The More the Merrier? Diversity and Private Equity Performance

This paper explores how diversity among lead partner teams (LPTs) of private equity (PE) funds affects buyout performance. We argue that there is a trade-off between the “bright side” of diversity, i.e., improved decision-making due to a broader set of perspectives, and the “dark side”, i.e., deteriorated decision-making due to a potential for clashes and a lack of cooperation. Our theoretical framework suggests that the net effect on performance depends on whether LPTs are diverse in socio-demographic or occupational aspects. To test this hypothesis, we develop a comprehensive index that measures LPT diversity along six dimensions. Using a sample of 241 buyouts and 547 involved PE partners, we find that higher scores in the socio-demographic component (gender, age, nationality) are associated with higher deal returns and multiple expansion. The opposite is true for higher scores in the occupational component (professional experience, educational background, university affiliation). Further results suggest that the “bright side” of diversity gets relatively more important in case of complex buyouts and uncertain deal environments.


Introduction
The alternative investment (AI) industry has become a cornerstone of global wealth management.
Including hedge funds, private equity (PE), and venture capital (VC), it has about USD 12 trillion in assets under management (AuM), which represents 16% of global AuM (Boston Consulting Group, 2019). Despite its economic importance, however, the AI industry is still relatively young (Kaplan and Strömberg, 2009;Wood and Wright, 2010), and continues to be dominated by a homogeneous group of people, i.e., white men who attended elite business schools and came from investment banking or consulting (Dezelan, 2019;Cumming et al., 2019b). Women account for only 20% of AI professionals today, and represent less than 12% of senior positions globally (Preqin, 2019). Even more sobering are numbers from the U.S., the largest AI market, where women own 5% of PE firms and account for 3% of AuM in the PE industry (Lerner et al., 2019).
The situation is no better for race or ethnicity, with only 2% Hispanic and 1% black VC investors in the U.S. (Gompers and Kovvali, 2018).
The bias in the PE and VC industry has only recently begun to crumble 1 due to empirical evidence on diversity as a driver of performance. For example, several studies show that more diverse boards are less prone to financial restatements and fraud, and that female board representation is positively correlated with improved governance, environmental sustainability, and corporate social responsibility (Ben-Amar et al., 2017;Cumming et al., 2015;Wahid, 2019).
Research over the [2004][2005][2006][2007][2008] period has also shown that Fortune 500 boards with female directors have higher returns on sales and invested capital than their all-male counterparts (Wagner, 2011).
Consequently, Goldman Sachs CEO David Solomon recently announced they would not take a company public unless it has at least one "diverse" board member (Goldman Sachs, 2020).
Evidence on the impact of diversity on PE buyouts is scarce, despite the fact that they provide an interesting laboratory to study performance implications for several reasons. First, the lead partner teams (LPTs) combine the job elements of a top management team (TMT), as they are involved in strategic decision-making, with a supervisory board, as they closely monitor management (Metrick and Yasuda, 2011). Second, LPTs have powerful incentives to maximize their portfolio company's value within a short holding period, since large parts of compensation are tied to deal performance (Hammer et al., 2018). This creates ample channels for the diffusion of LPT characteristics to everyday decision-making at a portfolio-firm level.
A priori, the impact of diversity among LPTs on deal performance is ambiguous. Extant literature suggests a trade-off between the "bright side" and the "dark side" of diversity. The "bright side" highlights the value of additional perspectives and broader mindsets, which lead to more nuanced decision-making and improved information processing. The "dark side" suggests that diversity of group characteristics can create a potential for clashes among individuals and uncooperative behavior. The resulting inefficiencies may hamper information processing and impede decision-making.
We argue that the net effect of the "bright side" and "dark side" on PE performance depends on the concrete type of LPT diversity. We present an empirically testable taxonomy of diversity characteristics that differentiates between socio-demographic characteristics, such as gender, age, and ethnicity/nationality, as well as occupational characteristics, such as professional experience, education, and university affiliation. We argue that the net effect is positive when a group of individuals is diverse in terms of socio-demographic factors, because different perspectives are not the result of deliberate career choices. Thus, while LPTs may benefit from different mindsets based on "endowed" attributes, they may still share large parts of their cognitive bases and values, as reflected in common life choices such as work experiences or education. This creates a broad pool of opinions among LPTs while causing few diversity-related transaction costs. The opposite holds for diversity in terms of occupational characteristics. Because individuals voluntarily assume professional skills and attributes throughout their careers, they end up with distinct cognitive bases and values. Thus, LPTs with great occupational diversity may lack common ground and face barriers for communication and collaboration that potentially outweigh the value of additional perspectives.
Based on a sample of 241 PE buyouts with full demographic information on 547 PE partners, we find evidence for these conjectures. In our baseline regressions, socio-demographic diversity of LPTs is associated with significantly higher deal performance, as measured by enterprise-value growth rates as well as multiple expansion. The opposite holds true for occupational diversity in LPTs, which is associated with lower deal performance. Thus, when using a composite index including both dimensions, the negative impact of occupational diversity offsets the positive impact of socio-demographic diversity, leading to an insignificant relationship between overall diversity and deal performance. These findings are consistent with the idea that there can be too much of a good thing when diversity leads to a lack of shared cognitive bases and "common language" owing to different fields of expertise Pelled et al., 1999).
Finally, we investigate moderators for the impact of diversity on performance. Prior literature suggests that complexity and uncertainty requires a particularly broad set of individual backgrounds, expertise, and experiences to manage the deal environment so that diversity among LPTs may matter most in such situations (e.g., Carpenter and Fredrickson, 2001). Consistent with this idea, we find that the positive effect of socio-demographic diversity is more pronounced, and that the negative effect of occupational diversity is mitigated, when deals are complex or occur in uncertain environments.
Although there is some prior literature on buyout performance and the impact of governance as well as legal aspects (Cumming et al., 2007;Cumming et al., 2010;Wood and Wright, 2009), research on team characteristics and diversity is rather limited. We attempt to close this gap by taking advantage of a clean research design that analyzes the backgrounds and diversity dynamics of different LPTs. In particular, our contributions to the literature are threefold: First, we develop a theoretical framework on the relationship between diversity and performance suitable for the PE space. Second, we develop a novel and comprehensive diversity index that spans LPT characteristics along six dimensions, and which differentiates between socio-demographic and occupational LPT characteristics. The index is empirically testable and can be transferred to other key "decision-making units" inside and outside the AI industry, e.g., to boards and TMTs. Third, we provide an empirical examination of the impact of LPT diversity on PE deal performance, and investigate the channels behind those results as well as moderating factors.

