1 Introduction

Job applicants with a criminal record are much less likely than others to receive an offer of employment. Recent audit studies suggest that lower human capital does not fully explain this difference, and that employers apply a hiring penalty to job applicants with a criminal background. Recent legislation and initiatives have attempted to improve the labor market prospects of individuals with a criminal record. One approach, known as Ban the Box, would restrict or prohibit employer inquiries about an applicant’s criminal record until a conditional offer of employment has been made. Alternative approaches include attempts to encourage voluntary efforts by employers either through persuasion or tax incentives.

The success of any approach to improve the employment prospects of those with a criminal record depends on why firms impose a hiring penalty. Are employers primarily concerned with potential workplace misconduct or are they using a criminal record as a proxy for the personality characteristics associated with job instability or poor performance? Is either fear founded on an accurate view of how individuals with records behave on the job if they are hired?

This paper addresses these questions leveraging a dataset drawn from the client firms of a hiring consultancy whose data has been previously used in academic work (Burks et al. 2015). The data consists of observations of individual applicants to, and employees holding, low-skill white-collar jobs, primarily at call centers. The data includes information available at the time of hiring, such as criminal record and job qualifications, and for those hired, it contains tenure-related outcomes such as length of service and, if the employee separated, the circumstances of separation.

In positions in which turnover is a major labor cost determinant, we find that workers with criminal records have a longer tenure and are less likely to quit their jobs voluntarily than other workers. This finding suggests that individuals with criminal records  represent an untapped productivity pool. However, when we disaggregate by job, we find that the association between criminal record and termination for misconduct is heterogeneous and differs between customer service and sales, suggesting the need for caution in drawing general conclusions.

We examine the differences in outcomes between jobs using psychometric test results of the type commonly used in hiring. First, we find striking differences between the psychometric predictors of poor outcomes in sales jobs and customer service jobs. Second, we find psychometric differences between sales people and customer service workers with criminal records. These results suggest that bad outcomes for sales people with records may be driven by differences either in worker or in job characteristics.

We then use detailed psychometric data on workers to show that criminal background predicts longer tenure even after controlling for personality traits. In theory, psychometric testing might provide a mechanism to allow employers to ignore criminal records entirely by testing directly for the characteristics that cause difficulties in a subset of employees with criminal records. In our data, we find that the psychometric variables have a reasonable degree of association with both a criminal record and job outcomes. However, they do not reduce the value of a criminal record in predicting misconduct, and further research is required to determine whether they could do so.

In considering the policy implications of these findings, one qualification is critical. Especially since employers apply a hiring penalty to individuals with criminal records, those who are actually hired are presumably more qualified than those who are not, and the instances in which we find better performance of employees with a criminal record cannot be generalized to the entire applicant pool. However, we believe we can make some reasonable inferences about marginal hires with a criminal record, and our results suggest that some employers could be missing profitable opportunities to hire low turnover workers. Our contribution to the existing literature is as follows. To the best of our knowledge this is the first paper to study the correlation between criminal background and productivity in a civilian setting. Unusually, we have data on the criminal background and psychological characteristics of all applicants merged with tenure data. Consistent with the literature that focuses on the hiring process, we find that applicants with criminal records are penalized at the hiring stage conditional on observable characteristics. Furthermore, we find that criminal history seems to be associated with better performance overall in customer service positions and is ambiguous in sales positions. These associations persist even after conditioning on psychological characteristics and other observables that employers use in the hiring process. These findings suggest that, at least with regard to customer service positions, there are unexploited opportunities to expand marginally the hiring toward applicants with a criminal background in a way that makes sense both on efficiency and on moral grounds.

2 Related literature

Observational studies have repeatedly found that job applicants with criminal records are much less likely than others to obtain legitimate employment (Western et al. 2001). Six months or so after release, 50 to 80% of the formerly incarcerated are not employed in the legitimate labor market (Visher et al. 2011; Petersilia 2003).

Establishing whether a causal relation exists between a criminal record and poor employment prospects is difficult, but studies using a variety of methods suggest that traditional human capital measures alone do not explain the hiring penalty associated with a criminal record, and that employers consider a criminal record a liability in job applicants.Footnote 1 Survey evidence confirms these findings and indicates that 69% of organizations conduct criminal background checks on all of their job candidates (Holzer et al. 2004; Society for Human Resource Management 2012).

The poor employment prospects of individuals with criminal records are of considerable policy concern. Technological changes continue to make background checking easier, and evidence suggests that lower costs have driven the increased use of these checks (Bushway 2004; Finlay 2009). The incarceration rate in the USA has grown considerably over the past several decades and now far exceeds that of any other country in the world. Over 650,000 offenders are released from prison each year (Carson and Golinelli 2013). The post-release employment prospects of inmates are of great practical consequence. Over half of released prisoners are reconvicted within 3 years (Durose et al. 2014). A failure to obtain legitimate employment is one of the strongest correlates of criminal recidivism, and recent evidence suggests that this relation may be causal (Uggen and Shannon 2014; Yang 2017).

