Predicting Family Violence Recidivism Using the DVSI-R

Abstract

Recent published predictive validation studies of risk assessment instruments commonly used the explanatory power of Receiver Operating Characteristics (ROC) analysis. The present study also tested the predictive validity of the Revised Domestic Violence Screening Instrument (DVSI-R) by conducting a ROC analysis of 18-month follow-up data post assessment from the entire state of Connecticut between January 1, 2010, and May 31, 2012. With a sample of 29,317, the analysis examined rearrests for new family violence incidents and new violations of protective or restraining orders. The sample was predominantly male (70%), with an average age of 34 years, and equally split between non-Latino White and minority perpetrators. Moving beyond the explanatory power of ROC analysis, the Kaplan–Meier (KM) estimator of the survival function was used to examine the cumulative probability of rearrest during the time at-risk period. In addition to determining the survival functions of perpetrators based on their level of risk as identified by their DVSI-R total numeric risk scores, the analysis examined differences by characteristics of perpetrators, an issue given scant attention in prior research. Survival analyses identified critical times after intake assessments when those who reoffended were at greatest risk and significant differences in timing of reoffending by gender and ethnicity. Implications for intervention are discussed.

Keywords

domestic violence screening family violence risk assessment survival analysis

In the wake of burgeoning arrests for family violence crimes (e.g., Hirschel, Buzawa, Pattayina, Faggiani, & Reuland, 2007; Simpson, Bouffard, Garner, & Hickman, 2006), investigators developed risk assessment instruments to guide the supervision of perpetrators and initiated predictive validation research on those assessment tools (e.g., Chan, 2012; Goodman, Dutton, & Bennett, 2000; Heckert & Gondolf, 2004; Williams, 2011; Williams & Grant, 2006; Williams & Houghton, 2004; and for reviews, see Dutton & Kropp, 2000; Hilton & Harris, 2005; Hilton, Harris, & Rice, 2010). These instruments generally assessed the risk of future offending through an assessment of past offending (but see Chan, 2012). More specifically, the research objective of these investigations was to determine whether assessed risk (typically, a total score based on summation of item ratings) predicted actual risk measured by behavioral recidivism (i.e., repeated violent behavior post assessment). The primary statistical procedure used to address this objective was Receiver Operating Characteristic (ROC) analysis. It became the state-of-the-art procedure for predictive validation studies (Chan, 2012; Hilton, Carter, Harris, & Sharpe, 2008; Hilton & Harris, 2009; Williams, 2011). Although useful for predicting reoffending, ROC analysis did not provide information about the timing of reoffending and differences in timing between groups.

This study addressed the limitations of ROC analysis by conducting survival analysis of 18-month follow-up data post assessment and risk assessment data derived from the application of the Revised Domestic Violence Screening Instrument (DVSI-R; Williams, 2011; Williams & Grant, 2006; and see Williams, 2008, for a discussion of the developmental history of the DVSI-R). More specifically, the present study made three important contributions to the literature on family violence risk assessment. First, previous studies of the predictive validity of the DVSI-R analyzed data from either the developmental sample used to construct the instrument (Williams & Grant, 2006) or from a statewide sample of perpetrators assessed during a 2-month period (Williams, 2011). The present study examined recidivism among all perpetrators arrested on family violence charges in Connecticut in 2010 (N = 29,317) during an 18-month follow-up period. Second, survival analysis techniques were utilized that empirically identified the point in time when perpetrators or groups of perpetrators failed (i.e., reoffended), that is, determined a time-specific estimate of recidivism (DeJong, 1997; Heagerty & Zheng, 2005). Third, this technique allowed a comparison of the distribution of time to reoffending between two or more different subgroups in the population and a test of the significance of differences between the groups (Rice & Harris, 1997). Specifically, the analysis estimated moderation effects by age, ethnicity, and gender of perpetrators, as well as household relationship between perpetrators and victims on the relation between the DVSI-R risk levels and the timing of recidivism.

Incorporating Survival Analysis

As mentioned, recent studies favored ROC analysis to estimate the predictive accuracy of risk assessment instruments pertaining to violent recidivism (Dolan & Doyle, 2000; Mossman, 1994; Rice & Harris, 1997; Williams, 2011). A ROC curve is a plot of test sensitivity against its false-positive rate (Lusted, 1971), with each point on the graph generated by different cut points on the prediction instrument. The area under the ROC curve (AUC), which takes on values between 0.0 and 1.0, measures the accuracy of the risk assessment instruments. An AUC coefficient of 1.0 is perfectly accurate because the sensitivity is 1.0 when the false-positive rate is 0.0. The ROC plot offers ease of interpretation and enables a direct visual comparison of two or more tests on a common set of scales at all possible cut points of a prediction instrument (Lusted, 1971). Furthermore, ROC analysis easily ascertains how altering the selection ratio (by adjusting the cutoff score on the prediction instrument) changes both the true-positive and false alarm rates (Rice & Harris, 1997). ROC curves also remain constant as the base rates of violence in the sample change (Mossman, 1994). Although these benefits rendered ROC analysis useful for predicting reoffending, this analytical technique did not empirically identify the timing of reoffending. Survival analysis is appropriate for this task.

Survival analysis, or event-history analysis, has roots in studies of mortality, estimating the probability of death over the life course. Although used sporadically, the technique has also been used to estimate the probability of offending and recidivism in criminology. Studies of family violence (Yoshima & Gillespie, 2002; Yoshihama & Horrocks, 2003), dating violence (Smith, White, & Holland, 2003), sexual offenders (Rice & Harris, 1997), and recidivism more generally (DeJong, 1997; McNeil & Binder, 2007) benefitted from survival curves and event-history analysis to examine the generality of findings and to compare subgroup differences, controlling for other factors (Rice & Harris, 1997). To this end, survival analysis was used successfully with short-term recall data of periods between 6 months and 2 years (DeJong, 1997; McNeil & Binder, 2007; Rice & Harris, 1997; Smith et al., 2003; Yoshihama & Gillespie, 2002; Yoshihama & Horrocks, 2003).

