Hierarchical time-oriented approaches to missing data inference

In practice clinical data are nearly always incomplete. When confronted with such data, a physician or investigator must make inferences about missing information. Possible strategies for inference include (1) interpolation, (2) extrapolation, (3) repeating the nearest value, (4) repeating the previous value, (5) patient-specific mean values, (6) patient-specific linear regression over time, (7) disease-specific mean values, (8) normal values, and (9) linear regression of correlated co-recorded variables. This study analyzes these strategies in a time-oriented data bank of patients with systemic lupus erythematosus, demonstrating that more accurate inferences of missing data are obtained when (1) strategies are tailored to the characteristics of the individual variable, (2) time-oriented strategies (e.g., interpolation) rather than non-time-oriented strategies (e.g., disease mean) are incorporated, (3) a ranked set of strategies is incorporated in a hierarchical stepwise fashion, and (4) the degree to which missing data are "nonrandomly" missing is assessed to allow estimation of bias. Interpolation is the best single technique with these data while linear regression of correlated co-recorded variables is a relatively weak technique. Inferences made by these hierarchical time-oriented approaches show significantly smaller mean differences from the actual values than do results from typical statistical package strategies.