Theoretical background and hypotheses
To structure our hypothesis development, we propose a framework reflecting the specifics of the PE industry that draws on prior literature on decision-making and diversity in TMTs and corporate boards. Our framework consists of five elements (Figure 1).

Figure 1. Research framework -Diversity and performance in private equity
The first element, the upper echelons theory (UET) of Hambrick and Mason (1984), provides the foundation for why the demographics of top managers matter. UET builds on the premise that most TMT decisions are the result of dealing with bounded rationality, i.e., situations of unlimited information but limited attention and selective perception (Cyert et al., 1963). Consequently, all strategic choices and thus company performance are a result of information filtering built on managers' cognitive bases and value sets (Hambrick and Finkelstein, 1987). Because these psychological dimensions cannot be measured precisely, UET proposes analyzing demographic characteristics, such as age, education, or work experience, as observable proxies instead (Finkelstein and Hambrick, 1996). Following this line of thought, group characteristics should matter when several individuals are involved in strategic decision-making (Hambrick and Mason, 1984).
The second element posits that UET is not just relevant for TMTs but also for other key decisionmaking units such as supervisory boards (Nielsen, 2010). Interestingly, LPTs in PE combine governance bodies with supervisory bodies, because as majority owners they are concerned with both strategic decision-making and monitoring. For example, LPTs frequently provide on-site support, and are involved in decisions such as capital structure management, M&A activity, and efficiency improvement programs. At the same time, their investment decisions may be predicated on contractual provisions such as board seats, veto rights, and contingent control rights (Metrick and Yasuda, 2011).
For the third element, prior literature on organizational diversity explains how group characteristics in "key decision-making units" affect performance and organizational outcomes (Nielsen, 2010;Mannix and Neale, 2005). We cite two key strands of earlier diversity research literature (Williams and O'Reilly, 1998). The first is based on Newcomb's (1956) similarity attraction theory and Turner's (1985) self-categorization theory, which suggest that a high diversity of group characteristics can lead to uncooperative subgroups, increased conflicts, and poor outcomes (e.g., Pelled et al., 1999;Jehn, 1997). The associated costs from these inefficiencies hamper information processing and decision-making (Auh and Menguc, 2005). The second consists of decision-making/information processing theory. It suggests a positive relationship between diversity and performance because heterogeneity in group characteristics provides valuable additional perspectives and helps to process information (Hoffman, 1959;Hoffman and Maier, 1961). van Knippenberg and Schippers (2007) suggest combining the two strands. They posit that all diversity dimensions can potentially disrupt group processes while simultaneously creating synergistic performance benefits.
The fourth element categorizes the impact of diversity via specific characteristics Anderson et al., 2011;Cannella, Jr. et al., 2008). Despite the expected advantages of diversity, research finds some downsides, such as communication or coordination costs (Williams and O'Reilly, 1998;Auh and Menguc, 2005). These may be interpreted as "transaction costs" (Lazear, 1999). Because there is no simplistic causal relation between higher overall diversity and performance, multiple studies have argued for a taxonomy of group characteristics to differentiate between non-chosen (socio-demographic) and chosen (occupational) attributes (Milliken and Martins, 1996;Richard and Shelor, 2002;Anderson et al., 2011;Gompers et al., 2016;Gompers and Kovvali, 2018;Cox and Blake, 1991;Mannix and Neale, 2005). 2 Our framework adopts this differentiation in order to link diversity to performance in a PE context.
We expect the impact of socio-demographic diversity to be positive, since it provides different perspectives that are not the result of deliberate career choices. LPTs may apply different "filters" based on non-chosen attributes, but may nevertheless share large parts of their cognitive bases and values. This creates a broad pool of opinions among LPTs, while causing few diversity-related transaction costs.
This assumption is in line with literature that analyzes individual aspects of socio-demographic diversity. Regarding gender diversity, for example, several studies focus on the performance impact of female representation (Farrell and Hersch, 2005;Joecks et al., 2013;Perryman et al., 2016;Carter et al., 2003). Perryman et al. (2016) conclude that increased gender diversity at a top management level delivers better firm performance (Tobin's Q) and reduces risk. Torchia et al. (2011) refer to critical mass theory, and suggest a minimum threshold of females on boards to overcome tokenism and ensure a significantly positive impact. There is also anecdotal evidence that gender-diverse LPTs outperform their all-male counterparts with regard to higher returns and reduced capital losses (Gottschalg, 2019). Research on nationality and age diversity shows that more diverse teams incorporate a wider range of experience from different environments (Nielsen and Nielsen, 2013;Pichler et al., 2019). They can gain mentoring benefits from having a balance between wisdom and youth (Anderson et al., 2011). In sum, this leads to our first hypothesis: H1: Socio-demographic diversity among LPTs is positively associated with LBO performance.
While occupational diversity also adds different perspectives, it comes at the expense of reduced execution speed. The key differentiating factor from socio-demographic characteristics is that people voluntarily acquire occupational attributes that reflect distinct cognitive bases and sets of values. These require coordination, and therefore may add transaction costs.
This assumption is in line with literature that analyzes individual aspects of occupational diversity. Regarding work experience, theory suggests that differences in functional areas of expertise (e.g., marketing vs. finance) may cause communication issues (Simons et al., 1999;Bunderson and Sutcliffe, 2002). Hambrick et al. (1996) link functionally heterogeneous teams to lower execution speed. Based on a sample of PE buyouts, Acharya et al. (2013) identify performance benefits for specific acquisition strategies (i.e., organic vs. inorganic) from a specialization of involved PE partners (i.e., with either industry/consulting or banking/accounting experience). Similarly, Jelic et al. (2019) analyze a sample of secondary management buyouts and find that PE directors' financial experience (i.e., banking/accounting experience) affects buyout profitability, while business education drives growth measures. Research on VC fund management teams gives further insight into educational backgrounds, and finds that specialization in a particular field is more valuable than diversification (Zarutskie, 2010). This leads to our next hypothesis: H2: Occupational diversity of LPTs is negatively associated with LBO performance.
The fifth element addresses the specific context of team decisions. Building on Hofer's (1975) contingency theory, Hambrick et al. (2005) identify organizational complexity and environmental uncertainty as conditions that make executive tasks more challenging. Board literature also leans on resource dependence theory (Pfeffer and Salancik, 2003) to highlight the potential benefits of board diversity in uncertain and complex environments (Hillman et al., 2000).
Regarding complexity, diversity theory suggests that knowledge-based differences are important for handling the overload of complex information (Hambrick, 1995). This logic holds in our LPT context, and is supported by PE industry reports suggesting that large and cross-border deals, as well as inorganic deal strategies, may increase complexity for PE firms (Preqin, 2015;Bain & Company, 2019). Distinct perspectives are helpful in mastering the deal environment and obviating additional transaction costs. If this logic holds, the positive impact of socio-demographic diversity should increase in the case of complex deals, and the negative impact of occupational diversity should revert. This leads to our third set of hypotheses: H3a: The positive effect of socio-demographic diversity on LBO performance is more pronounced in the case of complex deals.
H3b: The negative effect of occupational diversity on LBO performance is mitigated in the case of complex deals.
Regarding environmental uncertainty, the diversity literature suggests that decision makers should be matched with environmental context to realize their full potential (Cannella et al., 2008).
According to Matusik and Fitza (2012), this can be achieved through a diversity of knowledge stocks, i.e., diverse professional experiences and backgrounds. Transferring these results to a PE context, we expect that the ability to be flexible is critical during times of higher environmental uncertainty. LPTs with a diverse background may benefit from superior information assessment during opaque economic conditions (Carpenter and Fredrickson, 2001;Hammer et al., 2020). Thus, diverse perspectives, skills, and abilities should become relatively more valuable. If this logic holds, the positive impact of socio-demographic diversity will increase, while the negative impact of occupational diversity reverts in times of environmental uncertainty. This leads to our fourth set of hypotheses: H4a: The positive effect of socio-demographic diversity on LBO performance is more pronounced during times of environmental uncertainty.
H4b: The negative effect of occupational diversity on LBO performance is mitigated during times of environmental uncertainty.

