Relationships among Race,Education,Criminal Thinking,and Recidivism

Abstract

Moderator and mediator relationships linking variables from three different theoretical traditions—race (subcultural theory), education (life-course theory), and criminal thinking (social learning theory)—and recidivism were examined in 1,101 released male federal prison inmates. Preliminary regression analyses indicated that racial status (White, Black, Hispanic) moderated the relationship between criminal thinking, as measured by the General Criminal Thinking (GCT) score of the Psychological Inventory of Criminal Thinking Styles (PICTS), and recidivism. Further analysis, however, revealed that it was not racial status, per se, that moderated the relationship between the PICTS and recidivism, but educational attainment. Whereas the PICTS was largely effective in predicting recidivism in inmates with 12 or more years of education, it was largely ineffective in predicting recidivism in inmates with fewer than 12 years of education. When education and the GCT score were compared as possible mediators of the race–recidivism relationship only the GCT successfully mediated this relationship. Sensitivity testing showed that the GCT mediating effect was moderately robust to violations of the sequential ignorability assumption on which causal mediation analysis rests. Moderator and mediator analyses are potentially important avenues through which theoretical constructs can be integrated and assessment strategies devised.

Keywords

race education criminal thinking recidivism moderation mediation

Traditional social science theories frequently offer single-variable explanations for complex behaviors. This is particularly true of efforts to explain crime. Crime has been ascribed to social disadvantage, family structure, peer relations, opportunity, and a host of other factors but only occasionally (e.g., Thornberry, 1987) do we see serious attempts to integrate these concepts. For research, policy, and clinical reasons, the desire on the part of many in the field to have an integrated perspective is strong. In fact, three of the most currently popular theories of crime, Moffitt’s (1993) dual trajectory model, Gottfredson and Hirschi’s (1990) self-control model, and Sampson and Laub’s (1993) age-graded life-course model of informal social control, have sought to integrate traditional sociological concepts with principles from biology, psychology, and economics, respectively. Despite these promising developments, the field of criminology remains a field dominated by single-variable theories. As such, it has a long way to go before it can offer a broad-based integrated perspective on crime. It is my contention that greater progress could be made toward developing a broad-based integrated perspective on crime if scholars working in the fields of criminology and criminal justice would examine moderating and mediating effects between variables from different models.

When discussing moderator and mediator variables the first order of business is properly defining the terms because they are often confused with each other even though they represent two distinct processes. As outlined in a classic paper by Baron and Kenny (1986), a moderator affects the direction or strength of the relationship between an independent or predictor variable and a dependent or outcome variable, whereas a mediator accounts, in part or in whole, for the relationship between an independent or predictor variable and dependent or outcome variable. To say that race moderates the criminal thinking–recidivism relationship means that the criminal thinking–recidivism relationship is weaker or reversed for some races than others (e.g., Caucasian vs. African American). To say that education mediates the race–recidivism relationship means that education accounts, at least in part, for the race–recidivism relationship through a causal chain of events that runs from race to education to recidivism. The purpose of the current study was to examine moderator and mediator relationships between concepts from three different theories of crime—race (subcultural theory), education (life-course theory), and criminal thinking (social learning theory)—and recidivism as a means of demonstrating how these constructs might be integrated for both theoretical and research purposes.

Whether crime is assessed with official arrest records or offender self-report, African American men are more often arrested, convicted, and incarcerated for crime, and more often report criminal involvement than their European American counterparts (Federal Bureau of Investigation, 2010; Sampson & Wilson, 1995). Rates of recidivism tend to be higher and survival times lower in African American men when African American men and European American men are compared (Jung, Spjeldnes, & Yamatani, 2010; Sabol, Minton, & Harrison, 2007). Black–White differences in crime are particularly pronounced when analyses are restricted to more serious offenses (Piquero & Brame, 2008). Research further indicates that Hispanics commit fewer crimes than Blacks but significantly more crimes than Whites (Steffensmeier, Feldmeyer, Harris, & Ulmer, 2011). In all likelihood, the race–crime relationship is driven by a combination of cultural, structural, and social psychological factors, but it remains a central feature of several criminological theories, to include culture conflict theory (Sellin, 1938), social disorganization theory (Burgess & Bogue, 1967), and, most particularly, Wolfgang and Ferracuti’s (1967) subculture of violence perspective.

Education has also been found to correlate with recidivism and is considered a criminogenic need capable of significantly reducing an adult or juvenile offender’s future chances of offending (Vieira, Skilling, & Peterson-Badali, 2009). In their age-graded life-course theory of offending, Sampson and Laub (1993) discuss the role of marriage, military service, employment, and education in transitioning youthful offenders away from crime during late adolescence and early adulthood. According to Sampson and Laub (1993), these life-course changes have the power to divert a youthful offender from a criminal trajectory to a noncriminal trajectory. Hence, educational achievement should be capable of reducing future recidivism. Results from a series of recent studies, in fact, suggest that educational attainment during incarceration is associated with a significantly lower rate of recidivism as well as a significant reduction in the seriousness of subsequent offenses (Blomberg, Bales, Mann, Piquero, & Berk, 2011). School attendance was particularly beneficial to African American males but was effective in preventing future criminality across all racial and gender groups (Blomberg, Bales, & Piquero, 2012).

