Texas A&M University - Department of Statistics
In both observational and randomized studies, data are often missing either by chance or by design. In recent years, the parametric multiple imputation method proposed by Rubin (1978, 1987) has become one of the most popular methods for handling missing data. Unfortunately, Fay (1992, 1994, 1996), Meng (1994), Rubin (1996) and Clayton et al. (1998) have shown that, in certain settings, the variance estimator proposed by Rubin will be inconsistent with upward bias, resulting in conservative confidence intervals whose expected length is longer than necessary. In this talk, I will show a general formula for the large sample bias of the commonly used variance estimator. A simple example is used to illustrate a situation that the traditional approach can be anti-conservative. That is, the actual coverage rates of the resulting confidence intervals are less than nominal. I will also discuss the asymptotic variance structure of the general imputation estimator and it's usage as a basis to understand the behavior of two Monte Carlo iterative estimators, stochastic EM (Celeux & Diebolt, 1985; Wei & Tanner, 1990) and simulated EM (Ruud, 1991).