Does Regret Explain Why People Search too Little? A Model of Sequential Search with Anticipated Regret and Rejoicing Zhiquan(Darren) Weng March 04, 2009 Abstract We re-examine the sequential search problem where sellers search for the best price from a known distribution. Existing theory predicts the optimal strategy to be a unique constant reservation price. However, experimental evidence to date has found that people generally "search too little" compared to the theoretical benchmarks. We argue the behavioral anomaly is an artifact of the truncated information structure of the sequential search problem that gives rise to asymmetric regret and rejoicing. Embedding the problem within a behavioral model incorporating both anticipated regret and rejoicing based on the regret theory of Loomes & Sugden (1982) and Bell (1982), we …nd: First, people search too little if and only if they are more sensitive to regret than to rejoicing. Second, if we revise the feedback structure of the search problem so that people expect to see what the price would have been had they continued to search (after they stop), search behaviors become observationally indistinguishable from the benchmarks. Moreover, these results are found to hold even when the price distribution becomes unknown. By building a dynamic structure between anticipated regret/rejoicing and experienced regret/rejoicing, our model explains why people sometimes exercise recall. An empirical investigation of 673 separate searches from an experimental dataset con…rms that people’s (latent) reservation prices do shift with regret in the way predicted by the model. Estimation results show that regret about the last search being unsuccessful increases the probability of stopping from 18% to 31% in the current round. Competing explanations for "search too little", such as risk aversion and satis…cing behaviors, are evaluated in a three-way horserace and are rejected in favor of regret/rejoicing. One policy implication of the model is that promoting post-purchase price transparency may induce consumers to search more e¢ ciently. Key Words: Sequential search, Regret theory, Search too little, Recall, Experiment, Latent reservation price estimation. JEL Classi…cation: C91, D03, D11, D12, D83 PhD Candidate. 410 Arps Hall, Department of Economics, The Ohio State University, Columbus, Ohio, USA 43210. Email: weng.27@osu.edu. Phone: (614) 218-2045. I am grateful for Daniel Schunk at the University of Zurich for sharing his experimental data. All errors in the paper are my own. Preliminary Draft: Please do not cite or redistribute without the author’s knowledge. 1 1 Introduction We analyze search behaviors in the simplest possible setting: A seller (for example, a worker) sequentially searches for buyers’ bid prices (for example, o¤ered wages), which are known to be drawn independently from a pre-speci…ed distribution. The cost of each search is …xed at a positive constant c, and moreover, perfect recall for past prices is allowed. The seller can sample in…nitely many prices and there is no discounting. Established theory predicts that the optimal search rule in this environment is for the seller to follow an optimal stopping rule, i.e., to set a reservation price H and to accept the …rst price exceeding H : However, previous experimental studies have consistently found that (1) people search too little. In other words, sellers accept prices that are too low relative to the optimal reservation price derived under risk neutrality. 2 . (2) A sizable portion (usually 10 to 20 percent) of the searches result in the exercises of recall (whereas optimal stopping rule predicts none whatsoever), and (3) Searcher’s eventual accepted prices seem to be path dependent (as if they were uncertain of the price distribution after all.) Two possible explanations, namely, risk aversion and satis…cing behavior, have been advanced in the literature to account for the "search too little" results. However, neither of the two theories has been able to satisfactorily resolve the issue of recall, unless people’s risk attitudes or earning aspirations are assumed to evolve in a systematic way during the search process. The last assertion has not received any supporting evidence. In this paper, we o¤er a third, and more plausible, explanation that is able to explain both too little search and recall at the same time. And it is based on the information/feedback structure unique to the sequential search problem: Namely, regret and rejoicing are only possible when one continues the search; and the only source of regret is about "stopping too late". A …ne distinction should be made between the regret/rejoicing model and a simple learning model: In the regret/rejoicing model, the "search too little" result is derived from anticipation of future regret and rejoicing whereas in a simple learning model, people only look at the success or failure of past searches. In an experimental dataset that contains 673 separate searches, we investigate and compare the abilities of risk aversion, satis…cing behaviors and regret/rejoicing in predicting actual searches in a head-to-head horserace. Risk aversion and satis…cing behaviors are soundly rejected in favor of the regret/rejoicing explanation. 2 See, for instance, Schotter and Brounstein (1981), Hey (1982,1987), Kogut (1990), Cox and Oaxaca (1996), Sonnemans (1998), Einav (2005). 2 1.1 Motivation Bell (1982) and Loomes & Sugden (1982) pioneer the work of regret theory as an alternative paradigm to the expected utility theory in explaining choice under uncertainty. The premise of the regret theory is that once the uncertainty of the world is resolved, decision maker will conscientiously compare the outcome from the chosen alternative and the (counterfactual) outcome had the choice been made di¤erently, and will experience a cognitively based emotion as a result of this comparison. She feels pleased if what she gets is better than what she would have gotten, i.e. she rejoices. she feels displeased if what she gets is worse than what she would have gotten, i. e. she regrets. Thus even if a choice is ex ante optimal, it may not look as good after the fact. The regret theory futher hypothesizes that people have the ability to anticipate these positive or negative emotions so they adjust their choice accordingly prior to the resolution of the uncertainty. In order for anticipated regret and rejoicing to a¤ect choices, people must also hold the expectation that the resolution of the unchosen alternatives will be known. The sequential search problem presents an especially fertile ground to apply the regret theory. First, the information ‡ows in this problem dictates that regret and rejoicing are asymmetric. While no feedback (hence regret or rejoicing) is possible once she stops the search, she cannot escape from feeling either a sense of regret or a sense of rejoicing if she continues to search, because she can always compare the outcome of the latest search to her earning from the last round, which is the counterfactual payo¤ had she chosen not to search. In addition, as the distribution of prices one actually observes in the search process is truncated by her …nal accepted price, her only possible source of regret is based on prices from the unfavorable part of the distribution which translates into reactions such as "Price is bad again; I wish I had not wasted my time and energy in running around and searching!". In other words, one will only regret "stopping too late" but never "stopping too early". Conversely, one’s only source of rejoicing comes from drawing a good price which translates into reactions such as "I am glad that I stuck around for one more round because the price turns out to be really good", but the likelihood and extent of this happening is very limited due to the truncation from top by the accepted price. While largely ignored by the previous economic literature on search, the tension between regret and rejoicing turns out to be a very interesting dynamics that characterizes the actual search that goes on. A separate strand of literature has developed in the …eld of psychology arguing that regret is a more intense emotion than rejoicing. If this argument is true, it implies in sequential search that one might search more conservatively if she is more concerned with minimizing regret than with maximizing rejoicing. Secondly, given the sequential nature of the search task, every price that one encounters during the search has the e¤ect of reminding her whether or not the last search has paid o¤, making the concerns for regret and rejoicing more salient as the search continues and thus amplifying the impact of anticipated regret/rejoicing in subsequent search decisions. The sequential 3 nature of the task combined with the immediacy of the feedback makes it an ideal setting to test if anticipations of regret and rejoicing are driving the behaviors. Moreover, by making people’s sensitivities towards regret and rejoicing endogenous to the experienced regret and rejoicing during the search process through psychological reinforcing, the model can also shed light on why one would rationally exercise recall in certain situations. 1.2 Related Literature Sonnemans (1998) studies what information individuals use to form their search strategies by utilizing the electronic information board technique as well as by asking subjects to explicitly record their strategies. He discovers that subjects sought a combination of information including the last price, the highest price, and the total earning in the search. When explicitly formulating their strategies, 78% of the subjects admitted total earning or number of bids in their stopping criteria. However, these explicitly developed stopping rules, when executed by computer on the same price sequences, resulted in substantially fewer number of recalls than that of actual searches performed by the same subjects. This discrepancy suggests that either some additional stopping criteria adopted in the search are not incorporated into the formulated strategies, or there are dynamic inconsistency in the search behavior. Furthermore, Sonnemans points out the possibility of one-sided learning to be directing subjects to stopping too early, because late stoppers can learn much better than early stoppers. In essence, learning in his construct is similar to the concept of "experienced regret" in the present model. However, unlike in Sonnemans, the "search too little" result here is derived from the interplay of forward-looking anticipatory regret versus rejoicing, rather than the backward-looking experienced regret alone (although we do believe experiencing regret tends to heighten the sensitivity towards future feelings of regret). Finally, using a risk-return pro…le, Sonnemans concludes that 52% of the searches his subjects make cannot be explained by risk aversion alone. Closely related to Sonnemans, Einav (2005) invokes the idea that "observational regret" is driving the search too little result because sellers, if motivated to minimize this backward-looking type of regret, can only learn to adjust their reservation downwards. In addition, number of searches made and search e¢ ciency improve in an experiment treatment in which sellers can observe three more random prices after they complete the search. This is the …rst piece of evidence to our knowledge that shows modifying the information structure in the sequential search environment does mitigate the "too little search" behavior. Yet his idea of observational regret is again backwardlooking and is not embedded in any formal model. Moreover, subjects in Einav’s experiment are not allowed to revise their reservation prices during the search, preventing them from any chance of recall. One theoretical …nding in the present paper is knowing that just one more price, rather than three more, will be revealed following the end of the search is su¢ cient to lead subjects to search in the most e¢ cient manner. 