In contrast to the sociological bent of most criminological theories, the social learning model takes a decidedly psychological approach to crime. With an emphasis on individual psychology rather than group process, social learning theory considers crime a function of observational learning, covert reinforcement, and cognitive mediation. As part of his social learning–influenced criminal lifestyle model, Walters (2012) proposes the existence of six cognitive mediators (criminal thinking styles, attributions, outcome expectancies, efficacy expectancies, goals, and values). These cognitive variables are held to mediate important relationships between crime-relevant variables. Using the General Criminal Thinking (GCT) score from the Psychological Inventory of Criminal Thinking Styles (PICTS: Walters, 1995) as a measure of criminal thinking, Walters (2009, 2011b, 2011c) has shown that the PICTS is capable of predicting recidivism after age and criminal history had been controlled. It is important for both theoretical and practical reasons to understand if status variables such as race and gender moderate this relationship and whether the GCT mediates important relationships between social status variables and recidivism.

The first goal of this study was to evaluate whether race moderates the relationship between criminal thinking and recidivism. In other words, is the criminal thinking–recidivism relationship invariant across race? Although research has failed to identify consistent PICTS score differences between Black, White, and Hispanic inmates (Walters, 2002), there is some indication that race may moderate the relationship between the PICTS and release outcome because recidivism is most effectively predicted in White inmates and least effectively predicted in Hispanic inmates (Walters, 1997; Walters, Frederick, & Schlauch, 2007). As a self-report measure, the PICTS requires a sixth-grade reading level or higher to complete. Education—which tends to be lower in Black and Hispanic inmates than in White inmates (Harlow, 2003)—and not race, may therefore be responsible for moderating the relationship between the PICTS and recidivism. The second goal of this study was to evaluate whether education and the PICTS mediate the relationship between race and recidivism. Using causal mediation analysis and sensitivity testing (Imai, Keel, & Tingley, 2010), the mediating effect of both education and the GCT score on the race-recidivism relationship were evaluated.

Method

Participants

The current sample was composed of 1,101 male inmates released from prison less than 42 months after providing a valid PICTS at a medium security federal correctional institution. At the time of their release, participants ranged in age from 19 to 78 years (M = 35.71, SD = 9.81) and had between 3 and 20 years of education (M = 11.33, SD = 1.94). More than two thirds of the sample was Black (68.1%), with the remainder of the sample consisting of White (19.6%), Hispanic (11.5%), and Asian/Native American (0.8%) participants. Nearly three quarters of the sample listed their marital status as single (73.3%), whereas 19.5%, 6.6%, and 0.5% listed their marital status as married, divorced, and widowed, respectively. The two most common instant offenses in the current sample were drugs (27.5%) and parole/supervised release violations (27.1%); the remainder of the sample was serving time for weapons violations (16.3%), robbery (9.3%), violent crimes (5.3%), property crimes (4.3%), and miscellaneous offenses (10.2%).

Psychological Inventory of Criminal Thinking Styles

The PICTS (Walters, 1995) is an 80-item self-report measure designed to assess the eight thinking styles believed to support a criminal lifestyle. Each PICTS item is scored on a 4-point Likert-type scale (strongly agree, agree, uncertain, disagree), with strongly agree responses being assigned four points, agree responses three points, uncertain responses two points, and disagree responses one point on all scales except Defensiveness-revised (Df-r) where items are reverse-scored (strongly agree = 1, agree = 2, uncertain = 3, disagree = 4). The PICTS yields two 8-item validity scales—Confusion-revised (Cf-r) and Df-r—eight 8-item nonoverlapping thinking style scales—Mollification (Mo), Cutoff (Co), Entitlement (En), Power Orientation (Po), Sentimentality (Sn), Superoptimism (So), Cognitive Indolence (Ci), and Discontinuity (Ds)—four factor scales—Problem Avoidance (PRB), Infrequency (INF), Self-Assertion/Deception (AST), and Denial of Harm (DNH)—two composite scales—Proactive Criminal Thinking (P) and Reactive Criminal Thinking (R)—and a General Criminal Thinking (GCT) score. The GCT score, which is the sum of the raw scores for the eight PICTS thinking style scales, served as the criterion for criminal thinking in the current investigation. Research indicates that the GCT is an internally consistent (α = .95) measure with strong test-retest reliability (r = .84-.86, after 12 weeks), and moderate predictive validity (unweighted mean r with recidivism = .22: Walters, 2002).

Procedure

The base sample from which data used in the current study were drawn contained 3,039 male inmates who had completed the PICTS between 2003 and 2010 as part of a routine intake screening procedure held in a medium security federal correctional institution in the northeastern United States. Nearly half the base sample (n = 1,435, 47.2%) had been released from prison at some point during the follow-up. In light of research showing that the predictive efficacy of the PICTS drops off when the test-release interval (number of months between completion of the PICTS and release from prison) exceeds 41 months (Walters, 2011c), 132 inmates with test-release intervals greater than 41 months were removed from the sample. One hundred and forty inmates who spent fewer than 12 months in the community and showed no evidence of recidivism and 62 inmates who produced invalid PICTS profiles (20 or more unanswered items, Cf-r ≥ T-score of 95, or Df-r ≥ T-score of 68; Walters, 2011c) were also eliminated from the study. This resulted in a final sample of 1,101 male inmates who recidivated during the follow-up or were recidivism-free for at least 12 months in the community.

The independent (predictor) variable for the moderator analyses was the PICTS GCT score. Race (White and Asian/Native American = 1, Black = 2, Hispanic = 3) and education (<12 years = 1, ≥12 years = 2) served as moderator variables. Recidivism constituted the dependent (outcome) variable. A review of electronic files from the Federal Bureau of Investigation’s National Crime Information Center and Federal Bureau of Prisons’ inmate data base were used to construct the recidivism outcome (yes = 1, no = 0). Technical parole violations were counted as evidence of recidivism when they resulted in the individual’s return to prison. Nearly two thirds of the sample (65.8%) received at least one charge. For the purposes of comparison and incremental validity, two additional predictor variables were included in some of the analyses: age at time of release from prison and the violation factor score from the Lifestyle Criminality Screening Form (LCSF-V; Walters, White, & Denney, 1991). The LCSF-V was employed as a proxy for criminal history in this study.