Survival analysis in studies of violent recidivism has advantages. First, the information produced empirically identifies the exact point in time when perpetrators or groups of perpetrators failed (i.e., reoffended)—that is, a time-specific estimate of recidivism (Heagerty & Zheng, 2005). This analytical technique also allows examination of survival to the end of the study period and differences between immediate and delayed failure. As DeJong (1997) highlighted this matter, such estimation is important because after sanctioning, a perpetrator may reoffend quickly, reoffend after an extended period, or may never reoffend again. Such information is important for targeting intervention programs with the maximum possibility of effectiveness in preventing further reoffending (Yoshihama & Gillespie, 2002). Second, the technique delineates the distribution of time to reoffending between two or more subgroups in the population and tests the significance of the difference between the subgroups. For example, in their study of child molesters and rapists, Rice and Harris (1997) demonstrated the important extensions made by survival analyses by illustrating the prediction of future violence and the differential rates of failure between the two groups. Supplemented with Cox Regression Estimation, survival analysis determines group differences while controlling for other factors (Rice & Harris, 1997).

Method

Sample Selection

The research reported below used survival analysis techniques to examine 29,317 perpetrators arrested on family violence charges across the state of Connecticut for the entire year of 2010. Data on perpetrator characteristics and risk assessment scores were obtained from intake interviews conducted shortly after the arrest of a perpetrator by trained Family Relations Counselors (FRCs) who administrated the DVSI-R. They coded the DVSI-R items during those interviews, and a computerized system automatically generated the total numeric risk scores. Completing this risk assessment instrument required FRCs to draw from five sources of information: perpetrator interviews; victim interviews typically conducted by victim advocates, with information given to FRCs only after victim consent; review of the police report; criminal history records review; and protective–restraining order registry review. Perpetrator characteristics and risk assessment scores were entered via computerized instrumentation into a statewide data management system. These data for 2010 were matched with follow-up family violence recidivism data for new family violence offenses (NFVO) and violations of protective or restraining orders during an 18-month follow-up period post assessment to form the data file analyzed in the present study. As discussed below, the recidivism data were rearrests for these offenses and violations based on in-state records checks.

Risk Assessment Instrument

The DVSI-R includes 11 items, 7 of which primarily address the behavioral history of perpetrators (prior nonfamily assaults, arrest, or criminal conviction; prior family violence assaults, threats, or arrests; prior family violence intervention or treatment; prior violation of orders of protection or court supervision; prior or current verbal or emotional abuse; the frequency of family violence in the previous 6 months; and escalation of family violence in the previous 6 months). The other 4 items include substance abuse, objects used as weapons, children present during violent incidents, and employment status. Item scores range from 0 to 2 or 0 to 3, depending on the item, estimating the intensity of the risk factor, not just whether it is present for the perpetrator assessed. For example, one item on the instrument is “evidence of nonfamily assaults, arrests, or criminal convictions,” and the scoring categories are “0 = no evidence, 1 = one or two incidents, 2 = three or more incidents.” An example of an item with a 0 to 3 scoring category is “history of violation of orders of protection or court supervision,” and the scoring categories are “0 = no evidence, 1 = prior incident, 2 = current incident, 3 = prior and current incidents.” The possible range of the total numeric risk score generated by the administration of the DVSI-R is 0 to 28.

This instrument also includes two perceptual items that ask the evaluator to judge the imminent risk of future violence (i.e., within the next 6 months) to the victim and to some other persons known to the perpetrator and/or victim, scored 0 “low risk,” 1 “medium risk,” and 2 “high risk.” These perceptual items were included to allow evaluators the option of exercising a “clinical” override to the total numeric score in relatively rare cases where they strongly believed that score did not represent the imminent risk to victims or other persons. That issue was not central to the objectives of the present study, and thus, those perceptual items were not included in the analysis. Furthermore, imminent risk to victim and imminent risk to others were highly associated with risk level (DVSI-R total numeric risk scores collapsed into quartiles), as indicated by an ordinal measure of statistical association (γ = .86 and γ = .77, respectively). In addition, previous research showed that adding these perceptual risk items to logistic regression equations estimating the odds of recidivism did not add to predictive accuracy beyond the DVSI-R total numeric risk score (Williams, 2011).

The DVSI-R is a significantly revised version of the original DVSI developed and validated in Colorado (Williams & Houghton, 2004). It went through various revisions resulting from on-site observations by the researcher, input from FRCs through focus groups, and iterative analyses for quality assurance to identify any confusion on the part of FRCs or inaccuracies in administration. This process occurred between 2002 and 2004 when the final version of the DVSI-R went statewide. The researcher conducted trainings on the DVSI-R for FRCs that included the historical development of the instrument in Connecticut, the research upon which it was based, and administration of the instrument itself. Follow-up booster training sessions were also conducted either by the researcher or supervisors of FRCs. Three different case files were used by all those in training to practice the application of the DVSI-R to those cases, with an agreement in scoring of 80% “plus or minus one” maintained. A more formal follow-up testing of interrater reliability has yet to be completed.