4 An experimental study by Schunk and Winter (forthcoming)3 investigate the relationship between individuals’heterogeneity in risk attitudes and loss aversions and their behaviors in sequential search task. They …nd no connection at all between the risk measures developed independently through the lottery tasks and through the attributed reservation prices in the search task. On the other hand, they …nd that more loss averse (rather than risk averse) individuals tend to search less. They also consider the possibility that individuals in the experiment falsely believe that the search costs are not sunk. Even for these individuals (who try to buy a unit of item at the best price), optimal reservation prices are declining over time. Therefore, none of the search strategies that they consider is able to accommodate the possibility of recalls. In a follow-up paper, Schunk (2008) conjectures that subjects in a sequential search task would constantly update their reference point to equate their current earning in the search and act according to the prospect theory rather than base their decisions on the absolute payo¤ from the entire search, which might explain why individual’s measure of loss aversion is related to the extent of search. It is worth stressing that Schunk’s reference-updating utility speci…cation is very similar to our regret/rejoicing speci…cation in the sense that both utilities are relative to the current earning and that a measure of loss aversion is also very similar to sensitivity towards regret. However, Schunk does not derive any theoretical result that directly compares search under reference-updating preference and search under risk-neutral optimal search while we do under a regret/rejoicing framework. Due to a lack of dynamic structure, his model is also silent on why recall takes place. Methodologically, one recent paper by Filiz and Ozbay (2007) applies the regret theory and similar utility representation of regret to ours to a di¤erent setting and obtains some very interesting results. They investigate both theoretically and experimentally whether the over/under-bidding in …rst-price auctions are related to anticipated regrets stemming from varying information feedback structures: 1) bidders know that they will learn the winning bid if they lose (loser regret condition); 2) bidders know the second highest bid if they win (winner regret condition); 3) or they will receive no feedback regarding the other bids. They …nd that the loser regret condition does generate overbidding while the winner regret condition does not lead to underbidding, relative to the control condition. However, because rejoicing is always absent in the auction setting, they do not deliberately address the issue of the tension between anticipated regret and rejoicing in decision making. The rest of the paper will proceed as follows: Section 2 presents a formal model of the optimal search strategy under anticipatory regret and rejoicing, constructs a counterfactual search where feedback becomes possible even after search is stopped, and extends the results to situation in 3 The empirical analyses contained in the second half of the present paper utilize the same experimental dataset as in the Schunk and Winter study. 5 which price distribution is unknown; Section 3 examines some of the model’s predictions through an experimental dataset consisting 673 separate searches and conducts a number of robust checks; Section 4 discusses the distinction between the regret/rejoicing model and two other competing explanations for the "search too little" result, which are risk aversion and satis…cing behaviors, and administers a three-way horserace comparing the accuracy of the three in predicting the data; Section 5 discusses future directions and policy implications of this research. 2 2.1 The Model The Benchmark Model: Risk Neutral, Optimal Search The benchmark case throughout our analysis is the simplest variant of the sequential search environment for which formal theory predicts a constant reservation price property. We will …rst use the example of a seller searching for the best price: In the benchmark case, a seller (e.g. an unemployed worker) who attempts to sell a unit of good (e.g. her labor) with cost normalized to 0 has to sample bid prices (e.g. wage o¤ers) sequentially. The bid prices are drawn independently from a known distribution with CDF F (x) and P DF f (x) over interval (a; b). Each price sampled costs the seller a constant amount, c: The search process could last in…nitely and there is no discounting. Moreover, seller has the ability to recall bid prices she encountered previously. The capacity to recall is consistent with, for example, the ease with which people can return to past prices in an online shopping session. The benchmark case is simple enough yet retains the essence of the relationship between search and information structure that we are interested in. In the benchmark case, the seller is risk-neutral therefore she aims to maximize the monetary payo¤ from the search. The preceding search setting presents a dynamic programming problem where the only state variables are the number of searches that have been conducted, t, and the highest of those t encountered prices, which we call the candidate price and denote as Ht : Let V (Ht ; t) represent the value function, with the choice variable being a binary decision between stopping or continuing the search. (Ht ; t) represents the payo¤ from choosing to stop now. V (Ht ; t) = M ax { (Ht ; t); E[V (Ht+1 ; t + 1)jHt ]g (Ht ; t) = Ht c t (1) (2) Lippman & McCall (1976) show that the solution to the above in…nite horizon problem is identical to solving a one-period problem myopically, because the option value of searching in the next round will not be higher than in the current round. Disregarding the sunk search costs that have been incurred prior to the current decision results in the stationarity of the problem. As a result, from this point on we suppress the time dimension of the problem. To a seller, the payo¤ from choosing to stop now is: 6 Stop (H) =H (3) If she continues the search, her payo¤ when the new bid price equals x is: Search (x; H) = M axfH; xg c (4) Therefore, the expected payo¤ from continuing the search is: Search (H) = Z b Search (x; H)dF (x) a = [F (H)H + Z b = H + [ (x Z b xdF (x)] c H)dF (x) c] H (5) H Equation (5) shows that the decision of when to stop the search can be determined by comparing Rb the marginal bene…t ( H (x H)dF (x)) and marginal cost (c) from one more search given the candidate price H. Thus, de…ne a di¤erence function G(H) Search (H) Stop (H) Z b (x H)dF (x) c (6) H It is easy to verify that dG(H)=dH < 0: Assuming all the end conditions (i.e. G(a) > 0; G(b) < 0) are met, the optimal search strategy for a risk neutral seller is therefore of reservation price type. Speci…cally, she should set a reservation price equal to HRN ; and stop search and accept the …rst price exceeding HRN ; where HRN is de…ned implicitly by G(HRN ) = 0: In other words, Z b HRN (x HRN )dF (x) = c (7) Except for a few well-speci…ed density functions, the reservation price HRN will not have a closedform expression. HRN is also a function of the search cost c and the price distribution F (:): Two interesting comparative statics can be readily derived from taking the total derivatives of Equation (7). @HRN < 0 @c HRN (F1 ) > HRN (F2 ); if F1 (:) …rst-order dominates F2 (:) 7 (8) (9) If we do not restrict the seller’s risk attitudes, the preceding problem can still be solved in a very similar manner: The only di¤erence is that a properly de…ned utility function U (:) will replace all of the monetary reward/cost terms in Equation (3) and (5). The solution as in (10) still possesses the constant reservation price property (subscript RA is short for risk averse subjects). U (HRA ) = F (HRA )U (HRA c) + Z b HRA U (x HRA c)dF (x) (10) One important implication of the optimal stopping rules outlined above is that a seller conducting a sequential search should NEVER exercise the recall option. This assertion holds so long as the seller has a constant reservation price throughout the search, regardless of the value of the reservation price itself. But the previous experiments have revealed that subjects in the lab did frequently recall past prices, ranging anywhere from 10 to 20 percent of the total searches. This issue remains puzzling and unsatisfactory even if we take into account the full spectrum of subject’s risk attitudes in the search process such as in equation (10) 2.2 The Behavioral Model: Sequential Search with Anticipated Regret and rejoicing In the ensuing sections, we model the search process of a subject that while staying risk neutral, feels either a sense of regret or a sense of rejoicing after a search is conducted and a price is revealed. Note that due to the asymmetrical information structure of the sequential search problem, one only regrets about wasting the search cost when an unfavorable price is revealed, and one only rejoices at the discovery of a favorable price. There is neither regret about missing out a good price nor rejoicing at not wasting the search cost. First, we will make explicit our regret/rejoicing utility representation. 2.2.1 Regret/Rejoicing Utility Speci…cation Bell (1982) and Loomes & Sugden (1982) pioneered "Regret/Rejoicing Theory" as an alternative paradigm to the expected utility theory in explaining choice under uncertainty. Using an axiomatization approach, they proved that a utility function incorporating regret and rejoicing should be expressed in the following general form: v(x; y) = u(x) + f (u(x) u(y)); for some function f (11) where x is the choice being made and y is the (only) alternative, u(:) is comprehended as the standard utility function without concerns for regret/rejoicing. Therefore, f (u(x) u(y)) expresses the additional utility arising from regret or rejoicing. 8 Following Filiz and Ozbay (2007) in which they adopt the same utility representation to study the relationship between anticipated regret and over/under-bidding in …rst-price auctions, we decide to modify a subject’s utility function by incorporating a more explicit functional form for the regret/rejoicing component of (11): VSearch (x; H) = Search (x; H) where R(y) = y if y > 0, + R( Search (x; H) Stop (x; H)) (12) called the intensity of rejoicing = 0 if y = 0 = y if y < 0; Both ; called the intensity of regret are positive. Our regret/rejoicing utility formulation (labelled V to distinguish it from a standard utility) is comprised of two additive elements: The …rst element, a "choiceless" utility, is equal to the monetary payo¤ from the choice outcome, Search (x; H)4 . "Choiceless" is in the sense that the outcome is evaluated as if it were obtained without an internally processed decision, such as getting $100 as a result of an income tax credit. The second element, R( Search (x; H) Stop (x; H)), represents the subjective utility from the feelings of either regret or rejoicing (R(:) stands for the regret/rejoicing function). This second element of the utility function depends on the decision maker’s available choice set and the resolved state of nature. We further simplify the functional form by assuming it to be linear in the di¤erence in payo¤s between the chosen and forgone alternatives in every state of the world. But the slopes, which we call the intensities for regret and rejoicing, are set di¤erently for gains (i.e. rejoicing) and losses (i.e. regret). The di¤erentiation in slopes for regret and for rejoicing not only allows more ‡exibility but also is more realistic in light of the sizable psychology literature arguing that regret generates more cognitive and emotional arousal than rejoicing does. It also makes good sense to make R(0) = 0 when the chosen and forgone alternatives fair equally. Additionally, in our model the regret and rejoicing intensities are not stationary. In fact, they are reinforced by the seller’s experienced regret and rejoicing at each search round as a result of the past search outcomes. Hence they are functions of the price history in the current search, i.e. t = (p1; p2 ; :::; pt ); t = (p1; p2 ; ::::; pt ) (The exact structure of the relationship will be made more clear in later sections). This dynamic relationship expresses the notion that the more one experiences regret or rejoicing in the past, the more sensitive (or mindful) she is to future regret and rejoicing. The history dependence of (:); (:) o¤ers plausible insights into why sellers might exercise recall and why their behaviors appear to be path-dependent, even though in fact the sellers in our model still adhere to a reservation price strategy. 