The Journal of Risk and Uncertainty, 30:3; 241259, 2005

c 2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.

Informational Asymmetries and Observational Learning
in Search
LIRAN EINAV [email protected]
Department of Economics, Stanford University, Stanford, CA 94305-6072, USA
Abstract
As economics modeling moves from super rational decision makers to considering boundedly rational agents,
some economic problems deserve a second look. This paper studies the effects of learning on the efciency of
search. Once learning is taken into account, the structure of information ow becomes important. In particular,
I highlight the truncated information structure in the search problem. Agents stop searching once a sufciently
attractive price is found. Therefore, they observe the performance of shorter searches, but do not directly observe
the performance of longer searches. I design and conduct an experiment to test the hypothesis that this asymmetric
ow of information leads agents to search too little. I nd strong evidence in its favor. This suggests that in the
presence of learning, the provision of a more symmetric information structure will make search more efcient.
Keywords: search, learning, experiment, regret
JEL Classication: C91, D83
Nowadays it is widely agreed that learning and other behavioral regularities are important
in many economic settings. The importance of learning raises a full set of new questions
related to the speed at which learning takes place. In a repeated interaction, if learning is
fast, equilibrium analysis is of great importance. If learning is slow, however, many real
world observations may be of off-equilibriumactions, and hence our prediction and analysis
should take a different path.
1
As economics modelingmoves fromsuper rational decisionmakers toconsideringbound-
edly rational agents, some economic problems deserve a second look. This paper studies
the effects of learning on the efciency of search. The solution to the search problem is
straightforward for super rational decision makers, who compute the optimal reservation
price and search accordingly.
2
Boundedly rational agents, however, cannot compute the
optimal reservation price in advance, and hence adjust their search strategy over time and
learn from experience.
Once learningis takenintoaccount, the structure of informationowbecomes important. I
suggest that strategies used by agents are more likely to be reevaluated once it was actually
observed that a different strategy could have done better. A more frequent reevaluation
leads, in turn, to faster learning. Hence, learning is likely to be faster when it is possible
to observe hypothetical earnings from strategies that were not actually played. Whether
or not agents observe these hypothetical earnings depends on the specic information
ow.
242 EINAV
This paper highlights the truncated information structure in the search problem. When
agents search for the highest price, the search ends once a sufciently high price is found.
The typical cutoff stopping rule makes agents observe only part of the price distribution,
truncatedfromabove bythe nal transactionprice. Suchinformationdoes not allowagents to
symmetrically assess and evaluate their strategies, and hence slows their learning process.
Therefore, as agents search, they obtain information about the performance of shorter
searches, but not about the performance of longer searches. Consequently, agents who
search too much will observe the benets of shorter searches; they will be able to adjust
their behavior and to search less in subsequent searches. In contrast, agents who search
too little do not get the same opportunity to observe their errors, and will not be as likely
to correct it and search more in subsequent searches. The hypothesis suggests that a more
symmetric information ow will signicantly increase the speed of learning and therefore
improve the efciency of search. This idea is consistent with directional learning theory, as
suggested by Selten and Buchta (1994).
I test this hypothesis and its implications in an experimental setup. As discussed in more
detail in Section 1, there is one common phenomenon that stands out from the vast majority
of experimental papers about search behavior: subjects in experiments search too little when
compared to the optimal search strategy.
3
Section 1 discusses three potential reasons for this
result. The rst is risk aversion, which provides an incentive to avoid risky search and stop
too early. The second is an asymmetric cost structure around the optimal search strategy:
searching too much would typically be much more costly than searching too little. The
third potential reason for the too little search result is a direct implication of the hypothesis
discussed above, which is the focus of the paper. It suggests that the truncated information
structure of the search problem may be a driving force behind the robust too little search
result.
Section 2 describes the experimental design that tests this possible explanation. The
control group uses the standard search environment, in which subjects do not get any further
feedback once they stop searching.
4
In contrast, subjects in the other experimental treatment
face a more symmetric informationstructure byobservingpost-purchase information. While
post-purchase information is useless for subjects who search optimally, it may be used as a
learning device for boundedly rational agents. In particular, if an asymmetric information
ow is important, post-purchase information may provide a more symmetric feedback,
thereby generating longer searches.
The results are described in Section 3. They showthat subjects who face a more symmetric
information ow search signicantly more, although still less than the optimal level. In
addition, I document the important effects of information ow on learning and on strategy
adjustments. I nd strong evidence for the learning mechanismdescribed above: subjects are
signicantly more likely to adjust their strategies in the direction of the information they
actually observe. In particular, after subjects observe that a shorter (longer) search would
have done better, they are more likely to shorten (lengthen) their search in a subsequent
search. This result is robust, statistically and economically signicant, and is present in
both experimental treatments. This is precisely the mechanism which makes post-purchase
information so effective: subjects in the control group cannot observe such information
and therefore can never actually observe the performance of longer searches. In contrast,
INFORMATIONAL ASYMMETRIES AND OBSERVATIONAL LEARNING IN SEARCH 243
subjects in the experimental treatment, who face a more symmetric information structure,
will sometime observe that longer search would have done better. Thus, they are more likely
to revise their search strategy towards longer searches.
As discussed in Section 4, these results have important implications regarding the ef-
ciency of search. They suggest that once these behavioral regularities are taken into account,
search rules should be revised to adjust for the truncation of information ow. These im-
plications may be more important for search markets in which post-purchase information
is less common, i.e. in which it is uncommon for participants to observe prices when they
are not actively searching. For example, it implies that post-purchase, supposedly irrele-
vant information may be actually valuable in improving search strategies in subsequent
purchases; by making mistakes more salient, economic agents are more likely to reconsider
their search strategy and search more efciently in the future.
1. Motivation and background
The literature on consumers search and its implication on market equilibrium dates back
to Stiglers (1961) seminal paper The Economics of Information. Most of the literature
deals with homogeneous good markets, in which there exists price dispersion. The general
setting is that consumers have information only about the distribution of prices and they
search for the lowest price. Since search is costly (xed cost per search) and consumers
cannot afford to search as much as they would like, higher prices survive in equilibrium.
In this paper I focus on the simple sequential search problem, in which an agent sequen-
tially obtains prices from a known price distribution. The agent searches for the highest
price at which to sell a good, and has to decide when to stop; this is the prominent setup
in the literature. In such a case, we can think of the agent as a seller (or a job seeker), who
searches for the highest bid (wage) for her product (services).
I rst outline the theoretical predictions for the simple search model. If the agent perfectly
knows the (stationary) distribution from which prices are sampled, then her optimal search
employs a strategy of a reservation price. The risk-neutral agent continues to search as long
as the expected value of an additional bid is higher than the cost of obtaining this bid, i.e.
the reservation price R is the solution to
c = (1 F(R))E
F
( p R | p R) (1)
where c is the search cost, p is the sampled price, and F() is the cumulative distribution
function of prices, which is known to the agent.
5
Throughout I do not allow the possibility of recall, i.e. whether the agent can change her
mind and accept previously rejected offers. While the possibility of recall is favorable to
the agent, note that in this standard settingstationary price distribution and identical cost
per searchreservation price strategy is optimal, so a rational agent would never recall. In
addition, by construction, there is no recall when using any other, sub optimal, reservation
price, so recall is not an issue for the present study, as will become clear later on.
Experimenters in economics started studying individual search behavior in the early
1980s. The main question they tried to answer is how individuals search, and how close
244 EINAV
they are to optimal behavior. As already mentioned, the vast majority of the experiments
share a common result: subjects tend to search too little compared to the optimal rule.
6
This
result is the main motivation for this paper.
