Code creation in endogenous merger experiments.

Feiler, Lauren ; Camerer, Colin F.

I. INTRODUCTION

Groups and organizations often use specialized language, or "code," for coordinating economic activity. In this paper, we focus on the experimental creation of code among small laboratory groups, and what happens when groups are combined in a "merger."

Code is one facet of corporate culture, which is the value system, symbols, ideas, icons, stories, and language that express the informal contracts between a firm and its employees, customers, and suppliers--"how business is done" at a company. There is much popular writing about corporate culture, and interest among businesspeople, but little careful scientific research on how cultures are created, are changed, and affect performance. For example, many businesspeople say that conflicts in corporate culture contribute to merger failures (and there are prominent examples like the AOL-Time Warner merger). Firms are also interested in creating cultural systems that can motivate employees and create brand loyalty among customers.

While culture seems important, it is difficult to operationalize the vague concept of corporate culture precisely and measure it. Organizational economists writing about culture have singled out the role of code and categorization for firm productivity (e.g., Cremer, Garicano, and Prat 2004). We adopt the same emphasis on code, which is relatively easy to measure and create in the short span of a typical laboratory experiment. Code also has some properties shared by other components of culture, like values--for example, codes are often path dependent and difficult to recreate. Other components of culture besides code are set aside for future research.

To study code creation, we create simple "firms" in a laboratory setting and use a picture-naming paradigm in which they develop code. We then examine the impact on performance when two firms with specialized codes for different pictures are merged together. We are also interested in other features of a merger, such as perceptions of how difficult mergers are and the endogenous choice of whether to join the merger, and we measure these phenomena by having subjects bid for extra payments to join the merger. Our paper extends work by Weber and Camerer (2003), using different picture sets and endogenizing the subjects' choices of whether to join a merged group. Endogenizing their choices enables us to see whether subjects systematically underestimate the difficulty of merging and its economic consequences.

These experimental organizations are highly stylized and simple. Their simplicity gives us precision in measuring variables and understanding the determinants of performance. In the experimental world, the code words subjects use are their culture and the money they earn is their organizational performance. Furthermore, the experiments are just a starting platform onto which complications can be added. The long-run goal is to learn from a series of experiments, and just as theory develops from simple to complex, it is usually efficient to start with simple experiments and gradually complicate them. Because the design is obviously much simpler than large corporate behavior, concerns about the simplicity of the design are most useful when they come in the form of a suggestion of how to enrich the design, and a conjecture about how the enrichment will change the behavior.

A. Merger Failure

Although mergers are often met with excitement on Wall Street and in the boardroom, there is ample evidence that acquiring firm shareholders often lose from mergers (e.g., Andrade, Mitchell, and Stafford 2001). Most of these studies use stock market returns in a short window of time around the merger announcement, so they rely on the hypothesis that the stock market guesses future merger success accurately. By using only returns, it is impossible to tell whether mergers actually generate anticipated economic synergies years later and whether acquiring firms overpay even if synergies do result.

In careful studies using Federal Trade Commission (FTC) Line-of-Business data to compare acquired firms with similar firms that did not merge, Ravenscraft and Scherer (1987, 1989) found that on average, acquired firms are less profitable than before the acquisition and mergers often result in later divestitures. The high turnover rates that accompany mergers are frequently attributed to firing poor managers in the acquired firm, but the best managers are actually the ones who are most likely to leave in the first year after a merger (Walsh 1988; Walsh and Ellwood 1991). Using longitudinal data on Danish firms, Smeets, Ierulli, and Gibbs (2007) found more turnover when merging firms are from different industries than when employees share the same industry background. This evidence stands in contrast to explanations of post-merger turnover as purely a result of economies of scale or scope, and instead points to culture-based explanations.

Business people involved in mergers often say that many difficulties in mergers have to do with clashes in the cultures of the merging firms. A 1992 Coopers and Lybrand study on the largest acquisitions in the United Kingdom from the late 1980s to the early 1990s found that the executives surveyed considered 54% of the acquisitions to be failures. Target management attitudes and cultural difference was the most widely cited cause of merger failure; 85% of those surveyed listed it as a key factor (Sudarsanam 1995). (1)

Some examples indicate that firms can under-forecast how differing ways of doing business can prevent the imagined synergies of a merger. Sony, known for its innovative consumer electronics, branched into the entertainment industry in the late 1980s and early 1990s, acquiring firms like CBS Records and Columbia Pictures. Now, conflicts of interest between the electronics and entertainment divisions stymie the development of popular products, such as portable music players that allow audio files to be easily transferred from computers. Sony's electronics division is home to such revolutionary products as the Walkman and the most successful transistor radio, and its developers would be expected to lead the way in creating groundbreaking digital media devices. However, the entertainment division is part of an industry with a long history of battling anything that could potentially foster piracy, such as CDs and VCRs. The entertainment side of Sony insists on copy protection mechanisms that make the products developed by the electronics division too cumbersome to be widely sold. Sony has even produced a music CD that could not be read on its own Sony computers! The opposing value systems at the two warring sides of Sony prevent each division from profiting as much as they likely could if the values could be reconciled within the firm (Rose 2003).

The AOL-Time Warner merger is a "perfect storm" in which many factors, including culture, apparently combined to create an unsuccessful merger (see the carefully researched book by Klein 2003, who covered AOL as a Washington Post reporter for many years). Time Warner employees valued a "best in class" structure, in which each division strove to be the best compared with other companies in its field. Under this structure, each division essentially operated as its own separate firm. There was little communication between different divisions and few successful joint projects. Being unused to this independence, AOL employees were appalled at being charged for services by a division of AOL-Time Warner. In turn, genteel Time Warner workers were disgusted at the crass language AOL employees would use in meetings and with clients. AOL employees had developed a reputation for shouting at and exploiting clients (actually defrauding advertisers on AOL and creating accounting frauds), in sharp contrast to Time Warner's polite business etiquette and old-school decorum. While it is apparent in retrospect (and perhaps prospectively) that the cultures of AOL and Time Warner were bound to clash, the CEOs who engineered the merger and the boards who approved it did not seem to acknowledge this possible clash or its effects. In April 2002, accounting rules requiring firms to recognize declines in stock values after mergers led to a $54 billion write-off, the largest corporate loss in U.S. history. In December 2004, Time Warner (which shed the "AOL" part of its name in 2003) paid $510 million to settle accounting fraud allegations at AOL.

