Absconding Among Juvenile Parolees in Arizona

Abstract

Absconding refers to the active or passive avoidance of contact with correctional supervisory agencies by offenders. Absconders are problematic because their whereabouts are unknown and their threat to the public is elevated. The aim of this study was to construct and validate an actuarial instrument designed to assess risk for absconding among juvenile parolees that accounts for gendered differences. The data were gathered from 1,063 juveniles released from the Arizona Department of Juvenile Corrections into community settings in 2008 and 2009. Juveniles were randomly subdivided into construction and validation samples to assess the validity of the instrument. Twelve risk factors were identified to construct the instrument, three of which were found to operate differently for male and female juveniles. Upon application to the validation sample, the instrument correctly classified 70% of juvenile parolees, and a corresponding r value of .37 was observed. The author discusses the substantive and practical implications for assessing absconding risk and modeling gender differences when supervising offenders in community settings.

Keywords

absconding risk assessment gender parole juvenile offenders

The placement of criminal and delinquent offenders under correctional supervision (e.g., probation, parole) can result in two possible outcomes: success or revocation. A large body of research has examined risk and protective factors associated with the occurrence of these two outcomes (Cottle, Lee, & Heilbrun, 2001; Gendreau, Little, & Goggin, 1996; Grattet, Petersilia, & Lin, 2008), resulting in the development of risk instruments that have been implemented in various correctional settings (Bonta, 2002; Latessa, Smith, Lemke, Makarios, & Lowenkamp, 2009). An additional outcome, however, has received less attention from the research community but is of equal importance to practitioners interested in alleviating problems associated with offenders under correctional supervision: absconding.

Absconding refers to the active or passive avoidance of contact with correctional supervisory agencies by offenders. Absconders, although technically falling into the camp of revocation, differ from the other types of outcomes in that their whereabouts are unknown to supervising agencies, and their threat to the general public is elevated. Although all offenders under correctional supervision are at risk for committing new offenses against the public when not incarcerated, absconders pose a unique problem for these agencies. The difference for absconding offenders is that high-profile crimes can evoke public outrage should it be discovered that the supervising agency lost track of its subjects (Williams, McShane, & Dolny, 2000). For this reason, high-risk absconders, especially sex offenders, appear regularly on the “most wanted” bulletins of policing and correctional agencies throughout the United States. If these high-risk individuals could be identified prospectively, this information would hold considerable utility for correctional agencies in developing supervisory strategies. More broadly, this logic serves as the basis for risk assessment classification.

Two problems emerge from this discussion for absconding in particular and risk assessment in general. First, validated risk instruments for absconding are not present in the research literature. The application of existing instruments designed for other outcomes has demonstrated only modest success (see Lowenkamp, Holsinger, & Latessa, 2001). Furthermore, research in this area has been mostly descriptive, with few studies having examined absconding as an outcome in a multivariate context (Mayzer, Gray, & Maxwell, 2004; Williams et al., 2000). Second, the risk assessment literature in general tends to treat risk and protective factors equally for male and female offenders, despite a body of evidence that would suggest otherwise (Belknap & Holsinger, 2006; Holtfreter & Cupp, 2007; Holtfreter, Reisig, & Morash, 2004; Reisig, Holtfreter, & Morash, 2006; Salisbury, Van Voorhis, & Spiropoulos, 2009; Van Voohris, Wright, Salisbury, & Bauman, 2010). If risk instruments are not validated on female populations, this could introduce bias and imprecision into risk models, with the most serious implication being that male and female offenders could be misclassified.

This article details the construction and validation of the Abscond Risk Instrument (ARI), developed for the Arizona Department of Juvenile Corrections (ADJC). The ARI is an actuarial tool that was created using data from the population of juveniles (N = 1,063) paroled from ADJC facilities into the community during 2008 and 2009 and observed for a 1-year period after release. The population was subdivided randomly into construction (n = 525) and validation (n = 538) samples. Static and dynamic factors drawn from various risk domains—demographic, familial, social, psychological, and criminal justice—were used to construct the instrument on the former sample. Multiplicative interactions were incorporated to account for gender-specific processes. The final model was then implemented on the latter sample to evaluate the validity of the risk instrument.

The Extent and Nature of Absconding

Incidents in New Jersey and California underscore the concern that absconders pose to correctional agencies, in which paroled felons were suspected of murder after having avoided authorities (Blankstein, 2010; Megerian, 2010). Tragedies such as these are what prompted Williams et al. (2000) to refer to absconders as “political nightmares” (p. 24). Cases in which absconding offenders taunt authorities can be embarrassing for correctional agencies, especially if the stories are picked up by the news media (O’Donnell & Jewkes, 2011). Although absconders pose practical difficulties for supervising agents, Mayzer et al. (2004) also pointed out that absconders “negate sentences, avoid punishment, and deny any sense of justice” (p. 138). Traditionally, very little can be done to curb absconding behavior, although studies have shown electronic monitoring and specialized probation teams to be successful (Padgett, Bales, & Blomberg, 2006; Taxman & Byrne, 1994).

