Missing Data
*Joseph L. Schafer, Department of Statistics
Keywords: latent class, multiple imputation, nonignorable, selection model
Data are said to be nonignorably missing if the probabilities of missingness depend on unobserved quantities. Traditional election models for nonignorable nonresponse are outcomebased, tying these probabilities to the partially observed values directly e.g., by a logistic regression) and can be quite unstable. With multivariate responses, the number of distinct missingness patterns ecomes large, making outcomebased selection modeling unattractive. In some examples, however, the information in the binary missingdata indicators is well summarized by a simple latentclass structure, suggesting that relationships between these indicators and the partially observed items might be explainable by the latent classes. In this talk, we present a latentclass selection model for multivariate incomplete data and investigate its properties. We then compare multiple imputations drawn from
