Peer Effects and Recidivism: The Role of Race and Age

Abstract

1. Introduction

The rate of incarceration and recidivism in the United States is high and costly. Keeping a prisoner incarcerated costs the state prison system approximately $26,000 a year, more than a full time minimum wage earner in many states (Schmitt and Warner 2010). Importantly, many members of the prison population are repeat offenders, cycling in and out of the prison system. This paper investigates a potentially key determinant of recidivism, namely peer effects in prisons. If, in fact, elements of the prison experience such as peer exposure results in higher rates of recidivism, this would create a vicious circle trapping offenders into the prison system. This not only could lead to increased prison costs, it also raises crime-related social costs.

Our study contributes to the literature by uncovering novel features of prison peer effects for a sample of adult prisoners in the United States. Studies suggest that social interactions are particularly important in the criminal sector. As a response to the lack of formal institutions, informal networks facilitate the transfer of knowledge and development of criminal skills.¹ While previous peer effects literature establishes the existence of peer effects among fellow inmates in prison(s) and juvenile facilities, more research is needed to formulate policies of prison assignments to reduce recidivism (Doleac 2019). With this motivation, our study adds to the literature by providing new evidence that peers in prison can impact recidivism of adults. In particular, we must identify the key characteristics that drive these peer effects. We show that both race and age are important factors. Further, we find that the network effects are present conditional on defining peers as fellow inmates of the same race. This suggests that individuals sort into networks that exhibit racial homophily and that same-race and same-age peers are more influential.

The composition of prisoners in a correctional facility is an important prison characteristic that is part of the role that prisons play in determining recidivism, and there is a small but growing literature examining these peer effects. A few papers study peer effects among juveniles (Bayer et al. 2009; Patacchini and Zenou 2009; Stevenson 2017) and young adults (Damm and Gorinas 2013) in juvenile centers in the United States and in prisons in Denmark, respectively. Stevenson (2017) argues that social contagion is the most likely channel through which such peer effects operate, but does not explore whether race plays a role. Damm and Gorinas (2013) show that peers similar in age have a larger effect in the context of Denmark’s young adults prisoner population. However, their sample is more homogenous in terms of race which may explain a lack of precision on their race-specific peer effects estimates. More broadly, this paper contributes to the literature that links peer influences within schools and neighborhoods with criminal behavior (Case and Katz 1991; Ludwig et al. 2001; Kling et al. 2005; Ludwig and Kling 2007; Deming 2011; Billings et al. 2013; Kirk 2015; Corno 2016; Kim and Fletcher 2018; Steinberg et al. 2019).

This paper builds on the existing literature by investigating peer effects in a more general context. We examine adult prisoners who make up the majority of the prison population. While Bayer et al. (2009) study juveniles in the United States and Damm and Gorinas (2013) study young adults in Denmark, we study a rich dataset from the state of Georgia containing the universe of prisoners in the state prison system from 1995 to 2005. We focus on male prisoners and estimate peer effects within prison on recidivism for violent, drug-related, property, and sex crimes. A complementary recent paper by Billings and Schnepel (2017) also studies peer effects and recidivism among adult former inmates in North Carolina. However, the peer group analyzed is that of the residential neighborhood rather than fellow prison inmates. This is a considerably broader definition of a peer group and perhaps less easily manipulated by policy makers. Nevertheless, they estimate the effect of neighborhood criminal peers on recidivism. They find that the more residential peers are incarcerated at the time of release from prison, the less likely it is for the released prisoner to reoffend, with the effect being larger for more similar peers in terms of demographics.

Our data therefore allows us the distinct opportunity to study adult prisoners in the United States within the prison system setting, facing a well-defined set of potential peers (their fellow inmates). However, there remains the challenge of endogenous sorting in the assignment of prisoners to facilities, likely generating a positive correlation between the types of fellow inmates a prisoner faces and the probability of recidivating. Our key identification assumption to address this challenge is that the changes in prison composition over time for a given prison are random, so that at any particular month, one’s prison mates are randomly determined relative to other months in the same facility. Conditional on observable individual characteristics and criminal history, two inmates at the same prison may be exposed to considerably different peers depending on the timing of their stay. Similar to Bayer et al. (2009), we argue that this variation is quasi-random after dealing with non-random prison assignment rules by including facility-by-prior-offense fixed effects.²

