In development economics, we speak often of durable change — the idea that a well-designed intervention does not merely shift outcomes while it runs, but catalyzes something that persists after it ends. The concept is doing a lot of work in the literature on gender-based violence prevention. Programs argue that by changing norms, not just behaviors, their effects will outlast the program period. Donors fund on that basis. Evaluators design follow-ups to test it.
The study I contributed to — a nine-year post-program follow-up of a combined cash transfer and behavior change communication intervention in rural Bangladesh — gave me an opportunity to look seriously at that claim. What we found was complicated, and it has shaped how I think about the whole enterprise.
The intervention and the original findings
The Targeting the Ultra-Poor (TUP) program in Bangladesh combined a productive asset transfer, consumption support, training, and a savings component. An additional arm layered on behavior change communication (BCC) targeting gender attitudes and intimate partner violence. The original RCT, evaluated at endline in 2012, found reductions in physical and sexual IPV in the BCC arm. This was held up — reasonably — as evidence that norm-change programming could work.
The nine-year follow-up, conducted in 2021, asked what had survived.
What persists, and why the answer is complicated
The short answer: some things did, some didn’t, and the pattern is hard to explain with a simple story.
Physical IPV remained lower in the BCC arm relative to control at nine years. That is encouraging. But sexual IPV — which had shown the largest reductions at endline — had converged back toward control by the follow-up. Psychological IPV showed mixed patterns by sub-measure.
What should we make of this? A few possibilities:
The mechanisms were different for different forms of violence. Physical IPV may have been more responsive to material conditions — assets, consumption stability, women’s bargaining power — that persisted through the asset transfer. Sexual IPV may have been more directly tied to the BCC content, which receded once facilitated sessions ended.
Norm change is leaky. The social environment into which households were embedded did not change; only the households in the study did. Over nine years, the treated households may have been re-embedded in norms that gradually pulled behavior back toward community means. This would suggest that individual-level BCC, without community-level norm shift, is working against structural headwinds.
The measurement is doing something. IPV measurement is notoriously sensitive to recall windows, enumerator effects, and women’s willingness to disclose. Nine years is a long time. The surviving sample is not random — attrition is correlated with household characteristics that also predict IPV. We attempted to address this through IPW and bounding, but uncertainty remains.
What I took from it
The honest takeaway is not that the program failed. The persistent reduction in physical IPV at nine years is meaningful. But it complicates the version of this story that gets told in policy briefs — the one where rigorous implementation plus sufficient dosage equals durable norm change.
The leakage of sexual IPV effects over nine years suggests that behavior change communication, however well-designed, is operating in a context it cannot itself transform. The program changed how individual men and women in treated households related to each other during the program period, and for some outcomes, that change persisted. For others, the social surround slowly reasserted itself.
This is not a counsel of despair. It is an argument for context — for understanding what a program can hold in place and what it cannot, for thinking more carefully about the community-level conditions under which individual-level norm change becomes self-sustaining.
I keep coming back to a line from a paper I read years ago in a different context, about literary interpretation rather than impact evaluation. The argument was that meaning is not stored in a text but produced at the site of encounter between the text and its reader — and that the reader always brings the world with them. Something similar might be true of interventions: their effects are not stored in the program but produced at the site of encounter between the program and the social environment into which it is delivered. When the program ends, the social environment continues. And the social environment has a longer memory.
This essay reflects my own interpretation of the findings. See Lokot et al. (2025) for the full analysis.