4 Risk neutrality is maintained throughout the discussion surrounding the regret/rejoicing utilities in order to facilitate comparison with the benchmark case. 9 Before the seller decides to make the investment to search, she is assumed to be able to anticipate her feelings of regret and rejoicing based on the realized outcomes in every possible state of nature. So her ex ante utility from continuing the search given the current candidate price H is: VSearch (H) = Z b VSearch (x; H)f (x)dx a = Z H [H c + ( c)]f (x)dx + a + Z Z H+c [x c + (x c H)]f (x)dx H b [x c + (x c H)]f (x)dx (13) H+c Z b H)dF (x) c] [expected monetary payo¤] Z H+c [cF (H) + (c + H x)f (x)dx] [anticipated regret] H Z b + (x c H)f (x)dx [anticipated rejoicing] = [H + (x H (14) H+c Equation (14) shows that, VSearch (H) consists of three separate terms: the term in the …rst square bracket measures the expected monetary payo¤ from the search and thus is identical to Search (H) in the benchmark model. The second term captures the e¤ect of anticipated regret, The third term captures the e¤ect of anticipated rejoicing. Generally speaking, anticipated regret makes search less attractive while anticipated rejoicing makes the search more attractive; therefore, a priori it is not obvious how the interplay between regret and rejoicing would a¤ect the seller’s search strategy. On the other hand, the expected utility of stopping the search remains at H. This is because stopping the search will preclude any information feedback as to what the price would have been, implying R(:) will not be operative here. Hence, VStop (H) = H (15) De…ne L(H) to be equal to the di¤erence between the expected utilities from continuing the search and from stopping the search given H: L(H) VSearch (H) VStop (H) Z b Z H+c = [ (x H)dF (x) c] [cF (H) + (c + H H H Z b + (x c H)f (x)dx H+c 10 x)f (x)dx] (16) Therefore, the seller’s decision rule is to continue to search () L(H) 2.2.2 0 Optimal Search Rule under Anticipated Regret/Rejoicing It is easy to see that in Equation (16) L0 (H) < 0:This is intuitive because not only does the monetary return to search decreases with H; but also the likelihood and the extent of regret increase with H while those of rejoicing decrease with H: It implies that a seller, in anticipation of regret and rejoicing, should once again adhere to a unique reservation price strategy with the reservation price HRR (c; F; ; ) which solves L(HRR (c; F; ; ) ) = 0: We can further simplify the three components in L(H): 1) Anticipated Regret: cF (H) Z Z H+c (c + H x)f (x)dx = H 2) Anticipated Rejoicing: Z b (x H+c F (x)dx (17) H c H)f (x)dx = (b c Z H) H+c b F (x)dx (18) H+c In aggregate, we now have Z b L(H) = [ (x H)dF (x) H (c + H) + b c] Z b Z H+c F (x)dx H F (x)dx (19) H+c Ultimately our goal is to diagnose whether and how concerns for regret and rejoicing will drive the seller to revise her search strategy relative to the risk-neutral benchmark case. Alternatively, the benchmark case can be viewed as a special case under the regret/rejoicing model where = = 0: Because it is impossible to derive the precise explicit solutions for HRN and HRR ; a direct comparison between the two cuto¤ points is infeasible. To circumvent this problem, an alternative strategy is devised in which we examine the value of L(H) evaluated at HRN : To that end, we …rst utilize the property that HRN is the implicit solution to the Equation (7). Hence, Z H +c Z b RN L(HRN ) = G(HRN ) F (x)dx (c + HRN ) + b F (x)dx = Z HRN HRN +c HRN F (x)dx (c + HRN ) + b 11 Z HRN +c b HRN +c F (x)dx (20) Further simplifying Equation (20) (detailed derivation contained in the Appendix), we obtain the following crucial equation: L(HRN ) = ( ) Z HRN +c F (x)dx (21) HRN Equation (21) shows that if > > 0, then L(HRN ) < 0: Then it must be true that a < HRR < HRN because L(a) > 0; L0 < 0: On the other hand, if 0 < < ; then it must be true that HRN < HRR < b: If = ; then HRR = HRN : A graphic summary of the results is in Figure 1 in which H1 =.HRR and H0 = HRN :Therefore the following proposition holds. Figure 1 A Comparison Between Reservation Prices Proposition 1: When both anticipated regret and anticipated rejoicing are present in the search process, the seller continues to adopt a reservation price strategy. Nevertheless, the value of the reservation price will be lower than that in the benchmark model if and only if the intensity of regret is stronger than the intensity of rejoicing, i.e. > ; 12 resulting in a higher hazard rate of stopping the search as well as lower expected number of searches than the benchmark case. Conversely, the seller will search more than the benchmark case if and only if the intensity of rejoicing is stronger than the intensity of regret, i.e. < . When the two intensities are equal; a seller will search in the exact same way as a risk-neutral seller without any concerns for regret and rejoicing. The preceding proposition makes it clear that anticipatory concerns for regret would indeed lead to "too little search", provided that it is stronger than the counteracting force of rejoicing. Next, we derive some comparative statics of how the reservation prices under the regret/rejoicing framework would respond to the changes in the underlying parameters c; ; : Setting L(HRR (c; F; ; )) = 0 as in Equation (19) and take the total derivatives with respect to the parameters of interest will give: @HRR (c; F; ; ) @c @HRR (c; F; ; ) @ @HRR (c; F; ; ) @ < 0; (22) < 0; (23) > 0 (24) Note if a purchasing framework is considered wherein a buyer is trying to shop at the lowest price, all the comparative statics are exactly reversed. Therefore, under a regret/rejoicing model, a seller’s reservation price will decrease if her regret coe¢ cient gets larger, and vice versa for rejoicing. Put di¤erently, a seller will search for prices less aggressively en ante if she becomes more sensitive to the feeling of regret ex post, other things being equal. A further rami…cation of the statement is that, because as we have hypothesized, the values of the regret and rejoicing intensities (:); (:) are subject to the reinforcement from the experienced regret and rejoicing based on the history of realized prices in the search, the reservation prices HRR will also be dependent on past prices. For example, a consecutive string of bad prices might cause the seller to be more sensitive to regret and therefore lower her reservation prices over time, possibly causing her to recall a past price. This o¤ers one plausible explanation why actual search behavior in the lab will appear history-dependent and why recall is often opted for, even though the distribution is known. Thus a complete behavioral model encompassing anticipatory regret and rejoicing can coherently explain the three empirical anomalies that the benchmark model cannot not explain. 13 2.2.3 Counterfactual Search with Availability of Feedback even when Search is Stopped In most of the real life search settings including the one we have just modelled, stopping the search will preclude any information feedback. In other words, a seller never gets to or expects to observe what the next price would have been. Consequently, regret or rejoicing only occurs when one continues the search. Hence, if regret is the more prominent force here than rejoicing (e.g. let = ! 1), stopping the search becomes the preferred, regret-minimizing choice. Nevertheless, there are certain situations in life where you may …nd out what price you have missed out after you stop the search. Suppose that you are shopping for a new Sony Playstation at a couple of local electronic stores for your son’s birthday this coming Wednesday. You happen to also have a plan to go to the local Walmart to run some errands this Saturday so there is a chance that you may see Walmart carry the same game console at a lower price. How does this piece of information a¤ect your search for the console in the …rst place? In fact, you might decide to check out the price at Walmart …rst before you buy it at any other store such as Bestbuy. Or you may decide to purposefully avoid looking up the price of the console at Walmart on your weekend shopping trip. Can this kind of behavior be predicted by our regret/rejoicing search model? In this section, we simulate the preceding situation by allowing information feedback even if the search is stopped. More concretely, in the new "counterfactual" (not all that "counterfactual" considering the real-life scenario that we just presented) search environment, the seller expects to observe one more price quote even after she stops the search and sells the item. Given this new information structure, the seller will have the opportunity to experience another kind of regret: This happens following the end of the search when the next price exceeds the current price by more than the search cost. Correspondingly, there is also the opportunity of experiencing another kind of rejoicing when the price revealed after the search turns out to be unfavorable. As opposed to the original search problem, the expected information feedback on the Stop the Search option will augment its expected utility to account for the additional elements of regret and rejoicing: VStop (H) = H Z b (x H c)f (x)dx + cF (H) + H+c Z H+c (c + H x)f (x)dx (25) H The expected utility of the Continue the Search option, however, is unchanged. Clearly, VStop (H) H increases with the value of H because of diminished regret and heightened rejoicing. Similarly, VSearch (H) H decreases with the value of H: De…ne a di¤erence function M (H) VSearch (H) VStop (H): We have M 0 (H) < 0: Simplifying the expression for M (H), we have: 14 M (H) VSearch (H) VStop (H) Z b Z H+c (x H)dF (x) c ( + )[cF (H) + (c + H H H Z b (x c H)f (x)dx +( + ) x)f (x)dx] (26) H+c Note that the functional form of M (H) in Equation (26) is identical to that of L(H) in Equation (16), except that in M (H) the intensity of regret and the intensity of rejoicing are identically equal to the sum of and : According to Proposition 1, whenever the intensities of regret and rejoicing are equal, reservation price under regret/rejoicing is identical to that of a risk-neutral optimal search. Therefore, the search outcomes in the search setting with two additional elements: 1) anticipation of regret and rejoicing 2) counterfactual information feedback after search is stopped become behaviorally indistinguishable from those of the benchmark model where both are absent. This forms the basis of our second proposition: PROPOSITION 2: If, after stopping the search and selling an item, a seller with anticipated regret/rejoicing freely observes one more price quote, then her behaviors in the search will be observationally equivalent to those in the benchmark case. In other words, HRN is the chosen reservation price in this hypothetical search problem, no matter what the intensities of regret and rejoicing are. The implication is, behaviors in this counterfactual search environment should be path independent, and recall should never happen. 