Although unemployment rates are at historical lows, there is still a persistent gap between unemployment rates in black and white population. Some have proposed that part of the gap for men can be explained by the higher rate of criminal records in the black population.
This analysis aims to use negative binomial regressions and the detailed crime data available from the National Longitudinal Survey of Youth 1997 survey to determine if black men with criminal records appear to be the driving force behind the gap.
The author finds that there are significant deviations in labor market outcomes depending on race and ethnicity, even when controlling for a criminal record and premarket skills.
Lowering the disproportionate rate at which black men are incarcerated will not in itself eliminate the unemployment gap between white and black men.
Despite the tightening of the labor market, with the aggregate unemployment rate currently below its natural rate, racial disparities continue to exist. Recent Fed research (Cajner et al., 2017) has shown that observables explain very little of the persistent black–white unemployment and wage gaps. There are economists who believe that some of the gap is caused by the relatively higher proportion of black men who have criminal records (Holzer et al., 2005). This analysis aims to use the detailed crime data available from the National Longitudinal Survey of Youth (NLSY) 1997 (NLSY97) survey to determine if black men with criminal records appear to be one of the driving forces behind the gap.
Ritter and Taylor (2011) attempted to control for many observables, including having a criminal record, and then ran regressions to estimate racial disparities in unemployment. Using the NLSY79, they construct work histories for each individual and then count the number of weeks when the individual was unemployed and the number of weeks when he was not working. They then estimate regressions that use weeks unemployed or weeks not working as the dependent variable. Their independent variables are as follows: black, Hispanic, age at start of work history, age2, Armed Forces Qualifications Test (AFQT), AFQT2, own education, mother’s education, father’s education, self-esteem scale, Rotter scale, suspensions, expulsions, charged with illegal activity, and convicted of illegal activity. They use different specifications of these variables to come to their conclusions. Ritter and Taylor found that the gap cannot be explained. “Black men and women experience substantially higher lifetime unemployment than white individuals with similar levels of premarket skills.” In their sample, black men experience a 176% higher incidence of weeks of unemployment than white respondents.
Ritter and Taylor used the NLSY79 for their analysis, but the NLSY97 has more detailed crime data. In the NLSY79, the only way to identify if a respondent went to prison is if they have ever been interviewed for the survey while housed in a prison or jail. Therefore, someone who served a shorter sentence is unlikely to be captured as a former prisoner. The NLSY97 cohort is a longitudinal project that follows the lives of a sample of American youth born between 1980 and 1984; 8,984 respondents were aged 12–17 when first interviewed in 1997. This ongoing cohort has been surveyed 17 times to date and is now interviewed biennially. The NLSY97 includes detailed crime data, including total cumulative incarcerations and age at first incarceration. Both the NLSY79 and NLSY97 categorize race into three main categories: black, Hispanic, and nonblack/non-Hispanic. The nonblack/non-Hispanic group includes both white and Asian. However, the group is mostly white, and for simplicity, like Ritter and Taylor, this group is referred to as white in the analysis.
In the literature, there are few people who have performed analyses with the NLSY97 data, possibly because the cohort is still relatively young. The latest NLSY97 data is from 2016, when the respondents were 30–36 years. Although the quantity of data is still limited, the NLSY97 respondents are all prime age workers, which is a demographic of interest when analyzing labor market outcomes.
For this analysis, exclusion criteria were not included for those with little work history. Although the respondents have differing amounts of observed work history, in the regressions, the offset log (time in labor force) or log (observed history) has been accounted for this. By specifying this offset in Stata, the regressions have mathematically taken into account that those who were in the labor force longer had more opportunities to become unemployed or nonemployed, so this should not impact the coefficients on the regressors. It is important not to exclude based on time observed because it is possible that those with the worst labor outcomes have lower survey retention rates. By round 17 (most recent survey), the overall retention rate is around 79%. However, negative values, which indicate different skip or no response reasons, have been recoded to be blank so they do not influence summing of weeks or other variables later in the analysis.
With the NLSY97 data, a work year variable was constructed using the school_grad variable. Work year is set to the year after they left or graduated from school; 0 = 2000 and 17 = 2017. Through the work year, those who are only marginally attached to the labor force, before analyzing weeks unemployed or nonemployed, are weeded out. Out of 8,984 observations, 203 had no information recorded for when they left school and 921 left school before 2000, which is the year chosen to start the analysis. For these people, their work year is set to when they would have been at least 19 years old (Figure 1).
Other important variables used later in the analysis include: Armed Services Vocational Aptitude Test (ASVAB), own education, number of suspensions, number of charges, total employment, total unemployment, and total nonemployment (Table 1). Details about these variables and how they were constructed are available in Appendix.