There are different competing explanations for the too little search result. The rst,
which is often discussed in the literature, is risk aversion. Once a subject is faced with the
decision whether to stop or to continue, the choice is equivalent to choosing a lottery. The
higher the level of risk aversion, the less likely the subject will accept the lottery and
continue searching. There are reasons to believe, however, that risk aversion is not enough
to explain the too little search result. First, all experiments offer very lowmonetary prizes,
over which one may assume that subjects are locally risk neutral (Rabin, 2000). In addition,
in a repeated search problem(which is the case of most experiments) subjects are diversied
over time and bear lower risk.
7
The risk aversion problem is further reduced once subjects
are restricted to use predened search strategies. Indeed, Sonnemans (1998) shows that risk
aversion can be a possible explanation for only twenty percent of his subjects who search
too little.
8
The second possible explanation, which is suggested by Sonnemans (1998), is that stop-
ping rules that give rise to too little search perform rather well in most cases. On the other
hand, stopping rules that give rise to too much search perform quite poorly. Hence, subjects
who search too much are more likely to revise their rules downwards, so on average subjects
will search too little. In Einav (2002) I test for this explanation, and show that, indeed, it
partially explains the too little search result.
The third explanation, which has not been tested in the literature, arises fromthe structure
of information owin the search problem, and the notion of observational regret as a learning
device. Subjects can experience regret in this setting when they observe a previously rejected
price that could have made thembetter off.
9
When the stopping rule is of a reservation price,
then only prices that are lower than the reservation price can be a source for regret. A price
higher than the reservation price cannot be a source for regret, because the only such price
observed will be the transaction price, which is the last in the sequence of offers (after which
search ends). Consequently, if observational regret enhances learning and affects subjects
when they revise their stopping rules, then we should observe more revisions downward,
so subjects will end up searching too little. It is important to note that the mechanism
described is not based on anticipated regret that enters into the individuals utility function,
but on observational regret that serves as a learning device that reminds the individual
to reevaluate her strategy.
10
This explanation is the focus of the experiment described
below.
2. Experimental design
The experimental design has two different experimental conditions, a control group and a
treatment group. Subjects in the control group face the experimental design of Sonnemans
(1998), while subjects in the treatment group face a slightly different information structure.
Differences between the groups will allow me to test my central hypothesis.
The setup uses a repeated search for the highest price (bid), in which subjects are restricted
to use reservation price strategies. The reservation price can be revised fromperiod to period,
INFORMATIONAL ASYMMETRIES AND OBSERVATIONAL LEARNING IN SEARCH 245
but not between bids within the same period. Subjects have full information about the price
distribution, and they pay a xed amount at the beginning of each period.
11
There are several reasons why this setting is attractive. First, the distribution is known
to the subjects, so there is no room for prior beliefs that subjects might form, for which
the experimenter cannot control. Complete information about the distribution also con-
trols for heterogeneity among subjects in information-gathering techniques, and provides
a unique optimal strategy that can serve as a benchmark for comparison. If informa-
tion is not complete, such optimal strategy depends on the prior, which may vary across
subjects.
12
Second, the search problem is repeated, so search is observed for consecutive periods
and learning can take place. As has been widely shown in the experimental literature,
13
learning is key in many experiments, and many subjects need experience in order to better
understand the setup and the environment.
Third, while making the problem somewhat less realistic, the use of predened strategies
allows the experimenter toavoiddealingwithdynamicallyinconsistent behavior bysubjects,
which is likely to occur in a search problem. Finally, strategies are simple, and consist only
of a reservation price. This allows for a simple and convenient quantitative prediction and
analysis, without losing too much generality.
The price distribution is uniformover a discrete support, which includes integers between
1 and 100 (cents), so each of these integers has a probability of one percent. The search cost
is 2 cents for each new bid. The xed cost per period is 50 cents. The optimal reservation
price in such a setting is 81 cents. Sonnemans (1998) replicates the too little search result
in such a setting, so it is reasonable to explore this result even with these restrictions on
search strategies.
14
Let us dene the concept of (observational) regret: subjects observe regret if their prots
could have been higher by stopping earlier.
15
Using this denition, Sonnemans results
show that although regret is not very frequent, it has a very important role: subjects are
twice as likely to revise their reservation price after experiencing regret, and when they
do, the absolute magnitude of change is much higher. Similar results are replicated in this
paper and provide the motivation for a deeper look at the role of observational regret as a
learning-enhancing device.
As mentioned earlier, this basic setup results in an asymmetric information structure,
whichleads toa one-sidedregret. For that, I distinguishbetweentwotypes of regret. The rst,
upward regret, is a case in which a subject observes that she could have done better by setting
a higher reservation price and continuing to search. The second, downward regret, is the
original case of regret in which a subject could have done better by setting a lower reservation
price and stopping the search earlier. In the control group, by construction, subjects can only
experience downward regret, because they do not observe bids after they stop searching.
Therefore, the second experimental treatment allows for a more symmetric information
ow by making subjects observe post-purchase bids. This allows for experiencing regret in
both directions. In particular, I let subjects in the second experimental treatment observe
three additional bids that they would have gotten had they continued searching (by setting
higher reservation price). Other than that, subjects follow exactly the same procedure as
subjects in the control group. In such a setup, a downward regret is still possible, as in the
246 EINAV
control group, but now there is also the possibility of an upward regret. This may be the
case if one of the three additional bids is sufciently high.
The three additional bids are, on average, about 60%more bids than typically observed, as
the average number of bids around the optimal strategy is about ve. It allows for sufciently
high probability (about 45% at the optimal strategy) of observing upward regret, without
making it too tedious for the experimental subjects. Given the results described below,
which suggest that these additional bids affect learning and efciency, it may be interesting
to investigate how learning changes when the number of additional bids varies. I leave this
for future research.
3. Implementation and results
After a small-scale pilot run that led to a few technical changes in the experimental design,
I ran ve identical sessions of the experiment. The subjects were chosen (on a rst-come-
rst-served basis) from the subject pool of the Computer Lab of Experimental Research
(CLER) at the Harvard Business School. The majority (but not all) of the subjects in this
pool are undergraduate students from Harvard University and Boston University.
16
Each
of the runs was performed in a single session that included subjects from both treatments.
All sessions were held at the CLER and lasted about 90 minutes each. The random draws
from the distribution were drawn in advance, and were identical for all subjects. Subjects
were not aware that they all faced identical draws.
17
The number of subjects was 35 for the
control group and 40 for the experimental treatment.
In what follows I describe the experimental results. First I report the means and medians
of earnings and reservation prices over time and across experimental treatments. These
results give a general idea of the behavior of subjects, and are most easily compared to
results of previous experiments. Next, I continue by looking at other dimensions of the
experimental results, focusing on observational regret.
Using regression analysis, I obtain strong evidence in support of the hypothesis that
observational regret serves as a learning-enhancing device (Tables 3 and 4). The results
show that an observed downward regret makes subjects signicantly more likely to de-
crease their reservation price, and that an observed upward regret makes subjects signif-
icantly more likely to increase their reservation price. Probabilistic regret, which is not
actually observed, has no signicant effect. In the end of this section I summarize my
ndings.
3.1. Descriptive statistics
Table 1 presents the minimum, maximum, and average earnings for each treatment. As can
be seen, subjects earned on average a little less than $30, with earnings varying between
$20 and about $31, except for few subjects who earned very little. The earning distributions
for the two groups are very similar.
At the end of each session, after subjects had received their earnings, they were asked to
describe in writing the way they approached the problem. Despite the big effort to make
INFORMATIONAL ASYMMETRIES AND OBSERVATIONAL LEARNING IN SEARCH 247
Table 1. Earnings.
Control Treatment
N 35 40
Min 11.98 7.51
Mean 26.17 26.23
Median 28.18 28.74
Max 31.29 31.02
Std. dev. 9.36 10.30
Optimal

