Unit 2

THE PROBLEMS AND PROMISE OF ENTREPRENEURIAL PARTNERSHIPS: DECISION-MAKING,

OVERCONFIDENCE, AND LEARNING IN FOUNDING TEAMS

JOHN S. CHEN University of Florida

DANIEL W. ELFENBEIN Washington University in St. Louis

HART E. POSEN University of Wisconsin–Madison

MING ZHU WANG Washington University in St. Louis

How should decision-making be organized in entrepreneurial teams when founders exhibit confidence biases? New ventures are commonly founded by teams of entrepre- neurs, who must employ a decision-making structure that implicitly or explicitly defines how individual beliefs are aggregated into team decisions. We consider this issue through the lens of organizational economics, which focuses on decision-making governance. Using a computational model, we consider three archetypal decision-making structures: partnership voting, a boss with employees, and a buyout option (partnership convertible to boss structure). We highlight the conditions under which partnership voting is an effec- tive means of governing market entry and exit decisions when teams’ decision-making is informed by efforts to learn about the merits of uncertain opportunities. The promise of partnership voting is realized when entrepreneurs are either unbiased or optimistic about their likelihood of success. Partnership voting is problematic when entrepreneurs differ in their biases or respond too rapidly to new information, in which case a buyout option is better. From a policy perspective, we show that confidence biases may be man- aged by selectively matching the decision-making structure to entrepreneurs’ biases, and that doing so may substantially improve the performance of new ventures.

When a new venture is founded by a team of entre- preneurs, how should they organize decision-making? A founding team must employ a decision-making structure that implicitly or explicitly defines how individuals’ beliefs about, for instance, market entry and exit are aggregated into team decisions. Should the founders be voting partners, jointly responsible for critical decisions, or should one of the founders be the boss? How should the decision-making

structure change when the founders exhibit confi- dence biases? Although it has long been known that new ventures are frequently founded by teams (e.g., Cooper, Woo, & Dunkelberg, 1989), the question of how decision-making should be organized remains largely unaddressed in the extant literature, which has instead focused on issues such as team formation, composition, and affect.1 To shed light on these

Equal coauthors listed in alphabetical order. We thank Alphonso Gambardella, Tarek Ghani, Lamar Pierce, Eric Van Den Steen, seminar participants at the Bocconi Uni- versity, the University of Minnesota, conference partici- pants at the Strategy Science, AOM, and SMS 2019, and Associate Editor Joseph Mahoney for their helpful insights in developing this paper.

1 For recent reviews of these bodies of research, see the following: process of team formation (Lazar, Miron- Spektor, Agarwal, Erez, Goldfarb, & Chen, 2020); equity division, autonomy, and entitativity (Knight, Greer, & de Jong, 2020); knowledge, experience, and relational resour- ces of team members (Ganco, Honor�e, & Raffiee, 2019); and goal setting, demographic diversity, and affect (Klotz, Hmieleski, Bradley, & Busenitz, 2014).

489

Copyright of the Academy of Management, all rights reserved. Contents may not be copied, emailed, posted to a listserv, or otherwise transmitted without the copyright holder's express written permission. Users may print, download, or email articles for individual use only.

r Academy of Management Review 2022, Vol. 47, No. 3, 489–520. https://doi.org/10.5465/amr.2019.0119

issues, we view founding teams through the lens of organizational economics, which focuses on the design of decision-making governance (e.g., Klein, Mahoney, McGahan, & Pitelis, 2019). We examine the conditions under which partnership voting is an effec- tive means of governing market entry and exit deci- sions when entrepreneurs exhibit confidence biases. Our key insight is that the structure of decision- making must follow from the characteristics of entre- preneurs’ biases. The promise of partnership voting is realized when entrepreneurs are optimistic about their likelihood of success, but partnership voting is prob- lematic when entrepreneurs differ in their biases or update their beliefs too rapidly in response to new information.

In the popular imagination, entrepreneurial teams are often viewed as decision-making partnerships. Familiar examples include Gates and Allen at Micro- soft or Jobs and Wozniak at Apple. Entrepreneurial opportunities are highly uncertain “unique business investments for which it is difficult … to assign meaningful probabilities to outcomes … [therefore] individuals will reach different decisions, even if they share the same objectives” (Foss & Klein, 2012: 78), and failure is highly likely (Kerr & Nanda, 2010; Knott & Posen, 2005). In many cases, the decisions that arise from two heads may be better than those arising from just one. Recent empirical evidence, however, suggests that partnership decision-making may be less common than we might expect.2 Data from a nationally representative survey of the U.S. population found that only 18% of new firms are governed by equity partnerships (Lee, 2020), and firms with two to four owners represent a mere 33% of all private businesses in operation for two years or less (United States Census Bureau, 2016). This may be explained, in part, by the fact that partnerships can be difficult to manage and sustain for a variety of interpersonal reasons (Klotz et al., 2014). As we argue in this paper, it might also be the case that the governance of entrepreneurial decision-making is not always effectively managed by a pure partner- ship decision-making structure. Indeed, while we may think of partnership voting as the default struc- ture, the relative paucity of partnerships suggests

that there are costs of partnerships that are not fully understood.

To advance our understanding of how alternative decision-making structures impact critical decisions such as entry and exit and the performance of found- ing teams, we must account for the defining features of entrepreneurship. These features include the exis- tence of substantial uncertainty about the merits of opportunities, the need to learn about opportunities by gathering additional information, and the pres- ence of behavioral biases that hamper learning and decision-making. We develop a computational model that builds on the entrepreneurial learning framework of Chen, Croson, Elfenbein, and Posen (2018), which conceptualized entrepreneurship as an unfolding feedback-learning process through which teams discover the viability of their opportu- nities and make market entry and exit decisions. In our model, a two-person team is endowed with an entrepreneurial idea of uncertain merit. Each team member is essential to implementing the idea, in the sense that each supplies unique human capital that is critical for operating the business. At all times, these agents maximize their individual profits con- ditional on their beliefs, but their beliefs may be inaccurate due to bias. Entry into the marketplace is costly for both individuals, so, prior to making an entry decision and investing substantial resources, they seek to learn more about the merits of the oppor- tunity (e.g., engaging focus groups or developing a prototype). This feedback is subject to uncertainties that make learning difficult, and each entrepreneur may interpret the feedback somewhat differently. At the end of this pre-entry learning period, the team must decide whether to enter the market. If they do choose to enter, they will continue to learn about their opportunity through the profits and losses their firm accrues, and, if they come to conclude that they entered erroneously, they may decide to exit.

We consider three archetypal and practically grounded decision-making structures for founding teams of two individuals. These structures corre- spond to the classic decision-making structures examined in Sah and Stiglitz’s (1985, 1986) theory of economic organization and recent extensions of this theory in the management literature (Christensen & Knudsen, 2010; Csaszar, 2013; Csaszar & Eggers, 2013). First, a partnership voting structure is the case in which both team members split post-entry returns and must agree on entry and continuation decisions. Second, a boss-employee arrangement is the case in which a single agent—that is, the boss— has full decision rights. Third, a buyout option is the

2 Note that we use the term “partnership” not to refer to the legal form (e.g., a partnership vs. a corporation), but, rather, to refer to an arrangement by multiple parties to manage a business, regardless of its legal form. Moreover, the absence of a legal partnership need not preclude voting-type decision-making in a team that includes non- equity owners.

490 Academy of Management Review July

case in which the pair begins learning about the opportunity as an equal partnership. If both individ- uals agree about entering the market, the structure remains a partnership post-entry, but, if there is dis- agreement, a pro-entry entrepreneur may buy out the other individual, such that the governance structure changes to a boss-employee arrangement post-entry.

Our central concern is how confidence biases moderate the effectiveness of alternative team decision-making structures. A critical challenge in investigating the learning process of entrepreneurs is that they can be subject to decision-making biases such as overconfidence (Åstebro, Herz, Nanda, & Weber, 2014; Cain, Moore, & Haran, 2015; Chen et al., 2018; Wu & Knott, 2006). These biases may underlie common mistakes such as excess entry into the mar- ket (Camerer & Lovallo, 1999; Koellinger, Minniti, & Schade, 2007) and delayed exit from the market when a new venture’s prospects are poor (Åstebro, Jeffrey, & Adomdza, 2007; DeTienne, Shepherd, & De Castro, 2008; Elfenbein, Knott, & Croson, 2017; Gimeno, Folta, Cooper, & Woo, 1997). In particular, we follow Moore and Healy (2008) in examining both “estimation bias” (overly optimistic or pessi- mistic beliefs) and “precision bias” (over- or under-responsiveness of individuals’ beliefs to new information).

A core insight from our model is that confidence biases may be managed by selectively matching the decision-making structure to the biases of the entre- preneurs involved in the venture. Traditional policies for mitigating bias are based on screening out the most biased individuals (e.g., Elfenbein et al., 2017; Gutier- rez, Åstebro, & Obloj, 2020), de-biasing them via edu- cation (e.g., Elfenbein & Knott, 2015; Lovallo & Sibony, 2010; Shane, 2008; Shepherd & Patzelt, 2017), or structuring the processes through which they learn (Camuffo, Cordova, Gambardella, & Spina, 2020). Our model suggests a complementary approach: selecting the decision-making structure that most effectively ameliorates the consequences of bias and identifies the structures that perform best when the potential biases of entrepreneurs are known (e.g., via diagnostic testing) or when only the population distribution of biases is known.

We highlight three results related to the promise and problems of a partnership decision-making structure relative to the boss-employee or buyout option arrange- ments. First, when prospective entrepreneurs are unbi- ased, the partnership structure, in which both founders must agree on both entry and continuation decisions, yields better results than the alternative decision- making structures. With unbiased team members,

partnership effectively balances errors of omission and commission in entry and exit decisions. Teams of unbi- ased agents with high-quality opportunities fail to enter the market at a moderate rate under a partnership, while those with low-quality opportunities are much less likely to enter than with alternative structures. Addi- tionally, when an unbiased team with a low-quality opportunity mistakenly enters, the partnership struc- ture leads to more rapid exit.

Second, when prospective entrepreneurs come from populations with a distribution of “neutral” biases (i.e., individuals are unbiased on average, although each individual may be biased), the buyout option structure provides better results than does partnership. The strength of the buyout option in this case is that it enables a team to overcome the problem of a single pessimistic partner who can uni- laterally veto entry, leading to fewer “failures to enter” for teams with high-quality opportunities. This “mistaken nonentry” results in irreversible foregone gains, while entry mistakes by teams with low-quality opportunities, the vulnerability of the buyout option, can still be mitigated via exit. When we relax the assumption of neutral biases, however, the relative effectiveness of alternative structures changes: partnership generates greater returns in a population of optimists, as its structure reduces entry of low-quality enterprises (that individual optimists would launch). Furthermore, the length of time allotted for pre-entry learning can change the rank order of the efficacy of these decision-making structures. For instance, when potential entrepre- neurs are, on average, optimistic about the quality of their opportunities, buyout can be the best structure when the pre-entry learning period is short. Yet, for longer learning periods, partnership is best, and boss-employee also outperforms the buyout option. Taken together, these first two insights show the contingent nature of commitment (partnerships) and flexibility (buyout option) among possibly biased entrepreneurial teams.

Finally, decision-making structures impact the dynamics of entry and exit patterns. When entrepre- neurs are drawn from a population characterized by a distribution of biases, the decision-making struc- tures shape both the types of biases exhibited by firms that choose to enter the market and entrants’ performance. The partnership structure may appear to be more successful from the perspective of an econometrician who can only observe firms that have elected to enter the market, even when the buy- out structure is superior in expectation (i.e., at the start of the learning process). Furthermore, the

2022 Chen, Elfenbein, Posen, and Wang 491

model predicts that firms with partnership struc- tures will be less likely to enter, so they will be underrepresented among entrants relative to their frequency among the set of teams that start learning about an opportunity. This observation, along with the other problems related to partnership outlined above, suggests a potential explanation for the pau- city of partnerships. Moreover, our model yields testable predictions that boss-employee founders will exhibit both greater overestimation and overpre- cision biases than founders in partnerships, and that both estimation and precision biases in partnerships will be positively correlated among founders.

THEORETICAL BACKGROUND

Scholars of entrepreneurship increasingly recog- nize that new ventures are often founded by teams rather than lone entrepreneurs (Klotz et al., 2014). Beckman (2006), for example, constructed a sample of 173 high-tech firms in Silicon Valley founded between 1982 and 1995, and discovered that only 18 had been launched by solo entrepreneurs. Foss, Klein, Kor, and Mahoney (2008: 74) focused their inquiry on entrepreneurial judgment, which is neces- sarily “influenced by the composition and … dynam- ics of the entrepreneurial management team.”3

Taking a “team perspective,” then, is critical to deep- ening our knowledge of entrepreneurial dynamics.

Once entrepreneurial endeavors are conceptual- ized as team (rather than individual) pursuits, sev- eral new and distinct theoretical questions emerge. For example, how do entrepreneurial teams form? Do different formation processes lead to systematic differences in team composition (e.g., Agarwal, Campbell, Franco, & Ganco, 2016; Aldrich & Kim, 2007)? Does similarity between team members— whether demographic, experience related, or in social ties—enhance or reduce new venture perfor- mance (e.g., Ganco, Campbell, & Raffiee, 2016; Zenger & Lawrence, 1989)? Similarity between team members may facilitate communication and execu- tion, but at the expense of searching for novel solu- tions (e.g., Beckman, 2006; West, 2007); it may also reflect redundant external networks (e.g., Beckman, Burton, & O’Reilly, 2007). Another important ques- tion arises regarding performance: do solo-founded firms outperform team-founded ones? Greenberg

and Mollick (2018), for example, found that solo ventures involved in crowdfunding last longer than team-founded ones; however, as much literature on persistence in entrepreneurship suggests, survival does not necessarily equate with success.

