Abstract:
|
Missing values in the explanatory variables of a mixed-effects regression typically cannot be ignored, as is well-known. Such missingness is of particular concern for individual participant meta-analysis, where explanatory variables may be missing for entire studies - a pattern known as systematic missingness. Multiple imputation (MI) is an attractive approach to handling such missing data patterns, since MI-based analyses are widely accessible. However, multilevel and systematic missing data structure present challenges for creating imputations. We compare a joint imputation model with a fully conditional approach that is simpler to implement. These proper imputation methods are rarely equivalent, even for normally distributed multilevel data. However, we found that practical differences between the methods in the meta-analysis setting largely reflect differences in typical 'default' choices for prior distributions rather than differences in the likelihoods. In particular, the two methods differ in their estimation of between-study variability. We illustrate our findings via simulation and an analysis of mental health outcomes after pediatric traumatic brain injury.
|