30.90 30.90
All gures in US Dollars.
Optimal refers to the earnings obtained by
setting 81 at all periods.
Note that subjects were guaranteed $20
(so the few subjects who obtained less than
that received $20 in the end of the experiment).
subjects understand the setup, it seems from the answers that some of them had a very
weak idea of what they were supposed to do during the experiment. For comparing the
means and medians of reservation prices across the two experimental conditions (and only
for that), these subjects introduce biases and increase the variance, so I chose to drop them
from the sample for this specic analysis.
18
The reason is that a random draw from the
distribution has a mean of about 50, while the majority of the observed reservation prices
are in the range of 70 to 85. It turned out that subjects in the experimental treatment were
more likely to get confused and to think that they do not understand the instructions. This is
due to the additional observed draws, which had no practical effect on the outcome (beyond
their effect through learning). Therefore, comparing means with these random subjects
in the sample is likely to bias the results downward in this treatment, when compared to
the control group, and hence create problems in the interpretation of the simple comparison
of means across groups. In all other statistical analyses, such random subjects can be
treated as a white noise, so I keep them in the sample, and they have little effect on the
results.
Table 2 and Figure 1 show the average and median reservation price for each group,
as it evolved over the 100 periods of the experiment. As in previous experiments, which
are discussed in Section 1, it is shown that, on average, subjects search too little. For both
groups in all periods the average reservation price is below the optimal reservation price
of 81, making subjects stop searching too early. The results conrm the hypothesis that
subjects who may experience symmetric regret by observing higher prices after they have
stopped searching end up searching longer and setting higher reservation prices, closer to
the optimal one. Once the random subjects are dropped, there is much less noise in the
averages, so I am able to obtain statistically signicant differences between the two groups.
There are a fewadditional interesting results that should be mentioned. First, even subjects
in the experimental treatment, who search more, still search too little by setting average
reservation prices that are signicantly lower than the optimal one, 81. This can be explained
248 EINAV
Table 2. Averages of reservation prices.
Control Treatment
All subjects
First period 63.3 (2.76) 63.1 (3.18)
All periods 68.7 (2.33) 69.4 (2.26)
Last 50 periods 70.8 (2.41) 70.8 (2.32)
Last 25 periods 71.3 (2.37) 70.8 (2.43)
Last 10 periods 72.0 (2.46) 71.3 (2.43)
Last period 72.7 (3.02) 70.8 (2.93)
No. of subjects 35 40
All Non-Random subjects
First period 66.5 (2.17) 66.7 (3.38)
All periods 72.6 (1.83) 75.2 (1.45)

