Causal Inference for Multilevel Observational Data with Application to the Kindergarten Retention Study

• Activity Number: 321
• Type: Topic Contributed
• Date/Time: Wednesday, August 6, 2003 : 8:30 AM to 10:20 AM
• Sponsor: Social Statistics Section
• Author(s): Guanglei Hong*+
• Companies: University of Michigan
• Keywords: causal inference ; propensity score ; multilevel modeling

Abstract

The standard causal theories and methods rest on the "stable unit treatment value assumption (SUTVA)." This assumption is hardly tenable in multilevel settings where interference between units and treatment enactment variation are likely. The primary objectives of this study are to extend current theory in statistical science about causal inference to encompass multilevel educational data, and to investigate the implications of this theoretical extension for various propensity score-based causal inference techniques. I redefine the causal effects of treatments in multilevel settings by replacing the SUTVA with the exchangeability assumption, clarify the treatment assignment mechanisms, and sort out the potential sources of bias for nonexperimental data. I examine in multilevel settings the applicability of various propensity score-based causal inference techniques including propensity score matching, stratification, and "inverse-probability-of-treatment" weighting. For illustration, I apply the extended framework and these different methods to an empirical estimation of the causal effects of kindergarten retention using multilevel observational data.