A number of studies have reported on the prevalence of absconding in correctional supervisory settings, comparing rates of absconding to rates of success and revocation. A study of nearly 76,000 offenders in Florida placed on home confinement reported that 43% of the offenders completed successfully, 41% experienced revocation for technical violations or new offenses, and 16% absconded from supervision (Padgett et al., 2006). A probation study in Michigan revealed that 53% of the probationers completed successfully, 39% were revoked, and 8% absconded (Mayzer et al., 2004). Among a sample of 442 felony offenders sentenced to programming in lieu of prison time, 20% reportedly absconded prior to termination of the sentence (Lowenkamp et al., 2001). Finally, a study of offenders released on parole in California found that 21% absconded within 1 year (Williams et al., 2000). More recent figures from California indicated that 17% of the parolee population, or over 20,000 parolees, were considered as having absconded (Grattet et al., 2008). Rates of absconding, however, often depend on various factors, including the correctional supervisory context (probation vs. parole), geographical context (rural vs. urban), jurisdictional policy (tolerance of technical violations), and political and economic climate (budget issues). Nevertheless, these figures suggest that absconders constitute a sizable proportion of the population that correctional agencies are accountable for supervising.

The nature of absconding behavior has been captured over the course of in-depth interviews with absconders (Hanrahan, Gibbs, & Zimmerman, 2005; Schwaner, 1997; Schwaner, McGaughey, & Tewksbury, 1998; Wolfe & Vivian, 2008). Supervisory conditions present offenders with a stringent set of rules that restrict freedom and activities. Many offenders were at one time dependent on alcohol or drugs prior to incarceration and are presented with ubiquitous opportunities for use upon their return to the street. Absconding offenders, especially juveniles, are often found in unsupportive households, in which strains such as physical and emotional abuse are not uncommon. Weakened ties to social institutions such as family, education, and employment reduce the capacity for informal social controls to influence offenders’ behavior positively. Moreover, absconders are commonly subject to negative peer influences, notably gang membership, and are embedded in criminal contexts. Many offenders recognize that their deviant behavior has come to the attention of authorities, leaving them with the option of running or returning to prison. The factors described above present offenders with inducements to “escape” from the restrictive conditions of supervision. As a result, absconding is less a product of careful, future-oriented decision making than it is of shortsightedness, situational influences, and poor impulse control. To the extent that the qualitative characteristics of absconding behavior can be accounted for empirically, these factors should be instrumental in assessing risk and identifying absconders prospectively.

Identifying Absconding Offenders

Most of the research examining supervised offending populations has focused on recidivism or success as outcomes. Williams et al. (2000) remarked that the absconding population “represent[s] such a large group of problem people, the virtually nonexistent amount of research is almost inexplicable” (p. 37). The body of research that has focused on absconding is limited because it has (a) a descriptive focus, therefore providing no formal analysis, or (b) omitted variable bias due to the bivariate nature of study designs. The studies of Williams et al. and Mayzer et al. (2004) are exceptions to the limitations mentioned above and have documented the correlates of absconding in a multivariate context.

Williams et al.’s (2000) study was based on a randomly drawn sample of 4,047 adult parolees in California in the late 1990s. Parolees were considered absconders if they were “absent” during the 1st year of parole. Drawing from data gathered by the California Parole Division, Williams et al. identified seven variables that were associated statistically with absconding. They found that unstable living environments, frequent unemployment, previous parole violations, low-stakes offenders, number of prior arrests, previous felonies, and single marital status increased the likelihood that offenders would abscond. Mayzer et al.’s (2004) study was based on a randomly drawn sample of 1,157 probationers in Michigan. They compared absconders with those who completed probation successfully and with those who experienced revocation over a 30-month period. They found that probation absconders, compared with “successes,” were more likely to be non-White, have lower supervision levels, inconsistent employment, and stable residency. They also found that probation absconders, compared with “revocations,” were more likely to be female, have fewer misdemeanors, lower supervision levels, and report fewer residential address changes. On the whole, the researchers concluded that absconders were more similar to those experiencing revocation than those successfully completing probation, with criminal history being the main characteristic distinguishing the two groups.

The key implication of this research is that there are factors capable of identifying absconders among correctionally supervised groups. Enough evidence exists indicating that absconders operate as a unique risk group among monitored populations, and the research literature accordingly treats criminal, absconding, and technical violators differently (Mayzer et al., 2004, p. 148; Steen & Opsal, 2007; Williams et al., 2000, p. 36). Risk assessment is based on the capability to discriminate empirically between qualitatively different groups. Risk is typically established through clinical or actuarial approaches. Clinical approaches involve assessments that are based on practitioner or professional subjective judgments or estimates based on interviews with subjects. Actuarial approaches involve assessments that are based on standardized static and dynamic background factors that are used to gauge risk. On the basis of reviews of the risk assessment literature, Bonta (2002) stated that the “weight of the evidence clearly favors actuarial assessments of risk” (p. 358; see also Andrews, Bonta, & Wormith, 2006; Van Voorhis & Brown, 1997). Actuarial assessments serve as the basis for risk instruments such as Level of Service Inventory–Revised (LSI-R; Andrews & Bonta, 1995), the Psychopathy Checklist–Revised (Hare, 1991), and the Violent Risk Appraisal Guide (Harris, Rice, & Quinsey, 1993). Lowenkamp et al. (2001) applied an established risk instrument, the LSI-R, to absconding, finding a modest effect size of .14. This modest effect should not come as a surprise, as the instrument was not designed to predict absconding as an outcome, a limitation the present study seeks to address.