While previous work has contributed greatly to our understanding of how founding team characteristics and size relate to entrepreneurial performance, it has not examined what we believe to be a critical feature of entrepreneurial teams: the structure of decision- making inside the organization. Given the empirical nature of much prior research, this is not wholly sur- prising, as decision rights are not systematically reported by organizations. We assert that theory related to the structure of decision-making is particu- larly valuable in this setting given the paucity of reported data about these structures, both at a nascent (pre-entry) phase and subsequent to entry. Conceptu- ally, decision-making can be concentrated in the hands of a few (or one) or it can be more decentralized through the use of a voting rule that demands, for example, plurality or consensus. These different decision-making structures aggregate individuals’ het- erogeneous information, and potentially conflicting objectives, differently. As we argue below, the process through which information is aggregated and deci- sions are made critically impacts how team character- istics translate into organizational performance.

Entrepreneurial Decision-Making Structures and Learning

While entrepreneurship scholars have not exten- sively examined the decision-making structures of new ventures, scholars of organizations have long been interested in team decision-making, and, in par- ticular, in how organizational decision-making struc- tures shape the properties of decisions. Underlying this view is the classical conception of the organiza- tion as an information processing system (Galbraith, 1974; March & Simon, 1958). Indeed, for Simon (1947/1997: 19), information processing is the basis for the existence of the organization, reflecting “the pat- tern of communications and relations among a group of human beings, including the processes for making and implementing decisions.” More recently, there has been renewed interest in the aggregation of distrib- uted information into a single organization-level deci- sion, due to the popularity of concepts such as the wisdom of crowds (e.g., Page, 2008).

A salient question in the organizational context is “how does organizational design affect decision- making?” (Christensen & Knudsen, 2010: 71). Sah

3 Foss et al. (2008) focused on heterogeneous mental models among team members, whereas we focus our atten- tion on heterogeneous initial beliefs, signals, and informa- tion processing.

492 Academy of Management Review July

and Stiglitz (1985, 1986) developed a formal treat- ment of decision screening in organizations. In their model, individuals may be organized into different decision-making structures. A project is revealed to the organization, and a decision is made to accept or reject the project based on rules of the particular decision-making structure that aggregates individu- als’ “screening functions” (individuals’ beliefs about the merits of the project) into an organizational deci- sion. Sah and Stiglitz (1985, 1986) explicitly con- trasted hierarchies with polyarchies, which are related to “AND” and “OR” logical functions. The objective in this model is to balance errors of com- mission against errors of omission in a manner that maximizes performance. Christensen and Knudsen (2010) extended this model to the full range of orga- nizational structures between hierarchy and poly- archy. Csaszar and Eggers (2013) examined the optimal choice of structure, given properties of the external environment and knowledge of individuals. Piezunka, Aggarwal, and Posen (2020) extended the model to include individual learning.

Applying the logic of the Sah and Stiglitz (1985, 1986) model to issues of entrepreneurship requires further articulation of the theory to account for the process of learning about an opportunity that pre- cedes decision-making. The Sah and Stiglitz (1985, 1986) model focuses on one-shot, or independent, screening decisions in which there is no individual learning about the merits of the alternatives being screened. With the exception of recent work by Pie- zunka et al. (2020), the literature building on Sah and Stiglitz (1985, 1986) has not considered the role of ongoing feedback in shaping team-level outcomes.

Yet, as we think of it, learning is fundamental to entrepreneurship. Entrepreneurs engage in a process of learning over time about the merits of an opportu- nity because such opportunities are often highly uncertain (e.g., Foss & Klein, 2012; Kerr, Nanda, & Rhodes-Kropf, 2014; Knight, 1921; Packard, Clark, & Klein, 2017; Wu & Knott, 2006), and each entrepre- neur may interpret the feedback somewhat differ- ently (Foss & Klein, 2012). The importance of learning has been increasingly recognized in the liter- ature (e.g., Minniti & Bygrave, 2001; Woo, Daellen- bach, & Nicholls-Nixon, 1994), although, as Cope (2005: 373) noted, “in terms of theory building, many aspects of entrepreneurial learning remain poorly understood.” Extant research on entrepreneurial learning can be classified on the basis of whether learning occurs pre-entry (e.g., Elfenbein et al., 2010; Moeen & Agarwal, 2017; Wu & Knott, 2006) or post-entry (e.g., Frank, 1988; Jovanovic, 1982; Parker,

2006). Thus, the learning process starts prior to the market entry decision. Before substantial entry costs are incurred, be they opportunity or financial costs, entrepreneurs gather data about the merits of the opportunity. This learning underpins the critical entry/termination decision. If the decision is made to pay the cost to enter the market, learning continues as market-based feedback (profits and losses) is ongo- ing and exit/continuation decisions are made repeat- edly over time. The process of learning integrates the pre- and post-entry periods because what is learned pre-entry influences both which firms enter and their decisions post-entry (Chen et al., 2018).

The discussion of learning above is silent on how decisions are made when the new venture is founded by a team of entrepreneurs. Consider an entrepreneurial team consisting of two individuals who learn about the merits of an opportunity from independent sources pre- and post-entry. At the time when entry/termination and exit/continuation decisions are made, the theoretical logic outlined by Sah and Stiglitz (1986) suggests a conceptually sim- ple mapping to three ideal types of decision-making structures in entrepreneurial ventures. First, an equal partnership voting structure has characteris- tics of hierarchy as outlined in Sah and Stiglitz (1986), as both entrepreneurs must agree to enter the market and, post-entry, they must both agree to remain in the market (i.e., not exit). In contrast, the boss-employee structure gives one individual fiat in decision-making. Intermediate to these two func- tions is a buyout option structure. The buyout option structure starts with a partnership but can convert to a boss-employee structure at the time of the entry decision. If team members disagree on the merits of an opportunity, the positively inclined entrepreneur buys out the negatively inclined entrepreneur by paying them at least their opportunity costs, effec- tively hiring them as an employee, and, in doing so, assumes full decision rights regarding market entry and exit. The boss-employee and buyout option structures share characteristics of the polyarchy out- lined in Sah and Stiglitz (1986) because only the consent of one individual is needed to enact a deci- sion. Of course, real-world decision-making struc- tures may deviate from these ideal types.

Confidence Biases in Learning and Decision-Making

The entrepreneurship literature has identified confidence bias as a key mechanism underlying two well-known empirical regularities associated with

2022 Chen, Elfenbein, Posen, and Wang 493

erroneous decision-making. The first is that of excess entry, such that too many entrepreneurs enter the market (e.g., Camerer & Lovallo, 1999; Koellinger et al., 2007), and the second is that of delayed exit, such that entrepreneurs persist in the market even when faced with strong evidence that they should terminate operations (Åstebro et al., 2007; DeTienne et al., 2008; Gimeno et al., 1997). More generally, overconfidence in one form or another has come to be viewed as a defining trait of entrepreneurs. In their pioneering survey of 3,000 entrepreneurs, Coo- per, Woo, and Dunkelberg (1988) reported that approximately half of entrepreneurs included in their study believed their success likelihood to exceed 90%, while, at the same time, they believed that most others had a much lower likelihood of suc- cess. Busenitz and Barney (1997: 10) compared a sample of entrepreneurs to a sample of managers and found that their “overconfidence and represen- tativeness variables correctly categorized entrepre- neurs and managers more than 70% of the time.” Wu and Knott (2006) studied entry into commercial banking and highlighted the distinction between overconfidence in the face of uncertainty about mar- ket demand, versus one’s own ability. Other manage- ment research focuses on types of confidence biases that may have important implications for entrepre- neurial decisions and outcomes (e.g., Camerer & Lovallo, 1999; Dushnitsky, 2010; Hayward, Shep- herd, & Griffin, 2006; Lowe & Ziedonis, 2006; San- dri, Schade, Musshoff, & Odening, 2010).4

We consider behavioral bias in the context of the entrepreneurial learning processes outlined above. Following Cyert and DeGroot (1974), we conceptual- ize a learning process as one of Bayesian belief updating. Learning results from feedback signals due to actions taken by firms, be it pre-entry via the development of a prototype, market research, or talk- ing to potential customers, or post-entry from prod- uct sales. These signals are inherently noisy, causing uncertainty to diminish rather slowly over time as additional feedback is garnered. While one can

employ the adaptively rational Bayesian formulation of Cyert and DeGroot (1974), it is also possible to relax these assumptions and explicitly introduce behavioral bias into the learning process. Bias may be manifest both during the pre-entry period, such that it impacts the entry/termination decision, and post-entry, such that it impacts the exit/continuation decision.

Recently, psychology researchers have begun to recognize that the overconfidence construct has mul- tiple dimensions and that these dimensions impact learning in different ways. The literature distin- guishes between (at least) two forms of overconfi- dence: “overestimation,” which may manifest as beliefs that the likelihood of success is higher than it actually is, and “overprecision,” which may manifest as too-narrow confidence intervals around a given set of beliefs (Moore & Healy, 2008).5 Empirical research has sometimes confounded these forms of overconfidence. For example, a study with a common research design that seeks to elicit data about over- confidence asks respondents how confident they are that they picked the correct answer, in which case “overestimation and overprecision are one and the same” (Moore & Healy, 2008: 503). We assert that dis- ambiguating these distinct forms of overconfidence is essential to understanding the implications of alter- native decision-making structures for performance.6

In the context of learning, overestimation and overprecision are clearly distinct constructs that map onto well-known characteristics of entrepre- neurs. Overestimation reflects a positive bias in an entrepreneur’s initial beliefs about the merits of an opportunity (and underestimation reflects a negative bias). Such an entrepreneur would start the learning process believing the opportunity is better than it is in reality. All else being equal, such an entrepreneur is more likely to enter the market and, conditional on entry, is more likely to persist in the market. We designate bias that results from overestimation as “optimism” and its opposite as “pessimism.”

4 There is also substantial research examining entrepre- neurs’ risk preferences as distinct from overconfidence. Åstebro et al. (2014: 57, 61) discussed evidence supporting the claim that risk-seeking and overconfidence bias are dis- tinct theoretical constructs, and argued that overconfi- dence is a better fit in terms of explaining patterns of entrepreneurial behavior. Along the same lines, Wu and Knott (2006) argued that risk-neutral entrepreneurs may exhibit “apparent risk-seeking” since ventures tend to have large (but risky) upsides.

5 Moore and Healy (2008) identified a third type of con- fidence bias, “overplacement,” where an entrepreneur views themself too favorably compared to others. Over- placement seems closely related to overestimation, where the former is an exaggerated understanding of quality rela- tive to others and the latter is an exaggerated understand- ing of absolute quality.

6 Recent research has also sought to distinguish over- confidence from other forms of bias, such as ambiguity aversion (Gutierrez et al., 2020; Posen, Leiblein, & Chen, 2018).

494 Academy of Management Review July

Overprecision is reflected in the confidence interval around an estimated belief that is smaller than necessi- tated by the uncertainty inherent in the opportunity. This overprecision has important dynamic consequen- ces in a model of learning (Chen et al., 2018; Posen et al., 2018). We follow Moore and Healy (2008) by defining over- and underprecision relative to rational belief updating as determined by Bayes’s rule. In a Bayesian learning model, the optimal rate of belief updating is a function of the noisiness of the feedback signal. When an entrepreneur exhibits overprecision bias, they are too sure that their beliefs are correct and, as a consequence, update their beliefs in the presence of new information less than is warranted by the uncertainty inherent in the signal. In the extreme case, the entrepreneur completely ignores new information. Likewise, underprecision bias is the case in which the entrepreneur responds too strongly (i.e., updates their beliefs too much in response to new information).

It is not at all obvious how individual biases inter- act in a team decision-making setting. In principle, there may be conditions under which partners’ biases amplify one another and other situations in which these biases cancel each other out. It may also be the case that alternative decision-making struc- tures are differentially effective at mitigating entre- preneurial bias, such that the most appropriate decision-making structure may be a function of team members’ bias types.

Opportunity Costs, Teams, and Entrepreneurial Entry

Much scholarly literature examines the impact of opportunity costs on entrepreneurial entry (e.g., Shane, 2000). In this literature, an individual enters the market if their expected utility from doing so— that is, the sum of both anticipated pecuniary and nonpecuniary rewards—exceeds the entry costs of setting up the new organization and the opportunity costs of foregone wages in paid employment (Parker, 2009b; Wu & Knott, 2006). Numerous studies build on this framework, implicitly or explicitly, to explain the differential patterns of entrepreneurial entry across the wage spectrum (e.g., Åstebro & Thompson, 2011; Braguinsky, Klepper, & Ohyama, 2012; Elfen- bein, Hamilton, & Zenger, 2010; Hamilton, 2000), to examine the role of liquidity constraints on entrepre- neurial entry (e.g., Evans & Jovanovic, 1989; Hurst & Lusardi, 2004), and to model the entry decisions of diversifying incumbents (Levinthal & Wu, 2010).

When the entry decision is preceded by a period of learning, opportunity costs influence which firms

enter the market. This has two implications (Chen et al., 2018). First, the extent of the opportunity costs truncates the distribution of entrants’ beliefs about the merits of their opportunities. The higher the costs of entry, the more positively a firm must view its opportunity to cross the market entry threshold. Second, there is a positive correlation between the level of opportunity costs and the time needed for a firm with a poor-quality opportunity to exit.

Although previous literature has generated signifi- cant scholarly and practical insights, it is premised on depictions of decision-makers as individuals or firms as unitary actors. When teams of individuals are involved in entrepreneurship, their joint access to resources may increase, but the opportunity costs to each individual team member must be consid- ered. Each member of the team must ultimately ben- efit from entry. To our knowledge, no prior work examines the effect of this requirement on market entry or the tendencies of teams that have similar versus heterogeneous opportunity costs, nor do we know of prior work that examines how different decision-making structures affect entry and exit decisions when opportunity costs among team mem- bers are heterogeneous.7

In the next section, we lay out the foundations of our computational model. First, the entrepreneur- ship process involves learning about an uncertain opportunity: learning pre-entry that informs entry/ termination decisions and learning post-entry that leads to ongoing exit/continuation decisions. Sec- ond, at the time of a positive entry decision, entrepre- neurs each incur nonrecoverable fixed opportunity costs. Third, entrepreneurs may exhibit Bayesian rationality or confidence biases that cause their beliefs about the merits of their opportunities to devi- ate from the rational baseline; however, entrepre- neurs act rationally conditional on their (potentially biased) beliefs. Finally, entrepreneurial teams make entry and exit decisions based on a decision-making structure that aggregates the beliefs and opportunity costs of all team members.

