Comparison of Li-Wong and loglinear mixed models for the statistical analysis of oligonucleotide arrays

Li and Wong have described some useful statistical models for probe-level, oligonucleotide array data based on a multiplicative parametrization. In earlier work, we proposed similar analysis-of-variance-style mixed models fit on a log scale. With only subtle differences in the specification of their mean and stochastic error components, a question arises as to whether these models could lead to varying conclusions in practical application.

In this paper, we provide an empirical comparison of the two models using a real data set, and find the models perform quite similarly across most genes, but with some interesting and important distinctions. We also present results from a simulation study designed to assess inferential properties of the models, and propose a modified test statistic for the Li-Wong model that provides an improvement in Type 1 error control. Advantages of both methods include the ability to directly assess and account for key sources of variability in the chip data and a means to automate statistical quality control.