The Current Study

The aim of this study was to develop and validate the ARI for juvenile parolees in Arizona. Absconding is a correctional supervisory outcome that is not well understood empirically. The general lack of research attention devoted to absconding is problematic, because absconders are an important risk group. Although some studies have demonstrated that absconders can be identified prospectively, validated risk instruments are not present in the research literature. Furthermore, a serious limitation observed among risk instruments in general is that male and female offenders are treated similarly, despite evidence suggesting that they differ. That is, certain factors are weighted equally across genders when creating risk scores when there is reason to believe they should not be, thus introducing the potential for bias and imprecision into the final model. This study explores these issues and incorporates multiplicative interactions into the risk instrument to account for gender-specific processes.

Methods

Data

The data were drawn from juvenile records extracted by the research and development staff at the ADJC. When juveniles are received from Arizona’s 15 committing counties, they go through a reception, assessment, and classification process. This process involves collecting admission information (e.g., medical records), diagnostic evaluations (e.g., psychological information), and secure care criteria that assist in developing an initial placement category recommendation. During this process, and within no longer than 14 days of admission, juveniles are administered the Criminogenic and Protective Factors Assessment (CAPFA). The CAPFA gathers information from 12 domains with respect to various familial, social, psychological, environmental, and attitudinal factors. Various ADJC personnel administer this assessment every 90 days after intake. The items outlined below were drawn from juveniles’ CAPFA at the time of release.

All juveniles released from ADJC facilities into community settings during calendar years 2008 and 2009 were included in the present study (N = 1,063). For the purpose of validating the risk instrument, it was necessary to split the subjects randomly into construction and validation samples. The construction sample was used to build the risk instrument and obtain weighted estimates, and the validation sample was used to test the predictive validity of the risk instrument. Actuarial assessment research suggests a sample size of at least 500 subjects when constructing a risk tool (Gottfredson & Snyder, 2005). Using a random sample generator in Stata 12.0 (StataCorp LP, College Station, TX), approximately 50% of the original cases were partitioned into a construction sample and the remaining cases into a validation sample. The final samples consisted of 525 (construction) and 538 (validation) juveniles released from secure facilities in community settings during calendar years 2008 and 2009.

Dependent Variable

Absconding is the outcome in this study. For a juvenile to be classified as absconded, a parole officer makes a determination on the basis of his or her contact with the juvenile. If the juvenile cannot be located at his or her parole placement residence (e.g., foster home, contracted provider, family residence), the parole officer then requests a warrant to be issued within 24 hours for juveniles in community residential program settings or 72 hours for juveniles in noncommunity residential program settings (i.e., family home). On the basis of the criteria submitted by the parole officer, a supervisor then approves or denies the warrant request, and the warrant proceeds along the appropriate legal channels. For the purposes of the current study, juveniles who were recorded as having absconded were coded 1 and juveniles who did not abscond were coded 0. The window for absconding was 1 year from the time of release.¹

Independent Variables

Drawing from questions in the CAPFA domains, 12 independent variables were included in the analysis on the basis of the procedures described in the analytic strategy section below.

Demographic

Age is scaled continuously in years and was recorded at the time of release from secure care. Female is a dichotomous variable coded 1 if the juvenile was female and 0 if the juvenile was male.

Familial

Ward is a dichotomous variable coded 1 if the juvenile was a ward of the state and 0 otherwise. Family incarceration is an ordered categorical variable that taps into a juvenile’s familial experience with incarceration. This variable is a count of the number of family members (e.g., mother, stepfather, brother, grandfather) with current and/or previous experiences with incarceration. Parenthood is a dichotomous variable determined via self-report or, in the case of female offenders, during a pregnancy test at intake. Juveniles were coded 1 if they were parents and 0 otherwise.

Social

Religion is a dichotomous variable coded 1 if the juvenile is involved in religious activities (e.g., attends church) and 0 if otherwise. Gang membership is a dichotomous variable determined by a gang intelligence officer at intake through official records or other information (e.g., self-proclamation, tattoos). Juveniles were coded 1 if they were gang members and 0 otherwise. Runaway is an interval variable based on the number of times a youth ran away or was ejected from his or her residence. The variable was collected via self-report and scaled from 0 to 8. The original coding was mean adjusted according to the coding scheme (e.g., 1 to 3 events = 2 events) to approximate a meaningful distribution.

Psychological

Substance dependent is a dichotomous variable determined by a behavioral health specialist, whereby juveniles were coded 1 if they were considered substance dependent and 0 otherwise. Low self-control is composed of 11 items drawn from the Behavioral Assessment System for Children, Second Edition (Reynolds & Kamphaus, 2004). These items (e.g., self-regulation, concern for others, future orientation) tap into the construct of low self-control put forth by Gottfredson and Hirschi (1990). The items were summed to create a composite score with a possible range of values from 0 to 44, with higher scores indicating lower levels of self-control and lower scores indicating higher levels of self-control. The scale exhibited a distribution that approximated normality and maintained acceptable psychometric properties (Cronbach’s α = .91, mean interitem r = .45).