Moderator analyses were conducted in two stages. In the first stage, the dichotomous recidivism outcome measure (Y) was regressed onto the GCT score (X), moderator variable (Z; either race or education), and GCT-moderator interaction (XZ) as part of a Cox survival regression analysis that considers both the recidivism event (0/1) and time until recidivism. If the interaction was significant, nonlinear quadratic terms (X², Z²) were added to the equation and inserted before the interaction term and the analysis recomputed (Cortina, 1993; Lubinski & Humphreys, 1990). In the second stage, receiver operating characteristic (ROC) curves were calculated separately for White, Black, and Hispanic inmates and for inmates with less (<12 years) and more (≥12 years) education. ROC analyses were also calculated for inmates with less and more education within each of the three racial groups. Cox regression analyses were also used to test the incremental validity of the GCT relative to age at time of release and criminal history (LCSF-V). These analyses were conducted separately for each educational level (<12 years, ≥12 years) in the full sample as well as in each of the three racial subsamples.

A mediation analysis was conducted with race (White and Asian/Native American = 1, Black and Hispanic = 2) as the independent variable, recidivism as the dependent (outcome) variable, and education and GCT each considered as mediator variables. Given the prospective nature of the current design (i.e., race preceded education/GCT and education/GCT preceded recidivism) a mediation analysis was justified. Causal mediation analysis was performed with algorithms devised by Imai and his colleagues (Imai, Keel, & Tingley, 2010; Imai, Keel, Tingley, & Yamamoto, 2010). The (continuous) mediator models (education, criminal thinking) were fit with linear least squares regression and the (binary) outcome model (recidivism) was fit with probit regression. After completing Step 1 of the procedure, output objects were bootstrapped 1,000 times with replacement using a nonparametric mediational analysis. Sensitivity to violations of the sequential ignorability assumption (absence of confounders that could explain the mediation results) was tested with a sensitivity analysis of the mediation and outcome models using two covariate confounders: age (in years) and the LCSF-V score.

Results

Descriptive statistics for and intercorrelations between the three theoretical variables (race, education, GCT) and the single-outcome measure (recidivism) are provided in Table 1. As the results indicate, all three theoretical variables correlated significantly among themselves and each predicted recidivism.

Table 1.

Descriptive Statistics and Correlations for the Three Predictors and Recidivism Outcome.

				Correlations
Predictor	Mean	SD	Range	Education	GCT	Recidivism
Race	1.80	0.40	1-2	−.15**	.11**	.12**
Education	11.33	1.94	3-20		−.12**	−.08*
PICTS GCT	113.85	27.20	67-217			.17**
Recidivism	0.66	0.48	0-1

Note. Race = White (1) versus Black or Hispanic (2); Education = years of education; PICTS GCT = General Criminal Thinking score of the Psychological Inventory of Criminal Thinking Styles; SD = standard deviation; Range = high and low scores on this particular measure in the current sample; N = 1101.

p < .01. **p < .001.

Moderator Analyses

Table 2 summarizes the results of the first set of moderator analyses. Cox regression analysis of race as a moderator variable revealed a significant interaction between the PICTS GCT score and race (left column of Table 2). When nonlinear quadratic terms (X², Z²) were added to the equation, the PICTS GCT effect remained significant (Wald = 5.43, p < .05) but the interaction became nonsignificant (Wald = 3.60, p = .06). Replacing race with education as the moderator of the GCT–recidivism relationship resulted in a Cox regression analysis in which none of the main effects (GCT, education) or the interaction effect (education × GCT) achieved statistical significance (right column of Table 2). It should be noted, however, that before the interaction term was introduced into the model, both the GCT (Wald = 33.19, p < .001) and education (Wald = 4.27, p < .05) main effects were significant. At first glance, these results seem consistent with the conclusion that race exerted a moderating effect on the GCT–recidivism relationship in this study. Further analysis, however, revealed that the significant interaction between race and the GCT became nonsignificant when quadratic terms were added to the estimated model and that although education did not achieve a statistically significant interaction with GCT, inclusion of the education × GCT interaction term eliminated a highly significant GCT effect, something the race × GCT interaction failed to do even when quadratic terms were added to the model.

Table 2.

Summary of Cox Regression Results With Race and Education as Moderator Variables.

	Race moderator		Education moderator
Predictor	Wald	$\exp (β_{x}^{s})$ [95% CI]	Wald	$\exp (β_{x}^{s})$ [95% CI]
PICTS GCT	21.85**	1.02 [1.012, 1.029]	0.26	1.00 [0.993, 1.011]
Moderator	8.51*	2.09 [1.275, 3.442]	2.79	0.58 [0.308, 1.099]
Moderator × GCT	8.73*	0.99 [0.990, 0.998]	2.38	1.00 [0.998, 1.009]

Note. Figures reported are the final logistic regression results at the end of Block 2; Predictor = predictor variables; PICTS GCT = General Criminal Thinking score of the Psychological Inventory of Criminal Thinking Styles; Moderator = moderator variable, either race (White = 1, Black = 2, Hispanic = 3) or education (1 = less than 12 years, 2 = 12 or more years); Moderator × GCT = moderator (either race or education) by GCT interaction; Wald = Wald statistic with a χ² distribution and one degree of freedom; exp(β) = exponent of the unstandardized coefficient in the form of an odds ratio (numbers below 1.00 indicate a negative relationship with the criterion and numbers above 1.00 indicate a positive relationship with the criterion); 95% CI = 95th percentile confidence interval for the exponent of the estimated coefficient; N = 1101.

p < .01. **p < .001.

