Criminals, bankruptcy, and cost of debt

Abstract

We examine whether criminal records of CEOs and rank-and-file employees are associated with firms’ likelihood of bankruptcy, and whether lenders adjust their required cost of debt accordingly. We use a nationwide sample of private firms and criminal registers covering all firm employees. We find that the likelihood of bankruptcy is positively associated with the CEO’s criminal record and the proportion of employees with criminal records. We find some, though less robust, evidence that lenders price a firm’s loan higher when the firm’s CEO has a criminal record and when more of the employees have criminal records. The results suggest that the characteristics of firm employees represent a risk that, to some extent, is priced by lenders.

1 Introduction

Do characteristics of rank-and-file employees provide information about a firm’s risk? If so, are these characteristics associated with the cost of debt? Mounting evidence shows that lenders assess the management of borrowing firms and that this assessment influences lending decisions (Grunert et al. 2005; De Franco et al. 2017; Donelson et al. 2017; Bui et al. 2018). This attention devoted by lenders to top managers makes sense, given the extensive research that examines how managers influence firm outcomes such as financial reporting (Davidson et al. 2015), performance (Bennedsen et al. 2020), and risk-taking (Kallunki and Pyykkö 2013). However, recent research moves beyond the characteristics of top managers to examine the association between the characteristics of rank-and-file employees and firm outcomes. Due to data availability constraints, this evidence is typically based on indirect proxies, such as educational level (Call et al. 2017), religiosity (McGuire et al. 2012; Dyreng et al. 2012), or attitudes about gambling (Christensen et al. 2018) among people near firms’ headquarters, or it is limited to industries for which data are readily available (Amir et al. 2014a; Law and Mills 2019). In addition, practitioners have recently expressed interest in the value of information about a firm’s employees (SEC 2017).

This paper examines whether traits of both CEOs and rank-and-file employees are associated with firm risk and the cost of debt. To measure these traits, we rely on proprietary access to comprehensive criminal registers from Denmark, which cover all criminal charges in the country, dating back to 1980, on top managers and rank-and-file employees in our sample firms. We access each employee’s full criminal record, including convictions and investigations for crimes that led to case dismissals or acquittals. The records comprise felonies, misdemeanors, and legal infractions and thus cover both serious and petty crimes. We link individual employees and their criminal records to their employers and test how employee characteristics relate to firm outcomes in a much broader setting than has been used elsewhere.

The criminology literature predicts that crime is caused by a lack of self-control (Gottfredson and Hirschi 1990) or exposure to criminal peers (Akers 1973). Individuals lacking self-control are impulsive, risk-seeking, and shortsighted (Gottfredson and Hirschi 1990), characteristics that can lead to risk-taking.

Employees can influence firm risk in several ways. First, they can affect firm actions, such as investment decisions—a view supported in the literature. Graham et al. (2015) survey CEOs and CFOs and find that decisions about investments are commonly delegated to employees below the CEO or CFO level. In addition, research on employees in the financial industry finds that lending officers influence loan contracts (Campbell et al. 2019; Bushman et al. 2021) and that financial advisors with criminal records imperil their clients’ well-being (Law and Mills 2019). Second, employees can influence firm decisions indirectly through their influence on coworkers (called peer effects). Peer effects are documented across many academic disciplines (e.g., Sunstein 2002). For example, Dimmock et al. (2018) show that fraud is contagious among coworkers in financial advisory firms. Finally, employees can provide internal governance (Dyck et al. 2010; Acharya et al. 2011; Li 2019) by disciplining (or not) managers from making risky decisions.

We estimate three bankruptcy prediction models one at a time, to empirically test whether the criminal records of CEOs and employees relate to firm risk. Specifically, we estimate the models of Altman (1968), Ohlson (1980), and Beaver et al. (2005), which we complement with additional control variables motivated by the literature. We include additional firm-specific controls, including the wealth of a firm’s owners, earnings volatility, and employee counts. We further include personal controls for the CEOs and employees, such as their education, gender, and age. Finally, we add a variable for a CEO’s criminal record (an indicator of whether the CEO has a record) and a variable for employees’ criminal records (which measures the proportion of firm employees with criminal records). Incremental to all the control variables, we find that the criminal records of CEOs and employees help predict bankruptcies. We estimate that a CEO with a criminal record is associated with an increase in the likelihood of bankruptcy of 45–47 basis points or about 35%–36% of the unconditional mean. A one standard deviation increase in the percentage of employees with criminal records is associated with an increase in the likelihood of bankruptcy of 31–34 basis points or about 20%–22% of the unconditional sample mean.

We then examine the out-of-sample prediction accuracy, as measured by the area under the curve (AUC) statistic. Our results are as follows: (1) In addition to all the variables described above, the criminal records of CEOs and employees significantly improve the prediction accuracy. (2) The personal control variables do not collectively improve the prediction accuracy. And (3) a specification that includes criminal records, model-specific accounting variables, and firm controls but excludes personal control variables leads to the highest prediction accuracy. The two variables of criminal records increase the AUC statistic by 22–45 basis points, depending on the specification. The economic magnitudes are meaningful, although the increase in the AUC statistic is modest compared to related research.

We find that our measure based on the percentage of all employees with criminal records outperforms alternative measures, such as those limited to employees with the highest salary (highest within-firm quartile) or to non-CEO top managers, in terms of the out-of-sample bankruptcy prediction accuracy. We view this finding as consistent with the predictions regarding peer effects, in which employees with decision-making authority are influenced by coworkers (e.g., Dimmock et al. 2018). Limiting our measure to those at the top of companies thus omits this information, leading to impaired predictions. Although we cannot directly observe the influence exerted by coworkers, we can observe whether they are associated with other decisions in terms of committing new crime. Consistent with peer influence, we find that people are more likely to commit new crime when they start working in a company that employs more criminals. This holds both for individuals with and without a record prior to the employment.

We then condition our analysis on different types of crime. First, we examine the nature of crime and find that the prediction accuracy is larger when we use only white-collar crime than when we use other crime. We infer that white-collar crime drives our results, although it strongly correlates with other types of crime.¹ Second, we condition by the severity of crime. We do not find that crime penalized by imprisonment (the most serious category considered) leads to better prediction accuracy than less serious crimes. Third, we condition by whether crime is disclosed on the certificate of criminal record at hiring.² We find that undisclosed crime predicts bankruptcies and leads to the largest prediction accuracy. Finally, we condition by recent versus nonrecent crimes. Both predict bankruptcy, although nonrecent crime does so more accurately. This is consistent with the notion that crime is an observable outcome of an inherent trait that persists throughout life, as proposed by Gottfredson and Hirschi (1990).

We conduct several exploratory analyses. (1) We find some evidence that our results are concentrated among small firms with weak governance and among firms managed by a CEO without a criminal record. (2) Current changes in the percentage of employees with criminal records positively predict future changes in measures of firm risk (investments, growth, and debt), suggesting that employees’ criminal records convey information that later manifests in the accounting figures. (3) Criminal records of CEOs and employees are not associated with firm efficiency, on average. However, criminal records are positively associated with the likelihood of winning the “Gazelle Prize,” which is awarded to fast-growing and successful firms, thus indicating more right-skewed extreme firm outcomes. And (4) criminal records of CEOs and employees, to some extent, predict bankruptcies over longer horizons.

