ABSTRACT
The demand for running imputation diagnostics and checking is high given the popularity of multiple imputation analysis nowadays. Multiple methods and options are available in imputation software packages so researchers and practitioners also need to make some decisions based on imputation diagnostics. We provide some relate discussion in this chapter. Imputation diagnostics is a comprehensive process. Adhering to important principles such as the inclusive imputation strategy is effective in forming good imputation models. We present several strategies in checking the quality of imputed values. This can be done by comparing observed values with imputed values through comparisons on marginal and conditional distributions. The latter can be implemented using the balancing property of the propensity score. Another strategy is to compare completed data with their replicates generated from the same model following the idea of posterior predictive checking used in Bayesian model diagnostics. We can also monitor and compare the fraction of missing information, which quantifies the ratio of between-imputation variance to total variance, of certain estimates from different imputation models because it is related to model predictability. Besides checking imputed values, it is also important to compare analysis results from alternative imputation and other missing data methods. We also briefly discuss about the use of prediction accuracy in multiple imputation analysis. Examples based on both real data and simulation studies are given.