To further investigate the ability of race and education to moderate the GCT–recidivism relationship, ROC analyses were computed by race (White, Black, Hispanic) and education (<12 years, ≥12 years). As the results outlined in Table 3 indicate, the GCT score achieved the most accurate results with White participants, the least accurate results with Hispanic participants, and intermediate results with Black participants. A z-test procedure designed to compare independent ROC curves, nonetheless, failed to detect any significant differences in the area under the curve (AUC) results between Blacks and Whites, z = 0.87, p > .10, Blacks and Hispanics, z = 1.04, p > .10, or Hispanics and Whites, z = 1.58, p > .10. A significant difference in accuracy did arise, however, when higher and lower educated participants were compared, z = 2.20, p < .05. A z-test comparison of dependent ROC curves revealed that the PICTS GCT score was significantly more accurate in predicting recidivism in higher educated participants than age at time of release, z = 2.14, p < .05, and the LCSF-V was significantly more accurate than the GCT in predicting recidivism in lower educated participants, z = 2.40, p < .05.

Table 3.

Receiver Operating Characteristic Results by Race and Educational Level.

Description	N	% Invalid	GCT	Age	LCSF-V
Total sample	1,101	5.3	.604 [.569, .639]**	.595 [.560, .630]**;	.612 [.576, .647]**
Education ≥ 12	672	4.4	.628 [.584, .671]**	.565 [.521, .609]*	.592 [.548, .636]**
Education < 12	429	6.7	.546 [.488, .605]	.618 [.558, .677]**	.645 [.584, .706]**
White	224	4.8	.640 [.568, .713]**	.621 [.547, .695]*	.603 [.528, .678]*
Education ≥ 12	170	4.6	.658 [.577, .740]**	.579 [.492, .665]	.605 [.520, .690]*
Education < 12	54	5.7	.540 [.373, .707]	.691 [.534, .848]*	.591 [.425, .757]
Black	750	5.1	.602 [.557, .646]**	.572 [.528, .616]*	.591 [.546, .636]**
Education ≥ 12	448	4.5	.618 [562, .674]**	.553 [.498, .608]	.576 [.520, .632]*
Education < 12	302	5.9	.558 [.484, .632]	.578 [.502, .655]*	.613 [.533, .693]*
Hispanic	127	7.3	.543 [.442, .644]	.668 [.574, .762]**	.638 [.542, .735]*
Education ≥ 12	54	1.8	.619 [.467, .771]	.587 [.432, .741]	.523 [.366, .679]
Education < 12	73	11.0	.499 [.365, .633]	.714 [.592, .835]*	.706 [.584, .827]*

Note. Description = description of sample in terms of race (White, Black, Hispanic) and education (12 or more years, less than 12 years); N = sample or subsample size; % Invalid = proportion of Psychological Inventory of Criminal Thinking Styles (PICTS) that were removed because they were invalid; GCT = PICTS General Criminal Thinking score; Age = age at time of release (inverse scored); LCSF-V = Violation factor of the Lifestyle Criminality Screening Form; the first number in the GCT, Age, and LCSF-V columns is the area under the ROC curve (AUC) and the next set of numbers [in brackets] is the 95% confidence interval of the AUC.

p < .05. **p < .001.

A second round of Cox regression analyses were conducted using all three indicators (age at time of release, LCSF-V, PICTS GCT) as predictors for the purpose of determining whether the PICTS GCT score possessed incremental validity relative to age and the LCSF-V. When the analyses were restricted to inmates with higher educational levels, the GCT score displayed incremental validity relative to age at time of release and LCSF-V (left half of Table 4). When the analyses were restricted to inmates with lower educational levels, the GCT score failed to achieve incremental validity relative to age at time of release and the LCSF-V (right half of Table 4). When these analyses were conducted on the three racial groups separately, the GCT score achieved incremental validity relative to age and the LCSF-V in White (Wald = 8.39, p < .01) and Black (Wald = 14.24, p < .001) inmates with higher educational levels but not in White (Wald = 2.20, p >.10) and Black (Wald = 3.31, p =.07) inmates with lower educational levels. Although the GCT score failed to demonstrate incremental validity in either higher (Wald = 1.07, p > .10) or lower (Wald = 0.55, p > .10) educated Hispanic participants, the beta value was in the predict direction (positive) in the higher educated Hispanic group and in the nonpredicted direction (negative) in the lower educated Hispanic group.

Table 4.

Summary of Incremental Validity Cox Regression Analyses for Inmates With ≥12 Years of Education and for Inmates With <12 Years of Education.

	Education ≥ 12 years (n = 672)		Education < 12 years (n = 429)
Predictor	Wald	$\exp (β_{x}^{s})$ [95% CI]	Wald	$\exp (β_{x}^{s})$ [95% CI]
Age	6.95*	0.87 [0.78, 0.96]	15.45**	0.78 [0.69, 0.88]
LCSF-V	22.69**	1.27 [1.15, 1.40]	16.85**	1.25 [1.12, 1.38]
PICTS GCT	21.70**	1.24 [1.13, 1.36]	2.93	1.10 [0.99, 1.24]

Note. Figures reported are the final logistic regression results at the end of Block 2; Predictor = predictor variables; Age = age at time of release from prison; LCSF-V = violation factor of the Lifestyle Criminality Screening Form; PICTS GCT = General Criminal Thinking score of the Psychological Inventory of Criminal Thinking Styles; Wald = Wald statistic with a χ² distribution and one degree of freedom; exp(β) = exponent of the x-standardized (M = 0, SD = 1) coefficient in the form of an odds ratio (numbers below 1.00 indicate a negative relationship with the criterion and numbers above 1.00 indicate a positive relationship with the criterion); 95% CI = 95th percentile confidence interval for the exponent of the estimated coefficient.

p < .01. **p < .001.

