Outcome coding choice in randomized trials of programs to reduce violence

TL;DR

This paper examines how the choice of outcome coding in randomized violence reduction trials affects bias, power, and insights, proposing a generative model and re-analyzing recent trials to inform better practices.

Contribution

It introduces a generative model for violence reduction and analyzes the impact of outcome coding choices through simulations and re-analysis of recent trials.

Findings

Coding choices influence bias and efficiency of estimates.

Re-analysis reveals potential biases from standard coding methods.

Guidelines for improved outcome coding in violence trials.

Abstract

Over the last decade, the number of randomized trials of programs to reduce intimate partner violence (IPV) has grown precipitously. However, most trials continue to measure and code violence using standards originally designed for global prevalence surveys. This choice may have consequences in terms of bias, power, and efficiency of trial estimates and may limit what we can learn about how programs are working. In this paper, we return to first principles to develop a generative model for violence reduction. We then use this model to better understand trade-offs in outcome coding choices via simulation. We re-analyze results from seven recent trials in Southern and Eastern Africa to highlight some of our findings. We conclude with a discussion of key take-aways for trialists.