Family Violence Recidivism

The analysis used two measures of behavioral recidivism because finding similar results across different measures enhances confidence in the predictive validity of the DVSI-R. The indicator of recidivism was rearrest. The primary limitations of this indicator are well known. For example, not all reoccurring incidents of family violence are detected and reported to the police, and despite Connecticut having a mandatory arrest law, not all reported incidents always result in an arrest on an official offense charge. However, these data are commonly used in recidivism studies, with alternatives such as perpetrator or victim self-reports being too costly, ethically controversial, and having their own sources of error (for a more general discussion see Hilton et al., 2010; Williams, Tuthill, & Lio, 2008).

The two primary measures of recidivism during the 18-month follow-up period were rearrests for NFVO only and rearrests for violations of protective or restraining orders only (order violations). NFVO rearrests bear on the issue of continued behavioral involvement in family violence. Violations of protective or restraining orders bear on the issue of not honoring court orders intended to keep victims safe. Knowing that a risk assessment instrument has predictive validity of both issues provides guidance in identifying higher risk perpetrators and also guidance in taking steps to protect victims from such perpetrators. Utilizing these measures determined whether assessed recidivistic risk of family violence was associated with actual recidivistic risk of this behavior. The “only” term used with the two family violence recidivism measures denotes that each of these categories included “refined” types, meaning perpetrators rearrested for some type of family violence recidivism were sorted into the specific and mutually exclusive categories enumerated above. Dummy variables were calculated in which those falling into one of these categories were scored “1.” All others, including those rearrested for something besides a NFVO only or a violation of a protective or restraining order only, and those not rearrested, were scored “0.”

Analysis Plan

The analysis examined rearrests for NFVO incidents and new violations of protective or restraining orders during an 18-month follow-up period, between January 1, 2010, and May 31, 2012. This period gave a total examination frame of 882 days, assuming an intake on January 1, 2010, and a potential rearrest date on May 31, 2012, the final day of the follow-up period. In the data collected, rearrests occurred between 1 and 869 days after intake. Although the focus of the present study moved beyond the explanatory power of ROC analysis, the predictive accuracy of the DVSI-R was determined by estimating the relation between the total numeric risk scores and the two recidivism measures, using ROC analysis. As mentioned, recent published predictive validation studies of risk assessment instruments commonly used this technique (Hilton et al., 2008; Hilton & Harris, 2009; Williams, 2011). For a review of this technique, see Quinsey, Harris, Rice, and Cormier (2006).

Moving beyond the explanatory power of ROC analysis, the Kaplan–Meier (referred to as KM from now on) estimator of the survival function was used to examine the cumulative probability of rearrest during the time at-risk period. The estimator of Kaplan and Meier (1958) is an estimate of either the probability of survival past time t, or the probability of failing after t (Cleves, Gould, Gutierrez, & Marchenko, 2008). KM analysis allows estimation of population survival by simply calculating the fraction of the number of individuals at risk at each time point that fail by that time point. The probability of survival at any point is calculated from the cumulative probability of surviving each of the preceding time points (Cleves et al., 2008).

A survival plot illustrated the pattern of first rearrest over the course of this period, yielding useful information regarding the timely targeting of more intensive supervision and intervention. Survival analysis is particularly useful in recidivism studies where perpetrators enter the at-risk period at different points along a timeline and censored before the study period ends. At the end of that period, a majority of perpetrators “survived” (not rearrested) concerning NFVO, although this does not guarantee that some were rearrested at a later point in time. These cases are “censored.” Survival analysis correctly accounts for censored cases in estimating the probability of rearrest in days, while allowing for the possibility that some perpetrators will never be rearrested. With censored data, the KM estimator can be calculated and plotted to display the cumulative probability of rearrest by a certain point in time (Yoshihama & Gillespie, 2002).

The KM survival procedure does not require that perpetrators be under supervision by the court or family services by the start of the study period. It also does not require that they recidivate or successfully complete their supervision sometime during the study period. In this study, the number of days until the recidivism event (i.e., rearrest) was defined as the number of days since the initial intake date and the date of the first rearrest.

In addition to determining the survival functions of perpetrators based on their level of risk as identified by their DVSI-R total numeric risk scores, the analysis examined differences in the survival functions by characteristics of perpetrators, an issue given scant attention in prior research. The characteristics included gender, age split at the mean to form two categories of younger (15-35) and older (36-91) perpetrators, relationship type including partner victims and child or other victims, and ethnicity, dichotomized along the comparative line of ethnic majority (non-Latino Whites) and ethnic minorities (including primarily African Americans and Latinos).

Estimation of Cox proportional hazards regression models achieved the objective of determining differential effects by these characteristics (Cox, 1972). Proportional hazard models are widely used in survival analyses given their ability to relate the time to failure to one or more covariates associated with that time. The Cox (1972) proportional hazards regression model assumes that a unit increase in any given covariate multiplicatively shifts the baseline hazard function (Cleves et al., 2008). Thus, the use of a Cox regression model allowed estimation of the effects of perpetrator characteristics on one’s hazard for rearrest, and determine whether those characteristics moderated the empirical relation between the DVSI-R risk level (i.e., the total numeric risk scores categorized into quartiles) and rearrests for NFVO or order violations. Another advantage of Cox regression is that it permitted the estimation of both categorical and continuous variables. Thus, it allowed an empirical estimation of the importance of perpetrator characteristics (i.e., age, gender, ethnicity, and household relationships between perpetrators and victims) and risk level on these two measures of recidivism. This analytical procedure also ensured the identification of the variables important for examining survival curves (Allison, 1995; Rice & Harris, 1997).

Estimating this model entailed two important assumptions. First, it assumed that effects of covariates were additive and linear. Second, it assumed that perpetrator characteristics moderated the empirical relation between the risk level as determined by the total numeric risk scores and the two measures of family violence recidivism. Hence, interaction terms (i.e., cross products between each perpetrator characteristic and risk level) were calculated and included in the model estimated.