The Sony and AOL-Time Warner cases are only anecdotal evidence, but they indicate the extent to which cultural clashes can reduce the profitability of a merger, even on a large scale. Although culture is widely acknowledged as a key factor in the success or downfall of a merger, there is scant scientific evidence about the role culture plays. Many organizational studies are case studies based on an author's years of experience within a single firm. In such studies, it would be nearly impossible to isolate the effect of any single component of the firm's operations. By using a laboratory experiment, we are able to examine one specific facet of corporate culture to learn about its effects on a merger, and provide analyzable data rather than mere anecdotes. Note that because the groups in the experiment are small, their behavior might be considered an analogy to combinations of work groups within a firm (e.g., merging two divisions or small teams to work on a bigger project), rather than to corporate mergers.

B. Code

Codes are sets of words or phrases that are used to efficiently convey information within an organization. Code allows members of an organization to describe aspects of their complex environment and reduce the costs of communication (Arrow 1974; Cremer, Garicano, and Prat 2004; Wernerfelt 2004). Code typically develops as members of an organization interact over time, and it is therefore unlikely to be immediately understood by outsiders.

Codes and jargon are common in organizations. To a cop, "11-27" means "subject has a felony record but is not wanted"; in a fast-moving kitchen, "give it some radar love" is an instruction to microwave a dish; air traffic controllers call the holding area for planes whose arrival gates are occupied the "penalty box" (which is an actual box in hockey). In all these situations, the members of the organization know the code and use it to quickly communicate with each other. Besides coordinating activity rapidly and clearly, code serves other organizational purposes. Code can: inspire ("just do it!"); identify who is in a group and who is not; relieve tension ("circling the drain" is emergency room slang for a patient who has suddenly taken a turn for the worse, "code brown" for a bed with excrement); identify villains ("Larry Parker syndrome," a patient complaining of pain who is looking for an insurance settlement); and sanitize tragedy ("collateral damage," military slang for civilian casualties).

Although efficient code normally takes a while to develop, it can be created quickly. People working on a project together may choose to create a common set of definitions that they will use throughout the task (Cremer, Garicano, and Prat 2004). Even robots have been found to quickly develop code, in an artificial intelligence experiment in which robots equipped with video cameras took turns describing what they "saw" to the other robot, which had to guess the object being described. Words that had been used successfully with the most frequency became the ones that would always be used to describe a certain object (Steels 2003).

Organizational culture is obviously more than code, but code is a natural place to start because it can be created so rapidly and has many similarities to the larger concept of culture. Although definitions of organizational culture vary, it is generally described as a shared social understanding among members of an organization, resulting in commonly held assumptions and views of the world (Cremer 1993; Rick, Weber, and Camerer 2007; Schein 1985, 1990). In economic terms, culture allows members of an organization to tacitly coordinate actions and provides guidelines for behavior under unforeseen change (Kreps 1990). Language plays an important role in culture because it restricts what can be communicated and potentially shapes the way people think.

In recent theory, organizational code is created out of the need for rapid, clear communication between group members. More precise or longer messages provide more information, but they cost more to communicate than shorter messages do (Cremer, Garicano, and Prat 2004; Wernerfelt 2004). In our study, laboratory firms will have to develop code to do well in a task that rewards speed and accuracy.

II. DESIGN

We use the picture-naming paradigm introduced by Weber and Camerer (2003) to examine the problems that can arise in a merger based on lack of a common code. (2) To assess subjects' ability to anticipate merger difficulties, we elicit subjects' valuations of a merger using auctions and simple questions. In this section, we will detail the procedures of our experiment and then explain our design choices and how the experiment captures elements of organizational mergers.

Nine sessions were run between April and May of 2004. Another four sessions were run between October and December of 2004. (3) Each session was conducted simultaneously at UCLA and Caltech. All instructions were the same for both schools. Subjects were UCLA and Caltech undergraduate and graduate students.

The experiment contains three phases, though each phase is composed of a few distinct tasks (Figure 1). One of these tasks is the picture-naming task, in which subjects describe and identify pictures that they see on their computer screens, using a web-based program. (3) The basic structure of this task is the same in each phase. In every round, each person in a group sees eight pictures on her computer screen, which are chosen randomly from a larger set of 16 pictures.

Pictures are displayed in a random order on each person's screen. One person is designated as the "manager." On the manager's screen, four of the pictures are numbered 1-4. The manager must describe these four pictures to the "employees" by typing messages that are sent to all members of the group. The employees click on a picture to identify it. Once each employee has correctly identified the four pictures, or if time runs out (after 300 sec), the next round begins. The manager rotates cyclically in each round, so participant 1 is the manager in the first round, 2 is the manager in the second round, and so on.

Payments for this task are a function of time and accuracy. Employees earn $0.60 each round, minus $0.005 for every second it takes to complete the task (a $.30/minute penalty), and minus $0.50 for each incorrect guess. Managers earn the average of the employees' payoffs. This payment structure requires the laboratory "firms" to develop a shared language to quickly and accurately identify each picture--it rewards good code. The better participants are at creating short but clear descriptions of each picture, the better they will perform.

While the picture-naming paradigm was not developed to be a close replica of any particular organizational process, it may be helpful to draw an analogy to something concrete. The picture naming in our experimental organization is like a business where one worker has private information about which of many scenarios is occurring and must describe that scenario to another worker who can "see" possible scenarios only in the mind's eye. An example is emergency services like police dispatching. A 911 operator talks to witnesses to a crime and forms an image of central elements of the crime scene--are victims badly injured? Is a suspect armed and dangerous? The operator is like the "manager" in our experiment who knows which picture is the correct target picture. The operator must then translate that 'picture' into a code that allows police heading to the scene to imagine which of several scenarios they will confront. The police are like the employees in our experiment, except that experimental subjects see actual pictures and the police only envision scenarios, which are gradually narrowed down by the operator's coded description. The analogy between the lab and the field holds if one accepts the hypothesis that experimental subjects seeing pictures on a screen are roughly similar to highly expert employees who have clear images of possible scenarios in their minds.

Figure 1 is a diagram of the events that occur in each phase of the experiment. Before the first phase begins, participants receive a brief set of instructions, and then they participate in a short four-period practice session with different pictures. Subjects also complete a short quiz to ensure that they understand the task.

[FIGURE 1 OMITTED]

In the first phase, there are two separate groups--six UCLA students in one group and six Caltech students in the other group. Each group sees subsets of 16 pictures of buildings and other features from their own campus. There are 40 rounds of the picture-naming task in this stage. This stage is meant to familiarize subjects with the picture-naming task and allow them to create code in their "area (campus) of expertise." Following this stage, subjects take part in a public good game that yielded no interesting result and will not be discussed further.