Last 50 periods 75.1 (1.48) 77.1 (1.19)

Last 25 periods 75.1 (1.57) 77.6 (1.12)

Last 10 periods 75.2 (1.66) 77.9 (1.23)

Last period 76.0 (1.78) 79.1 (1.52)

No. of subjects 30 30

Signicantly different from the control group with p-value

> 5% for a one-sided t -test.

Signicantly different from the control group with p-value

> 10% for a one-sided t -test.

Signicantly different from the control group with p-value

> 20% for a one-sided t -test.
Standard deviations in parentheses.
All gures in all tables are signicantly (at 5%) different from
81, the optimal reservation price.
The criteria for dening a subject as random is given and mo-
tivated in endnote 18. This table and the corresponding Figure 1
are the only place in the paper where such selection is applied.
All other analysis of the paper is based on the full sample.
by the other factors that affect search strategy (see Section 1), such as the cost asymmetry
(see also Einav, 2002). Second, it is quite clear that the results are not driven by initial
conditions. The rst period is completely identical for both group,
19
and subjects indeed
start from almost the same average reservation price, as shown in the rst row of Table 2.
Finally, it is important to note that even after one hundred repetitions subjects continue to
adjust their reservation prices, albeit they do so less frequently.
3.2. Observational regret as a learning-enhancing device
In this section I focus on the effects of observational regret, which is the main focus of the
paper. In analyzing observational regret, it is useful to make comparisons with an articial
INFORMATIONAL ASYMMETRIES AND OBSERVATIONAL LEARNING IN SEARCH 249
Figure 1. Averages and medians over time.
250 EINAV
experimental group, the Pseudo control group. This comparison looks at the behavior
of subjects in the control group as if they were observing the three additional prices just
as the subjects in the experimental treatment.
20
This allows me to better control for the
results associated with the experimental treatment. That is, I control for the possibility that
observational regret itself has no impact, but is simply correlated with some other relevant
variables. For example, it is natural to assume that subjects will be more likely to adjust
their reservation price downwards (upwards) after setting a relatively high (low) reservation
price. Downward (upward) regret is also more likely to be observed after setting a high (low)
reservation price. Hence, an adjustment upward may be positively correlated with upward
regret only because of statistical reasons, without any causal effect of observational regret.
By using the pseudo control group, and comparing it to the experimental treatment, I am
able to obtain the perfect control, so that the only difference among subjects in the two
groups is the actual observation of regret.
While simple empirical probabilities may provide an idea of the effect of observational
regret on subjects behavior, the cleanest test can be done using a more rigorous statistical
analysis. In what follows, I use regression and multinomial logit analysis to test the idea of
directional learning through observational regret.
21
Tables 3 and 4 analyze the effects of observed regret on the subsequent adjustment of
reservation prices. In particular, for each subject i in period t (t > 1) I regress the change in
reservation price from period t 1 to period t (i.e. r p
t
r p
t 1
) as a function of observed
regret in period t 1 and other controls. The control variables, which include the reservation
price at period t 1, the number of bids in period t 1, and the dollar earnings (in cents) in
period t 1 are all trying to make sure that the estimated effect of observed regret is indeed
causal and driven by learning, and it is not due to some other statistical regularity, such as
the one mentioned above. All regressions are run separately for each group, and cluster by
subject, to correct for the likely correlation in each individuals behavior over time. The
most convincing result, in my view, is the comparison between the effect of upward regret
in the pseudo control group and the treatment group. The signicant difference in its effect
across the two groups can only be attributed to the actual observation of upward regret.
Table 3 reports a linear regression in which the dependent variable is a logistic transfor-
mation of the change in reservation price (so that the results are not driven by few extreme
changes, from 1 to 100 and vice versa). In all groups the effect of a downward regret is
negative and statistically signicant: ceteris paribus, when a subject observes a downward
regret, she is more likely to subsequently choose a lower reservation price. The effect of an
upward regret works at the other direction: everything else equal, subjects in the treatment
group who observe an upward regret are more likely to subsequently choose a higher reser-
vation price. While this effect loses some of its statistical signicance once more control
variables are added (column 6; its p-value is about 12%), it is in sharp contrast to the absence
of this effect (in terms of magnitude and signicance) in the pseudo control regressions.
Subjects in the control group are not affected by upward regret. This suggests that it is the
actual observation of upward regret, rather than some hypothetical regret, that makes sub-
jects adjust their strategies. Often, this adjustment would be in the right direction, towards
the optimal reservation price, as most subjects use too low reservation prices. This is the
primary factor for the efciency differences between the groups shown in Figure 1.
INFORMATIONAL ASYMMETRIES AND OBSERVATIONAL LEARNING IN SEARCH 251
Table 3. Linear regressions.
Control Pseudo control Treatment
Dependent variable: logistic transformation, exp(x)/(1 +exp(x)), of the change in reservation price from
period t 1 to t .
Downward regret
t 1
0.176

0.146

0.188

0.158

0.172

0.092

(4.75) (3.81) (5.08) (4.18) (5.71) (2.55)

Upward regret
t 1
0.006 0.0016 0.037

0.026
(0.35) (0.10) (2.46) (1.59)
Constant 0.767

0.795

0.760

0.795

0.806

0.851

(11.87) (9.54) (12.14) (9.64) (10.92) (9.49)

Reservation price
t 1
0.0036

0.0031

0.0035

0.0031

0.0043

0.0036

(4.18) (3.84) (4.19) (3.84) (4.26) (3.71)

No. of bids
t 1
0.0033* 0.0032* 0.0073

(2.41) (2.52) (4.15)

Earnings
t 1
0.0006 0.0006 0.0008
(0.83) (0.84) (1.30)
No. of subjects 35 35 35 35 40 40
No. of observations 3,465 3,465 3,465 3,465 3,960 3,960
R-squared 0.054 0.056 0.055 0.057 0.089 0.095
t -stats below coefcients, clustered by subject.