We then examine whether these risk factors are associated with the cost of debt, which we measure as the interest rate. While the literature demonstrates that lenders view the characteristics of the management of the borrowing firm as an important factor in the lending decision (Grunert et al. 2005; De Franco et al. 2017; Donelson et al. 2017; Bui et al. 2018), we are not aware of any research on the interplay of lending decisions and the attributes of borrowers’ employees. In our cross-sectional regressions, we find that firms pay higher interest rates when their CEOs have criminal records and when more of the employees have records. We then estimate panel models with firm fixed effects and find that the criminal records of CEOs are not significantly associated with interest rates, potentially due to the rarity of CEO turnovers in our sample. The criminal records of employees continue to be associated with firms’ interest rates, although the results are sensitive to our control variables.

In summary, our results indicate that the criminal records of CEOs and employees help explain the likelihood of bankruptcy. Our interpretation is that these characteristics represent a source of information about firms’ risk. Our estimations provide evidence that lenders charge more for debt when a borrower’s CEO has a criminal record and when more of the borrower’s employees have criminal records, although these results are not robust across different specifications.

We acknowledge that employees are not randomly assigned to firms (e.g., Van den Steen 2010), so we cannot rule out concerns about endogeneity.³ We conduct several tests to address these concerns. For example, we find that employees with undisclosed crime at hiring (no offenses appear on the certificate of criminal record) predict bankruptcies. That is, firms that unknowingly hire criminals are more likely to go bankrupt. This is consistent with matching of firms not conducting background checks (arguably a special type of firm) with record-holder employees not driving our results. Our results using propensity-score matching, a changes specification, and subsample estimations (where we condition on the CEOs having or not having a criminal record) help mitigate concerns about endogenous sorting driving our results.

We contribute to the literature in several ways. Our main contribution is to show that the traits of employees can be associated with firm outcomes. One strand of this literature approximates workforce characteristics using demographic variables of people living near firms’ headquarters (McGuire et al. 2012; Dyreng et al. 2012; Call et al. 2017; Christensen et al. 2018; Beck et al. 2018). Our results provide direct evidence, using the traits of actual employees.

Another strand of this literature examines professional services companies for which researchers can obtain sufficiently detailed data and in which the outcomes of the single employee are traceable (Amir et al. 2014a; Law and Mills 2019; Griffin et al. 2019). This stream documents that the traits of each employee are associated with that employee’s decisions. We make two contributions here. First, our results are consistent with the notion that employees influence their coworkers. Thus, they suggest how employees might affect firm outcomes. Second, we show that employee effects are not limited to professional services companies. For example, Law and Mills (2019) conclude that “financial advisors with pre-advisor criminal records … pose a greater risk to investors than those without” (p. 497). In this study, we show that the risk of criminal records permeates a large countrywide sample of firms in many different industries.

The remainder of this paper proceeds as follows. The next section discusses related research and develops testable hypotheses. Section 3 describes the sample and key measures. Section 4 outlines the research design and presents the results. Section 5 concludes and discusses possible limitations.

2 Related literature and hypothesis development

2.1 Overview of related literature

2.1.1 Individuals and corporate decisions

Since the formulation of the Upper Echelons Theory (Hambrick and Mason 1984), mounting evidence has emerged about the influence of top managers on corporate outcomes (e.g., review by Plöckinger et al. 2016). Several studies explore how manager characteristics are associated with risky corporate decisions. Researchers link these decisions to (1) observable off-the-job behavior, such as taking on leverage in personal real estate purchases (Cronqvist et al. 2012) and personal payment defaults (Kallunki and Pyykkö 2013); (2) experiences, such as military service (Benmelech and Frydman 2015) and exposure to natural disasters (Bonsall et al. 2017); (3) inherent characteristics, such as age (Li et al. 2017) or gender (Adhikari et al. 2019); and (4) proxies for psychological traits, such as overconfidence (Hirshleifer et al. 2012), risk-aversion (Graham et al. 2013), and sensation seeking (Cain and McKeon 2016; Sunder et al. 2017).

Recent research suggests that employee characteristics also can influence firm behavior. Due to data limitations, researchers use indirect geographic proxies, based on demographic variables of people surrounding a firm’s headquarters, including their gambling attitudes (Christensen et al. 2018), educational levels (Call et al. 2017; Beck et al. 2018), or religiosity (McGuire et al. 2012; Dyreng et al. 2012). Other researchers limit their studies to professional services companies for which they can obtain sufficiently detailed data, such as firms in the financial (Law and Mills 2019; Griffin et al. 2019; Campbell et al. 2019; Bushman et al. 2021; Honigsberg and Jacob 2021) or auditing (Amir et al. 2014a) industries.

We identify several channels through which employees can influence a firm’s risk. First, employees can directly influence corporate actions, such as investment decisions. Graham et al. (2015) survey CEOs and CFOs and find that investment decisions are most commonly delegated to employees below the CEO or CFO level. Relatedly, McElheran (2014) provides establishment-level empirical evidence on delegation of information technology (IT) investment decision rights and finds that 62% of the sample establishments have decision rights over nonpersonal computer IT investments.⁴ Second, employees can contribute to internal governance, as suggested by Dyck et al. (2010), Acharya et al. (2011), and Li (2019). Third, employees with certain traits (e.g., criminals) can influence firm decisions through their influence on coworkers, a phenomenon termed “peer effects.” Peer effects are documented across many academic disciplines (e.g., Sunstein 2002). Dimmock et al. (2018) show that financial advisors’ propensity to commit financial misconduct increases with the proportion of new coworkers with a history of misconduct following mergers and acquisitions. Murphy (2019) exploits the random assignment of US soldiers to units and finds that those assigned to units with more criminal peers are more likely to misbehave. Based on the Cambridge-Somerville Youth Study, with random assignment within pairs matched prior to treatment, Dishion et al. (1999) show that boys sent to summer camp (part of the treatment) were more likely to commit crime and experience other adverse life outcomes (pre-mature death, alcoholism, or psychiatric impairment).

Beyond the potential to influence firm behavior, employees might endogenously sort themselves into certain firms and thereby reflect corporate culture (Van den Steen 2010). That is, employees may self-select into firms that fit their traits, and firm managers may hire people who share their own traits.

2.1.2 Criminal records

The criminology literature provides several theories of the causes of crime. Two theories have received considerable attention. First, Gottfredson and Hirschi’s (1990) General Theory of Crime posits that a lack of self-control determines criminality, independent of the nature of the crime, and that crime provides easily accomplished and immediate gratification. The extent to which individuals lack self-control is determined in childhood and persists. Individuals lacking self-control are characterized as impulsive, risk-taking, and shortsighted. Second, Akers’ (1973) social learning theory argues that individuals learn criminality the same way they learn other behaviors—from peers.