COMPUTATIONAL MODEL

We developed a simulation model in which multi- ple agents independently learn about the success prospects of an entrepreneurial opportunity during a

7 Alvarez and Barney (2005) discussed heterogeneous opportunity costs as affecting bargaining power over deci- sion rights and cash-flow rights, but did not offer a con- crete assessment of their impact.

2022 Chen, Elfenbein, Posen, and Wang 495

pre-entry period. At the end of this period, the agents decide whether to engage in the opportunity by forming a single business and starting operations. The manner by which this decision is made is deter- mined in advance, and our model focuses on three possibilities that are described in more detail below: partnership, boss-employee, and buyout option.8

We begin this section by describing the structure of the model, starting with a single-agent model and moving to a multi-agent model. We then exercise our multi-agent model to assess the efficacy of alterna- tive decision-making structures and to highlight the effects of confidence biases across these structures.

Single-Agent Model

The simulation builds on the single-agent entre- preneurial learning simulation framework of Chen et al. (2018) (henceforth, “CCEP”), which itself adds pre-entry learning and behavioral bias to Ryan and Lippman’s (2003) model of optimal exit from a pro- ject with noisy returns. In CCEP, agents costlessly evaluate an opportunity during a pre-entry learning period of length L (i.e., from t 5 2L to 0). During this period, these agents may, for example, engage in market research, build a product prototype, or make contact with potential suppliers. Based on their updated beliefs about success prospects at the end of the pre-entry learning period, agents may opt to pay an entry cost, k, and enter the market to exploit the potential profits of the opportunity. Entry costs include both the cash costs associated with starting operations and the opportunity costs of foregoing alternative employment. Post-entry, agents accrue actual profits and losses and use these profits and losses to continue to update beliefs about the merits of the opportunity.

For parsimony, CCEP focuses on two discrete types of opportunities: with probability p, the oppor- tunity is profitable (type H), with mean profit rate m 5 mH . 0; and, with probability 1 2 p, the opportu- nity is unprofitable (type L), with mean profit rate m 5 mL , 0. In either case, profit variance in both the pre- and post-entry periods is s2. The signal of cumulative profits, Xt, that an agent receives follows a Brownian motion with drift m and variance s2. Fol- lowing entry, profits are discounted at a rate d . 0. Though an agent does not know the opportunity type ex ante, they do know the parameters of the

underlying Brownian motion. We use p̂t to denote an agent’s belief about the probability that the opportunity is type H at time t. As the agent accrues profit signals pre-entry, and actual profits post-entry, p̂t is continuously updated to reflect their evolving beliefs.9 The simulation initiates at t 5 2L, which is the start of the pre-entry learning period. A rational agent’s belief that the opportu- nity is type H at the beginning of the pre-entry period is p̂2L5p. Assuming the opportunity is type H, p̂t ! 1 as t ! 1.

The entry decision is a central element of CCEP. An agent enters if and only if the expected return from the opportunity, as informed by the pre-entry period, exceeds the entry cost, k. This expected return, which includes both expected operating profit as well as the option to cut losses by exiting, is a function of p̂0, the agent’s belief about the probabil- ity of a type H opportunity at k 5 0. Thus, an agent’s optimal policy is to enter if p̂0 maps to expected returns that exceed k. Conditional on entry, the deci- sion to continue is an optimal stopping problem; if the agent’s belief p̂t falls below p

�, itself a function of the key parameters of the model, the agent termi- nates the business and exits the market.10 Figure 1 depicts the timing of the model.

CCEP incorporates two types of confidence- related behavioral biases. First, agents may have biased initial beliefs and, as such, may exhibit estimation bias. Values of p̂2L.p reflect overesti- mation (optimism), while p̂2L,p reflects underesti- mation (pessimism). Second, agents may have biased beliefs about the noisiness of profit signals and accordingly may exhibit precision bias (Moore & Healy, 2008). Per standard convention, precision t is defined as the inverse variance in profit signals (i.e., t 5 1/s2). We use t̂ to denote the agent’s beliefs on t. By recognizing that t̂=t is simply the ratio of the true s2 to the agent’s beliefs on s2, t̂=t ,1 reflects overprecision, in that agents update beliefs too slowly (relative to a Bayesian learner) because they believe profit signals to be less informative (i.e., noisier) than they really are. Conversely, t̂=t .1 reflects underpre- cision, in that agents update beliefs too rapidly

8 We acknowledge that the choice of decision-making structure itself may be a function of agents’ bargaining power or bargaining ability.

9 Ryan and Lippman (2003) derived a number of closed form expressions, including an expression for an agent’s beliefs at time t, which we employ in our modeling exercise.

10 The closed form solution for p� can be found in equa- tion (7) in Ryan and Lippman (2003: 442).

496 Academy of Management Review July

because they believe profit signals to be more infor- mative (i.e., less noisy) than they really are.11

Two-Agent Model

We extend the single-agent CCEP framework to two risk-neutral agents.12 We assume that each agent is necessary for operating the enterprise—that is, that each agent has human capital that is essential for investigating and exploiting the opportunity. A firm, then, consists of two agents who form a team prior to entry to investigate a single opportunity. As in CCEP, we assume that, prior to entry, these agents receive signals with independent noise components, drawn from a normal distribution with mean m and variance s2. Each team member forms unique beliefs about the quality of the opportunity, behaving exactly as they would in CCEP. We assume that the private nature of the pre-entry signals, strategic con- siderations, and the potential for bias preclude agents from directly and credibly communicating their beliefs to one another; as a result, neither agent makes an inference based upon any information received by their teammate.13 Their decision to com- mit to joining the firm, however, is credible and con- tractible. Agents, by definition, commit only if they believe that, by doing so, they will be better off in expectation.

We further assume that each agent, j 5 1, or has their own opportunity cost of entry, kj. To agree to participate, each agent must expect to receive a reward equal to or greater than their opportunity cost. Following entry, entrants receive either a fixed payment or a share of the profits, depending on their decision-making structure, which we introduce below.14 As in CCEP, we abstract away from liquid- ity constraints in this model; thus, any agreement that is viewed in expectation as value creating for all parties can be costlessly financed.

With two agents, we must specify a decision- making structure that encompasses three related processes to describe firm behavior fully: (a) how the agents in the team decide to enter, (b) how they decide to exit, and (c) how they divide the proceeds of the business conditional on entry. The first two of these correspond to decision rights, and the third corresponds to cash flow rights. We identify three alternative decision-making structures that deter- mine learning aggregation, ownership, and entry/ exit decisions, which we describe below.

Boss-employee (boss). In the boss decision- making structure, a pre-identified agent, who we will call agent 1, has both decision rights and cash flow rights as the boss of the firm from the beginning of the pre-entry period. If the boss decides to enter, they must ensure agent 2 earns at least their opportu- nity costs, k2. This may come in the form of an upfront fixed fee or a periodic payment that, in expectation, is equal to k2, which may be thought of as a wage. The team enters only if agent 1 believes that the expected value of the opportunity exceeds k1 1 k2. In other words, only agent 1’s beliefs matter for the boss decision-making structure, and her entry threshold is defined by the sum of both agents’ opportunity costs. Conditional on entry, the firm

FIGURE 1 Timing of Model

Costless and profitless period to

evaluate entry

Entrants may choose to exit

t = −L t = 0 time

Notes: The first period of the model is a pre-entry learning period of length L, which concludes at t 5 0 when agents make the market-entry decision. If entry occurs, agents evaluate whether to remain in the market continuously following t 5 0.

11 For an agent who holds possibly erroneous beliefs on the noisiness of profit signals but is otherwise Bayesian, there is a one-to-one relationship between the rate at which the agent updates their beliefs to new profit signals and their beliefs about the variance of profit signals. This rela- tionship can be calculated through the analytical solution offered in Ryan and Lippman (2003).

12 In principle, the model can be expanded to incorpo- rate N agents. For simplicity, we focus on N 5 2.

13 In the Appendix we relax this assumption by allow- ing pre-entry signals to be correlated.

14 The fixed payment might be thought of as a wage in an employment relationship.

2022 Chen, Elfenbein, Posen, and Wang 497

exits if and only if agent 1’s beliefs about the value of the opportunity fall below p�. The boss’s exit deci- sion is binding and irrevocable for both agents.15

Equal partnership voting (partnership). In the partnership decision-making structure, agents com- mit at the beginning of the pre-entry learning period to split all post-entry profits equally, and cannot re-contract at any point onward. Entry occurs only if both agents believe that the expected value of a 50% share of the profits exceeds their individual opportu- nity costs. Conditional on entry, each agent receives exactly half of the profits. Since all agents are ne- cessary for the operation of the firm, exit occurs the first time any agent’s beliefs fall below p�. At that point, it is optimal for that agent to withdraw their effort.16

Buyout option in an equal partnership (buyout). We additionally model a decision-making structure in which agents begin the pre-entry learning period in an equal partnership, but may reallocate decision and cash flow rights at the end of the pre-entry learn- ing period if there is a disagreement. Under buyout, one of three situations follows the learning period: (a) if both agents 1 and 2 believe that the expected value of a 50% share of the profits exceeds their opportunity costs—k1 and k2, respectively—then the firm remains a partnership and enters; (b) if condi- tion (a) is not met, but one agent believes expected profits exceed k1 1 k2, then the firm enters as a boss structure, where the positively inclined agent pays the negatively inclined agent at least their opportu- nity costs to work for the firm;17 and (c) if neither agent believes the expected value of total profits exceeds k1 1 k2, the team does not enter. We note that, post-entry, this decision-making process leads

either to a boss structure, where decision-making and profits are consolidated in a single agent, or a part- nership structure, where decision-making and profits are divided between two agents. In other words, the buyout option may or may not be exercised.

We note that, when it comes to entry, the partner- ship structure relates closely to the “hierarchy” decision-making structure described by Sah and Sti- glitz (1986), insofar as all agents must agree for entry to proceed. By contrast, the buyout structure relates to Sah and Stiglitz’s (1986) “polyarchy” decision- making structure, insofar as only a single agent must take ownership for entry to proceed.

Simulation Output

The computational simulation produces a rich data stream at both the individual agent and firm lev- els, which enables us to compare the performance of different team structures and tie these performance differences to specific mechanisms. At each point in time, our model records what each agent believes about the probability that their opportunity is type H. For teams that decide to enter, the model addi- tionally records profits and losses during each period of operation, total profit net of opportunity costs, and the time of exit for those firms that choose to shut down.

We focus primarily on the expected value of new venture profits conditional on the team’s decision- making structure and agents’ opportunity costs. In all simulations, we fix p 5 .50, mH 5 150, mL 5 250, and d 5 0.1. Unless otherwise specified, we set k1 5 k2 5 50, and simulate one million two-agent teams. We vary the length of the pre-entry learning period (L), as well as agents’ initial beliefs ðp̂2LÞ and their beliefs about the precision of the signals they receive ðt̂=tÞ.

Given that p, mH, mL, and d are fixed, performance is determined by the cost and incidence of three types of errors: (a) mistaken entry, in which type L teams mistakenly enter; (b) mistaken nonentry, in which type H teams mistakenly fail to enter at all; and (c) mistaken exit, in which type H teams enter but subsequently leave the market mistakenly. The left-hand panel of Figure 2 shows how type L and type H firms sort into groups defined by these errors. Since type L teams that enter suffer both the oppor- tunity costs of entry and accumulated losses as long as they stay in the market, the cost of mistaken entry is a function of both k1 1 k2 and how long it takes type L entrants to exit the market. The cost of mis- taken nonentry represents the foregone stream of

15 Entry and exit behavior in the boss decision-making structure is identical to the single-agent model described in CCEP, with the modification that beliefs at t 5 0 must be sufficiently high that the expected value of operation exceeds the sum of both agents’ opportunity costs.

16 Post-entry, the Ryan and Lippman exit threshold for an individual owner is identical to the exit threshold for a partial owner, since the shutdown payoff is zero and a par- tial agent receives the same proportion of profits, regard- less of whether they are positive or negative.

17 Cramton, Gibbons, and Klemperer (1987) established that such a deal is feasible in N-player settings when the total surplus to be divided is high enough, a condition that is trivially satisfied by our entry requirements. The nega- tively inclined agent may be paid more than their opportu- nity costs in this scenario, as the surplus division in this case is not precisely determined by the parties’ outside options.

498 Academy of Management Review July

earnings from an opportunity that would have suc- ceeded, and is thus characterized by the unrealized perpetuity value of a type H team (mH/d) less the entry cost of the agents (k1 1 k2). The cost of mis- taken exit can be characterized the same way as mis- taken nonentry, treating the entry costs as sunk and starting from the time the type H team erroneously exits.

Figure 2 shows the relative incidence of each type of error by decision-making structure for two unbi- ased agents with equal opportunity costs k1 5 k2 5 50 and L 5 1. Partnership generates the fewest entry mistakes but the most nonentry mistakes, consistent with the conservatism borne by requiring both agents to agree to the entry decision. Boss has inter- mediate levels of both entry mistakes and nonentry mistakes. Buyout has the most entry mistakes but the fewest nonentry mistakes, which follows from requiring only one agent to have beliefs above the entry threshold for entry to occur. Furthermore, when the buyout option is exercised, exit will be

determined by the firm’s single owner, who, by vir- tue of deciding to buy out their negatively inclined partner, is likely to be optimistic. Therefore, the single-owner boss who exercises a buyout option suffers a winner’s curse, which exacerbates entry mistakes. These contrasting levels of each type of error point to the key mechanisms underlying the main findings in our experiments. We now provide detailed analyses of these experiments.