Criminal justice system

Sex offender is a dichotomous variable coded 1 if the juvenile was recorded as a sex offender and 0 otherwise. Community placement measures where the juvenile was placed after being released from secure care. Juveniles placed in community-based centers (e.g., foster homes, residential treatment) were coded 1, while juveniles placed in their families’ residences were treated as the reference category and coded 0.

Analysis

Many studies create risk assessments using forward or backward stepwise multivariate regression, allowing researchers to begin with all available variables. A model is then estimated using an algorithm that eliminates variables that do not make statistical contributions. The best-fitting model is eventually identified after a series of iterations. The main problem with the pure version of this approach is that variables that do not make statistical contributions, but may contribute substantively, are eliminated from the analysis (see Gottfredson & Snyder, 2005, p. 31). For the present study’s purpose, because the subjects consist of a population (i.e., all Arizona juvenile parolees), statistical significance is less of a concern than the magnitude of the effect. As a result, a screening process was used to identify both empirically and substantively relevant variables as follows. First, bivariate mean differences were examined between absconding and nonabsconding groups, within and across genders. As a result the likelihood of rejecting the null hypothesis was so high because of the sample size, mean differences that were not statistically indistinguishable from zero were removed from the analysis. Second, bivariate logistic models partitioned by gender were estimated to assess differences across the study variables (Clogg, Petkova, & Haritou, 1995). Coefficients that differed statistically were interacted multiplicatively with gender in the multivariate analysis. This approach permits accounting for both statistical and substantive contributions of the independent variables.

The results of this study are presented in two sections: construction and validation. In the first stage, the descriptive characteristics of the sample and bivariate relationships among the study variables are discussed. This portion of the study involved the construction of the instrument and detailed the incorporation of gender-based risk factors. Weighted estimates were drawn from the multivariate logistic regression model, and predicted probabilities were generated. The second stage of the analysis focused on the validation of the ARI. Weighted estimates were applied to the validation sample and converted into predicted probabilities. The distributional properties were then evaluated between samples. Predictive validity was assessed by evaluating sensitivity, specificity, and the overall correct rate of prediction in the validation, or blinded, sample. The concluding section of the analysis examined the effect sizes, or r values, across categories of risk classification associated with the ARI.

Results

Descriptive Statistics and Bivariate Relationships

Table 1 details the descriptive statistics for the study variables. The second column from the left displays the characteristics of the construction sample. The sample consisted mostly of male offenders, and very few of the juveniles were wards of the state of Arizona or were parents. On average, at least one family member had a history of incarceration. A large number of respondents participated in religious activities and were officially recorded as gang members. Leaving home, whether voluntary or involuntary, occurred frequently, with juveniles reporting a mean of three instances in which they ran away or were ejected from their residences. Approximately half of juvenile parolees were evaluated as being substance dependent, and 8% were recorded as sex offenders. Finally, when juveniles were released from secure care, 12% were placed in community-oriented facilities (e.g., group homes or treatment centers), and the remainder went to their original residences or the residences of family members.

Table 1:

Descriptive Statistics for the Study Variables in the Construction Sample

	Full Sample	Nonabscond	Abscond
Variable	M (SD) or %	M (SD) or %	M (SD) or %
Abscond	31	—	—
Age (years)	16.79 (0.89)	16.91 (0.89)**	16.54 (0.82)
Female	15	13**	21
Ward	6	3**	11
Family incarceration	0.88 (1.05)	0.85 (1.05)	0.93 (1.18)
Parenthood	6	6	7
Religion	32	34*	27
Gang membership	28	24**	39
Runaway	2.94 (3.19)	2.37 (2.96)**	4.17 (3.32)
Low self-control	18.77 (6.42)	17.96 (6.28)**	20.53 (6.39)
Substance dependent	52	50*	57
Sex offender	8	10**	4
Community placement	12	11	14
n	525	360	165

Note. Statistical significance was assessed using t tests for interval variables and chi-square tests for discrete variables.

p < .10. **p < .05.

Nearly one third of the sample, or 165 juveniles, absconded within 1 year of release from secure care. These juveniles differed across a variety of factors in comparison with juveniles who did not abscond. These differences are important to take into consideration because they permit the development of risk instruments. Absconders tended to be younger in age, female, and wards of the state compared with those who did not abscond. Absconders were more involved in gang activity and less involved in religious activities than their counterparts. Those who absconded averaged two more instances of running away or being ejected from their homes. Absconders had poorer self-control and experienced substance dependency at higher rates than nonabsconders. No statistically significant differences emerged for juveniles who were parents, those with familial incarceration histories, sex offenders, and those released to community-based facilities, although fewer absconders were sex offenders and were placed in the care of their families.