Recent legislation and initiatives have attempted to improve the labor market prospects of applicants with criminal records. Some approaches restrict the use that employers can make of a criminal record. Ban the Box statutes prohibit employer inquiries about an applicant’s criminal record until a conditional offer of employment has been made and sometimes restrict the type of records that can be used or prohibit the employer from withdrawing the offer unless certain conditions are met.Footnote 2 The Equal Employment Opportunity Commission (EEOC) has challenged criminal background checks on the grounds that they have a disparate impact on African Americans and encourages employers to meet the requirements of disparate impact doctrine by procedures similar to those in Ban the Box laws.Footnote 3

Alternative policy approaches attempt to encourage voluntary employer efforts. The Obama administration launched an initiative called “Take the Fair Chance Pledge.” Businesses and educational institutions have been asked to commit to employing fairly those with a criminal record. Over 100 organizations in a variety of industries signed on, including such companies as American Airlines, Coca-Cola, Koch Industries, Google, Starbucks, and Walmart.Footnote 4 Related policies attempt to provide incentives or remove disincentives for hiring people with criminal records. The Work Opportunity Tax Credit allows employers to reduce their federal income tax liability by $2400 for hiring ex-felons within 1 year after their conviction or release from prison.Footnote 5

An employer who hires an applicant with a criminal record faces a double risk: the employee’s criminal record will preclude the employer from obtaining private insurance against misconduct, and if the employee commits a wrongdoing on the job that harms another individual, the criminal record is generally admissible as evidence of negligence. To address this, the Federal Bonding Program provides limited bonding for some employers during the first 6 months of an eligible employee’s employment.Footnote 6 Some jurisdictions have limited the extent to which an employee’s criminal record can be the basis of employer liability for negligent hiring.Footnote 7

Yet another approach attempts to improve the employability of people with criminal records by identifying those at low risk for recidivismFootnote 8 or improving human capital through job training or services directed toward individual change (Visher et al. 2017).

These policies are based on widely varying assumptions about the nature and accuracy of the business rationale for the hiring penalty. Employers often claim to use criminal records primarily from concern about liability for negligent hiring.Footnote 9 Fewer state, when asked, that they are using criminal records as evidence regarding the personality characteristics they seek. However, studies have found that employers overstate their overall willingness to hire applicants with a criminal record (Pager and Quillian 2005), and in surveys, they may well de-emphasize reasons for non-hiring that they perceived as socially disfavored, such as using criminal records as a personality proxy.Footnote 10

If little is known about exactly why employers use criminal records, still less is known about whether this use is based on accurate assumptions about how those with criminal records perform once hired or whether there are general characteristics that explain why employees with criminal records are at higher risk of bad job outcomes. Our paper aims to shed some light on these issues.

Only one other paper examines the job performance of people with criminal records. Lundquist et al. (2018) compare the performance of felons and non-felons using data drawn from the military. Like us, Lundquist et al. (2018) find that military personnel with a felony record are more attached to the job than other personnel and appear better on some performance dimensions, though different ones than we use. In contrast to our results, they find weaker performance and retention among those with a history of lesser offenses. On the other hand, Lundquist et al. (2018) provide an extensive qualitative analysis of selection procedures but not an estimate of the hiring penalty, while we provide an estimate of the hiring penalty with less institutional detail regarding the hiring process itself. We thus view the two papers as complementary, examining different labor markets, civilian, and military, and providing different perspectives on the selection process.

Somewhat more evidence bears on the personality characteristics of those with criminal records and the value of personality in predicting work outcomes. Psychologists define personality as “enduring patterns of perceiving, relating to, and thinking about the environment and oneself that are exhibited in a wide range of social and personal contexts” (American Psychiatric Association and others 2013, Glossary). In the last 20 years, the most frequently used framework for personality psychology has been trait theory, which typically begins with the responses of adult lay subjects to self-descriptive words (Goldberg 1990) or sentences (Costa and McCrae 1992a, 1992b). The personality evidence in our data consists of responses to 15 sentences of the type used in trait methodology.

The descriptive categories of trait analysis were developed by grouping the responses of a pool of subjects using factor analysis (Block 2010).Footnote 11 The most specific, lowest-level category is called a facet, of which there are generally thought to be about 25–30. Researchers agree that there is at least one higher-level category, factors, and the most common number of top levels is five, notably in the Five Factor Approach (FFA), sometimes called the Big Five (Lee and Ashton 2004). A commonly used version of the FFA describes the five factors as conscientiousness, agreeableness, neuroticism, extraversion, and openness to experience (Costa and McCrae 1992a, 1992b).

Although early work focused on higher-level factor analysis, more recent work has found that the less aggregated facet level is more predictive (Judge et al. 2013; Paunonen and Ashton 2001). Factors do remain useful because of data constraints as well as their use by much of the existing literature. For instance, each of the 15 questions in our data can be roughly associated with a factor, although, as will always be the case, most sentences load on more than one factor.

