Implications of Different Outcome Measures for an Understanding of Inmate Misconduct

Abstract

Quantitative studies geared toward understanding differences among prison inmates in their odds of committing rule infractions have grown over the last decade but with little consistency in the models examined, especially regarding the types of rule violations examined. These differences have, in turn, contributed to an increasingly complex picture of inmate misconduct that appears counterproductive for both theory and practice. The study described here was designed to assess the ramifications of examining different outcome measures for quantitative analyses of the subject. Findings revealed that three of the nine models examined produced unique information regarding the effects of various inmate predictors, including the models of physical assaults (on inmates and/or staff), drug/alcohol use, and other nonviolent misconduct. Analyses also uncovered several new substantive findings on the topic. Findings are discussed in light of their relevance for practice as well as theories of inmate behavior.

Keywords

prison inmates rule infractions inmate misconduct

Rule violations committed by inmates during incarceration pose a significant barrier to the safety of both inmates and staff (Bottoms, 1999; DiIulio, 1987; Gendreau, Goggin, & Law, 1997), and understanding the factors that increase inmates’ likelihood of misconduct is critical for order maintenance. Although a considerable amount of research on inmate misconduct exists, reviews of this literature have generally revealed inconsistent findings regarding the correlates/causes of rule infractions with the exception of a few predictors such as an inmate’s age and criminal history (e.g., Bottoms, 1999; Gendreau et al., 1997; Wooldredge, 1991).

Some of the differences in findings between studies could be due to the examination of different facilities and systems, the use of official versus self-report data on misconduct, and differences in model specification. Yet differences in findings across studies could also be due, in part, to the types of rule violations examined. Some researchers have focused on specific categories of misconduct such as inmate assaults or drug offenses (e.g., Camp, Gaes, Langan, & Saylor, 2003; Cunningham & Sorensen, 2007; Harer & Steffensmeier, 1996; Huebner, 2003; Sorensen, Wrinkle, & Gutierrez, 1998), whereas others have examined more general measures that combine different types of offenses (Goetting & Howsen, 1986; Ruback & Carr, 1993; Wooldredge, Griffin, & Pratt, 2001). The extent to which the findings for any one outcome measure can be generalized to other outcome measures is unknown, and the assumptions underlying the decision to treat different types of misconduct as either the same or distinct have important implications for theories of inmate deviance and correctional policy. Using data on men incarcerated in state facilities compiled from two waves of the survey of state inmates conducted by the U.S. Census Bureau, we modeled several different outcome measures to determine whether different outcomes produce substantively different results and also to delineate those outcomes that offer the most unique (or least redundant) information across models.

Modeling Inmate Misconduct

Determining the primary influences on the prevalence and incidence of inmate misconduct is necessary for developing practical methods for enhancing the safety of prison/correctional facilities (Bottoms, 1999; DiIulio, 1987; Gendreau et al., 1997). Both inmate and facility characteristics have been examined in studies of the topic, and recent empirical findings have underscored the utility of examining bi-level effects on misconduct (Camp et al., 2003; Huebner, 2003; Wooldredge et al., 2001). As these bi-level studies increase in number, efforts to synthesize results across these studies to estimate effect sizes and to identify common themes may benefit from focusing on similar models. Although different methodological approaches are important for several goals of scientific inquiry, there are also advantages to adopting more similar approaches when studying a policy-relevant question such as the safety of a prison environment.

Rule violations by prison inmates are considered to be valid indicators of maladjustment to incarceration, and the prevalence and incidence of rule violations are the most commonly examined outcomes in studies of inmate deviance (e.g., Camp et al., 2003; Jiang & Winfree, 2006; Toch, Adams, & Grant, 1989; Van Voorhis, 1994). Yet despite the preference to examine misconduct, little agreement exists as to whether measures of different types of misconduct (e.g., drug/alcohol offenses vs. physical assaults) should be examined separately from one another. Scholars of the topic have examined pooled measures of all rule violations (Goetting & Howsen, 1986; Jiang & Winfree, 2006; Ruback & Carr, 1993; Wooldredge et al., 2001), whereas others have specified their analyses by more specific categories of misconduct such as violence, property offenses, assaults on inmates or staff, security offenses, or drug or alcohol violations (see, e.g., Camp et al., 2003; Cunningham & Sorensen, 2007; Gaes & McGuire, 1985; Gaes, Wallace, Gilman, Klein-Saffran, & Suppa, 2002; Harer & Steffensmeier, 1996; Huebner, 2003; McCorkle, Miethe, & Drass, 1995; Wooldredge, 1994). These contributions have yielded potentially important information regarding possible links between assorted predictors and misconduct, yet the statistical significance of findings across these studies and across models within the same studies have often been mixed for various predictors tapping inmate background characteristics (e.g., race, education), situational characteristics/routines during incarceration (e.g., program participation, hours spent at a work assignment), and environmental characteristics (e.g., crowding, security level).

Findings from a study conducted by Camp et al. (2003) are particularly relevant to our focus. Examining data on inmates housed in the Federal Bureau of Prisons, Camp et al. observed that the statistical significance of the same predictors varied across models of different types of rule violations. Further inspection of Camp et al.’s tables, however, revealed very similar coefficient estimates for any one predictor across the models examined (although equality of coefficients tests could not be performed with the information presented). The similarity across coefficient estimates suggests that pooling different types of misconduct might be useful from the standpoint of data reduction; however, the underlying assumption of this approach is that different types of offenses can be treated the “same” in empirical studies. From a practical standpoint, this may not be a credible assumption from the perspective of facility administrators, because they are likely to perceive very real differences in threats to safety between offenses such as assault with a deadly weapon versus possession of contraband. Still, it is possible that less serious offenses are highly correlated with more serious offenses in a prison environment, and inmates who are more likely to engage in the former are also at higher risk of engaging in the latter. Before moving in either direction, however, it is important to examine whether the differences across predictors vary significantly across different types of misconduct. Scholars have made similar arguments regarding gender differences in offending (e.g., Smith & Paternoster, 1987) and offense specialization in criminal career research (e.g., Britt, 1996), underscoring the need to focus on differences in the magnitude of empirical relationships across models of different outcomes rather than differences in the statistical significance of these relationships.

Comparing results from models predicting different types of offenses could aid in model specification by shedding considerable light on the degree of unique information that is offered by analyses of different outcomes. Regarding different measures of misconduct, a review of extant studies reveals similar groupings to those commonly examined in the broader criminological research including violence, nonviolent offenses aside from drugs and alcohol, drug/alcohol offenses, and pooled measures of all misconduct. Based on the unique population examined, researchers of inmate misconduct have typically narrowed their focus on violence to physical assaults and/or threats and have (sometimes) separated inmate-on-inmate from inmate-on-staff assaults/threats. Also potentially relevant to a study of prison inmates is a focus on security threats.

