Minimum description length methods of medium-scale simultaneous inference

Nonparametric statistical methods developed for analyzing data for high numbers of genes, SNPs, or other biological features tend to overfit data with smaller numbers of features such as proteins, metabolites, or, when expression is measured with conventional instruments, genes. For this medium-scale inference problem, the minimum description length (MDL) framework quantifies the amount of information in the data supporting a null or alternative hypothesis for each feature in terms of parametric model selection. Two new MDL techniques are proposed. First, using test statistics that are highly informative about the parameter of interest, the data are reduced to a single statistic per feature. This simplifying step is already implicit in conventional hypothesis testing and has been found effective in empirical Bayes applications to genomics data. Second, the codelength difference between the alternative and null hypotheses of any given feature can take advantage of information in the measurements from all other features by using those measurements to find the overall code of minimum length summed over those features. The techniques are applied to protein abundance data, demonstrating that a computationally efficient approximation that is close for a sufficiently large number of features works well even when the number of features is as low as 20. More generally, the MDL-based information for discrimination does not suffer from the asymmetry of the p-value as a measure of evidence for one hypothesis over another.