Our study protocol is described elsewhere [13]. Briefly, the isafe study consisted of a two-arm parallel RCT of a Web-based safety decision aid for women experiencing IPV. Efficacy of the decision aid was assessed primarily using participants’ exposure to IPV and mental health after 12 months of access to the decision aid. Participation was open to English-speaking women aged 16 years or older in New Zealand who reported exposure to current IPV. They also needed access to a safe email address (to which only they had password-protected access) to send and receive study-related information. The trial was approved by the Auckland University of Technology Ethics Committee (reference number 12/51).

Complete details regarding participant recruitment and engagement can be found elsewhere [14]. Participants enrolled in the study by accessing a secure New Zealand registration website; the most common referral source was a Web classified advertisement. A participant was considered enrolled once they had visited the registration website and (1) met eligibility criteria, (2) consented to participate, (3) provided contact information, and (4) passed a validation check. The validation check to reduce the risk of fraudulent enrollment involved automated matching against the New Zealand electoral roll (based on name and address) or a manual process of logic checking against information available from Web searching.

Once enrolled, participants were randomized by software to the control or intervention arm. Randomization was based on a minimization scheme. Two stratification factors (severity of violence and children) and two random factors each with two equiprobable levels were used to achieve the minimization. Severity of violence factor was dichotomous, based on one positive response to the IPV eligibility items versus two or more. The children factor was also dichotomous, identifying whether the participant had one or more children versus none.

After randomization, participants were sent an email with a username, password, and URL access for the website. The participant’s allocation was kept secret from herself and the study team in New Zealand. There was no procedural difference between arms until after baseline measures were obtained. After baseline measures collection, participants in different arms were exposed to different screen contents, corresponding to the control and active interventions described below (see eg, Figure 1).

Women randomly assigned to the intervention group were able to access a Web-based safety decision aid throughout the 1-year postbaseline follow-up period via the secure password-protected trial website. The decision aid intervention included three components. The first component was a safety priority setting activity based on a multicriteria decision model developed in the United States [8]. The five criteria (priorities) were evaluated for the New Zealand context [12] and minor wording changes made to the descriptions [13]. Women moved a sliding bar toward the priority that was most important to them for all pairwise combinations. Through a series of matrix computations using the analytic hierarchy process [15], the program provided feedback to the participant summarizing her priorities. The second component was the Danger Assessment (DA) [16] or Danger Assessment-Revised (DA-R; for female same-sex relationships) [17]. Women completed the DA or DA-R and received immediate scored feedback on their level of danger for severe or lethal violence in the intimate relationship, ranging from variable danger to extreme danger. The third and final component was an interactive process using an underlying matrix of resources to help women develop an individually tailored action plan. The matrix included local, regional, and national resources; and tips about safety for the participant and her children based on her safety priorities and DA or DA-R scores.

Women randomly assigned to the control group were able to access a standardized list of resources and a standardized emergency safety plan throughout the 1-year postbaseline follow-up period via the secure password-protected trial website. They were not provided with individualized feedback or tailored action plan.

Participants were followed up for 1 year following completion of their baseline measures. Assessments were scheduled to take place at baseline, then 3, 6, and 12 months after completion of baseline. Women were invited to complete assessments at each time period regardless of prior assessment completions. Retention mechanisms included automated emails sent to participants at 1, 2, 4, 8, and 10 months. If a participant was 6 weeks late for a scheduled follow-up visit, the research team sent email or phone reminders (following the woman’s safety instructions).

Logs were maintained by study personnel documenting phone and email contacts with participants, potential unanticipated events, and other issues affecting the trial (such as server interruptions). Study safety protocols addressed computer safety procedures; confidentiality measures; support and referral procedures in case of violence or distress, including suicidality; responding to calls to the study site by partners or children; and investigation and reporting of unanticipated events. In addition to safety review during data monitoring committee (DMC) meetings, study logs and media reports of serious IPV events were monitored during the trial. The study protocol outlined the safety review process, which for study-related serious adverse events would involve escalation to the principal investigator and chairs of the ethics and data monitoring committees and reporting to the funder and university legal counsel.

