Lutz P Breitling1, Ewout W Steyerberg, Hermann Brenner. 1. Division of Clinical Epidemiology and Aging Research, German Cancer Research Center, Bergheimer St. 20, D-69115 Heidelberg, Germany. L.Breitling@dkfz.de
Abstract
BACKGROUND: Examination of the genetic structure of complex diseases by a "genomic pathway approach"--which applies stepwise model selection to sets of more than 1000 polymorphisms in studies of several hundred subjects--has recently been proposed. Models constructed through extensive selection procedures may yield misleading test statistics and measures of predictive performance; we aimed to quantify the extent of such problems inherent to stepwise regression on the genomic pathway scale. METHODS: We performed permutation analyses and data-splitting approaches using one of the datasets examined in the paper that originally suggested this approach (n = 536; 1195 SNPs in 22 genes) (Lesnick et al. PLoS Genet. 2007;3:e98). RESULTS: The P values for the genetic effects produced by standard testing severely overestimated the significance, resulting in our example in a standard P value of 3.5 x 10(-69) and a permutation P of 0.003 (95% confidence interval = 0.001 to 0.009). Furthermore, the apparent validity as measured by the area under the receiver operating characteristic curve in 90% training datasets (0.935 [interquartile range = 0.918-0.951]) was extremely overoptimistic when compared with the validity estimated from the excluded 10% validation subsets (0.564 [0.518-0.614]). This validated area under the receiver operating characteristic curve was lower than for models predicting case/control status solely from age and sex while excluding any genetic effects (median difference = -0.040 [95% confidence interval = -0.049 to -0.031]). CONCLUSIONS: The application of stepwise model selection on the genomic pathway scale--at least in the simple form currently put forward--is prone to yield highly misleading results. We provide pointers to some promising alternatives.
BACKGROUND: Examination of the genetic structure of complex diseases by a "genomic pathway approach"--which applies stepwise model selection to sets of more than 1000 polymorphisms in studies of several hundred subjects--has recently been proposed. Models constructed through extensive selection procedures may yield misleading test statistics and measures of predictive performance; we aimed to quantify the extent of such problems inherent to stepwise regression on the genomic pathway scale. METHODS: We performed permutation analyses and data-splitting approaches using one of the datasets examined in the paper that originally suggested this approach (n = 536; 1195 SNPs in 22 genes) (Lesnick et al. PLoS Genet. 2007;3:e98). RESULTS: The P values for the genetic effects produced by standard testing severely overestimated the significance, resulting in our example in a standard P value of 3.5 x 10(-69) and a permutation P of 0.003 (95% confidence interval = 0.001 to 0.009). Furthermore, the apparent validity as measured by the area under the receiver operating characteristic curve in 90% training datasets (0.935 [interquartile range = 0.918-0.951]) was extremely overoptimistic when compared with the validity estimated from the excluded 10% validation subsets (0.564 [0.518-0.614]). This validated area under the receiver operating characteristic curve was lower than for models predicting case/control status solely from age and sex while excluding any genetic effects (median difference = -0.040 [95% confidence interval = -0.049 to -0.031]). CONCLUSIONS: The application of stepwise model selection on the genomic pathway scale--at least in the simple form currently put forward--is prone to yield highly misleading results. We provide pointers to some promising alternatives.