A prison, however, can be a large place. Unfortunately, we do not have direct data on the peer networks within prisons. In this paper, peers are therefore defined as all other inmates in the same prison at the same time. One way to improve resolution on peers is to rely on the principle of homophily. The networks literature has long noted a propensity for people with similar demographics to form ties more frequently (see McPherson et al. [2001] for an overview). We thus expect prisoners to have more social interactions with inmates from the same race, and we expect the share of inmates of the same race to have a larger effect on the individual’s probability of recidivism than the share of inmates in general. If that is the case, a potentially fruitful policy would avoid segregation within prisons by race or crime. Alternatively, policies that assign prisoners by both their demographics and crime type may have the unintended effect of increasing exposure to peers who can facilitate a criminal career. We therefore construct peer measures irrespective of demographics, as well as taking into account both race and age. This approach is in a similar spirit to a number of papers in the literature suggesting that social interactions are strong conditional on socio-economic demographics (Bayer et al. 2008; Cohen-Cole and Fletcher 2008; Billings et al. 2019).

We find that exposure between peers and prisoners with histories of violent crime has a deterrent effect on future violent crime. Conversely, we see a link to higher future violent crimes when peers with and prisoners without histories of such offenses interact. This effect is persistent both for race-specific peer effects as well as general peer effects. On the other hand, we find strong reinforcing effects for criminals convicted with property, drug sale, or drug possession crimes only when peer groups are defined by prisoners of the same age-and-race and convicted of the same crime. We do not find these effects if we look at general peer exposure. A standard deviation increase of 0.12 (0.10) [0.11] in the number of prison peers with a property (drug sale) [drug possession] history is associated with an increase in the probability of recidivism with a property (drug sale) [drug possession] offense from 11.5% (2.8%) [7.5%] to 12.6% (4.1%) [8.8%], for individuals with a criminal background in these crimes. Finally, we also document important heterogeneity in the peer effects and find that younger prisoners are particularly susceptible to peer influence for recidivism in drug-related crimes.

We note that our main results should be interpreted broadly as “peer effects” and our identification strategy is unable to decompose these effects into more specific mechanisms such as direct social interactions or social learning. That said, the difference in the estimated peer effects when peer measures are defined by demographic characteristics is strongly suggestive that the internal dynamics of social interactions in the prison system are key to understanding how the prison experience can impact recidivism. We interpret our findings of stronger reinforcing results among peers who have similar demographics to indicate that self-segregation plays a key role in determining exposure rates to other inmates within a prison. This, in turn, opens up interesting possibilities for policy makers to influence recidivism through the assignment of prisoners to prisons, as well as enforced segregation (or de-segregation) within prisons.

2. Data

2.1 Sources and Descriptive Statistics

We take advantage of an administrative dataset provided by the Georgia Department of Corrections (GDC) to examine the effect of prison characteristics and composition on recidivism. We observe the universe of people released from the Georgia prison system from 1995 to 2008. These data provide rich socio-demographic, criminal history, facility assignment, prisoner’s diagnostic evaluation, and recidivism information for each inmate. We focus on Georgia for two main reasons. First, the richness of the data allows us to pin down all inmates housed in each prison facility over time, which is crucial in identifying any peer effects that might occur within an institution. Second, at least on key observable characteristics, the prison population in Georgia seems to be representative of that nationwide.³

We restrict our analysis to male inmates who are not housed in bootcamps, local jails, out-of-state or unknown facilities.⁴ We further drop individuals who are missing security assignment or any other demographic or criminal characteristics that are crucial for our analysis. We present summary statistics of the individuals in our sample in Table  1.

Our data provide information about the last institution in which the inmate was housed as well as his most recent three previous institutions. In addition to that, we have information on how many days he spent in each institution and the reason for transfer (if any). For this analysis, we use only one of the four institutions we observe. We define this institution to be the one in which the inmate spends at least 50% of the total time he serves in the Georgia state prison system.⁵ In the data, we observe that 86% of all inmates in Georgia spent at least 50% of the total time they spent in the prison system in a single institution.⁶Online Appendix Figure A1 lists all the institutions used in the analysis along with the number of observations per institution. On average, we observe around 1000 inmates released from each one of the 67 institutions in our analysis. Furthermore, in Table  2, we observe that the average inmate spends 394 days in this institution.

