A Bayesian nonparametric method for model evaluation: application to genetic studies

Authors: Shahbaba, Babak¹; Gentles, Andrew²; Beyene, Joseph³; Plevritis, Sylvia²; Greenwood, Celia³

Source: Journal of Nonparametric Statistics, Volume 21, Number 3, April 2009 , pp. 379-396(18)

Buy & download fulltext article:

Price: $56.94 plus tax (Refund Policy)

Abstract:

Statistical models applied to genetic studies commonly assume linear relationships (between disease and risk factors) and simple distributional forms (by relying on asymptotic methods) for inference. However, when the sample size is small, inference using traditional asymptotic models can be problematic. Moreover, the gene-disease relationship is not always linear. In this article, we present a new nonparametric Bayesian method for model assessment, and we demonstrate the advantages of this approach particularly when the sample size is small and/or the true model is non-linear. We evaluate our approach on simulated data and find that it performs substantially better than alternative models. We also apply our method to two real studies: diagnosis of conventional high-grade non-metastatic osteosarcoma, and survival in Burkitt's lymphoma.

Keywords: non-linear models; Dirichlet process mixtures; model evaluation

Document Type: Research article

DOI: http://dx.doi.org/10.1080/10485250802613558

Affiliations: 1: Department of Statistics, University of California, Irvine, USA 2: Department of Radiology, Stanford University, Stanford, USA 3: Dalla Lana School of Public Health, University of Toronto, Toronto, Canada

Publication date: 2009-04-01