CAUSAL INFERENCE BY SEMIPARAMETRIC IMPUTATION

Open Access
- Author:
- Kang, Doh Yung
- Graduate Program:
- Statistics
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- December 19, 2006
- Committee Members:
- Joseph Francis Schafer, Committee Chair/Co-Chair
Rana Arnold, Committee Member
Vernon Michael Chinchilli, Committee Member
Murali Haran, Committee Member - Keywords:
- estimating function
survey inference
imputation
Causal inference
missing data - Abstract:
- Causal effects are comparisons among the outcomes that a study subject would have under different treatment conditions. Because no subject can receive multiple treatments at the same time, causal inference may be regarded as a missing-data problem. Inferences about causal effects are challenging in the presence of confounders, which may distort estimated effects due to their mutual associations with the treatment assignment and with the outcomes. To deal with this problem, we propose a marginal causal model (MCM), a regression that allows average causal effects to vary with respect to one or more variables of interest. We estimate the parameters of the MCM by constructing imputation models and replacing the missing potential outcomes with predicted values. The causal effects are then estimated by solving a set of estimating equations based on the observed and imputed outcomes. To mitigate the biases that may result when the imputation models are misspecified, we augment the imputation models with covariates derived from estimated propensity scores. We apply these methods to data from real and simulated observational studies of the causal effects of dieting on body weight among adolescent girls.