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

Area of Science:

• Longitudinal research methodology
• Statistical modeling

Background:

• Missing data is a common challenge in longitudinal research.
• Accurate handling of missing data is crucial for valid statistical inference, particularly in complex models like latent growth models.

Purpose of the Study:

• To compare the performance of maximum likelihood (ML) and multiple imputation (MI) for handling missing data in latent growth models.
• To evaluate these methods under conditions of small sample size, intermittent missingness, and nonnormality.

Main Methods:

• A Monte Carlo simulation study was employed.
• Conditions included varying sample sizes, missing data patterns (MCAR, MAR, MNAR), and data distributions (nonnormal).
• Performance was assessed based on bias, rejection rates of test statistics, and influence of prior information in MI.

Main Results:

• Maximum likelihood (ML) exhibited slightly less bias than multiple imputation (MI) across missing completely at random (MCAR) and missing at random (MAR) conditions.
• While prior information specification in MI influenced performance, especially with nonnormality and missing not at random (MNAR), the effect was not substantial.
• Corrected ML test statistics demonstrated appropriate rejection rates, whereas MI's posterior predictive p-values were sensitive to distribution shape and showed higher rejection rates in MCAR/MAR than MNAR.

Conclusions:

• Maximum likelihood (ML) is generally preferable to multiple imputation (MI) for latent growth models with small samples and multivariate nonnormality.
• ML offers a more robust approach regardless of the availability of strong prior information for MI imputation-posterior (I-P) phase.