A shortcoming in much of the risk assessment research is the lack of attention given to females (Holtfreter & Cupp, 2007; Salisbury et al., 2009; Van Voorhis et al., 2010). Often, instruments are developed without considering that variables operate differently for male and female offenders. To circumvent this issue, this study examined the extent to which variables should be modeled differently for female offenders. Following the above screening procedure, separate bivariate absconding models were estimated for male and female offenders, and coefficient differences were examined (Clogg et al., 1995). Most of the differences are trivial, which is good news for risk assessment research in general. There were, however, three variables that exhibited at least modest differences across genders. First, parenthood mattered more for female offenders than it did for male offenders. That is, parenthood was a protective factor for female offenders and positively, but unrelated statistically, associated with absconding for male offenders. In fact, it appears that male offenders were suppressing the parenting effect, because of their disproportionate prevalence in the sample. Second, these results indicated that running away or getting ejected from their homes mattered more for male offenders than it did for female offenders. This is the strongest effect observed among all of the factors included in the study, yet remains a trivial factor for female offenders, likely because of the already elevated base rates for female (M = 4.47) compared to male (M = 2.67) offenders. Third, low self-control acted as a risk factor for male offenders, increasing their likelihood of absconding, while remaining a weak and negative factor for female offenders. In sum, there are important differences when partitioning the sample by gender. To incorporate these differences, covariates that differed substantively or statistically were then interacted multiplicatively in the multivariate logistic models to allow for gender-weighted estimates (Jaccard, 2001).

Multivariate Logit Model

Table 2 details the results of a multivariate logistic regression model predicting absconding within the construction sample.² Diagnostics indicated that the model contained adequate explanatory capability, with the independent variables explaining the outcome better than a constant-only model. A likelihood ratio test revealed that the addition of gender-based interactions resulted in a statistically significant improvement of model fit (χ² = 12.61, p =.006). Using the mean rate of absconding as the cut point, 72% of the subjects were correctly classified, with sensitivity and specificity rates of 69% and 73%, respectively.

Table 2:

Logistic Regression Model Predicting Absconding Within the Construction Sample (n = 525)

Variable	b	SE	OR	95% CI
Age	−0.56	0.12	0.57	0.45 to 0.72
Female	3.03	0.97	20.70	3.10 to 138.3
Ward	0.98	0.50	2.67	1.01 to 7.07
Family incarceration	−0.003	0.10	1.00	0.83 to 1.20
Parenthood	0.67	0.46	1.96	0.80 to 4.80
Religion	−0.11	0.25	0.89	0.55 to 1.45
Gang membership	0.55	0.23	1.73	1.10 to 2.70
Runaway	0.19	0.04	1.21	1.13 to 1.30
Low self-control	0.06	0.02	1.06	1.02 to 1.10
Substance dependent	0.13	0.22	1.14	0.75 to 1.75
Sex offender	−0.34	0.52	0.71	0.26 to 1.97
Community placement	−0.19	0.36	0.83	0.41 to 1.68
Female, parenthood	−1.30	1.02	0.27	0.04 to 2.04
Female, runaway	−0.32	0.09	0.80	0.68 to 0.94
Female, low self-control	−0.09	0.04	0.92	0.84 to 0.99
Constant	6.59

Note. CI = confidence interval; OR = odds ratio.

Over half of the 15 covariates identified in the screening process meet conventional standards of statistical significance. The covariates related to absconding include gender, age, ward, and gang membership. Although older subjects were less likely to abscond, wards of the state and officially identified gang members were 2.7 and 1.7 times more likely to abscond, respectively. Although the confidence intervals for the odds ratios fell on both sides of 1.00, subjects who were religious, placed in community facilities, and sex offenders were less likely to abscond, while subjects dependent on substances were more likely to abscond.

Among the gender-specific covariates, for male offenders, 1-unit increases in low self-control and runaway corresponded with 6% and 21% increases in the likelihood for absconding. Alternatively, for female offenders, runaway and low self-control performed in the opposite direction, although at trivial rates. Notably, parenthood acted as a risk factor for male offenders but a protective factor for female offenders. The predicted probabilities, or translated log odds, for absconding are informative in this regard. When holding all covariates constant and varying only gender and parenthood, for male offenders, parenthood increases the odds of absconding from 27% to 41% but decreases the odds of absconding for female offenders from 42% to 28%.

Validation

Derived from the above logistic model, the weighted coefficients were then applied to the validation sample. Because the estimates that were obtained were modeled on the construction sample, they are essentially “blind” to the validation sample. Thus, if error is present in the model, it should be identified at this stage of the process. The predicted logged odds were translated into predicted probabilities, ranging from 0 to 1, that vary across juveniles according to their scores associated with the independent variables. The predictive validity of the ARI was assessed by evaluating rates of correct classification and r values according to risk categories.

First, rates of correct classification were obtained by comparing actual and predicted rates of absconding in the validation sample. As the convergence between actual and predicted rates increases, so too does support for the validity of the instrument. A two-by-two table categorized the subjects by true and false positives and negatives on the basis of the mean rate of absconding as the cut point. The overall rate of correct prediction was 70%, with a sensitivity rate (i.e., true positives) of 67% and a specificity rate of 71%. Receiver-operating characteristic curve analysis revealed that the area under the curve was .76 (95% confidence interval = .71 to .80). Second, r values, or effect sizes, refer to the bivariate Pearson’s correlations between actual absconding rates and categories of risk (Latessa et al., 2009). Subjects were subdivided into four categories (minimum to −1 SD, −1 to 0 SD, 0 to 1 SD, and +1 to maximum SD) on the basis of the distribution of the predicted probabilities, which permit correctional officials to make decisions related to risk. The observed r value for the validation sample was .37, compared with .43 for the construction sample, which is consistent with the effects found in the Ohio study on recidivism risk assessment across five stages of the criminal justice system (Latessa et al., 2009). As expected, slippage was observed in the validation process both at correct classification and r value stages, but this is a common occurrence in the literature (Borum, 1996; Gottfredson & Moriarty, 2006). Nevertheless, the instrument adequately predicted absconding behavior in the validation sample. In summary, the ARI passed a series of important tests indicating that it maintains predictive validity.