Although the two theories offer opposing predictions, the literature provides empirical support for both. Pratt and Cullen (2006) conduct a meta-analysis of 21 studies and 126 size effects and conclude that both sets of variables, one set from each theory, strongly predict crime.⁵ In sum, we conjecture that the presence of a criminal record—an empirically observable outcome of a certain personal trait—affects decision-making either directly or through peer effects.

Several researchers within the finance and accounting literature associate criminal records with outcomes related to the firm or to individuals’ actions within the firm. Top managers with criminal records are associated with corporate outcomes such as the propensity to commit fraud, financial reporting risk (Davidson et al. 2015), earnings volatility, goodwill impairments (Amir et al. 2014b), and insider trading (Kallunki et al. 2018; Davidson et al. 2020). On a more granular level, some researchers focus on professional services companies in which each employee’s criminal record and production outcomes are traceable. For example, Amir et al. (2014a) find that audit partners with criminal records have riskier clients, and Law and Mills (2019) find that financial advisors with criminal records are more likely to receive future customer complaints along with other adverse outcomes. Honigsberg and Jacob (2021) show that financial advisors with adjudicated expungement requests (a process allowing brokers to remove financial misconduct from their public records) are more likely to misbehave in the future.

2.1.3 Bankruptcy prediction

Classic bankruptcy prediction models, such as Altman’s (1968) Z-score and Ohlson’s (1980) O-score, rely on accounting figures. Accounting-based econometric models are widely accepted due to their relatively high predictive power. However, researchers have complemented these models with factors based on stock return data⁶ (e.g., Shumway 2001; Chava and Jarrow 2004; Beaver et al. 2005) and macroeconomic information (Hillegeist et al. 2004) and find that doing so improves predictive accuracy. A limited amount of research investigates how observable manager effects may provide incremental information (e.g., Kallunki and Pyykkö 2013). However, to the best of our knowledge, no research examines the informational value of employee characteristics for bankruptcy prediction.

2.1.4 Cost of debt

Lenders rely on both hard and soft information when evaluating loan applicants or loan extensions (Liberti and Petersen 2019). Hard information, such as financial statement data, is undoubtedly important for lenders’ credit assessments (Agarwal and Hauswald 2010; Donelson et al. 2017). Lenders also collect and use soft information in their assessments. Grunert et al. (2005) analyze internal credit files of four German banks and find that nonfinancial (soft) factors, incremental to financial (hard) factors, improve the accuracy of probability-of-default estimations. Interestingly, they find that a factor capturing the lending officer’s subjective assessment of management quality significantly improves the prediction model. In a similar vein, the majority of the survey respondents of Donelson et al. (2017) indicate that, when they evaluate credit extensions, “character and reputation and experience of management” (Table 3, p. 2062) are among the most important factors, above “leverage and financial condition,” “guarantees,” and “liquidity.”

Agarwal and Hauswald (2010) describe the decisions of a large US bank lending to small firms. They note that each branch has considerable autonomy in its decisions but “has to justify any deviation from bank-wide practices on the basis of predefined subjective criteria, such as impression of management quality” (p. 2763). This suggests that the quality of borrower firm management is an important component of the lending decision. The results of De Franco et al. (2017) and Bui et al. (2018) suggest that managers of higher ability obtain lower bank-loan prices.

2.2 Hypotheses

Based on the extensive body of research on top managers and corporate outcomes—including studies that link criminal records to several corporate outcomes—we expect that firms whose CEOs have criminal records will have a higher likelihood of bankruptcy than other firms. We formally state this hypothesis as follows.

• Hypothesis 1a: A firm has a higher likelihood of bankruptcy when the CEO has a criminal record.

We expect that the percentage of employees with criminal records is also associated with firm risks that reflect the likelihood of bankruptcy. As outlined in Section 2.1.1, employees can affect firm outcomes through their influence on corporate policies and investment decisions, their internal governance role, and their sway with coworkers. Alternatively, employee characteristics can explain firm outcomes through sorting mechanisms, whereby employees opt to work for firms that share their traits. We expect employees’ criminal records to provide information about a firm’s risk, independent of the channel. This leads to our next hypothesis.

• Hypothesis 1b: A firm’s likelihood of future bankruptcy increases with the proportion of employees with criminal records. The effect is incremental to that of the CEO’s criminal records.

Lenders use evaluations of the management of a borrower firm in their credit assessments. We do not expect lenders to require the criminal records of borrower management.⁷ However, to the extent that the presence of a criminal record is an observable outcome of a certain personal trait, we expect that lenders can discover the type and traits of the borrower’s CEO. Therefore we expect that lenders charge a higher price when a CEO has a criminal record. This leads to our next hypothesis.

• Hypothesis 2a: A firm has a higher cost of debt when the CEO has a criminal record.

Lastly, we aim to explore the extent to which lenders adjust the cost of debt to the criminal records of the workforce of borrower firms. We are not aware of any studies assessing lenders’ pricing of borrowers’ workforce characteristics. We provide two sets of opposing arguments. The first set implies that lenders do not price the criminal records of a borrower’s workforce. Firms are not required to disclose workforce information beyond the number of people employed. And under the current Danish regulation, lenders can only access criminal records if all employees consent to share them with their employer’s lender. Our interviews with banks indicate they do not collect this information. The second set of arguments implies that lenders do price the criminal records of a borrower’s workforce. Lenders could indirectly learn about the records if they are reflected in firm behavior that lenders can observe. Our last hypothesis is therefore explorative, stated as follows.

• Hypothesis 2b: The cost of debt is increasing in the proportion of the borrower’s workforce with criminal records.

3 Sample construction, key variables, and descriptive statistics

3.1 Data sources and data description

Throughout our data sampling, we use unique firm identification numbers (CVR numbers) and unique personal identification numbers (CPR numbers) to merge datasets across sources. We use proprietary employment spells (employer-employee links) provided by Statistics Denmark to link the individuals, including their personal information, to the firms in which they work.

3.1.1 Firm-specific data

We gather financial statement data for all limited liability firms incorporated in Denmark for the period of 1998–2016 with total assets above DKK 1 million (EUR 0.13 million). We obtain data from Orbis, managed by Bureau Van Dijk, and complement that with data from Experian. The data include income statement items, balance sheet items, industry membership (NACE codes), full-time equivalent employee counts, and report publication dates. We hand-collect data on firm bankruptcies from Auktioner P/S, including firm identification numbers and filing dates.⁸ The bankruptcy data cover the period of 2004–2016.