Predictors of Misconduct

The purposes of this study were (a) to assess whether predictors of inmate misconduct differed significantly across different types of misconduct and, if so, (b) to determine which types of misconduct merit separate examination in related studies. To enhance the reliability of the analyses, we considered a large number of possible predictors for model specification. These predictors were identified based on prior research as well as from the broader literature on inmate adaptation to incarceration. Most measures were excluded because they offered no improvement in the prediction of misconduct and did not improve coefficient estimates for the predictors included in the models described here. To maintain consistency across time periods, we also excluded measures that were not common to both data sets. When multiple measures tapping the same concept were available in the data sets, zero-order correlations between each measure and the outcome measures were used to determine the strongest predictor of misconduct. Finally, we do not believe that the measures examined here are the only factors that influence inmate misconduct, but they are the predictors we and other reviews have identified to be consistently included in studies of adaptation/misconduct (e.g., Bottoms, 1999; Goodstein & Wright, 1989).¹ The following discussion highlights the significant predictors that have emerged from many of these studies. In light of the study objectives, we do not link the predictors considered here to a particular theoretical perspective per se. Suffice it to say that most of these predictors could be framed within one or more of the existing theories of inmate behavior (e.g., importation, deprivation).

Background variables that have emerged as significant predictors of misconduct include age, race or ethnicity, criminal history, history of drug abuse, and type of offense leading to incarceration (for discussions of one or more of these variables, see Bottoms, 1999; Camp et al., 2003; Cao, Zhao, & Van Dine, 1997; Cunningham & Sorensen, 2007; Gaes et al., 2002; Goetting & Howsen, 1986; Harer & Steffensmeier, 1996; Jiang & Winfree, 2006; Toch et al., 1989; Wooldredge, 1994; Wooldredge et al., 2001).

Situational and lifestyle factors have also been found to be relevant to the subject. Variables such as sentence length and participation in programs focused on counseling, education, and/or vocational training have been demonstrated to improve prediction of misconduct and/or victimization during incarceration (Cao et al., 1997; Cunningham & Sorensen, 2007; Thomas, 1977; Wooldredge, 1994).

Differences between facilities often coincide with differences in a facility’s level of misconduct. An especially relevant measure across studies, possibly because it proxies so many other environmental characteristics (e.g., level of supervision, average custody score of the inmate population), has been a facility’s security level. Extant studies have revealed that more secure facilities are associated with higher levels of misconduct (e.g., Huebner, 2003; Jiang & Winfree, 2006; McCorkle et al., 1995).

Methods

The target population for the study included all males incarcerated in state-operated correctional facilities for men in the United States (excluding community-based facilities and facilities housing females). Female inmates were excluded from the analysis due to unmeasured structural and cultural differences that exist between many of the state-operated facilities for men versus facilities that house women that might also correlate with misconduct.

Samples and Data

Samples for the analysis were obtained from two waves of the survey of state inmates conducted by the U.S. Census Bureau (1997 Survey of Inmates in State and Federal Correctional Facilities and 1991 Survey of Inmates in State Correctional Facilities). The surveys offer nationally representative data on inmates’ offenses, sentences, criminal histories, backgrounds, substance abuse, and daily routines while in prison. To enhance the reliability of the results, the samples were examined separately, instead of combining the two groups, facilitating a comparison of any differences in the findings from the analyses of the 1991 data versus those derived from the 1997 data.

The Census Bureau used similar sampling designs for each time period. Facilities for men were selected from a sampling frame of all state-operated correctional facilities housing men. The largest facilities were selected with certainty, and the remaining facilities were separated into substrata defined by region of the country and facility type (confinement vs. community-based). The facilities were then grouped by security level and population size within each substratum to ensure adequate representation of these characteristics. Systematic sampling was used after calculating probabilities proportionate to size to determine the sampling interval within each substratum.

In the second stage of sample selection, inmates were selected randomly from a list provided by each facility that included all persons who occupied a bed the previous night. The total number of inmates selected for interviews in each facility was based on facility size. These procedures yielded roughly 11,000 inmates from 226 facilities in 1991 and about 11,500 inmates from 223 facilities in 1997.

In addition to our exclusion of inmates from community-based facilities, facilities housing females, and federal facilities (a unique addition to the Census Bureau’s sample in 1997), we also eliminated inmates if they were not sentenced (N₁₉₉₁ = 123, N₁₉₉₇ = 136) or if they had missing data on sentence length, time served, committing offense, or number of prior arrests (N₁₉₉₁ = 258, N₁₉₉₇ = 388). Removing these cases reduced the 1991 sample to 9,680 male inmates within 205 facilities across 41 states and the 1997 sample to 9,663 male inmates within 203 confinement facilities across 39 states.

Although missing data were only a minor problem in both data sets, we compared descriptive statistics for all measures without missing data between the reduced samples and the larger samples. These descriptives did not differ significantly between the two groups during either period, based on t tests conducted for each measure described in Table 1 and in the next section (p < .05).

Table 1.

Sample Means and Standard Deviations for the Inmate and Facility Measures.

	1991		1997
Measures	M	SD	M	SD
Level 1: Inmates
Outcome measures
All misconduct	.44	.50	.54	.50
Staff assault	.04	.20	.04	.20
Inmate assault	.14	.34	.15	.35
Assault	.16	.36	.16	.37
Drug/alcohol misconduct	.08	.27	.08	.28
Property misconduct	.08	.27	.10	.29
Security misconduct	.16	.37	.18	.38
Other misconduct	.28	.45	.35	.48
Nonviolent misconduct	.38	.49	.47	.50
Predictor variables
Age	31.86	9.52	33.29	9.78
African American non-Latino	.47	.50	.47	.50
Latino	.16	.37	.17	.38
Violent offense	.49	.50	.49	.50
Drug offense	.20	.40	.20	.40
Sentence length (months)	60.01	174.79	203.53	234.50
Number of prior arrests	4.61	4.62	4.15	3.81
Used drugs month before arrest	.50	.50	.56	.50
Program participation	2.17	1.43	1.86	1.35
Time served (months)	34.44	42.86	51.84	53.65
N₁	9,680		9,663
Level 2: Facilities
Maximum security	.38	.49	.39	.49
Minimum security	.14	.34	.13	.34
N₂	205		203

Note. All measures dummy coded except age, sentence length, time served, program participation, and number of prior arrests.

