Random Forests Analysis: A modern statistical method for screening in high-dimensional studies and its application in a population-based genetic association study

Abstract: Technology advances in this century, especially, in molecular generics yields high volume, high dimensional data. This creates many unprecedented challenges for statisticians who are responsible for analysis of such data. Although logistic regression method is quite popular in association analysis in medical researches but it has some serious limitations in handling high dimensional data. In present study, our goal is introduce a modern model-free statistical method called random forest that we believe is able to overcome difficulties of the classical statistical methods in finding association between predictors and a trait. Material and Methods: In this study, the nonparametric random forest technique was employed to determine the important factors associated with ankylosing spondylitis (AS) disease. Genetic materials including information on HLA-B27 status (positive/negative) and 12 polymorphisms of the ERAP-1 gene were collected on 401 patients and 316 healthy controls. The data were analyzed both with the logistic regression method and random forests technique and the results were compared. Results: Based on a stepwise logistic regression, HLA-B27 and rs28096 polymorphism were significantly associated with the disease. However, using the random forests technique, we found that HLA-B27 and rs1065407 were the main factors associated with diseases and in fact rs28096 polymorphism becomes the third in importance ranking. Conclusion: The results from our study indicate some discrepancies between logistic regression and random forest analyses of high-dimensional data such as the genetic data that we are dealing here. Although logistic regression is quite popular, easy to employ, and is a predominant statistical method among researchers, but it has some serious limitations. On the other hand, more modern statistical such random forest enjoy a more methodological sophistication and yield more accurate and reliable results. Therefore, researchers should be aware of such alternatives and should use these alternatives accordingly and as situation arise in screening tests especially in genetic data analyses. Key words: random forests, High-dimensional data, interaction, logistic regression, CART