Detection of influential observations in longitudinal mixed effects regression models.

Mixed effects models for longitudinal data with fixed as well as random parameters are often used to describe average profiles. Influence measures are usually constructed to detect influential subjects and observations for the fixed regression parameters, treating the subject-specific parameters as nuisance parameters. One of these measures is the well-known Cook's distance. We show that this statistic may fail to detect or may incorrectly detect influential observations due to the random-effects variances and covariances. A conditional version of Cook's distance is proposed to assess the influence of observations on the estimated regression parameters.