Measures

All of the measures included in the analyses are described in Table 1. A total of nine outcome measures were examined for each time period. Each measure was created from a series of survey questions asking inmates whether they had been written up or formally charged with committing particular crimes or rule violations. Due to the wording of the questions (i.e., the questions only asked about behaviors inmates were written up or formally charged for), the measures may be more similar to indicators of official misconduct as opposed to inmates’ self-reported behaviors. Limitations of self-report data include systematic errors resulting from poor recall and/or underreporting by certain groups of respondents (e.g., Hindelang, Hirschi, & Weis, 1981). However, official reports of misconduct by inmates have also been criticized in that they may underestimate actual deviance within an institution (see, e.g., Hewitt, Poole, & Regoli, 1984). Official indicators of misconduct have also been criticized because prisons and states may differ in how they define and enforce misconducts (Reisig, 2002). Although the extent of this problem is unknown, variation between facilities and states in how misconducts are classified could make meaningful comparisons across prisons problematic. By restricting the analyses to examining differences between inmates within facilities and differences across facilities within states, the analyses that were performed here, and are subsequently described, should reduce the likelihood of bias resulting from differences in how facilities classify these two types of violent misconduct. Nonetheless and even though both self-report and official measures of misconduct have been considered valid indicators of adjustment (Van Voorhis, 1994), the limitations of the outcome measures examined here should be kept in mind when interpreting the findings.

Researchers of inmate deviance have generally focused on either the prevalence (e.g., Griffin & Hepburn, 2006; Harer & Steffensmeier, 1996; Wooldredge et al., 2001) or incidence (Huebner, 2003; Jiang & Winfree, 2006) of misconduct. Although incidence measures are potentially important for tapping into different aspects of deviance and/or criminality (Blumstein, Cohen, & Nagin, 1978), we focused on the prevalence of misconduct due to the analysis of self-reported officially detected incidents. The magnitude of undetected misconduct among inmates is unknown (Hewitt et al., 1984), and so the incidence of self-reported misconduct that was officially detected may not adequately reflect the inmate’s level of involvement in related behaviors. Prevalence measures of official misconduct are not without their limitations, but they are inherently less biased than incidence measures because inmates who engage in rule infractions are more likely to be accurately categorized as such, assuming a study period of reasonable length.² More unique to the analyses performed here, this study focused on determining whether the effects of potential predictors of misconduct differed across different types of misconduct. Because the data used in this study are cross-sectional, prevalence measures are just as appropriate as incidence measures for answering the primary research question posed in this study.

The measure all misconduct tapped whether an inmate engaged in any rule violation. Also examined were three measures of physical assaults with or without weapons, including staff assaults, inmate assaults, and a combined measure of assaults on staff or inmates. The measure drug/alcohol misconduct captured the possession, sale, or use of either drugs or alcohol. Property misconduct was defined as possession of stolen property or an unauthorized item (excluding weapons and drugs). Following Camp et al. (2003), security misconduct encompassed escape, possession of a weapon, or verbally assaulting a staff member or another inmate. All other less serious facility rule violations were categorized under other misconduct (Camp et al., 2003). Finally, nonviolent misconduct combines the offenses captured in the three measures of security, property, and other misconduct.³

As previously discussed, predictors were selected for the models based on previous research findings. We included the number of prior arrests as an indicator of the inmates’ criminal history. An examination of the magnitude of the zero-order relationships between several different indicators of criminal history (e.g., prior incarceration, number of prior incarcerations) and the outcomes examined here revealed that the number of prior arrests maintained the strongest effect on most of the types of misconduct. Furthermore, the inclusion of other measures of criminal history in the same model as number of prior arrests did not substantively improve prediction of any of the types of misconduct.⁴

Other predictors of misconduct included demographic information regarding an inmate’s age and whether an inmate was African American non-Latino or Latino. Caucasian Anglos, along with other race/ethnic groups (<3 percent in both data sets), were left out of the models as the reference category.⁵ Also included were whether an inmate was incarcerated for a violent offense, whether an inmate was incarcerated for a drug offense, whether an inmate indicated he had used drugs in the month before arrest, a composite index of program participation (which was a count of the number of programs an inmate had participated in since admission, including vocational training, education aside from vocational training, counseling, work assignments, and/or substance abuse treatment), and an inmate’s sentence length (in months).⁶ The potential drawback to the index of program participation is that it could obscure the unique effect of any one type of program. We did not distinguish between program types, however, because of substantial variation in the quality of these programs across facilities (see French & Gendreau, 2006).⁷ At the facility level, dichotomous measures of security level were also included tapping whether the facility was classified as a maximum security or a minimum security institution. The designation of medium security was left out as the reference category.

A limitation of the survey data is that the outcome measures and some of the predictors were derived from questions that did not refer to behaviors occurring within fixed periods of time. These questions were preceded by the phrase, “Since your admission, have you . . .,” which would increase the likelihood of misconduct as well as program participation among inmates who have been incarcerated for longer periods of time. To help adjust for this limitation, we included the measure time served as a statistical control variable.⁸

Statistical Analysis

All steps in the analysis involved multilevel modeling techniques due to the need to recognize the hierarchical structure of the data (inmates nested within facilities and facilities nested within states). Therefore, three-level data sets were created for each time period with inmates at Level 1, facilities at Level 2, and states at Level 3.⁹ Creating the trilevel data files allowed us to (a) adjust for correlated error among inmates “nested” within the same facility as well as for correlated error among facilities within the same state, (b) base the hypothesis tests on the appropriate sample sizes (for inmates versus facilities), (c) control for compositional differences in inmate samples across facilities in the same state (providing more accurate assessments of security level effects on misconduct), and (d) control for unmeasured state-level differences that could affect facility misconduct rates across states (possibly due to differences in budgets, classification procedures, management practices, crime rates, etc.). Consistent with objectives (c) and (d), the inmate-level measures were centered on their means for each facility, and the facility measures of security level were centered on their means for each state. This approach controlled for compositional differences in inmate samples across facilities in the same state that might account for differences in misconduct rates across facilities, and it also limited variation in misconduct rates to within-state differences only (Raudenbush & Bryk, 2002). More specific to our focus, the technique allowed us to adjust for differences in the opportunities for misconduct and reporting of misconduct across facilities and across state systems. Although we created a three-level data set to address all of these issues, it is important to understand that the models displayed here are technically two-level models because they only include measures at the first two levels of analysis (inmates and prisons). This is why findings are displayed for only the first two levels of analysis.

The dichotomous outcome measures were examined with hierarchical Bernoulli regression using the correction for the overdispersion of outcome variances, available in HLM 6.0 (see Table 1 for the mean and standard deviation of each outcome). Due to the limited numbers of inmates within facilities, and of facilities within states, we could not examine full random effects models of either the inmate-level measures across facilities or the facility-level measures across states (i.e., these effects were “fixed” across all aggregates). The Level 1 intercepts were allowed to vary randomly across facilities, however, permitting an analysis of the main effects of a facility’s security level on the aggregate outcome at Level 2. Each Level 1 intercept reflects an “adjusted” mean of the prevalence of misconduct for all inmates within a particular facility, controlling for all other predictors in a model. Analysis of variance tests revealed significant variation in each of the Level 2 outcome measures across facilities during both time periods, introducing the possibility that differences in a facility’s security level might account for some of the between-facility differences in the level of misconduct.