Although a number of previous studies have examined the predictive value of FFA in employment settings, most of these studies are small and study highly specific outcomes. Useful conclusions, therefore, require meta-analyses, of which the most comprehensive and recent are Judge et al. (2013) and Barrick et al. (2001). Both find that neuroticism usually has a negative effect on work outcomes and that, broadly speaking, all other factors have on average a positive effect, with conscientiousness the most important. However, there is a great deal of occupation and task specificity, with different factors predictive in different settings. The more recent of the two meta-analyses stresses that lower-level traits like facets are much more predictive than higher-level factors (Judge et al. 2013).

Relatively little work has been done on the relation between FFA and criminal behavior due both to the relative newness of the FFA and to the unpopularity among criminologists of personality-based theories compared with theories based on factors such as social class (Andrews and Bonta 2014; Jones et al. 2011).

Within this literature, most studies do find that the population of interest differs from the population as a whole in some personality traits. Most commonly, people with criminal records or related traits score high in extraversion and neuroticism and low in conscientiousness and agreeableness (Jones et al. 2011; O’Riordan and O’Connell 2014). Since low neuroticism and high scores on other factors tend to predict good job outcomes, these findings suggest that people with criminal records may not perform well on the job.

3 Data description and summary statistics

The dataset contains information on all the applicants to low-skilled white-collar jobs—typically, customer service or sales representatives in a call center. Unsuccessful applicants show up only once in the data. Successful applicants will typically occur repeatedly in the dataset, with reoccurrences indicating key HR events such as changes in position or termination. The data cover the period May 2008 to January 2014. The data are provided by a hiring consultancy whose business model was to provide a number of corporate clients with hiring recommendations. In the process, the consultancy administered pre-employment exams, including the psychological questions examined here.

After dropping repeated observations referring to the same worker, we are left with 1,163,384 observations, each of which refers to a unique applicant. We further drop a comparatively small number of observations that pertain to establishments located outside of the USA. We are left with 1,144,575 observations and will refer to this dataset as the “applicant pool.” Table 1, panel A, provides summary statistics about the available variables for the applicant pool. Note that fields are often missing. We believe this occurred because different clients requested that the consultancy collect different data and provided different elements of their own data for merging. For most applicants, we have data on schooling and prior jobs. The school variable is an indicator that equals 1 if the applicant has schooling above high school.Footnote 12 Two variables indicate the applicant’s stability on previous jobs. The fewer_short_jobs variable codes the answer to the question “In the last five years, how many full-time jobs have you held for less than six months, other than jobs you had while in school?” This variable takes value − 1 when the answer was “None,” − 2 when the answer was “One job,” all the way down to − 5 when the answer was “More than 6 jobs.” The variable longest_job codes the answer to the question “What is the longest amount of time that you ever worked for a single company?” This variable takes value 1 when the answer was “Not applicable/Less than 3 months,” all the way up to 6 when the answer was “More than 5 years.” The variable hired records whether the applicant was in fact hired.

Table 1 Summary statistics for the applicant pool

The variable pre_crim is a field that can be zero or one, depending on whether the applicant is recorded as having a criminal history. This field is recorded for only about 264,000 observations out of the entire sample. Because this is the key variable in this paper, we investigated possible reasons why it could be missing. We concluded, based on a cross-analysis with other missing fields, that the occurrences in which pre_crim is missing reflect a deliberate decision by an employer not to collect criminal records information, rather than a choice by an applicant to not respond.Footnote 13 We are somewhat reassured by this conclusion, but still we acknowledge that a selection bias might occur if, even within our narrow occupational range, some jobs entailed a higher risk or cost of crime. Such employers would be less likely than average to hire applicants with a criminal record. However, no evidence of such selection is evident in our data.Footnote 14 The applicant pool contains 110,023 observations that have data on criminal record, schooling, and job stability. The data do not contain the position_type, firm, or location for applicants who were not hired, but for many non-hired applicants we were able to reconstruct the values for position and location.Footnote 15 We were not able to reconstruct values for the firm applied to. However, the data indicate a good though not perfect association between firms and location. As discussed below, our sample of the hired was limited to sales and customer service workers, so for consistency, we also limited our applicant sample to these jobs. After dropping observations for other positions and where key fields were missing, we were left with 73,885 observations described in Table 1, panel B.