Empirical Results

Sample Characteristics

Table 1 displays descriptive statistics of the DVSI-R total numeric risk scores, perpetrator characteristics, and different types of household relationships between perpetrators and victims by each recidivism measure used in the analysis. The majority of the sample was male (69.68%), with an average age of 34 years as of January 1, 2010. Almost half of the sample was non-Latino White, and half was Other ethnic minorities. The mean score of the DVSI-R was 9.46, indicating a rather high level of risk of recidivism, compared with the developmental sample (7.75) or the cross-validation sample (8.28) previously published (Williams & Grant, 2006; Williams, 2011, respectively). Internal consistency was α = .73, similar to the coefficient reported in a cross-validation study of the DVSI-R (α = .75; Williams, 2011). Of the 29,317 cases, rearrests for a NFVO occurred for 6,864 (23.41%) perpetrators, with 14.15% rearrested for a violation of protective or restraining orders. Table 1 also shows that among those rearrested, assessed risk was higher, and rearrests were higher for younger perpetrators, male perpetrators, and incidents involving partners as victims.

Table 1:

Sample Characteristics for the Total Sample and by Rearrest Type

	Total	NFVO^a	VPO/VRO^b
	(N = 29,317)	(n = 6,864)	(n = 4,148)
DVSI total score (SD)	9.460 (5.597)	11.694 (5.583)	12.664 (5.434)
Risk level, 1-4 (SD)	2.591 (1.130)	3.017 (1.050)	3.204 (.989)
Age (SD)	34 (11.868)	32.5 (11.321)	33 (11.105)
% age 15-34	56.65	63.80	61.91
% age 35 +	43.35	36.20	38.09
% male	69.68	75.25%	79.73%
% female	30.32	24.75	20.27
% non-Latino White	49.31	46.95	49.06
% minority	50.69	53.05	50.94
% partner victim	64.28	67.24	74.17
% child victim	12.48	11.00	7.86
% other victim	23.24	21.26	17.97

Note. Standard Deviations (SD) in parentheses. NFVO = new family violence offenses; VPO/VRO = violation of protective or restraining orders.

Rearrests for a new family violence offense only.

Rearrests for violations of protective or restraining orders only.

Predictive Accuracy of the DVSI-R

Table 2 presents the results of the ROC analysis. Use of this technique ensured that the statewide sample produced estimates consistent with prior research. Comparing data from the entire state of Connecticut with a smaller subsample of the population used in a prior cross-validation study (Williams, 2011), a large degree of consistency was demonstrated. The DVSI-R total numeric risk scores were significantly associated with rearrests for NFVO and order violations, with AUC coefficients of .649 and .690, respectively. The AUC coefficient for both types combined was .674.

Table 2:

Comparisons of Area Under the Curve Coefficients

	Present Study (N = 29,317)			Williams, 2011 (N = 3,569)
	NFVO^a	VPO/VRO^b	Both	NFVO^a	VPO/VRO^b	Both
DVSI-R total	.649 [.642, .656]	.690 [.681, .698]	.674 [.668, .681]	.621 [.598, .644]	.715 [.662, .768]	.659 [.639, .678]

Note. Confidence Intervals [CI] in parentheses. NFVO = new family violence offenses; VPO/VRO = violation of protective or restraining orders.

Rearrests for a new family violence offense only.

Rearrests for violations of protective or restraining orders only.

Survival Analysis

Consider now the results of the survival analysis. The survival estimate displays the probability of surviving beyond time t, ranging from 1.00 to 0.00, with 1.00 representing survival for everyone in the sample. Figure 1 shows that in Connecticut, the first 200 days represented a particularly key window within which the majority of perpetrators who reoffended were rearrested. Life tables provided information to specify key points in time for NFVO and order violations.

Figure 1:

Kaplan–Meier Survival Curves for NFVO and VPO/VRO

Examining NFVO only, the failure rate was fastest in the first 100 days after intake, when 1,891 individuals (6.45% of cases) failed. NFVO rearrests occurred for 10% of all cases by the end of day 217. By the end of 595 days after intake, 20% of the cases failed. Almost 78% of cases (22,850) survived until the end of 869 days. Rearrests for order violations constituted a smaller percentage of cases, with 86.50% of cases surviving until the end of the study period. Similar to NFVO, however, perpetrators failed faster within the first 200 days. By the end of 200 days, 9% of all cases failed, more than half of which failed within the first 63 days. Although 10% of cases failed by the end of 257 days, only an additional 3.45% of the total sample failed by the end of the study period, 612 days later.

Perpetrator Characteristics and Moderation Effects

Table 3 displays the result of the Cox regression analysis with a dichotomous dependent variable indicating the probability of rearrest for NFVO. Table 4 displays the same information but for the probability of order violations. Results for both dependent variables indicated that all perpetrator characteristics and partner as victim had statistically significant estimated effects. Examining the hazard ratios, the model indicated that for male perpetrators, the rate of rearrest for NFVO was 54.6% higher compared with females, holding all other variables constant. For non-Latino White perpetrators, the rate of rearrest for NFVO was 28.9% lower (1.00 -.711), compared with ethnic minority perpetrators. If the risk level of the perpetrator increases by one unit, while holding all other variables constant, the rate of rearrest for NFVO increases by 55.9%. These results were based on the reported hazard ratios. Using the raw coefficients is instructive in discussing the variables involved in the interaction terms (Cleves et al., 2008; UCLA Statistical Consulting Group, 2011), provided in the appendix.

