The Effects of Misspecification of Level-I errors Structure in Multilevel Models for Longitudinal Design.

Multilevel models have become popular models for analyzing longitudinal data over past two decades. Longitudinal designs are common specially in medical science, where data is recorded on patients more than two times. As the data is collected on the same patients repeatedly over time, it is more likely that observations are correlated with each other. When the observations are correlated and the data have nested structure, traditional methods give biased estimates of the parameters, as they require the assumption of independence. The present study is concerned with exploring the effect of misspecification of level-I errors covariance structure in multilevel model for longitudinal data. The fixed effects were estimated with little to no bias, and accurate type-I error rates were observed under all the specifications of the covariance structure for the test of fixed effects. Random effects were estimated poorly for most of the conditions. Overestimated variances of the level-II random effects were accompanied by the underestimated level-I variance. The covariance between intercept and slope was underestimated in all conditions. The over specification of the covariance matrix of level-I errors gave better estimates than correct specification, under-specification and general misspecification. [ABSTRACT FROM AUTHOR]