The second phase is the first merger phase. At the beginning of this phase, subjects are told that another group of six people is participating in the same experiment at another school, which we specify as UCLA or Caltech. Subjects will participate in the picture-naming task for ten more periods, but instead of being in two groups of six, they will be divided into three groups of four. We explain that of the six people at each school, two will be chosen to be in a "mixed" group. The mixed group will see eight pictures at a time, randomly drawn from a set containing pictures from both original groups, so on average half the pictures each person sees will be new to her. However, in addition to the normal earnings, she will be paid an additional amount to join the mixed group. The four people at each school who are not in the mixed group will do exactly the same task as they did in the first phase, except in a four-person "same" group rather than a six-person group.

After explaining the basic structure of the merger, subjects are asked to guess how much the average person in the mixed group will earn over the next ten periods, excluding the additional payments from bids. They are told the average earnings in each group in the original picture-naming task, over all 40 periods and also over the last ten periods, to help subjects use these data in making their estimates. They also see sample pictures from the two-campus pool of pictures the mixed group will see. Subjects who are not in the mixed group are paid $2.00 if they guess an amount within $0.50 of the actual average earnings of the mixed group. (4)

When subjects have written down their guesses, they are given instructions on the merger bidding process (with simple practice bidding). Subjects are asked to write down the additional amount they would need to be paid to join the mixed group. They write this as a total for the ten periods. The two people from each original group who bid the lowest amounts are the ones who are placed in the mixed group. Then the mixed group and two unmixed groups participate in ten rounds of the picture-naming task. Members of the unmixed group at each school again see pictures from their own campus, while members of the mixed group see pictures from a set containing eight UCLA campus pictures and eight Caltech campus pictures. The eight pictures from each campus were selected randomly from each original set of 16. Members of the mixed group earn money from their performance in the picture-naming task plus the amount they bid to join the group.

The second merger stage (or third phase of the experiment) is run exactly like the preceding phase, with two exceptions. The first is that before making their guesses, subjects are told the average earnings of the mixed group in the prior ten-period picture-naming task. The other difference is that people may leave or enter the mixed group in this stage, based on their bids. At the end of the second merger stage, each subject is paid according to his or her total earnings from all parts of the experiment.

The last four sessions used a different auction mechanism than the first nine sessions (and excluded the public good phase). While the first-price auction we used in the first nine sessions is easy to explain to subjects, in theory, subjects will bid more than their valuations. To try to make subjects ask for their lowest acceptable payment, we conducted the last four sessions with a Vickrey, or second-price, auction. As before, the two subjects from each school who bid the lowest amounts joined the mixed group, but in these sessions, they were all paid the third-lowest bid. The experimental instructions explained why the auction is incentive compatible: Subjects should bid exactly the minimum amount they would accept, because their bid does not determine the amount they actually get paid.

[FIGURE 2 OMITTED]

This design is meant to have central features of a typical horizontal or diversifying merger, in which two firms who make related products, but have differences in culture and code, merge and must coordinate activity. To create differences, and challenging mergers, we make use of the fact that a typical student is an expert on his school environment. By using students at both UCLA and Caltech, we create natural cultural differences (and potentially, conflict). The picture sets contain images of buildings, fountains, and other features of both the UCLA and Caltech campuses. Some of the code groups use has already been created, because there are proper names or slang for most of the buildings and landmarks seen in the pictures. However, separate groups at Caltech and at UCLA do not all use the same code for each picture. For example, the image from the Caltech campus pictured in Figure 2A was coded as "Millikan bridge," "Millikan pond," "arch bridge," and "bridge over pond" by various groups at Caltech.

Within each group, the code used for a picture would evolve over many rounds. For one group at UCLA, the first description of the image in Figure 2B was "rectangle; its a flat building; with one tree; looks like haines; grass in front; its not that hard." [sic, semi-colons indicate separate lines of type] The second description of this image (from a different manager in the group) was shorter: "rectangular with grass in front; looks like castle; two trees." The third description was "rectangle closer up [to contrast with another building described as a rectangle]; with trees." Eventually, the group quickly identified this image as "rect, castle" or simply "castle." Although code varied both among and within groups, it often reflected expertise present at each school that students at the other campus would not have (e.g., a UCLA student would not describe the Figure 2A picture with the name "Millikan," and may not even recognize a bridge over a pond).

The formation of the mixed group in our experiment is like a horizontal merger. The 40-round training phase would allow the "firm" at each school to cement a campus-specific code. Then employees from two distinct firms are forced to work together on a common task. Each person can contribute some expertise, but everyone must learn how to communicate with each other, because the code they were previously able to use freely in their own organization may not be understood by others.

A. Hypotheses

While subjects in the unmixed groups see pictures they are already familiar with, the merged groups face a more difficult task. About half the pictures a member of the merged group sees will be from the other school, and members of this group are expected to have trouble communicating with each other. However, because merged groups only participate in a small number of rounds in the picture-naming task, they are unlikely to gain sufficient experience. Even though members of merged groups will be paid an additional sum of money, we do not think it will compensate for the greater difficulty the merged groups face. This leads to an obvious first hypothesis:

H1: Subjects in the mixed group will earn less than subjects in the unmixed group, even when their additional payments are counted as part of their earnings.

If performance is worse in the mixed group, it is interesting to know whether subjects can correctly anticipate the difficulty the merged group will have. Expectations of post-merger performance are measured in two ways: through the guesses about average earnings in the mixed group and the bids for additional payments to join the mixed group.

If subjects fail to account for the difficulty of the picture naming in the mixed group, guesses of average earnings will be higher than the actual earnings and the bids will be too low. These overvaluations would be akin to a situation in which two firms guess that a merger will be easy and it is actually harder than they expect.

In our design, participants may underestimate the difficulty of the merger because of a judgment bias, the "curse of knowledge," in which people act as if their expertise and knowledge is more widely shared than it is (Camerer, Loewenstein, and Weber 1989). People have also shown a tendency to assume that their strong emotions and deceptions are more transparent to others than they really are (Gilovich, Medvec, and Savitsky 1998; Griffin and Ross 1991).