Table 3:

Results of Cox Regression Predicting the Likelihood of Rearrest for a New Family Violence Offense

	Hazard Ratio	SE	z
Gender (female = 0, male = 1)	1.546***	0.127	5.31
Age (15-35 = 0, 36-91 = 1)	0.607***	0.049	−6.17
Ethnicity (minority = 0, White = 1)	0.711***	0.054	−4.47
Victim (child = 0, partner = 1)	1.179*	0.092	2.09
Risk level	1.559***	0.048	14.39
Age × Risk level	1.059*	0.027	2.28
Ethnicity × Risk level	1.118***	0.027	4.63
Partner victim × Risk level	0.981	0.025	−0.76
Gender × Risk level	0.902***	0.024	−3.85

p < .05. **p < .01 *** p < .001.

Table 4:

Results of Cox Regression Predicting the Likelihood of Rearrest for Violations of a Protective or Restraining Order

	Hazard Ratio	SE	z
Gender (female = 0, male = 1)	2.160***	0.273	6.10
Age (15-35 = 0, 36-91 = 1)	0.703**	0.082	−3.04
Ethnicity (minority = 0, White = 1)	0.605***	0.068	−4.50
Victim (child = 0, partner = 1)	1.623***	0.197	4.00
Risk level	1.824***	0.085	12.83
Age × Risk	1.021	0.035	0.60
Ethnicity × Risk	1.214***	0.041	5.80
Partner victim × Risk	0.973	0.036	−0.74
Gender × Risk	0.855***	0.034	−4.01

p < .05. **p < .01. ***p < .001.

The interactions were expressed as the cross product of the risk level (ranging from 1 representing lowest risk to 4 representing highest risk) and the categorical variables of gender or ethnicity. The results suggested that if risk is increased by one unit, the perpetrator is male (gender = 1), and all other variables are held constant, the hazard ratio is equal to exp (−.104 + .435) = 1.392. Thus, the rate of rearrest is increased by 39.2% for males with each additional unit of risk. The interaction implies substantial moderation of the effect of risk level by gender. Similarly, if risk is increased by one unit, the perpetrator is non-Latino White (ethnicity = 1), and when all other variables are held constant, the hazard ratio is equal to exp (.111 + −.342) = .794. Thus, the rate of rearrest is decreased by 20.6% with an increase in risk for non-Latino White perpetrators. This interaction effect represented a narrowing between ethnic groups in rate of rearrest when accounting for risk. These moderation effects of risk level by gender and ethnicity also held when examining the likelihood of rearrest for order violations.

Survival Curves by Gender and Ethnicity

The KM estimator examined the cumulative probability of rearrest for NFVO and order violations. Figure 2 shows the KM survivor curves for these two measures of recidivism for female and male perpetrators. Males failed faster and more often than females. Details from life tables confirmed that 95% of females avoided rearrest for NFVO by the end of 100 days, compared with only 93% of males. This gap continued to widen with time, with males failing faster and more frequently through the study period. By the end of the study period, 14,087 males survived (75.97%) compared with 6,898 females (82.37%). Similar percentages held with rearrests for order violations. After 100 days, 95.5% of females remained compared with 92.4% of males, and this gap widened so that at the end of the study period, 91.2% of females survived, compared with 84.5% of males. Although significant gender differences in survival existed, both males and females shared two facts: More than half of the rearrests for order violations occurred within the first 200 days for both genders and almost half of all failures for NFVO occurred within the first 200 days.

Figure 2:

Kaplan–Meier Survival Curve for NFVO and VPO/VRO

When additional analyses examined the survival curves of males and females by risk level, the disparities decreased as risk level increased (Figure 3). That is to say, among the lowest risk category of perpetrators, men fail with a rearrest for a NFVO faster and more frequently than their female counterparts. Among high-risk perpetrators (with a DVSI-R total score of 13-28), males and females were more similar in their recidivism for NFVO. A slightly different picture was revealed when examining order violations in that the lowest risk category of male perpetrators fail faster than both of the two lowest risk female categories. Among higher risk categories, the highest risk males fail significantly faster and more often than the highest risk females.

Figure 3:

Kaplan–Meier Survival Curves for NFVO and VPO/VRO by Gender and Risk Level

Figure 4 shows the KM survival curves for NFVO and order violations for non-Latino Whites as compared with ethnic minority perpetrators. For NFVO, non-Latino White perpetrators and minority perpetrators failed at even rates across the first 200 days. Specifically, 90.8% of minorities survived the end of the 200th day, compared with 90.9% of non-Latino Whites. The two ethnic groups began to diverge significantly shortly thereafter, leaving a higher percentage of non-Latino White perpetrators without NFVO. The results by ethnicity were even more interesting with order violations. Non-Latino Whites actually failed faster than ethnic minority perpetrators during the first 200 days, with 9.2% rearrested, compared with just 8% of minorities. After 200 days, however, non-Latino White violations steadied more so than ethnic minorities, whose likelihood of rearrest surpassed that of their non-Latino White counterparts. Survival curves for NFVO categorized by both ethnicity and risk (Figure 5) revealed a more complete picture. Specifically, as was the case for gender differences, the disparities decreased as risk level increased. That is to say, among high-risk perpetrators (with a DVSI-R total score of 13-28), ethnic minorities and non-Latino Whites were more similar in their recidivism of NFVO. Furthermore, high-risk non-Latino Whites and ethnic minorities not only merged but also actually recidivated faster than high-risk minorities. These same patterns also emerged with order violations. Among higher risk perpetrators, the two groups not only merged, but high-risk non-Latino Whites also failed faster and more often than high-risk minorities.