To determine which types of misconduct (if any) should be examined separately, we tested for significant differences in the effects of each Level 1 predictor across all nine models of misconduct. Using an extreme example, if none of the regression coefficients for any one predictor were significantly larger or smaller across all nine models, this would lead to the conclusion that any one of the nine outcome measures could be examined as a proxy for the other eight outcomes. This stage of the analysis involved the use of an equality of coefficients test introduced by Clogg, Petkova, and Haritou (1995).¹⁰ Although other methods could have been used to compare differences in the estimates across models (for a review of several of these, see Mazerolle, Brame, Paternoster, Piquero, & Deane, 2000), the equality of coefficients test is well suited to the research questions of this study because it allows comparisons of the maximum likelihood coefficients examined here (Brame, Paternoster, Mazerolle, & Piquero, 1998; Paternoster, Brame, Mazerolle, & Piquero, 1998). The test determines whether two unequal coefficients should be treated as significantly different in value, as opposed to treating any raw difference as meaningful even though the difference could be attributable to sampling error. After delineating the “unique” outcome measures, based on the differences in effects across the nine models, we then compared the estimates derived from the 1991 models to those from the 1997 models. For these analyses, we used the same equality of coefficients test.

Results and Discussion

Tables 2 and 3 present the findings from the analyses of the 1991 and 1997 data, respectively. Several predictors were related to each type of misconduct including an inmate’s age, prior incarceration, prearrest drug use, and whether a facility was maximum security. (Time served was also consistently significant, although it was treated solely as a necessary statistical control based on the absence of a fixed time period for misconduct.) All remaining measures were only associated with particular outcome measures. Also worth noting is the fact that the directions of the relationships between some of the predictors and misconduct depended on the type of misconduct examined (e.g., African American non-Hispanic). These latter two findings seem to support Camp et al.’s (2003) conclusion regarding a more specific focus on inmate misconduct. Before making this claim, however, we needed to evaluate the magnitude of the differences between the coefficients generated for the various types of misconduct.

Table 2.

Maximum Likelihood Coefficients (With Standard Errors) for Inmate and Facility Effects on Misconduct—1991.

	All	Staff Assault	Inmate Assault	Assault	Drug/Alcohol	Property	Security	Other	Nonviolent
Level 1: Inmates
Intercept	−0.19	−3.64	−2.19	−2.03	−2.62	−2.54	−1.87	−1.00	−.49
Age	−.06***	−.08***	−.09***	−.09***	−.05***	−.05***	−.08***	−.05***	−.06***
	(.004)	(.01)	(.01)	(.01)	(.01)	(.004)	(.004)	(.004)	(.003)
African American non-Latino	.05	.14	.25***	.28***	−.61***	−.47***	−.05	−.02	.004
	(.04)	(.08)	(.07)	(.06)	(.10)	(.10)	(.08)	(.05)	(.05)
Latino	−.10	−.29	.13	.13	−.07	−.18	−.11	−.23***	−.17
	(.07)	(.19)	(.11)	(.10)	(.11)	(.09)	(.11)	(.06)	(.07)
Violent offense	.02	.13	.28***	.24**	.16	.005	.13	.01	.001
	(.07)	(.14)	(.07)	(.07)	(.08)	(.08)	(.08)	(.06)	(.06)
Drug offense	−.20***	−.26	−.17	−.20**	−.06	−.27**	−.11	−.19***	−.24***
	(.05)	(.14)	(.07)	(.07)	(.12)	(.10)	(.08)	(.05)	(.05)
Sentence length (months)	−.0001	−.001	−.001**	−.001***	.000003	−.00002	−.0002	−.0002	−.0001
	(.0002)	(.0003)	(.0002)	(.0002)	(.0003)	(.0002)	(.0002)	(.0001)	(.0002)
Number of prior arrests	.05***	.05***	.06***	.06***	.05***	.05***	.07***	.04***	.05***
	(.005)	(.01)	(.005)	(.005)	(.01)	(.01)	(.01)	(.01)	(.01)
Used drugs month before arrest	.30***	.33***	.34***	.34***	.76***	.40***	.32***	.30***	.29***
	(.05)	(.08)	(.06)	(.06)	(.08)	(.09)	(.07)	(.05)	(.04)
Program participation	.10***	−.09	.02	.01	.14***	.10***	.03	.10***	.09***
	(.02)	(.05)	(.02)	(.03)	(.02)	(.02)	(.03)	(.02)	(.02)
Time served (months)	.02***	.02***	.02***	.02***	.02***	.01***	.02***	.01***	.01***
	(.001)	(.002)	(.001)	(.001)	(.001)	(.001)	(.001)	(.001)	(.001)
Level 2: Facilities
Maximum security	.77***	1.50***	.97***	1.07***	.97***	.62***	.97***	.50***	.69***
	(.13)	(.13)	(.12)	(.10)	(.10)	(.11)	(.14)	(.11)	(.13)
Minimum security	−.34*	−1.31***	−.75***	−.82***	−.25	−.45**	−.69***	−.11	−.30
	(.14)	(.20)	(.15)	(.15)	(.17)	(.16)	(.19)	(.18)	(.16)

Note. Level 1 random intercepts as outcomes at Level 2; all other effects “fixed” across facilities and states.

p < .05 (Level 2 only). **p < .01 (both levels). ***p < .001 (both levels).

Table 3.

Maximum Likelihood Coefficients (With Standard Errors) for Inmate and Facility Effects on Misconduct—1997.

