Harrison H. Li

Experimentation is an extremely powerful tool to ascertain the causal impact of one or more treatments of interest, as proper randomization of treatment assignments eliminates the possibility of confounding variables that prevent causal associations from being made. However, the effectiveness of an experiment depends on how the treatments are allocated (the design). A poorly designed experiment can yield an imprecise or biased estimate of the true treatment effect. My research focuses on various design problems that arise in modern experimentation under limited resources, motivated substantially by problems in the social sciences. For example, to evaluate the effectiveness of a government unemployment insurance program in increasing future wages, one might randomize who is required to participate in such a program. However, a socially responsible government will want to preferentially allocate their limited resources for this program to certain groups that might benefit more. I am interested in how to formulate and optimize such a trade-off between statistical and non-statistical objectives in experiment design