Figure 4:

Kaplan–Meier Survival Curve for NFVO and Violations of a Protective or Restraining Order

Figure 5:

Kaplan–Meier Survival Curves for NFVO and VPO/VRO by Ethnicity and Risk Level

Discussion

Previous research on the predictive validity of the DVSI-R was unable to provide time-specific information. This study addressed that issue through a survival analysis involving all perpetrators arrested on family violence charges and assessed in Connecticut during 2010. Once assessed for risk of persistence in such behavior, perpetrators were followed for 18 months, until the end of May 2012. The analysis focused on two refined measures of recidivism and determined whether perpetrator characteristics moderated the effects of the DVSI-R on recidivism.

In the first stage of the analysis, the DVSI-R performed similar to previous research in its predictive validity for both measures of family violence reoffending, providing further supportive evidence of its predictive validity. Of primary interest in this study, the survival analysis added a time-specific dimension of recidivism beyond previous studies. The survival analyses of the total sample were consistent with prior survival analyses (Klein, Wilson, Crowe, & DeMichele, 2005; Rice & Harris, 1997) in that risk of rearrest was greatest in the early time at risk. Specifically, the results showed a heightened risk within the first 200 days in which rearrests for half of NFVO occurred. Despite a deceleration in rearrest for NFVO after the first 200 days, the steady downward trend of the survival curve beyond this point showed that those rearrested on family violence charges continued to show significant risk over at least the first 800 days. The risk of rearrest for order violations after 200 days was less appreciable, evidenced by a more sudden leveling of the survival curve after 200 days.

The survival analysis provided significant evidence that male and ethnic minority perpetrators were at higher risk of being rearrested for NFVO and order violations, compared with females and non-Latino White perpetrators, respectively. The results further showed, however, interaction with risk level. For both gender and ethnicity, the disparities decreased as risk level increased. In general, men were still far more likely to be rearrested than women, a finding consistent with extant prior literature on domestic violence recidivism (Maxwell, Garner, & Fagan, 2002; Menard, Anderson, & Godbolt, 2009). Importantly, however, among high-risk perpetrators (with a DVSI-R total score of 13-28), males and females were similar in their recidivism of NFVO. A recent study of intimate partner violence recidivism by Menard and colleagues (2009) not only found significant gender differences in their predictive models, but they also found significant overlap. For example, comparing across gender models, the authors found that a history of drug problems had the largest effect for men, followed by race, a history of probation and parole, and unemployment. Variables of race and drug use were also highly predictive of recidivism for women, although they were both less important than prior termination of the relationship. These results were interpreted as evidence that women who do recidivate may be more like their male counterparts, without forgetting that men recidivate at a much higher level. Contrary to research suggesting gendered pathways to recidivism (Daly & Chesney-Lind, 1988; Rettinger & Andrews, 2010), the results of the present study suggested that future research might benefit from examining whether such pathways might be limited to lower risk offenders.

Regarding the predictive effects of ethnicity, higher survival rates were found for non-Latino Whites in the sample, similar to the findings of other studies examining racial and ethnic differences in recidivism (Bachman & Coker, 1995; Kingsnorth, 2006; Maxwell et al., 2002). The use of survival analytical techniques to examine rearrest by ethnicity illuminated an interesting picture. The results showed that ethnic minorities did not depart significantly from non-Latino Whites in their recidivism during the first 200 days, when the risk for recidivism for both groups was high. Rather, toward the end of the year after the initial intake date when non-Latino Whites were slowing in their failure rate, the ethnic minority failure rate became higher. This finding possibly represented potential evidence of the impact of community differences between non-Latino Whites and ethnic minorities. Further research might investigate whether the long-term impact of conditions of community disadvantage, more likely faced by racial and ethnic minorities (Peterson & Krivo, 2005; Quillian, 2012) make rehabilitation and desistance from family violence increasingly more difficult for these groups. With the exception of gender and ethnic differences in the empirical relation between the DVSI-R risk levels and rearrests for order violations and NFVO, no evidence was found that the DVSI-R differentially predicted recidivism depending on the age characteristics of the perpetrator or the type of household relationship involved in the family violence (e.g., partner as victim).

Whether or not the analyses would produce similar results in other states may depend on a range of bureaucratic and legal differences in the processing of family violence cases between states. Findings from a study of the Rhode Island probation model might provide some insights. In a study of 552 misdemeanor family violence perpetrators, Klein et al. (2005) found that more than one third of the sample was rearrested for NFVO before they had been on probation for 2 months. Interestingly, however, the amount of time it took for most perpetrators to report to probation after disposition was about 2 months, highlighting the lack of opportunity for intervention during that time.

Beyond questions of generalizeability, other limitations to the current study included limiting analyses to certain perpetrator characteristics. This research was an initial time-specific analysis of some potential moderators of the empirical relation between the DVSI-R and family violence recidivism, using survival analytical techniques. Future analyses may expand the literature by considering other characteristics of perpetrators, variations in criminal charges, and recommendations made by the evaluators. After all, the ability of professionals responsible for addressing the problem of such violence to link valid risk assessments to well-crafted strategies of supervision and treatment is the most important element, so that the victimized or other potential victims are protected and perpetrators are held accountable for their actions.