In our study, participants in the merged group who see a picture from their own campus will generally be inclined to describe it with its proper name or with the short code that had previously been used for that picture. People from the other school should not be expected to know the proper names, but participants may believe that short descriptions will be quickly understood. For example, when a manager from UCLA is looking at a picture of Kerckhoff Hall, which he tries to describe as "brick building," he may fail to realize that the other participants are looking at a series of pictures that contain three different brick buildings. The curse of knowledge, combined with the scramble to create new code, can cause difficulties for group members at the manager's school as well as at the other school.

These ideas lead to a second hypothesis:

H2: Subjects will overvalue the merger, guessing average earnings for the mixed group that are higher than the actual earnings of that group.

The alternative hypothesis to H2 is rationality of expectations: guesses and bids correctly anticipate, in the sense of a zero mean forecast error, the earnings differential between groups.

III. RESULTS

A. Auction Treatments

We first compare data from the first-price and Vickrey sessions. Bids do differ between the two treatments, as predicted by theory. In the first merger stage, Vickrey bids are significantly lower than the first-price bids (Kolmogorov--Smirnov [K--S] test, p = .007). (5) In the second merger stage, there is no significant difference between bids across the two treatments (K--S test, p = .877). As we will see later, bids in the first-price sessions remained stable across the two merger stages, but those in the Vickrey treatment rose significantly between the two stages.

Keep in mind that in first-price low-bid auctions of this type, with six bidders competing for two "objects," equilibrium bidding requires subjects to bid more than their reservation price (trading off the risk of losing the auction for additional surplus), just as in high-bid auctions, bidders should underbid. Most experiments do show strategic bidding of this sort (e.g., Kagel 1995). Extrapolating from these earlier experiments (and theory), it is likely that first-price bids are inflated relative to true reservation prices. The gap between the median Vickrey-auction bid ($.30 in merger phase I) and the median first-price auction bid ($.50 in merger phase I) is consistent with strategic inflation of bids in the first-price auctions.

The Vickrey treatment was added because, in theory, subjects should just bid their guess about the earnings differential--that is, bidding one's valuation is a dominant strategy, so there is no strategic incentive to inflate bids. (6) The fact that bids are lower in the Vickrey auctions than in the first-price auctions suggests that switching auction mechanisms reduces the extent of strategic inflation, which makes the Vickrey data, in theory, a better test of whether participants are underestimating the difficulty of the mergers. We also used an alternative econometric procedure to infer unobserved valuations from actual first-price bids (assuming that subjects are making equilibrium bids; Guerre, Perrigne, and Vuong 2000).

B. Overall Performance

Table 1A and 1B gives average (per period) statistics for performance, guesses, and bids. Figure 3 shows average earnings across sessions, before additional payments for the mixed group are included. The earnings for Caltech and UCLA students are pooled. (7)

During the initial 40 rounds, earnings grew as subjects became faster at identifying pictures and made fewer mistakes. (See Appendix (8) for average completion times and mistakes by round.) From round to round, there is a great deal of variability in subjects' earnings. Often a large dip in earnings was caused by a manager who did not use the established code for a picture or who confused employees by describing several pictures at once. The random draws of pictures each period could also make the task more or less difficult in a given round, because some images are fairly similar and others are distinct. (9) Despite variance in Figure 3, a trend is readily apparent. Earnings increased over time for the unmixed groups, eventually asymptoting at around $0.45 per round. The mixed groups also increased performance over time, nearly reaching the earnings of the unmixed groups by the end of the two merger phases of 20 total rounds.

Figure 4 depicts average code length (the number of characters used by the manager to describe a single picture) across rounds. (10) There are a few interesting features of Figure 4. The first to note is the relationship between increasing earnings and decreasing code length (i.e., the code length in Figure 4 looks like an inverse of the earnings in Figure 3). Groups completed each round quickly once they had established clear, short codes for each picture. Another feature is the similarity of the decrease in code length, for the first mixed group in rounds 41-50, and for the original groups in early rounds 1-10. The mixed group initially used slightly longer descriptions than the original groups at each school had used, but both the mixed and the original group reduced their code to about 20 characters per description by the end of ten rounds. This shows some "learning to learn" across the first and second merger stages.

[FIGURE 3 OMITTED]

[FIGURE 4 OMITTED]

C. First Merger Stage

The mixed groups clearly have more trouble with the picture-naming task than the unmixed groups. They use longer code, take more time, and make more costly mistakes. The next question is whether subjects accurately anticipate the worse performance of the mixed groups. The evidence indicates subjects are generally too optimistic, but the results are somewhat sensitive to the type of bids.

There are three ways to measure whether subjects correctly forecasted the lower earnings in the mixed group--earnings guesses, Vickrey bids to merge, and inferred values for first-price bids to merge. Table 2 summarizes the differences between these statistics and the actual mixed-group earnings. (Negative numbers mean that guesses or bids were too optimistic.) Guesses in the first merger phase were generally too optimistic. Eighty-four percent of the forecast errors (actual earnings minus guesses) were negative, and 69% of the earnings guesses were more than $1 above the mixed group's actual earnings.

Because forecast errors are correlated (due to shared dependence on the mixed-group average earnings), we use a random resampling bootstrapping method to estimate a confidence interval around the mean forecast error. Pooling both auction conditions, the mean error is -$2.42 (for all ten periods), with a 95% confidence interval of [-2.85, -1.99], so the difference between the forecast errors and zero is highly significant.

Analysis of bids is less conclusive, but it is suggestive that bidders who join the mixed group underestimate the earnings differential. One test focuses on mixed-group earnings with bids added in, minus unmixed-group earnings. If this difference is negative, then subjects who make low bids and join the mixed group earn less than their counterparts in the unmixed group.

First, note that there are three possible levels of analysis--taking each session as a data point, taking each group as a data point, and taking each subject as a data point. Because each subject shares a common group experience, their earnings differentials are correlated so that traditional tests assuming independence will overstate significance. Taking each session as a data point creates a small sample but is an overly conservative test. We will generally compromise and report the group-level analysis, noting the least conservative individual analysis and the most conservative session analysis in footnotes. In both auction sessions, there is a clear difference between earnings of unmixed groups and mixed groups before bids are added in, as expected, which is highly significant at all levels of analysis. (11)

The crucial comparison is between mixed-group earnings with bids added in (i.e., the amount the mixed-group subjects actually earned) and unmixed-group earnings. The results differ across the two auction treatments. In the Vickrey treatment, the mixed-plus-bid earnings average $1.17 and unmixed groups average $4.32. The difference is significantly negative for groups ([t.sub.l0] = 2.344, p = .021). (12) In the first-price treatment, the mixed-plus-bid earnings average $4.45 and unmixed groups average $4.21. This small difference is not significant in the group-level analysis ([t.sub.25] = 0.528, p = .301). (13)

