The impact of covariance misspecification in multivariate Gaussian mixtures on estimation and inference: an application to longitudinal modeling