Table 2, panel A, provides summary statistics for the subset of applicants who were in fact hired. We refer to this subset as the “hired pool.” A number of additional variables are available for hired workers. For example, hired employees have an anonymized identifier of their employer called firm_id; a location field that encodes the city and state in which the employee was hiredFootnote 16; a position_type field describing the type of job held by the worker (agent, customer service, sales, technical support, or other)Footnote 17; a variable LOE recording the length of employment, in daysFootnote 18; and the cause of termination either voluntary (TERM_V) or involuntary (TERM_I) when known.Footnote 19 For 4.5% of our employees, the cause of termination was “misconduct.” All applicants took a personality test that including three proprietary questions and 15 FFA questions. The 15 FFA questions were grouped by the consultancy’s industrial psychologist into the standard FFA categories of conscientiousness, agreeableness, neuroticism, extraversion, and openness to experience.Footnote 20 Our variables neurtot, opentot, extratot, contot, and agreetot represent the sum total of answers chosen that support the respective personality factor of the FFA. These values can range from zero to three for each of these variables. Our variables neur1-neur3 and so on represent answers to individual questions that support the respective personality factor of the FFA. The three proprietary variables—badservice, confidence_regress, and rulebreaker—will be described in greater detail in Section 6.

Table 2 Summary statistics for the hired pool

The total number of hired applicants is 58,977. Criminal record was only available for a subset of the sample, leaving 18,142 observations of hired workers. In most of our estimates, the employee’s position was an important control, and the limited availability of pre_crim reduced the number of observations for jobs other than sales and customer support below the usable level.Footnote 21 Retaining only customer service and sales left 17,256 observations, or over 95% of all those hired for whom we had criminal record information. Finally, we eliminated observations which did not have all of our explanatory variables (job stability, school, and the psychometric variables) or outcome variables (related to turnover), leaving a total sample of 10,699 hired workers described in Table 2, panel B.

4 Hiring penalty attached to criminal record and selection bias

We examine the hiring penalty attached to a criminal record and the resulting selection bias in the pool of employees with a criminal record. We restrict attention to the sample of applicants for sales or customer service jobs for whom we have information about criminal record and all controls.

Our model of hiring, whose results are shown in Table 3, is as follows:

$$ {y}_{ijl}={\alpha}_l+\beta \bullet \mathrm{pre}\_{\mathrm{crim}}_i+{\mathbf{X}}_i\bullet \gamma +\delta \bullet j+\zeta \cdotp \left(\mathrm{pre}\_{\mathrm{crim}}_i\ast j\right)+{\varepsilon}_{ijl}, $$

where yijl is a dummy which equals 1 if applicant i was hired in any job (sales or customer service) and in any location; αl is a location fixed effect; pre _ crimi is a dummy that equals 1 if individual i has a criminal record; Χi is a vector of applicant-specific characteristics including education and job history; j is a dummy which equals 1 if the job to which the worker applied was a sales job; and (pre _ crimi ∗ j) is an interaction term. Following a referee’s suggestion, we cluster standard errors at the location level as a proxy for clustering at the firm level. We were not able to reconstruct firm for the applicant sample, but regard location, which we could reconstruct, as a reasonable substitute, since it captures some of the variation associated with firms,Footnote 22 and since the residual might also be correlated with the local labor market conditions. Without any controls, we find that having a criminal record per se does not have a hiring penalty: a criminal history is actually positively correlated with the probability of being hired (Table 3, column 1). This average effect can be gleaned directly from a comparison with pre_crim means in Tables 1 and 2, which shows that there are more individuals with a criminal background among the hired than in the applicant pool. The human capital controls alone do not change the positive effect of the criminal record (Table 3, column 2).

Table 3 Correlates of hiring rates (customer service and sales)

However, after controlling for either location or the position to which the applicant applied, the effect of a criminal record becomes negative (Table 3, columns 3–4), as suggested by previous studies (Pager 2003; Holzer et al. 2004, 2006; Uggen et al. 2014; Agan and Starr 2018; Doleac and Hansen 2016). With full controls other than the position interaction term (column 5), a criminal record lowers the probability of being hired about 4% from the 20% absolute probability of being hired in our sample. Comparable results for the sign and magnitude of the effect of a criminal record are found when a logit model of the same variables is estimated (Table 15 in Appendix 1).

This coefficient reversal from columns 1 and 2 to 3 through 6 suggests that those with a criminal history in our sample are applying to jobs and in locations with better hiring rates than those of the average position. To explore whether these higher hiring rates are for the job generally or represent a greater willingness to hire applicants with criminal records, we add an interaction of sales with pre_crim. The results, which are reported in column 6, suggest possible heterogeneity of the hiring penalty across positions. The baseline customer service positions have a penalty of 2.6% while sales positions have a hiring penalty of an additional 2.8% for a total of 5.4%, although neither coefficient is significant. The logit version does not show a significant difference between the hiring penalty for the two jobs (Table 15 in Appendix 1, column 6).

5 Tenure and separation of employees with a criminal record

Job performance can include many potential measures, which may vary depending on the type of position. Tenure is an important measure of employment outcomes, as finding and training a new worker can be very expensive.Footnote 23 Of course, there are many other potentially important measures of job performance. However, due to data availability, we focus on job tenure, separation, and reason for separation. We restrict analysis to the 10,698 observations for which all of our explanatory variables and outcome measures are available.