However, keep in mind that bids in the first-price treatment are, in theory, strategically inflated. Therefore, even if bidders underestimate the difficulty of merging, they could still show a result in which net earnings for mixed groups (with bids added in) and unmixed groups are close together. We use an econometric procedure to estimate underlying valuations from first-price bids nonparametrically (see Guerre, Perrigne, and Vuong [2000] and Appendix). This procedure basically runs a bidding function--which maps values into bids--in reverse: By assuming independent private values and equilibrium bidding, unobserved values can be guessed from bids. (14) Through estimates of the distribution of bids across all sessions, this technique allows us to generate an estimated value for each bidder. This value should be interpreted as the bidder's guess about the expected cost of joining the mixed group. For the purpose of estimating values, we excluded bids above 20. While the estimated value distribution on the whole looked reasonable, the lowest bidders had some fairly large negative inferred values, which were excluded for finding the cumulative distribution. (15)

The only other experimental application of the procedure (Bajari and Hortacsu 2005) used data in which valuations were known because they were controlled by the experimenter. Their application pretends that the values are not known, inters them econometrically, and compares the inferred values to the actual ones under various models of bidding. For values uniformly distributed in the interval [0, $30], the average estimation error is only $1.39 for a Nash bidding model including risk-aversion, so the procedure shows some promise in this simple case.

The value-inference procedure can be used to answer two questions.

First, are inferred underlying values from first-price bids closer to the Vickrey bids than the first-price bids are? If the subjects randomly sampled into the first-price and Vickrey sessions have the same distribution of underlying values (up to sampling error), and the Vickrey subjects are bidding around their values (as they should in theory), then the inferred-value and Vickrey bid distributions should be closer together than the Vickrey and first-price bids are. Figure 5A shows the three cumulative distribution functions (cdfs) of Vickrey bids, first-price bids, and inferred values, pooling across subjects, for phase I. (The actual cost distribution--that is, the gap between mixed and unmixed earnings--is also plotted as a benchmark.) The value cdf inferred from first-price bids is indeed closer to the Vickrey distribution than the original first-price bids are, though the inferred values are still somewhat higher than the Vickrey bids (K--S test, p = .016). This gives some assurance that the value inference procedure is on the right track.

Second, do the Vickrey bids, and first-price inferred values, of subjects who entered the mixed group (those with the lowest bids) approximate the actual costs of merging? Figure 5B shows a distribution of the lowest two values for six draws from the inferred-value distribution and Vickrey distribution. We are interested in these artificial distributions because even if the mean inferred valuation is approximately correct (as it appears to be from Figure 5A), if the two lowest bidders' valuations are too optimistic, then they will join mixed groups and be surprised at their poor relative performance. Figure 5B shows that there does appear to be such a winner's curse: The two-lowest-value distributions (inferred from first-price bids, and using raw Vickrey bids) are well below the actual cost distribution.

Thus, in the first merger stage, it appears that guesses about mixed-group earnings are generally too optimistic. Furthermore, the average Vickrey bid and the average value inferred from first-price bids are pretty good forecasts of actual costs (the mixed-group minus unmixed earnings difference), but the values of those who joined the mixed group (the low bidders) are too optimistic. It appears that in these simple experiments, subjects have some tendency to be surprised at how difficult assimilation is.

D. Second Merger Stage

In the second merger stage, guesses and bids are more accurate but are also consistent with anchoring on the results of the first merger stage. Earnings in the second merger phase rise steadily across the ten periods. By analogy to actual corporate mergers, there appears to be a learning curve in "learning to merge" and create common code rapidly and accurately. The mixed groups suffered an initial dip in earnings compared with the last round of the previous picture-naming task, but within about four periods the mixed groups surpassed the mixed-group performance in the preceding phase and continued to earn more. However, guesses and bids both suggest that subjects did not anticipate this improvement--they are too pessimistic about mixed-group performance. Two-thirds of the guesses about mixed-group earnings (67.3%) were below the average earnings of the mixed group. The mean difference between average earnings and guesses was $.60 (the 95% confidence interval is [$.36, $.83] based on the bootstrap). The fact that 76% of the guesses were within $1 of the mixed-group previous average earnings (which subjects were told) suggests that subjects are anchoring on the mixed-group experience in the first stage to forecast the mixed group's experience in the second stage.

[FIGURE 5 OMITTED]

Bids in this stage also indicate anchoring. Bids in the first-price auction treatment were statistically similar in both merger stages. (16) The average of the winning bids across the two stages is $3.40. When bids are added in, the mixed groups in the second merger stage received higher earnings than the unmixed groups did ([t.sub.25] = 4.16, p < .001 for the group analysis). (17) Mixed-group members earned an average of $1.17 more than members of the unmixed group.

In the Vickrey auction treatment, subjects greatly increased their bids after the first merger stage. (18) Winning bids rose from an average of $1.69 in the first merger phase to $3.79 in the second phase. When they are included as part of mixed-group members' earnings, these increased bids, as well as an improvement in performance, cause the mixed-group members to make significantly more than members of the unmixed groups in the second phase, at the group level ([t.sub.10] = 4.589, p < .001). (19) Mixed-group members earned an average of $.36 more than the unmixed-group members.

The inferred values from the first-price session nearly match the Vickrey values in this second merger stage (K-S test, p = .165), so it is not surprising that they are significantly higher than the distribution of actual costs (K-S test, p = .000). Figure 5C shows that the bootstrapped distributions of low values for both the Vickrey and first-price treatments are very close to the distribution of actual costs. (20) Thus, while the average subject is too pessimistic about second-stage merger performance, the lowest Vickrey and first-price auction bids (i.e., those subjects most likely to join the merger) are rather accurate.

It is possible that subjects had trouble estimating the performance of the mixed group because they did not know how many of the same people would be returning to the mixed group for the second merger phase. A mixed group that retained its original members could use the code developed in the first merger stage, whereas a group with a lot of new members would need to create a new code. In fact, only one mixed group remained intact in both phases. On average, about 1.8 of the four players in the original mixed group were replaced in the second merger stage. Learning effects were strong despite large turnover in the group membership from the first to the second merger phase.

Taken together, the two merger stages illustrate a common underestimation of the effect of structural changes. In the first stage, the lowest bidders systematically underestimate how much earnings will suffer when groups are mixed. However, in the second stage, they overestimate the earnings decline--or put differently, they underestimate the improvement from the first stage to the second.