Sample construction and distribution
For our main sample, we include all deals labelled as "private equity" or "leveraged buyouts" in Bureau van Dijk's (BvD) Zephyr database between January 1, 1997, andDecember 31, 2015. 3 This results in 17,401 global LBOs. Since we are interested in the impact of diversity on the performance of exited LBOs, we also require an "exit date" in Zephyr. This reduces the sample to 7,087 deals.
Following  and Arcot et al. (2015), we choose the compound annual growth rate of the target's enterprise value (EV CAGR) from entry to exit as the first proxy for deal-level performance. Compared to the realized equity internal rate of return (IRR), the EV CAGR has few data requirements, is not affected by leverage levels, and can be compared to the enterprise value development of non-PE-backed peers. Restricting our sample to LBOs with available information on entry and exit enterprise values further shrinks our dataset to 2,079.
Next, we account for multiple expansion as one of the most important value drivers Guo et al., 2011). We rely on EV/EBITDA multiple expansion (Achleitner et al., 2010;Achleitner et al., 2011) and on EV/sales multiple expansion that are frequently used in the literature Gilligan and Wright, 2014). To obtain these variables at entry and exit, we add accounting data from BvD's ORBIS database. Due to a lack of disclosure requirements in many countries, our sample shrinks to 686 observations when we require the availability of all three performance indicators.
Next, we need biographical information on the PE partners for each LBO, so we match our performance sample from BvD with PE partner information from the Preqin database. We are left with 263 LBOs that we can find in Preqin and for which LPT names are available. We then handcollect demographic information for each partner (if available) from LinkedIn, Bloomberg Executive Information Systems, and company websites. 4 This leaves us with a sample of 241 LBOs from 25 different countries that have full demographic information about the 547 PE partners (see Table A2 for sample construction details). Note that this sample is larger than those of most other comparable empirical diversity studies (Table A3). Table 1 presents the distribution of our sample. It shows buyouts and their involved partners by deal entry period (panel A) and geographic region of portfolio firm headquarters (panel B). As expected, most LBOs (55.2%) are from the buyout boom period (panel A of Table 1). The number of PE partners per deal steadily increases from an average of 1.5 during the new economy period (1997)(1998)(1999)(2000) to 2.6 during the post-financial crisis period (2011)(2012)(2013)(2014)(2015). The regional distribution (panel B) shows that 95.4% of deals origin from Europe (including the U.K.) and North America.
We observe more European (84.2%) than North American (11.2%) deals because disclosure requirements are higher than in the U.S. We compare our distribution to the whole sample of 17,401 buyouts that we initially pulled from BvD's Zephyr database (step one of our sample construction), and to Strömberg's (2008)