The GDC has two main ways to measure recidivism—as the rate of return to prison (RTP) rates or as the rate of felony reconviction (FR). The RTP classifies an inmate as a recidivist if he returns to a Georgia prison for any reason, including technical violations of parole or probation. On the other hand, FR classifies an inmate as recidivist if he is subsequently convicted in Georgia of a new felony offense for which he receives either probation, prison, or a split sentence. We use the FR rate, measured within 3 years of release, as a proxy for recidivism because of two main reasons. First, we want to examine whether one’s peers have any effect on one’s decision to recidivate with a crime that he has already committed or with a crime that he has not committed but his peers have.⁷ Second, parole (probation) are revoked based on the judgment of the parole (probation) officer, and therefore people who RTP because of revocation might be different from those returning because of a new conviction. Since we observe all releases in Georgia through 2008 and we want each released prisoner to have a well-defined three-year window period to potentially recidivate, we restrict our estimation sample to prisoners who were released before 2005. Table  2 presents summary statistics for our recidivism measure. Within three years of release, 30% of the sample is reconvicted of a new crime. Property crimes are the most frequent types of recidivating crimes accounting for 11.5%, followed by drug possession with 7.5%. We focus our main analysis on five of eight crime categories listed in Table  2; violent personal, property, drug sale, drug possession and sex crime. We exclude non-violent personal, habitual driving under the influence of alcohol, and alcohol crime for two reasons. First, the recidivism rates for these crimes are not high enough for a precise estimation. For instance, recidivism rates for these crimes are less than 1%. Second, due to the low severity of these crimes, there are very few people incarcerated for them. Specifically, only 12.8% and 2.6% of the inmates we observe have a history of either DUI or alcohol crime and non-violent crime, respectively.⁸

Finally, to motivate our detailed analysis of peer composition in this paper, we report the race and age composition of prisoners in our sample (see Table  3). We see that the prisoners are fairly heterogeneous in age, and that both blacks and whites are well represented across the age distribution. There are slightly more black prisoners in the younger age category. This heterogeneity supports our assertion that peer measures should account for age and race.

2.2 Facility Assignment

Upon prison admission in Georgia, offenders are put through a series of evaluations, including medical and mental health screenings at one of the GDC evaluation prisons. In Georgia, there are four such prisons for male convicts (Georgia Diagnostic and Classification Center, Arrendale, Scott, and Bostick) and one for female convicts (Metro Women’s State Prison).

When assigning individuals to prisons, the GDC develops an individual comprehensive profile that includes the offender’s crime, social background, education, job skills and work history, health, addiction problems, and criminal record. Primarily based on the current and past criminal history, the offenders are also assigned to the most appropriate security level classification—trusty, minimum, medium, close, and maximum. The classification levels are in ascending order of perceived public safety risks of the inmate. Our estimation sample consists of predominantly people housed in minimum or medium security institutions. In particular, 59% of observations in Online Appendix Figure A1 are in minimum security prisons, while 29% are in medium security ones. Since we observe most of the elements from the evaluation process, we can account for the non-random prison assignment to some extent by controlling for observable individual characteristics that are likely to influence prison assignment.

From conversations with the Director of Offender Administration at the GDC, security classification and program needs of the inmate are the two main factors the GDC takes into account in determining prison assignment. We control for all individual characteristics which are essential in determining security classification such as offender’s sentence, nature of the crime, criminal history, and history of violence. We also include indicators for drug or alcohol problems, education attainment, and the Intelligence Quotient score of the inmate in order to account for selection based on inmate’s prison program needs.

2.3 Peer Measures

Online Appendix Table A1 presents a summary description of the demographic and criminal peer measures. We note that 51% of inmates peers have had a property crime conviction. The other most common crime categories for one’s peers are violent crimes (47%), drug possession (32%), and drug sale (19%). Most inmates have higher peer exposure to property and drug offenders as the spread of the distribution is much bigger, while less exposure to sex and violent criminals. This is not surprising given that violent (and to some extent, sex) offenders are usually assigned to close or maximum security prisons.

Online Appendix Tables A2–A4 present summary descriptions of the demographic and criminal peer measures by race, age, and race-and-age.

3. Peer Effects

Peer effects in prison can operate through multiple, not necessarily mutually exclusive, channels. Prisons may facilitate the transmission of information and skills and make individuals better criminals. They can likewise decrease the frictions of meeting potential co-offenders in order to build informal social networks (e.g. prison gangs). Prisons may also serve as catalysts for increasing or even adopting antisocial behavior. This paper, however, will not speak to the exact mechanism by which peers affect individual reoffending decisions. Rather, we are more interested in establishing the existence (or lack thereof) of peer effects in the US adult prison population.

3.1 Empirical Methodology

Here, $Yijtc$ is an indicator equal to 1 if inmate i, who is released in period t from prison facility j, is reconvicted of crime c within three years of release. As describe in Section 2.3, $PeerOffijtc$ is the weighted average of exposure to peers who have a history of committing a crime type c. The indicator $Offic$ is equal to 1 if inmate i has previously committed crime type c and 0 otherwise. Similarly, the indicator $NoOffic$ is equal to 1 if inmate i has not previously committed crime type c and 0 otherwise. X_i is a vector of individual criminal and personal characteristics, while Peer_jt are the counterpart vectors containing the same characteristics but for inmate i’s peers. ν_j and $μj*Offic$ are facility and facility-by-crime history fixed effects, respectively, and τ_t is a vector of indicators with quarter of release.