Mediator Analyses

In causal mediation analysis a significant mediating effect is defined by a 95% confidence interval that does not include zero (Imai, Keel, & Tingley, 2010; Imai, Keel, Tingley, et al., 2010). As indicated by the results in Table 5, education failed to mediate the relationship between race and recidivism. The GCT score, on the other hand, successfully mediated the race-recidivism relationship (see Table 6). Further analysis revealed that the mediating effect of the GCT score on the race-recidivism relationship was only partial (direct effect still significant) and that it accounted for approximately 15% of the total variance in this relationship. When the GCT score was used to mediate the education–recidivism relationship the total model effect failed to achieve significance (95% confidence interval = −0.0142 to 0.0045).

Table 5.

Results of a Causal Mediation Analysis in Which Education Served as a Mediator of the Race–Recidivism Relationship.

Effect type	Point estimate	95% CI
Mediation effect (Race → Educ → Recid)	0.0051	[−0.0043, 0.0157]
Direct effect (Race → Recid)	0.0968	[0.0244, 0.1699]
Total effect	0.1019	[0.0299, 0.1713]
Proportion of total effect via mediation	0.0494	[0.0297, 0.1627]

Note. Race = White (1) versus non-White (2). Educ = education (in years); Recid = recidivism (yes = 1, no = 0); Point Estimate = estimate of the size of the effect; 95% CI = 95% confidence interval of the point estimate; N = 1101.

Table 6.

Results of a Causal Mediation Analysis in which General Criminal Thinking Served as a Mediator of the Race-Recidivism Relationship.

Effect type	Point estimate	95% CI
Mediation effect (Race → GCT → Recid)	0.0159	[0.0053, 0.0294]
Direct effect (Race → Recid)	0.0906	[0.0173, 0.1592]
Total effect	0.1066	[0.0362, 0.1783]
Proportion of total effect via mediation	0.1477	[0.0887, 0.4288]

Note. Race = White (1) vs. non-White (2). Educ = education (in years); Recid = recidivism (yes = 1, no = 0); Point estimate = estimate of the size of the effect; 95% CI = 95% confidence interval of the point estimate; N = 1,101.

Sensitivity testing was performed on the mediation and outcome models using age and the LCSF-V score as confounding covariates. The results indicated a rho (ρ) at which mediation equals zero of .16 (see Figure 1). Coefficients of determination (R²) for the mediator and outcome models were used to construct a graph of the amount of variance that an unobserved confounder would have to explain to totally eliminate the mediation effect of GCT on the race–recidivism relationship (see Figure 2). According to the graph, an unobserved confounding variable or set of variables would need to account for approximately 14% of the variance in the mediator and 14% of the variance in the outcome to reduce the mediation effect to zero.

Figure 1.

Sensitivity analysis of the binary recidivism outcome and continuous criminal thinking mediator.

Figure 2.

Sensitivity analysis of the binary recidivism outcome and continuous criminal thinking mediator.

Regression analyses in which GCT was regressed onto age and the LCSF-V score revealed an R² of .018 (mediation model) whereas logistic regression analyses in which recidivism was regressed onto age and the LCSF-V produced a pseudo-R² of .067 (outcome model). Only the binomial logistic regression results (outcome model) exceeded the minimum requirements for explaining recidivism (R² ≥ .035) or GCT (R² ≥ .04), and even then the amount of mediator variance required to bring the mediation effect down to zero would have had to have been more than 45%. Instead, the two covariate confounders accounted for only 1.8% of the variance in the mediator variable.

Discussion

The results of this study suggest that testing for moderation and mediation can be helpful in integrating concepts from different theories. Patterns of moderation and mediation may provide clues as to how different concepts might fit together in an integrated theory. A concept central to life-course theory (educational attainment), for instance, moderated the criminal thinking–recidivism relationship whereas a concept central to subcultural theory (race) did not. Likewise, a concept central to social learning theory (criminal thinking) mediated the race–recidivism relationship whereas educational attainment did not. In the moderator analyses, the GCT score was able to predict recidivism in individuals with 12 or more years of education but it failed to predict recidivism in individuals with fewer than 12 years of education. Because 61% of the current sample had 12 years or more of education and more than two thirds of samples such as the present one dropped out of high school prior to completing the 12th grade (Walters et al., 1991), a substantial portion of participants in the current study probably earned their GEDs in prison. Educational attainment in prison has been found to exert a protective effect against future recidivism (Blomberg et al., 2011); and in the current study, it displayed a moderating effect on the GCT–recidivism relationship but failed to mediate the race–recidivism relationship.