Variable measurement
We derive our main dependent variables abnormal EV CAGR, abnormal EV/sales multiple expansion (ME) and abnormal EV/EBITDA multiple expansion (ME) in three steps. First, we calculate the CAGRs of EV, EV/sales, and EV/EBITDA for each LBO from deal entry to exit.
Second, we calculate the respective industry median CAGRs based on listed industry peers within the same holding period from the largest available stock index of each country or region. We obtain respective country/region peers from Thomson Reuters' Datastream. In the few cases with insufficient data, we use global median levels. Third, we deduct the median of the industry-level performance from each buyout's performance metric.
Our three main explanatory variables are socio-demographic diversity, occupational diversity and total diversity. They are calculated as weighted diversity indices as per Blau (1977): where L represents the different manifestations of each diversity component, and p is the share within the PE partner team that is involved in the respective LBO. 5 For the total diversity index, we include six components: 1) gender Blau diversity -female/male ratio ( Farrell and Hersch, 2005;Anderson et al., 2011); 2) national Blau diversity -local/foreigner ratio based on target company's home country (Nielsen and Nielsen, 2013); 3) age diversityaverage age difference for each deal (Richard and Shelor, 2002); 6 4) work experience Blau diversity -operational/non-operational ratio (where we define operational as consulting experience and/or relevant industry experience in the same Fama-French 10-industry classification as the target company) (Jelic et al., 2019;Acharya et al., 2013;Zarutskie, 2010); 5) educational background Blau diversity -business only/other ratio field of study (Zarutskie, 2010); and 6) university network Blau diversity -number of distinct universities per partner (Fuchs et al., 2018). 7 We then split total diversity into socio-demographic diversity and occupational diversity. We divide each score by the maximum value to obtain a normalized value, ranging from 0 (complete homogeneity) to 1 (maximum diversity). 8 Further explanatory variables describe the context of each LBO. Complexity bases on dealrelated and firm-specific factors such as inorganic growth strategies , crossborder transactions (Meuleman and Wright, 2011;Russo and Perrini, 2006), and company size (Cornelli and Karakas, 2008). Uncertainty describes the degree of volatility/risk based on four drivers: crisis period (Cumming and Zambelli, 2013), economic policy uncertainty (Baker et al., 2016;Nagar et al., 2019), industry volatility (Boutchkova et al., 2012), and company age/maturity (Grilli et al., 2010).
We also include variables from Zephyr to control for deal-, firm-, and PE sponsor-specific characteristics: Inorganic deal reflects the post-buyout value creation strategy Hammer et al., 2017); holding period measures time from LBO entry to exit in years Achleitner et al., 2011); firm size uses the natural logarithm of the target company's EV ((ln) deal value) Achleitner et al., 2011); PEI 100 indicates the reputation of the PE sponsor Boone and Mulherin, 2011); and involved PE sponsors' prior experience is the natural logarithm of the lead sponsor's prior LBO experience ((ln) sponsor experience) (Hammer et al., 2017).
We also introduce a set of PE partner controls to account for the different quality levels of teams: share top tier work experience (Fuchs et al., 2017(Fuchs et al., , 2018; average work experience (Bottazzi et al., 2008); share chartered accountants ; share higher academic degree (Zarutskie, 2010); and average university ranking (Fuchs et al., 2017). See Table A1 for details on variable construction.

Methodology
To test H1 and H2, we estimate the following cross-sectional regression on our sample of 241 LBOs: where APVm represents abnormal EV CAGR, abnormal EV/sales ME, or abnormal EV/EBITDA ME; DVn are our main variables of interest (socio-demographic diversity, occupational diversity and total diversity); is an LBO control vector of deal-, firm-, and PE sponsor-specific characteristics (inorganic deal, holding period, (ln) deal value, PEI 100, (ln) sponsor experience); is the control vector of the PE partner characteristics (share top tier work experience, average work experience, share chartered accountants, share higher academic degree, average university ranking); and is a set of fixed effects (entry channel FE, 9 entry period FE, 10 and team size FE 11 ). If socio-demographic diversity is positively associated with performance, we expect the corresponding coefficient on DVn ( 1 ) to be positive. If occupational diversity is negatively associated, we expect the respective coefficient on DVn ( 1 ) to be negative.
To test H3a and H3b, we estimate the following cross-sectional regression: where complexity is our continuous context variable; DVn x Complexity is an interaction term between diversity variables DVn and context dimension complexity; and all other variables are as defined in Equation (2). If socio-demographic diversity is positively associated with LBO performance in the context of complexity, we expect the coefficient on DVn x Complexity ( 1 ) to be positive. If the negative effect of occupational diversity on LBO performance is mitigated in the context of complexity, we expect the coefficient on DVn x Complexity ( 1 ) to be positive.
To test H4a and H4b, we estimate the following cross-sectional regression: where uncertainty is our continuous context variable; DVn x Uncertainty is an interaction term between DVn and context dimension uncertainty; and all other variables are as defined in Equation (2). If socio-demographic diversity is positively associated with LBO performance in the context of uncertainty, we expect the coefficient on DVn x Uncertainty ( 1 ) to be positive. If the negative effect of occupational diversity on LBO performance is mitigated in the context of uncertainty, we expect the coefficient on DVn x Uncertainty ( 1 ) to be positive. for abnormal EV CAGRs, and 29.7% (median: 21.7%) for non-peer-adjusted EV CAGRs. These are in line with  and Acharya et al. (2013). 12 The basis for our multiple expansion calculations is also in line with the PE literature. 13 All of our main explanatory variables range from 0 to 1. Note that median socio-demographic diversity is significantly lower than median occupational diversity (10.9% vs. 27.7%), 14 reflecting the low degree of diversity. Table 4a shows the correlation matrix of our three diversity indices and their components. As the diversity components show relatively high correlations, our approach of grouping the components into socio-demographic and occupational diversity traits is important for avoiding multicollinearity.

Descriptive statistics
We do not discuss our control variables here for the sake of brevity. Note that our deal, portfolio company, and PE firm controls are very similar to those in recent PE literature. 15 For details, see Table 3.