2.2.4 Extension to Search under Unknown Distribution We …rst examine search behaviors in which the underlying distribution is completely known because theorists and to a larger extent experimentalists wish to avoid the additional complexity based upon people’s heterogeneity in their ability in forming a prior regarding the price distribution and in updating the distribution by processing new price information. However, this simpli…cation also makes the model possess much less external validity. The seminar paper in this area is Rothschild (1974) where he used a simple Bayesian updating process on a …nite set of possible prices to model search from unknown distribution. In this section, we would try to make a very limited attempt 15 in shedding some lights on search from an unknown distribution under the anticipation of regret and rejoicing. Our direct concern, once again, is how the search behaviors with regret and rejoicing compare with the behaviors in absence of these elements. Will people also search too little? The answer seems to be yes. Our overarching strategy here is to approximate a Bayesian updating process from an unknown distribution by a Bayesian updating process between the probabilities of making draws from n known distributions. In other words, the problem is turned into one in which the seller is sure that the price comes from one of the n known distributions, but in the meantime she is unsure as to which one it is. Therefore with each new price revelation she updates her belief about which one of the n distributions she is choosing from. Our sense is that this approximation is always possible given the appropriate choices of the number of distributions n, the cumulative probability functions of each of the distributions F1 (x); F2 (x); :::; Fn (x) (with Fi (x) de…ned on [ai; bi ]; respectively) and the seller’s initial belief 0 but we still need to locate the precise statistical theory. The process of a seller sequentially searching for the highest prices from an unknown distribution is equivalent to the process of her searching from n known distributions with the probabilities of each distribution being represented by her belief t ; a n-element simplex, at each decision period P t; t = (p1t; p2t; p3t;::::; pnt ):& i pit 1: t is updated according to the Bayes rule after each new price observation xt 1 :We’ll …rst derive the results for the benchmark case where the concerns for regret and rejoicing are not present. 2.2.5 Search from an Unknown Distribution with No Anticipated Regret/Rejoicing First, we would customarily use subscript t to denote the decision and information variables at decision period t, because now the problem is no longer time invariant, due to the fact that each realized price on the path would change the perceived price distribution from which the search is made. Nonetheless, it is clear that the return to no search remains is still the highest standing price whether or not regret/rejoicing is part of the utility function. t VStop = t Stop = Ht Here Ht denotes the current candidate price after t price draws. The expected payo¤ from search will be a probabilistically weighted sum of the expected payo¤ from conducting one more search from each of the possible distribution as in the following expression: 16 t Search n X = i=1 n X = pit [(Ht c)Fi (Ht ) + Z bi (x c)dFi (x)] Ht pit [(Ht c) + Z bi (x Ht )dFi (x)] Ht i=1 = Ht + n X pit i=1 Z bi (x Ht )dFi (x) c (27) Ht As before, the value of conducting one more search consists of the option value of R bt ; and Precall H the cost of search c, and the expected net return to search in probabilistic form, ni=1 pit Hit (x Ht )dFi (x): R bi Pn Ht )dFi (x) c: Apparently Gt (Ht ) still Let us set Gt (Ht ) = search stop = i=1 pit Ht (x retains the property dGt (Ht )=dHt < 0: Therefore, the solution to the seller in search of a best price will still be of the optimal stopping form. The major di¤erence between the known and unknown distribution cases, notably, is that now the reservation price is path dependent due to the fact that all the (p1t; p2t; p3t;::::; pnt )0 s are subject to Bayesian updating. Hence, the seller should sell at search period t whenever the revealed price exceeds HRN;t ; and HRN;t is the solution to the following equation: 0= n X pit Z bi (x HRN;t i=1 HRN;t )dFi (x) c (28) where Pit is subject to the following Bayesian updating process, Pit = Pi(t n P 1) fi (xt 1 ) Pj(t ; with 0 given. (29) 1) fj (xt 1 ) j=1 The optimal stopping rule and the beliefs of the seller should be su¢ cient to completely pin down her search behaviors in this no regret/rejoicing problem. Moreover, since the reservation price at any search period HRN;t is just the weighted average of the reservation prices the seller would have adopted if she knew for sure that she was searching from one of the n distributions weighted by the relative probabilities of the n distributions, then all the …rst-order comparative statics are also retained as in the search from known distribution case. 17 2.2.6 Search from an Unknown Distribution with Anticipated Regret/Rejoicing In essence, the value of search from an unknown distribution with the incorporation of anticipatory regret and rejoicing is just a probabilistic version of the one we had before for the case when the distribution is known. To see this, …rst, because the feedback on the no search option is still absent, t VStop = Ht: In contrast, the expected value of search should encompass anticipation of regret and rejoicing experienced on searching from each of the n possible distributions: t VSearch n X = i=1 Z pit [Ht + bi (x Ht )dFi (x) c x)dFi (x) + Z Ht Z Ht +c (c + Ht cFi (Ht ) bi (x c Ht )dFi (x)] (30) Ht +c Ht Simplifying the preceding expression, we have t VSearch = Ht + n X pit i=1 n X pit [ Z bi (x Ht )dFi (x) c+ Ht Z Z Ht +c Fi (x)dx + (bi c Ht ) Fi (x)dx] (31) Ht +c Ht i=1 bi t t VStop Let HRR;t satis…es Lt (HRR;t ) 0: Again, since HRR;t is a As before, set Lt (Ht ) = VSearch weighted average of the reservation wages that a regret/rejoicing driven seller would have chosen if she was searching from just one of the n possible distributions, all the …rst-order comparative statics would still hold. To compare the values of Ht0 and Ht1 ; we again calculate the value of Lt (Ht ) evaluated at HRN;t : Lt (HtRN ) = 0+ n X pit [ i=1 = ( ) n X Z = ( ) i=1 Fi (x)dx + (bi Ht0 pit i=1 n X 0 +c Htt pit Z 0 +c Htt Fi (x)dx + Ht0 Z HtRN < ( )0 if and only if n X i=1 HtRN +c Fi (x)dx >( ) 18 c Z Ht0 ) pit (bi bi Ht0 Ht0 Z Fi (x)dx + Z Ht0 +c Ht0 Fi (x)dx] bi Ht0 Fi (x)dx c) (32) The preceding result says that the reservation prices will be lower and search is less than the benchmark case if and only if the regret intensity, in anticipatory terms, is stronger than the intensity of rejoicing. Proposition 3 summarizes the major results in this section. PROPOSITION 3: When a seller is searching for the best price from an unknown distribution, at every period she still adopts a reservation price type strategy: stop search if and only if Ht HRN;t (or HRR;t if regret/rejoicing are present), but her reservation price is going to be dependent on the price history Hto = Hto (x1; x2;:::; xt 1 ) [analogously for Ht1 ] due to Bayesian updating. Nevertheless, on the same realized price path, a seller with anticipated regret & rejoicing will search less than someone without them if and only if she feels regret more strongly than rejoicing. Therefore, the expected number of searches will be fewer for a seller with anticipated regret/rejoicing if and only if > . 2.3 Competing Explanations for Why People Search Too Little Previously, researchers puzzled by the observed "search too little" results have o¤ered two chief explanations: 1) human subjects may display risk aversion on monetary payo¤s, 2) human subjects may not be able to make the precise optimization calculation (i.e. fully rationality is costly), and therefore they will satis…ce at reaching a vicinity of the optimum. 2.3.1 Risk Aversion: Risk aversion is a natural …rst instinct in explaining why people accepts reservation prices that are too low. Essentially, conducting one more search is like taking a gamble, waging the cost of search in order to win potential higher price than the current price. If people are risk averse, on the margin the uncertain prospect of higher earning looks less enticing than the certain loss in search cost. Presumably the larger the search cost, the more people will shave their reservation price. For example, in the empirical investigation section to follow, we use a sequential search dataset that asks subjects to search for the lowest possible price to purchase a particular item that they value at $500. The price distribution is a discrete normal distribution with mean=$500 and standard deviation=$10 (truncated at end points $460 and $540). If we also assume a CARA utility function for the subject on the monetary outcomes, the following graph depicts the relationship between a subject’s risk coe¢ cient and her constant reservation price. 5 = 0 corresponds to risk neutrality. Apparently, higher does lead to higher reservation price and more conservative search here. 5 CARA utility of the following form is used: X Xmin u(x) = 1 (1 exp( )) Xmax Xmin 19 Figure 2: 2.3.2 Satis…cing Behaviors The other concern is: People may not be able to calculate the optimal reservation price accurately due to limited cognitive processing abilities. But is there a reason why they tend to search less rather than more? For the same distribution as above, we calculate the expected payo¤ from adopting reservation prices of di¤erent values and translate them into a percentage term of the maximum expected payo¤ (obtained at R=490 when = 0) in the Figure 3: 20 Figure 3: Expected Payo¤s from Various Constant Reservation Strategies as % of the Maximum Expected Payoffs (R=480:R=499) 10 49 8 49 6 49 4 49 2 49 0 48 8 48 6 48 4 -5 48 2 0 48 0 Expected Payoff 5 -10 -15 -20 -25 Reservation Price The expected values are calculated with the following equation: E (L) = $500 Ep (pjp L) c F (L) (33) In this purchasing task, because the e¢ ciency of choosing reservation prices higher than the optimum is relatively high compared to that of choosing reservation prices lower than the optimum, a satis…cing seller on average will choose to search too little. 