Sample summary statistics
|Work year||8,984||3.89||4.00||0 (2000)||17 (2017)|
|Total number of suspensions||8,984||6.12||20.65||0||600|
|Total number of charges||8,984||0.67||1.57||0||17|
|Total weeks of unemployment||6,240||49.87||69.65||0||651|
|Total weeks of nonemployment||6,240||129.38||168.55||0||828|
|Total weeks of employment||6,240||441.71||243.83||0||835|
|Total weeks in labor force||6,240||491.58||242.44||0||835|
|Total work history weeks||5,316||665.63||169.12||3||987|
With the work years, the weekly employment status data was taken and Ritter and Taylor’s summary statistics for unemployment and nonemployment table was constructed (Table 2) using the NLSY97 instead of the NLSY79. Table 2 shows those who have ever been incarcerated and those who have never been incarcerated.
Summary statistics for unemployment and nonemployment
|Fraction of men with > 0 weeks||0.718||0.808||0.596|
|Mean fraction of weeks, given > 0||0.159||0.267||0.188|
|Median fraction of weeks, given > 0||0.081||0.193||0.095|
|Fraction of men with > 0 weeks||0.775||0.836||0.685|
|Mean fraction of weeks, given > 0||0.158||0.278||0.150|
|Median fraction of weeks, given > 0||0.078||0.173||0.065|
|2000–2016, Ever incarcerated||Hispanic||Black||Nonblack/Non-Hispanic|
|Fraction of men with > 0 weeks||0.897||0.945||0.886|
|Mean fraction of weeks, given > 0||0.167||0.334||0.189|
|Median fraction of weeks, given > 0||0.117||0.266||0.124|
|Fraction of men with > 0 weeks||0.959||0.988||0.936|
|Mean fraction of weeks, given > 0||0.275||0.465||0.233|
|Median fraction of weeks, given > 0||0.208||0.477||0.166|
|2000–2016, Never incarcerated||Hispanic||Black||Nonblack/Non-Hispanic|
|Fraction of men with > 0 weeks||0.686||0.770||0.561|
|Mean fraction of weeks, given > 0||0.158||0.245||0.188|
|Median fraction of weeks, given > 0||0.072||0.172||0.088|
|Fraction of men with > 0 weeks||0.743||0.793||0.654|
|Mean fraction of weeks, given > 0||0.129||0.206||0.132|
|Median fraction of weeks, given > 0||0.064||0.116||0.050|
4 Ritter–Taylor summary statistics
The summary statistics using the sample show lower proportions than Ritter and Taylor found, possibly because of the shorter time sample (Figure 2). Ritter and Taylor (2011) found 82% of Hispanic men, 88% of black men, and 71% of white men reported at least one week of unemployment. For men’s nonemployment, they found 93% of Hispanics, 94% of blacks, and 86% of whites report at least one week. Using the NLSY97, the respective numbers are 72% of Hispanic men, 81% of black men, and 60% of white men reporting at least a week of unemployment. For men’s nonemployment, 78% of Hispanics, 84% of blacks, and 69% of whites report at least one week. Although the numbers are different, the orders are the same. Blacks report the highest unemployment/nonemployment, followed by Hispanics, and whites fare the best.
Of the men surveyed for the NLSY97, ~15% have been incarcerated at some point. The average age of first incarceration is 22.29 years old, with a little deviation by race. Black men are first incarcerated at a mean age of 21.96, Hispanic men at 22.2, and white men at 22.56. The age of first incarceration being so young indicates that many of the men in the sample have a criminal record very early in their working history. The average length of first incarceration is 14.4 months, with high deviation by race. Black men mean first incarceration is 22.7 months, Hispanic men is 11.5 months, and white men is 8.1 months. For longest incarceration, mean duration is 20.4 months for all men, 31 months for black men, 16.5 months for Hispanic men, and 12.6 months for whites. Cumulatively, for men who have ever been incarcerated, mean total months incarcerated is 28.1 months, for black men it is 41.4 months, Hispanic men 22.2 months, and for white men 18.9 months. Could this high deviation in incarceration lengths by race be worsening labor market outcomes for black men because of skill atrophy while incarcerated? Regressing total unemployment and total nonemployment by total months incarcerated shows that the relationship between incarceration length and labor market outcomes is not significant, so there is more at work causing the gap.