To measure whether employees with a criminal record have longer tenure, we construct a variable called length of employment (LOE). This variable measures the number of days that elapse between the hire date and the termination date or the date of final data entry, whichever is smaller. The average length of employment in our sample is 169 days (Table 2, panel B). Note that this variable is subject to right-censoring.

Table 4 takes a first pass at the data by regressing LOE on pre_crim. We run the following model:

$$ {LOE}_{ijfl}={\alpha}_f+{\alpha}_l+\beta \bullet \mathrm{pre}\_{\mathrm{crim}}_i+{\mathbf{X}}_i\bullet \gamma +\delta \bullet j+\zeta \cdotp \left(\mathrm{pre}\_{\mathrm{crim}}_i\ast j\right)+{\varepsilon}_{ijfl}, $$

where LOEijfl is the employment duration of employee i who was hired to perform job j (sales or customer service) by firm f at location l; αf is a fixed effect for firm; αl is a fixed effect for location; pre _ crimi is a dummy that equals 1 if individual i has a criminal record; j is a dummy which equals 1 if the job to which the worker applied was a sales job; and Χi is a vector of employee-specific characteristics which, in addition to education and job history, now includes psychometric measurements. Following a referee’s suggestion we cluster standard errors at the firm level.

Table 4 Correlates of length of employment, OLS

For the sample of both sales and customer service employees and with controls only for location, position, and firm, the estimates indicate that employees with a criminal background stay employed on average 19 days longer than those who do not have a criminal background (Table 4, column 1).

LOE combines the effect of voluntary and involuntary terminations. Involuntary termination is by definition associated with weaker performance, while studies indicate that voluntary termination is most common among highest and lowest performers.Footnote 24 The expected sign of the human capital variables is thus ambiguous, since departing employees include both the best and the worst. Schooling does not predict LOE, but the coefficients on both job history variables are positive and statistically significant in all specifications (Table 4, columns 2–4). Job history and school together reduce the LOE associated with pre_crim by about 2 days (Table 4, columns 1–2). Controlling for proprietary psychometric variables and FFA reduce the additional LOE of employees with criminal records to about 16.5 and 17.7 days, respectively (Table 4, columns 3–4). The effect of the psychometric variables will be considered at more length in the next section.

To quantify the economic significance of longer LOE, we obtained estimates from the consultancy on the average cost of replacing a worker found in our dataset. This figure amounted to $4000 per termination. At average values of other variables, a worker without a criminal record lasts 167 days while a worker with a record lasts 183 days.Footnote 25 Since the average wage for call center employees is about $30,000 per year, this amounts to a savings of about 2.5% of wages per year for these employees.Footnote 26 However, if applicants with records were hired in greater numbers, their quality, and therefore the associated employer savings, might well drop.

The LOE of employees varies strikingly with the job held (Table 4, columns 1–4). Because the effect is so large, we examine the possibility that different models underlie each job (Table 4, columns 5–12). With only location and firm as controls, sales people with criminal records last about 19 days longer than others, while customer service agents with a record last about 21 days longer. The coefficient in the sales estimate is not significant; the p values are between .11 and .14. Schooling has no effect in any estimate. In most specifications, the job stability and school variables together reduce the effect of a criminal record slightly. The psychometric variables will be discussed in the next section.

We next conduct a more refined analysis to account both for different types of separation and for the censoring that results from the unknown ultimate length of employment of those workers who were employed when the data collection ended. About 77% of all workers had separated by the end of data collection, and about 55% of all separations were voluntary. We run the following Cox proportional hazards model to predict employment tenure:

$$ {h}_{ijfl}={h}_0(t)\cdotp \exp \left({\alpha}_f+{\alpha}_l+\beta \cdotp \mathrm{pre}\_{\mathrm{crim}}_i+{\mathbf{X}}_i\cdotp \gamma +\delta \cdotp j\right), $$

where hijfl is the hazard that employee i, who was hired to perform job j (sales or customer service) by firm f at location l, separates t days after being hired; h0(t) is the baseline hazard rate; αf is a fixed effect for firm; αl is a fixed effect for location; pre _ crimi is a dummy that equals 1 if individual i has a criminal record; Χi is a vector of employee-specific characteristics; and j is a dummy which equals 1 if the job to which the worker applied was a sales job.

