Mixed-effect models with variance groups in R and the Trace Information Criteria (TRIC)

Microarray technology has permitted the analysis of thousands of cDNA hybridizations on a single chip, yielding a picture of the expression and co-expression of thousands of genes. The goal of statistical analysis is to appropriately account for sources of error in the experiment, while estimating genetic variability to produce interpretable and meaningful results. Current linear mixed-effects models analyses either estimate a single variance for all genes or a unique variance for each gene. We propose a method to estimate groups of gene variance using a mixed-effects model, using Ward's clustering method to estimate gene variance groups in R. The use of variance groups allows for a compromise between an over-fit and complex model and the number of estimates needed to accurately model gene variability. However, model comparison for mixed-effects models using the standard Akaike's Information Criteria Corrected (AICC) model selection criteria can be difficult and time consuming due to large datasets and computing limitations. We propose an alternative to standard model selection criteria, the Trace Information Criteria (TRIC), alleviating the estimation of numerous mixed-effects models and avoiding the common computational restrictions.