In conclusion, the present study provided additional evidence in support of the DVSI-R in assessing risk among family violence perpetrators. It also suggested that risk is greater within the first 200 days after intake, especially for ethnic minority and male perpetrators. However, the differences by gender and ethnicity diminish among higher risk categories of perpetrators. This is an important finding, suggesting that the interaction effects occurred at the lower end of risk levels. Among lower risk perpetrators, ethnic minorities and males were significantly more likely to be rearrested and to fail faster than their non-Latino White and female counterparts do. Differences based on the gender and ethnicity of perpetrators diminished as the risk level increased. Higher risk perpetrators continued to be of the highest risk of reoffending, regardless of their gender, race, or ethnicity. These findings thus added complexity to the literature stressing the need for gender-specific and ethnicity-specific instruments (e.g., Reisig, Holtfreter, & Morash, 2006): It may depend on the level of risk.

With the extensions made to existing predictive validation studies through survival analytical techniques, the results suggested further implications from a public policy perspective. The speed with which reoffending occurred suggests that more intense supervision and more targeted intervention should occur earlier in the at-risk period. Implementation of such supervision and intervention should occur as closely as possible following disposition and assessment.

Footnotes

Appendix

Results of Cox Regression Predicting the Likelihood of Rearrest for a New Family Violence Offense and Violations of Protective and Restraining Orders, Raw Coefficients

NFVO^a	Coefficient	SE	z
Gender	0.435	0.082	5.31
Age_Group	−0.499	0.081	−6.17
Ethnicity	−0.342	0.076	−4.47
Victim	0.164	0.078	2.09
Risk_Level	0.444	0.031	14.39
Age_risk	0.057	0.025	2.28
Ethnicity_risk	0.111	0.024	4.63
Victim_risk	−0.019	0.025	−.76
Gender_risk	−0.104	0.027	−3.85
VPO/VRO^b	Coefficient	SE	z
Gender	0.770	0.126	6.10
Age_group	−0.353	0.116	−3.04
Ethnicity	−0.502	0.112	−4.50
Victim	0.484	0.121	4.00
Risk_level	0.601	0.047	12.83
Age_risk	0.021	0.035	.60
Ethnicity_risk	0.194	0.034	5.80
Victim_risk	−0.027	0.037	−.74
Gender_risk	−0.157	0.039	−4.01

Note. NFVO = new family violence offenses; VPO/VRO = violation of protective or restraining orders.

Rearrests for a new family violence offense only.

Rearrests for violations of protective or restraining orders only.

Authors’ Note:

This research was conducted with assistance from the Judicial Branch, Court Support Services Division (CSSD) of the State of Connecticut. This agency owns the copyright for the Revised Domestic Violence Screening Instrument (DVSI-R). Special appreciation is expressed to William Carbone, Executive Director of CSSD. Appreciation is also expressed (in alphabetical order) to Kathy Ceruti, Joe DiTunno, Steve Grant, Brian Hill, Deb Kulak, and Brian Sperry. The authors also gratefully acknowledge the anonymous reviewers and the editor for their feedback on previous drafts of this manuscript.

Richard Stansfield is a doctoral candidate at the University of Delaware. His research interests include structural predictors of crime, family and intimate partner violence, homicide, and race, ethnicity, and immigration. His recent work has appeared in Social Science Research, Sociological Perspectives, the Journal of Urban Affairs, and Homicide Studies.

Kirk R. Williams is the chair of the Department of Sociology and Criminal Justice at the University of Delaware. He has published widely on the causes and prevention of violence, particularly involving youth or adult intimate partners, with the most recent publications addressing bullying, juvenile offending, and domestic violence risk assessment. He has received numerous grants from federal and state funding sources, in addition to private foundations to support his research. He also has worked extensively with community-based groups, schools, and agencies in violence prevention planning, implementation, and evaluation.

References

Allison

(1995). Survival analysis using the SAS system: A practical guide. Cary, NC: SAS Institute.

Bachman

Coker

A. L.

(1995). Police involvement in domestic violence: The interactive effects of victim injury, offender’s history of violence, and race. Violence and Victims, 10, 91-106.

Chan

K. L.

(2012). Predicting the risk of intimate partner violence: The Chinese Risk Assessment Tool for victims. Journal of Family Violence, 27, 157-164.

Cleves

Gould

Gutierrez

Marchenko

(2008). An introduction to survival analysis using Stata. College Station, TX: Stata Press.

Cox

D. R.

(1972). Regression models and life-tables. Journal of the Royal Statistical Society: Series B, 34, 187-220.

Daly

Chesney-Lind

(1988). Feminism and criminology. Justice Quarterly, 5, 497-538.

DeJong

(1997). Survival analysis and specific deterrence: Integrating theoretical and empirical models of recidivism. Criminology, 35, 561-576.

Dolan

Doyle

(2000). Violence risk prediction: Clinical and actuarial measures and the role of the Psychopathy Checklist. The British Journal of Psychiatry, 177, 303-311.

Dutton

D. G.

Kropp

P. R.

(2000). A review of domestic violence risk instruments. Trauma, Violence, & Abuse, 1, 171-181. doi:10.1177/1524838000001002004

10.

Goodman

L. A.

Dutton

M. A.

Bennett

(2000). Predicting repeat abuse among arrested batterers: Use of the Danger Assessment Scale in the criminal justice system. Journal of Interpersonal Violence, 15, 63-74. doi:10.1177/088626000015001005

11.

Heagerty

P. J.

Zheng

(2005). Survival model predictive accuracy and ROC curves. Biometrics, 61(1), 92-105.

12.

Heckert

D. A.

Gondolf

E. W.

(2004). Battered women’s perceptions of risk versus risk factors and instruments in predicting repeat reassault. Journal of Interpersonal Violence, 19, 778-800.

13.

Hilton

N. Z.

Carter

A. M.

Harris

G. T.

Sharpe

A. J.

(2008). Does using non-numerical terms to describe risk aid violence risk communication? Clinician preference, agreement, and decision-making. Journal of Interpersonal Violence, 23, 171-188. doi:10.1177/0886260507309337

14.