Table 2 indicates that even among like groups, racial disparities exist. Among all the men in the sample, 81% of the black men report at least one week of unemployment, compared to 72% of Hispanic men and 60% of white men. If they have ever been incarcerated, the racial disparities are smaller but still exist. For black men who have ever been incarcerated, 95% report at least one week of unemployment, and 99% report at least one week of nonemployment. Additionally, this 99% of black former prisoners report that a mean 47% of their work history has been spent in nonemployment. This is significantly higher than the 28% and 23% reported by Hispanic and white men. If medians are compared, black men who have been incarcerated are nonemployed, almost three times longer than white men who have been incarcerated. The never incarcerated table indicates that it is not necessarily the higher proportion of black men with criminal records pushing up the black unemployment rate. About 77% of black men who have never been incarcerated report at least one week of unemployment, compared to 69% of Hispanic men and 56% of white men. In all the groupings, the nonblack/non-Hispanic group always has the smallest proportion having ever experienced unemployment or nonemployment.
In addition to the summary tables, the Ritter–Taylor analysis includes negative binomial regressions. The negative binomial regressions are offset by log (weeks in labor force) for unemployment or log (recorded work history weeks) for nonemployment. The offset function simply controls for the varying amounts of work history by mathematically specifying that log (work history weeks or weeks in labor force) equals 1. The negative binomial regression models count data with variability different from its mean. For each one-unit change in the predictor variable, the difference in the logs of expected counts of the response variable is expected to be the respective regression coefficient, holding everything else constant.
Table 3 reports results from negative binomial regressions for six different specifications for men’s weeks unemployed. The large and always significant coefficient of the dummy variable for black shows that black men spend more weeks unemployed, regardless of what controls are included in the columns to account for “human capital.” The Ritter–Taylor analysis also found large and always significant effects for the blacks. The sixth column has the strongest pseudo R2, and according to this model being black increases weeks unemployed by 2.2 weeks. The most significant coefficient for Hispanic in column 1 also shows an increase of 1.3 weeks unemployed relative to white men. Including variables controlling for criminal history increases the weeks they spend unemployed but has less of an impact than being black. The fourth specification shows one more week of unemployment for those who have ever been incarcerated. Including more fine-tuned crime measures in columns 5 and 6 in place of the incarceration dummy variable show effects in the direction that the author would expect, but only the coefficients for number of suspensions and number of charges are significant. Each suspension increases weeks of unemployment by one week. It is possible that number of suspensions is significant and not number of arrests or incarcerations because the NLSY97 sample is so young. Complete information on their primary schooling have been obtained, so the suspensions variable might be capturing other effects from those who are more likely to commit crimes.
Negative binomial regression, weeks unemployed, men
|Coefficient (P value)|
Table 4 reports results from negative binomial regressions for six different specifications for men’s weeks not working. Like the unemployment results for men, being black sometimes has a significant impact on time spent nonemployed. The sixth specification is the strongest, and it predicts being black increases weeks nonemployed by 1.1 weeks. The incarceration dummy is never significant, but interestingly it shows a negative impact on weeks of nonemployment. This could be because those who have been incarcerated upon their release are less likely to engage in black market economic activity that could be classified as nonemployment out of fear of returning to prison. We see a similar negative pattern with arrests in men’s nonemployment. For Hispanic men, the coefficients are almost never significant. This differs from Ritter and Taylor’s regressions using the NLSY79, which had significant coefficients for Hispanics. In columns 5 and 6, most of the fine-tuned crime data does not have a significant impact on weeks nonemployed for men. Number of charges is the only strongly significant crime measure, predicting an increase of about 1.1 weeks of nonemployment with each charge.
Negative binomial regression, weeks not working, men
|Coefficient (P value)|
There is a large literature on ex-offenders and the labor market. Hjalmarsson (2008) used the NLSY97 data to determine if juvenile justice system interactions affect high school graduation rates and found arrested and incarcerated individuals are about 11 and 26 percentage points less likely to graduate high school than nonarrested individuals. Apel and Sweeten (2010) used the NLSY97 to discover the impact of incarceration during late adolescence and early adulthood on short- and long-term employment outcomes. Their results suggest that incarceration decreases the probability of employment and increases the duration of labor force nonparticipation. Other studies do not find that incarceration diminishes employment prospects. Western et al.’s (2001)review of the literature concluded that while serving time in prison can diminish an individual’s earnings, administrative data tends to show that employment prospects do not necessarily decrease. However, many recent papers claim employment prospects are diminished, and it could be having an aggregate effect on the economy. Schmitt and Warner (2011) estimate that the ex-offender population lowered the male employment rate in 2008 by 1.5 to 1.7 percentage points, costing the United States about $57–$65 billion. A more recent paper by Abraham and Kearney (2018) used 2014 NLSY97 data as a proxy to estimate that there are 5.13 million working age individuals with a prior prison term of more than 2 years and that the 2016 employment-to-population-ratio was 0.13 percentage points lower because of this.
