Causal Inference Using Semi-parametric Imputation

• Activity Number: 143
• Type: Contributed
• Date/Time: Monday, August 2, 2010 : 8:30 AM to 10:20 AM
• Sponsor: Section on Survey Research Methods
• Author(s): Andrea Piesse*+ and David Judkins and Laura Alvarez-Rojas and William R. Shadish
• Companies: Westat and Westat and Westat and University of California, Merced
• Keywords: Causal inference ; potential outcomes ; counterfactual ; hot deck ; multiple imputation ; bootstrap

Abstract

In non-randomized observational studies, selection bias may confound the relationship between treatment and outcome. Imputation is one method for addressing selection bias. With this approach, a potential outcome is imputed for each treatment level not received and the association between treatment and outcome is estimated using both reported and imputed outcomes. Multiple imputations may be used to account for the impact of the imputation process on variances. This paper analyzes data from a four-arm comparison study (Shadish et al., 2008) where students were first randomly assigned to a randomized experiment or an observational study. Using the randomized experiment as a benchmark, we examine treatment effects estimated by applying semi-parametric multiple imputation, ANCOVA, and propensity scoring methods to the observational study. Issues with variance estimation are discussed.