Abstract:
|
Missing data often cause undesired properties such as bias and loss of efficiency. By modeling the distribution of complete data and the response mechanism, incorporating them into the likelihood can solve the problem. In nonignorable missing data analysis, it is difficult to check (i) verification of models for complete data and (ii) model identifiability. In this talk, we model the distribution of the observed data and derive sufficient conditions for the model identifiability by specifying that the model of the response mechanism is logistic and that of the distribution for the observed data is the generalized linear model. For analyzing missing data, multiple imputation is a popular technique. One of its advantages is ease of the variance estimation without troublesome computation. We propose a multiple imputation method with models whose identifiability can be verified with observed data. We also propose a fractional imputation method in the same setup as an alternative method to solve estimating equations. We present results of numerical simulations and a real data analysis with both the multiple imputation and the fractional imputation methods.
|