,

implies signicance at the 1 and 5% condence level, respectively.
The regressions are run for each experimental treatment separately. The control group and the pseudo control
group use the same data. The only difference is that in the pseudo control an upward regret regressor is
computed, as if subjects could observe post-purchase information. The coefcient on Downward regret
suggests the existence of observational learning. The most important observation, however, is the difference
in the magnitude and signicance in the coefcient on Upward regret in the pseudo control group and the
treatment group regressions. Its signicance only in the treatment group regression implies that it is only the
actual experience of regret that affects behavior. The rest of the regressors are controls. They all obtain the
predicted signs. The regressions include all subjects.
To gain some intuition for the magnitude of the effect, a subject who would otherwise not
change her reservation price will decrease (increase) it by about 0.7 (0.15) after experiencing
downward (upward) regret.
22
This magnitude is quite large considering the fact that, on
average, each subject increases her reservation price by about 0.08 per period. Loosely
speaking, a downward regret offsets about nine periods without regret, while an upward
regret is equivalent to three periods without regret.
The control variables all have the predicted signs. Higher reservation prices make it more
likely to adjust downwards, and higher number of bids have similar effect. Earnings do
not seem to have a signicant effect on the strategy adjustment. These coefcients on the
control variables illustrate why it is important to use them. Downward regret is more likely
after a high reservation price and high number of bids. Thus, if we did not control for these
variables in the regression, we would obtain stronger estimated effect of downward regret,
which would only be driven by statistical properties of regret. After using these variables
252 EINAV
Table 4. Multinomial logit regressions.
Control Pseudo control Treatment
Downward Upward Downward Upward Downward Upward
Dependent variable change change change change change change
Downward regret
t 1
4.34

1.46 4.98

1.56 4.53

1.28
(3.98) (1.05) (4.27) (1.18) (4.71) (0.60)
Upward regret
t 1
1.10 1.15 1.02 1.42

(1.09) (1.41) (0.22) (3.64)

Reservation price
t 1
0.993 0.972

0.994 0.973 0.987 0.962

(0.46) (1.97) (0.42) (1.87) (0.99) (3.11)

No. of subjects (observations) 35 (3,465) 35 (3,465) 40 (3,960)
Log-likelihood 3,645.7 3,642.2 4,037.1
Pseudo R-squared 0.032 0.033 0.058
Downward regret
t 1
2.08

1.12 2.38

1.20 1.85

1.58
(2.00) (0.36) (2.23) (0.56) (2.00) (1.20)
Upward regret
t 1
1.05 1.06 0.95 1.24

(0.44) (0.48) (0.46) (2.17)

Reservation price
t 1
0.992 0.974 0.992 0.974 0.985 0.968

(0.53) (1.80) (0.53) (1.81) (1.19) (2.73)

No. of bids
t 1
1.052

1.005 1.051

1.006 1.054

0.948

(2.87) (0.25) (2.75) (0.32) (4.49) (2.96)

Earnings
t 1
0.993 0.991 0.994 0.992 0.991

0.988

(1.45) (1.59) (1.02) (1.11) (2.08) (2.47)

No. of subjects (observations) 35 (3,465) 35 (3,465) 40 (3,960)
Log-likelihood 3,631.3 3,630.1 4,006.8
Pseudo R-squared 0.036 0.036 0.066
Estimates are reported as odds ratios. z-stats are reported in parentheses, clustered by subject.

,

implies signicance at the 1% and 5% condence level, respectively.
In all regressions, the base category is no change in the reservation price.
The regressions reported below are similar to those reported in Table 3. Rather than using the magnitude of the
change in reservation price, these regressions replicate the same pattern of Table 3, when the dependent variable is
the discrete choice of whether to change the reservation price, and if so in which direction. The top panel runs the
regressions with no controls, while the bottom panel adds several controls. The results are qualitatively similar.
The specication is of a multinomial logit, where the reported gures are of odds ratios.
As in Table 3, the Downward regret works as predicted: it increases the probability of changing the reservation
price downwards, but has no signicant effect on a change upwards. As before, the most interesting contrast is
between the effect of Upward regret in the pseudo control group, where it is not signicant, vis-` a-vis its signicant
effect in the treatment group. Again, this implies that it is only the actual experience of regret that affects behavior.
INFORMATIONAL ASYMMETRIES AND OBSERVATIONAL LEARNING IN SEARCH 253
as controls, it seems much more likely that the remaining effect of regret is indeed due to
learning.
Table 4 repeats the exercise, using multinomial logit regression. Rather than using the
dependent variable as a continuous measure, Table 4 only analyzes the direction of strategy
adjustment, tofocus onthe directional learningidea. Thus, the dependent variable is discrete:
upward adjustment of the reservation price (r p
t
r p
t 1
> 0), no adjustment (r p
t

