Title: Model-assisted Calibration for Missing Data Analysis and Causal Inference
Calibration is an idea originated from survey sampling to incorporate known population information for efficiency improvement. We look at model-assisted calibration for missing data problems and causal inference, where the working models are for the propensity score and the data distribution. Using these models as calibration variables, under missing-at-random or no-unmeasured-confounding assumption, estimators become more robust against working model misspecifications, especially when no working model is correct. In addition, for causal inference, exact covariate balancing under finite sample size can be achieved when moments of covariates are also used as calibration variables.
Peisong Han is an Assistant Professor in the Department of Biostatistics at the University of Michigan. His primary research interests include (i) missing data problems in public health studies and survey sampling, (ii) causal inference, and (iii) data integration in the presence of data heterogeneity, especially when both summary information and individual-level data is available from multiple data sources. For missing data and causal inference problems, he has developed methods based on the calibration idea. Such methods help to reduce the estimation bias in the presence of working model misspecification for missing data problems and improve covariate balancing for causal inference problems.