3 Empirical Investigation The most crucial prediction from the behavioral model in the preceding section is that, once we model for a search agent’s auxiliary utilities related to her ex ante anticipation of regret and rejoicing which are in turn based on the ex post evaluation of her decision making abilities, she has a tendency to search more conservatively than someone who is both risk-neutral and unmotivated by these behavioral concerns, provided that the intensity of regret in the behavioral utility model we de…ned is larger than the intensity of rejoicing . Although the ability of anticipated regret/rejoicing to explain the "search too little" result due to the asymmetric information structure of the sequential search problem is a theoretical novelty in its own merit, what is more important 21 is to show these concerns do underpin the actual search behaviors. If anticipation of regret and rejoicing is in fact recognized to play a role in the actual search process, does it have a more prominent role than other competing factors, such as risk attitudes and satis…cing behaviors? In order to answer these questions, we borrowed an experimental dataset that contains 673 individual search sequences with random price o¤ers.6 One key testable proposition we have from the model is that search agent’s reference price shifts according to the intensities of regret and rejoicing. Furthermore, we will make one additional behavioral assumption that people’s sensitivities to future regret or rejoicing are reinforced by their past experiences of regret or rejoicing which could happen within the same search sequence. Therefore, our empirical methodology is focused on estimating the latent reservation prices at a particular point of the search as a function of experienced regret and rejoicing up to that point during the same search. Moreover, the estimation results also make a convincing case why anticipation of regret and rejoicing matters more than individual’s risk attitudes or satis…cing behaviors. Finally, it is worth underscoring that our theoretical model is based on forward-looking regret and rejoicing which enables us to make the connection to backward-looking regret and rejoicing. Hence, our explanations are fundamentally di¤erent from a simple learning model such as a directional learning model. 3.1 Data Description The dataset we examine is from a search experiment conducted at the University of Mannheim in the fall 2003.7 A total of 64 subjects participated in the experiment divided into two sessions, with each of them completing either 10 or 11 separate sequential search tasks. The search setting is identical to our model except that it is phrased in terms of purchasing at the lowest price rather than selling at the highest price: Each subject can sequentially search, for an unlimited number of times, prices from a known distribution in order to purchase an item that they value at $500. There is no discounting, and the ability to recall past prices. Subjects are informed that the prices will be drawn randomly from a discrete truncated normal distribution with mean $500 and standard deviation $10, truncated at $460 and $540. Each price that they sample costs $1. In addition, the subjects are permitted to practice the search task for an unlimited number of times during a certain time period before going into the actual paid tasks. They are paid for their earning in one of the 10 or 11 search tasks One distinguishing feature of this dataset is that it is one of the few in which the price sequences are randomized across tasks and across individuals, which gives us more estimation power. In total, there are 673 individual searches, and 3,408 stop-and-go decisions in the dataset. Each stop-and-go decision will serve as one observation in our estimation. 6 We thank Prof. Daniel Schunk from University of Zurich for generously o¤ering to make his experimental data available to our use. 7 See Schunk & Winter (2005) for more details regarding the dataset. 22 But …rst, we will o¤er some descriptive analysis related to the question: Do people search too little (yet again!) in our dataset? 3.1.1 Do subjects search too little? In order to see if subjects in the experiment follow optimal search reasonably well or otherwise, Figure 4 depicts the distribution of actual number of searches in bars against the density of number of searches if an optimal rule is followed. In the latter case, the optimal stopping price will be 491, and the number of searches would exhibit a geometrically declining pattern. In 134 of the 673 tasks, buyers accept the …rst price that they obtained and in 95 of them they searched for just two prices. In contrast, on a few occasions, buyers make as many as 30 or 40 searches, which are clearly unpro…table. The di¤erence between actual and optimal searches are apparent from Figure 4: exhibited search behaviors are concentrated more on shorter searches (t < 7) and less on longer searches (t 7) than what is optimal. Similarly, under optimal risk-neutral assumption, the expected number of prices searched would be 6.26. The actual number of prices searched is 5.06 on average. So here buyers search for roughly one fewer price. The expected payo¤ from each search under optimality is 8.53. The actual average payo¤ is 5.71, an average e¢ ciency loss of 33%. Figure 4 23 It is also interesting to look for any pattern of convergence or learning among subjects over time. Figure 5 presents the average and median number of price searches across the 64 subjects over number of tasks completed. Neither evidence of convergence nor of learning is discernible from the picture. The number of searches is persistently lower than the optimal level up to the …nal task. Median number of search is even lower due to a few outliers who go through lengthy and costly searches. E¢ ciency of actual search does not improve with experiences either. It goes from 58% in Task 1 to 77% in Task 11 with no consistent upward trend (Figure 6). Figure 5 24 Figure 6 3.2 Estimation In our behavioral model, agents in a sequential search task optimizes based on a utility function that incorporates both monetary payo¤ and the anticipation of the psychological a¤ect of regret and rejoicing. Consequently, a search agent’s reservation price is in‡uenced by her own sensitivities to regret and rejoicing in a manner predicted by the model. That is, she would search less aggressively if she tends to worry more about not obtaining a good price and therefore wasting her search costs. Conversely, anticipated rejoicing would motivate her to search more aggressively. In the last section we also propose that the intensity coe¢ cients t and t are not time invariant. Instead, they are functions of the history of prices that the searcher has encountered in the current search period. t = (p1; p2 ; :::; pt ); t = (p1; p2 ; ::::; pt ) The logic here is simple: past regrets from observing unfavorable prices would reinforce current regret intensity t ; and past rejoicing from observing favorable prices would reinforce current rejoicing intensity t . In turn, reservation price becomes a function of past experiences of regret and rejoicing through t ; t . This statement forms our main testable implication from the model. This also separates a regret/rejoicing driven search from a risk-aversion driven search: Although both are able to predict less search than that in a risk-neutral model, reservation price is supposedly held constant throughout the search process for a risk-averse searcher (it’s hard to imagine one’s risk attitude changes as the search unfolds) whereas a subject anticipating regret and rejoicing holds a shifting reservation price based on the realization of prices. In addition, because the distribution 25 of observed prices is truncated by the …nal accepted price, searchers are more likely to come across unfavorable prices during the search than favorable prices, resulting in higher frequency of regrets than rejoicing. In response, reservation prices will often change in a way that makes recall a tenable option as the search goes on. 3.2.1 De…ning Measures of Experienced Regret/Rejoicing In order to test whether the reservation prices move with experienced regret and rejoicing, we’ll operationalize the regret/rejoicing reinforcing functions t = (p1; p2 ; :::; pt ); t = (p1; p2 ; ::::; pt ) in the following way. First, we de…ne two possible types of backward regret. In a framework where a buyer is trying to purchase a good value at V at the lowest possible price by comparison shopping, let LT represents the lowest price of all the ones the buyer has obtained up to the current decision period T ; i.e. Lt = min fp1; p2 ; :::; pt g: If the buyer stops search now, her payo¤ would be T = V LT T c: Her counterfactual payo¤, had she chosen to stop search last period, would have been T 1 = V LT 1 (T 1) c: If T 1 > T ; she would feel a sense of regret about continuing the search into this round. Note this preceding condition is equivalent to LT 1 LT < c; indicating whether the last round of search has paid o¤ or not. This forms our …rst measure of regret, which is called one-step regret. It is either 1 or 0 in our dataset due to the fact that the search cost c is equal to 1. One-Step Regret: reg_1_stepT = M axf T ; 0g T 1 = M ax f(V = M axfLT LT LT (T 1 1 (34) 1) c) (V LT T c) , 0g + c; 0g (35) (36) Analogously, to evaluate the merit of her decisions, a buyer can also compare her current payo¤ against the highest (counterfactual) payo¤ she could have earned, by stopping earlier at any of the previous rounds from t = 1 to T 1. This counterfactual payo¤ can be computed as max V Lt t c. So if the current payo¤ is lower than this counterfactual payo¤, the 1 t T 1 buyer also feels a sense of regret about continuing the search into the current round. We will call this multi-step regret. The distinction between one-step regret and multi-step regret helps us learn against which reference outcome the subject evaluates her payo¤ in the decision making process. Multi-Step Regret: reg_t_stepT = M axf M ax (V 1 t T 1 Lt 26 t c) (V LT T c) ; 0g (37) Two measures of backward rejoicing can similarly be developed. They are positive whenever current payo¤ is favored over the counterfactual payo¤s which corresponds to a sense of rejoicing over the decision to search to the current round. One-Step Rejoicing: rej_1_stepT = M ax { (V = M axfLT LT 1 T LT c) [V LT (T 1 1) c], 0g c; 0g (38) Multi-Step Rejoicing rej_t_stepT = M axf(V LT T c) M ax (V 1 t T 1 Lt t c) ; 0g (39) According to how the terms are de…ned, the following properties can easily be veri…ed to hold amongst there four measures: reg_i_step > 0 ) rej_i_step = 0, i = 1 or t rej_i_step > 0 ) reg_i_step = 0; i = 1 or t (40) rej_t_step > 0 ) rej_1_step (41) reg_1_step > 0 ) reg_t_step 3.2.2 reg_1_step > 0 rej_t_step > 0 Postulated Relationship between Reservation Price and Regret/Rejoicing Measures With all the backward regret and rejoicing measures, the intensity functions can be expressed more concretely as functions of these measures. Moreover, it’s natural to assume that all the partial derivatives here are positive: t = (reg_1_stept ; reg_t_stept ); and @ t =@reg1t > 0; @ t =@regtt > 0 (42) t = (rej_1_stept ; rej_t_stept ); and @ t=@rej1t > 0; @ t =@rejtt > 0 (43) Moreover, inserting Equation (42) and Equation (43) into the reservation price function LRR;t = LRR (c; F; t ; t ) (in a purchasing context), we discover the following relationship hold between reservation price and the regret/rejoicing measures. 27 LRR;t = LRR (c; F; (rej_1_stept ; rej_t_stept ); (reg_1_stept ; reg_t_stept )) = (reg_1_stept ; reg_t_stept ; rej_1_stept ; rej_t_stept ; c; F ) (44) The following …rst-order partial e¤ects from Equation(44) o¤er some very interesting testable hypotheses: @LRR;t @reg_i_stept @LRR;t @rej_i_stept 3.3 @LRR;t = @ t > 0; for i = 1; t @reg_i_stept (+) @ t < 0; for i = 1; t @rej_i_stept (+) @ t (+) @LRR;t = @ t ( ) (45) (46) Latent Reservation Price Estimation Every continue-or-stop decision in each of the 673 price search sequences can be represented by a vector (Yij1; :::; YijT ) = (0; :::; 0; 1); with Y =0 denoting search and Y =1 denoting stopping. Search tasks in which the buyers accept the …rst price (134 of them) will be excluded from the estimation, leaving us with a total of 2735 observations. For simplicity, assume the latent reservation price at each decision point of the search LRR;t has a simple linear relationship with all the likely covariates in the decision process, including all the regret and rejoicing measures as in the following equation (We are going to suppress the subscript RR from this point on unless a digression from the regret/rejoicing model is explicitly pointed out) : Lijt = 0 Xijt + 0 ijt + 1 ijt + 2 Searchijt + Ii + "ijt (47) in which, i = 1; ::::; 64 indexes subjects participating in the search tasks; j = 1; :::::; 11 indexes the search tasks; t = 1; :::::; T indexes the search rounds; Xijt = (reg_1_stepijt ; reg_t_stepijt ; rej_1_stepijt ; rej_t_stepijt )0 is the vector of regret/rejoicing measures; ijt is a vector of other factors that may a¤ect search such as subject’s total earning in the experiment at t; ijt is subject’s payo¤ in the current search if she stops now; searchijt is the number of prices encountered so far in the current search; 28 Ii ’s are individual-speci…c dummies, re‡ecting subject’s innate di¤erences in a variety of factors that in‡uence the search outcomes, including but not limited to di¤erences in their risk attitudes and regret/rejoicing intensities and abilities to process the task at hand; "ijt are i.i.d error terms, distributed N ormal(0; ! 