Regression results
Table 5 reports our base results to test H1 and H2. First, we introduce diversity with its two major components, socio-demographic and occupational diversity, along with the relevant control variables and fixed effects. We analyze their influence on abnormal EV CAGR (specification 1), abnormal EV/sales ME (specification 3), and abnormal EV/EBITDA ME (specification 5). In specifications 2, 4, and 6, we investigate the influence of our total diversity index on abnormal EV CAGR, EV/sales ME, and EV/EBITDA ME, respectively.
In support of H1 and H2, we find that socio-demographic diversity has a significantly positive correlation (p < 0.06 and below) with all three performance indicators, while occupational diversity has a significantly negative correlation (p < 0.07 and below). Our results suggest that a 1-standard deviation increase in socio-demographic diversity at the mean corresponds to a 4. share higher academic degree, and average university ranking, we observe a positive but nonsignificant relationship with all performance variables.  (2)). The sample comprises of 241 worldwide LBOs from 1997 to 2015. The dependent variables are abnormal returns, i.e., compound annual growth rates of enterprise value (abnormal EV CAGR; specifications 1 and 2), EV/sales multiple (abnormal EV/sales ME; specifications 3 and 4), and EV/EBITDA multiple (abnormal EV/EBITDA ME; specifications 5 and 6) between deal entry and deal exit minus listed industry peer performance. Socio-demographic diversity, occupational diversity and total diversity, are the main explanatory variables. We control for partner, deal, and PE characteristics, as well as for entry channel, entry period, and team size fixed effects. See variable definitions in Table A1. We use robust clustering of standard errors that are displayed in parentheses. *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.  Table 6 reports our results for H3a and H3b. We interact our complexity score with the diversity dimension socio-demographic diversity, and explore its influence on abnormal EV CAGR (specification 1), abnormal EV/sales ME (specification 4), and abnormal EV/EBITDA ME (specification 7). In specifications 2, 5, and 8, we do the same for occupational diversity. In specifications 3, 6, and 9, we interact the complexity score with total diversity, and then analyze the influence on abnormal EV CAGR, EV/sales ME, and EV/EBITDA ME, respectively.
Consistent with H3a and H3b, the results show that the interaction term on socio-demographic diversity x complexity remains positive for all performance indicators (significantly positive (p < 0.06) in specification 7). Interestingly, occupational diversity x complexity now has a significantly positive coefficient (p < 0.10 and below) based on all three performance indicators. This indicates that, in complex situations, diversity becomes more beneficial in terms of both socio-demographic and occupational traits. Thus, specialization benefits decrease in challenging situations when a holistic evaluation becomes more valuable . In line with this, total diversity x complexity has a significantly positive relationship (p < 0.10 and below) with all performance variables; total diversity alone exhibits a significantly negative coefficient (p < 0.10) only for EV/sales (specification 6); and complexity on a standalone basis is statistically unrelated to deal performance. Similar results have been found in the management and board literature (e.g., Richard and Shelor, 2002;Anderson et al., 2011). The control variables are qualitatively similar to our previous findings. 16 Table 6. Abnormal performance relative to partner diversity and deal complexity The table presents the results of multiple regressions of abnormal deal performance on PE lead partner team diversity in the context of complex deals (see Equation (3)). The sample comprises of 241 worldwide LBOs from 1997 to 2015. The dependent variable are abnormal returns, i.e., compound annual growth rates of enterprise value (abnormal EV CAGR; specifications 1-3), EV/sales multiple (abnormal EV/sales ME; specifications 4-6), and EV/EBITDA multiple (abnormal EV/EBITDA ME; specifications 7-9) between deal entry and deal exit minus listed industry peer performance. In addition to the explanatory variables and controls used in our base regression of Table 5, we interact total diversity and its sub-indices with deal complexity. See variable definitions in Table A1. We do not display PE partner controls for brevity. We use robust clustering of standard errors that are displayed in parentheses. *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.
Confirming H4a and H4b, results show that all the interaction terms on socio-demographic diversity x uncertainty and occupational diversity x uncertainty display positive coefficients on abnormal performance with specifications 2, 4, and 7 showing a significantly positive coefficient (p < 0.10). In line with the literature, the multiple perspectives provided by greater diversity appear relevant for companies during uncertain times (e.g., Carpenter and Fredrickson, 2001;Cannella, Jr. et al., 2008). The interaction term total diversity x uncertainty has a significantly positive coefficient (p < 0.06) for EV CAGR and EV/sales. Total diversity alone exhibits a significantly negative coefficient (p < 0.10) only for EV/sales (specification 6). Uncertainty on a standalone basis has a negative correlation with EV CAGR (p < 0.02 and below). Control variables are qualitatively similar to our previous findings. 17 Table 7. Abnormal performance relative to partner diversity and deal uncertainty This table presents the results of multiple regressions of abnormal deal performance on PE lead partner team diversity in the context of uncertain deals (see Equation (4)). The sample comprises of 241 worldwide LBOs from 1997 to 2015. The dependent variable are abnormal returns, i.e., compound annual growth rates of enterprise value (abnormal EV CAGR; specifications 1-3), EV/sales multiple (abnormal EV/sales ME; specifications 4-6), and EV/EBITDA multiple (abnormal EV/EBITDA ME; specifications 7-9) between deal entry and deal exit minus listed industry peer performance. In addition to the explanatory variables and controls used in our base regression of Table 5, we interact total diversity and its sub-indices with deal uncertainty. See variable definitions in Table A1. We do not display PE partner controls for brevity. We use robust clustering of standard errors that are displayed in parentheses. *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.    Table A5). 18 Similar to our base results (repeated in specifications 3, 6, 9 in Table A5), we find that socio-demographic diversity has a positive correlation (p < 0.02 and below) with all abnormal performance indicators. In addition, occupational diversity exhibits a negative sign for abnormal performance (and a significantly negative coefficient regarding EV/EBITDA (p < 0.06); specification 7). Results also hold when excluding holding period in specifications 2, 5, 8 in Table A5 as in Acharya et al. (2013).
Second, we test for a potential omitted variable bias by following Frank (2000) and comparing impact thresholds. We calculate the impact threshold as the minimum product of the partial correlations between an unobserved variable and the predictor, and between an unobserved variable and the dependent variable that would lead to a non-significant relation. We apply this test for our base results (specification 2 in Table 5) and find that an unobservable variable would need to have a relatively high impact of 0.0419 to invalidate our significantly positive relationship between socio-demographic diversity and abnormal EV CAGR. We compare this value to the thresholds of our other independent variables (specification 3 in Table A6). The results show that the impact of an unobservable variable would need to be much higher than the highest impact (in absolute terms) of our other independent variables, i.e., occupational diversity (impact of -0.029). Thus, the significantly positive correlation between socio-demographic diversity and abnormal EV CAGR is strongly robust regarding any unobservable factor. We conduct these impact threshold tests for all specifications between each explanatory variable and each dependent variable, but do not include the results for brevity. In sum, we find a strong robustness regarding any uncontrolled external effects.
Third, we provide an alternate set of control variables (Table A7) to check our base results for robustness. We apply the same model specification as Acharya et al. (2013), and now control for share operation, organic deal, share science background, share MBA degree, share chartered accountants, average PE experience (tenure), (ln) deal value, and holding period, and we only include sponsor (adjusted) and entry period fixed effects as further controls (specifications 1, 3, 5 in Table A7). We again exclude holding period to validate these results in a further step (specifications 2, 4, 6 in Table A7). Similar to our base results in Table 5, socio-demographic diversity has a positive correlation (p < 0.02 and below) and occupational diversity has a negative correlation (p < 0.07 and below) with all abnormal performance indicators. We also find that the control variables (i.e., holding period and (ln) deal value) are very similar in sign and significance to our base results. 19 Furthermore, we perform sensitivity tests on our dependent variables and our main explanatory variables to address possible measurement errors. We replace abnormal performance, i.e., the difference between LBO performance and the median performance of listed industry peers from the target company's home country, with absolute LBO performance. To control for industry and regional effects, we add entry region 20 and industry 21 fixed effects in Table A8. For our main explanatory variables (Table A9), we apply the natural logarithms ((ln) total diversity, (ln) sociodemographic diversity, (ln) occupational diversity) to ensure that results are not driven by outliers.
In sum, we find that our base results are not sensitive to the choice of dependent variable, and are not driven by outliers in our diversity indices.