r p
t 1
= 0), or a downward adjustment (r p
t
r p
t 1
< 0). The results are similar to
those presented in Table 3. In all groups, observing a downward regret makes it more likely
that subjects adjust their reservation price downwards, while observing an upward regret
makes it more likely that subjects in the treatment group adjust their reservation prices
upwards. In sharp contrast, (hypothetical) upward regret has no signicant effect in the
pseudo control group, suggesting that it is only the actual observation of regret that affects
behavior.
Again, the magnitude of the effect is quite substantial. On average, each subject ad-
justs her reservation price downwards about 28 percent of the times. The odds ratio of
4.34 in the top panel of Table 4 implies that after experiencing a downward regret, this
probability is much higher, and is about 64 percent. Similarly, the odds ratio of 1.42 for
upward regret (the last column of Table 4) implies that the probability of adjusting the
reservation price upwards become about 40 percent, compared to an average probability of
32 percent.
3.3. Summary
The results replicate well previous results that found that subjects search too little. In this
setup it is shown that subjects in both groups set on average a reservation price which is
signicantly lower than the optimal reservation price of 81. In addition, I showthat subjects
who are able to obtain more symmetric information structure tend to search better, closer
to optimal, albeit still too little. Finally, the results document quite strongly the effect of
actual observation of regret, in both directions, upward as well as downward. The analysis
suggests that observing regret can be thought of as a feedback to the subjects, which enhance
their learning. Once the feedback is objective, symmetric in both directions, subjects react
accordingly and better converge towards the optimal reservation price.
4. Conclusions and further research
Previous experimental search papers have consistently found that subjects in experiments
search too little. In this paper I test for a possible explanation for this robust result, and nd
evidence in its support. Subjects, who can evaluate their strategies more symmetrically by
observing post-purchase information, search more intensively.
In addition, the results show the importance of observational learning in the form of
observational regret as a learning-enhancing device. Using the individual search problem it
is shown that actual observation of regret leads to a directed adjustment of played strategies,
while probabilistic regret has no statistically signicant effect. It is evident that subjects
254 EINAV
who are able to actually observe regret end up searching in a more efcient way by setting
reservation prices that are closer to the optimal one.
We can try to extend these results to strategic interactions. For example, one could look
at a sequential two-player game, using both the action method and the strategy method.
23
In the strategy method subjects are informed about the full strategy of their opponents, so
they can calculate their unrealized payoffs. In this sense, the strategy method is similar to
the experimental treatment and allows for observational regret, while the action method is
similar to the control group, in which subjects may only speculate what could have been
their earnings had they chosen a different action. The results from this paper suggest that
learning and convergence to equilibrium would be faster when the strategy method is used.
The observation of post-purchase information is not only a mechanical device to test the
explanation for too little search. It may also be a closer description of certain markets in the
real world. It is quite often that a consumer buys a certain good and later observes the same
good being sold at a lower price. This new information can be obtained from a friend who
bought the same good, fromadvertisement, or fromvisiting the same retail store for another
reason. Moreover, it seems plausible that consumers who bought a certain product recently
are more likely to internalize newinformation about it. Consequently, the results obtained in
this paper suggest an important empirical implication: search is going to be closer to optimal
in markets in which consumers are more likely to obtain post-purchase information. This
prediction may be testable. This can be done by comparing search behavior in markets
for similar products that differ in their information structure.
24
In markets where search is
repeated but post-purchase information is not frequently available, a more efcient search
can be achieved by designing search strategies that will make information ow in a more
symmetric manner, making post-purchase information more readily available.
Appendix A: Experiment instructions
You are participating in an experiment about individual decision making. Hence, nothing
that you do can affect the other participants, and nothing they do will affect you. In addition,
the setup is not completely identical for all of you (and a lot of it is random), so trying to
learn from others choices does not make much sense.
The experiment consists of 100 identical periods. At the start of each period you will get
an articial object for 50 cents. You will have to sell this object to the computer. Each period,
before the computer makes bids for the object, you will have to announce the minimum
price at which you are willing to sell the object. Then, the computer will make bids for the
object, until there is a bid greater than or equal to the minimum price you have announced.
The selling price will be this last computer bid.
You can think of this setup as a search for a new job. In this analogy, each computer bid is
a wage offer for the same job, and your search criterion denes a minimum wage at which
you are willing to take the job.
There are three things that are very important to know:
1. Every bid costs 2 tokens. Using the wage analogy, these costs can stand for the time you
spend looking for a job. Hence, your costs in a period are two tokens times the number
of bids in that period, plus the 50 tokens that you had to pay to get the object.
INFORMATIONAL ASYMMETRIES AND OBSERVATIONAL LEARNING IN SEARCH 255
2. Your earnings in a period are the price at which you make the sale (the rst bid that is
greater than or equal to the minimum price announced at the beginning of that period),
minus the costs in the same period (i.e. Earning = Sale Price 2 Number of Bids
50). Note that it is possible to have negative earnings in a period. If that happens, the
loss will be deducted from the earnings in other periods.
3. All bids are integers between 1 to 100 tokens (1 and 100 included). Each bid is indepen-
dent of previous bids and of the minimum price you announced. The bids have already
been generated randomly, prior to the experiment, so any decision that you make during
the experiment is not going to affect the bids you will get later on. In page 4 of the
instructions you can take a look at a histogram of 100,000 bids that were generated by
