|
Abstract:
|
Multilevel regression and poststratification (MRP) is a flexible modeling technique that has been used in a broad range of small-area estimation problems. Traditionally, MRP studies have been focused on estimating a single value, like a conditional mean, from a nonrepresentative sample. Nonrepresentative samples often occur in experimental data, and adjusting for nonrepresentativeness is commonly done using causal inference techniques, many of which have analogies in survey sampling literature. In this talk, I will extend this analogy to using MRP-type methods for estimating causal effects. We simulate a large-scale randomized control trial based on a complex experimental design, and compare traditional and nonparametric treatment effect estimation methods with newly developed MRP methodology for experimental causal inference. The context of the simulation study is to evaluate how effective an education intervention is on the GPA of high-school students in the US, when the experiment is run using a stratified cluster design. Using MRP, we produce treatment effect estimates for small-areas in the population that have lower bias and variance than standard causal inference methods.
|
