The majority of papers reporting findings from GWAS list SNPs that are significantly associated with a trait, when analyzed one at a time, and do not attempt to integrate them into a risk prediction model [5,6,26,27,116]. Exceptions are the few efforts to develop risk scores that are based on simple linear functions of the genetic profiles [100]. Although several investigators see this initial selection as the first step to prognostic modeling [102], the reductionist approach has two limitations: it may identify too many associations because of dependencies between genetic variants that are the results of evolution [117] and it is unable to discover associations that are due to interdependent multiple genotypes [118]. For example, Hoh and Ott [118] describe a situation in which the simultaneous presence of three genotypes at different loci leads to a disease. The three genotypes have the same marginal penetrance and would not be found associated with the disease in a one-at-a-time search but only when examined simultaneously. Multivariate statistical models, such as linear or logistic regression, can circumvent these limitations by examining the overall dependency
