|
Abstract:
|
Sequential Regression Imputation (Raghunathan et al. (2001)) employs a set of regression models to impute missing values in each variable, conditional on all other variables. The imputations, however, are susceptible to model misspecification. We investigate a robust alternative approach. For each variable with missing values, we obtain a predicted value of the variable and a response propensity for missing this variable, conditional on observed and imputed values of other variables. We partition the data set into strata based on the joint distribution of the propensity score and predicted value. A hierarchical Bayesian model across all strata is used to impute the vector of missing values in each stratum. The approach is evaluated using both actual and simulated datasets.
|
