Study 1
Rahman and Poynton (2018) used a quasi-experimental design to assess the effect of the EQUIPS Domestic Abuse Program on both general and domestic violence recidivism after 12 months for men in New South Wales, Australia.
The cohort of interest was all males who had committed an offense and were referred to the program in 2015. Details for all individuals who were referred to the program in 2015 were extracted from the Corrective Services of New South Wales Offender Information Management System, and additional data were provided on all individuals referred to the program between 2008 and 2014. As individuals could be referred more than once between 2008 and 2015, each individual’s earliest referral to the EQUIPS Domestic Abuse Program was selected. This resulted in a sample of 1,312 unique individuals referred in 2015. This dataset was then linked to the New South Wales Bureau of Crime Statistics Research Re-Offending Database to obtain demographic, offending, and reoffending data, and information on all court appearances finalized by March 31, 2017. The index appearance was identified by choosing the closest index appearance to the earliest sentence date on all active supervised orders that the individual was on at the time of referral to the program.
The treatment group comprised individuals who had committed an offense and who started the program within 12 months of referral (n = 487). The CrimeSolutions review of this study focused on outcomes for individuals who started the program within six months of referral, where starting was defined as completing four or more sessions. The treatment group received the EQUIPS Domestic Abuse Program as described in the Program Description. Individuals who were referred but did not start the program within 12 months comprised the control group (n = 786).
An instrumental variables approach was used to separate the effect of treatment on offending from the effect of other factors that were correlated with treatment participation. The instrumental variable reflected the EQUIPS Domestic Abuse Program commencement rate at the same time and location at which an individual was referred to the program (referred to as the peer non-commencement rate). This measured the availability of other eligible individuals to start the program at the time of a participant’s referral to the program. Location fixed effects (which would be relatively stable over the time period examined) were also included in the instrumental variable models to account for confounding effects. A range of variables that could potentially influence treatment allocation (referral to the program) and/or the likelihood of reoffending were considered for inclusion in the instrumental variable model. These included characteristics of the individuals, including age; Indigenous status; socioeconomic disadvantage categorized into quartiles (quartile 1 was the most disadvantaged, and quartile 4 was the least disadvantaged); remoteness of residence of the individual (categorized into major cities, inner regional, outer regional or remote or very remote, or unknown); and the Level of Service Inventory-Revised score in the most recent risk category (low or medium-low, medium, medium-high or high, or unknown). Other control variables included the characteristics of index finalization, prior criminal history, and characteristics of the participants’ active supervised orders.
Men in the treatment group were mostly in the 25–34 age group (38.6 percent) and the 35–44 age group (29.6 percent), had non-Indigenous or unknown status (63.2 percent) and scored in the medium-risk category on the Level of Service Inventory-Revised (62.4 percent). In terms of socioeconomic disadvantage, 29 percent of men in the treatment group were quartile 1, most disadvantaged; and 34.9 percent were quartile 2, disadvantaged. Men in the treatment group had an average of 3.62 proven concurrent charges at index contact (including principal offense). The control group men were mostly in the 25–34 age group (37.3 percent) and the 35-44 age group (30.5 percent), had non-Indigenous or unknown status (60.3 percent) and scored in the medium-risk category on the Level of Service Inventory-Revised (56.6 percent). In terms of socioeconomic disadvantage, 31.9 percent of men in the control group were in quartile 1, most disadvantaged; and 33.6 percent were in quartile 2, disadvantaged. Men in the control group had an average of 4.05 proven concurrent charges at index contact (including principal offense). There were statistically significant differences between groups on demographic characteristics (age, socioeconomic disadvantage, and remoteness of residence), Level of Service Inventory-Revised risk category, the number of proven concurrent charges at index contact, the time remaining on supervised orders, the number of finalized court appearances (with proven offense/s) as a juvenile or adult, and the penalties received in the five years prior to finalization. These differences were controlled for in the analysis.
Recidivism was the outcome examined and was defined as any new proven offense or a proven domestic violence-related offense committed after the program referral date and prior to December 31, 2016, which allowed for a minimum of 12 months of follow up after the recorded referral date for all individuals in the sample and a 3-month lag for new offenses to be finalized in court. The referral date was used as the starting date for measuring reoffending as it was available for individuals in both the treatment and control groups.
The models were estimated as follows: first, a two-stage least-squares estimation method was used for the instrumental variables analysis. The instrumental variable (the number of program starters, by peer non-commencement rate) was included in a first-stage linear regression (the treatment equation), to obtain the predicted probability of starting the program for those in the referral group. These predicted probabilities were then used in place of starting the program in a second-stage linear regression predicting reoffending (the outcome equation). A recursive bivariate probit approach was also used. Both models were estimated using clustered standard errors for the community correction offices from which individuals were referred, to take into account any possible correlation in reoffending rates within sites. No subgroup analysis was conducted.
Study 2
Blatch and colleagues (2016) conducted a quasi-experimental design with propensity score matching to assess the effect of the Corrective Services of New South Wales’s Domestic Abuse Program on time to reconviction and reconviction rates for participating men after approximately 33 months.