3.1.2 Individual data and criminal records

We identify CEOs through firms’ filings with the Danish Business Authority. Through Statistics Denmark, we obtain access to the Integrated Database for Labor Market Research (IDAN database), which keeps data on employment spells, including annual data on salary received from the firm as well as starting and ending dates of employment.⁹

Statistics Denmark further provides access to the Danish Criminal Registry (Kriminalstatistik Afgørelse), which covers all criminal decisions from 1980. The dataset provides information on (1) judicial decisions, including criminal convictions and investigations for crimes that led to dismissals and not guilty verdicts, (2) penalties imposed on offenders, such as imprisonment, suspended sentences, and fines above DKK 1500 (EUR 200),¹⁰ and (3) the nature of the crime, based on seven-digit crime codes used by the Danish police. (The digit system has a tree structure, similar to industry classifications.) The offenses include felonies, misdemeanors, and legal infractions. The data thus cover serious crimes, such as sexual, violent, or drug-related offenses, and petty crimes, such as shoplifting. We use the crime codes to map the nature of crime reported in the Danish registers to the Federal Bureau of Investigation (FBI) definitions of general crime categories and white-collar crime, based on the conversion tables reported by Andersen et al. (2020), and present these mappings in Online Appendix C. We also use (4) the year of the criminal decision and (5) other information, such as length of incarceration.

Criminal records are not publicly available in Denmark. The Danish police can issue certificates of criminal records to individuals, who can then share them with employers (e.g., when applying for a job). The certificates include information on offenses of the Danish penal code and certain other offenses. Fines and suspended sentences appear on the certificates for two and three years following a conviction, respectively. Prison sentences appear for five years following release.¹¹ After this period, the crime is considered spent (comparable to sealing in the United States); that is, it is automatically removed from the certificate but appears in the police’s databases and, thus, in our proprietary dataset.

We estimate that employers ask for criminal records of new employees in less than 63% of new employments.¹² To investigate whether banks request criminal records of borrowers, we called several of the largest Danish banks and asked about their practices. These conversations revealed that lenders do not routinely collect criminal records of managers or employees in borrower firms, although Danish legislation does not prevent this. The lenders do sometimes request the criminal record of the CEO of a potential borrower as part of the “Know Your Customer” procedure in cases where the lender suspects that the firm is seeking to become a customer for financial-crime purposes (such as money laundering and terror financing).

3.1.3 Sample selection

We keep firm-years for the period of 2003–2015 to allow for a year’s lag between the last annual report and the bankruptcy filing. We merge the datasets and impose several screens. We exclude firm-year observations that do not cover 12 months, to make the observations comparable across firms and time. Consistent with the literature, we exclude certain industries (financial, utilities, and state-owned). To avoid double counting, we exclude subsidiaries for which the parent firm reports on a consolidated basis. We also impose several size thresholds. Based on the current auditing thresholds as outlined by Bernard et al. (2018), we keep firm-year observations with total assets of at least DKK 4 million (EUR 533,000) and at least 12 full-time equivalent employees. The minimum thresholds ensure that all of our sample firms are audited, prevent mom-and-pop stores from driving our results, and allow for variation in employee traits. We also impose an upper size threshold and keep only firms that conform to the small and medium-sized enterprises (SME) definition of the European Commission.¹³ Finally, we exclude observations for which data are missing in estimating the bankruptcy prediction models. Table 1 outlines the sample selection procedure. The final sample comprises 15,697 unique firms, 103,774 firm-years, 1,429,368 unique individuals, and 6,103,074 individual-firm-years.

Table 1 Sample selection

3.2 Key variables

3.2.1 Criminal records of executives and employees

On the individual level, we set an indicator variable, Record, equal to one if an individual has a criminal record and zero otherwise.¹⁴ As is standard in the literature, we include both convictions and criminal charges that led to dismissals or acquittals in our measure of criminal records.¹⁵ We do not include traffic-related offenses for two reasons. First, this is consistent with the literature. (See Bennett (2018) and Breining et al. (2020) for examples with Danish data, and Kallunki et al. (2018) for an example with Swedish data.) Second, many individuals in our sample have traffic-related records: 70% (37%) of CEOs (employees).¹⁶ At the firm level, we define the variable CEO_record as an indicator variable that takes the value one if the CEO of the firm has a criminal record (if Record = 1 for the CEO) and zero otherwise. We define the variable %EMPL_record as the percentage of a firm’s employees with criminal records.¹⁷ For each firm-year, we calculate the percentage of employees for whom Record = 1.

3.2.2 Bankruptcy variable and firm risk

We use the legal definition of bankruptcy to identify firms in financial distress, which is likely due to excessive risk-taking. Appiah et al. (2015) review the literature on corporate failure prediction and find that 84% of studies use the legal definition of bankruptcy to classify firms as failing or nonfailing. Hayden (2003) compares credit-scoring models with different default criteria (bankruptcy, restructuring, and delay-in-payment) and finds that models with bankruptcy as the dependent variable are as powerful in predicting credit losses as models with the alternative criteria as dependent variables, suggesting that the proxy for financial distress is of minor concern.

Our data contain bankruptcy notice dates—the dates when a bankruptcy court has ruled that a company must undergo bankruptcy proceedings. Under Danish regulation, a bankruptcy filing leads to firm termination (i.e., liquidation), similar to a Chapter 7 filing in the United States. Following the bankruptcy notice, a trustee is appointed, and the firm’s management loses control. The trustee sells off the assets and distributes the collected funds to creditors.¹⁸

We define an indicator variable, Bankrupt, which takes the value of one if the annual report is the last published report preceding the bankruptcy notice and zero otherwise.¹⁹ We use Bankrupt as the dependent variable in the bankruptcy prediction estimations.

3.2.3 Cost of debt

We use the interest rate to capture the cost of debt. We measure the interest rate as financial expenses divided by interest-bearing debt. We measure a firm’s interest-bearing debt as total liabilities net of trade payables. We then define the variable CoD as financial expenses scaled by the average interest-bearing debt for year t and year t-1. Related research uses comparable approaches of dividing interest expenses with debt (e.g., Minnis 2011; Vander Bauwhede et al. 2015; Gassen and Fülbier 2015). However, while data on actual debt and interest expenses are very limited in our dataset, related studies use actual debt and interest expense data in their estimations.²⁰ We acknowledge that our approach could contain noise. To mitigate the effect of outliers, we follow Minnis (2011) and truncate the CoD measure at the 5th and 95th percentiles and truncate observations more than 10 percentage points over the interest rate of Danish government bonds for the year.²¹

3.3 Descriptive statistics

We present descriptive statistics in Table 2. Columns 1 and 2 describe the sample. Columns 3–5 condition the sample by Bankrupt and compare the samples. The average sample firm is relatively small, with total assets of about EUR 6 million, a headcount of about 58, which in full-time equivalent employees corresponds to about 37. On average across firm-years, 18.8% of the CEOs and 17.1% of the employees have criminal records. The percentage of CEOs with criminal records is slightly lower than that reported by Kallunki et al. (2018), likely because our study employs Danish data, whereas Kallunki et al. (2018) use Swedish data. The average interest rate in the sample is 4.0%, which conforms closely to the officially reported average interest rate charged to Danish SMEs for the period 2007–2015 (4.4%) (OECD 2017 Table 3.10).

Table 2 Descriptive statistics

We observe that 29.3% of bankrupt firms have a CEO with a criminal record, compared to 18.7% of nonbankrupt firms. In Fig. 1, we further depict how the number of CEO crimes relates to the bankruptcy rate and generally find that the bankruptcy rate increases by the number of CEO crimes. The univariate statistics provide initial evidence supporting H1a.

