Improving Privacy for Respondents in Randomized Controlled Trials: A Differential Privacy Approach

Randomized controlled trials (RCTs) have become a powerful tool for assessing the impact of interventions and policies in many contexts. Researchers have published an increasing number of studies that rely on RCTs for at least part of the inference, and these studies typically include the response data collected, de-identified, and sometimes protected through traditional disclosure limitation methods. In our presentation and extended paper, we explore the impact of applying differentially private methods on inference, computational feasibility, and accuracy as a case study on a published article.

We find, not surprisingly, that robust but naïve methods yield strong protection, but at great loss of inference ability. However, we also explore how a partial targeted relaxation as well as a model-specific protection method can alleviate those concerns, at least in this one case study.

The case study is part of a larger research program exploring the feasibility of stronger privacy methods in real-world contexts. We briefly outline some of the consequences of applying these methods for data openness, research transparency, and privacy of respondents.