Modeling can have a much more profound impact in more complex associations than go beyond single markers, e.g. with approaches that try to model dozens and hundreds of gene variants that form a “pathway.” [42, 43] Such complex models may be built by MDR, kernel machines, stepwise logistic regression, or a diversity of other methods and it is important for the replication process to use the same exact steps as the model building. Even then, because these models are so flexible, it is unclear whether a “significant” finding in a second data set constitutes replication; the association may be driven by different sets of SNPs in the different studies. Researchers who conduct complex model-selection/model-building analyses should report their “final” model in as much detail as possible, so other investigators can judge the fit of that model in other data sets.