The interactions terms, $PeerOffijtc*NoOffic$ and $PeerOffijtc*Offic$ capture the differential impact peers can have on individuals who do and do not have any previous history of crime type c. On the one hand, the influence of peers could affect individuals who already have some experience in a particular crime category, this is captured by β₂. We refer to this specific peer effect as “reinforcing.” On the other hand, we call the peer effect “converting” if peers influence individuals who have no history in a particular crime category. The converting peer effect is captured by the coefficient β₁. If the estimates are negative, we refer to the peer effect as “deterring.”

Table  4 shows evidence that the distinction between converting and deterring peer effects is important to our analysis. In this table, we regress the individual’s probability to recidivate with one of the eight crime categories on having a history of committing this particular crime or any of the other seven crimes. The first row reports the coefficient estimate on the indicator for whether an individual has previously been convicted of the crime for which he is recidivating with. For brevity, the second row reports the average of the off-diagonal coefficients for the crimes with which the individual is not recidivating with. The results show a clear pattern: having a history of a particular crime is statistically significantly associated with the odds of reoffending with the same crime.¹⁰ The effect is the largest for property criminals—having had a property offense on one’s criminal history record is positively linked to one’s chance to commit a property offense after being released from prison by almost 14 percentage points.

Note that the average of the off-diagonal coefficients in all crime categories is more than 10 times smaller than that on the diagonal coefficients. This implies that the correlation between recidivating with a particular crime when one has a history in the same crime is much higher than of having a history in any other crime type. These results present suggestive evidence that it is important to separate the effect of peers by individuals’ criminal history.

3.2 Identification

The main identifying assumption of both the “converting” and “reinforcing” peer effect is that exposure to peers within a given prison is as good as random. To guarantee that this assumption holds, we need the timing of the assignment of individuals to facilities with respect to the already existing composition of the prison to be as good as random. This argument relies on the intuition that while the stock composition of each prison is clearly not random with respect to the type of offense committed by prisoners, the flow is random. We check that the identifying assumption holds in two ways.

First, we include quarter of release indicators to account for any criminal trends. Further, we test the existence of time trends by regressing the peer crime indicators on quarter of release dummies. We present our results in Online Appendix Figure A2. We find some weak evidence that peer crime characteristics (with exception of property crimes) increase over time, providing some evidence of an upward trend in criminality. However, almost all of the coefficients on the quarter dummies are not statistically significant suggesting that any trend in criminality is very weak. If crime trends were significantly changing over our study period, the prison compositions would be less comparable over time, and threaten the assumption that variation in composition over time is random. Additionally, we perform a robustness check by exploiting only within-facility variation over a three-year span (instead of the entire 10-year period) by interacting the facility fixed effects with three-year dummies. We find that the trends are small and do not affect the estimated coefficients significantly (Online Appendix Table A6).

Second, if the variation in peer composition within a facility is strongly correlated with individual characteristics, our identification strategy for peer effects would be invalid. To test if such a correlation exists in the data, we employ the two-step test discussed by Bayer et al. (2009). Overall, we find no evidence that this is the case.

In the first step, we run an OLS regression of the individual’s probability to recidivate with one of the five crimes (violent, property, drug sale, drug possession, and sex) on the individual’s demographic characteristics, drug and alcohol addiction, and any crime history characteristics listed in Table  1.¹¹ Then, we estimate the predicted probability of recidivism based on these observable characteristics. The estimated predicted recidivism thus encapsulates the role of observable individual characteristics in predicting recidivism.

In the second stage, we test whether the “converting” (⁠$PeerOffijtc*NoOffic$⁠) and “reinforcing” (⁠$PeerOffijtc*Offic$⁠) peer effects have any predictive power to explain the predicted recidivism probability from the first stage. In particular, we regress the predicted recidivism on the two interaction terms via a seemingly unrelated regressions (SURs) model and present our results in Table  5, Panel A. We observe that the coefficients on the two interaction terms are large and statistically significant for all five crime categories. These results suggest that there exists a strong relationship between peer exposure and observable individual characteristics that are very likely to determine both facility assignment and an individual’s propensity to recidivate.

