We used existing data from a cohort study of male perpetrators with female partners identified through the Seattle Police Department Domestic Violence Unit database. The original cohort study was designed to assess the impact of criminal justice system responses on IPV recidivism and health outcomes. Eligible incidents were those in which: (1) the Seattle Police responded to a call within the Seattle city limits during the study period of January 11,999 through December 31, 2001; (2) the call resulted in a police incident report indicating IPV in a current or former heterosexual partnership with the male partner identified as the perpetrator of abuse; and (3) only the male was identified as the perpetrator of abuse. The first police-reported IPV incident that met these criteria during the recruitment period was classified as the index incident.
A total of 6266 male/female partners met the initial inclusion criteria. We further restricted our study population to include only those with an arrestable offense at the index incident. Arrestable offenses as defined by the Washington State laws included: assault, burglary, criminal trespass, custodial interference, harassment, menacing, property damage, reckless endangerment, stalking, theft, threats and violation of court orders (Washington State Legislature 2016). Thus, we excluded 800 incidents with index offenses classified as “disturbance,” “suspicious circumstances,” “outstanding warrants” or “other offense.” The final sample consisted of 5466 former or current intimate partner couples. Perpetrator and survivor names and dates of birth were used to identify Seattle Police Department Domestic Violence Unit records for historical and recidivistic IPV incidents.
We defined recurrent IPV, the outcome of interest, as a police-reported IPV incident between the index couple in the 12 months post-index event and classified each event as physically abusive (with or without psychologically abusive behavior) or psychologically abusive behavior only using primary and secondary offense codes for the incident. Physical IPV recurrent events were incidents with offense code(s) for assault, reckless endangerment or unlawful imprisonment. Psychological IPV only recurrent events were incidents with offense code(s) for harassment, menacing, stalking, threats, disturbance, criminal trespass, custodial interference, interfering with IPV reporting, or property damage and no recorded physical IPV offense codes. Index incidents that included sexual abuse were referred to the Seattle Police Department Sexual Assault Unit rather than the Domestic Violence Unit and were unavailable for study. Arrest of the perpetrator for the index incident served as the exposure of interest, and was obtained from police reports, as were all demographic, index event characteristics and historical IPV incident history between each index couple for the 12 months prior to the index incident. Race of the perpetrator and survivor were assessed by police officers and recorded in the incident report as either: White, African American/Black, Asian/Pacific Islander, Hispanic, American Indian/Alaskan Native or Unknown. Since racial disparities in police response have been documented in prior literature, race was included as a covariate as it may play a role in police decisions to arrest (Hepburn 1978).
Multivariable survival analyses were used to quantify the association between arrest at the index event and the risk of police reported IPV incidents following the index incident. Arrest was coded as a time-dependent variable, with a perpetrator considered exposed only from the day of arrest onward. Analyses were conducted for the entire 12-month follow-up. Psychological and physical abuse were analyzed as separate outcomes.
We studied time to first recurrent IPV event and multiple recurrent IPV events (failures) separately. For the time to first recurrent IPV event, we used multivariable extended Cox regression analysis and allowed couples to be at risk for a first recurrence until the end of the analysis period or until there was a recurrent police-reported IPV incident (physical and psychological IPV recurrence were analyzed separately). In analyzing physical abuse outcomes, events comprising psychological abuse only were ignored rather than treated as competing events as the observation of psychological IPV does not preclude the observation of physical IPV. Models for psychological IPV likewise ignored physical IPV outcomes.
Our interest in characterizing the relative difference in the instantaneous risk of a recurrent event between subpopulations with varying number of prior IPV events suggested the use of a multiplicative intensity-based model. Alternative models have been proposed for the study of recurrent events in survival analysis, including additive regression models (Amorim and Cai 2015; Ullah et al. 2014; Schaubel et al. 2006; Cook and Lawless 2007). A model class may be deemed appropriate in the context of a given application if 1) its form is consistent with what the available science suggests about the underlying process under study (e.g. are successive events of the same type, and is their order relevant?), and 2) model coefficients can be easily interpreted on the desired scale (e.g. multiplicative versus additive contrasts) (Amorim and Cai 2015; Cook and Lawless 2007). The specific assumptions, strengths and limitations of various models available in the literature have been described in detail elsewhere (Amorim and Cai 2015; Cook and Lawless 2007; StataCorp 2015a).
Based on our knowledge and understanding of IPV perpetration, we selected the gap-time Prentice, Williams and Peterson (PWP) model as we believed IPV recidivism to be best modeled with ordered events of the same type, with time reset to zero following each event to capture the potential acceleration of IPV perpetration. The PWP model incorporates stratification based on couples’ history (number) of prior IPV events. This approach allows the baseline risk of recurrence to differ arbitrarily based on number of prior events since the index IPV event as a function of time (Prentice et al. 1982). The PWP model can be implemented using either total time, where time to each event is measured from the study baseline, irrespective of the timing of prior events, or gap time, which resets the time scale after each recurrent event. In our study, information on arrests at subsequent police-reported IPV was not available. However, had this information been available, it could have easily been incorporated into the PWP model as a time-varying covariate (e.g., as a cumulative count of arrests at prior IPV reports, or as a binary variable indicating if the latest police-reported IPV event resulted in arrest).
This model postulates that the recurrence hazard ratio (HR) comparing instantaneous risk of recurrence at any given time since the previous event in couples with and without an arrest at the index event, but otherwise similar characteristics, to be constant and common across risk set groups. This allowed the calculation of a single, overall recurrence hazard ratio (HR) for the exposure of interest.
In order to allow the HR to vary based on risk set groups defined by prior number of IPV events since baseline, we also fit a more flexible PWP model including interaction terms between arrest status and a trichotomization of prior number of events (see Appendix for statistical models used). Using this model, we estimated a HR for recurrence of IPV events separately for couples having experienced zero, one, or at least two, three or four or more follow-up IPV events. Since only 2.5% of all couples had more than four follow-up events, risk set groups were not refined further to avoid unreliable estimates. A test of null interaction in the latter PWP model suggested that these risk set group-specific HRs differed significantly. We also examined effect modification by whether the perpetrator was contacted by the police at the index incident by testing the significance of the inclusion of an interaction term between arrest and police contact. The interaction term was not significant; consequently, results are not presented separately by police contact status.
All analyses were adjusted for confounders determined a priori based on associations established in existing literature: index IPV abuse type, cohabitation, weapon use, police reported survivor injury at the index incident and police-recorded perpetrator and survivor race (white, African-American or Other) (Black et al. 2011; Hepburn 1978; Frantzen et al. 2011; Sanchez-Lorente et al. 2012; Abramsky et al. 2011; Campbell et al. 2003). Stata 14 was used for all analyses presented using the st suite of commands, although we additionally conducted our analyses in R to verify results (StataCorp 2015b).