We begin by applying this model to voluntary separations. Table 5 provides the estimated coefficients of a Cox proportional hazard model in which only voluntary separations are counted as “failures.” For the whole pool, a criminal background has a consistently negative and statistically significant impact on voluntary separations, suggesting that having a criminal background makes an employee less likely to leave voluntarily (Table 5, columns 1–4). This result could be due to several factors. Workers with a criminal record presumably have fewer external labor market opportunities. They may also feel a sense of loyalty or gratitude to an employer who has given them a second chance. Again, the coefficient on job type is highly significant (Table 5, columns 1–4), and we therefore also examine each job separately (Table 5, columns 5–12). A criminal record decreases the voluntary separation rates of customer service employees more than that of sales employees, though both effects are significant and negative (Table 5, columns 5–12). Like LOE, voluntary departures are driven by a mix of high and low performers, and thus, the expected signs on the human capital coefficients are ambiguous. Schooling seldom predicts voluntary separation for either the entire sample or for the sales sample, but for the customer service sample is positive and significant at the 5 or 10% level (Table 5, columns 2–4, 6–8, and 10–12). One of the job history variables, longest_job, is significant and negative for the whole sample and sales, but is insignificant for customer service; the other, fewer_short_jobs is significant (and positive) for customer service only and insignificant otherwise (Table 5, columns 2–4, 6–8, and 10–12). Two psychometric variables are predictive for sales and one is predictive for customer service––we discuss these in more detail later.

Table 5 Correlates of voluntary separation, Cox Proportional Hazard Model

Table 6 provides the estimated coefficients of the Cox proportional hazard model in which only involuntary terminations are counted as failures. Since involuntary termination is associated with lower quality workers, coefficients should now have the sign associated with lower quality. For the whole sample, a criminal background predict involuntary terminations at the 10% significance level in two of four specifications. (Table 6, columns 1–4). Other variables are much more significant. As expected, involuntary terminations are more likely among workers who would be regarded as lower quality by traditional measures: better schooling (Table 6, columns 2–4) and higher job stability (Table 6, columns 2–4) reduce involuntary termination in all four specifications at the 1% significance level. Again, the coefficient on a sales position is highly significant (Table 6, columns 1–4), so we examine the two jobs separately. A criminal record does not predict involuntary termination for customer service jobs (Table 6, columns 9–12) but is highly predictive for sales jobs in all specifications (Table 6, columns 5–8). School and one job stability variable are, as expected, negative and significant in all specifications in both positions. The other job stability variable is negative for all specifications, but significant only for customer service. Again, the psychometric predictors are somewhat different for the two jobs, a finding we discuss further in the next section.

Table 6 Correlates of involuntary separation, Cox Proportional Hazard Model

So far, the evidence on tenure has shown that having a criminal background makes an employee less likely to leave voluntarily and likely to have a longer tenure. Employees with a criminal record are no more likely to be terminated involuntarily in customer service positions, but more likely in a sales position. Since involuntary turnover is associated with weaker performance and turnover is costly, this evidence taken together suggests that customer service employees with a criminal background are, at least at the current margin, a promising pool for employers. Sales employees present a slightly more mixed picture, since they do appear to have higher rates of involuntary terminations.

To further explore the cost of involuntary terminations, we examine a subset of these terminations, those that involve misconduct. Note that the concept of “misconduct,” as used in the human resource setting, corresponds to the definition found in unemployment insurance (UI) law. Employers keep records of misconduct discharges for the purpose of UI reporting: if an employer discharges an employee for misconduct, the employee receives reduced UI or none at all, and thus, the discharge has a lesser effect on the employer’s UI premiums. Although the term “misconduct” seems to imply severe misbehavior, it may also include much lesser failings such as excessive absenteeism or even the use of profanity.Footnote 27

Misconduct discharges are a relatively rare event, occurring in 4.5% of our sample of the hired (Table 2, panel B). Termination for misconduct is generally more common in sales, occurring in 5.9% of sales positions compared with 3.1% of customer service positions. For the sample of both positions, a criminal record is associated with a higher risk of misconduct (Table 7, columns 1–4). Because of the apparently different underlying models, we again examine each job separately. A criminal record is significantly associated with an increased risk of misconduct in sales jobs (Table 7, columns 5–12). Sales workers with a criminal record are about 34% more likely to be terminated for misconduct than those without.Footnote 28 In customer service jobs, the coefficient on a criminal record is about 40% smaller than is sales and borderline significant, suggesting a possible connection but weaker than that in sales.

Table 7 Correlates of misconduct separation, Cox Proportional Hazard Model

Our results suggest that all employees with a criminal record have longer tenure and lower voluntary turnover than other employees. Customer service employees with a criminal record are also not significantly more likely to be terminated involuntarily or for misconduct, though a higher misconduct rate cannot be ruled out. Sales employees with a record display a more complicated pattern. The value of their longer tenure is at least partly offset by their significant and slightly higher rates of involuntary discharge and their significant and clearly higher rates of misconduct discharge. The discrepancy between sales and customer service jobs is particularly striking since the hiring penalty seems to be greater for sales jobs (Table 3, column 6): despite the higher degree of selection, more misconduct is observed. We consider our finding a cautionary tale of the risks of drawing broad conclusions based on one type of position or industry.