All individuals were eligible for treatment if they were referred to the Domestic Abuse Program at a Corrective Services of New South Wales community corrections office from October 1, 2007, until June 30, 2010 (known as the “census date,” the final date for observation of potential reoffending). The treatment group of 953 program participants was identified from the Corrective Services of New South Wales electronic Offender Information Management System, and a pool of potential control group individuals was also extracted from this database using the following criteria: considered eligible for referral; identified in case notes as having domestic abuse or violence as a treatment need due to past behaviors, convictions, or family reports; had not been previously enrolled in a domestic abuse program; and served a community supervision or parole order between July 1, 2007, and the census date of June 30, 2010. The final pool was 5,734 potentially eligible controls for the propensity score matching procedure. Risk factors empirically associated with reoffending were modeled using binary logistic regression to predict membership of the treatment group. Nine significant risk factors were identified, including type of order (community service, parole, supervision); most serious offense committed (other, assault or homicide, property damage, regulatory offenses/order breaches); custodial sentences and conviction counts in the previous five years; and Aboriginal and Torres Strait Islander status. Using a stepwise logistic regression model, these risk factors, including age, were used to calculate a single propensity score (between 0 and 1) for each participant. A nearest neighbor, one-to-one matching procedure without replacement was employed. The list of treated participants was sorted randomly and matched in sequence to the control group member with a propensity score closest to their own, leaving a final matched cohort of 953 in each group. The treatment group received the EQUIPS Domestic Abuse Program as described in the Program Description, and the control group did not participate in the program during the study period.
In the matched sample, men in the treatment group were an average of 33.4 years old, had a mean total Level of Service Inventory-Revised score of 23.8, and 19 percent were Aboriginal and Torres Strait Islander. Men in the treatment group had an average of 0.66 times in custody in the last five years, 76 percent had an assault or homicide as their most serious offense, and 81 percent were on supervision for their order type. Men in the control group were an average of 33.5 years old, had a mean total Level of Service Inventory-Revised score of 23.8, and 19 percent were Aboriginal and Torres Strait Islander. Men in the control group had an average of 0.65 times in custody over the last five years, 76 percent had an assault or homicide as their most serious offense, and 80 percent were on supervision for their order type. After matching, there were no statistically significant differences between the treatment and control groups.
Prior convictions and reconviction data for all potential study participants, before matching, came from the New South Wales Bureau of Crime Statistics and Research Re-Offending database, which recorded details of all convictions and sentences. Reconvictions were counted up to December 31, 2010, allowing six months after the study census date of June 30, 2010, for any outstanding charges to be finalized. The “follow-up start date” was defined as the start date of the supervised order, associated with the most serious offense active at the time of Domestic Abuse Program enrollment. For individuals on supervised parole, the release date from custody defined their follow-up start date. For men in the control group, the follow-up start date was also the start date of the supervision order during which they were identified as having a domestic violence treatment need from their Offender Information Management System case management plan. “Days at risk” were calculated from an individual’s follow-up start date to the study census date, subtracting any days spent in custody or police cells, to provide an accurate measure of the number of days free in the community for any reoffending to occur. Finalized convictions for each individual, from their order start date up to December 30, 2010, were totaled to determine the reconviction rate. If several charges were finalized on the same date, only the most serious offense was recorded. Some of the treatment cohort reoffended before receiving treatment, and these reconvictions were included to ensure equivalent follow-up start date definitions for both the treatment and control groups.
An intent-to-treat design was used in the analysis. Cox regressions and one Poisson regression model, fitted with risk factors associated with reoffending fitted as covariates, were run. These included Level of Service Inventory-Revised risk scores; age (at start of order); most serious offense of current order; Aboriginal or Torres Strait Islander status; and most serious offense and number of incarcerations in the five years prior to the order date, with concurrent sentences counted as one custodial episode. Cox regressions were used to compare survival time to first general reconviction (which included all types of crime) and first violent reconviction (violent offenses were classified using the Australian Standard Offence Classification Codes and included homicide, assaults against a person, sexual assault, or violence order breaches). Poisson regression compared the number of finalized reconvictions (most serious offense only) in the follow-up period, adjusted for time at risk, with those of controls. Poisson regression analysis was also conducted for time to first reconviction by program completion status.
Study 3
Blatch and colleagues (2020) used the full dataset from Study 2 (Blatch et al. 2016) with a retrospective quasi-experimental design, to assess the effectiveness of the Domestic Abuse Program on reconviction rates. While this study aimed to examine outcomes specifically for the subgroup of Australian Indigenous men, the authors included results for the full sample, which was the focus for this CrimeSolutions review.
The methodology was the same as described in Study 2 (Blatch et al. 2016). In the propensity score matching procedure, 10 significant factors predicted membership in the Domestic Abuse Program treatment group: Level of Service Inventory-Revised risk score, type of supervision order (community service, parole, or supervision), most serious offense type (other, assault or homicide, property damage, regulatory offenses/order breaches), Indigenous status, and the number of custodial sentences in the five years preceding the index offense. Using stepwise logistic regression, these risk factors were used to calculate a single propensity score (between 0 and 1) for the 957 men in the treatment group and all potential (untreated) controls (n = 5,734), resulting in a matched sample of 953 in each group. As in Study 2, after matching there were no statistically significant differences between the treatment and control groups.
The outcome of interest was reconvictions, summed from the index offense date for participants on community orders, or the custody release date for participants released on supervised parole, up to December 31, 2010, six months after the final observation date of June 30, 2010, to allow offenses committed before that date to be prosecuted and finalized. Treated participants’ reconvictions before the program start date were included to ensure equivalent reconviction start dates for the treatment (which included nonstarters and dropouts) and control groups. Cohen’s d determined effects size for the full treatment and control groups, using standardized mean reconviction differences. A subgroup analysis was conducted by Australian Indigenous status.
