The influence of informative priors in the estimation of MIMIC model parameters with small sample sizes and outliers
The influence of informative priors in the estimation of MIMIC model parameters with small sample sizes and outliers
The influence of informative priors in the estimation of MIMIC model parameters with small sample sizes and outliers
Abstract
This study examined the performance of Bayesian and maximum likelihood estimation in MIMIC models when the latent factor contained outliers but the fitted model assumed latent normality. Using a simulation study with varying sample sizes and contamination proportions, we evaluated absolute relative bias, coverage probability, interval width, and power for the direct effect parameter.
The results showed that the correctly centered informative Bayes approach consistently performed best, yielding the smallest bias, nominal or perfect coverage, and the highest power across conditions. A correctly centered informative coefficient prior with a naïve loading prior also performed well, although slightly worse than the fully correctly specified Bayesian model. In contrast, misspecified priors, especially misspecified loading priors and severely misspecified coefficient priors, led to poorer performance, and maximum likelihood estimation showed the weakest robustness overall.
These findings suggest that informative priors can substantially improve estimation and inference in MIMIC models with outliers, but their benefits depend on accurate prior specification.