Discussion

This study developed an instrument to assess risk for absconding among juvenile parolees in Arizona. Drawing on five domains, 12 risk factors were identified and used to construct the instrument on the basis of information drawn from 1,063 juveniles paroled from ADJC facilities into the community in 2008 and 2009. Following the principles of risk assessment (Bonta, 2002; Gottfredson & Snyder, 2005), this study randomly subdivided juveniles into construction and validation samples, accounting for gender-specific factors, to develop and validate the ARI. Upon application of the estimates derived from construction sample to the validation sample, 70% of the subjects were classified correctly, and an effect size of r =.37 was observed when the subjects were categorized by risk levels. The ARI demonstrated sufficient predictive and discriminant validity in the identification and classification of absconding juvenile parolees. On the basis of these results, two main points guide this section.

First, gender differences matter. Three covariates—parenthood, runaway, and low self-control—were found to operate differently for male and female offenders. Parenthood acted as a protective factor for female offenders and a risk factor for male offenders. There are gendered responses to teenage parenthood, whereby females tend to assume greater responsibility over their children’s welfare and childbearing acts as a turning point in the life course (Kreager, Matsueda, & Erosheva, 2010; Luker, 1996). If female offenders are forcibly separated from their children, this relationship might perform in the opposite direction. Runaway behavior and low self-control, alternatively, were strong risk factors for male offenders but negligible for female offenders. The runaway effect was likely observed because base rates of runaway behavior were nearly 70% greater for female offenders, consistent with what is known about gender and runaway behavior (Kempf-Leonard & Johansson, 2007). In other words, there is likely a selection process at play for female offenders in terms of runaway and absconding behavior, whereas for male offenders, it taps the propensity and mechanisms to abscond. Supplementary analysis exploring interaction with household experiences that might condition this relationship did not reveal additional differences. For low self-control, the gendered finding is not entirely unexpected. Although female offenders were more likely to abscond, the literature demonstrates that male offenders engage in delinquency and risk-taking at higher rates than female offenders (Burton, Cullen, Evans, Alarid, & Dunaway, 1998; Byrnes, Miller, & Shafer, 1999). Violating parole conditions demonstrates shortsightedness and risk-taking behavior. There are gendered, substantive differences in absconding behavior.

Static characteristics such as gender (and race/ethnicity) present issues to risk assessment instruments that are both serious and sensitive in nature. An example of this can be found in Smith, Cullen, & Latessa’s (2009) meta-analysis of the LSI-R and recidivism, for which they asked if a sample of over 14,700 women was enough to convince critics that the tool was gender neutral (see also Schwalbe, 2008). Rather than approaching the issue from either side of the risk assessment “gender aisle,” this study sought to let the data speak for themselves. As discussed above, three of the covariates differed in meaningful ways. Researchers and practitioners should recognize that the incorporation of gender differences improve model fit and explanatory power, reducing the likelihood of misclassification. A theoretically informed screening process for gendered differences is not a labor-intensive task. Analytically, multiplicative interactions can account for these differences and estimates are thus weighted accordingly. These weights result in one equation that can then be pilot tested according to a subject’s information and absconding probabilities can be obtained. This is no more difficult than adding up scores using the Burgess (1928) method and allows agencies to interpret scores in terms of chance, or percent likelihood of absconding, while accounting for gendered processes.

Second, although the results indicate that absconding can be modeled empirically, several questions emerge about the efficacy of absconding risk assessment instruments, given the limited state of the literature. There are a number of tools that seek to assess risk for law-violating behaviors among those under supervision. With few exceptions, the problem is that very few of these tools have been applied to the context of absconding. Although it is premature to conclude absconding behavior requires a separate risk tool, the existing literature demonstrates only modest success in predicting absconding using existing tools (e.g., Lowenkamp et al., 2001) compared with what can be described as a large r value in the current study. The larger question that this study does not explore is how absconding and reoffending behavior interact. In other words, are subjects at high risk for absconding also those at high risk for new violations? Many of the factors used in the current study are consistent predictors of criminal, delinquent, and deviant behavior. Future research may consider assessing the degree to which the coefficients for such factors vary in magnitude across outcomes of relevance to supervisory agencies. However, given what is known about the nature of absconding behavior and existing empirical evidence (i.e., Mayzer et al., 2004), it is likely that predictors differ in nontrivial ways, thus supporting the efficacy of an absconding risk assessment instrument. Nevertheless, an operational issue emerges for agencies if they are confronted with divergent risk classifications (e.g., high abscond/low offending risk or low abscond/high offending risk). As it stands, without evaluations of discriminant validity, each tool would be recognizing the potential for at least some problem behavior that would be of interest to agencies responsible for monitoring the behavior of probationers or parolees.

