Abstract

GARCIA, María del Carmen et al. Information and  predictive  criteria  for  selection of mixed linear model. SaberEs [online]. 2014, vol.6, n.2, pp.00-00. ISSN 1852-4418.

In longitudinal studies the experimental units are observed repeatedly in several occasions. Linear mixed models are an important tool for analyzing this type of data because allows modeling separately the multiple response variable measurements (as a function of the covariates) and the correlation between them. The model-building process includes: the selection of covariates, the definition of the number of random and fixed effects, and the specification of the random error correlation structure. For the selection of the "optimal" model a wide range of information and predictive criteria are available. The statistical packages use the marginal model to calculate them, prioritizing the inference about population parameters. However, some authors argue that individual random effects are often of interest and introduce the conditional versions of them. The purpose of this paper is to introduce some of these criteria considering both the conditional and marginal version and illustrate its use with data from the "Argentina Household Permanent Survey". In the application it is observed that the performance of the criteria is dissimilar in terms of their choice behavior.

Keywords : Longitudinal data; Marginal and conditional criteria; Conditional Akaike.

        · abstract in Spanish     · text in Spanish     · Spanish ( pdf )