2 ): The estimation is carried out via Maximum Likelihood in a way that is akin to a standard Probit model. There are two kinds of possible outcomes, each implying a certain relationship between the latent reservation price Lijt and the lowest price observed Lijt : Pr(Yijt = 0) = Pr(Lijt > Lijt ) Pr(Yijt = Pr(Lijt > 0 Xijt + 0 1 = ( (Lijt ( 0 Xijt + ! = 1) = Pr(Lijt H1ijt ) ijt + 1 0 ijt + ijt + 1 ijt 2 Searchijt + + Ii + "ijt ) 2 Searchijt + Ii )) 0 = Pr(Lijt Xijt + 0 ijt + 1 ijt + 2 Searchijt + Ii + "ijt ) 1 = ( ( 0 Xijt + 0 ijt + 1 ijt + 2 Searchesijt + Ii Lijt )) ! where (:) denotes the CDF of standard normal distribution. (48) (49) 0 s. The objective is to identify (b; b; b) that maximizes the probability of observing all the Yijt (b; b; b) = arg max P log Pr(Yijt = 1) + Y ijt=1 P log Pr(Yijt = 0) (50) Y ijt=0 Equations (45)-(46) lead us to be particularly interested in testing if the following hypotheses will hold in the estimation results: 1. Buyers’ reservation prices will be shifting through the search process, rather than staying constant. This is true even during a single search sequence, as a result of experienced regret and rejoicing. 8 If it holds, simple risk aversion cannot explain satisfactorily why people tend to search too little. In addition, we have direct evidence that search is history-dependent. 2. In a purchasing context, buyer’s reservation price increases (i.e. search less aggressively) with greater regrets and decreases (i.e. search more aggressively) with greater rejoicing. 3. Quantitatively, regret moves the reservation price more than rejoicing does. 8 To my knowledge, previous experimental studies have only examined the changes in reservation prices between two consecutive searches by having the subjects directly report their reservations before the search begins. See Sonnemans (1998) and Einav (2005). 29 4. The occurrence of recall is more frequent in search sequences where regrets are more present. 5. To the extent that satis…cing behavior exists, reservation prices should fall at roughly the rate of c with each additional search. 3.4 Estimation Results 3.4.1 Descriptive Statistics Table 1 below presents the summary statistics of the main variables of interest to be used in the MLE estimations. Table 1: Summary Statistics Variables Mean Std. dev. Endogenous Variables Last Round 0.20 0.40 Explanatory Variables Search 6.42 5.53 Prices 499.73 9.81 Lowest Prices 493.10 6.28 Search Pro…t if stops now 0.48 6.76 Balance 41.21 31.70 Regret-not stopping last round 0.63 0.48 Regret-not stopping earlier 2.77 4.18 rejoicing-not stopping last round 2.76 5.44 rejoicing-not stopping earlier 2.34 5.17 Sample Size (N=2,735) Note: 773 observations are deleted from the sample in which the search sequence length equals 1. Min Max 0 1 2 467 467 -27 -14 0 0 0 0 40 535 519 31 149 1 32 38 38 After eliminating searches with stops after only observing the …rst price, 2,735 stop-or-go decisions from 539 individual searches remain in our sample. The number of searches in a task ranges from 2 to 40. The random prices appear to follow the distribution very well. Search pro…t if stopping now tracks buyer’s payo¤ from stopping at each of the decision point. In the event that all the subjects in the experiment try to reach a certain aspired level of payo¤, this measure should exert a very large impact on their decisions as to whether to stop or continue the search. The variable, balance, measures the subject’s total earning from previously completed search tasks which is inserted here as another possible control variable. 30 The …nal four explanatory variables on the list have the preponderance of our interest. First, notice that one step regret literally acts as a dummy variable due to the fact that if current round payo¤ is worse than last round payo¤, the di¤erence will be exactly equal to the search cost c that’s wasted. In 63% of the cases, the realized price is no better than the previous price, resulting in one-step regrets. On the other hand, in roughly 35% of the cases, the buyer does discover a better price than the last round, resulting in one step rejoicing. The magnitude of one step rejoicing is sometimes much larger (the highest being 38). In roughly 28% of the cases, the buyer discovers a much better price in the sense that her current payo¤ beats any counterfactual payo¤ had she stopped in any of the earlier rounds. This translates into a positive multi-step rejoicing. Nevertheless, over 50% of times after she has obtained such a good price as de…ned by positive multi-step rejoicing, the buyer continues to search for at least one more price, demonstrating a tendency to become more aggressive while experiencing rejoicing. We’d also like to point out that one step rejoicing and multi-step rejoicing are very close in values in our sample. Regarding the question of when buyers decide to stop in the search process, our sample shows that 73% of them ultimately decide to stop when getting a better price than the last round, of which 85% also have a positive multi-step rejoicing when they stop. In 26% of the stopping rounds, the buyer actually does not do as well as if she has stopped at one of the earlier rounds, of which 70% have a lower payo¤ than they had in the last round. The overall recall percentage in our sample is 17.4%, which is similar to what’s found in previous experiments. Table 2 reports the Pearson correlation matrix of all the major explanatory variables within our estimation. All of the correlations in the table are found to be signi…cantly di¤erent from zero at the 5 % level. In particular, the variable Number of Search has very positive correlation with multi-step regret at 0.88. This may arise from the fact that buyers usually prolong their searches confronted with a string of unfavorable prices which also results in large mutli-step regrets. Also, it is noteworthy that the positive correlation between one-step regret and multi-step regret is 0.41 whereas the positive correlation between one-step rejoicing and multi-step rejoicing runs as high as 0.98. The pair of rejoicing variables seem to both arise from the discovery of a very low price in the sequence. All of the regret and rejoicing variables are negatively correlated, as expected. 31 Table 2: Pearson Correlations between Independent Variables Variables Search t Balance Reg_1 Reg_t Rej_1 Rej_t Note: All 3.4.2 Search Balance Reg_1 Reg_t Rej_1 Rej_t t 1 -0.492 1 -0.118 0.193 1 0.216 -0.388 -0.087 1 0.881 -0.586 -0.131 0.414 1 -0.205 0.524 0.103 -0.658 -0.317 1 -0.246 0.506 0.101 -0.589 -0.300 0.978 1 correlations reported here are signi…cantly di¤erent from zero at p=.05 Main Results All of the succeeding estimations are run using Maximum Likelihood algorithm based on Equations (48) through (50). The algorithm is very similar to a standard Probit estimation with the caveat that throughout the estimations, the coe¢ cient on the variable current (lowest) price is …xed at 1: Pr(Yijt = 1) = ( 0 Xijt + 0 ijt + 1 ijt + 2 Searchijt + Ii + ( 1) Pijt ) where Pijt denotes the lowest price including the current period price. 32 (51) Table 3: Maximum Likelihood Estimation (1) Ind. variables prelim.a coef. Std. Err. Const. 491.10** 0.48 Reg: 1 step 5.05** 0.77 Reg: t step -0.26** 0.05 Rej: 1 step -0.42** 0.12 Rej: t step 0.34** 0.12 Search t of the Latent Reservation Prices (2) (3) Complete b w/ Spline Fn.c coef. Std. Err. coef. Std. Err. 494.24** 1.77 493.72** 1.89 4.02** 0.69 5.07** 0.59 0.08 0.16 0.03 0.19 -0.14** 0.04 -0.12* 0.05 -0.41** -0.13 0.00 0.06 0.12 0.10 0.13 0.03 0.23 0.09 -0.35** -0.11 0.00 0.03 0.13 0.13 0.03 0.23 Balance Task Search*Reg1step Reg: t step-10 0.33 0.26 Reg: t step-20 -0.62 0.42 Reg: t step-30 9.42** 1.55 Rej: 1 step-10 0.09 0.12 Rej: 1 step-20 -0.77* 0.32 Sample Size 2,735 2,735 2,735 Notes: a :Error terms clustered at the subject level. b : 63 individual dummies inserted. Coe¢ cients on dummies not shown here. c : 63 individual dummies inserted. Spline functions have knots at 10,20,30. a;b;c : robust std err applied. **: signi…cant at 1% level. * signi…cant at 5% level. Table 3 reports the main results from the estimations. The speci…cation in the estimation has enabled us to interpret the coe¢ cients in the result as unit changes in the latent reservation prices given marginal changes in the set of independent variables. Column (1) presents a preliminary estimation by including only the regret and rejoicing variables. The result shows that all of these four measures have signi…cant e¤ects on deciding the reservation price for the buyer. Moreover, one step regret and one step rejoicing are found to shift the reservation price in the directions predicted by our model. Most striking of all is the very strong quantitative e¤ect of one step regret in making the buyers revise their reservation prices upwards by as many as 5.05, which is a very sizable adjustment given that the standard deviation of the price distribution is 10. Compared to one step regret, multi-step regret, one step rejoicing and multi-step rejoicing all seem to be much less relevant in the decision, although keep in mind that the average values of these measures are much larger than one step regret so the overall e¤ect needs …ner calculation. Finally, since one step rejoicing and multi-step rejoicing are so highly correlated and their coe¢ cients in (1) are of 33 opposite signs, it is worthwhile to ask if they are jointly signi…cant. A likelihood ratio test soundly rejects the null hypothesis that they are jointly insigni…cant (p=0.0012). Column(2) presents the full-‡edged version of the estimation. It inserts an additional set of control variables into the estimation, including task number, search rounds, current search payo¤, total earning balance, and an interaction term between search and one-step regret in order to test whether regret matters more in longer search sequence. Besides, individual dummies are added to control for subject level variations in determining reservation price that are unrelated to the observed prices. Multi-step rejoicing is omitted from the estimation to avoid any collinearity issue with one-step rejoicing. One-step regret remains very signi…cant in the complete version of the estimation. Other things equal, the regret generated by seeing this round’s search to be less pro…table than not searching at all will trigger the buyer to revise her reservation price upwards by 4.02. At the average values of other covariates, the marginal e¤ect of one step regret will raise the probability of stopping in the next round from 0.18 to 0.31, a 72 percent increase. Mutli-step regret raises the reservation price as well, although not very signi…cantly. Rejoicing is found to signi…cantly reduce one’s reservation price, yet the marginal e¤ect is much smaller in scale. All these are in line with the behavioral model’s predictions. In the meantime, a risk-aversion based story of sequential search appear all but dead, because the constant reservation property does not appear to hold at all. In Column (2) current