Although this study is among the first to develop a risk instrument for absconding, limitations of this research should be considered. This study is limited to juvenile parolees in Arizona. There may be political, economic, or jurisdictional influences and policies for recording absconding behavior that might limit the applicability of the findings to other settings (see Grattet et al., 2008). Similarly, there may be risk factors for absconding that are unique to Arizona youth that may not operate correspondingly in other contexts. Because most of the factors included in this study maintain a theoretical orientation (e.g., social learning theory, self-control theory, strain theory), this burden rests on theoretical generalizability and measurement more than the accuracy of the instrument. Also, although the predictive ability of the independent variables was adequate, with 76% of subjects falling in the area under the curve, there were a number of false positives and negatives that were not identified by the model. Furthermore, although this study was motivated empirically in its gender modeling strategy, the continued conceptualization and modeling of gendered pathways to success and failure in correctional supervisory contexts should be a high priority for future risk assessment research. As Holtfreter and Cupp (2007) pointed out, “‘add gender and stir’ in tests of criminological theories and actuarial risk assessment . . . may be far more costly over the long run” (p. 378).

Future research should address the limitations identified above and expand on models developed herein. Four areas of study should be pursued which would hold utility to the correctional supervision literature in general and absconding in particular. First, much like risk instruments for recidivism in general, revalidation in the Arizona context and extension to other geographical contexts would be a worthwhile endeavor. Second, qualitative interviews with recommitted absconders should be carried out to identify more immediate factors related to decisions to avoid supervising agents, with attention directed toward factors identified in the ARI and differential patterns of absconding behavior. This may help identify unobserved or unmodeled factors that could be considered in further validation. Third, differential follow-up periods should be used as an outcome, because not all juveniles absconded within 12-month time frames. A survival analytic approach could identify factors associated with short-term (i.e., immediately after release) and long-term (i.e., 6, 12, and 18 months after release) time patterns to failure. Finally, experimental and quasi-experimental designs should be used with risk categories to determine the practical effectiveness of the instrument. These approaches can be combined with practitioner and professional clinical approaches.

Footnotes

Acknowledgements

I would like to acknowledge Michelle Anderson, Stella Vasquez, and John Vivian of the Arizona Department of Juvenile Corrections for their technical support of this project and Scott Decker, Mike Reisig, and John Vivian for their helpful comments on earlier versions of the manuscript. Portions of this research were presented at the 2011 annual meeting of the Western Society of Criminology in Vancouver, Canada. The opinions and conclusions expressed in this document are solely my own and do not necessarily reflect the views of the Arizona Department of Juvenile Corrections.

Notes

David C. Pyrooz is an assistant professor in the College of Criminal Justice at Sam Houston State University. He received his PhD in Criminology and Criminal Justice in 2012 from Arizona State University. His research interests concentrate on gangs and deviant networks, developmental and life-course criminology, and criminal justice system responses. His research has appeared in Crime & Delinquency, Homicide Studies, the Journal of Research in Crime and Delinquency, Justice Quarterly, and Social Science Research.

References

Andrews

D. A.

Bonta

J. L.

(1995). The Level of Service Inventory–Revised. Toronto, Canada: Multihealth Systems.

Andrews

D. A.

Bonta

Wormith

J. S.

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

Belknap

Holsinger

(2006). The gendered nature of risk factors for delinquency. Feminist Criminology, 1, 48-71.

Blankstein

(2010, September 14). Absconded parolee named prime suspect in slaying of Valley Village woman. The Los Angeles Times. Retrieved October 26, 2010, from http://latimesblogs.latimes.com/lanow/2010/09/absconded-parolee-named-prime-suspect-in-slaying-of-valley-village-woman.html

Bonta

(2002). Offender risk assessment: Guidelines for selection and use. Criminal Justice and Behavior, 29, 355-379.

Borum

(1996). Improving the clinical practice of violence risk assessment. American Psychologist, 51, 945-956.

Burgess

E. W.

(1928). Factors determining success or failure on parole. In Bruce

A. A.

Burgess

E. W.

Landesco

Harno

A. J.

(Eds.), The workings of the indeterminate sentence law and the parole system in Illinois (pp. 221-234). Springfield: Illinois State Board of Parole.

Burton

V. S.

Jr. Cullen

F. T.

Evans

T. D.

Alarid

L. F.

Dunaway

R. G.

(1998). Gender, self-control, and crime. Journal of Research in Crime and Delinquency, 35, 123-147.

Byrnes

J. P.

Miller

D. C.

Shafer

W. D.

(1999). Gender differences in risk taking: A meta-analysis. Psychological Bulletin, 125, 367-383.

10.

Clogg

C. C.

Petkova

Haritou

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

11.

Cottle

C. C.

Lee

R. J.

Heilbrun

(2001). The prediction of criminal recidivism in juveniles: A meta-analysis. Criminal Justice and Behavior, 3, 367-394.

12.

Gendreau

Little

Goggin

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

13.

Gottfredson

D. M.

Snyder

H. N.

(2005). The mathematics of risk calculation: Changing data into valid instruments for juvenile courts. Washington DC: Office of Juvenile Justice and Delinquency Prevention.

14.

Gottfredson

M. R.

Hirschi

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

15.

Gottfredson

Moriarty

(2006). Statistical risk assessment: Old problems and new applications. Crime & Delinquency, 52, 178-200.

16.