Fig. 1

Bankruptcy frequency per number of CEO convictions. This figure depicts the bankruptcy frequency on the y-axis over the number of CEO convictions (CEO_conv) on the x-axis

Bankrupt firms on average have more employees with criminal records. Specifically, 22.3% of employees in bankrupt firms have criminal records, versus 17.0% in nonbankrupt firms. In Fig. 2, we plot bankruptcy rates per criminal employee quintile (within-year quintiles based on %EMPL_record), conditioned by the CEO’s having a criminal record (i.e., by CEO_record = 1 and CEO_record = 0). We observe, in both subsamples, that the bankruptcy rate increases with the percentage of employees with criminal records. This suggests that employee characteristics are incremental to CEO characteristics in explaining bankruptcy rates, and provides initial evidence supporting H1b.

Fig. 2

Bankruptcy frequency per quintile based on within-year %EMPL_record. This figure depicts the bankruptcy frequency on the y-axis over the quintile based on within-year %EMPL_record. To isolate the association between employees and bankruptcy frequency, we condition by the criminal record of the CEO (CEO_record = 1 and CEO_record = 0, respectively)

We tabulate a correlation matrix in Table 3. Bankrupt and CoD relate positively to both CEO_record and %EMPL_record. Table 4 provides information on the types of crime and the associated bankruptcy rate. Interestingly, the bankruptcy rate is higher across all CEO criminal record categories. Specifically, in column 2, we observe bankruptcy rates of 0.016–0.031, which are all higher than the unconditional mean bankruptcy rate at 0.013. We observe a similar pattern when we examine criminal records of employees in columns 4 and 5.

Table 3 Correlation matrix

Table 4 Criminal offense distribution and bankruptcy frequency by category of crime

4 Empirical design and results

4.1 Bankruptcy prediction models

To test the relation between the likelihood of bankruptcy and the criminal records of the CEO and employees, we estimate Eq. (1) with a hazard estimation (Shumway 2001), which equals a logistic regression with adjusted standard errors. Specifically, chi-squared statistics are divided by the average number of years per firm to correct the standard logit estimates. We estimate the following model.

$$ {\displaystyle \begin{array}{c}{Bankrupt}_{it}={\alpha}_0+{\beta}_1 CEO\_ recor{d}_{it}+{\beta}_2\% EMPL\_ recor{d}_{it}+{\beta}_3 AC{C}_{it}\\ {}+{\beta}_4 Firm\ \mathit{\operatorname{var}}{iables}_{it}+{\beta}_5 Person\ variable{s}_{it}+{\varepsilon}_{it}\end{array}} $$

(1)

for firm i in year t. Bankrupt, CEO_record, and %EMPL_record are defined in Section 3.2 above. ACC represents accounting-based variables used to predict bankruptcy by Beaver et al. (2005) (henceforth, the BMR model), Altman (1968), and Ohlson (1980), respectively.²²Firm variables is additional firm-level control variables motivated by the literature. These include the relative wealth of a firm’s owner(s) (Beaver et al. 2019), earnings volatility (Amir et al. 2014b), and the logarithm of employee counts. Person variables represents person-specific control variables for CEOs’ and employees’ other personal characteristics that the literature suggests are associated with firm outcomes. These include educational level (Call et al. 2017), gender (Adhikari et al. 2019), age (Belenzon et al. 2019), marital status (Roussanov and Savor 2014), and the corruption index at the country of ancestry (Liu 2016). We define all variables in Appendix A. We estimate Eq. (1) with three sets of ACC control variables (one model at a time) and further control for year and industry fixed effects.

The β₁ and β₂ slopes measure the extent to which CEOs’ and employees’ criminal records, respectively, provide information on a firm’s likelihood of bankruptcy (beyond what is explained by accounting variables). They are the coefficients used to test H1a and H1b.

We estimate the models and present the results in Table 5. Consistent with our expectations (H1a and H1b), across three different estimation models, we find that the likelihood of bankruptcy increases with the CEO having a criminal record (CEO_record) and more of a firm’s employees having criminal records (%EMPL_record). Although CEO_record and %EMPL_record are positively correlated (see Table 3), our estimations suggest that both variables predict bankruptcies incremental to each other. The economic significance is sizable. Using all the controls, the bankruptcy likelihood increases by 45–47 basis points (bps) when the CEO has a criminal record, or about 35%–36% of the unconditional sample mean. A one standard deviation (interquartile) change of %EMPL_record is associated with a change in the likelihood of bankruptcy of 26–28 (31–34) bps or about 20%–22% (24%–26%) of the unconditional sample mean.²³

Table 5 Likelihood of bankruptcy estimation

The proportion of employees with bachelor’s degrees or higher (%EMPL_HighEduc) is marginally significant in two of three estimations, but none of the other person-specific variables predict bankruptcies. The bankruptcy likelihood decreases with the wealth of the firm’s owners (EquityFirmOwner/TL and EquityPersOwner/TL). The accounting variables generally relate to the likelihood of bankruptcy as expected, although some variables are not statistically significant, likely due to high correlations between the variables. (For example, the correlation between EBIT/TA and EBITDA/TL is 0.78.)

4.2 Out-of-sample tests

We then analyze the out-of-sample predictive ability of the bankruptcy likelihood scores based on five different specifications. First, we present the results using only the prediction model variables (ACC) and the extra firm variables (Firm variables) (Specification A). The literature documents that these variables predict bankruptcies. We then stepwise add the personal variables (Specification B) and the criminal records of CEOs (Specification C) and employees (Specification D). For completeness, we also present our results excluding the personal variables (Person variables) (Specifications E and F). Within each specification, we compare the predictive accuracy of the estimations described above using area under the ROC curve (AUC) fit statistics. For each year t, we estimate the respective model and use the estimated coefficients to predict the out-of-sample bankruptcy likelihood for year t + 1.

We present the results in Table 6 with each of the three bankruptcy prediction models (BMR, Altman, and Ohlson). Compared to the specifications that include all the control variables (ACC, Firm variables, and Person variables), the AUC increases by 22–27 bps when we include the criminal records of CEOs and employees (Specification D versus B). The incremental AUC associated with employees’ criminal records is only marginally statistically significant in the Beaver and Ohlson models and insignificant in the Altman model. Note that the AUC statistic decreases when we include personal variables (Specification B versus A), likely because of the inclusion of many low correlating estimators, leading to spurious patterns being picked up in the learning sample (Beaver et al. 2019). However, this decrease is only statistically significant in the Altman model.

Table 6 Out-of-sample tests

When we exclude the personal variables from the estimation, the incremental AUC associated with CEOs’ and employees’ criminal records increases to 35–45 bps (Specification F versus A). With this specification, the criminal records of employees, incremental to the CEO’s criminal record, significantly improve the out-of-sample prediction accuracy (Specification F versus E). In addition, we find that specifications that exclude the personal variables but include the criminal records of CEOs and employees (Specification F) outperform all other specifications.