RESULTS

We begin by examining the performance of the three different decision-making structures—boss, partnership, and buyout—on populations of agents who differ in the nature of their biases. We first use the model to describe the relative performance with identical unbiased, Bayesian agents. We show that, with these agents, partnership is the optimal struc- ture, so long as team members’ opportunity costs of entry are not too different from one another. With

FIGURE 2 Comparison of Decision Errors for Two-Agent Teams of Unbiased Agents

P ro

p o rt

io n

o f

A ge

n ts

Type L

Nonentrant

10000 Boss Partnership Buyout

9000

8000

7000

6000

5000

4000

3000

2000

1000

0 0 5 10 15 20 0 5 10 15 20

10000

9000

8000

7000

6000

5000

4000

3000

2000

1000

0 0 5 10 15 20

10000

9000

8000

7000

6000

5000

4000

3000

2000

1000

0

Entry Mistakes

Entry Mistakes

Exited

In Market

Nonentry Mistakes

Type H

Notes: Simulation of the 10,000 teams of two unbiased agents with opportunity costs equal to 50 for each agent. The x-axis on each graph is the number of periods post-entry. The top half of each graph represents the decisions of 5,000 type L teams, and the bottom half represents the decisions of 5,000 type H teams. The left-most panel contains labels for each section of the graphs. “Nonentrant” indicates the proportion of type L teams that did not enter. “Entry mistakes” indicates the proportion of type L teams in the market, and “Exited” represents the proportion of type L teams that have exited the market. “In market” represents type H teams in the market. “Exit mistakes” represents type H teams that entered the market but subsequently exited. “Nonentry mistakes” represents the proportion of type H teams that fail to enter altogether. For each panel, the sum of the shaded portions corresponds to the total number of firms making entry or exit errors under the decision-making structure of interest.

2022 Chen, Elfenbein, Posen, and Wang 499

unbiased Bayesian agents, we additionally investigate how the relative advantage of partnership over the other structures varies with the amount of pre-entry learning. Next, we fix entry costs and examine the rela- tive performance of the three decision-making struc- tures with a population of agents that is, on average, unbiased, but contains agents with a distribution of estimation and precision biases. With a popula- tion of on-average-unbiased agents, buyout becomes the optimal structure. Third, we relax the assump- tion that the population has a zero mean bias and identify the boundary conditions in terms of population-level estimation and precision biases that define when buyout dominates partnership and vice versa.

We augment this analysis with two sets of exten- sions that highlight the utility of our computational model. First, we compare the relative post-entry per- formance and bias characteristics of firms that enter as a single-owner decision-making structure versus a partnership. Here, we examine how selection influ- ences the population of entrants upon which empiri- cal analyses would typically be based and use our model to generate novel predictions for empirical study. Second, we hold the biases of the potential entrepreneurs fixed, and determine the optimal structure for each combination of biases (i.e., agent 1’s initial estimation bias and precision bias; agent 2’s initial estimation bias and precision bias). This further demonstrates the utility of our framework in generating managerially relevant recommendations to teams of prospective entrepreneurs.

Comparison of Decision-Making Structures with Identical Bayesian-Rational Agents

Our first computational experiment employs only Bayesian-rational agents—agents who have the cor- rect initial beliefs (i.e., rational with p̂2L5 0.5) and who update their beliefs accurately (i.e., Bayesian with t̂/t 5 1)—and identical opportunity costs k1 5 k2 5 50. We vary the length of the pre-entry learning period L from 0.1 to 5. We interpret L 5 0.1 as indi- cating that the pre-entry learning period is very short, leading the pre-entry information collected to be relatively noisy, and L 5 5 as indicating that the pre-entry learning period is very long, meaning that the pre-entry information collected is relatively pre- cise. Our focus throughout is on how total expected profits of the team respond to the decision-making structure in question. We abstract away from bar- gaining power considerations, which might lead to the selection of one structure over the other or alter

the profit split between agents. Figure 3 shows the relative performance for boss, partnership, and buy- out at each level of L. As L increases, prospective entrepreneurs make fewer entry and exit mistakes, and the average value created by entrepreneurial teams increases. For Bayesian-rational agents with equal opportunity costs, teams organized as partner- ships generate the greatest value, regardless of the length of the pre-entry learning period. Partnership’s relative advantage over boss is 1% with low pre-entry information (L 5 0.1), rises to nearly 5% at moderate information levels, and declines back to about 1% at high levels (L 5 5). Relative to boss, the reduction of mistaken entry produced by partner- ship more than compensates for the increase in mis- taken nonentry, although entry behavior converges as L approaches infinity.

Somewhat paradoxically, adding the buyout option destroys value when all agents are Bayesian- rational. For these agents, entry mistakes resulting from the winner’s curse (i.e., the “curse” from the likelihood that at least one agent will overestimate the probability of success) are particularly costly. With small amounts of pre-entry information, the winner’s curse is modest: buyout performs 4% worse than boss when L 5 0.1. With moderate amounts of pre-entry information, however, the winner’s curse is at its most severe: buyout performs 10% worse than boss from about L 5 0.3 to 1. For Bayesian- rational agents, then, the flexibility afforded by buy- out comes with a cost that varies in severity as a function of pre-entry information.

We next examine the relative performance of boss, partnership, and buyout when opportunity costs vary. Our main concern here is with the distribution rather than the total level of opportunity costs. To examine this, we hold constant k1 1 k2 5 100 and vary k1 from 0 to 50. Figure 4 depicts the results of asymmetric opportunity costs for differing lengths of pre-entry learning.

As Figure 4 indicates, the supremacy of partner- ship breaks down when the agents’ opportunity costs differ significantly; that is, when k2 . k1. For modest amounts of pre-entry learning—that is, L , 2—partnership provides the best organizational form only when the two agents’ opportunity costs are relatively similar. As the two agents’ opportunity costs become more asymmetric, partnership’s per- formance worsens, boss generates the highest expected profits, and buyout also overtakes partner- ship. The mechanism for this result is intuitive: when a team includes a high-opportunity cost agent and pre-entry learning is limited, mistaken nonentry

500 Academy of Management Review July

by type H teams is exacerbated by the partnership structure. This is because, with a short L, the high-opportunity cost agent is less likely to be sure that their payoff from entry will be positive, and therefore more likely to veto entry even when the opportunity is type H. By contrast, single-agent own- ership structures allow the boss to pay the employee their opportunity costs directly if the boss believes expected profits are large enough to cover k1 1 k2. In other words, when opportunity costs are asymmet- ric, boss and buyout decision-making structures are relatively insulated against the heightened nonentry mistakes that plague a partnership. Finally, Figure 4 shows that, for unbiased agents, longer pre-entry

learning periods erase the problems of partnership when opportunity costs are heterogeneous. When pre-entry knowledge is sufficiently comprehensive, teams make correct decisions irrespective of oppor- tunity cost asymmetry. Pre-entry learning also miti- gates the buyout’s winner’s curse; at L 5 5, structure has limited impact on performance with rational agents, regardless of opportunity cost distribution.

As the discussion above indicates, our model highlights two prospective solutions to remedy part- nership’s problem in handling asymmetric opportu- nity costs: (a) changing the decision-making structure or (b) maintaining the partnership struc- ture but extending pre-entry learning. A third

FIGURE 3 Partnership Is Superior to Buyout and Boss with Deterministically Unbiased Agents

150

100

50

15

Boss Buyout

P er

fo rm

an ce

R el

at iv

e to

P ar

tn er

sh ip

( %

) 10

5

0

−5

−10

−15

Length of pre-entry learning period, L 0.1 0.3 0.5 1 2 3 4 5

P er

fo rm

an ce

o f

P ar

tn er

sh ip

Notes: The top panel plots the average performance of deterministically unbiased agents under the partnership decision-making structure as a function of pre-entry learning duration L. The bottom panel plots the average performances of deterministically unbiased agents each under the boss and buyout decision-making structures relative to partnership performance (top panel) (%), as a function of L. For example, given L 5 2, performance under buyout is on average �7.5% worse than performance under partnership. A log-scale is applied to the x-axis.

2022 Chen, Elfenbein, Posen, and Wang 501

solution also exists: (c) changing the equity distribu- tion to mimic the ratio of opportunity costs.18

In other words, set the equity share of agent 1 to k1/(k1 1 k2) and the equity share of agent 2 to k2/ (k1 1 k2). Figure A1 in the Appendix shows how this

remedy restores partnership’s supremacy with unbi- ased agents. While, theoretically, this is a rather obvious solution, we note that empirical evidence suggests that relatively few entrepreneurial teams adopt this approach. For instance, the vast majority (89%) of two-owner ventures in the Kauffman Firm Survey (KFS) report an approximate 50:50 (within 5%) split in returns, suggesting that partnerships

FIGURE 4 Partnership Performance Declines with Unequal Opportunity Costs

160 L = 0.1 L = 0.5 L = 1

L = 2 L = 5

140

120

100

80

P ro

fi t

60

40

20

0 0.1 0.2 0.3 0.4 0.5

0 0.1 0.2 0.3 0.4 0.5

Proportion of team entry costs assumed by agent 1

0 0.1 0.2 0.3 0.4

Boss Partnership Buyout

0.5

0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5 0

160

140

120

100

80

P ro

fi t

60

40

20

0

Notes: For each pre-entry learning duration L e {0.1, 0.5, 1, 2, 5}, we plot the average performance under each decision-making structure as a function of the proportion of total opportunity costs assumed by agent 1. Total opportunity costs of entry for both agents in all simulations equal 100. The x-axis represents the proportion of total entry costs possessed by agent 1. For example, at x 5 0.3, agent 1’s opportunity costs are k1 5 30, and agent 2’s opportunity costs are k2 5 70. The graphs are each symmetric around x 5 0.5.

18 We thank an anonymous reviewer for this suggestion.

502 Academy of Management Review July

tend to default to a 50:50 division of equity irrespec- tive of individuals’ opportunity costs (Ewing Marion Kauffman Foundation, 2013).19

Comparison of Decision-Making Structures with a Population of Mean-Unbiased Agents

In the prior section, we assumed that potential entrepreneurs were known not to have decision- making biases. In this section, we relax this assump- tion and assume that potential entrepreneurs’ decision-making biases are unknown, but are drawn from “neutral” distributions with zero mean bias. In other words, we consider a population of potential entrepreneurs that is unbiased on average, but whose agents individually exhibit a range of estimation and precision biases. Specifically, for each agent i, we draw initial beliefs about the likelihood of being type H, p̂2L, from a b(2,2) distribution, which par- tially mimics the attributes of a normal distribution but over the range 0 to 1. We independently draw each agent’s precision bias, t̂=t52X, where X � 4 � b(2,2) 2 2, setting the range of precision from 0.25 to 4.20 We plot these distributions in Figure A2 in the Appendix. In choosing these distributions, we seek to represent (a) a population of nascent entrepre- neurs that has neutral bias, on average and at the median, but does have variance in both initial esti- mation and precision biases; and (b) a process whereby individual entrepreneurs do not con-

sciously match according to similarities or differ- ences in estimation or precision bias when forming teams.21 As our interest is in understanding the rela- tive performance of partnership, boss, and buyout in this and in subsequent sections, we equalize each agent’s opportunity costs and hold them constant (k1 5 k2 5 50). We again vary the length of the pre- entry learning period L from 0.1 to 5.

We plot the relative performance of the three decision-making structures in Figure 5. With on- average-unbiased agents drawn from the distribu- tions detailed above, partnership is no longer the structure that generates the best results. In this case, buyout and boss are both superior or equal to part- nership across the full range of pre-entry durations that we examine. The disparity is greatest at L 5 0.1, when buyout and boss outperform partnership by 55.5% and 14.6%, respectively. The performance difference among all three decision-making struc- tures becomes small beyond L 5 2.22

These results indicate that buyout is clearly more robust in handling agents with a range of (uncorre- lated) biases than the other decision-making struc- tures we examine. Why is this the case? Intuitively, the reason is that biases introduce two particularly problematic types of teams that are nonexistent with deterministically unbiased agents: pessimistic part- nerships, wherein entrepreneurial teams including at least one pessimist are organized as partnerships; and optimistic buyouts, wherein teams including at least one optimist are organized as buyouts. Pessi- mists in a partnership exacerbate that structure’s tendency to make nonentry mistakes, while opti- mists in a buyout exacerbate its tendency to make entry mistakes. While pessimistic partnerships and optimistic buyouts are equally prevalent, by as- sumption, in a mean-unbiased population, they are not equal in terms of impact on profits. Pessimistic partnerships are worse, because they induce mis- taken nonentry that leads to irreversible foregone gains, in contrast to optimistic buyouts whose entry mistakes can be mitigated by exit. Thus, buyout is the superior decision-making structure when the

19 The KFS data are available from www.kauffman. org/kfs.

20 Given these functional forms, the median agent in our simulation will have initial beliefs equal to 0.5 and preci- sion equal to 1. Note that precision acts multiplicatively on beliefs, such that an overprecision of 0.5 is the counter- part to an underprecision of 2, and a precision of 1 is unbi- ased (unlike estimation, in which, e.g., a bias of 20.2 is the counterpart of 0.2, and 0 is unbiased). Thus, the log of pre- cision should follow a symmetric distribution around zero, which the log b, suitably scaled and translated, accomplishes. While initial beliefs are naturally restricted to the [0,1] interval, the distribution of agents’ precision has no natural support. We restrict our attention to preci- sion values (t̂=t) ranging from 0.25 (overprecision) to 4 (underprecision). Correct precision is found at t̂=t 5 1, at which point agents update in an unbiased (or Bayesian) manner; with t̂=t 5 0.25, agents put half as much weight on incoming information as a Bayesian updater, and, with t̂=t 5 4, they put twice as much weight on incoming infor- mation as a Bayesian updater.