Grattet

Petersilia

Lin

(2008). Parole violations and revocations in California. Washington DC: U.S. Department of Justice. Retrieved March 10, 2011, from http://www.ncjrs.gov/pdffiles1/nij/grants/224521.pdf

17.

Hanrahan

Gibbs

J. J.

Zimmerman

S. E.

(2005). Parole and revocation: Perspectives of young adult offenders. Prison Journal, 85, 251-269.

18.

Hare

R. D.

(1991). The Hare Psychopathy Checklist–Revised (PCL-R). Toronto, Canada: Multihealth Systems.

19.

Harris

G. T.

Rice

M. E.

Quinsey

V. L.

(1993). Violent recidivism of mentally disordered offenders: The development of a statistical prediction instrument. Criminal Justice and Behavior, 20, 315-335.

20.

Holtfreter

Cupp

(2007). Gender and risk assessment: The empirical status of the LSI-R for women. Journal of Contemporary Criminal Justice, 23, 363-382.

21.

Holtfreter

Reisig

Morash

(2004). Poverty, state capital, and recidivism among women offenders. Criminology and Public Policy, 3, 185-208.

22.

Jaccard

(2001). Interaction effects in logistic regression. Newbury Park, CA: Sage.

23.

Kempf-Leonard

Johansson

(2007). Gender and runaways: Risk factors, delinquency, and juvenile justice experiences. Youth Violence and Juvenile Justice, 5, 308-327.

24.

Kreager

D. A.

Matsueda

R. L.

Erosheva

E. A.

(2010). Motherhood and criminal desistance in disadvantaged neighborhoods. Criminology, 49, 221-258.

25.

Latessa

Smith

Lemke

Makarios

Lowenkamp

(2009). Creation and validation of the Ohio risk assessment system: Final report. Cincinnati, OH: University of Cincinnati.

26.

Lowenkamp

Holsinger

A. M.

Latessa

E. J.

(2001). Risk/need assessment, offender classification, and the role of childhood abuse. Criminal Justice and Behavior, 28, 543-563.

27.

Luker

(1996). Dubious conceptions: The politics of teenage pregnancy. Cambridge, MA: Harvard University Press.

28.

Mayzer

Gray

K. M.

Maxwell

S. R.

(2004). Probation absconders: A unique risk group? Journal of Criminal Justice, 32, 137-150.

29.

Megerian

(2010, September 14). More than 1,000 fugitives, parole absconders arrested in N.J. summer sweep. Retrieved October 26, 2010, from http://www.northjersey.com/news/091410_More_than_1000_fugitives_parole_absconders_in_NJ_summer_sweep.html

30.

O’Donnell

Jewkes

(2011). Going home for Christmas: Prisoners, a taste of freedom. Howard Journal of Criminal Justice, 50, 75-91.

31.

Padgett

K. G.

Bales

W. D.

Blomberg

T. G.

(2006). Under surveillance: An empirical test of the effectiveness and consequences of electronic monitoring. Criminology and Public Policy, 5, 61-92.

32.

Reisig

M. D.

Holtfreter

Morash

(2006). Assessing recidivism risk across female pathways to crime. Justice Quarterly, 23, 384-405.

33.

Reynolds

C. R.

Kamphaus

R. W.

(2004). Behavior Assessment System for Children (2nd ed.). Circle Pines, MN: AGS.

34.

Salisbury

E. J.

Van Voorhis

Spiropoulos

G. V.

(2009). The predictive validity of a gender-responsive needs assessment: An exploratory study. Crime & Delinquency, 55, 550-585.

35.

Schwalbe

C. S.

(2008). A meta-analysis of juvenile justice risk assessment instruments: Predictive validity by gender. Criminal Justice and Behavior, 35, 1367-1381.

36.

Schwaner

S. L.

(1997). They can run, but can they hide? A profile of parole violators at large. Journal of Crime and Justice, 20, 19-32.

37.

Schwaner

S. L.

McGaughey

Tewksbury

(1998). Situational constraints and absconding behavior: Toward a typology of parole fugitives. Journal of Offender Rehabilitation, 27, 37-55.

38.

Smith

Cullen

F. T.

Latessa

E. J.

(2009). Can 14,737 women be wrong? A meta-analysis of the LSI-R and recidivism for female offenders. Criminology and Public Policy, 8, 183-208.

39.

Steen

Opsal

(2007). Punishment on the installment plan: Individual-level predictors of parole revocation in four states. Prison Journal, 87, 344-366.

40.

Taxman

F. S.

Byrne

J. M.

(1994). Locating absconders: Results from a randomized field experiment. Federal Probation, 58, 13-23.

41.

Van Voorhis

Brown

(1997). Risk classification in the 1990s. Washington DC: National Institute of Corrections.

42.

Van Voorhis

Wright

Salisbury

Bauman

(2010). Women’s risk factors and their contributions to existing risk/needs assessment. Criminal Justice and Behavior, 37, 261-288.

43.

Williams

F. P.

III McShane

M. D.

Dolny

H. M.

(2000). Predicting parole absconders. Prison Journal, 80, 24-38.

44.

Wolfe

Vivian

(2008). Follow up reports on absconds (Memorandum). Phoenix: Arizona Department of Juvenile Corrections.