Hilton

N. Z.

Harris

G. T.

(2005). Predicting recidivism among serious wife assaulters: A critical review and implications for policy and practice. Trauma, Violence, & Abuse, 6, 3-23. doi:10.1177/1524838004272463

15.

Hilton

N. Z.

Harris

G. T.

(2009). How nonrecidivism affects predictive accuracy: Evidence from a cross-validation of the Ontario Domestic Assault Risk Assessment (ODARA). Journal of Interpersonal Violence, 24, 326-337. doi:10.1177/0886260508316478

16.

Hilton

N. Z.

Harris

G. T.

Rice

M. R.

(2010). Risk assessment for domestically violent men: Tools for criminal justice, offender intervention, and victim services. Washington, DC: American Psychological Association. doi:10.1037/12066–000

17.

Hirschel

Buzawa

Pattavina

Faggiani

Reuland

(2007). Explaining the prevalence, context, and consequences of dual arrests in intimate partner cases. Washington, DC: U.S Department of Justice.

18.

Kaplan

E. L.

Meier

(1958). Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53, 457-481.

19.

Kingsnorth

(2006). Intimate partner violence: Predictors of recidivism in a sample of arrestees. Violence Against Women, 12, 917-935.

20.

Klein

A. R.

Wilson

Crowe

A. H.

DeMichele

(2005). Evaluation of the Rhode Island Probation Specialized Domestic Violence Supervision Unit. Washington, DC: National Institute of Justice, Office of Justice Programs.

21.

Lusted

L. B.

(1971). Signal detectability and medical decision-making. Science, 171, 1217-1219.

22.

Maxwell

C. D.

Garner

J. H.

Fagan

J. A.

(2002). The preventive effects of arrest on intimate partner violence: Research, policy, and theory. Criminology & Public Policy, 2, 51-80.

23.

McNeil

D. E.

Binder

R. L.

(2007). Effectiveness of a mental health court in reducing criminal recidivism and violence. American Journal of Psychiatry, 164, 1395-1403.

24.

Menard

K. S.

Anderson

A. L.

Godbolt

S. M.

(2009). Gender differences in intimate partner recidivism: A 5 year follow-up. Criminal Justice and Behavior, 36, 61-76.

25.

Mossman

(1994). Assessing predictions of violence: Being accurate about accuracy. Journal of Consulting and Clinical Psychology, 62, 783-792.

26.

Peterson

R. D.

Krivo

L. J.

(2005). Macrostructural analyses of race, ethnicity, and violent crime: Recent lessons and new directions for research. Annual Review of Sociology, 31, 331-356.

27.

Quillian

(2012). Segregation and poverty concentration: The role of three segregations. American Sociological Review, 77, 354-379.

28.

Quinsey

V. L.

Harris

G. T.

Rice

M. E.

Cormier

M. A.

(2006). Violent perpetrators: Appraising and managing risk. Washington, DC: American Psychological Association. doi:10.1037/11367–000

29.

Reisig

M. D.

Holtfreter

Morash

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

30.

Rettinger

L. J.

Andrews

D. A.

(2010). General risk and need, gender specificity, and the recidivism of female offenders. Criminal Justice and Behavior, 37, 29-46.

31.

Rice

M. E.

Harris

G. T.

(1997). Cross-validation and extension of the violence risk appraisal guide for child molesters and rapists. Law and Human Behavior, 21, 231-241.

32.

Simpson

S. S.

Bouffard

L. A.

Garner

Hickman

(2006). The influence of legal reform on the probability of arrest in domestic violence cases. Justice Quarterly, 23, 297-316. doi:10.1080/07418820600869087

33.

Smith

P. H.

White

J. W.

Holland

L. J.

(2003). A longitudinal perspective on dating violence among adolescent and college-age women. American Journal of Public Health, 93, 1104-1109.

34.

UCLA Statistical Consulting Group. (2011). Statistical computing seminars: Survival Analysis with Stata. Retrieved from http://www.ats.ucla.edu/stat/stata/seminars/stata_survival/

35.

Williams

K. R.

(2008). Domestic Violence Screening Instrument (DVSI). In Cutler

B. L.

(Ed.), Encyclopedia of psychology and law (pp. 240-242). Newbury Park, CA: SAGE.

36.

Williams

K. R.

(2011). Family violence risk assessment: A predictive cross-validation study of the Domestic Violence Screening Instrument-revised (DVSI-R). Law and Human Behavior. Advance online publication. doi:10.1007/s10979-011-9272-6

37.

Williams

K. R.

Grant

S. R.

(2006). Empirically examining the risk of intimate partner violence: The revised Domestic Violence Screening Instrument (DVSI-R). Public Health Reports, 121, 400-408.

38.

Williams

K. R.

Houghton

(2004). Assessing the risk of domestic violence re-offending: A validation study. Law and Human Behavior, 28, 437-455. doi:10.1023/B:LAHU.0000039334.59297.f0

39.

Williams

K. R.

Tuthill

Lio

(2008). A portrait of juvenile offending in the United States. In Hoge

R. D.

Guerra

N. G.

Boxer

(Eds.), Treating the juvenile offender (pp. 15-32). New York, NY: Guilford.

40.

Yoshihama

Gillespie

B. W.

(2002). Age adjustment and recall bias in the analysis of domestic violence data: Methodological improvements through the application of survival analysis methods. Journal of Family Violence, 17, 199-222.

41.

Yoshihama

Horrocks

(2003). The relationship between intimate partner violence and PTSD: An application of Cox Regression with time-varying covariates. Journal of Traumatic Stress, 16, 371-380.