The observational nature of our data means that we cannot draw conclusions about the pool of the non-hired, and applicants who were not hired may have been those at greater risk of poor outcomes. Instead, since our estimates describe average values for those hired, they are not conclusive of how marginal applicants with a criminal record might perform. Although we have no direct evidence of which employees were marginal, we do have some indirect evidence. For each firm, we can calculate the percent of the workforce with a criminal record (crimper) and examine the association of this with the likelihood of misconduct by individual workers (Table 8). For the whole sample and the sales and customer service jobs estimated separately, we find that the likelihood of misconduct is actually lower in firms with a larger number of workers with a criminal record. There are at least two possible explanations for this surprising result. First, firms with a comparative advantage in managing misconduct are more likely to hire workers with a higher propensity to misconduct. Second, firms with high levels of misconduct-prone workers may learn to manage misconduct better. The second explanation is a plausible explanation of the present findings because the workplaces in our sample are relatively homogeneous. In either case, these results do not suggest a story in which rising levels of workers with records rapidly drive up the overall misconduct level.

Table 8 Percent of employees with a criminal record as a correlate of misconduct, Cox Proportional Hazard Model

Our findings provide an interesting comparison to those of the only other study of the job performance of employees with criminal records, which was conducted on armed forces data (Lundquist et al. 2018). Military personnel with the most serious criminal records appeared to be no more likely than personnel without a criminal record to leave for performance-related reasons, including misconduct, and to be superior to others on several performance dimensions. Interestingly, military personnel with less serious criminal records performed worse than other personnel across various measures of attrition and promotion, including misconduct. As the authors of that study note, this initially puzzling difference may be explained by military screening procedures, which add extra scrutiny for applicants with criminal records, and more stringent checks for more serious offenses. Perhaps, the civilian firms in our study should consider using different screening methods for customer service and sales jobs.

We next compare the level of misconduct among firms that request criminal records with the level among those that do not. In Table 9, our sample includes both firms that do and do not ask for criminal information, and we include a variable indicating whether the employer had this information at the time of hiring. We find that whether the employer had information about an employee’s criminal record does not predict the likelihood of employee misconduct.

Table 9 Availability of criminal records as a correlate of misconduct separation, Cox Proportional Hazard Model (customer service and sales)

This result may seem surprising, since a criminal record is predictive of misconduct, at least for those entering sales positions. Table 10 examines the hiring of firms who do not use information about criminal records. Columns 1 and 2 examine correlates of a criminal record within the applicant pool. Those with criminal records have more short jobs, though a longer longest job than those without records. They also have less schooling than those without records, although this association is weaker controlling for location. Columns 3 and 4 show the correlates of being hired without criminal record information: these employees tend to have more school, a shorter tenure in their longest job, and fewer short jobs, all of which are associated with not having a criminal record. Although any firms in our sample that declined to use criminal records information probably did so voluntarily, this finding is consistent with the result of studies that find that Ban the Box legislation increases statistical discrimination based on characteristics associated with a criminal record (Agan and Starr 2018; Doleac and Hansen 2016). However, we are not able to examine the effect of non-use of criminal record information on characteristics such as race, since these factors were not provided to us.

Table 10 Correlates of criminal records in applicant pool compared with correlates of missing records in hired pool (customer service and sales)

6 Personality, criminal background, and job performance

The key findings of the previous section are that employees with a criminal record are less likely to quit any job, while exhibiting varying levels of misconduct that are clearly higher in some jobs and possibly not higher in others. These results raise a number of further questions that we now address using the psychometric questions in our data, which are similar to those now commonly used in hiring. The 18 questions in our data are sufficient to be suggestive of how useful psychometric testing could be in this context but are not intended as an exhaustive consideration of the issues.Footnote 29

We first examine the extent to which the two jobs, sales and customer service, may differ in ways other than the outcomes for workers with a criminal record. To do this, we examine whether in each of our jobs the psychometric variables have the same association with our two measures of poor performance, misconduct and involuntary termination.Footnote 30 The psychometric variables perform well in comparison with other explanatory variables,Footnote 31 but show entirely different patterns in each job, summarized in Table 11. Footnote 32These differences raise the possibility that differences in job characteristics rather than differences in the applicant pool might drive the different outcomes observed for employees with a criminal record. Perhaps something in the sales environment affects everyone, but those with criminal records more so.

Table 11 Psychometric predictors in different jobs

We next examine whether employees with criminal records in sales have any specific psychometric characteristics that might explain why their outcomes differ from both sales people of all types and customer service people with criminal records. Regressing each psychometric variable on position, criminal record, and the interaction of the two, we find that sales people and customer service workers clearly have different psychological profiles (Table 16 in Appendix 1).Footnote 33 The different outcomes we observe for sales and customer service representatives with criminal records might therefore be explained by either differences in worker characteristics or differences in job characteristics. However, the observational nature of our data prevents us from drawing any conclusions about causality, since selection both by workers and by firms affects the assignment of workers with various personality traits to each job.

