4.8 Conclusion

Multivariate missing data lead to analytic problems caused by mutual dependencies between incomplete variables. The missing data pattern provides important information for the imputation model. The influx and outflux measures are useful to sift out variables that cannot contribute to the imputations. For general missing data patterns, both JM and FCS approaches can be used to impute multivariate missing data. JM is the model of choice if the data conform to the modeling assumptions because it has better theoretical properties. The FCS approach is much more flexible, easier to understand and allows for imputations close to the data. Automatic tile imputation algorithms with simultaneous partitions of rows and columns of the data form a vast and unexplored field.