Develop new methods for analyzing large-scale observational studies and field experiments, building on approaches from machine learning in order to obtain personalized estimates of treatment effects, and apply the methods to analyze behavioral nudges in the digital environment.
In both experimental and observational studies of the impact of “behavioral nudges,” the impact of the nudges is likely to vary across individuals. However, attempting to estimate this heterogeneity in treatment effects raises myriad problems, including problems of multiple hypothesis testing as well as computational issues in attempting to explore all possible forms of heterogeneity. I propose to develop new methods for estimating personalized treatment effects and optimal personalized treatment assignment policies, building on tools from machine learning, where the methods enable the researcher to systematically explore heterogeneity while preserving the ability to conduct valid hypothesis tests. The methods will be applied in original studies of behavioral nudges in internet search, prominence of news, driver safety incentives in ride-sharing, and nudges in online education. Existing behavioral nudge studies will also be re-analyzed.