Identification and Estimation of Long-term Treatment Effects via Data Combination
Abstract:
In this talk, I will talk about the problem of identifying and estimating the treatment effect of a certain intervention (e.g., new product design or new therapy) on some long-term outcome of interest (e.g., users’ long-term satisfaction or patients’ long-term health). This problem is very challenging: randomized experiments are often expensive and have short durations, so long-term outcome observations may not be available; observational studies can be cheaper and more likely to collect observations for long-term outcomes, but they are susceptible to confounding bias. In the first part of this talk, I will review some very recent literature that address this challenge by combining experimental and observational data and leveraging their complementary strengths. I will discuss major assumptions in the literature and discuss their strengths and limitations. In the second part of this talk, I will present my latest work on long-term causal inference under a confounding model more general than those in the existing literature.