	All	Staff Assault	Inmate Assault	Assault	Drug/Alcohol	Property	Security	Other	Nonviolent
Level 1: Inmates
Intercept	0.32	−3.64	−2.08	−1.93	−2.44	−2.33	−1.66	−0.57	−0.05
Age	−.06***	−.08***	−.08***	−.08***	−.04***	−.04***	−.07***	−.05***	−.05***
	(.004)	(.01)	(.01)	(.005)	(.004)	(.005)	(.004)	(.004)	(.004)
African American non-Latino	.07	.47**	.03	.13	−.42***	−.63***	.17	.03	.04
	(.05)	(.18)	(.07)	(.07)	(.08)	(.13)	(.07)	(.06)	(.06)
Latino	−.13**	−.05	.11	.07	−.06	−.33	−.05	−.09	−.15***
	(.04)	(.17)	(.10)	(.10)	(.08)	(.15)	(.08)	(.05)	(.03)
Violent offense	.06	.04	.08	.08	.10	−.02	.06	.03	.04
	(.06)	(.08)	(.10)	(.08)	(.08)	(.09)	(.06)	(.05)	(.05)
Drug offense	−.13	−.56***	−.25**	−.27***	−.10	−.07	−.26**	−.09	−.08
	(.06)	(.16)	(.08)	(.08)	(.10)	(.11)	(.09)	(.06)	(.07)
Sentence length (months)	−.0001	−.001***	−.0002	−.0003	−.0002	.00004	−.0002	−.0001	.00002
	(.0001)	(.0002)	(.0001)	(.0001)	(.0002)	(.0001)	(.0001)	(.0001)	(.0001)
Number of prior arrests	.05***	.07***	.05***	.05***	.06***	.03***	.05***	.04***	.04***
	(.01)	(.01)	(.01)	(.01)	(.01)	(.01)	(.01)	(.01)	(.05)
Used drugs month before arrest	.26***	.29***	.34***	.32***	.85***	.38***	.31***	.18***	.20***
	(.03)	(.07)	(.06)	(.05)	(.09)	(.07)	(.05)	(.04)	(.03)
Program participation	.16***	−.08	.08***	.05	.15***	.06	.01	.14***	.14***
	(.03)	(.03)	(.02)	(.02)	(.02)	(.03)	(.03)	(.03)	(.03)
Time served (months)	.01***	.01***	.01***	.01***	.01***	.01***	.01***	.01***	.01***
	(.001)	(.001)	(.001)	(.001)	(.001)	(.001)	(.001)	(.001)	(.001)
Level 2: Facilities
Maximum security	.77***	1.35***	.93***	1.01***	.57***	.80***	.62***	.31**	.63***
	(.16)	(.12)	(.13)	(.12)	(.11)	(.09)	(.11)	(.11)	(.13)
Minimum security	−.58***	−1.80***	−.84***	−.95***	−.83***	−.41	−.56**	−.29*	−.50***
	(.14)	(.27)	(.19)	(.19)	(.23)	(.20)	(.18)	(.11)	(.14)

Note. Level 1 random intercepts as outcomes at Level 2; all other effects “fixed” across facilities and states.

p < .05 (Level 2 only). **p < .01 (both levels). ***p < .001 (both levels).

Comparing the coefficients across the different models of misconduct involved 860 separate z tests (430 per study period), and so they are not displayed in Tables 2 and 3 (although the statistically significant tests for the “unique” models are subsequently presented).¹¹ Here we discuss the key themes that emerged from these tests as they relate to targeting the most relevant outcome measures for analyses of inmate misconduct. Overall, the equality of coefficients tests confirmed Camp et al.’s (2003) argument favoring the analysis of more specific, rather than more general, measures of misconduct. Several of the models predicting more specific forms of misconduct produced parameter estimates for many of the predictors that were significantly different (p < .01) from the estimates for corresponding predictors from the model of all misconduct. (The criterion of p < .01 was used based on the large inmate samples examined, consistent with the hypothesis tests for each predictor displayed in Tables 2 and 3). The largest differences in magnitude emerged from the comparison of all misconduct with staff assault, inmate assault, and staff/inmate assault (i.e., all three assault models).

By contrast, when comparing differences across the three assault models, no significant differences emerged for 1991; and for 1997, only the coefficients for sentence length and program participation differed significantly between the models of staff assaults versus staff/inmate assaults combined. Yet when comparing each of the assault models with all other models, each model comparison produced at least three (more often four or five) significant differences in coefficients during each study period. Based on these findings, we concluded that physical assaults should be examined separately from other types of misconduct, although it does not matter whether physical assaults are measured as assaults on staff only, inmates only, or both staff and inmates combined.

Similarly, several estimates from the models of drug/alcohol offenses differed significantly from estimates for the corresponding predictors in the remaining models (property offenses, security threats, other offenses, and nonviolent offenses). In each model comparison, at least three coefficients (more commonly four or five) differed significantly across the 1991 models. The same was found for 1997, except that one particular comparison produced only two significantly different coefficients.

Based on the above observations, it appears that assaults (on inmates and/or staff) as well as drug/alcohol offenses should be examined separately. When comparing the effects on assaults to those on drug/alcohol offenses for 1991 (Table 2), note that an inmate’s race maintains a positive effect on the prevalence of assaults versus a negative effect on the prevalence of drug/alcohol offenses. Offense incarcerated for (whether violent or drug) matters for predicting assaults, but not for predicting drug/alcohol offenses. The same also applies to sentence length and whether a facility was minimum security. Program participation, on the other hand, maintained a significant and positive relationship with the odds of drug/alcohol offenses but not with the odds of engaging in assaults. In short, half of the predictors maintain different effects on the prevalence of each form of misconduct.

Of the remaining comparisons for both 1991 and 1997, the models predicting property, security, “other” misconduct, and nonviolent misconduct revealed similar effects across the four models, with no more than two significantly different effects between any two models (more typically one significantly different effect between models). From these results, we surmised that any one of the four models could be examined in place of the other three with only a minor loss in the “unique” information that would be obtained by examining all four models. Given that the category of nonviolent misconduct captures the offenses included in these other three outcomes, we chose nonviolent misconduct as one of the “independent” models for the remainder of the analysis.

To demonstrate the unique effects across the three models identified above, Tables 4 and 5 display the coefficient estimates for each model as well as the equality of coefficients tests for each pair-by-pair comparison, for 1991 and 1997 (respectively).

Table 4.

Comparing Effects Across Models of Assaults, Drug/Alcohol Offenses, and Other Nonviolent Offenses—1991.

	Maximum Likelihood Coefficients			z Tests
	Assault	Drug/Alcohol	Other Nonviolent	Assault vs. Drug/Alcohol	Assault vs. Other Nonviolent	Drug/Alcohol vs. Other Nonviolent
Level 1: Inmates
Age	−.09***	−.05***	−.06***	−2.82**	−2.79**
African American non-Latino	.28***	−.61***	.004	7.63***	3.53***	−5.49***
Latino	.13	−.07	−.17
Violent offense	.24**	.16	.001		2.59**
Drug offense	−.20**	−.06	−.24***
Sentence length (months)	−.001***	.000003	−.0001	−2.78**	−3.18***
Number of prior arrests	.06***	.05***	.05***
Used drugs month before arrest	.34***	.76***	.29***	−4.56***		5.83***
Program participation	.01	.14***	.09***	−3.61***
Time served (months)	.02***	.02***	.01***		7.07***	7.07***
Level 2: Facilities
Maximum security	1.07***	.97***	.69***		2.15*
Minimum security	−.82***	−.25	−.30	−2.51*	−2.37*

Note. Blank entries for z tests indicate no significant differences between coefficients.

p < .05 (Level 2 only). **p < .01 (both levels). ***p < .001 (both levels).

Table 5.

Comparing Effects Across Models of Assaults, Drug/Alcohol Offenses, and Other Nonviolent Offenses—1997.