the same process that generated the bids you will get during the experiment. The height
of each bar is the probability of getting the corresponding bid.
At the beginning of each period you will start by having a sample of computer bids
(arranged in a circle shape) which were generated by the same process as the real ones.
These sample bids are independent of the real ones, and are provided only in order to
give you a better understanding of the distribution of bids. There are no tricksif you
feel that you understand the distribution from which the bids are generated, it is totally
sensible to ignore these sample bids. The sample bids are provided only in order to
give you a better feel of the distribution of bids. After you have looked at the sample
bids, you can choose your minimum selling price for that period and bidding will take
place.
At the end of each period, some of you will get the opportunity to observe the next three
computer bids (you get it the same way you received the previous bids). These additional
bids cannot be used for sale. They are provided in order to let you see how more bids would
have looked like in case the minimum selling price you had set would not have made you
stop where you stopped, so bidding would have continued.
The above will go on for 100 periods. Your total earnings will be the sum of the earnings
in each period.
Note that by choosing your minimum selling price you choose both the expected price
you will receive and the expected number of bids you will pay for. Therefore, by choosing
your minimum selling price thoughtfully, you can increase your earnings.
Appendix B: Computer program description
You begin the experiment by having one of the instructors activating the program. This
will make the welcome window disappear, and you will be able to start the experiment by
clicking on the Next Period button, which appears in the right bottomcorner of the screen
(please raise your hand if you do not see this button).
After clicking the Next Period button, the minimum price window (shown below) will
pop up. The circle that appears in the lower part of the window is the sample bids that are
discussed in the instructions. You can then choose the minimum price for that period from
the ten menus in the top of the window. After you made your choice, click on the OK
button, which appears in the top right corner of the window.
256 EINAV
Now bidding will take place, as can be seen in the screen below. You can get new bids by
clicking the Next Bid button (left bottom corner) for each new bid. When bids take place,
they appear on the left side of your screen. There are three columnsthe bid number, the bid,
and the amount you would have earned if the bid was the last one for the period (calculated
by: bid 2 bid number 50). On the right side of your screen, your performance in
past periods is recordedthe period, the minimum price you announced for that period,
the number of bids you got, the sale price and the total earnings you had for the period. At
the top of the screen, you can see also your subject number, what period it is, the minimum
price you announced for the period, and your total and average earnings so far. Once a
period ends, click the Next Period to go on.
INFORMATIONAL ASYMMETRIES AND OBSERVATIONAL LEARNING IN SEARCH 257
After you are done, the last screen will pop up. Fill in your name and social security
number and click the Print and Quit button. We will collect the printouts from the printer,
prepare your payment and call you to pick it up. The receipt is the only place in which your
name and social security number are recorded. Other than that, only the random number
assigned to you is recorded.
If there are no further questions, you are ready to start. The instructors will approach
your desk and activate the program.
Enjoy!
Acknowledgments
This paper is a revised version of chapter 3 of my 2002 Harvard University disserta-
tion. I am particularly grateful to Al Roth for continuous guidance, helpful discussions,
and nancial support of the pilot studies. I thank Drew Fudenberg, David Laibson, an
anonymous referee, and the editor for helpful comments and suggestions. Estelle Cantilon,
Guillermo Caruana, Yair Eilat, Muriel Niederle, Robert Slonim, Manuel Trajtenberg, Leeat
Yariv, Elena Zoido, and seminar participants at Harvard University and at the 1999 Econo-
metrics in Tel-Aviv provided suggestions and comments on earlier drafts. I thank Joep
Sonnemans for sharing with me his experimental data. Financial support by the Russell Sage
Foundation (RSF Project #98-00-07) is gratefully acknowledged. All remaining errors are
my own.
Notes
1. See also the introduction to Erev and Roth (1998) and the references therein.
2. See Stigler (1961) and Kohn and Shavell (1974).
3. See, for example, Schotter and Brounstein (1981), Hey (1982, 1987), Kogut (1990), Cox and Oaxaca (1996),
Moon and Martin (1996), Butler and Loomes (1997), and Sonnemans (1998).
4. The control group is identical to the design of Sonnemans (1998).
5. Kohn and Shavell (1974) and Telser (1973) discuss the implications of extensions to this setup.
6. See the references listed in endnote 3.
7. More formally, suppose a search strategy yields expected payoff of e with payoff variance of v. Now, instead
of running one such search problem, we run N i.i.d identical problems, such that the sum of the expected
payoffs remains the same (so the size of each problem is N
1
times the original problem). Thus, it is easy
to see that the variance of the same strategy in the repeated problem is N(vN
2
) = vN
1
.
8. Although not in the scope of this paper, a possible way to directly test the power of the risk-aversion explanation
is by using binary lottery techniques. These techniques use offers as probabilities of winning a binary lottery,
instead of direct monetary payments. See Roth and Malouf (1979), Berg et al. (1986), and Kagel and Roth
(1995) for more details. As mentioned before, Rabin (2000) suggests that this may not be necessary.
9. Sonnemans (1998) mentions this explanation, and documents that the behavior of subjects after experiencing
regret is different fromtheir behavior in other periods. He does not, however, formally analyzes this hypothesis,
as it is not the focus of his paper.
10. Such ideas are consistent with directional learning theory (Selten and Buchta, 1994), and, to a lesser extent,
with regret matching (Hart and Mas-Colell, 2000).
11. The xed cost per period is not a choice and has no effect on the optimal strategy. It is useful because it
increases the salience of net payoffs and directs subjects towards setting higher reservation prices, speeding
258 EINAV
up learning and convergence. Pilot studies without xed cost have indeed shown that the xed cost is quite
useful for this purpose.
12. Even with full information, one may wonder how well subjects understand the information they are given. In
the experiment described below I try to help subjects understand what the distribution is by showing them a
