CIRJE-F-584. Kubokawa, Tatsuya and Muni S. Srivastava, "Modified Bayesian Information Criterion in Linear Mixed Models", September 2008.

In this paper, we consider the problem of selecting variables from the fixed effects as well as from the random effects when observations from several clusters are available to provide consistent estimators of some unknown parameters. We obtain Bayesian Information Criterion (BIC) using the prior on regression parameters and random effects but all other parameters as unknown. To improve the performances of the BIC in small samples, we use the idea of decomposing the likelihood into two parts, called 'within' and 'between' analysis of variance. A modified BIC (mBIC) is then obtained by combining the BIC obtained from the two parts. It is numerically shown that mBIC is superior to the marginal and conditional AIC's.