	Maximum Likelihood Coefficients			z Tests
	Assault	Drug/Alcohol	Other Nonviolent	Assault vs. Drug/Alcohol	Assault vs. Other Nonviolent	Drug/Alcohol vs. Other Nonviolent
Level 1: Inmates
Age	−.08***	−.04***	−.05***	−6.25***	−4.69***
African American non-Latino	.13	−.42***	.04	5.17***		−4.60***
Latino	.07	−.06	−.15***
Violent offense	.08	.10	.04
Drug offense	−.27***	−.10	−.08
Sentence length (months)	−.0003	−.0002	.00002
Number of prior arrests	.05***	.06***	.04***
Used drugs month before arrest	.32***	.85***	.20***	−5.15***		6.85***
Program participation	.05	.15***	.14***	−3.54***
Time served (months)	.01***	.01***	.01***
Level 2: Facilities
Maximum security	1.01***	.57***	.63***	2.70**	2.15*
Minimum security	−.95***	−.83***	−.50***

Note. Blank entries for z tests indicate no significant differences between coefficients.

p < .05 (Level 2 only). **p < .01 (both levels). ***p < .001 (both levels).

The results presented in Tables 4 and 5 underscore the types of differences noted above between the models of assaults versus drug/alcohol offenses. At least half of the Level 1 parameter estimates for assaults were different from the estimates produced for either drug/alcohol or nonviolent misconduct across the 1991 models. Additionally, three of the Level 1 estimates for drug/alcohol and nonviolent misconduct differed across the 1991 models. Although not as many of the coefficients differed between 1997 models, enough differences emerged that when considered with the results from the z tests not reported here, we were relatively confident in concluding that future studies of inmate misconduct should examine these forms of misconduct separately.

As a reliability check on our observations regarding unique outcome measures, we also conducted a principal components analysis of the original nine outcome measures for each time period. Each analysis produced three primary factors accounting for roughly 70% of the variance in all nine measures (the Kaiser-Meyer-Olkin measures of sampling adequacy were .61 and .59 for 1991 and 1997, respectively). An oblique rotation of the three factors revealed that the three unique outcome measures identified above loaded on separate factors. For both time periods, outcome measures with the highest loadings on the first factor included all misconduct, other misconduct, and nonviolent misconduct. The second factor included staff assault, inmate assault, and assault. The third factor included drug/alcohol misconduct and property misconduct. The loadings for security misconduct were almost evenly split between the first two factors during both study periods.¹²

Stability of Empirical Relationships Between 1991 and 1997

We also examined the extent to which findings differed between 1991 and 1997. Across all three models, only three of the equality of coefficients tests were statistically significant, and two of the three tests involved time served (our statistical control variable), whereas the third test involved sentence length. In all three cases, the effects were stronger in 1991 versus 1997.¹³ These differences might have reflected the nationwide trend toward mandatory and determinant sentencing schemes (Irwin & Austin, 2001), as more inmates in general would be serving longer sentences in 1997 than in 1991. An inspection of the descriptive statistics in Table 1 supports this finding, with equal or greater dispersion in the vast majority of inmate level predictors for 1997 compared to 1991. On the other hand, the differing effects of these two predictors might also be due to methodological differences across studies, as these differences may be too broad to be attributed to major shifts in sentencing practices over a 6-year period. Unfortunately, we could not determine which of these potential influences contributed to the differences with the data available here.

The display of information in Tables 4 and 5 helps to underscore several potentially important observations about particular effects that are not as easily seen in Tables 2 and 3. First, age was inversely related to all three types of misconduct, yet the magnitude of the age effect was stronger for assaults (for both 1991 and 1997). Second, African American non-Latino inmates were more likely to commit assaults in the 1991 model, yet less likely to commit drug/alcohol misconduct. These empirical relationships were similar in the 1997 model, and although the estimates for assaults did not reach statistical significance, the similarity of the coefficients between the two time points (not significantly different) afford greater confidence in our observation. This result is also consistent with one of Harer and Steffensmeier’s (1996) observations from their analysis of data from the Federal Bureau of Prisons.

Findings for the type of offense an inmate was incarcerated for lead to a third observation about the importance of including these types of measures for model specification. Incarceration for either violence or drugs was relevant for predicting the odds of assault, and in opposite directions. For 1991 only, incarceration for drugs was also related to other nonviolent offenses.

Our fourth observation relates to the null effects of sentence length across all of the models for both study periods except the 1991 model predicting assaults. Before dismissing this as a relevant predictor, however, it is important to weigh these null effects against the consistently significant effects of time served. The two variables were not collinear although they were correlated (r ≈ .25 for each period). Thus, with regard to measures tapping an inmate’s sentence, it may be more important to examine how long an inmate has served up to the point of study (perhaps tapping an inmate’s opportunity for misconduct) as opposed to his sentence length.

The findings for an inmate’s prior drug use and program participation are the focus of our next observation. Prior drug use was consistently significant across all of the models examined for both time periods, but its effect was strongest in the models predicting drug/alcohol offenses. Although program participation was also a significant predictor of drug/alcohol offenses and other nonviolent offenses for both periods (although it was not significant for predicting assaults during either period), the effect on the prevalence of drug/alcohol offenses was stronger relative to the effect on other nonviolent offenses. In short, prior drug use and program participation are important for these types of models in general, but they may be even more salient for understanding drug/alcohol offenses during incarceration.

Finally, the effects of a facility’s security level were relatively stable across the models. No matter what type of misconduct was examined, inmates who were in maximum-security facilities were more likely to engage in misconduct, and the magnitude of the coefficient for assaults was larger than the parameter estimate for nonviolent misconduct during both periods. Compatible with these effects, minimum-security facilities were associated with lower levels of each type of misconduct during 1997, although the measure was only a significant predictor of assaults for 1991. Nonetheless, the nonsignificant differences in the effects of both maximum security and minimum security across the two periods underscore the relevance of distinguishing facility security levels in future studies of the topic.

Conclusions

Although our primary focus in this study was to contribute to model specification by determining whether the effects of a number of routinely examined predictors of inmate deviance differed across specific types of inmate misconduct, the analyses that were conducted toward this end also provided important substantive information on the topic. These observations include (a) the stronger effects of an inmate’s age on assaults relative to any other form of misconduct, (b) the significant yet opposite effects of an inmate’s race on assaults versus drug/alcohol offenses, (c) the importance of including type of offense for which incarcerated to improve model specification, (d) the stronger effects of how much time an inmate has served on all forms of misconduct relative to the more commonly examined measure of sentence length, (e) the overall importance of an inmate’s drug use immediately prior to incarceration in conjunction with its strongest effect on drug/alcohol offenses during incarceration, and (f) the utility of including multiple measures tapping a facility’s security level in these models to separate the significant effects of maximum versus minimum versus “other” security facilities on all forms of misconduct.

