Adaptive False Discovery Rate Control under Independence and Dependence
The Journal of Machine Learning Research
Editorial: Second Issue for Computational Statistics for Clinical Research
Computational Statistics & Data Analysis
A hierarchical Bayes model for biomarker subset effects in clinical trials
Computational Statistics & Data Analysis
| Hi-index | 0.03 |
For personalised medicine the identification of predictive biomarkers is of great interest. These could guide the choice of therapy and could therefore optimise the benefits of patients of such treatments. The technology of gene expression microarrays allows one to scan thousands of potentially predictive biomarkers simultaneously. In clinical trials it has nowadays become common to use microarrays to collect gene expression data of the patients before treatment. The identification of predictive biomarkers can be statistically addressed by inference of gene-wise generalised linear models (GLM) including an interaction term gene expression times treatment. Inference for such GLMs is then often based on likelihood-ratio (LR) or Wald test statistics to test the influence of interaction of gene expression and treatment on the clinical treatment response. For multiple testing scenarios coming along with these gene-wise GLMs the control of the false discovery rate (FDR) would be appropriate; some false positives can be tolerated within a list of potential candidate genes which deserve further investigation. In a simulation study the utility of various FDR controlling multiple testing procedures for the identification of predictive genes is examined. Since the usual experiment on microarray data deals with small numbers of observations due to financial or probe limitations special interest lies on the behaviour of small sample sizes. Results reveal that a permutation of regressor residuals (PRR) test is superior to standard LR and Wald tests in terms of FDR control.