IV. CONCLUSION

This study used simple artificial "firms" to study the development of organizational code, and what happens when firms with different codes merge. The firms' task is to name target pictures from a set, by developing natural language descriptions of the pictures--the "code"--which enable an observer to know which of many pictures was the target. Code use of this sort is like an organization where one person sees a situation or object clearly and must convey it to another person (like police dispatching, journalism, or a busy restaurant kitchen). Good code is short and distinctive, because firms are penalized for going slowly and for choosing the wrong pictures. The goal was to learn about development of code, and whether employees could accurately judge how difficult mergers would be because of "cultural conflict" due to differences in code and familiarity. The experiments do not attempt to recreate all the complexities of naturally occurring organizations. We simply tried to take one element of firms, which can be created and measured, and has economic value, and study some of its properties.

The experimental sessions began with students in separate six-person groups on the UCLA and Caltech campuses, creating code for pictures of their own campuses. After a common training phase, the mergers were created by asking subjects how much they would demand to be paid to join a mixed four-person group, with two people from their own-campus group, and two from the opposite campus. This gives us a measure of how well the subjects think the mergers will go, and also sorts them into mixed (i.e., merger) and unmixed groups endogeneously.

Not surprisingly, the mixed groups were slower than the unmixed groups and made more mistakes. Guesses about the performance of the mixed groups, and bids in an incentive-compatible Vickrey auction, underestimated the difficulty of the merger. First-price bids priced the merger difficulty accurately, but are also likely to be biased upward by rational strategic inflation. To correct this, we used an econometric procedure to infer unobserved valuations from bids. The procedure basically takes a bidding function, which maps a value onto a bid, and runs the process in reverse, inferring what unobserved value must have led to an observed bid. The one study that applied this procedure to experimental data (where inferred values and actual values can be compared) found a rather accurate correspondence, after adjustment for risk-aversion. This procedure shows that the values inferred from all first-price bids are rather accurate guesses of mixed group earnings differentials, as are Vickrey bids on the whole. However, the lowest bidders--those who actually join the mixed groups--are systematically too optimistic. We consider this an "organizational winner's curse," in which the most optimistic organization members will be most willing to make structural changes and are too optimistic even when the average person is accurate.

Estimates of the mixed group's earnings switched from being overly optimistic in the first merger stage to being a little too pessimistic in the second merger stage. Estimates seemed to reflect a belief that the mixed group's performance had flattened out and would not improve in the second merger stage, although it did.

Inaccurate guesses in both stages are also consistent with anchoring effects. Participants seemed to extrapolate from the experience in the previous phase to the next phase, even though they were clearly instructed about the structural adjustment between phases. The reversal in the first and second stages also means people are not being irrational or optimistic in general--instead, those who participate in the changes do not have rational expectations about the impact of the change. The fact that subjects are much more accurate in the second stage, however, indicates a process of learning.

A. Future Research

One line of future research is alternative mechanisms for sorting employees into the merged firm, such as requiring employees to pay to stay in an unmixed group (corresponding to having to incur relocation or switching costs to stay in their "old job"). Another line of research is turnover: A merger can be intended to trim duplicated human resources from the two merging firms, but in practice the best workers often leave rather than the worst ones. Structures in which managers can decide who stays or goes and workers can choose to leave could look for such selection effects and their impact.

A common problem in mergers is that employees from one firm are marginalized. One firm often takes the dominant role and attempts to impose its culture on members of the other firm. (See Weber and Camerer 2003, for an example of dominance in the Daimler-Chrysler "merger of equals.") Using the picture-naming paradigm, we could see whether one firm imposes its code on the other when it is larger or has a longer history, and study how it does this. Members of the dominant firm might not explain their code to the new members, or they might train the other firm on their code, as we saw in some sessions involving a person who was unfamiliar with the campus.

Future studies could create firms around a different kind of expertise. The subjects in this study were "experts" on a topic they would not expect members of the other group to know anything about. The curse of knowledge might play a more pronounced role if the original firms are composed of experts on areas that others might also know about, such as artwork or movies. One could also study "hypercode" (how short do codes become with very long training?), and the role of emotions, humor, and memory in code formation and transmission with time interruptions and organizational turnover (cf. Heath and Seidel, 2005).

Any of these ideas could be examined by slightly altering the flexible paradigm used in this study. Although we cannot capture ali the intricacies of culture in simple experiments, this paradigm allowed us to examine one important facet of culture and create precise measures to test our hypotheses on the cultural difficulties involved in mergers.

doi: 10.1111/j.1465-7295.2009.00200.x

REFERENCES

Andrade, G., M. Mitchell, and E. Stafford. "New Evidence and Perspectives on Mergers." The Journal of Economic Perspectives, 15(2), 2001, 103-20.

Arrow, K. J. The Limits of Organization. New York: W.W. Norton and Company, 1974.

Bajari, P., and A. Hortacsu. "Are Structural Estimates of Auction Models Reasonable?: Evidence From Experimental Data." The Journal of Political Economy, 113(4), 2005, 703-41.

Camerer, C., G. Loewenstein, and M. Weber. "The Curse of Knowledge in Economic Settings: An Experimental Analysis." The Journal of Political Economy, 97(5), 1989, 1232-54.

Cremer, J. "Corporate Culture and Shared Knowledge." Industrial and Corporate Change, 2(3), 1993, 351-86.

Cremer, J., L. Garicano, and A. Prat. "Codes in Organizations." CEPR Discussion Paper No. 4205, 2004.

Gilovich, T., V. H. Medvec, and K. Savitsky. "The Illusion of Transparency: Biased Assessments of Others' Ability to Read One's Emotional States." Journal of Personality and Social Psychology, 75(2), 1998, 332-46.

Grether, D. M., C. R. Plott, D. B. Rowe, M. Sereno, and J. M. Allman. "Mental Processes and Strategic Equilibration: An fMRI Study of Selling Strategies in Second Price Auctions." Experimental Economics, 10(2), 2007, 105-22.

Griffin, D. W., and L. Ross. "Subjective Construal, Social Inference, and Human Misunderstanding," in Advances in Experimental Social Psychology, Vol. 24, edited by M. P. Zanna. San Diego, CA: Academic Press, 1991, 319-59.

Guerre, E., I. Perrigne, and Q. Vuong. "Optimal Nonparametric Estimation of First-Price Auctions." Econometrica, 68(3), 2000, 525-74.