Recapping the findings pertinent to our primary focus, the analyses described here generated useful observations that may help to make research on inmate misconduct more efficient, facilitate a synthesis of related findings, and inform correctional practice. Findings generally supported Camp et al.’s (2003) claim that a single predictor may have different effects on various forms of misconduct. Specifically, we found that models should be estimated separately for physical assaults, drug/alcohol offenses, and all other nonviolent offenses. Findings for physical assaults were virtually identical to those for assaults on inmates only and assaults on staff only; results for nonviolent offenses were redundant with property offenses, security offenses, “other” violations, and “all” violations; and the model of drug/alcohol offenses produced the most unique information when compared to all other models. Accordingly, we recommend examining these groups separately in future research. An analysis of all three outcomes also provides a fuller picture of inmate “crime” by encompassing personal, property, and drug crimes.

An important caveat to the study findings is that additional research using official as opposed to self-report data on inmate misconduct would be worthwhile to assess whether our findings are robust across different data sources. Even if these findings are robust across different data sources, scholars may still want to examine other types of misconduct separately from the three outcomes identified here due to interest in theoretically and/or policy-relevant predictors that we were unable to include in our models. Some of these predictors may, in turn, force reconsideration of our conclusions if these other measures differ in their effects across two outcomes that we defined as “similar.”

An additional limitation of the study worth reiterating is that the outcome measures and measure of program participation used here were not restricted to a fixed period of time, which increased the odds of misconduct for individuals who had served more time (via more opportunity for misconduct). Similarly, the positive effects of program participation on nonviolent misconduct may reflect an inmate’s need for counseling, and those who “need” counseling and education might be at higher risk for violating facility rules, at least in the early phases of their incarceration. Although we adjusted for this problem in part by controlling for time served, the limitations of the outcome measures and measures of program participation should be kept in mind when interpreting the findings.

The limitations noted above notwithstanding, the findings described here also have implications for theories of inmate deviance. Similar to findings from the broader criminological literature (e.g., Blumstein, Cohen, & Farrington, 1988; McGloin, Sullivan, Piquero, & Pratt, 2007; Sullivan, McGloin, Pratt, & Piquero, 2006), the results of this study suggest that some inmates who engage in misconduct may exhibit a degree of offending specialization. To substantiate this claim, however, more research (both cross-sectional and longitudinal) is needed to examine the effects of potential predictors across the three outcomes recommended here. These specifications will allow for a determination of whether inmates specialize in their institutional offending, whether such specialization persists, and whether different causal processes or factors account for why inmates commit particular types of offenses. If further support for offending specialization is observed, theories of inmate behavior will need to account for specialization.

Researchers have often selected possible predictors of inmate misconduct based on their compatibility with importation and/or deprivation theories of inmate adaptation to incarceration (e.g., Gaes & McGuire, 1985; Harer & Steffensmeier, 1996; Thomas, 1977). Some researchers have also selected predictors based on their links to, for example, opportunity theory or theories of informal social controls (e.g., Wooldredge, 1998; Wooldredge et al., 2001). Findings in support of offending specialization in prisons could make some of these perspectives more relevant than others. Importation theory, for example, suggests that inmates’ behavior within prisons is a manifestation of latent culture or preincarceration experiences (Irwin, 1980; Irwin & Cressey, 1962). It could be that distinct adaptations to imprisonment emerge from inmates who are pulled from different neighborhoods (more generally, see Cloward & Ohlin, 1960; Mazerolle et al., 2000). Inmates pulled from neighborhoods where violence is tolerated and expected as a part of everyday life may be more likely to engage in violence in prisons (see, e.g., Harer & Steffensmeier, 1996), whereas other neighborhoods could encourage drug use or property crime, increasing the likelihood that inmates from those areas might engage in related behaviors in correctional facilities.

Theories of social control or opportunity perspectives may also be relevant. Inmates’ local life circumstances or the factors that influence the strength of their bond to conventional society can affect the amount and types of criminal opportunities inmates are exposed to (more generally, see McGloin et al., 2007). Changes in opportunity structures may lead some individuals to specialize in their offending, if only for a short term such as a prison sentence (McGloin et al., 2007; Sullivan et al., 2006). Indeed, such a perspective seems reasonable in correctional institutions that contain a pool of previously motivated offenders (Wooldredge, 1998). For inmates, opportunities may be influenced by their daily routines in conjunction with environmental constraints placed upon their behaviors, as well as interactions between routines and environments. These and other possibilities may be worthy of consideration in future studies. As Goodstein and Wright (1989) noted, it is only by examining the interaction between the inmate and the environment in which he or she is confined that we can gain a better understanding of the processes that influence levels of misconduct.

Footnotes

Authors’ Note

This project was supported by a grant from the American Statistical Association and the Bureau of Justice Statistics as part of their Special Data Collections and Statistical Methodological Studies Research Program. The data sets examined for this article were made available by the Interuniversity Consortium for Political and Social Research (ICPSR). The data for both the 1991 Survey of Inmates in State Correctional Facilities (ICPRS 6068), the 1990 (ICPSR 9908), 1995 (ICPRS 6953) and 2000 (ICPSR 4021) Census of State and Federal Adult Correctional Facilities the 1997 Survey of Inmates in State and Federal Correctional Facilities (ICPSR 2598) were collected by the U.S. Census Bureau. Neither the U.S. Census Bureau nor the ICPSR bear any responsibility for the analyses presented here.

Notes

Author Biographies

Benjamin Steiner is an assistant professor in the School of Criminology and Criminal Justice at the University of Nebraska at Omaha. He holds a PhD from the University of Cincinnati.

John Wooldredge is a professor in the College of Criminal Justice at the University of Cincinnati. His research and publications focus on institutional corrections (crowding, inmate violence, inmate adaptation) and criminal case processing (sentencing and recidivism, extralegal disparities in case processing and outcomes). He is currently involved in research on interjudge variability in sentencing, short- versus long-term effects of sentencing guidelines on extralegal disparities in prison sentences, and official responses to prison inmate rule violations.

References

Blumstein

Cohen

Farrington

(1988). Criminal career research: Its value for criminology. Criminology, 26, 1-36.

Blumstein

Cohen

Nagin

(1978). Deterrence and incapacitation: Estimating the effects of criminal sanctions on crime. Report of the Panel on Research on Deterrent and Incapacitative Effects. Washington, DC: National Academy of Sciences, National Research Council.

Bottoms

(1999). Interpersonal violence and social order in prison. In Tonry

Petersilia

(Eds.), Crime and justice: A review of research (Vol. 26, pp. 205-282). Chicago: University of Chicago Press.

