Causal inference for multi -level observational data with application to kindergarten retention.
Guanglei Hong · Deep Blue (University of Michigan) · 2004
The purpose of this dissertation is to extend the potential-outcomes causal framework to encompass multi-level data. I handled the multiplicity of potential outcomes associated with each treatment for each individual unit in a multi-level setting by replacing the stable unit treatment value assumption with the exchangeability assumption. I defined the causal effects of treatments for three basic types of multi-level experimental designs---multi-site randomized designs, cluster randomized designs, and joint multi-level randomized designs. For the corresponding multi-level observational designs, I investigated the applicability of various propensity score-based approaches to causal inference. Using the national Early Childhood Longitudinal Study kindergarten cohort data, I applied the extended causal framework and the propensity score-based causal inference techniques to an empirical study of the causal effect of kindergarten retention and that of kindergarten retention policy on children's literacy and math learning. While the kindergarten retention policy demonstrated no average effect, there was clear evidence that the kindergarten retention treatment leaves most retainees even further behind.