The AUC improvements that we document are modest compared to related research. Gutierrez et al. (2020, Section 7.4) show an increase in the AUC of 110 bps by including the auditor’s going-concern opinion, and Kallunki and Pyykkö (2013, Figs. 5 and 6) show increases in the AUC of 147–198 bps by including past payment defaults of the CEO and the firm’s directors. However, even small increments matter to a firm’s stakeholders. For example, Iyer et al. (2016) note that “a 0.01 (100 bps) improvement in the AUC is considered a noteworthy gain in the credit scoring industry” (p. 1565).

4.3 Types of employees

We then examine which employees’ criminal records are associated with bankruptcy. First, we assess which individuals are associated with the bankruptcy likelihood. As reported in Online Appendix A, we find that only the criminal records of one person, the CEO, are significantly associated with the bankruptcy likelihood. The findings suggest that the CEO is unique, consistent with the conclusions of Bennedsen et al. (2020).

We then examine which groups of employees predict bankruptcies, besides the CEO. If, on the one hand, employee groups with the authority to make decisions do so based solely on their own traits (as measured by their criminal records), we would expect that only their records would predict bankruptcies. If, on the other hand, employee groups with decision-making authority are influenced by coworkers (i.e., are subject to peer effects), we would expect the criminal records of all employees to illuminate bankruptcy likelihood.

We use the salary received from the firm to identify employees with decision-making authority. For each firm-year, we sort the employees into quartiles based on their salaries and calculate the percentage with criminal records within each quartile. We then use the quartiles to re-estimate modified versions of Eq. (1) (one estimation for each quartile) and benchmark the prediction accuracies with our main model using all employees. To preserve space and avoid repetition, we report only the results using the Ohlson model and note, in the text below, cases where the results are sensitive to using the Altman and the Beaver models. We use the Ohlson model because it predicts the most accurately. Panel A of Table 7 shows the marginal effects at the mean for each quartile. The marginal effects of each quartile are statistically indifferent. Panel B shows the out-of-sample AUC. Our main model, based on all employees, outperforms each of the models based on salary quartiles.²⁴ We find similar results when we use job titles (non-CEO top managers) instead of salary to identify employees with decision-making authority (results reported in Online Appendix A). We interpret our results as consistent with the predictions regarding peer effects.

Table 7 Types of employees

We cannot observe the corporate decisions by individuals in the company, as, for example, do Amir et al. (2014a) and Law and Mills (2019). However, we can observe individuals’ new crime. To test whether the behavior of employees is consistent with the predicted peer effect, we identify a sample of job changers and examine whether their propensity to commit crime is associated with the percentage of employees with criminal records and the criminal record of the CEO at the new employer.²⁵ Table 8 shows that individuals are more likely to commit new crime when they start working in a company where more employees have criminal records, consistent with employees being subject to peer effects. This relation holds for both employees with and without prior records.

Table 8 New employments and propensity to commit new crime

4.4 Types of crime

We condition our results by several types of crime. For each type of crime, we re-estimate Eq. (1) including only the type of crime in question. As in Section 4.3 above, we report only the results using the Ohlson model and note, in the text below, cases where the results are sensitive to using the Altman and the Beaver models. We present all results in Table 9. We describe the different types of crime and the results regarding the out-of-sample AUC below. We do not find that any of the marginal effects of CEO_record and %EMPL_record, using only the type of crime in question, differ significantly from the marginal effects reported in our main analysis (Table 5). The marginal effects of some crime types are, however, insignificantly different from zero, possibly because these crimes are rare. For example, the marginal effects of both CEO_record and %EMPL_record for crimes against society are insignificant at the 10% level. Just 2.2% (3.7%) of CEOs (employees) have committed such crimes.

Table 9 Types of crime

Nature of crime

We conduct our tests regarding the nature of crime using two FBI classification systems: (1) white-collar crime and its subcategories (fraud, corporate, legal), and (2) National Incident-Based Reporting System (NIBRS) classifications, which separate offenses into crime against persons, property, society, and other. The AUC of our main analysis is statistically larger than the AUC of any crime category except white-collar crime and fraud (a subcategory to white-collar crime). That is, these categories predict firm bankruptcies as well as the use of all crimes does.

We analyze, in more detail, the out-of-sample AUC for white-collar crime and its subcategories. The AUC is statistically larger for white-collar crime than for nonwhite-collar crime (p value <0.01). Within the category of white-collar crime, we find that the AUC of fraud is statistically higher than that of corporate and legal (p values in the range 0.02–0.03). Using the Altman model, the AUC of fraud is not larger than that of corporate white-collar crimes (p value = 0.12). The results provide some evidence that our main results are driven by employees whose criminal records pertain to white-collar crime, specifically fraud.

Nonwhite-collar crimes still predict bankruptcies, although less accurately. Moreover, the measures of the employees’ criminal records pertaining to different types of crimes are highly correlated, which impedes disentangling the effects of each type of crime.²⁶

Severity of crime

We condition individuals’ criminal records by the most serious crime on the record and determine the severity based on whether the crime is penalized by imprisonment, suspended sentences, and other outcomes. Other outcomes include mainly fines but also diversion or deferred adjudications, military penalties, treatment sentences, and other penalties, as well as dismissals and acquittals. Our results show that the marginal effects of CEO_record and %EMPL_record using serious crimes are statistically indistinguishable from those reported in our main analysis (Table 5). This is consistent with the findings of Law and Mills (Law and Mills 2019, Table 6.1 of the online appendix), who also observe coefficients that are statistically indistinguishable across the seriousness of crimes. We do not find that serious crime leads to better prediction accuracy. In contrast, we find that nonserious crime (captured by the “other” category) leads to the largest prediction accuracy. This is expected, because 2.7% of white-collar crimes lead to imprisonment, compared to 8.0% of all crimes in our dataset. Nonserious crime hence overlaps with white-collar crime—the type of crime that leads to the largest prediction accuracy—as described above. In addition, the majority of the criminal records include only nonserious crime. On the firm-year level, 85% (67%) of the CEOs’ (employees’) records do not include prior prison time or suspended sentences.

Disclosed and undisclosed crime

We condition by crime disclosed on the certificate of criminal records at hiring (individuals for which at least one offense appears on the certificate), undisclosed crime (individuals with prior criminal actions for which no offenses appear on the certificate), and crime committed following a hire (individuals who had not committed any crime before the employment). We find that undisclosed crime leads to the highest prediction accuracy, as measured by the out-of-sample AUC. This is consistent with the notion that crime is an observable outcome of a trait and persists throughout life, as proposed by Gottfredson and Hirschi (1990). Moreover, this is consistent with the fact that matching of firms not conducting background checks (arguably a special type of firm) with record-holder employees does not drive our results.

Timing

We partition on crime committed before the end of year t – 3 (indicates that an individual had a criminal record at the end of year t – 3) and crime committed after the end of year t – 3 (indicates that an individual had committed crime after the end of year t – 3 and before the end of year t). We find that both types of criminal records are significantly associated with bankruptcy (%EMPL_record is significantly different from zero). However, using criminal records before the end of year t – 3 leads to a significantly larger AUC. This is consistent with our results regarding disclosed and undisclosed crime; that is, crime is an observable outcome of a trait.