The mental health primary outcome consisted of depression measured by the Center for Epidemiologic Studies Depression Scale-Revised (CESD-R) [18-20] at 12 months. The CESD-R score, ranging from 0 to 80, was used. Higher scores are associated with more depression symptomatology, with scores of at least 16 indicative of depression. The IPV exposure primary outcome consisted of the Severity of Violence Against Women Scale (SVAWS) [21] at 12 months. For the SVAWS we used a 1 (never) to 4 (many times) subjective frequency scale for all items; total SVAWS score ranges from 46 to 196, with higher scores associated with greater exposure to IPV. Secondary violence domain outcomes included SVAWS total score at 3 and 6 months, SVAWS subscales (threats of violence, acts of violence, and sexual violence), and Women’s Experience with Battering (WEB) [22]. Secondary mental health domain outcomes included CESD-R score at 3 and 6 months, Post-Traumatic Stress Disorder Checklist-Civilian Version [23,24], Alcohol Use Disorder Identification Test dichotomized [25,26], and Drug Abuse Screening Tool [27,28]. Decisional process secondary outcome was measured using the Decisional Conflict Scale [29], safety-seeking behavior using the Safety Checklist [30,31], and Safety Checklist Helpfulness. Assessment schedules and time points are included in Multimedia Appendix 1.

As described in the published protocol [13], the figure of 340 women accounts for an upper limit of 35% dropout by 12 months based on attrition rates for previous Web-based studies (eg, the Recovery via Internet from Depression trial). Results from a New Zealand trial support the planned study to have 80% power to detect a 37% reduction in depression (CESD-R). Data from an international IPV study indicated that the proposed study would have power superior to 80% to detect a difference of −11.2 in IPV exposure (corresponding to ≈ ½ standard deviation baseline Severity Violence Against Women Scale).

Plans for all inferential analyses were finalized before allocation unblinding in a full statistical analysis plan. Although randomization was technically carried out before baseline assessment, the allocation remained entirely concealed until all baseline information was obtained. Accordingly, the intention-to-treat (ITT) analysis set consisted of all participants who provided data at baseline on a primary outcome. The main analyses were carried out in a modified ITT analysis set, consisting of the ITT analysis set from which participants with no postbaseline assessments were excluded. Sensitivity analyses were carried out in the ITT set to assess the effect of this exclusion on treatment effect estimates [32]. A per-protocol analysis set was also defined, excluding from the modified ITT set any participant identified with an eligibility violation (such as from repeat enrollment) or a major protocol violation (not receiving the intervention) and assigning the participant to the allocation arm corresponding to her actual uptake of the intervention.

All postbaseline assessments for a given outcome were entered in the regression models, initially assumed to be normal with a participant-specific random effect, adjusted for the baseline value, other covariates as described below, and placing the intervention in interaction with the assessment time, taken as a factor. A blind review, absent any information regarding allocation, was undertaken for each outcome for an assessment of missingness, a visual assessment of residual normality, an assessment of the covariance structure of the repeated measures, and an assessment of candidate covariates (ethnicity, children, paid employment, and age group) to include in the final analysis. In a model including all covariates, assessment time but no information regarding allocation, only covariates achieving a partial coefficient of determination larger than 1% were retained to adjust the final model. Subgroup analyses regarding children in their care versus not and Māori ethnicity versus not were planned for the primary outcomes. These proceeded along the same model as above, including a three-way interaction between treatment, assessment time, and subgroup indicator. Although women could select multiple ethnicities, a priori hierarchy prioritized Māori ethnicity.

Missing data at baseline was imputed using the mode of the variable in the observed ITT values [33]. Missingness in both primary outcomes was found to be significantly related to the last observed value of the outcome in question, to the assessment time being at 12 months in the case of CESD-R but to no other baseline variable, indicating that adjusting for baseline and fitting the available repeated measures with a suitable (nondiagonal) covariance structure appropriately removed the risk of bias from missingness under a missing at random assumption [34].