We finally examine whether these psychometric variables can be used to reduce the value of a criminal record in predicting job outcomes and find that they cannot.Footnote 34 Indeed, they do not even have the same association with misconduct and involuntary termination. In the misconduct estimate for sales, the inclusion of the proprietary variables lowers the coefficient on pre_crim slightly while the inclusion of the FFA variables increases the coefficient on pre_crim also very slightly (Table 7, columns 7–8). In the involuntary termination estimate for sales, the reverse is true: the inclusion of the proprietary variables raises the coefficient on pre_crim slightly while the inclusion of the FFA lowers slightly the coefficient on pre_crim (Table 6, columns 7–8). Overall, controlling for the psychometric variables does not sizably affect the relationship between criminal background and productivity. Moreover, although the psychometric questions appear well correlated with criminal background,Footnote 35 the psychometric factors associated with either set of poor job outcomes were not consistently related to those associated with a criminal record.Footnote 36 This was surprising, since prior work often found that traits associated with criminal behavior were those that tended to produce worse work outcomes.Footnote 37 The pessimistic implication of these findings is that psychometrics is not yet useful for eliminating the predictive value of a criminal record. The optimistic implication is that even simple psychometric tests have some predictive value, and the development of better tests may eventually eliminate the incentive for employers to use criminal records in all jobs.

In sum, our analysis of psychometric variables suggests two interesting lines of future inquiry. First, our evidence suggests that either differences in worker characteristics or differences in job characteristics might explain the different misconduct levels observed for employees with criminal records in sales and customer service positions. Second, the psychometric variables have a reasonable degree of association with both a criminal record and job outcomes. However, in our data, they do not reduce the value of a criminal record, and further research is required to determine whether they could do so.

7 Conclusions

Using a unique source of data, we find that employees with a criminal record have a much longer tenure and are less likely to quit their jobs voluntarily than other workers. We further find that in certain jobs, employees with a criminal record are no more likely than those without a record to leave their job involuntarily or for reasons of misconduct. These workers with a criminal background appear to be no worse than, and possibly even better than, workers without such a background. In our data, this low-risk job is customer service. In other jobs, however, employees with a criminal record do appear more likely to leave for reasons of misconduct. In our data, the high-risk job is sales, and we conjecture that whatever factors create the overall high misconduct rate observed in sales jobs may have an even greater effect on employees with criminal records.

The precise cost of this excess risk is highly speculative: the term “misconduct” encompasses behavior ranging from excessive absenteeism to a variety of criminal conduct. Surveys suggest that employers are primarily concerned about large negligent hiring judgments for violent acts (Walker and Miller 2009; Platt 1993), but no systematic evidence supports this concern.Footnote 38 The primary employer losses from misconduct are probably more pedestrian. Only one published study has any bearing on this and, though small and highly specific, suggests that the work-related misconduct of workers with criminal records is on average less serious than that of other workers.Footnote 39 With this in mind, we use dishonesty, a serious but non-catastrophic type of employee misconduct, as a basis for a rough estimate of the cost-benefit calculation facing an employer. The National Retail Federation estimates the loss from each dishonest employee case is about $1546 (National Retail Federation 2015). Only sales workers with a criminal record pose an excess misconduct risk: about 5.9% of sales workers with a criminal record are discharged for misconduct compared with 3.1% of other sales workers, a difference of 2.8%.Footnote 40 An employer who hired a worker with a criminal record rather than a worker without a record increased its expected theft-related costs by about 2.8% of $1546, or $43. The same employer saved about $746 in turnover costs on that worker.

Our results are subject to an important qualification: our estimates were made on employees, in other words, on those applicants who had been filtered through a hiring process that discriminated based on their criminal record. If discrimination against these applicants decreased, either by employer choice or through legislation, the employee group would change, and the new group of employees with records might not exhibit tenures as long as those we observe. All we can say is that, at the current margin, employers may be missing opportunities to hire quality employees by applying a hiring penalty of the current magnitude to a criminal record.

Our findings are not simple, and neither are their policy implications. Finding gainful employment for individuals with criminal records is an important public priority: without such employment, recidivism is almost inevitable, at great cost to both the individual and the community. At the same time, employers are concerned that employing individuals with records may carry risks, and our study does not entirely dispel those fears. On the whole, our results provide support for efforts to expand hiring of applicants with a criminal record at the margin, but one qualification is critical. The instances in which we find better performance of employees with a criminal record cannot be generalized to the entire pool.

Ban the Box laws have come to dominate the policy discussion of how to improve the employment prospects of people with criminal records. These laws apply uniform rules to all employees and employers. Yet a clear takeaway from our study is that not all workforces are the same. Employers should be encouraged to re-examine their assumptions about applicants with criminal records by studying their own workforce. A wide variety of measures could promote this self-examination. For example, the current Work Opportunity Tax Credit is only available to employers who hire ex-felons within 1 year after their conviction or release from prison even though employers appear to discriminate against applicants with criminal records long after release or conviction, and some employers apply a hiring penalty to those with a misdemeanor or even arrest record. Policy to reintegrate individuals with criminal records should consider the variety of job applicants, of jobs, and of employer motivation.