Multiple imputation and pooling strategies for handling wide-format missing data in latent growth curve modeling

Latent growth curve modeling (LGCM), commonly employed in psychological sciences, often encounters the challenge of missing data, which introduce difficulties into the modeling process. While the full information maximum likelihood (FIML) is the dominant missing data handling technique in practice, an alternative class of techniques, multiple imputation (MI), may be preferable in certain scenarios. With the recent increasing attention shine lights on multilevel MI, we conducted a simulation study examining both standard wide-format (MI-wEMB and MI-wFCS) and multilevel long-format MI methods (MI-llMLM and MI-lqMLM) in comparison with FIML under various missing data conditions and model specifications. Our results indicate that while FIML generally produced unbiased estimates and accurate confidence intervals (CIs), most MI methods demonstrated comparable performance across most conditions. However, the proportion of missing data played a notable role, affecting the performance of the methods to varying extents. We also examined the variation in pooling strategies for the likelihood ratio test (LRT) statistic. The results showed that pooling strategies did not exhibit significant differences in performance, but long-format MI methods displayed concerningly conservative Type I error rates (near zero), regardless of the pooling strategy used. The study further revealed that when the analysis model was misspecified, FIML still maintained the highest power, followed by MI-wEMB, MI-wFCS, and MI-lqMLM. Sample size and missing data proportion had the most significant impact on power. Our findings provide practical guidance for researchers in selecting appropriate missing data approaches for LGCM.