4.4.1 Bankruptcy prediction: Additional tests

We perform several additional tests. In the following, we briefly describe each of these tests as well as the results. The Online Appendix B elaborates on results of these additional tests.

• Subsample analyses. We find some evidence that our results are concentrated among small firms with poor governance and among firms with a CEO with no criminal record.

• Financial performance. Using data envelopment analysis (Demerjian et al. 2012), we do not find that criminal records of CEOs and employees are associated with better firm efficiency on average (consistent with Law and Mills 2019). However, criminal records of CEOs and employees are positively associated with firms’ likelihood of wining the “Gazelle Prize,” which is awarded to successful, fast-growing firms. This indicates that criminal records correlate with extreme right-skewed outcomes (e.g., Levine and Rubinstein 2017).²⁷

• Changes in employees. Current changes in the percentage of employees with criminal records positively predict future changes in investments, growth, and debt. (We assess three periods: year t – 1 to year t, t + 1, and t + 2.) An increase in these variables should indicate an increase in the firm risk, suggesting that the criminal records of employees convey information that manifests later in the accounting figures. In addition, we find some evidence that changes in the percentage of employees with criminal records predict bankruptcies, although this only holds for changes over three years (p value = 0.07).

• Long-term prediction. We explore the extent to which information about criminal records of CEOs and employees helps predict bankruptcies for longer horizons. Using all the control variables, we do not find that criminal records significantly predict bankruptcies at extended horizons. In specifications without the person-specific control variables, criminal records of employees predict bankruptcies when we extend the horizon by up to three years. (At horizons extended by two or more years, the results are only marginally significant at the 10% level.) Criminal records of CEOs marginally predict bankruptcies when we extend the prediction horizon by one year (marginally significant at the 10% level). They lose their predictive power for longer horizons.

• Propensity-score matching. Our main findings are robust to using propensity-score matching, where we match bankrupt firms with nonbankrupt firms using all the control variables. This suggests that bankrupt firms’ being significantly different on these variables from nonbankrupt firms does not drive our results. We find similar results when we match on having a record-holder CEO or having a large proportion of employees with criminal records (above within-year median %EMPL_record).

4.5 Cost of debt

After having examined whether criminal records of CEOs and employees provide information about the likelihood of future bankruptcy, we turn to examine the consequences for the cost of debt.

We estimate Eq. (2) with OLS and cluster standard errors by firm and year (Gow et al. 2010) as follows.

$$ {\displaystyle \begin{array}{c}{CoD}_{it+1}={\alpha}_0+{\beta}_1 CEO\_ recor{d}_{it}+{\beta}_2\% EMPL\_ recor{d}_{it}+{\beta}_3 AC{C}_{it}\\ {}+{\beta}_4 Fir\mathrm{m}\ \mathit{\operatorname{var}}{iables}_{it}+{\beta}_5 Person\ variable{s}_{it}+{\varepsilon}_{it+1}\end{array}} $$

(2)

for firm i in year t. CoD is the cost of debt, which we measure with the interest rate. CoD is calculated as financial expenses scaled by average total liabilities net of trade payables (described further in Section 3.2.3). As with the bankruptcy prediction estimations, ACC represents accounting-based variables that are used to predict bankruptcy in the BMR, Altman, and Ohlson models. Firm variables are the extra firm controls, and Person variables are the controls for personal characteristics. That is, we apply the variables used to predict bankruptcy as control variables in our estimation of cost of debt, and thus estimate Eq. (2) with three sets of ACC control variables (one model at a time). We estimate the models with either of two fixed effect specifications: (1) year and industry fixed effects or (2) year and firm fixed effects (Gormley and Matsa 2014).

We present the estimation results in Table 10. For brevity and to avoid repetition, we report only the results using the Ohlson ACC variables and note, in the text, whether our results using the BMR or the Altman model provide different inferences. In our cross-sectional tests in columns 1 and 2, we find that firms whose CEOs have criminal records (CEO_record) experience higher interest rates, consistent with H2a. Economically, these CEOs pay higher interest rates of 14–18 bps, corresponding to about 3.5%–4.5% of the unconditional sample mean. In the cross-sectional estimations, we find mixed evidence with respect to the criminal records of employees. In column 1, where we exclude the personal control variables, we do not find that the criminal records of employees are significantly associated with the interest rate. (Using the Altman model, this coefficient estimate is positive and marginally significant, p value = 0.065.) When we include these controls, however, we find that firms where more employees have criminal records pay a higher interest rate, which provides some evidence for H2b.²⁸ The economic magnitude is small. A one standard deviation change in %EMPL_record is associated with a higher interest rate of about 4 bps (1% of the unconditional sample mean).

Table 10 Cost of debt

Columns 3 and 4 estimate Eq. (2) with firm fixed effects to test whether the associations above are mainly driven by cross-sectional variation in CEO_record and %EMPL_record. In these estimations, CEO_record does not explain variation in the firm’s interest rate. Few of the sample firms experience CEO turnover during our 13-year sample period, and even fewer change from a criminal CEO to a noncriminal CEO or vice versa. Only 1164 firms (7.4% of all sample firms or about 1.5% of the firm-year observations) experience the latter type of CEO turnover. We acknowledge that our tests have low power due to the rarity of CEO turnover. The results regarding %EMPL_record are comparable to the cross-sectional estimations.

Collectively, these findings provide some evidence that lenders require a higher cost of debt, in the form of higher interest rates, when lending to firms with a record-holder CEO and firms where a high proportion of employees have criminal records. Although our estimate of cost of debt is a noisy proxy, we generally find that the control variables predictably relate to cost of debt.²⁹

5 Conclusion

This paper examines whether criminal records of CEOs and employees provide information about private firms’ likelihood of bankruptcy and cost of debt. We conclude that firms whose CEOs have criminal records and firms where more employees have criminal records exhibit a higher likelihood of bankruptcy. We find some evidence that lenders price these criminal records, although these results are sensitive to different specifications. The likelihood of bankruptcy increases by 45–47 bps (35%–36% of sample mean) when a firm has a CEO with a criminal record and by 26–28 bps (20%–22% of sample mean) when the proportion of employees with criminal records increases by one standard deviation. Lenders require higher interest rates of 14–18 bps (3.5%–4.5% of sample mean) when firms have a criminal CEO and 4 bps (1% of sample mean) when the proportion of employees with criminal records increases by one standard deviation.