If one follows the logic that interaction between an independent variable and putative moderator variable in a regression equation equals a moderator effect, then race moderated the criminal thinking–recidivism relationship in this study. However, when nonlinear monotonic terms for the independent and moderator variables (X² and Z²) were included in the regression equation (Cortina, 1993; Lubinski & Humphreys, 1990), the previously significant interaction effect turned nonsignificant. In addition, even though the education × GCT interaction failed to achieve statistical significance in this study, the inclusion of this interaction term in the regression equation completely eliminated an effect for GCT that had been significant without the interaction. Further analyses were accordingly conducted for the purpose of disentangling the possible confounding effect of education on race. In these analyses, AUC values were calculated for subgroups divided by race (White, Black, Hispanic) and education (<12 years, ≥ 12 years). Rice and Harris (2005) have determined that an AUC of .556 is comparable to a small effect size, an AUC of .639 is comparable to a moderate effect size, and an AUC of .714 is comparable to a large effect size. Effect size estimates for individuals with 12 or more years of education were of low-moderate to moderate magnitude across all three racial conditions, whereas effect size estimates for individuals with fewer than 12 years education were of zero to small magnitude for White, Black, and Hispanic respondents. These results indicate that education was the moderating variable in this study and that differences in efficacy across the three racial groups were an artifact of a greater proportion of Black (40.3%) and Hispanic (57.5%) than White (24.1%) inmates with fewer than 12 years of education.

Whereas moderation provides information on the relative stability or invariance of a relationship, mediation provides information on the nature of a relationship. In the current study, criminal thinking was found to partially mediate the relationship between race and recidivism. Hence, criminal thinking was differentially associated with race and recidivism such that a portion of the relationship between race and recidivism was explained by criminal thinking. Other cognitive and noncognitive variables probably also mediate the race–recidivism relationship. One recent study, in fact, confirmed that Black inmates had significantly more positive outcome expectancies for crime than White inmates (Walters, 2011a). Unfortunately, because there was no outcome variable (e.g., recidivism) in that study no conclusions about mediation can be drawn. Even so, Walters (2012) includes outcome expectancies and criminal thinking styles in his social learning theory of cognitive mediation. A study in which all six cognitive mediators (criminal thinking styles, attributions, outcome expectancies, efficacy expectancies, goals, and values) are considered simultaneously would provide more information on the role of cognitive mediation in the race–recidivism relationship.

There are methodological, theoretical, practical, and policy implications to the current findings that need to be discussed. Methodologically, the current results indicate that equating interaction with moderation is problematic in several respects. First, there is the problem of Type I errors. As reported by Lubinski and Humphreys (1990) and others (Cortina, 1993; Ganzach, 1997), a spurious interaction effect can surface if quadratic terms are not added to the regression equation. In the current study, the significant race × GCT interaction disappeared once quadratic terms for the independent and moderator variables were added to the model. Second, there is the problem of Type II errors. Whereas a Type I error entails rejecting a null hypothesis that is true, a Type II error involves failing to reject a null hypothesis that is, in fact, false. Interaction has been known to suffer from low power (Cronbach, 1987; Lubinski & Humphreys, 1990). Adding quadratic terms to the estimated model can reduce the power of the interaction term even further (Cortina, 1993) and may be why the race × GCT interaction went from significant (p < .01) to nonsignificant (p = .06) after the quadratic terms had been added. Be this as it may, dividing the sample along racial and educational lines made it clear that the chief moderating variable in the current study, even though it did not produce a significant interaction effect, was education. In testing mediation of the race–recidivism relationship I relied on a relatively new procedure development by Imai, Keel, & Tingley (2010) and Imai, Keel, Tingley, et al. (2010). Despite its recent arrival on the methodological scene, Imai et al.’s procedure is well documented and supported by mathematical proofs. In addition, it has the capacity for sensitivity testing, something that is missing from most structured equation modeling programs currently used to assess mediation effects.

The current study also has implications for theory development. Given the plethora of single variable models in criminology the task of organizing and integrating these models into a single comprehensive theory is truly formidable. Integration could nonetheless be made easier with the aid of moderation and mediation analysis. In the current study, concepts central to three different criminological models—race (subcultural theories), education (life-course theories), and criminal thinking (social learning theories)—were integrated using moderation and mediation methodologies. The results indicated that education but not race moderated the criminal thinking–recidivism relationship and that criminal thinking but not education mediated the race-recidivism relationship. Moderation and mediation analysis can be very useful in determining which constructs from various models belong together and which ones do not. The mediation analyses were particularly helpful in clarifying important criminological relationships. From the mediation results we can surmise that the effect of race on recidivism is at least partially mediated by cognitive factors such as criminal thinking, as well as, perhaps, by positive outcome expectancies for crime (Walters, 2011a), lower investment in marriage (King & South, 2011), and social-environmental differences between Black and White Americans (Phillips, 2002; Sampson, Morenoff, & Raudenbush, 2005). Future research could further clarify important theoretical relationships by examining how these and other potential mediating variables interact to shape and influence the race–recidivism relationship.

The current results, in addition to their methodological and theoretical implications, have potentially important practical implications, particularly for assessment. A self-report inventory is only as good as its ability to be comprehended by those who complete it. The PICTS items were originally written to reflect a 6th-grade reading level but subsequent analyses have shown that some of the items require as much as a 9th- or 10th-grade reading level (Walters, 2002). Completing 12 years of education or receiving a GED were proxies for a sufficient level of reading ability in the current study. Most of the inmates who participated in this study (61%) attained one of these two markers and were effectively classified for recidivism risk at a low-moderate to moderate level by the PICTS GCT score. The remaining 39% of the sample, however, could not be effectively classified for recidivism risk based on the results of the PICTS GCT score. These findings lead to the following two conclusions. First, we must rely more on demographic and historical measures such as age and the LCSF-V and less on self-report measures such as the PICTS when evaluating individuals with limited educational attainment and reading ability. Second, for the PICTS to be effective with a larger segment of the inmate population, the item content may need to be revised so that individuals with less than a 12th-grade education or 9th-grade reading level can comprehend the items sufficiently well enough to construct a valid protocol. Classification of offenders on the basis of race would be both illegal and unethical. The current results, however, point to a factor (i.e., criminal thinking) that may be partially responsible for racial differences in recidivism and which could lay the groundwork for a more rationale and effective offender classification system.