Discussion and conclusions
This study investigates the concept of diversity in a PE context, and relates it to performance.
Our study is most related to three studies. Acharya et al. (2013) identify performance benefits for specific acquisition strategies (i.e. organic vs. inorganic) from a specialization of involved PE partners (i.e. with either industry/consulting or banking/accounting experience). Jelic et al. (2019) analize a sample of secondary management buyouts and find that the PE directors' financial experience (i.e. banking/accounting experience) affects buyout profitability, while business education is driving growth measures. Siming (2014) shows that social networks arising from labor market movements of PE lead partners affect private equity firms' choices of financial advisors, as well as the sourcing, pricing, and performance of deals.
We add to these studies in various ways. First, we develop a novel theoretical framework tailored to the performance-driven context of PE. Second, we create a comprehensive diversity index based on LPT characteristics that is empirically testable inside and outside the AI industry, and allows for a taxonomy of various diversity characteristics. Third, for the first time, we show that socio-demographic diversity is positively associated with buyout performance, while occupational diversity has a negatively association. This supports the notion that the trade-off between the "bright side" and "dark side" of diversity depends on the concrete type of diversity.
We also show that the "bright side" of diversity becomes relatively more important in case of complex deals and uncertain deal environments, when a holistic assessment and adaptive thinking is required.

Managerial implications
Our paper has important implications for the PE and AI industry. As we find that sociodemographic diversity is positively associated with deal performance, our results may help to convince more PE firms that it pays off to employ a diverse work force. We expect that this is highly important in a number-driven industry such as PE. Our findings may also support non-profit initiatives such as Level20 that try to increase diversity in the notoriously homogenous PE industry.
More specifically, our results may help PE firms to find an adequate balance between too little and too much diversity in their hiring and staffing policies. Our results clearly show that it is beneficial to have as much socio-demographic diversity as possible in the work force, while it depends on the particular deal and its environment whether occupational diversity pays off. These findings suggest that it may be optimal to have a pool of individuals that are diverse in both aspects, but also that not every deal requires the maximum degree of diversity. Staffing policies should therefore carefully consider whether the deal and its environment require occupational diversity on top of socio-demographic diversity, as there can be too much of a good thing in an average deal.

Limitations and implications for future research
Although we carefully hand-collected and validated our dataset and carried out multiple robustness tests, our study has certain limitations, which could provide opportunities for future research. First, our research design does not allow for causal inferences between the diversity indices and LBO performance. 22 In a future qualitative research setting, it would be useful to investigate team performance using detailed questionnaires (through self-assessment of team performance before and after a lead partner was added to the team).
Second, our primary research focus was on the relation between diversity and performance on an individual buyout level. We showed that socio-demographic and occupational diversity matter for LBOs. We also find that, in the context of complexity and uncertainty, a holistic diversity perspective is more valuable. Future research could investigate how the positive association of diversity with LBO-level performance translates into fund-level performance. It would also be instructive to explore whether diversity and its performance impact changes hiring policies on a fund level over the medium to long term.  (KKR, 2020). 2 Although the terminology differs among studies, the key differentiation between the two dimensions remains largely the same. Some diversity traits are determined (e.g., gender, ethnicity, age), while others are developed by choice over time (e.g., work experience, education). We apply the "socio-demographic/occupational" terminology. Other studies use "visible/non-visible" (Milliken and Martins, 1996;Richard and Shelor, 2002;Cox and Blake, 1991;Mannix and Neale, 2005), "social/occupational" (Anderson et al., 2011), "endowed/acquired" (Gompers and Kovvali, 2018), or affinity/ability (Gompers et al., 2016). 3 Note that BvD's coverage determines the start date of our sample. We manually check deal completions with publicly available sources, because Zephyr does not always report deal status or type correctly. 4 We validate our full sample, including LBO performance and full partner information, using proprietary data from a PE fund-of-funds investor with $500 million-$1 billion in AuM. This proprietary data is especially valuable, because a key success factor for funds-of-funds is to assess detailed information on potential general partners, i.e., PE firms, their LBOs, and involved PE partners. 5 For example, there are two manifestations for gender Blau diversity: L = 2 (male/female). If a team consists of two female and four male lead partners, the respective diversity score is: 1 -((  Compound annual growth rate of target company's enterprise value (EV) from deal entry to deal exit minus the median compound annual growth rate of the EV of listed industry peers within both the same target country and time horizon. We follow Acharya et al.'s (2013) logic and apply excess performance variables. Values are corrected for obvious outliers and winsorized at the 5th and 95th percentiles.
LBO values: Bureau van Dijk (BvD); industry peers: Datastream Abnormal EV/EBITDA ME Compound annual growth rate of target company's EV/EBITDA multiple expansion (EV/EBITDA ME) from deal entry to deal exit, minus median compound annual growth rate of EV/EBITDA multiple expansion of listed industry peers within both the same target country and time horizon. EV/EBITDA is defined as enterprise value divided by respective earnings before interest, taxes, depreciation, and amortization (EBITDA). If accounting values at entry are not available, we apply the respective value from up to two years prior to or one year after the transaction, provided that the holding period is greater than two years. If accounting values at exit are not available, we apply the respective value from up to two years after or one year prior to the transaction, provided that the holding period is greater than two years. Values are corrected for obvious outliers and winsorized at the 5th and 95th percentiles.