The sensitivity analyses to be carried out in the full ITT set consisted in the production of point estimates only with missing values singly imputed at each assessment time according to four different schemes: (1) ITT extension: missing values because of attrition imputed as the mean of the observed values in the control arm; missing values from nondropouts imputed as the mean of the arm to which the participant was allocated; (2) Return to baseline: all missing values imputed as the baseline value; (3) Worst case for intervention: missing values imputed as the worst intervention arm outcome in the intervention arm, and the best control arm outcome in the control arm; and (4) Best case for intervention: missing values imputed as the best intervention arm outcome in the intervention arm, and the worst control arm outcome in the control arm.

Mediation analyses based on multivariate regression were to be completed only if the intervention proved to be significantly related to the hypothesized mediators at a 0.15 level. All other tests of hypotheses were carried out using a significance level of 0.05 and two-sided alternatives. There was no adjustment for multiple testing. All analyses were carried out in the R software environment, version 3.1 (R Foundation for Statistical Computing, Vienna), using packages lme4 [35] and nlme. An independent DMC was formed and met for the first time in September 2013, thereafter meeting four more times at approximately 6-month intervals. The role of the DMC was detailed in a charter.

The 412 participating women (Figure 2) were typically young (interquartile range [IQR]=23, 36 years) and lived in a main urban area (83%); 45% (185/412) were responsible for the care of one or more children living in the household, 40% (165/412) were in paid employment, 27% of participants self-identified as Māori compared with 12% among women 15 years of age or older living in New Zealanders (2013 census, NZ Stats table), and 30% (124/412) of participants lived in neighborhoods in the highest deprivation quintile compared with the expected 20% (82/412).

At baseline, partners were often living with the participant (227/412, 55%). Women’s decisional conflict about their partner relationship was evident: half of the participants were “unsure” of the future of their relationships, with one-quarter planning to end the relationship and one-quarter planning to remain in the relationship. Most women accessed the isafe study on a computer in their home (259/412, 63%) or the home of a friend or family member (37/412, 9%). The median SVAWS score was 84 (IQR 67, 103). The median CESD-R score was 32 (IQR 18, 50). A large proportion of isafe participants (309/412, 75%) evidenced depressive symptoms (score greater than 16); one in 12 reported suicidal thoughts (“I wished I was dead”) “nearly every day” in the prior 2 weeks.

Baseline characteristics, including violence severity and depression, were similar for women allocated to control or intervention (Table 1). The known inequities of race and socioeconomic deprivation were evident among our study participants: among the 113 Māori participants, 42.5% lived in the highest deprivation quintile; this compared with 25.1% among the 299 non-Maori participants. Although Māori were overrepresented in high deprivation neighborhoods, deprivation was not associated with group or outcome (and thus, not a confounder). Seventy-five percent of participating women completed all three follow-up assessments (73% of women assigned to control group and 76% of women assigned to intervention group). Retention rates at 3, 6, and 12 months were 81%, 83%, and 87%, respectively.

Intervention estimates for primary outcomes across all time periods favored the intervention (Table 2) but were not statistically or clinically significant. The SVAWS 12-month adjusted intervention estimate was −2.47 (95% CI −7.95 to 3.02) and the CESD-R 1-month adjusted intervention estimate was −0.98 (95% CI −4.89 to 2.94). No study-related adverse events were reported. Intervention estimates for secondary outcomes were also not significant (see Multimedia Appendix 2).

The statistical analysis plan included subgroup analysis of primary outcomes by ethnicity (Māori vs non-Māori) and children (responsible for one or more child vs none). There was no differential intervention effect for violence or depression symptoms based on whether women were caring for children or not.

There was a significant intervention effect for reducing violence for Māori women at 6 months (adjusted intervention estimate −14.19; 95% CI −24 to −4.37) and at 12 months (adjusted intervention estimate −12.44; 95% CI −23.35 to −1.54; Table 3 and Figure 3). There was also a significant intervention effect for reducing depression symptoms for Māori women at 3 months (adjusted intervention effect −8.7; 95% CI −15.9 to −1.6) but not at 6 or 12 months. Both violence (SVAWS) and depression (CESD-R) statistically significant changes exceeded standard error of measurement-based minimal clinically important differences [36] of 8 and 5, respectively.