Both the moderation and mediation results from this study have implications for criminal justice policy and correctional decision making. The policy implications of the moderating effect of education on the efficacy of the PICTS GCT score in predicting recidivism is that guidelines need to be established for the proper use of the PICTS in identifying criminogenic needs and assigning offenders to programs and levels of supervision. Depending on the offender’s education or reading level, alternative procedures (criminal history and other static risk factors) may need to be employed for the purpose of making program and classification decisions. The policy implications of the race → criminal thinking → recidivism mediation effect is that it emphasizes the necessity of including a criminal thinking component in secondary and tertiary prevention programs for minority youth. Although criminal thinking is a criminogenic need in offenders of all races and at all ages (Andrews, Bonta, & Wormith, 2006), its mediating role in the race–recidivism relationship suggests that it may be a particularly salient goal for interventions with African American and Hispanic youth.

Limitations

A principal limitation of the moderator results of this study is that reading ability was not tested. If reading ability rather than education moderates the relationship between criminal thinking, as measured by the PICTS GCT score, and recidivism, then additional research is required using a validated adult measure of reading comprehension such as the Adult Basic Learning Examination (ABLE) or Tests of Adult Basic Education (TABE). In addition, factors other than or in addition to reading level may moderate the GCT–recidivism relationship. Motivation may be one such factor. Lack of motivation could lead someone to drop out of school as well as exert less than maximum effort on a psychological inventory. The GCT–recidivism relationship could consequently be moderated by reading ability, motivation, and other correlates of education, but not race. Understanding moderation is a little like peeling back the layers of an onion. Race represents the outer layer of the onion, education the next layer, reading level and motivation the layer below this, and so on until we reach the core moderator variable(s).

The mediation results also suffer from a significant limitation. The confounding covariates in the sensitivity analysis should predate the treatment (independent) variable. In the current study, the treatment or independent variable was race. Because race is present at birth it was not possible to find confounding covariates in the current database that preceded the independent variable. This is a limitation but not a fatal flaw of the mediation analysis. The use of pretreatment confounders is based on the assumption that the independent and mediator variables were assigned randomly to participants. Using a nonmanipulated independent variable such as race will always raise questions about sequential ignorability. In the current study, the conditions for prospective analysis were satisfied—that is, the independent variable (race) preceded the mediating variable (GCT) that then preceded the dependent variable (recidivism)—and the two potential confounders (age and prior criminal history as represented by the LCSF-V), although not predating the independent variable, did predate the mediator. Moreover, in light of the fact that age and criminal history are often considered the two best predictors of recidivism (Gendreau, Little, & Goggin, 1996), potentially significant confounders of the race → GCT → recidivism mediated pathway were examined and found not to present a serious challenge to the robustness of the mediation effect.

Sampling was a limitation that affected both the moderation and mediation analyses. Because participants in this study were incarcerated federal male offenders we do not know how well these results apply to nonincarcerated, state, or female offenders. Additional research is consequently required to test the generalizability of the current results to populations different from those included in the current study. Generalizability is also a limitation of the PICTS in the sense that it was only capable of effectively predicting recidivism in participants with 12 or more years of education. Lowering the reading level of some of the PICTS items should make the test more applicable to a larger portion of the inmate population, but it will never make it accessible to the entire inmate population. There will always be inmates who will need to be evaluated with demographic, historical, or behavioral indicators. This is a limitation inherent to all self-report measures of criminality, but it is a limitation we must endure if our goal is to assess criminal thinking.

Footnotes

Author’s Note

The assertions and opinions contained herein are the private views of the author and should not be construed as official or as reflecting the views of the Federal Bureau of Prisons or the United States Department of Justice.

Declaration of Conflicting Interests

The author declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: Glenn D. Walters is the author of the Psychological Inventory of Criminal Thinking Styles (PICTS) and receives remuneration from sales of the PICTS manual.

Funding

The author received no financial support for the research, authorship, and/or publication of this article.

References

Andrews

D. A.

Bonta

Wormith

S. J.

(2006). The recent past and near future of risk and/or need assessment. Crime & Delinquency, 52, 7-27.

Baron

R. M.

Kenny

D. A.

(1986). The moderator-mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology, 51, 1173-1182.

Blomberg

T. G.

Bales

W. D.

Mann

Piquero

A. R.

Berk

R. A.

(2011). Incarceration, education and transition from delinquency. Journal of Criminal Justice, 39, 355-365.

Blomberg

T. G.

Bales

W. D.

Piquero

A. R.

(2012). Is educational achievement a turning point for incarcerated offenders across race and sex? Journal of Youth and Adolescence, 41, 202-216.

Burgess

Bogue

D. J.

(Eds.). (1967). Urban sociology. Chicago, IL: University of Chicago Press.

Cortina

J. M.

(1993). Interaction, nonlinearity, and multicollinearity: Implications for multiple regression. Journal of Management, 19, 915-922.

Cronbach

L. J.

(1987). Statistical tests for moderator variables: Flaws in analyses recently proposed. Psychological Bulletin, 102, 414-417.

Federal Bureau of Investigation. (2010). Crime in the United States: 2009. Washington, DC: Author.

Ganzach

(1997). Misleading interaction and curvilinear terms. Psychological Methods, 2, 235-247.

10.

