To be sure, "Yikes" is far too understated when writing about violence against women. But I opted to avoid some of the harsher words that were in my head as I wrote this blog.
In their new working paper, Cullen, Alik-Lagrange, Ngatia, and Vaillant (2025) present startling evidence from an impact evaluation of a couples-training program in Rwanda that aimed to reduce intimate partner violence (IPV). Startling because they find that the program resulted in increases in rates of IPV, not only for the women in the program but also (and more so) for fellow women in their VSLA group that were not in the program. Double yikes: An earlier study of the program by Dunkle, Stern, Chatterji, and Heise (2020), also in Rwanda AND by the same implementing partners, found significant reductions in IPV. What to make of these contradictory findings?
First a short description of the intervention and results
Cullen et al is based on a large-scale randomized controlled trial that was government-managed with local and international nongovernmental organizations implementing. The curriculum was developed for the Rwandan context and the program had 22 weekly facilitated group discussions of about 19 couples in each group (described as focusing on “improving family relationships”). Sessions included group activities and discussions, as well as take-home exercises. The focus was on reducing IPV by improving marital communication skills, shifting conservative gender views, and promoting more progressive gender norms in the community. The economic literature on IPV tends to lean on bargaining models – with a focus on material resources and outside options. But this program (modeled after others) proposes an alternative pathway to reducing IPV by changing dynamics within couples. Outcomes were measured 6 months after the program ended.
Among 98 villages, half were randomly selected for the program. All villages also had a week-long training for local opinion leaders on gender equality and introduced safe spaces providing IPV support services. The study sample consists of 3 types of couples in village savings and loan associations: couples in the program (643), couples in untreated villages (812), and couples in treated villages not selected for the program (the spillover group) which informs the indirect program effects (594). This last group is especially important because here they find the largest negative effects of the program.
There is balance across treatment groups. Follow-up rates from the baseline were high (above 97%), with no differential attrition across groups. Take-up rates are not reported but among couples who started, attendance was 95% (they attended 21 out of 22 sessions). Sensitive survey questions were asked using audio computer-assisted self-interviewing (ACASI) so respondents did not have to answer an enumerator face-to-face (see Cullen 2023).
The major headline is succinctly “Relative to pure control couples, women who attended the couples training report more emotional, physical, and sexual IPV (7, 5, and 10 percentage points more, respectively). Spillover couples have even larger increases, of 7, 11, and 17 percentage points, respectively.” The community-level response exceeded the individual program’s (negative) effects. Without paying attention to spillovers, if only studying within the village, one would conclude that the program reduced IPV. The size of effects is meaningful; for example, there is a 33% increase in physical violence among spillover couples.
Why the backlash?
The authors describe an increase in intra-household friction, stemming from diverging in attitudes and desire for power and agency within couples, with women wanting a greater decision-making role in the household. The divergence between husbands and wives is significantly larger in spillover than control couples. In treated couples, while men had modest increases in progressive attitudes, it was far surpassed by the increase among wives – thus resulting in increased divergence.
Critical to their interpretation of these results is the supplemental qualitative reports conducted a year after the program ended. Among the findings of these reports, treated men faced social sanctions for sharing domestic work. And treated couples shared insights at community events which “provoked resistance”.
So what to make of the conflicting findings
Two studies, basically the same intervention, same implementing partners, but opposite results. Well, “basically the same” and “the same” does not mean the same. One can easily go down the rabbit hole on where the studies differ, which I sort of did. But here I will highlight a few areas where the two studies diverge.
(1) Capturing spillovers. Cullen et al (2025) have a spillover sample, whereas Dunkle et al (2020) do not, and I think we learn a lot from the spillovers. Of course, the divergent findings are still in place when looking at the results for the simple treated-untreated samples. In a companion paper to Dunkle et al, Chatterji, Stern, Dunkle, and Heise (2020) report results for a random sample of couples in treatment communities and control, and also show a consistent increase in IPV across types in treatment communities with little change in control communities, but a difference that is not statistically significant – though they are underpowered. Still, it seems suggestive of some real negative spillovers happening.
(2) Measurement of IPV. Cullen et al (2025) uses ACASI adapted for illiterate respondents, using illustrations than words (“if X happened to you, please press on the banana, if it did not press on the pineapple”) (authors insight, not in the paper). In Dunkle et al (2020), a quarter of women were either not sufficiently literate to use ACASI or otherwise requested to be interviewed face-to-face. The presumption that motivates ACASI is that one gets more truthful responses, and, in the case of IPV, high prevalence rates will be reported. So less ACASI can mean underreported IPV but enough to drive reductions in IPV, maybe not.
(3) Balance. Dunkle et al (2020) do not discuss balance. But it caught my eye that the IPV outcomes were higher at baseline than the controls. For example, their primary IPV measure, at baseline, the treatment mean is 49.8% and the control mean is 41.1%. Again, not clear imbalance would explain their reduced IPV results.
(4) From treatment sample to activist. The Dunkle et al (2020) study included an added feature, briefly mentioned. After completing the program, 25% of couples has additional training as community activists, the scope of this added training activity is not detailed in the paper. Not clear whether this a large enough share to shift results to a reduction in IPV but worth looking into.
So what to make of it all?
This case study report compares the two interventions and somewhat limply concludes that the conflicting evaluation findings are largely due to lower implementation quality in the Cullen et al intervention, driven by shorter timelines and different contracting arrangements. I think this report critically misses two key points in its summary on what we can learn from the conflicting results:
Beware of backlash in scaling. Assuming both studies are “true”, we have learned that scaling up such an intensive program comes with risks in terms of lower implementation quality. In the case study report it is noted that in terms of backlash in the earlier intervention “this type of resistance was also observed early in the implementation…, but it was carefully managed and largely dispelled by the more experienced programme staff.” As programs go big, the agility needed to address such “resistance” will decline.
Beware of spillover effects. We have learned that programs that target members of communities to change traditional social structures can prompt backlash, especially for non-participating women who may try to put the program to action in their own relationship without the benefit of direct exposure for them and their husbands. Oddly the case study report blames the “sampling strategy” for increased backlash and dismisses the spillover findings which perplexed me.
Join the Conversation