Multiple imputation is a recommended method for handling missing data problems. One of the barriers to the successful use of multiple imputation is the non-convergence of estimation algorithms that are used to produce the imputations. In particular, problems with model failure are common with the popular approach of fully conditional specification or "chained equations".
This presentation will provide an overview of methods for handling problems with imputation model failure. We will describe approaches for diagnosing problems with imputation models, including checks for collinearity and sparse data. Strategies for overcoming these issues include data reduction methods and augmented regression for perfect prediction. These strategies will be reviewed and compared using a case-study evaluation based on data from the Longitudinal Study of Australian Children. Given that non-convergence of imputation algorithms is a common issue that hampers the implementation of multiple imputation, these proposed strategies will provide practical guidance for users of multiple imputation.