CAREER: Adaptive experiments towards learning treatment effect heterogeneity

University Of California-Berkeley, Berkeley CA

Investigators

Abstract

Understanding and characterizing differential and heterogeneous causal effects have become increasingly important in many scientific fields. For example, in precision health, identifying differential treatment effects serves as an essential step towards materializing the benefits of precision health, because it provides evidence regarding how individuals with specific characteristics respond to a given treatment either in efficacy or in adverse effects. In social science research, evaluations of the effectiveness of government programs or public policies across different individuals inform more effective policy-making. As reliably designed randomized experiments often provide evidence of the highest grade for verifying the effectiveness of a treatment or an intervention, this research project aims to develop three randomized experimental design strategies for better learning causal effect heterogeneity. These design strategies are broad and will be applicable in clinical trials, social experiments in biomedical sciences, public health sectors, and online controlled experiments in technological enterprises. Since the project will develop modern experimental design strategies and new statistical methods with many applications, this research project will provide opportunities for integrating research with teaching and training students across different stages. The project will impact STEM education through the training of undergraduate and graduate students and the recruitment of students from underrepresented groups into (bio)statistical fields. Activities to achieve these education-related goals include introductory reading groups, course developments, university undergraduate research programs, and outreach activities to underrepresented minorities. This research project will develop three novel response adaptive experimental design strategies and theoretical insights toward learning treatment effect heterogeneity from a frequentist viewpoint. The first strategy will focus on designing randomized experiments to sequentially allocate experimental efforts so that subpopulations mostly harmed or benefited from a particular treatment can be efficiently identified. The second strategy will focus on learning treatment effect heterogeneity measured by the variability of the conditional causal effect variability. The learned heterogeneity allows the development of a new efficient covariate-adjusted response adaptive framework whose estimator may attain the best achievable efficiency. The third strategy aims to further materialize the benefit of treatment effect heterogeneity by designing randomized experiments to maximize participants' overall welfare. The research project is thus expected to open a new research connection between adaptive experiments and social welfare improvement. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.