An Improved Algorithm for Finding Linked Subsets in the Presence of Heterogeneity or Interactions

Identifying regions of DNA that contribute to disease becomes more difficult as the diseases become more complex. Traits such as blood pressure or heart disease have complex etiologies involving multiple genes or gene by environment interactions. In these cases, many current models are too simple to detect the effects of genes. Recently, Shannon et al. proposed the use of recursive partitioning (RP) techniques to handle such situations. These methods can separate a sample into increasingly similar subgroups based on some criterion – typically an improved fit to a regression model. The “tree linkage” method employed by Shannon et al. was based on improving the regression fit for standard Haseman-Elston regression. In HasemanElston regression, the magnitude of the regression slope is the indicator of association between a trait and a region of DNA. Therefore, it was proposed that tree linkage algorithm would be improved by changing the method of partitioning the data to reflect this. Specifically, basing the partitioning criterion on the regression slope (and ignoring the intercept) should improve type I error rates. Simulation results comparing the original algorithm with one based on slope terms showed greater bias in the distribution of empirical p-values in the original algorithm. Furthermore, a comparison of performance under a specific alternative showed that the original algorithm had slightly less favorable distributions of linkage group impurity and tree specificity. Both algorithms performed similarly with respect to sensitivity and kappa, although there was a slight improvement in kappa for the improved algorithm.