search pro…t is not a signi…cant predictor of one’s reservation price, contradicting the premise of a satis…cing model in which search agent wants to reach certain aspiration payo¤ level. If higher payo¤ will make one more likely to quit the search, we would expect a positive coe¢ cient on t: The other possibility is that if we assume heterogeneous aspiration levels among subjects, the higher one’s aspiration payo¤ is, the more aggressively she would have to search, which would also create results that appear similar to Column (2). Therefore, it is too early to completely discount the "satis…cing" explanation of too little search. In the next section, we’ll run a direct horserace to demonstrate why regret matters more than satis…cing behaviors. Finally, the coe¢ cient on the number of search turns out to be signi…cantly negative, implying that buyers become more aggressive searchers the longer the search goes on . This directly contradicts a type of behavioral strategies that some experimenters have previously thought would apply in this setting, which is "keep searching until the number of searches exceeds XX". In contrast, the negative relation between number of searches and reservation price tends to lend support to either an escalating commitment story (i.e. keep searching until one …nds some favorable price that would justify her lengthy search) or a satis…cing story in which one’s reservation price falls proportionally with more searches. On the other hand, the variables, task, balance, and the interaction between one-step regret and search do not appear to be important here. 34 Since people seem to be more concerned with one-step regret than with multi-step regret whose values come from a much wider range, the reason that the e¤ects of multi-step regret or rejoicing appear to be weak could possibly be attributed to people’s increasing insensitivities towards larger values of regret or rejoicing. In other words, the curvatures of the regret/rejoicing intensity functions may matter also. Consequently, in Column (3) we divide multi-step regret and multi-step rejoicing into several numerical intervals and institute a spline function on each of them. For example, the variable, Reg: t step-10, subtracts 10 from the original value of multi-step regret (if >10) and helps to test whether there is some additional change in slope once the value of multi-step regret exceeds 10. In the results, we discovered: First, adding the spline functions do not change the e¤ects of remaining variables. Secondly, regtstep_10, regstep_20 as well as rej1step_10 do not appear to have signi…cant impact on the search process. In other words, intermediate values of regret and rejoicing do not change the reservation price much. However, once the regret value exceeds 30 and the rejoicing value exceeds 20, their e¤ects grow much stronger and signi…cant. In particular, one unit increase in regret once it exceeds 30 will raise the reservation price by 9.42, literally halting the search right away. The implication is simple: People really hate very extended searches, and perhaps they are prepared to give up immediately once they have drawn many disappointing prices (even though the underlying price distribution has not changed a bit!). 3.5 Robustness Checks There are two main types of arguments that could cast into doubt whether the revisions of reservation prices are truly caused by concerns for regret and rejoicing in the sequential search task, one behavioral and one statistical. First, whether or not the subjects in the experiment truly trust that prices are drawn randomly from the distribution they are told or whether or not they have a basic comprehension of the underlying (truncated normal) distribution could be an issue. In either of the two scenarios, subjects will be undertaking some extent of learning of the price distribution while searching for the best price. In that case, regrets (rejoicing) that are associated with unfavorable (favorable) price draws would also signal to the buyers of an unfavorable (favorable) distribution. Consequently, it is only natural for the buyers to decide to search less (more) extensively. Bounded rationality on the subject’s part causes a possible spurious relation between regret/rejoicing and the reservation price in the observed manner. Secondly, if due to the computational demands of the search task, people are unable to correctly pinpoint the optimal reservation price. Sometimes they set it too low; sometimes they set it too high. Facing the same (unfavorable) price sequences, those who set the reservation too low will always undertake longer searches than those who set it too high, resulting in a higher frequency of regrets based on higher expenditure on search costs. These subjects will also have tendency 35 to adjust their reservation upwards over time when they learn to correct the computation errors. Again, computational de…ciency together with statistical reasons causes a possible spurious relation between regret/rejoicing and the reservation price in the observed manner. In order to clarify these two issues, we run three additional robustness checks in the hope of corroborating with the basic …nding. 1. Prior to the start of the …rst paid search task, each subject in the experiment is allowed to practice freely with the software within a certain time limit. They can generate from zero up to an unlimited number of price searches during this time without any concerns for committing mistakes. We tabulated the number of prices drawn by each subject from the same distribution as in the later paid tasks. The numbers di¤er widely, ranging from a low of 7 prices to a high of 756 prices. The median number of prices searched during the trial period is 80.5. So for anyone who has seen at least 80 price draws during the trial, familiarity with the distribution should not be an issue. In Robustness Check (1), only subjects who have made above median number of free price searches during the practice will be included in the re-estimation. Figure 7: 2. Experimental subjects’concerns about the distribution and their inability to …nd the optimal reservation prices should be mitigated with more search tasks completed. An average subject would have observed around 30 price draws by the time she …nished half of the search tasks. As a result, we are going to repeat the estimation with data from Task 6 to Task 11 to see if the results hold (Robustness Check 2) 3. We are also going to separate the data into two groups: subjects who have recalled at least once through the 11 tasks, and subjects who have never recalled. If search sequences in which recalls happen in the end are more likely to identify those who set reservations too low and who stumble upon bad price draws (See Figure 8 for an illustration of this point), estimating exclusively 36 on the group who has never recalled would be instructive as to whether the relationship between reservation price and regret/rejoicing is barely statistical or not (Robustness Check 3). Figure8 Table 4: Results from the three robustness checks Ind. variables Const. Reg: 1 step Reg: t step Rej: 1 step Search t (1) Practices>80.5 a coef. Std. Err. 492.72** 1.96 4.19** 0.98 0.07 0.23 -0.14** 0.05 -0.28* 0.13 -0.04 0.14 0.02 0.04 0.01 0.30 0.13 0.12 1,419 (2) Task>5 b coef. Std. Err. 494.48** 3.51 4.83** 1.09 0.01 0.30 -0.19** 0.07 -0.25 0.19 0.00 0.23 0.08 0.07 -0.45 0.23 0.12 0.09 1,401 Balance Task Search*Reg1step Sample Size Notes: a;b;c : 63 individual dummies inserted. robust std err applied. **: signi…cant at 1% level. * signi…cant at 5% level. 37 (3) Never Recalled c coef. Std. Err. 496.58** 0.50 1.69* 0.87 0.03 0.08 -0.02 0.03 -0.45** 0.09 -0.15 0.11 0.02 0.02 -0.16 0.14 0.12 0.09 1,066 All the three robustness tests clearly con…rm that the strong relationship between buyer’s reservation price and one-step regret is neither the result of behavioral uncertainty regarding the price distribution nor the result of sheer statistical coincidence. All of the quantitative e¤ects resemble very closely those derived in the main estimation. In particular, in the …rst two subsamples containing experienced subjects, buyers are found to adjust their reservation prices upwards by roughly $4 once they experience a one-step regret. In the third subsample, buyers who never recall a past price would adjust reservation price upwards by a slightly smaller amount, nevertheless, the e¤ect is still robust at 5 percent level. 4 4.1 Discussions and Extensions Why risk aversion does not explain "too little search"? One of the most publicized explanations in the present literature to explain "too little search" is based on risk aversion in the search agent’s utility function. In essence, risk averse searchers view each search decision as a gamble, trading o¤ potential gains from …nding a better price against a safer option which is to save on the search cost. The more risk averse one person is, the more conservative her search pattern will be. However, as the early discussion in this paper has pointed out, regardless of the degree of risk aversion, there should be no dynamic aspects to the search problem at all. Basically, a constant reservation price prevail throughout the search process. However, this is what we have observed in the current dataset. On the contrary, it is found that the subject’s reservation price during the search ‡uctuates in a predicted manner with her experienced regret and rejoicing. And this could explain why subjects search too little in connection with our theory incorporating anticipated regret and rejoicing into the search. Other authors have also raised questions about the plausibility of risk aversion in explaining search. Schunk and Winter (2005) in their paper explicitly compared two sets of risk coe¢ cients: the …rst set is elicited through a pre-search questionnaire directly surveying subject’s risk preferences; the second set is inferred from their actual search behavior assuming risk aversion is the only driver of search. Surprisingly the two sets of risk coe¢ cients are found to have very little correlation. Sonnemans (1998) also provided evidence that rejects risk aversion in explaining actual search behaviors. His argument is that if risk aversion is interpreted as a willingness to accept lower average payo¤ in exchange for lower variability of payo¤, then pure risk aversion is only able to explain about 20 percent of the situations where people search too little. 