Gendreau

Little

Goggin

(1996). A meta-analysis of the predictors of adult offender recidivism: What works! Criminology, 34, 575-607.

11.

Gottfredson

Hirschi

(1990). A general theory of crime. Stanford, CA: Stanford University Press.

12.

Harlow

C. W.

(2003). Education and correctional populations [Bureau of Justice Statistics Special Report]. Washington, DC: U.S. Department of Justice.

13.

Imai

Keele

Tingley

(2010). A general approach to causal mediation analysis. Psychological Methods, 15, 309-334.

14.

Imai

Keele

Tingley

Yamamoto

(2010). Causal mediation analysis using R. In Vinod

H. D.

(Ed.), Advances in social science research using R (pp. 129-154). New York, NY: Springer.

15.

Jung

Spjeldnes

Yamatani

(2010). Recidivism and survival time: Racial disparity among jail ex-inmates. Social Work Research, 34, 181-189.

16.

King

R. D.

South

S. J.

(2011). Crime, race, and the transition to marriage. Journal of Family Issues, 32, 99-126.

17.

Lubinski

Humphreys

L. G.

(1990). Assessing spurious “moderator effects”: Illustrated substantively with the hypothesized (“synergistic”) relation between spatial and mathematical ability. Psychological Bulletin, 107, 385-393.

18.

Moffitt

T. E.

(1993). Adolescence-limited and life-course persistent antisocial behavior: A developmental taxonomy. Psychological Review, 100, 674-701.

19.

Phillips

J. A.

(2002). White, black, and Latino homicide rates: Why the difference? Social Problems, 49, 349-374.

20.

Piquero

A. R.

Brame

R. W.

(2008). Race-crime and ethnicity-crime relationship in a sample of serious adolescent delinquents. Crime & Delinquency, 54, 390-422.

21.

Rice

M. E.

Harris

G. T.

(2005). Comparing effect sizes in follow-up studies: ROC area, Cohen’s d, and r. Law and Human Behavior, 29, 615-620.

22.

Sabol

W. J.

Minton

T. D.

Harrison

P. M.

(2007). Prison and jail inmates at midyear 2006 (NCJ 217675). Bureau of Justice Statistics Bulletin. Washington, DC: Bureau of Justice Statistics.

23.

Sampson

R. J.

Laub

J. H.

(1993). Crime in the making: Pathways and turning points through life. Cambridge, MA: Harvard University Press.

24.

Sampson

R. J.

Morenoff

J. D.

Raudenbush

(2005). Social anatomy of racial and ethnic disparities in violence. American Journal of Public Health, 95, 224-232.

25.

Sampson

R. J.

Wilson

W. J.

(1995). Toward a theory of race, crime, and urban inequality. In Hagan

Peterson

R. D.

(Eds.), Crime and inequality (pp. 37-54). Stanford, CA: Stanford University Press.

26.

Sellin

(1938). Culture and conflict in crime. New York, NY: Social Science Research Council.

27.

Steffensmeier

Feldmeyer

Harris

C. T.

Ulmer

J. T.

(2011). Reassessing trends in black violent crime, 1980-2008: Sorting out the “Hispanic effect” in Uniform Crime Reports arrests, National Crime Victimization Survey offender estimates, and U.S. prisoner counts. Criminology, 49, 197-251.

28.

Thornberry

T. P.

(1987). Toward an interactional theory of delinquency. Criminology, 25, 863-892.

29.

Vieira

T. A.

Skilling

T. A.

Peterson-Badali

(2009). Matching court-ordered services with offender needs: Predicting treatment success with young offenders. Criminal Justice and Behavior, 36, 385-401.

30.

Walters

G. D.

(1995). The Psychological Inventory of Criminal Thinking Styles: Part I. Reliability and preliminary validity. Criminal Justice and Behavior, 22, 307-325.

31.

Walters

G. D.

(1997). Predicting short-term release outcome using the LCSF and PICTS. Journal of Mental Health in Corrections Consortium, 43, 18-25.

32.

Walters

G. D.

(2002). The Psychological Inventory of Criminal Thinking Styles (PICTS): A review and meta-analysis. Assessment, 9, 278-291.

33.

Walters

G. D.

(2009). The Psychological Inventory of Criminal Thinking Styles and Psychopathy Checklist: Screening Version as incrementally valid predictors of recidivism. Law and Human Behavior, 33, 497-505.

34.

Walters

G. D.

(2011a). Black-white differences in positive outcome expectancies for crime: A study of male federal prison inmates. Journal of Criminal Justice, 39, 192-197.

35.

Walters

G. D.

(2011b). Predicting recidivism with the Psychological Inventory of Criminal Thinking Styles and Level of Service Inventory–Revised: Screening Version. Law and Human Behavior, 35, 211-220.

36.

Walters

G. D.

(2011c). Taking the next step: Combining incrementally valid indicators to improve recidivism prediction. Assessment, 18, 227-233.

37.

Walters

G. D.

(2012). Crime in a psychological context: From career criminals to criminal careers. Thousand Oaks, CA: SAGE.

38.

Walters

G. D.

Frederick

A. A.

Schlauch

(2007). Postdicting arrests for proactive and reactive aggression with the PICTS proactive and reactive scales. Journal of Interpersonal Violence, 22, 1415-1430.

39.

Walters

G. D.

White

T. W.

Denney

(1991). The Lifestyle Criminality Screening Form: Preliminary data. Criminal Justice and Behavior, 18, 406-418.

40.

Wolfgang

M. E.

Ferracuti

(1967). Subculture of violence: Towards an integrated theory of criminology. London, England: Tavistock.