In Table  5, Panel B, we exploit only the within prison variation in peer exposure by adding facility-by-prior-offense fixed effects. We find almost no evidence of the strong relationship between peer exposure and observable individual characteristics that we show in Panel A. Most coefficients are more than 10 times smaller than those in Panel A and are not statistically significant.¹² From the results presented in Table  5, we can draw two main conclusions. First, if we estimate the peer effects across prisons, our estimates are likely to be biased (as evident from Panel A). Second, Panel B provides little evidence that the variation in peer composition within an institution is correlated with observable personal and criminal characteristics that are very likely to influence recidivism rates, and thus lends support to the conclusion that our data are consistent with our identifying assumption. When we run the same specification test for race-specific and age-specific peer effects in Online Appendix Table A5, we draw qualitatively similar conclusions—namely, we find almost no evidence of the strong relationship between peer exposure and observable individual characteristics when we include facility-by-offense fixed effects.

3.3 Main Results

Table  6 presents the estimation of Equation (3) as a SUR. For brevity, we report only the coefficients of interest, β₁ and β₂, but we also include all individual demographic and criminal characteristics from Table  1, all peer characteristics from Online Appendix Table A1, quarter of release dummies, and facility fixed effects. Note that we are exploring only the within facility variation in peers as we are also including facility-by-offense fixed effects in all specifications.

The estimated reinforcing peer effect (β₁) is statistically significant for individuals who have already committed and re-offend with a violent crime, while we find no evidence of this relationship in any of the other types of crime. Greater exposure to peers with violent criminal experience decreases the probability that an individual with a violent criminal record commits another violent crime within three years after prison release. This result can be rationalized if the prison experience of interacting with other violent crime offenders acts as a deterrent, or causes the individual to become better at committing such crimes, or better at avoiding being caught with another violent crime.¹³ The opposite is true if the individual has never committed a violent crime—namely, greater peer exposure to violent crime offenders results in a higher probability that the individual recidivates with a violent crime.

However, once we examine race-specific, age-specific, and race-and-age-specific peer effects, the results are quite different. In addition to the patterns seen for violent crime, we find that there are significant reinforcing effects for property and drug-related crimes. Accounting for race similarities gives rise to “reinforcing” effects for property crime and drug possession, accounting for age similarities gives rise to effects for property, and accounting for the combination gives rise to property crime, drug possession, and drug sale effects.

Both Bayer et al. (2009) and Damm and Gorinas (2013) find reinforcing peer effects for drug offenders. Our results are consistent with their findings, although in our data the peer effects only appear when race (drug possession) and race-and-age (drug sale) are accounted for. The need to account for homophily in our data is similar to Billings et al. (2019), where the authors only find peer effects on youth crime outcomes only when race, gender, and school grade are factored into the analysis. We find that prisoners with or without prior conviction of drug possession, when exposed to peers who have already committed a drug possession crime, will be more likely to recidivate with that crime.

The estimate on the reinforcing effect is 0.13 in Panel B of Table  6, Column (4). Thus, a one standard deviation increase in the exposure to peers with a history of a drug possession crime (0.10 as seen from Online Appendix Table A2) increases the individual’s probability to recidivate with a drug possession crime at the mean from 7.5% to 8.8%. To put this into context, recall from Table  4 that inmates with prior drug possession history have a 8.2% higher chance to recidivate with another drug possession crime. Having drug-possession offenders as peers can therefore increase the probability of recidivating with a drug possession crime significantly.

Similarly, from Panel D, we see that the estimated reinforcing effect for drug sales is 0.13. Given a standard deviation increase in exposure to peers with a drug sale history (0.10, see Online Appendix Table A4), the mean probability of recidivating with a drug sale crime increases from 2.8% to 4.1%. Again, for context, inmates with prior drug sale history have a 5.2% higher chance to recidivate with another drug sale crime.

These peer effects are large, especially for drug sales, where the size of the peer effect is roughly 25% of the effect of having an own-history with a crime-type. However, they are consistent with other papers in the peer effects literature. For instance, Bayer et al. (2009) find that for felony drug offenses among juveniles, the reinforcing effect is 0.25. With a peer-exposure standard deviation of 0.1, the mean probability of recidivating with a felony drug offense increases by 2.5 p.p., about double our effect. The broader literature on peer effects also finds large effects. Kremer and Levy (2008), for example, study the peer effects of college students who frequently consumed alcohol prior to college on the GPA of their roommates, and estimate peer effects that are quite large. Further, Carrell et al. (2008) argue that each high school cheater admitted to a US military service academy spread academic dishonesty to as many as 0.47 more students during their college years.

3.4 Heterogeneous Effects

In addition to the main results, we explore heterogeneity in the peer effect by ex ante prisoner characteristics. We focus on the age of the prisoner, the presence of mental health problems, and sentence length. All regressions in this section focus on race-by-age specific peer measures. The results are reported in Table  7.