Family-based association tests and genetic risk scores for survival data.

摘要

Several methods exist for examining within-family association between genotype and risk of adverse clinical outcome for survival data. In this thesis, we examine the performance of several family-based survival methods within the context of nuclear families. Using two-sib sibships, family-based association tests (FBAT) that are extensions of the logrank test and Wilcoxon test are compared to Cox regression based tests of association that account for family correlation. Type I error and statistical power are compared under several assumptions and conditions, including presence versus absence of parental genotypes and sibship frailty or not. Overall, the FBAT methods tend to be conservative, while the Cox regression methods tend to be liberal, especially when a robust variance estimator is used to adjust for residual sibship correlation. Cox regression methods have type I error closest to the nominal level when the model includes a sibship frailty component, and generally have the highest power among the tests compared.;Recent developments in the field of genetics have prompted researchers to include not only traditional risk factors but also genetic information in risk prediction models. Using simulation, we study several approaches for summarizing genetic information as a risk score predictive of survival. We find that commonly-used simple genetic risk scores, which sum the number of risk alleles across single nucleotide polymorphisms (SNPs), do not perform as well as weighted risk scores in terms of discrimination or goodness of fit, especially when the SNP effects are unequal or interactions among SNPs exist. Weighted genetic risk scores, in which the weights are the regression coefficients from an independent sample, closely approach the performance of using all risk polymorphisms independently in the model.;We examine approaches for selecting polymorphisms for inclusion in risk score algorithms when only a subset of available polymorphisms are truly associated with the outcome of interest, to see how this affects prediction models and risk score performance. We find that there is no clear difference between stepwise selection and an independent p-value selection process, and that the genetic risk scores perform less well when unassociated polymorphisms are included in simple or weighted scores.;Data from the Framingham Heart Study illustrate the methods for genetic risk scores.