Ibd.
Abnormal EV/Sales ME Compound annual growth rate of target company's EV/sales multiple expansion (EV/sales ME) from deal entry to deal exit, minus median compound annual growth rate of EV/sales multiple expansion of listed industry peers within both the same target country and time horizon. EV/sales is defined as enterprise value divided by respective sales. For treatment of not available accounting information see details in definition of Abnormal EV/EBITDA ME. Values are corrected for obvious outliers and winsorized at the 5th and 95th percentiles.

Ibd.
Non-peeradjusted EV CAGR Compound annual growth rate of target company's EV from deal entry to deal exit. Values are corrected for obvious outliers and winsorized at the 5th and 95th percentiles.

BvD
Non-peeradjusted EV/EBITDA ME Compound annual growth rate of target company's EV/EBITDA multiple expansion (EV/EBITDA ME) from deal entry to deal exit. EV/EBITDA is defined as enterprise value divided by respective earnings before interest, taxes, depreciation, and amortization (EBITDA). For treatment of not available accounting information see details in definition of Abnormal EV/EBITDA ME. Values are corrected for obvious outliers and winsorized at the 5th and 95th percentiles.

Ibd.
Non-peeradjusted EV/Sales ME Compound annual growth rate of target company's EV/sales multiple expansion (EV/sales ME) from deal entry to deal exit. EV/sales is defined as enterprise value divided by respective sales. For treatment of not available accounting information see details in definition of Abnormal EV/EBITDA ME. Values are corrected for obvious outliers and winsorized at the 5th and 95th percentiles. Ibd.

Variable
Description Source Occupational diversity Developed index measuring the diversity of "occupational" traits of the partner team involved in the respective LBO, calculated as per the Blau (1977) index on three components: work experience Blau diversity index, educational background Blau diversity index, and university network Blau diversity index (see Total Diversity definition for details). Index is normalized ranging from 0 (no diversity/single partner deal) to 1 (maximum diversity).
Average PE experience (tenure) Average experience in years of all involved private equity partners in the respective LBO. Ibd.

Average university ranking
Average university ranking of all involved private equity partners in the respective LBO. &&For each university that a partner obtained a degree in, we calculate the relative ranking compared to other universities in the respective country to account for local bias in university choice. We apply three different rankings to cover as many universities as possible: Preqin, Bloomberg, LinkedIn, company info Complexity Developed index measuring the complexity of an LBO based on three components: inorganic transaction -equals 1 if target company invests in another company and/or divests part of the company after the buyout, and 0 otherwise; cross-border transaction -equals 1 if target company's and PE firm's home countries are different and the PE firm does not have an office in the respective country; 0.5 if target company's and PE firm's home countries are different and the PE firm does have an office in the respective country, and 0 otherwise; and deal size -based on target company's EV at entry. Index is normalized ranging from 0 (least complex LBO) to 1 (most complex LBO).

Socio-demographic diversity
Developed index measuring the diversity of "socio-demographic" traits of the partner team involved in the respective LBO, calculated as per the Blau (1977) index on three components: gender Blau diversity index, national Blau diversity index, and age diversity index (see Total Diversity definition for details). Index is normalized ranging from 0 (no diversity/single partner deal) to 1 (maximum diversity).

Panel B: Independent variables/fixed effects (FE) (continued)
Variable Description Source Entry region Region in which the respective target company has its headquarters: "U.S. and Canada", "U.K.", "Western Europe", "Northern Europe", "Southern and Eastern Europe" and "Rest of the World". Ibd.

Holding period (duration)
Time in years from entry to exit of the respective LBO. BvD

Industry
Industry classification based on Fama-French 10 system. Ibd.
Inorganic deal Classification indicating the main strategy of the LBO (organic/inorganic). Binary variable that equals 1 if the target company conducted an add-on transaction/ divestment after the buyout, and 0 otherwise. Ibd.
(ln) Deal value Natural logarithm of deal value in millions USD. Ibd.
(ln) Sponsor experience Natural logarithm of previous LBO transactions executed by the lead private equity investor. Ibd.
Organic deal Classification indicating the main strategy of the LBO (organic/inorganic). Binary variable that equals 1 if the respective target company conducted no add-on transaction or a divestment after the buyout, and 0 otherwise (see also Inorganic deal). Ibd.

PE partner controls
All PE partner control variables that we do not display in certain regression tables for brevity: share top tier work experience, average work experience, share chartered accountants, share higher academic degree, and average university ranking (see respective definitions).
Preqin, Bloomberg, LinkedIn, company info PEI 100 2018 ranking of the 100 largest worldwide private equity firms. Binary variable that equals 1 if one of the private equity sponsors involved in the respective LBO is on this list, and 0 otherwise.

BvD; Private Equity International
Share chartered accountants Share of partners with a "chartered accountant" special qualification in relation to all private equity partners involved in the respective LBO.
Preqin, Bloomberg, LinkedIn, company info Share higher academic degree Share of partners with a "higher academic degree" in relation to all private equity partners involved in the respective LBO. "Higher academic degree" is defined as JD/MD/PhD, MBA, and/or Master/Diploma.

Ibd.
Share MBA degree Share of partners with an MBA degree in relation to all private equity partners involved in the respective LBO. Ibd.

Share operation
Share of partners with operational experience (defined as consulting experience and/or significant industry experience, i.e., multiple industry experience and/or experience in the same Fama-French 10-industry as the target company) in relation to all private equity partners involved in the respective LBO. Ibd.

Share science background
Share of partners with a "science" academic background (e.g., math, natural sciences, engineering, etc.) in relation to all private equity partners involved in the respective LBO. Ibd.