Brame

Paternoster

Mazerolle

Piquero

(1998). Testing for the equality of maximum-likelihood regression coefficients between two independent equations. Journal of Quantitative Criminology, 14, 245-261.

Britt

(1996). The measurement of specialization and escalation in the criminal career: An alternative modeling strategy. Journal of Quantitative Criminology, 12, 193-222.

Camp

Gaes

Langan

Saylor

(2003). The influence of prisons on inmate misconduct: A multilevel investigation. Justice Quarterly, 20, 501-533.

Cao

Zhao

Van Dine

(1997). Prison disciplinary tickets: A test of the deprivation and importation models. Journal of Criminal Justice, 25, 103-113.

Clogg

Petkova

Haritou

(1995). Statistical methods for comparing regression coefficients between models. American Journal of Sociology, 100, 1261-1293.

Cloward

Ohlin

(1960). Delinquency and opportunity: A theory of delinquent gangs. New York: Free Press.

10.

Cunningham

Sorensen

(2006). Nothing to lose? A comparative examination of prison misconduct rates among life-without-parole and other long-term high-security inmates. Criminal Justice and Behavior, 33, 683-705.

11.

Cunningham

Sorensen

(2007). Predictive factors for violent misconduct in close custody. The Prison Journal, 87, 241-253.

12.

DiIulio

(1987). Governing prisons: A comparative study of correctional management. New York: Free Press.

13.

French

Gendreau

(2006). Reducing prison misconducts: What works! Criminal Justice and Behavior, 33, 185-218.

14.

Gaes

McGuire

(1985). Prison violence: The contribution of crowding versus other determinants in prison assault rates. Journal of Research in Crime and Delinquency, 22, 41-65.

15.

Gaes

Wallace

Gilman

Klein-Saffran

Suppa

(2002). The influence of prison gang affiliation on violence and other prison misconduct. The Prison Journal, 82, 359-385.

16.

Gendreau

Goggin

Law

(1997). Predicting prison misconducts. Criminal Justice and Behavior, 24, 414-431.

17.

Goetting

Howsen

(1986). Correlates of prisoner misconduct. Journal of Quantitative Criminology, 2, 49-67.

18.

Goodstein

Wright

(1989). Inmate adjustment to prison. In Goodstein

MacKenzie

(Eds.), The American prison: Issues in research and policy (pp. 229-252). New York: Plenum.

19.

Griffin

Hepburn

(2006). The effect of gang affiliation on violent misconduct among inmates during the early years of confinement. Criminal Justice and Behavior, 33, 419-448.

20.

Harer

Langan

(2001). Gender differences in predictors of prison violence: Assessing the predictive validity of a risk classification system. Crime & Delinquency, 47, 513-536.

21.

Harer

Steffensmeier

(1996). Race and prison violence. Criminology, 34, 323-355.

22.

Hewitt

Poole

Regoli

(1984). Self-reported and observed rule breaking in prison: A look at disciplinary response. Justice Quarterly, 1, 437-447.

23.

Hindelang

Hirschi

Weis

(1981). Measuring delinquency. Beverly Hills, CA: Sage.

24.

Huebner

(2003). Administrative determinants of inmate violence: A multilevel analysis. Journal of Criminal Justice, 31, 107-117.

25.

Irwin

(1980). Prisons in turmoil. Boston: Little, Brown.

26.

Irwin

Austin

(2001). It’s about time: America’s imprisonment binge (3rd ed.). Belmont, CA: Wadsworth.

27.

Irwin

Cressey

D. R.

(1962). Thieves, convicts, and the inmate culture. Social Problems, 10, 142-155.

28.

Jiang

Winfree

L. T.

(2006). Social support, gender, and inmate adjustment to prison life: Insights from a national sample. The Prison Journal, 86, 32-55.

29.

Mazerolle

Brame

Paternoster

Piquero

Deane

(2000). Onset age, persistence, and offending versatility: Comparisons across gender. Criminology, 38, 1143-1172.

30.

McCorkle

Miethe

Drass

(1995). The roots of prison violence: A test of the deprivation, management, and “not-so-total” institution models. Crime & Delinquency, 41, 317-331.

31.

McGloin

J. M.

Sullivan

C. J.

Piquero

A. R.

Pratt

T. C.

(2007). Local life circumstances and offending specialization/versatility: Comparing opportunity and propensity models. Journal of Research in Crime and Delinquency, 44(3), 321-346.

32.

Paternoster

Brame

Mazerolle

Piquero

A. R.

(1998). Using the correct statistical test for the equality of regression coefficients. Criminology, 36, 859-866.

33.

Raudenbush

Bryk

(2002). Hierarchical linear models: Applications and data analysis methods (2nd ed.). Thousand Oaks, CA: Sage.

34.

Reisig

(2002). Administrative control and inmate homicide. Homicide Studies, 6, 84-103.

35.

Ruback

R. B.

Carr

(1993). Prison crowding over time: The relationship of density and changes in density to infraction rates. Criminal Justice and Behavior, 20, 130-148.

36.

Smith

Paternoster

(1987). The gender gap in theories of deviance: Issues and evidence. Journal of Research in Crime and Delinquency, 24, 140-172.

37.

Sorensen

Wrinkle

Gutierrez

(1998). Patterns of rule-violating behaviors and adjustment to incarceration among murderers. The Prison Journal, 78, 222-231.

38.

Steiner

Wooldredge

(2007). Understanding micro- and macro-level influences on inmate misconduct across state operated prisons and correctional facilities. American Statistical Association/Bureau of Justice Statistics.

39.

Sullivan

McGloin

Pratt

Piquero

(2006). Rethinking the “norm” of offender generality: Investigating specialization in the short-term. Criminology, 44, 199-234.

40.

Thomas

(1977). Theoretical perspectives on prisonization: A comparison of the importation and deprivation models. Journal of Criminal Law and Criminology, 68, 135-145.

41.

Toch

Adams

Grant

J. D.

(1989). Coping: Maladaptation in prisons. New Brunswick, NJ: Transaction Publishers.

42.

Van Voorhis

(1994). Measuring prison disciplinary problems: A multiple indicators approach to understanding prison adjustment. Justice Quarterly, 11, 679-710.

43.

Wooldredge

(1991). Correlates of deviant behavior among inmates of U.S. correctional facilities. Journal of Crime and Justice, 14, 1-25.

44.

Wooldredge

(1994). Inmate crime and victimization in a southwestern correctional facility. Journal of Criminal Justice, 22, 367-381.

45.

Wooldredge

(1998). Inmate lifestyles and opportunities for victimization. Journal of Research in Crime and Delinquency, 35, 480-502.

46.

Wooldredge

Griffin

Pratt

(2001). Considering hierarchical models for research on inmate behavior: Predicting misconduct with multilevel data. Justice Quarterly, 18, 203-231.