21 Parker (2009a) constructed a model in which agents match consciously on dimensions of cognitive bias, as this bias affected effort in his model.

22 In unreported analyses, we examine estimation biases that are distributed Unif(0,1) and precision biases that remain “neutral” but exhibit more dispersion than in b(2,2). The dominance of buyout remains under these assumptions. Results are available upon request.

2022 Chen, Elfenbein, Posen, and Wang 503

population has a distribution of biases that is mean- unbiased.

Figure 6 supports this intuition that the buyout advantage arises from the costs generated by pessi- mistic agents in partnerships relative to the costs of optimistic agents in buyouts. Panel A contrasts the profit contribution of type L and type H firms orga- nized via buyout and partnership, respectively, as a function of agent 1’s initial beliefs. (Note that the biases of agents 1 and 2 are independent, so a sym- metric figure exists for agent 2 as well). Profit contri- bution reflects the mean profit of a population of entrepreneurial teams according to teams with a

given profile (e.g., type H and buyout) and a given agent 1 bias (see Note following Figure 6 for a detailed explanation). Profit contribution values may be positive or negative. Panel B plots the differ- ences in these contributions. Per Panel B, when the team includes a pessimist with p̂2L5 0.3, there is a buyout advantage of 3.62. By contrast, when the team includes an optimist with p̂2L5 0.7, partner- ship generates an advantage of 2.57. Since our distri- bution contains an identical number of optimistic and pessimistic agents, by construction, the net effect is an overall buyout advantage of 1.05 when comparing this matched pair (10.2, 20.2) of biased

FIGURE 5 Buyout Is Superior to Partnership and Boss in an On-Average-Unbiased Population of Biased Agents

150

100

50

50

Boss Buyout

25

0

−25

–50 0.1 0.3 0.5

Length of pre-entry learning period, L 1 2 3 4 5

P er

fo rm

an ce

o f

p ar

tn er

sh ip

P er

fo rm

an ce

R

el at

iv e

to p

ar tn

er sh

ip (

% )

Notes: We refer to agents with estimation bias drawn from a b(2,2) distribution and precision bias drawn from a log2b(2,2) distribution as “on average, unbiased.” The top panel plots the average performance of 1 million simulated teams of on-average-unbiased agents under the partnership decision-making structure as a function of pre-entry learning duration, L. The bottom panel plots the average performances of 1 million simulated teams of on-average-unbiased agents each under the boss and buyout decision-making structures relative to partnership performance (top panel) (%), as a function of L. For example, given L 5 0.5, performance under buyout is, on average, �15% better than perfor- mance under partnership. A log-scale is applied to the x-axis.

504 Academy of Management Review July

initial beliefs. The total advantage is significantly larger because we integrate over the entire range of pessimist/optimist pairs.

The results of this section, contrasted with those of the previous section, highlight the effects of intro- ducing bias into decision-making agreements. When agents have a range of biases, buyout generates the best performance, due to the greater cost of irrevers- ible mistaken nonentry relative to mistaken entry that can be corrected through exit. This makes pessi- mistic teams more damaging when organized via partnerships than are optimistic teams when orga- nized via buyouts; as a result, buyout is the more robust decision-making structure when agents are symmetrically distributed around neutral biases. Moreover, partnership generates especially poor performance when pre-entry information is lim- ited or of low quality, because there is insufficient time for learning to correct the biases that lead to mistakes.

Comparison of Decision-Making Structures with Populations of Biased Agents

In our computational model, with equal opportu- nity costs, partnership generates the best results when all prospective entrepreneurs are unbiased

(Bayesian-rational), whereas buyout generates the best results when prospective entrepreneurs have a range of biases but remain unbiased on average. We next turn to an examination of relative outcomes when agents come from a distribution that exhibits bias on average. Here, our focus is on understanding the conditions under which buyout dominates partnership.

We construct nine sets of population distributions corresponding to three mean levels of initial beliefs—an optimistic population, where p̂2L� b(3,2); a realistic population, where p̂2L� b(2,2); and a pessimistic population, where p̂2L� b(2,3)— and three mean levels of precision—an underprecise population, where log t̂=t � b(3,2); a correctly pre- cise (in aggregate) population, where log t̂=t � b(2,2); and an over-precise population, where log t̂=t � b(2,3). Figure A3 in the Appendix illus- trates these nine pairs of distributions that define the estimation and confidence biases of each population under study.

Figure 7 depicts the boss and buyout profits rela- tive to partnership across these nine types of popula- tions with L ranging from 0.1 to 5. The left and center panels of Figure 7 show that, when agents are drawn from underestimating (pessimistic) or cor- rectly estimating (realistic) populations, buyout

FIGURE 6 Why Buyout Is Superior to Partnership in an On-Average-Unbiased Population of Biased Agents

30 Type H buyout

+11.17 6.62

−9.19

−7.55

Type H buyout – Type H partner

Net loss = –2.57

Net gain = 3.62

All buyout – All partner

Type L buyout – Type L partner

Partner and buyout individually Difference between buyout and partnerA B

Type L buyout

Type H partner

Type L partner

20

0

−10

−10

−5

0

5

10

−20 0.1 0.3 0.5 0.7 0.9

Initial optimism of agent 1 0.1 0.3 0.5 0.7 0.9

Initial optimism of agent 1

10

C o n

tr ib

u ti

o n

t o p

ro fi

t

C o n

tr ib

u ti

o n

t o p

ro fi

t

Notes: In both panels, L 5 1. Agent 2’s initial beliefs are drawn from a b(2,2) distribution and both agents’ precision biases are drawn from a log2b(2,2) distribution. The x-axis plots initial beliefs of agent 1. The y-axis provides mean profit by team characteristics and agent 1 bias. For example, the left panel indicates that type L buyout teams for which agent 1’s initial optimism is 0.3 contribute 18 to the mean profit of the pop- ulation. Thus, to obtain the overall mean profit for type L buyouts, we would sum over the plotted values for the type L buyout line in 0.1 incre- ments. Panel B shows that the net gain is larger in magnitude than net loss because pessimistic partnerships are more problematic than optimistic buyouts.

2022 Chen, Elfenbein, Posen, and Wang 505

delivers higher average performance than boss or partnership, regardless of the precision biases of the population.

In overestimating (optimistic) populations, how- ever, the patterns are more intriguing. When pre- entry learning is very limited, buyout provides the best structure. By contrast, with ample pre-entry learning, partnership becomes the best structure.

This is depicted by the buyout and boss lines falling below zero for modest levels of L in the rightmost panels. Although initial optimism in the population promotes entry for all structures, thereby limiting the overall incidence of mistaken nonentry, the “AND” structure of partnership additionally reduces the relative incidence of mistaken entry so long as the agents have had enough learning to overcome

FIGURE 7 The Optimal Decision-Making Structure in a Biased Population Depends on Bias and Pre-Entry Learning

Duration

0.1 0.3 0.5 1 2 3 4 5 0.1 0.3 Length of pre-entry learning period, L

0.5 1 2 3 4 5 0.1 0.3 0.5

Boss Buyout

1 2 3 4 5

100

Underestimation (Pessimistic)

Unbiased estimation (Realistic)

Overstimation (Optimistic)

U n

d er

p re

ci si

o n

U n

b ia

se d

p re

ci si

o n

O v er

p re

ci si

o n

P er

fo rm

an ce

R el

at iv

e to

p ar

tn er

sh ip

( %

)

50

0

100

50

0

100

50

0

Notes: For each combination of estimation and precision bias population distributions, we plot the average performances under the boss and buyout decision-making structures relative to partnership performance as a function of pre-entry learning duration (L). “Underestimation,” “Unbiased estimation,” and “Overestimation” refer to estimation biases drawn from a population distribution of b(2,3), b(2,2), and b(3,2), respectively. “Underprecision,” “Unbiased precision,” and “Overprecision” refer to precision biases drawn from a population distribution of log2b(2,3), log2b(2,2), and log2b(3,2), respectively. For example, when the population is, on average, underprecise and the initial probability of success is underestimated (i.e., top-left panel), performance under buyout is, on average, �50% better than performance under partnership for a pre-entry learning duration of L 5 2. A log-scale is applied to the x-axis.

506 Academy of Management Review July

their initial biases. Figure 8 depicts this dynamic: while the cost of mistaken entry does not change much for buyout as L increases (Panel A), it falls pre- cipitously for partnership (Panel B). When prospec- tive entrepreneurs are drawn from a population that is optimistic, then “two heads are better than one” in partnership, as its “AND” structure limits mistaken entry that would otherwise occur under boss or buy- out with one overly optimistic agent. In fact, partner- ship’s relative advantage is greatest when agents display both initial overestimation and overpreci- sion (see bottom right panel of Figure 7). In this case, partnership generates the best of both worlds, with optimism reducing nonentry mistakes and overpre- cision limiting the partnership’s propensity to make exit mistakes (partnerships are vulnerable to exit mistakes because exit is triggered by the agent with the smaller belief p̂t).

Our results, then, show that, although buyout handles populations with distributions of biases effectively, its advantage disappears when popula- tions display sufficient optimism bias about their projects’ likelihood of success. For optimistic populations—and especially optimistic, overprecise populations—the partnership structure yields the highest relative profits in expectation when pre-entry learning is sufficiently long.

Selection, Observed Bias, and Post-Entry Performance

The analysis in the prior section focused on total profits generated by all firms that begin the learning process. In this section, we focus our results on dif- ferences among entrants. In most empirical studies, only data about entrants will be available for analy- sis. We thus offer some of our model’s implications for these empirical studies. We focus on three meas- ures—operating profits, initial estimation bias, and precision bias—and examine how they differ in the population of surviving entrants over time. Each of the measures and their correlations changes dyna- mically over time as a result of patterns of exit shaped by pre-entry information, agents’ biases, and decision-making structure. We return to the assump- tions in “Comparison of Decision-Making Structures with a Population of Mean-Unbiased Agents” (sec- tion above) of mean-unbiased agents with equal opportunity costs, and hold L 5 1 to examine these patterns.

Entrants in our model will either take the form of a single owner or a two-owner partnership. Firms owned by a single individual may either come from the boss structure or from the buyout structure when the buyout option is exercised. Similarly, jointly owned firms come from either the partnership

FIGURE 8 Why Partnership Overtakes Buyout in Overestimating Populations When Pre-Entry Learning Is Sufficiently

Long

40 A Buyout Partnership

C o n

tr ib

u ti

o n

t o p

ro fi

t

30

Ty pe

H , L

= 1

Ty pe

H , L

= 1

Type L, L = 1

Type L, L = 1 Type L, L = 0.1

Type L, L = 0.1

Ty pe

H , L

= 0.1

Ty pe

H , L

= 0

.120

10

0

−10

−20

−30

40 B

C o n

tr ib

u ti

o n

t o p

ro fi

t

30

20

10

0

−10

−20

−30 0.1 0.50.3

Initial optimism of agent 1 Initial optimism of agent 1

0.7 0.9 0.1 0.50.3 0.7 0.9

Notes: Agent 2 is, on average, optimistic with initial beliefs drawn from b(3,2), and both agents are mean-unbiased in precision with preci- sion drawn from a log2b(2,2) distribution. The y-axis reports mean profit by team characteristics and agent 1 bias. The left panel indicates that increasing L from 0.1 to 1.0 in buyout reduces the losses from type L teams and increases the gains from type H teams modestly when agent 1’s initial beliefs are between 0.2 and 0.8. By contrast, the right panel shows that increasing L from 0.1 to 1.0 in partnership substantially reduces the losses incurred by type L teams when agent 1’s initial beliefs range from 0.4 to 0.9, while moderately increasing the gains from type H teams when agent 1’s initial beliefs range from 0.2 through 0.6.

2022 Chen, Elfenbein, Posen, and Wang 507

structure or the buyout structure when the team members enter as partners. Figure 9 plots the results over time for these four categories of entrants: two single-owner lines (boss or buyout!boss) and two joint-owner lines (partnership and buyout!part- ner). Panel A shows that, for all decision-making structures, average operating profits rise over time as (predominantly) type L firms drop out of the market. Unsurprisingly, Panel A also shows that jointly owned firms generate higher operating profits than individually owned firms. This occurs because the “AND” decision-making structure reduces entry by type L firms. Additionally, Panel A shows that buyout!boss realizes the lowest operating profits; among these entrants, the winner’s curse is severe. Additionally, for type L entrants, jointly owned firms (average time to exit 5 4.54 periods) exit

earlier than singly owned firms (average time to exit 5 6.36 periods). While this survival prediction is consistent with the results reported in a working paper by Greenberg and Mollick (2018), the perfor- mance results are not, perhaps suggesting that the underlying distribution of opportunities examined by Greenberg and Mollick’s data set differs between individual and jointly founded firms.