random sample from it.
13. See, for example, Roth and Erev (1995) and the references they cite.
14. Sonnemans (1998) had 19 subjects who played for 45 periods. They started (period 1) with an average
reservation price of 67.5 and ended (period 45) with an average reservation price of 77.9, which is still below
the optimal reservation price of 81.
15. As emphasized in the previous section, it is important to distinguish between the concept of observational
regret, which is dened above, and the more common use of (anticipated) regret, which enters into the utility
function. Throughout the paper, regret plays a role in enhancing learning, but not through any direct effect on
utility.
16. The complete instructions can be found in Appendix A. Appendix B provides a description of the computer
program, which was also given to the subjects before the beginning of the experiment. These, and a short
pre-experiment exercise, were read aloud to subjects prior to the experiment.
17. The identical draws eliminate some of the noise that can arise fromthe randomdifferences in the draws among
subjects. It is also true, however, that identical draws may cause some bias in the results (if, for example, the
very rst rounds are crucial for learning). With a relatively small sample, the trade off is best resolved by
having identical draws. With large enough sample, however, it would be denitely better to allow different
draws for each individual.
18. The criterion for dropping subjects was very conservative. I drop any subject who chose reservation price
lower than 50, for at least 5 times during the last 40 (out of 100) periods. In general, a reservation price lower
than 50 seems somewhat unreasonable given the xed cost of 50. I look only at the last 40 periods because it
makes sense that subjects may not realize immediately that less than 50 is a bad strategy, but only later on.
It goes without saying that the criterion was chosen without looking at its affect on the results. This criterion
led to dropping as many as 15 subjects (5 (out of 35) from the control group and 10 (out of 40) from the
treatment group). As could be predicted, more subjects got confused (and were dropped) in the somewhat
more confusing condition, namely the treatment group, in which they get to observe post-purchase bids,
which have no effect on the outcome.
19. At the very rst period subjects from either group do not know which condition they are in. Only in the end
of the rst period, once they observe or do not observe post-purchase prices, they realize to which one of the
groups they were randomly assigned.
20. These three additional prices are known to us, because the sequences of bids were previously generated. In
fact, for our purposes we could also generate new i.i.d draws from the same distribution, and think of them as
the three potential future bids.
21. Simpler descriptive statistics also lead to similar conclusions, and are reported in Einav (2002). The qualitative
results are not sensitive to the particular specication or to whether the random subjects are included or not.
Tables 3 and 4 provide results with all subjects included.
22. This is calculated by inverting the logistic transformation. For an estimated coefcient of a, its effect on the
dependent variable, which would otherwise be zero, is given by log(0.5 +a) log(0.5 a).
23. Take the ultimatum game for example, in the action method the responder would only accept or reject the
offer he actually received, while in the strategy method the responder would have to specify a fully contingent
strategy of whether he would have accepted or rejected any potential offer.
24. For example, we can think of two markets for the same product, a regular one and an online one. Alternatively,
one can think of two U.S. states that differ in the law regarding advertisement of tobacco products. Post-
purchase information is likely to be better in states that allow for such advertisements. It must be noted,
however, that a clean test of the prediction would require the two markets to differ in the post-purchase
information, but to be similar in the pre-purchase information, as well as in the mix of consumers who
participate in them. Such examples may be hard to nd in real world, which is one of the reasons that make
experiments so useful.
INFORMATIONAL ASYMMETRIES AND OBSERVATIONAL LEARNING IN SEARCH 259
References
Berg, J. E., L. A. Daley, J. W. Dickhaut, and J. R. OBrien. (1986). Controlling Preferences for Lotteries on Units
of Experimental Exchange, Quarterly Journal of Economics 101, 281306.
Butler, D. and G. Loomes. (1997). Quasi-Rational Search under Incomplete Information: Some Evidence from
Experiments, The Manchester School 65, 127144.
Cox, J. C. and R. L. Oaxaca. (1996). Testing Job Search Models: The Laboratory Approach, Research in Labor
Economics 15, 171207.
Einav, L. (2002). Essays in Industrial Organization, Ph.D. Dissertation, Harvard University, Chapter 3.
Erev, I. and A. E. Roth. (1998). Predicting How People Play Games: Reinforcement Learning in Experimental
Games with Unique, Mixed Strategy Equilibria, American Economic Review 88, 848881.
Goeree, J. K., C. A. Holt, and T. R. Palfrey. (2002). Quantal Response Equilibrium and Overbidding in Private-
Value Auctions, Journal of Economic Theory 104, 247272.
Hart, S. and A. Mas-Colell. (2000). A Simple Adaptive Procedure Leading to Correlated Equilibrium, Econo-
metrica 68, 11271150.
Hey, J. D. (1982). Search for Rules for Search, Journal of Economic Behavior and Organization 3, 6581.
Hey, J. D. (1987). Still Searching, Journal of Economic Behavior and Organization 8, 137144.
Kagel, J. H. and A. E. Roth. (1995). The Handbook of Experimental Economics. Princeton University Press.
Kogut, C. A. (1990). Consumer Search Behavior and Sunk Costs, Journal of Economic Behavior and Organi-
zation 14, 381392.
Kohn, M. G. and S. M. Shavell (1974). The Theory of Search, Journal of Economic Theory 9, 93123.
Moon, P. and A. Martin. (1996). The Search for Consistency in Economic Search, Journal of Economic Behavior
and Organization 29, 311321.
Rabin, M. (2000). Risk Aversion and Expected-Utility Theory: A Calibration Theorem, Econometrica 68,
12811292.
Roth, A. E. and I. Erev. (1995). Learning in Extensive-Form Games: Experimental Data and Simple Dynamic
Models in the Intermediate Term, Games and Economic Behavior 8, 164212.
Roth, A. E. and M. W. Malouf. (1979). Game-Theoretic Models and the Role of Information in Bargaining,
Psychological Review 86, 574594.
Schotter, A. and Y. D. Brounstein. (1981). Economic Search: An Experimental Study, Economic Inquiry 19,
125.
Selten, R. and J. Buchta. (1994). Experimental Sealed Bid First Price Auctions with Directly Observed Bid
Function, Discussion Paper No. B-270, University of Bonn.
Sonnemans, J. (1998). Strategies of Search, Journal of Economic Behavior and Organization 35, 309332.
Stigler, G. J. (1961). The Economics of Information, Journal of Political Economy 69, 213225.
Telser, L. G. (1973). Searching for the Lowest Price, American Economic Review 63, 4049.