Heath, C., and V. Seidel. "Language as a Coordinating Mechanism: How Linguistic Memes Help Direct Appropriate Action." Stanford working paper, 2005. http://www.si.umichi.edu/ICOS/Linguisticmemes4.2. pdf (accessed 07/22/2009).

Kagel, J.H. "Auctions: A Survey Of Experimental Research," in The Handbook of Experimental Economics, edited by J. H. Kagel and A. E. Roth. Princeton, NJ: Princeton University Press, 1995, 501-85.

Klein, A. Stealing Time. New York: Simon and Schuster, 2003.

Kreps, D. M. "Corporate Culture and Economic Theory," in Perspectives on Positive Political Economy, edited by J. E. Alt and K. A. Shepsle. Cambridge: Cambridge University Press, 1990, 90-143.

Li, T., I. Perrigne, and Q. Vuong. "Structural Estimation of the Affiliated Private Value Auction Model." The RAND Journal of Economics, 33(2), 2002, 171-93.

Milgrom, P. R., and R. J. Weber. "A Theory of Auctions and Competitive Bidding." Econometrica, 50(5), 1982, 1089-122.

Ravenscraft, D. J., and F.M. Scherer. "Life After Takeover." The Journal of Industrial Economics, 36(2), 1987, 147-56.

--. "The Profitability of Mergers." International Journal of Industrial Organization, 7, 1989, 101-16.

Rick, S., R. Weber, and C. F. Camerer. "Knowledge Transfer in Simple Laboratory Firms: The Role of Tacit vs. Explicit Knowledge." Working paper, 2007. http:// opim.wharton.upenn.edu/~srick/KnowledgeTransfer 7-5-07.doc (accessed 07/22/2009).

Rose, F. "The Civil War Inside Sony." Wired Magazine, Issue 11.02, 2003.

Schein, E. H. Organizational Culture and Leadership. San Francisco: Jossey-Bass, 1985.

--. "Organizational Culture." The American Psychologist, 45(2), 1990, 109-19.

Smeets, V., K. Ierulli, and M. Gibbs. "Merger of Equals and Unequals." Working paper, 2007. http://www.valeries meets.com/research.html (accessed 04/15/2008).

Steels, L. "Evolving Grounded Communication for Robots." Trends in Cognitive Sciences, 7(7), 2003, 308-12.

Sudarsanam, P. S. The Essence of Mergers and Acquisitions. London: Prentice Hall, 1995.

Walsh, J. P. "Top Management Turnover Following Mergers and Acquisitions." Strategic Management Journal, 9, 1988, 173-83.

Walsh, J. P., and J. W. Ellwood. "Mergers, Acquisitions, and the Pruning of Managerial Deadwood." Strategic Management Journal, 12, 1991, 201-17.

Weber, R. A., and C. F. Camerer. "Cultural Conflict and Merger Failure: An Experimental Approach." Management Science, 49(4), 2003, 400-15.

Wernerfelt, B. "Organizational Languages." Journal of Economics and Management Strategy, 13(3), 2004, 461-72.

(1.) This type of survey evidence is subject to criticisms of all survey responses, but it is the best available data from samples of working managers directly involved in mergers.

(2.) Our experiment has several innovations beyond Weber and Camerer's. They used a single set of pictures so differences in group code were modest. By using two groups of students from different campuses, we generated a more natural and large variation in code and expertise. They forced mergers while our mergers are endogenous. We allowed self-selection into the merged firms, which enables us to detect an optimistic bias in bids and a winner's curse among low bidders. Our computerized interface also allows more precise measurement of time and code language.

(3.) The experiments used CultureX software, developed by Charlie Hornberger and John Lin. Contact the authors for documentation on this software.

(4.) If subjects in the mixed group were also paid for accurate guesses, they would have a marginal incentive to guess very low amounts and then perform poorly. Not paying the mixed-group members for their guess accuracy removes this distortion.

(5.) Because some bids are very large, a nonparametric test is used to compare bidding behavior, rather than determining an arbitrary cut-off point to exclude outliers. The Kolmogorov--Smirnov test will also be used for other bid comparisons in the paper.

(6.) In experiments, subjects sometimes inflate bids strategically even in Vickrey auctions, where there is no incentive to do so. Kagel (1995) reviews studies in which subjects do not bid their values in second-price auctions. Grether et al. (2007) provide neural evidence on the difficulty people have in finding the dominant strategy in these auctions.

(7.) Although Caltech students earned more than UCLA students, no interaction was found between being a Caltech student and being in the mixed group. Because the differences between mixed and unmixed groups do not vary by school, we pool data here for ease of exposition.

(8.) The Appendix can be found at http://www.hss.caltech. edu/~camerer/camerer.html.

(9.) Software problems could also cause low earnings. For each of the eight rounds in which we were aware that a software glitch occurred, we interpolated data on earnings, completion time, and incorrect guesses from the two surrounding rounds.

(10.) One outlier was excluded from the analysis. In the first round of the second merger stage, one manager of a mixed group typed advice to the employees before starting to describe the pictures, causing that group to have a large number of characters per picture. With the outlier included, the average for the mixed group in round 5l is 49.9, and without the outlier, it is 34.1.

(11.) The results with individuals as the unit of analysis are [t.sub.l06] = 13.208, p < .001 in first-price treatment, and [t.sub.46] = 9.166, p < .001 in Vickrey. The results with groups as the unit of analysis are [t.sub.25] = 8.100, p < .001 in first-price and [t.sub.l0] = 4.680, p < .001 in Vickrey sessions. The results with session-level analysis are [t.sub.8] = 5.780, p < .001 in first-price treatment and [t.sub.3] = 3.304, p = .023 in Vickrey treatment.

(12.) In the Vickrey sessions, a subject in the mixed group would earn an additional amount greater than or equal to her bid, but her bid reflects her valuation of the merger, so the sum of performance-based earnings and bids is the proper measure to use here. The individual analysis result is [t.sub.46] = 4.623, p < .001; the session analysis result is [t.sub.3] = 1.632, p = .101 for the one-tailed, one-sample test.

(13.) The individual analysis gives [t.sub.106] = 0.806, p = .211 and the session analysis gives [t.sub.8] = 0.360, p = .364 for the one-sample, one-tailed test.

(14.) The procedure assumes independent values but values are estimates that are likely to be affiliated (Milgrom and Weber 1982). Li, Perrigne, and Vuong (2002) develop a procedure to infer affiliated values but it requires more data than we have and is beyond the scope of this experiment (especially given that we use Vickrey auctions which, in theory, should reveal values.) Furthermore, a regression of individual first-stage bids on group dummy variables gave only weak results (F = 1.98, p = .06), which indicates that the bids within a group are not highly clustered (as one would expect them to be if values are affiliated).