Share top tier work experience
Share of partners with "top tier" work experience in relation to all private equity partners involved in the respective LBO. Fuchs et al. (2017)  Private equity company that is involved in backing the respective LBO. Adjusted approach: We group PE companies that conducted less than two LBOs per year (≈ median LBO experience) in our database to minimize the loss of degrees of freedom. All other PE companies receive a binary variable that equals 1 if the respective PE company conducted the deal, and 0 otherwise.

Team size
Size of the partner team involved in each LBO: "Single partner" -one partner; "Duo" -two partners; "Medium team" -three partners; "Large team" -more than three partners.

Preqin
Total diversity Developed index measuring the diversity of the partner team involved in the respective LBO, calculated using the Blau (1977) index based on six components: gender Blau diversity index -female/male ratio; national Blau diversity indexlocal/foreigner ratio based on target company home country; age diversity (continuous variable, no Blau index) -average age difference for each deal divided by the maximum age difference of 23 years; work experience Blau diversity indexoperational/ non-operational ratio; operational defined as consulting experience and/or significant industry experience, i.e., multiple industry experience and/or experience in the same Fama-French 10-industry as the target company; educational background Blau diversity index -business only field of study/other ratio; and university network Blau diversity index -number of distinct universities per partner. The index is normalized ranging from 0 (no diversity/single partner deal) to 1 (maximum diversity within sample).
(ln) Total diversity Natural logarithm of Total diversity (see respective definition). Ibd.

Uncertainty
Developed index measuring the complexity of an LBO based on four components: crisis deal -binary variable that equals 1 if deal takes place during the New Economy crisis (2001) or Financial Crisis (2008, and 0 otherwise; policy uncertainty -normalized score ranging from 0 to 1 based on the Economic Policy Uncertainty (EPU) score of the target company's country at deal entry (Baker et al. (2016); uncertain industry -binary variable that equals 1 if peer industry (Fama-French 10 classification of the target company) experienced an above global average standard deviation of sales growth over the sample period (1997 to 2015), and 0 otherwise; and young companies -binary variable that equals 1 if target company was founded less than 15 years ago at deal entry (first quartile), and 0 otherwise.
BvD; Baker et al. (2016 -6), and EV/EBITDA multiple (abnormal EV/EBITDA ME; specifications 5-7) between deal entry and deal exit minus listed industry peer performance. We apply the same explanatory variables and controls as in our base regression in Table 5 (repeated in specifications 3, 6, and 9). For robustness, we now replace PEI 100 and (ln) sponsor experience with sponsor (adjusted) fixed effects (specifications 1, 4, and 7). In a further step, we follow Acharya et al. (2013) and exclude holding period (specifications 2, 5, and 8). See variable definitions in Table A1. We use robust clustering of standard errors that are displayed in parentheses. *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.
Dependent variable = CAGR of abnormal performance Enterprise Value EV/Sales EV/EBITDA Prime (1) Prime (2) Base (3) Prime (4) Prime (5) Base (6) Prime (7) Prime (8) (Table  5). The sample comprises of 241 worldwide LBOs from 1997 to 2015. The reference dependent variable is the abnormal compound annual growth rate of enterprise value (abnormal EV CAGR) and the reference explanatory variable, to be tested for omitted variable bias, is socio-demographic diversity (see specification 2 in Table 5). An omitted variable would have to be correlated at 0.205 with the predictor socio-demographic diversity (Correlation (v, X)) and at 0.205 with the dependent variable abnormal EV CAGR (Correlation (v, Y)) (conditioning on observed covariates) to invalidate an inference. Correspondingly, the impact of an omitted variable (as defined in Frank 2000) must be 0.205 x 0.205 = 0.0419 to invalidate an inference. To interpret the results, it is helpful to compare these thresholds to the correlation of the other independent variables with the predictor socio-demographic diversity (Correlation (v, X)) (column 1) and the dependent variable abnormal EV CAGR (Correlation (v, Y)) (column 2) as well as with their impact (column 3).
Partial correlations  -6), and EV/EBITDA multiple (abnormal EV/EBITDA ME; specifications 5-7) between deal entry and deal exit minus listed industry peer performance. We apply the same explanatory variables as in our base regression in Table 5. For robustness, we apply controls following the model specification of Acharya et al. (2013). We now control share operation of partners, organic deal, share science background of partners, share MBA degree of partners, share chartered accountants of partners, average PE experience (tenure), (ln) deal value, and holding period (specifications 1, 3, and 5). Following Acharya et al. (2013), we exclude holding period in a further step (specifications 2, 4, and 6). We now only control for entry period and sponsor (adjusted) fixed effects. See variable definitions in Table A1. We use robust clustering of standard errors that are displayed in parentheses. *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.  Table A8. Non-peer-adjusted performance variables relative to partner diversity This table presents the results of multiple regressions of deal performance on PE lead partner team diversity. The sample comprises of 241 worldwide LBOs from 1997 to 2015. For robustness, we now replace the dependent variables with alternative non-peer-adjusted compound annual growth rates of enterprise value, EV/sales multiple, and EV/EBITDA multiple. We apply the same explanatory variables and controls as in our base regression in Table 5 (repeated in specifications 4-6). In addition, we now apply entry region and industry fixed effects (specifications 1-3). See variable definitions in Table A1. We use robust clustering of standard errors that are displayed in parentheses. *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively. Holding period -0.066*** -0.040*** -0.026*** -0.068*** -0.046*** -0.031*** (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (ln) Deal value -0.078*** -0.039*** -0.029*** -0.057*** -0.023*** -0.024*** (0  , compound annual growth rates of enterprise value (abnormal EV CAGR; specifications 1 and 2), EV/sales multiple (abnormal EV/sales ME; specifications 3 and 4), and EV/EBITDA multiple (abnormal EV/EBITDA ME; specifications 5 and 6) between deal entry and deal exit minus listed industry peer performance. For robustness, we now apply the natural logarithm of our diversity variables, i.e. (ln) socio-demographic diversity, (ln) occupational diversity and (ln) total diversity. We apply the same controls as in our base results in Table 5. See variable definitions in Table A1. We use robust clustering of standard errors that are displayed in parentheses. *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.