Maximum likelihood versus multiple imputation for missing data in small longitudinal samples with nonnormality.

The study examined the performance of maximum likelihood (ML) and multiple imputation (MI) procedures for missing data in longitudinal research when fitting latent growth models. A Monte Carlo simulation study was conducted with conditions of small sample size, intermittent missing data, and nonnormality. The results indicated that ML tended to display slightly smaller degrees of bias than MI across missing completely at random (MCAR) and missing at random (MAR) conditions. Although specification of prior information in the MI imputation-posterior (I-P) phase influenced the performance of MI, especially with nonnormal small samples and missing not at random (MNAR), the impact of this tight specification was not dramatic. Several corrected ML test statistics showed proper rejections rates across research designs, whereas posterior predictive p values for MI methods were more likely to be influenced by distribution shape and yielded higher rejection rates in MCAR and MAR than in MNAR. In conclusion, ML appears to be preferable to MI in research conditions with small missing samples and multivariate nonnormality whether or not strong prior information for the I-P phase of MI analysis is available.