Panels B, C, and D yield predictions about the initial estimation and precision biases likely to be held by the entrant’s owners. Panel B shows that, while the median founders of all types of structures exhibit esti- mation bias, founders of single-owner (boss) entrants display greater initial optimism than founders of jointly owned (partnership) entrants. Buyout!boss firms’ owners are particularly optimistic. Panel C dis- plays the median levels of precision in the populations

FIGURE 9 Temporal Dynamics of Decision-Making Structures in Surviving Entrant Populations

400 0.13

1.02

1

0.98

0.96

0.94

0.92

0.9

0.12

0.11

0.1

A B CSurvivor profits Estimation bias Precision bias

300

200

T o ta

l p

ro fi

t C

o rr

el at

io n

C o rr

el at

io n

C o rr

el at

io n

M ed

ia n

e st

im at

io n

b ia

s

M ed

ia n

p re

ci si

o n

b ia

s

100

D E FCorr betw est bias and prec bias Intrateam est bias correlation Intrateam prec bias correlation

Boss

Partnership Buyout->Partner

Buyout->Boss

−0.15 0.02

0.01

0

−0.01

0

−0.1

−0.2

−0.3

−0.16

−0.17

−0.18

−0.19

−0.2

Period Period Period

−0.21

100 101 102 100 101 102 100 101 102

100 101 102 100 101 102 100 101 102

Notes: L 5 1 throughout. “Period” refers to time period post-entry. Each agent’s initial beliefs are drawn independently from a b(2,2) distri- bution and precision is drawn from a log2b(2,2) distribution. Opportunity costs ki 5 50 for both agents across all cases. In Panels B–D, the agent in whom bias is measured is the single-owner agent for buyout!boss and boss, and (arbitrarily) agent 1 in partnership and buyout!partner- ship. Panel A is total (i.e., full time horizon) profits for the set of surviving firms. In Panel B, values above 0 represent optimistic initial beliefs and values below 0 represent pessimistic initial beliefs. In Panel C, values above 1.0 represent underprecision and values below 1.0 represent overprecision. Panel D is the correlation between precision and estimation biases, and Panels E and F are the within-team correlations of esti- mation bias and precision bias, respectively.

508 Academy of Management Review July

of survivors over time. While partnership founders display slight underprecision, boss founders exhibit significant levels of overprecision, especially in peri- ods 1 through 10 post-entry. Panel D shows the expected correlation between estimation bias and pre- cision bias for new founders by decision-making struc- ture. While all founders display a positive correlation between optimism and overprecision (a negative corre- lation between estimation and precision bias), those operating jointly owned firms display a lower degree of correlation than those operating individually owned firms.

In Panels E and F, we look at the correlation between initial estimation bias and precision bias within entrepreneurial teams. By construction, the boss structure leads to no correlation between team members, as agent 2’s beliefs are irrelevant for operat- ing the firm. Buyout!boss, by contrast, yields nega- tive correlations between both the estimation and precision biases of team members; in other words, when the single-owner firm is the result of a buyout, the owner is likely to be optimistic (“overprecise”), while the nonowner employee is likely to be pessimis- tic (“underprecise”). The partnership structure yields the opposite results. In all jointly owned entrants, whether originating via partnership or buyout, team estimation and precision biases are positively corre- lated across team members. Our predictions about the combinations of cognitive biases in teams comple- ment those of Parker (2009a), are readily testable, and are, to the best of our knowledge, unique.

Finally, we note the differential entry patterns that result under our assumptions in this section. Among firms starting out as boss, 58.0% enter the market and 43.5% of these are still in operation after 5 periods. For partnership, the figures are 35.2% and 25.6%, respectively. Teams that begin with buyout enter the market 80.5% of the time, of which 43.8% enter as partnerships and 56.2% enter as boss firms. Collec- tively, the entry and persistence dynamics that we model offer some insight into the relative paucity of partnerships in reality, as highlighted in the introduc- tion. Conditional on equal numbers of teams beginning the pre-entry learning process in boss, partnership, and buyout decision-making structures, partnerships will represent a minority of the entrants.

Finding the Optimal Decision-Making Structure for Teams with Known Bias

The prior sections examined the relative perfor- mance of boss, partnership, and buyout decision- making structures on the expected profits of

populations of agents with varying characteris- tics, as well as the selection impact of decision- making structure on performance and bias character- istics of entrants and survivors in the marketplace. We next turn to a more prescriptive analysis by ask- ing: “Given two agents with known bias characteris- tics, which decision-making structure maximizes the expected profits for their venture?” The purpose of this prescriptive analysis is to suggest an optimal matching of decision-making structure to entrepre- neurs’ biases when these confidence biases are known.

Figure 10 displays the optimal decision-making structure for each pair of agents with varying levels of estimation and precision bias and L 5 1. The yel- low region depicts where buyout delivers the highest expected profits, the blue region where boss delivers the highest expected profits, and green indicates where partnership delivers the highest expected profits. The center block of the central panel pro- vides the optimal decision-making structure for unbiased-Bayesian agents, which is partnership. Focusing on the central panel, where precision biases are absent, partnership generates better results when agents’ initial beliefs are positively biased on average, and, conversely, buyout generates better results when agents’ initial beliefs are nega- tively biased on average. The results here are intui- tive: the “AND” structure of partnership corrects agents’ optimism by making both agents’ agreement a necessary condition for entry. Buyout option, by contrast, counteracts agents’ pessimism by requiring the assent of only one agent to enter. Furthermore, we also observe, in this central panel, that, if both agents have large and opposite estimation biases, the profit-maximizing structure is boss. In other words, in absence of precision bias, boss generates the high- est value only if the extreme and opposite estimation biases of the agents make both partnership and buy- out highly ineffective.

Moving beyond the central panel, we observe that the optimal decision-making structure for a pair of agents changes as we introduce their precision biases in addition to biases in estimation. Our prior work at the individual level shows that the severity of impact of an estimation bias depends on the preci- sion bias (Chen et al., 2018). Underprecision reduces the severity of initial estimation bias (either positive or negative), whereas overprecision exacerbates the severity of initial estimation bias, especially when agents are overly optimistic. These asymmetries, along with decision-making structure, shape the remaining patterns in Figure 10. Across most

2022 Chen, Elfenbein, Posen, and Wang 509

FIGURE 10 Optimal Decision-Making Structure as a Function of Both Estimation and Precision Biases

1

0.5

A ge

n t

1 i

n it

b el

ie f

0

1

0.5

A ge

n t

1 i

n it

b el

ie f

0

1

0.5

A ge

n t

1 i

n it

b el

ie f

0

1

0.5

A ge

n t

1 i

n it

b el

ie f

U n

d er

p re

ci se

Underprecise

Boss

Partnership

Zero entry

Indeterminate

Buyout

A ge

n t

1 P

re ci

si o n

Agent 2 Precision

O v er

p re

ci se

Overprecise

0

1

0.5

A ge

n t

1 i

n it

b el

ie f

0

1

0.5

A ge

n t

1 i

n it

b el

ie f

0

1

0.5

A ge

n t

1 i

n it

b el

ie f

0

1

0.5

A ge

n t

1 i

n it

b el

ie f

0

1

0.5

A ge

n t

1 i

n it

b el

ie f

Agent 2 init belief

0 0 0.5 1

Agent 2 init belief 0 0.5 1

Agent 2 init belief 0 0.5 1

Agent 2 init belief 0 0.5 1

Agent 2 init belief 0 0.5 1

Agent 2 init belief 0 0.5 1

Agent 2 init belief 0 0.5 1

Agent 2 init belief 0 0.5 1

Agent 2 init belief 0 0.5 1

Notes: Each plot displays the optimal decision-making structure, where applicable, for pre-entry learning duration L 5 1 at specified values of estimation and precision biases. The larger 9 3 9 grid varies precision bias from overprecise to underprecise values for agent 1 (vertical, bot- tom to top) and agent 2 (horizontal, left to right). Each smaller 9 3 9 grid varies the estimation bias from underestimated to overestimated values for agent 1 (vertical, bottom to top) and agent 2 (horizontal, left to right). In all simulations, each agent has opportunity costs of entry equal to 50. Performance values across decision-making structures are calculated as the average profit of 10,000 simulated two-agent teams. Blue, green, and yellow areas correspond to where boss, partnership, and buyout option is the optimal decision-making structure, respectively. For exam- ple, if agent 1 is overprecise and overestimates the initial probability of success (lower rows of larger 93 9, upper rows of smaller 9 3 9s) and agent 2 is underprecise but underestimates the initial probability of success (rows toward the right of larger 93 9, rows toward the left of smaller 9 3 9s), then buyout is likely to be the optimal decision-making structure (i.e., lower right of larger 93 9, upper left of smaller 9 3 9s are mostly dominated by buyout). Areas with a dot indicate zero entrants. Areas left blank and uncolored indicate two or more optimal decision-making structures.

510 Academy of Management Review July

combinations of precision bias, partnership tends to be the optimal structure when both agents have initial optimism, and buyout option tends to be the optimal structure when both agents have initial pessimism. Interestingly, the region in which buyout option dom- inates grows when at least one agent has underpreci- sion bias; it remains relatively fixed when both agents have overprecision bias. Underprecision makes indi- viduals’ beliefs more likely to fall below p�, which in turn differentially makes partnerships more likely to commit exit mistakes, since a single agent’s with- drawal of support terminates the operation. Finally, when both agents have strong, but opposing, under- and overprecision biases, boss is more likely to domi- nate. When team members draw very different con- clusions based on the same data (i.e., place very different weights on new information), then joint decision-making becomes ineffective, and it may sim- ply be better to ignore one agent’s perspective.

We expand upon these implications of these obser- vations for improving decision-making in the early stages entrepreneurship in the discussion below. Our insights complement traditional policies to address bias among entrepreneurs, which involve screening out highly biased individuals, de-biasing them through information and education, or structuring their learning processes. We posit that structure should follow bias. Specifically, our model suggests that, when individuals’ confidence biases are known, but are difficult or costly to correct, entrepreneurs should carefully select a decision-making structure that serves best to mitigate their biases.

Boundary Conditions and Limitations

Our model is a computational model rather than an analytical one, and we exercise it using a fixed set of parameters. We do not claim that our results hold for all possible values of these parameters, p, mH, mL, s, and d. We note, however, these parameters were cho- sen specifically to make the signal extraction problem in the model as difficult as possible—there is an equal ex ante likelihood of success and failure, and the mag- nitudes of the high and low signals are equal. Holding these assumptions fixed, in unreported work, we vary noise, s2, by a factor of four in either direction; the results remain qualitatively the same. Other changes to our parameters would make the signal extraction problem fundamentally easier (see Ryan & Lippman, 2003, for intuition). To our knowledge, this would not change the relative performance of the decision- making structures analyzed herein.

Furthermore, when we examine populations with biases, we apply a particular functional form to the distribution of biases. While we have also examined alternative distributions with greater dispersion than our main assumptions in the section titled “Comparison of Decision-Making Structures with a Population of Mean-Unbiased Agents” (see footnote 22), we cannot rule out the possibility that other distri- butions yield different conclusions. In “Comparison of Decision-Making Structures with Populations of Biased Agents” (also above), we explicitly focus on directly identifying the boundary conditions of our results by drawing from a series of beta distributions. Here, too, there is a possibility that alternative func- tional forms could yield different conclusions.

Finally, we assume that the pre-entry signals of our agents are uncorrelated—that is, that the mem- bers of the entrepreneurial team learn independently about the same opportunity prior to entry. This is an assumption that we can easily relax, by introducing a correlation in the noise team members receive. When agents are deterministically unbiased, the superiority of the partnership structure is robust to pre-entry noise correlation, though of course perfor- mance of all decision-making structures collapses to the single-agent case—that is, boss—when signals are identical (because, in that case, all agents’ beliefs are identical). When agents are drawn from popula- tions that include biases, the results remain qualita- tively the same and even quantitatively similar across the full range of noise correlation, because, in this case, the results are driven not only by evolving beliefs that are affected by the correlation, but also trait-like biases that are unaffected by it. (See Figures A4–A6 in the Appendix and the associated discus- sion of noise correlation between agents.)

Our study is not, of course, without multiple limita- tions, of which we can only include a partial list here. For tractability, we abstract away from issues of effort and investment through which an entrepreneur can affect their prospects for success. Our model thus rep- resents a process of passive, rather than active, learn- ing (see Pakes & Ericson, 1998, for a discussion). Relatedly, the notion of pivoting (changing to a new idea or approach), as exemplified in recent work on lean startups (Blank, 2013; Contigiani & Levinthal, 2019; Ries, 2011), is beyond the scope of our model. Moreover, teams are known to suffer from free-riding problems, which makes outcomes particularly sensi- tive to the distribution of rewards ex post, which we ignore. Additionally, we ignore (a) the possibility that optimism can affect not only whether the entrepre- neur exploits an opportunity, but also how that

2022 Chen, Elfenbein, Posen, and Wang 511

happens (e.g., Dushnitsky, 2010); (b) alternative mech- anisms that lead to overconfidence, such as asymmet- ric updating or confirmatory bias, which has been identified as empirically important in laboratory stud- ies of exit delay (Elfenbein et al., 2017); and (c) anchor- ing bias, which has been examined in survey-based work (Simon, Houghton, & Aquino, 2000). Our model is amenable to examining the impact of these and other biases, but, for simplicity, we have abstracted away from them. Furthermore, we have modeled the length of the pre-entry learning period as exogenous. Intuitively, endogenizing the entry timing against the backdrop of an opportunity that is declining in value will slow down the decision-making in a partnership relative to other structures, leading it to make fewer mistakes of all kinds. The net impact of this delay would depend critically on whether the partners face a single opportunity or a series of them. We believe that these interactions and extensions will provide fer- tile ground for future work.

CONCLUSION AND DISCUSSION

New ventures are commonly founded by teams of entrepreneurs, who must employ a decision-making structure that implicitly or explicitly defines how individual beliefs are aggregated into team deci- sions. How should decision-making be organized in entrepreneurial teams when founders exhibit confi- dence biases? We extend Chen et al. (2018)’s compu- tational model of entrepreneurial learning by (potentially) biased agents to include multiple team members who aggregate beliefs and make choices about entry and exit based upon different decision- making structures. We build on the work of Sah and Stiglitz (1986), Christensen and Knudsen (2010), and Csaszar (2013) in comparing the performance of structures in which all agents must agree to market entry and continuation decisions (partnership) to structures in which the affirmation of only one agent may be required to enter or exit the market (boss and buyout). Our model allows us to identify both the optimal decision-making structure when the estima- tion and precision biases of the team members are known ex ante and the relative performance of the different decision-making structures when the ex ante biases are unknown, but are drawn from known distributions.