Our main contribution is to show that the traits of employees predict firm outcomes. First, we contribute to the literature by using a direct measure of employee traits instead of relying on measures based on the populations surrounding firms’ headquarters.³⁰ Second, we contribute to the literature that examines effects of employees with criminal records by (1) providing evidence consistent with predictions of peer effects, hence illuminating one way employees can influence firm outcomes,³¹ and (2) showing that the risk of hiring employees with criminal records permeates a large countrywide sample of firms in many different industries and is not isolated to the financial industry.³² Finally, we contribute to the literature on top managers’ criminal records and firm outcomes by documenting that criminal records of employees are associated with firm outcomes in a way that is incremental to the criminal records of CEOs and other top managers.³³

We caution the reader to interpret the results with care. First, our cost of debt estimations pertain to the interest rate only. We thus cannot rule out that criminal records of CEOs and employees are associated with nonprice loan terms, such as collateral, the use of covenants, and the length of the loan. Second, the criminal record data do not cover criminal actions before 1980, crimes committed outside Denmark, or crimes committed by a person without a Danish personal identification number (persons who were not born in and never resided in Denmark). However, these data limitations bias against our results because some individuals are classified as noncriminals when they may have criminal records not covered by our dataset. Finally, we cannot rule out concerns about endogeneity, although we conduct tests to address this concern. Despite these limitations, our results document that the criminal records of CEOs and employees provide information about a firm’s risk.

Appendix A Variable definitions

Variable	Definition
Dependent variables and variables of interest
Bankrupt	Bankrupt is an indicator variable that takes the value one for the last annual report published preceding a bankruptcy notice, and zero otherwise
CoD	Approximation for a firm’s interest rate. Financial expenses to average debt net of trade payables. $Co{D}_t=\frac{Financial\ expense{s}_t}{\left( Deb{t}_t+ Deb{t}_{t-1}\right)/2}$ Where Debt = Total Liabilities – Trade Payables CoD is truncated at the 5th and 95th percentiles. CoD is truncated for observations larger than 10 percentage points above the interest rate of Danish government bonds.
CEO_record	CEO_record is an indicator variable that takes the value one if the CEO has a prior criminal record (traffic related offenses excluded), and zero otherwise.
CEO_conv	The CEO’s number of prior convictions
%EMPL_record	%EMPL_record measures the percentage of the employees with criminal records.
Record	On the individual level, Record takes the value one if an individual has a prior criminal record and zero otherwise.
ACC variables. Accounting variables used by the bankruptcy prediction models. BMR , Altman , and Ohlson in brackets indicates that this prediction model uses the variable.
EBIT/TA (BMR, Altman)	$\frac{Earnings\ Before\ Interest\ and\ {Tax}_t\ }{Total\ Asset{s}_{t-1}}$
TL/TA (BMR, Ohlson)	$\frac{Total\ Liabilitie{s}_t}{Total\ Asset{s}_t}$
EBITDA/TL (BMR, Ohlson)	$\frac{Earnings\ Before\ Interest, Tax, Depreciation, and\ {Amortization}_t}{Total\ Liabilitie{s}_t}$
NWC/TA (Altman, Ohlson)	$\frac{NW{C}_t}{Total\ Asset{s}_t}$ Where NWC=Net Working Capital = WCA-WCL Where WCA = Working Capital Assets =Current Assets -cash and cash equivalents -properties held for sale -receivables from closely related parties WCL = Working Capital Liabilities =Current Liabilities -current part of mortgage -current part of bank debt -liabilities to closely related parties -dividends if included in current liabilities
RE/TA (Altman)	$\frac{Retained\ Earning{s}_t}{Total\ Asset{s}_t}$
BV/TL (Altman)	$\frac{Book\ Value\ of\ Total\ Equit{y}_t}{Total\ Liabilitie{s}_t}$
GP/TA (Altman)	$\frac{Gross\ Profi{t}_t}{Total\ Asset{s}_{t-1}}$ Due to disclosure exemption rules, private firms below specific size benchmarks are not required to publicly disclose revenues, and therefore we use gross profits instead of revenues.
Log (TA) (Ohlson)	Measure of size. The logarithm of total assets. Ohlson (1980) uses a related size variable, calculated as log (total assets/price level index). In untabulated analyses, we substitute Log (TA) with Ohlson’s size variable and find largely unchanged slope coefficient and standard errors.
CL/CA (Ohlson)	$\frac{Current\ Liabilitie{s}_t}{Current\ Asset{s}_t}$
NITWO (Ohlson)	NITWO is an indicator variable that takes the value one if the sum of current year’s earnings and last year’s earnings is below zero, and zero otherwise.
OENEG (Ohlson)	OENEG is an indicator variable that takes the value one if the equity is negative (TL/TA > 1), and zero otherwise.
CHIN (Ohlson)	$\frac{\Delta Net\ Incom e}{\left| Net\ Incom{e}_{t-1}\right|+\mid Net\ Incom{e}_t\mid }$
Firm variables
EquityFirmOwner/TL	EquityFirmOwner/TL measures the equity of the parent company, for example a holding company, if the parent company does not report on consolidated basis (observations for these subsidiaries are not included in the sample), scaled by the firm’s total liabilities.
EquityPersOwner/TL	EquityPersOwner/TL measures the equity of the personal owner scaled by the firm’s total liabilities. We define an individual as a personal owner of the firm if (1) he/she owns 100% of the company, or (2) ownership data are not available, but he/she is the founder of the company. We get the data from the registers at Statistics Denmark and note that private company stocks are not included in the personal equity calculation.
Log (Employees)	Log (Employees) is the logarithm of the number of full-time equivalent employees.
StdROA	StdROA is the standard deviation of ROA, using the last five years’ annual reports, requiring observations for at least two years.
Person variables
CEO_HighEduc	CEO_HighEduc is an indicator variable that takes the value one if the CEO has a bachelor’s degree or higher, and zero otherwise.
%EMPL_HighEduc	%EMPL_HighEduc measures the percentage of the employees with a bachelor’s degree or higher.
CEO_Female	CEO_Female is an indicator variable that takes the value one if the CEO is a female, and zero otherwise.
%EMPL_Female	%EMPL_Female measures the percentage of female employees.
CEO_log(Age)	CEO_log(Age) is the logarithm of the CEO’s age in years.
%EMPL_log(Age)	%EMPL_log(Age) measures the logarithm of the average age in years of the employees.
CEO_Married	CEO_Married is an indicator variable that takes the value one if the CEO is married, and zero otherwise.
%EMPL_Married	%EMPL_Married measures the percentage of the employees who are married.
CEO_CorrupIndex	CEO_CorrupIndex is the CorruptionIndex of the CEO’s country of ancestry. CorruptionIndex is the average of Transparency International’s corruption perception indexes for the years 1995–2018 multiplied by −1. Our data on ancestry country cover only two generations back in time. That is, the individuals in our dataset are classified as foreigners only if they or their parents are immigrants.
%EMPL_CorrupIndex	%EMPL_CorrupIndex measures the average CorruptionIndex of the employees. CorruptionIndex is the average of Transparency International’s corruption perception indexes for the years 1995–2018 multiplied by −1. Our data on ancestry country cover only two generations back in time. That is, the individuals in our dataset are classified as foreigners only if they or their parents are immigrants.
Other variables
TA	Total Assets.
Firm age	Firm age in years since incorporation.
Headcount	Headcount. Headcount measures the number of non-CEO persons who have received salary from the firm during the year. The measure does not distinguish between full time and part time workers.
Employees	Full-time equivalent employees.