(15.) A negative inferred value implies that a bidder would pay a substantial sum to be in the mixed group. Because no one in the Vickrey sessions bid a negative amount, we felt comfortable excluding these values. It is known that the density estimator is biased near the boundaries of its support, and Guerre, Perrigne, and Vuong (2000) provide a formula for trimming. This formula does not work for our data, though. See Appendix for details.

LAUREN FELLER and COLIN F. CAMERER *

* This research was supported by an NSF and Betty and Gordon Moore Foundation grant to Colin Camerer. We are grateful to Galen Loram, Ming Hsu, Joseph Wang, and programmers Charlie Hornberger and John Lin for going beyond the call of duty to make sure this project worked. We also benefited from helpful discussions with Preston McAfee, Simon Wilkie, Jacob Goeree, and Paul Healy, two anonymous referees, and comments from audiences at Caltech, Chicago GSB, ESA. and the Behavioral Organizational Economics meeting at MIT.

Feiler: Assistant Professor, Department of Economics, Carleton College, Northfield, MN 55057. Phone 507-222-4119, Fax 507-222-4044, E-mail lfeiler@carleton.edu

Camerer: Rea A. and Lela G. Axline Professor of Business Economics, Division of Humanities and Social Sciences, California Institute of Technology, Pasadena, CA 91125. Phone 626-395-4054, Fax 626-793-4681, E-mail camerer@hss.caltech.edu

TABLE 1A
Summary Statistics for First-Price Sessions, per Period

 Average
 Average Number
 Earnings of
 Group ($) Mistakes

Rounds 1-40 All 0.246 0.29
 (n = 108) (0.475) (0.79)
Merger phase 1 Unmixed 0.421 0.13
 (n = 72) (0.206) (0.38)
 Mixed 0.106 0.40
 (n = 36) (0.468) (0.86)

Merger phase 2 Unmixed 0.470 0.06
 (n = 72) (0.130) (0.24)
 Mixed 0.246 0.24
 (n = 36) (0.339) (0.66)

 Average Median
 Completion Guesses
 Group Time (See) ($)

Rounds 1-40 All 42.1
 (n = 108) (32.6)
Merger phase 1 Unmixed 22.7 0.313
 (n = 72) (11.1) (0.249)
 Mixed 58.0 0.288
 (n = 36) (30.5) (0.171)
 All: 0.300
 (0.225)
Merger phase 2 Unmixed 19.7 0.138
 (n = 72) (7.6) (0.139)
 Mixed 48.6 0.150
 (n = 36) (24.8) (0.158)
 All: 0.150
 (0.145)

 Median
 Median Inferred
 Bids (a) Value ($)
 Group ($) (b)

Rounds 1-40 All
 (n = 108)
Merger phase 1 Unmixed 0.600 0.490
 (n = 72) (0.482) (0.495)
 Mixed 0.300 0.025
 (n = 36) (0.169) (0.191)
 All: 0.500 All: 0.381
 (0.461) (0.397)
Merger phase 2 Unmixed 0.525 0.445
 (n = 72) (0.347) (0.401)
 Mixed 0.315 0.045
 (n = 36) (0.183) (0.205)
 All: 0.500 All: 0.415
 (0.342) (0.310)

Note: Standard deviations are in parentheses.

TABLE 1B
Summary Statistics for Vickrey Sessions, per Period

 Average
 Number
 Average of

 Group Earnings ($) Mistakes

Rounds 1-40 All 0.210 0.29
 (it = 48) (0.161) (0.22)
Merger phase 1 Unmixed 0.432 0.07
 (n = 32) (0.049) (0.09)
 Mixed -0.052 0.56
 (n = 16) (0.294) (0.45)
 Third-lowest
 (n = 8)

Merger phase 2 Unmixed 0.463 0.04
 (n = 32) (0.036) (0.06
 Mixed 0.120 0.37
 (n = 16) (0.206) (0.27)
 Third-lowest
 (n = 8)

 Average Median
 Completion Guesses
 Group Time (Sec) ($)

Rounds 1-40 All 43.3
 (it = 48) (12.3)
Merger phase 1 Unmixed 23.5 0.258
 (n = 32) (4.0) (0.243)
 Mixed 54.9 0.355
 (n = 16) (11.3) (0.166
 Third-lowest
 (n = 8)
 All: 0.291
 (0.219)
Merger phase 2 Unmixed 21.2 0.150
 (n = 32) (3.6) (0.203)
 Mixed 46.4 0.108
 (n = 16) (7.7) (0.177)
 Third-lowest
 (n = 8)
 All: 0.135
 (0.194)

 Median
 Group Bids' ($)

Rounds 1-40 All
 (it = 48)
Merger phase 1 Unmixed 0.480
 (n = 32) (1.016)
 Mixed 0.138
 (n = 16) (0.122)
 Third-lowest 0.313
 (n = 8) (0.162)
 All: 0.350
 (0.902)
Merger phase 2 Unmixed 0.575
 (n = 32) (0.576
 Mixed 0.300
 (n = 16) (0.224)
 Third-lowest 0.448
 (n = 8) (0.228)
 All: 0.500
 (0.524)

Note: Standard deviations are in parentheses.

(a) Outlying bids above the 97th percentile are set to the highest
bid below this percentile for the purpose of computing standard
deviations.

(b) Only values between 0 and 20 are used for the purpose of
computing standard deviations.

TABLE 2
Summary of Per-Period Average Forecast Errors and Earnings
Differentials

 Mean Forecast
 Error (SE)

Merger Auction Mixed Earnings
Phase Type - Guesses

1 First price -0.194 (0.024)
1 Vickrey -0.343 (0.053)
2 First price 0.096 (0.013)
2 Vickrey -0.015 (0.019)

 Mean Net Earnings Mean Inferred
 Differential (SE) Net Earnings
 Differential (SE)

 (Bid + Mixed (Inferred Value (a)
Merger Earnings) - + Mixed Earnings)
Phase Unmixed Earnings - Unmixed Earnings

1 0.024 (0.024) -0.147 (0.263)
1 -0.315 (0.054)
2 0.117 (0.021) -0.036 (0.234)
2 0.036 (0.012)

Note: Errors should be negative if subjects are too optimistic and
positive if they are too pessimistic. (a) Negative values are set
to 0.