We find that the strength of equal partnership vot- ing in two-person teams is in handling prospective entrepreneurs who are biased in the sense that they initially overestimate their likelihood of success. A buyout option can create value by enabling the team

to deal with an agent who has a pessimistic viewpoint or an idiosyncratically high opportunity cost, either of which, in the absence of a buyout option, can prevent profitable ventures from getting off the ground. At the same time, the buyout option also carries with it the winner’s curse. Its net impact, as we have shown, is a function of both estimation and precision biases as well as the quality of pre-entry learning.

A key contribution of this study is that it unpacks the determinants of effective decision-making for entrepreneurial teams. When agents are behaviorally biased, there is not a “one size fits all” solution. Rather, the appropriate decision-making structure for the entrepreneurial venture is contingent on the nature and distribution of potential founders’ biases.

Our work connects with recent studies that explore decision-making in early-stage entrepre- neurship (e.g., Bennett & Chatterji, 2019; Cohen, Bingham, & Hallen, 2019) and contributes to an important ongoing effort to find effective solutions for reducing decision-making errors therein. The empirical work of Camuffo et al. (2020) suggested that entrepreneurs who rigorously test their expecta- tions/hypotheses would be less likely to pursue proj- ects with false-positive returns. The lean startup approach (e.g., Blank, 2013; Contigiani & Levinthal, 2019; Ries, 2011) emphasizes reducing the costs of decision-making errors. Other approaches to reduce these errors include screening out entrepreneurs whose biases make them most prone to mistakes (e.g., Elfenbein et al., 2017; Gutierrez et al., 2020), providing clear information about success rates (Shane, 2008) and market noise (Elfenbein & Knott, 2015) to speed and improve the accuracy of the learning process, and suggestions from the practi- tioner literature for de-biasing processes (Lovallo & Sibony, 2010; Shepherd & Patzelt, 2017). By con- trast, the model we develop suggests that—for entre- preneurial teams at least—careful attention to decision-making structure, potentially matched to knowledge of team members’ biases, together with conscious decisions about gathering pre-entry infor- mation can reduce costly decision-making errors in early-stage entrepreneurship. As such, our work offers a remedy when other de-biasing processes are ineffective. Future research may continue to contrib- ute toward a richer variety of solutions for minimiz- ing decision-making errors in the face of bias. Additionally, our model recommends extending a pre-entry learning phase to address bias and offers a lens through which to examine the work of Leather- bee and Katila (2020), who examined differences in how teams that may or may not include management

512 Academy of Management Review July

experts seek out and process information about uncertain new ventures.

Our modeling approach speaks to the literature exploring the nuances of confidence biases (e.g., Moore & Healy, 2008), and how different sources of bias affect entry in general (e.g., Du, Li, & Wu, 2019) or entrepreneurial entry specifically (e.g., Gutierrez et al., 2020). The model we develop takes into account the connected influences of estimation and precision biases as well as pre-entry learning on entry, exit, and performance implications of found- ing teams. In thinking specifically about how hetero- geneous beliefs are aggregated into team decisions, our work is connected to studies of knowledge aggre- gation within teams (e.g., Taylor & Greve, 2006) as well as behavioral strategy more generally (Levin- thal, 2011; Powell, Lovallo, & Fox, 2011).

REFERENCES

Agarwal, R., Campbell, B. A., Franco, A. M., & Ganco, M. 2016. What do I take with me? The mediating effect of spin-out team size and tenure on the founder–firm performance relationship. Academy of Management Journal, 59: 1060–1087.

Aldrich, H. E., & Kim, P. H. 2007. Small worlds, infinite possibilities? How social networks affect entrepre- neurial team formation and search. Strategic Entre- preneurship Journal, 1: 147–165.

Alvarez, S. A., & Barney, J. B. 2005. How do entrepreneurs organize firms under conditions of uncertainty? Jour- nal of Management, 31: 776–793.

Åstebro, T., Herz, H., Nanda, R., & Weber, R. A. 2014. Seek- ing the roots of entrepreneurship: Insights from behav- ioral economics. Journal of Economic Perspectives, 28: 49–69.

Åstebro, T., Jeffrey, S. A., & Adomdza, G. K. 2007. Inventor perseverance after being told to quit: The role of cogni- tive biases. Journal of Behavioral Decision Making, 20: 253–272.

Åstebro, T., & Thompson, P. 2011. Entrepreneurs: Jacks of all trades or hobos? Research Policy, 40: 637–649.

Beckman, C. M. 2006. The influence of founding team company affiliations on firm behavior. Academy of Management Journal, 49: 741–758.

Beckman, C. M., Burton, M. D., & O’Reilly, C. 2007. Early teams: The impact of team demography on VC financ- ing and going public. Journal of Business Venturing, 22: 147–173.

Bennett, V. M., & Chatterji, A. K. 2019. The entrepreneurial process: Evidence from a nationally representative survey. Strategic Management Journal. doi: 10. 1002/smj.3077

Blank, S. 2013. Why the lean startup changes everything. Harvard Business Review, 91: 63–72.

Braguinsky, S., Klepper, S., & Ohyama, A. 2012. High-tech entrepreneurship. Journal of Law & Economics, 55: 869–900.

Busenitz, L. W., & Barney, J. B. 1997. Differences between entrepreneurs and managers in large organizations: Biases and heuristics in strategic decision-making. Journal of Business Venturing, 12: 9–30.

Cain, D. M., Moore, D. A., & Haran, U. 2015. Making sense of overconfidence in market entry. Strategic Manage- ment Journal, 36: 1–18.

Camerer, C., & Lovallo, D. 1999. Overconfidence and excess entry: An experimental approach. American Economic Review, 89: 306–318.

Camuffo, A., Cordova, A., Gambardella, A., & Spina, C. 2020. A scientific approach to entrepreneurial deci- sion making: Evidence from a randomized control trial. Management Science, 66: 564–586.

Chen, J. S., Croson, D. C., Elfenbein, D. W., & Posen, H. E. 2018. The impact of learning and overconfidence on entrepreneurial entry and exit. Organization Science, 29: 989–1009.

Christensen, M., & Knudsen, T. 2010. Design of decision-making organizations. Management Sci- ence, 56: 71–89.

Cohen, S. L., Bingham, C. B., & Hallen, B. L. 2019. The role of accelerator designs in mitigating bounded rational- ity in new ventures. Administrative Science Quar- terly, 64: 810–854.

Contigiani, A., & Levinthal, D. A. 2019. Situating the con- struct of lean start-up: Adjacent conversations and possible future directions. Industrial and Corporate Change, 28: 551–564.

Cooper, A. C., Woo, C. Y., & Dunkelberg, W. C. 1988. Entre- preneurs’ perceived chances for success. Journal of Business Venturing, 3: 97–108.

Cooper, A. C., Woo, C. Y., & Dunkelberg, W. C. 1989. Entre- preneurship and the initial size of firms. Journal of Business Venturing, 4: 317–332.

Cope, J. 2005. Toward a dynamic learning perspective of entrepreneurship. Entrepreneurship Theory and Practice, 29: 373–397.

Cramton, P., Gibbons, R., & Klemperer, P. 1987. Dissolving a partnership efficiently. Econometrica, 55: 615–632.

Csaszar, F. A. 2013. An efficient frontier in organization design: Organizational structure as a determinant of exploration and exploitation. Organization Science, 24: 1083–1101.

Csaszar, F. A., & Eggers, J. P. 2013. Organizational decision making: An information aggregation view. Manage- ment Science, 59: 2257–2277.

2022 Chen, Elfenbein, Posen, and Wang 513

Cyert, R. M., & DeGroot, M. H. 1974. Rational expectations and Bayesian analysis. Journal of Political Economy, 82: 521–536.

DeTienne, D. R., Shepherd, D. A., & De Castro, J. O. 2008. The fallacy of “only the strong survive”: The effects of extrinsic motivation on the persistence decisions for under-performing firms. Journal of Business Ventur- ing, 23: 528–546.

Du, X., Li, M., & Wu, B. 2019. Incumbent repositioning with decision biases. Strategic Management Journal, 40: 1984–2010.

Dushnitsky, G. 2010. Entrepreneurial optimism in the mar- ket for technological inventions. Organization Sci- ence, 21: 150–167.

Elfenbein, D. W., Hamilton, B. H., & Zenger, T. R. 2010. The small firm effect and the entrepreneurial spawn- ing of scientists and engineers. Management Science, 56: 659–681.

Elfenbein, D. W., & Knott, A. M. 2015. Time to exit: Ratio- nal, behavioral, and organizational delays. Strategic Management Journal, 36: 957–975.

Elfenbein, D. W., Knott, A. M., & Croson, R. 2017. Equity stakes and exit: An experimental approach to decom- posing exit delay. Strategic Management Journal, 38: 278–299.

Evans, D. S., & Jovanovic, B. 1989. An estimated model of entrepreneurial choice under liquidity constraints. Journal of Political Economy, 97: 808–827.

Ewing Marion Kauffman Foundation. 2013, June 15. Kauff- man firm survey series. Retrieved from https://www. kauffman.org/entrepreneurship/reports/kauffman-firm- survey-series

Foss, N. J., & Klein, P. G. 2012. Organizing entrepreneur- ial judgment: A new approach to the firm. New York, NY: Cambridge University Press.

Foss, N. J., Klein, P. G., Kor, Y. Y., & Mahoney, J. T. 2008. Entrepreneurship, subjectivism, and the resource- based view: Toward a new synthesis. Strategic Entre- preneurship Journal, 2: 73–94.

Frank, M. Z. 1988. An intertemporal model of industrial exit. Quarterly Journal of Economics, 103: 333–344.

Galbraith, J. R. 1974. Organization design: An information processing view. Interfaces, 4: 28–36.

Ganco, M., Campbell, B. A., & Raffiee, J. 2017. Internal vs external markets: How the assembly of initial spin-out teams impacts spin-out survival. Academy of Management Proceedings, 2016. doi: 10.5465/ ambpp.2016.17556abstract

Ganco, M., Honor�e, F., & Raffiee, J. 2019. Entrepreneurial team assembly: Knowledge transfer, customer trans- fer, and matching models. In J. J. Reuer, S. F. Matusik, & J. Jones (Eds.), The Oxford handbook of

collaboration and entrepreneurship: 631–654. New York, NY: Oxford University Press.

Gimeno, J., Folta, T. B., Cooper, A. C., & Woo, C. Y. 1997. Survival of the fittest? Entrepreneurial human capital and the persistence of underperforming firms. Admin- istrative Science Quarterly, 42: 750–783.

Greenberg, J., & Mollick, E. R. 2018. Sole survivors: Solo ventures versus founding teams. Retrieved http://dx. doi.org/10.2139/ssrn.3107898

Gutierrez, C., Åstebro, T., & Obloj, T. 2020. The impact of overconfidence and attitudes towards ambiguity on market entry. Organization Science, 31: 308–329.

Hamilton, B. H. 2000. Does entrepreneurship pay? An empirical analysis of the returns to self-employment. Journal of Political Economy, 108: 604–631.

Hayward, M. L., Shepherd, D. A., & Griffin, D. 2006. A hubris theory of entrepreneurship. Management Sci- ence, 52: 160–172.

Hurst, E., & Lusardi, A. 2004. Liquidity constraints, house- hold wealth, and entrepreneurship. Journal of Politi- cal Economy, 112: 319–347.

Jovanovic, B. 1982. Selection and the evolution of indus- try. Econometrica, 50: 649–670.

Kerr, W. R., & Nanda, R. 2010. Banking deregulations, financing constraints, and firm entry size. Journal of the European Economic Association, 8: 582–593.

Kerr, W. R., Nanda, R., & Rhodes-Kropf, M. 2014. Entrepre- neurship as experimentation. Journal of Economic Perspectives, 28: 25–48.

Klein, P. G., Mahoney, J. T., McGahan, A. M., & Pitelis, A. N. 2019. Organizational governance adaptation: Who is in, who is out, and who gets what. Academy of Management Review, 44: 6–27.

Klotz, A. C., Hmieleski, K. M., Bradley, B. H., & Busenitz, L. W. 2014. New venture teams: A review of the litera- ture and roadmap for future research. Journal of Man- agement, 40: 226–255.

Knight, F. H. 1921. Risk, uncertainty, and profit. Boston, MA: Houghton Mifflin.

Knight, A. P., Greer, L. L., & de Jong, B. 2020. Start-up teams: A multidimensional conceptualization, integrative review of past research, and future research agenda. Academy of Management Annals, 14: 231–266.

Knott, A. M., & Posen, H. E. 2005. Is failure good? Strategic Management Journal, 26: 617–641.

Koellinger, P., Minniti, M., & Schade, C. 2007. “I think I can, I think I can”: Overconfidence and entrepreneur- ial behavior. Journal of Economic Psychology, 28: 502–527.

Lazar, M., Miron-Spektor, E., Agarwal, R., Erez, M., Gold- farb, B., & Chen, G. 2020. Entrepreneurial team forma- tion. Academy of Management Annals, 14: 29–59.

514 Academy of Management Review July

Leatherbee, M., & Katila, R. 2020. The lean startup method: Team composition, hypothesis-testing, and early- stage business models. Strategic Entrepreneurship Journal, 14: 570–593.

Lee, J. 2020. The benefits and costs of forming business partnerships. RAND Journal of Economics, 51: 531–562.

Levinthal, D. A. 2011. A behavioral approach to strategy— what’s the alternative? Strategic Management Jour- nal, 32: 1517–1523.

Levinthal, D. A., & Wu, B. 2010. Opportunity costs and non-scale free capabilities: Profit maximization, cor- porate scope, and profit margins. Strategic Manage- ment Journal, 31: 780–801.

Lovallo, D., & Sibony, O. 2010. The case for behavioral strategy. McKinsey Quarterly, 2: 30–43.

Lowe, R. A., & Ziedonis, A. A. 2006. Overoptimism and the performance of entrepreneurial firms. Manage- ment Science, 52: 173–186.

March, J. G., & Simon, H. A. 1958. Organizations. New York, NY: John Wiley & Sons.

