, A., Offerman. T. (forthcoming), Public good provision and public bad prevention: The
effect of framing, Journal of Economic Behavior and Organization.