The consensus from the literature seems to support that a criminal record negatively affects labor market outcomes, which is consistent with the results from this analysis. However, one should take into account that there are some limitations to the analysis. One weakness is the inability to separate white and others, regardless of the other group being small. Additionally, it is extremely difficult to prove causality with crime. It is possible that those who commit crimes were already more likely to have poor labor market outcomes, so they are not truly capturing an incarceration effect. Although the incarceration effect is unclear, racial disparities between black and white are apparent. About 22% of the black men in the NLSY97 have been incarcerated at some point, 35% have at least one conviction, and 57% have been arrested at least once. The regressions indicate that reducing the disproportionate rate of black male incarceration might narrow, but will not completely close, the gap between white and black employment outcomes. There are significant deviations in labor market outcomes depending on race and ethnicity, even when controlling for a criminal record and premarket skills. Therefore, the high proportion of black men with a criminal record does not seem to be the driving force responsible for the black–white employment gap.
Availability of data and material
The analysis in this paper was conducted using the publicly available NLSY97 dataset:
The author declares that she has no competing interests.
Armed Forces Qualifications Test
Armed Services Vocational Aptitude Test
National Longitudinal Survey of Youth.
This paper has benefitted from the input of Dr. Ivan Vidangos. Any remaining errors are my own.
Cajner, Tomaz, Tyler Radler, David Ratner, and Ivan Vidangos (2017). “Racial Gaps in Labor Market Outcomes in the Last Four Decades and over the Business Cycle,” Finance and Economics Discussion Series 2017-071. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2017.071
Holzer, Harry; Paul Offner; Elaine Sorensen (2005): Declining Employment Among Young Black Less Educated Men: The Role of Incarceration and Child Support. Journal of Policy Analysis and Management 24, 330-333.
Ritter, Joseph; Lowell Taylor (2011): Racial Disparity in Unemployment. Review of Economics and Statistics 93(1), 30-34.
Hjalmarsson, Randi (2008): Criminal Justice Involvement and High School Completion. Journal of Urban Economics 63, 613-630.
Apel, Robert; Gary Sweeten (2010): The Impact of Incarceration on Employment During the Transition to Adulthood. Social Problems 57(3), 448-479.
Western, Bruce; Jeffrey Kling; David Weiman (2001): The Labor Market Consequences of Incarceration: Crime and Delinquency 47, 410-427.
Schmitt, John; Kris Warner (2011): Ex-Offenders and the Labor Market. WorkingUSA: The Journal of Labor and Society 14, 87-109.
Abraham, Katharine; Melissa Kearney (2018): Explaining the Decline in the U.S. Employment-to-Population Ratio: A Review of the Evidence. Working Paper No. 24333. Cambridge, Mass.: National Bureau of Economic Research.
The ASVAB variable used in the regression is scaled in Stata using the below code:
Own education is CVC_HGC_EVER_XRND in the NLSY97 dataset, the highest grade ever completed. The NLSY97 does not have a cumulative suspensions or charges variable, but they can be created by taking the row total of SCH_SUSPENSIONS each year and YSAQ_455 each year. Similarly, total unemployment and total nonemployment are generated through row totals of employment statuses each week. After running a loop to replace the employment status indicating they were unemployed that week (4) with a 1 for 1 week of unemployment, the Stata code generates a total_unemp = -1 variable for all observations. Then, a total_unemp_1 variable was generated that sums up all the yearly weeks unemployed variables (UNEMP_ EMP_STATUS_YEAR_WK). Total_unemp is replaced with total_unemp_1 as long as it is not blank. There is a line of this code for each year, dropping the earliest year sequentially each time to account for the differing years the respondents formally enter the labor force.
Total employment is a row total of the yearly cumulative CVC_WKSWK_YR_ALL variables, using the same methodology used for unemployment above. Total unemployment and total employment are summed to make a time in labor force variable. The log of weeks in labor force is used as the offset in the weeks of unemployment negative binomial regressions to control for differing work histories. Specifying this offset in Stata theoretically adjusts for the opportunity of unemployment occurring by accounting for how many weeks they were in the labor force. Time in labor force and total nonemployment (calculated using the same methodology as unemployment with NONEMP_EMP_STATUS_YEAR_WK) are summed to make work history weeks for those who reported nonemployment at any point. Work history weeks is just valid weeks of work history observation. The log of work history weeks is used as the offset in the weeks of nonemployment regressions.
In addition to running negative binomial regressions, OLS regressions are ran to see what impact the variables have in a linear specification. Log of Time in Labor Force and Log of Weeks of Observed Worked History are used as a p weight for unemployed weeks (Table A1) and nonemployed weeks (Table A2), respectively. The results from the OLS regressions are consistent with those from the negative binomial regressions and show that being black still significantly negatively affects labor market outcomes even when controlling for a criminal record.
Ordinary least squares regression, weeks unemployed, men
Ordinary least squares regression, weeks not working, men
|Coefficients (P value)|