Minniti, M., & Bygrave, W. 2001. A dynamic model of entrepreneurial learning. Entrepreneurship Theory and Practice, 25: 5–16.

Moeen, M., & Agarwal, R. 2017. Incubation of an indus- try: Heterogeneous knowledge bases and modes of value capture. Strategic Management Journal, 38: 566–587.

Moore, D. A., & Healy, P. J. 2008. The trouble with over- confidence. Psychological Review, 115: 502–517.

Packard, M. D., Clark, B., & Klein, P. G. 2017. Uncertainty types and transitions in the entrepreneurial process. Organization Science, 28: 840–856.

Page, S. E. 2008. The difference: How the power of diver- sity creates better groups, firms, schools, and socie- ties. Princeton, NJ: Princeton University Press.

Pakes, A., & Ericson, R. 1998. Empirical implications of alternative models of firm dynamics. Journal of Eco- nomic Theory, 79: 1–45.

Parker, S. C. 2006. Learning about the unknown: How fast do entrepreneurs adjust their beliefs? Journal of Busi- ness Venturing, 21: 1–26.

Parker, S. C. 2009a. Can cognitive biases explain venture team homophily? Strategic Entrepreneurship Jour- nal, 3: 67–83.

Parker, S. C. 2009b. The economics of entrepreneurship. New York, NY: Cambridge University Press.

Piezunka, H., Aggarwal, V., & Posen, H. E. 2020. Learning- by-participating: The dynamics of information aggregation in organizations (INSEAD working paper no. 2019/29/EFE). Retrieved from http://dx.doi. org/10.2139/ssrn.3425696

Posen, H. E., Leiblein, M. J., & Chen, J. S. 2018. Toward a behavioral theory of real options: Noisy signals, bias, and learning. Strategic Management Journal, 39: 1112–1138.

Powell, T. C., Lovallo, D., & Fox, C. R. 2011. Behavioral strategy. Strategic Management Journal, 32: 1369–1386.

Ries, E. 2011. The lean startup. New York, NY: Crown Business.

Ryan, R., & Lippman, S. A. 2003. Optimal exit from a project with noisy returns. Probability in the Engineering and Informational Sciences, 17: 435–458.

Sah, R. K., & Stiglitz, J. E. 1985. Human fallibility and eco- nomic organization. American Economic Review, 75: 292–297.

Sah, R. K., & Stiglitz, J. E. 1986. The architecture of eco- nomic systems: Hierarchies and polyarchies. Ameri- can Economic Review, 76: 716–727.

Sandri, S., Schade, C., Musshoff, O., & Odening, M. 2010. Holding on for too long? An experimental study on inertia in entrepreneurs’ and non-entrepreneurs’ dis- investment choices. Journal of Economic Behavior & Organization, 76: 30–44.

Shane, S. 2000. Prior knowledge and the discovery of entrepreneurial opportunities. Organization Science, 11: 448–469.

Shane, S. 2008. The illusions of entrepreneurship: The costly myths that entrepreneurs, investors, and pol- icy makers live by. New Haven, CT: Yale University Press.

Shepherd, D. A., & Patzelt, H. 2017. Researching entrepre- neurial decision making. In D. Shepherd & H. Patzelt (Eds.), Trailblazing in entrepreneurship: 257–285. Cham, Switzerland: Springer.

Simon, H. A. 1997. Administrative behavior: A study of decision-making processes in administrative orga- nization (4th ed.). New York, NY: Free Press. (Original work published 1947)

Simon, M., Houghton, S. M., & Aquino, K. 2000. Cognitive biases, risk perception, and venture formation: How individuals decide to start companies. Journal of Business Venturing, 15: 113–134.

Taylor, A., & Greve, H. R. 2006. Superman or the fantastic four? Knowledge combination and experience in innovative teams. Academy of Management Journal, 49: 723–740.

United States Census Bureau. 2016. Annual survey of entrepreneurs (ASE): Characteristics of businesses: 2016 tables [Data set]. Retrieved from https://www. census.gov/data/tables/2016/econ/ase/2016-ase- characteristics-of-businesses.html

2022 Chen, Elfenbein, Posen, and Wang 515

West, G. P. 2007. Collective cognition: When entrepreneur- ial teams, not individuals, make decisions. Entrepre- neurship Theory and Practice, 31: 77–102.

Woo, C. Y., Daellenbach, U., & Nicholls-Nixon, C. 1994. Theory building in the presence of “randomness”: The case of venture creation and performance. Jour- nal of Management Studies, 31: 507–524.

Wu, B., & Knott, A. M. 2006. Entrepreneurial risk and mar- ket entry. Management Science, 52: 1315–1330.

Zenger, T. R., & Lawrence, B. S. 1989. Organizational demography: The differential effects of age and tenure distributions on technical communication. Academy of Management Journal, 32: 353–376.

John S. Chen ([email protected]) is Assistant Professor of Strategy at the University of Florida. John obtained his PhD from the University of Michigan. His research applies an organizational learning lens to entrepreneurship, real options, and technology standards. John employs both traditional empirical approaches and computational methods in his work.

Daniel W. Elfenbein ([email protected]) is Professor of Organization and Strategy, and Associate Dean EMBA- Shanghai, at Washington University in St. Louis. He received his PhD in business economics from Harvard University. His research focuses on how firms and employees create and capture value through reputations and relationships, on entrepreneurship, and on corporate social responsibility.

Hart E. Posen ([email protected]) is the Richard G. and Julie J. Diermeier Professor in Business at the University of Wisconsin–Madison. He received his PhD from the Wharton School, after a career as an entrepreneur in the technology and retail sectors. He studies strategy, innova- tion, and entrepreneurship from a behavioral perspective that focuses on the role of organizational learning.

Ming zhu Wang ([email protected]) is a PhD candidate in strategy and entrepreneurship at Washington University in St. Louis. Her research interests are in entrepreneurship and changes in the competitive landscape. She applies experience from her prior work as a mathematician to her current research.

APPENDIX

ADDITIONAL RESULTS

Profit split in proportion to opportunity costs

Figure A1 shows an alternative rule, applicable to each decision-making structure, in which profit is split in proportion to opportunity costs. As the figure shows, profit is invariant to opportunity cost split under this rule.

FIGURE A1 Profit under a Partnership Rule That Splits Profits

in Proportion to Opportunity Costs

100 Boss Partner

Buyout Option

80

P ro

fi t

70

60

50 0.1 0.2 0.3 0.4 0.5

Proportion of team entry costs assumed by agent 1

90

Notes: For L 5 0.5, we plot the average performance under each decision-making structure as a function of the proportion of the total opportunity costs assumed by agent 1. Here, profits are assumed to be split in proportion to opportunity costs. Total opportunity costs of entry for both agents in all simulations equal 100. The x-axis represents the proportion of total entry costs pos- sessed by agent 1. For example, at x 5 0.3, agent 1’s opportunity cost is k1 5 30 and agent 2’s opportunity cost is k2 5 70.

516 Academy of Management Review July

DISTRIBUTIONS OF BIAS IN MODELED POPULATIONS

Figures A2 and A3 show the probability density functions for the distributions used in the main paper.

FIGURE A2 Estimation and Precision Bias Under b and Log b Distributions

Initial Estimation Bias

b (2,2)

Pessimist Optimist

2

1.8

1.6

1.4

1.2

1

0.8

0.6

0.4

0.2

0 0 1 0 1 2 3 4 5 60.5

2

1.8

1.6

1.4

1.2

1

0.8

0.6

0.4

0.2

0

Precision Bias

log2 b (2,2)

Overprecise

(stubborn)

Underprecise (jump to conclusions based on recent data)

Notes: The left panel plots a b(2,2) probability density function as the on-average-unbiased population distribution for the initial estimation bias. The function is centered on p̂2L5 0.5, the unbiased initial estimate probability of success. Values of p̂2Le (0.5, 1) denote overestimation bias (i.e., optimism), while p̂2Le (0, 0.5) denotes underestimation bias (i.e., pessimism). The right panel plots a log2b(2,2) probability density function as the on-average-unbiased population distribution for precision bias. The function is centered on t̂/t 5 1 (i.e., t̂/t 5 2x when x 5 0), the unbiased value for precision. For any x e (0, 1), we note that t̂/t 5 2x reflects as much underprecision as t̂/t 5 2-x reflects overprecision. The log2b(2,2) probability density function is produced by translating each (x, y) from the b(2,2) probability density function (left panel) to (2(x 2 0.5)�5, y). For example, the value at 0.5 in the left panel is the value at 1 in the right panel.

2022 Chen, Elfenbein, Posen, and Wang 517

FIGURE A3 Distributions from Which Agent Biases Are Drawn

2 b (2,3)

log2 b (2,3) log2 b (2,2) log2 b (3,2)

b (2,2)

Panel A.Estimation Bias

Panel B.Estimation Bias

b (3,2)

1

1.8

0.8

1.6

0.6

0.4

0.2

0 0 0.5

0 1 2 3 4 5 6 0 1 2 3 4 5 6 0 1 2 3 4 5 6

1 0 0.5 1 0 0.5 1

1.4

1.2

2

1

1.8

0.8

1.6

0.6

0.4

0.2

0

1.4

1.2

2

1

1.8

0.8

1.6

0.6

0.4

0.2

0

1.4

1.2

2

1

1.8

0.8

1.6

0.6

0.4

0.2

0

1.4

1.2

2

1

1.8

0.8

1.6

0.6

0.4

0.2

0

1.4

1.2

2

1

1.8

0.8

1.6

0.6

0.4

0.2

0

1.4

1.2

Notes: The b(2,3) and b(3,2) probability density functions reflect biased population distributions for estimation bias, with the former indicat- ing a population that (on average) initially underestimates the probability of success and the latter indicating a population that (on average) ini- tially overestimates the probability of success. The log2b(2,3) and log2b(3,2) probability density functions reflect biased population distributions for precision bias, with the former indicating a population that is (on average) overprecise and the latter indicating a population that is (on average) underprecise. Each log2b(a, b) probability density function is produced by translating each (x, y) from the corresponding b(a, b) probability density function (top row) to (2(x - 0.5)�5, y). For example, the value at 0.5 in a b(a, b) panel is the value at 1 in the correspond- ing log2b(a, b) panel.

518 Academy of Management Review July

CORRELATED NOISE TERMS

Next, we present the results of a robustness analysis conducted to confirm that the key results of the paper are robust to signal correlation between members of an entrepreneurial team (see Figure A4). We begin with our result that partnership is superior to buyout when agents are deterministically unbiased. Notice that, as one might expect, the result weakens considerably as the correlation increases—as the value of the correlation approaches 1, all structures become functionally equivalent to boss, since beliefs of all team members become identical.

Interestingly, the results are much more resilient to correlation, and indeed even grow stronger at low pre-entry learning, when agents are individually biased but, on average, unbiased (see Figure A5). The reason is that, unlike the deterministic case in which evolving beliefs account for any performance heterogeneity across the different structures, differences in this case also arise from agent initial biases, which are invariant to correlation.

FIGURE A4 Performance Relative to Partnership Across Correlation Levels (Deterministically Unbiased Case)

Signal correlation = 0.5 Signal correlation = 0.9Signal correlation = 0.1 15

10

0

–5

5

P er

fo rm

an ce

R el

at iv

e to

P ar

tn er

sh ip

( %

)

–10

–15 0.1

Length of pre-entry learning period, Λ 0.3 0.5 1 2 3 4 5 0.1 0.3 0.5 1 2 3 4 5 0.1 0.3 0.5 1 2 3 4 5

Boss Buyout

Notes: For three different levels of correlation, we plot the average performances under the boss and buyout decision-making structures rela- tive to partnership performance as a function of pre-entry learning duration (L). A log-scale is applied to the x-axis.

FIGURE A5 Performance Relative to Partnership Across Correlation Levels (On Average Unbiased)

20

10

0

−10

P er

fo rm

an ce

R el

at iv

e to

P ar

tn er

sh ip

( %

)

−20

Signal correlation = 0.5 Signal correlation = 0.9Signal correlation = 0.1

0.1 Length of pre-entry learning period, Λ

0.3 0.5 1 2 3 4 5 0.1 0.3 0.5 1 2 3 4 5 0.1 0.3 0.5 1 2 3 4 5

Boss Buyout

Notes: For three different levels of correlation, we plot the average performances under the boss and buyout decision-making structures rela- tive to partnership performance as a function of pre-entry learning duration (L). Agents are, on average, unbiased with estimation drawn from a b(2,2) distribution and precision drawn from a log2b(2,2) distribution. A log-scale is applied to the x-axis.

2022 Chen, Elfenbein, Posen, and Wang 519

Lastly, our results on the interaction of pre-entry learning duration and structure are robust to correlation. Recall our finding that when agents are, on average, optimistic, the buyout option outperforms partnership at low pre-entry learning duration, but partnership outperforms the buyout option at high pre-entry learning duration (e.g., Figure 7 right panels in the paper). This finding is highly robust to signal correlation, as seen in Figure A6.

FIGURE A6 Performance Relative to Partnership Across Correlation Levels (On Average Optimistic)

20

Signal correlation = 0.5 Signal correlation = 0.9Signal correlation = 0.1

Boss Buyout

10

0

–10

P er

fo rm

an ce

R el

at iv

e to

P ar

tn er

sh ip

( %

)

–20

0.1

Length of pre-entry learning period, Λ 0.3 0.5 1 2 3 4 5 0.1 0.3 0.5 1 2 3 4 5 0.1 0.3 0.5 1 2 3 4 5

Notes: For three different levels of correlation, we plot the average performances under the boss and buyout decision-making structures rela- tive to partnership performance as a function of pre-entry learning duration (L). Agents are, on average, optimistic with estimation drawn from a b(3,2) distribution and precision drawn from a log2b(3,2) distribution. A log-scale is applied to the x-axis.

520 Academy of Management Review July

Copyright of Academy of Management Review is the property of Academy of Management and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.