38 4.2 4.2.1 Three-way Horserace between Regret/Rejoicing, Risk Aversion, and Satis…cing Models Methodology: A model’s explanatory power ultimately rests on its ability to predict actual behaviors. Of the three possible models we have to explain too little search, which one could predict the actual pattern of search the best? A three-way horserace among them helps reveal this point: First, our set of candidate strategies contains the range of all plausible satis…cing strategies, the range of all plausible reservation price strategies re‡ecting pure risk aversions, as well as the strategy of our principle interest, the moving reservation price strategy based on our estimation result from the last section. Speci…cally, the predicted reservation prices used in the horserace are predicted values from the main estimation equation (51). Let us use the example of one particular experiment subject so as to illustrate how the horserace is run. For instance, Subject 33 in the experiment performed a total of 11 separate searches. For any one of those 11 searches, if a search strategy predicts the same set of stop or continue decisions as what were made by the subject, then we’ll say the strategy predicts the search, or otherwise, the strategy fails to predict the search. Between a pair of strategies, the one that predicts a higher proportion of a subject’s total searches is said to win the horserace for that subject. Here is how one possible satis…cing strategy "Satis…ed7 "(i.e. stop whenever current search payo¤ exceeds $7), one possible risk-aversion strategy "RP490" (i.e. hold a constant reservation price equal to 490), and the Regret/Rejoicing predictions "RR" perform against each other in the horserace for Subject 33. Table 5: Horserace Example (Subject 33) Candidate Strategy RR Satis…ed7 RP490 Predicts 9/11 5/11 7/11 So RR wins the three-way race for Subject 33. Table 6 showcases the outcomes of the horserace run between the regret/rejoicing model and the whole range of plausible satis…cing strategies. To reiterate, in the regret/rejoicing model, a subject stops the search whenever the randomly received price is lower than the predicted reservation prices estimated from Equation (51). In the satis…cing model, a subject stops the search whenever the randomly received price results in a total search payo¤ in excess of a level prescribed by the strategy. The table makes a forceful case that the regret/rejoicing model performs a much better job in describing the data than the satis…cing model. The former beats every conceivable satis…cing strategy in terms of predicting correctly a higher proportion of search outcomes for more individuals in the sample. Even if we allow each individual to have di¤erent aspiration level and correspondingly pick the best satisfying strategy for every subject, the regret/rejoicing model still wins the horserace (see the last column in the table). 39 Table 6: Regret/Rejoicing Predictions vs. All Satis…cing Strategies (S2-S15) Satis…cing Strategies S2 S3 S4 S5 S6 S7 Wins 43 43 46 44 43 49 Ties 10 9 9 12 12 7 11 12 9 8 9 8 Losses Net W32 W31 W37 W36 W34 W41 Note: 1) out of a total of 64 subjects. 2) S-Best assumes individual heterogeneity by selecting individual respectively. S8 56 4 4 W52 S9 59 3 2 W57 S10 60 2 2 W58 S11 62 0 2 W60 S12 62 1 1 W61 S-Best 24 20 20 W4 the best satis…cing strategy for each Similarly, the horseracing results between the regret/rejoicing strategy and the whole range of risk aversion strategies are listed in Table 7. Various degrees of risk aversion is mapped into constant reservation price strategies with various cuto¤s. Once again, the regret/rejoicing model predicts searches better in more individuals in the data than any constant reservation price strategy when applied to the whole sample. When heterogeneity in risk attitudes is allowed and for each individual we select the best constant reservation price as her strategy, the regret/rejoicing strategy does fall behind in its predictive power (see the last column in Table #). Table 7: Regret/Rejoicing Predictions vs. Risk Aversion Strategies (RP488-RP498) Constant RP RP RP RP RP 489 490 491 Strategies 488 Wins 45 34 33 34 12 20 18 17 Ties Losses 7 10 13 13 Net W38 W24 W20 W21 Note: 1) out of a total of 64 subjects. 2) S-Best assumes individual heterogeneity for each individual respectively. RP 492 31 17 16 W15 RP 493 28 23 13 W15 RP 494 33 17 14 W19 RP 495 34 15 15 W19 RP 496 46 10 8 W38 RP 497 51 3 10 W41 RP 498 52 6 6 W46 by selecting the best constant reservation strategy Overall, the horserace results suggest clearly that the data in themselves favor the risk/rejoicing explanation over the risk aversion and the satis…cing behavior explanations. for any one of the 673 random price sequences in the dataset, if a particular search strategy is able to predict the exact same set of stop-or-go decisions as is actually made by the subject, that search strategy will score one point for that individual on that particular search. An individual will appear to favor one of any two competing search strategies more frequently if that search strategy 40 RP Best 2 19 43 L41 score more points than the other one among the 10 or 11 search tasks that the individual conducted in the data. Consequently a strategy that’s favored by a higher proportion of 64 total subjects will win the head-to-head horserace with another strategy. In particular, we are interested in examining if our regret/rejoicing model which possess a reservation price structure as estimated in equation (51) can defeat both the risk aversion and the satis…cing search strategies. 5 Conclusion This paper tries to answer the question why people in the lab have a tendency to search too little compared to the risk-neutral benchmarks by o¤ering an entirely new perspective: anticipatory regret and rejoicing could be a prominent factor here. In addition to deriving this analytically by embedding the original regret theory into the search model, we also show that the regret/rejoicing explanation organizes the experimental data better than risk aversion and satis…cing behavior. However, one issue that we shun away from is whether people do use a search strategy of the reservation form, or they use some other bounded rational rules. Nor do we have the con…dence to claim that anticipated regret/rejoicing is the only factor operating here. Our theory stresses that it is because of the feedback information structure of the sequential search problem that makes regret and rejoicing not only asymmetric but also one-sided. If people can anticipate these future feelings correctly and if regret dominate rejoicing, people will end up searching too little. Empirically, both our anticipatory regret/rejoicing model and a simple learning rationale put forward by Sonnemans (1998) and Einav (2005) will deliver similar results; nevertheless, Proposition 2 predicts that by modifying the information ‡ow in the search task, the two motives can be disentangled. More speci…cally, if revealing one more price following the end of search can e¤ectively induce subjects to search more aggressively, then it must be the anticipation of future regret and rejoicing that a¤ect people’s reservation prices. More strongly, Proposition 2 predicts that in the counterfactual search environment, people’s behaviors should be path-independent, and recall should never happen. How much of the gap between observed search behaviors and e¢ cient search can be closed by a change in the information ‡ow should be subject to future experimental investigation. Finally, it is worth highlighting that the regret/rejoicing model has rich empirical implications in the marketplace. For example, search is going to be closer to optimum in markets in which consumers are more likely to obtain post-purchase information. The di¤erences between consumer’s accessibility towards post-purchase price information can occur naturally or by government decree, such as local markets vs. online markets, states with or without banned advertising on certain goods. Policies that enhance post-purchase price transparency should induce more e¢ cient consumer search in the …rst place. Second, the "price-matching" clauses that a lot of retailers use could also mitigate 41 anticipatory regret and therefore lead to few searches and presumably higher market power by …rms that implement these marketing tactics. 42 REFERENCES Bell, David. 1982. "Regret in Decision Making under Uncertainty". Operation Research. Vol. 30, No.5, pp. 961-981. Cox, J. and R. Oaxaca. 1996. "Testing Job Search Models: The Laboratory Approach". Research in Labor Economics, Vol. 15, pp:171-207. Einav, Liran. 2005. "Information Asymmetries and Observational Learning in Search". The Journal of Risk and Uncertainty, Vol. 30, No.3, pp: 241-250 Filiz, E. & E. Ozbay. 2007. "Auctions with Anticipated Regret: Theory and Experiment". American Economic Review, Vol. 97(4), pp: 1407-1418 Hey, John. 1982. "Search for Rules for Search". Journal of Economic Behavior & Organization. Vol 3, pp: 65-81. Hey, John. 1987. "Still Searching". Journal of Economic Behavior & Organization. Vol 8, pp: 137-144. Houser, D. and J. Winter. 2004. "How do Behavioral Assumptions A¤ect Structural Inference? Evidence from a Laboratory Experiment." Journal of Business and Economic Statistics. Vo 22, No. 1, pp 64-79. Kogut, C. 1990. "Consumer Search Behavior and Sunk Costs." Journal of Economic Behavior and Organization. Vol. 14, pp. 311-321. Loomes, G. & R. Sugden. 1982. "Regret Theory: An Alternative Theory of Rational Choice Under Uncertainty" The Economic Journal, Vol. 92, No.368, pp:805-824. Lippman, S. and J. McCall. 1976. "The Economics of Job Search: A Survey". Economic Inquiry. Vol 14, pp 155-189. Rothschild, M. 1974. "Searching for the Lowest Price When the Distribution of Prices is Unknown". Journal of Political Economy, Vol. 82(4), pp: 689-711. Schotter, A. and Y. Brounstein. 1981. "Economic Search: An Experimental Study". Economic Inquiry. Vol. 19, 1-25. 43 Schunk, Daniel. 2008. "Sequential Decision Behavior with Reference Point Preferences". working paper, University of Zurich. Schunk, D. & J. Winter. 2009. "The Relationship Between Risk Attitudes and Heuristics in Search Tasks: A Laboratory Experiment". Journal of Economic Behavior & Organization, forthcoming. Sonnemans, Joep. 1998. "Strategies of Search". Journal of Economic Behavior & Organization. Vol. 35, pp. 309-332. 44 A Proof of Proposition One Lemma 1 L0 (H) < 0 This is intuitive. Not only does the monetary return to search decreases with H; but also the likelihood and the extent of regret increase with H while the likelihood and the extent of R H+c 0 0 rejoicing decrease with H: More concretely, we have L (H) = G (H) cf (H) [ H 1f (x)dx cf (H)]+ Rb H+c ( 1)f (x)dx = [F (H) 1] [F (H + c) F (H)] [1 F (H + c)]: Note that all the three terms in the preceding expression are negative which coincides with our previous intuition. Hence L0 (H) < 0 throughout its domain [a; b]: Furthermore, since in order to make the question interesting it must be the case that L(a) > 0; L(b) < 0; lemma 1 thus implies that a seller, in anticipation of regret and rejoicing, should once again adhere to a unique reservation price strategy with the reservation price equal to H1 (c; F; ; ) which solves L(H1 (c; F; ; ) ) = 0: We can further simplify the three components in the expressions for L(H) : 1) Monetary Payo¤: Z b (x H)dF (x) c Z b = xdF (x) H dF (x) c H H Z b = xF (x)jbH F (x)dx HF (x)jbH H Z b = b H c F (x)dx H Z b c (52) H 2) Anticipated Regret: cF (H) Z H+c (c + H x)f (x)dx Z H+c Z H+c [(c + H) dF (x) xf (x)dx] H H Z H+c [ cF (H) + F (x)dx] H = cF (H) = cF (H) Z H+c = H F (x)dx (53) H 45 3) Anticipated Rejoicing: Z b (x c H+c = (c + H) Z H)f (x)dx Z b dF (x) + (c + H)(1 = (b c xf (x)dx H+c H+c = b F (H + c)) + Z H) Z [xF (x)jbH+c b F (x)dx] H+c b F (x)dx (54) H+c In aggregate, we now have L(H) = [b H Z Z b F (x)dx c] H Z H+c F (x)dx (c + H) + b H b F (x)dx (55) H+c Ultimately our goal is to diagnose whether and how concerns for regret and rejoicing will drive the seller to revise her search strategy relative to the case without those concerns. Because it is di¢ cult to derive the precise explicit solutions for H0 and H1 ; a direct comparison between the two cut-o¤ points is unavailable. To circumvent this problem, an alternative strategy is devised in which we examine the value of L(H) evaluated at H0 : Whether L(H0 ) > 0 (< 0) will provide su¢ cient evidence as to whether H1 lies to the left or right of H0 ; To that end, we …rst utilize the fact that H0 is the implicit solution to the equation G(H0 ) = 0. That is, G(H0 ) = Z Z b (x H0 )dF (x) H0 c=b H0 b F (x)dx c=0 H0 Hence, L(H0 ) = G(H0 ) Z = H0 ) = = (b (b ( H0 +c F (x)dx H0 c c ) Z (c + H0 ) + b H0 +c F (x)dx H0 H0 ) ( ) Z F (x)dx Z b H0 +c F (x)dx b F (x)dx H0 +c F (x)dx H0 H0 H0 +c Z [ Z Z H0 +c F (x)dx] H0 (b c H0 ) (56) H0 46 The last equation shows that if > > 0, then L(H0 ) < 0: Then it must be true that a < H1 < H0 because L(a) > 0; L0 < 0: On the other hand, if 0 < < ; then it must be true that H0 < H1 < b: If = ; then H1 = H0 : 47