FACTORING VS LINEAR MODELING IN RATE ESTIMATION - A SIMULATION STUDY OF RELATIVE ACCURACY

A common strategy for modeling dose-response in epidemiology is to tra nsform ordered exposures and covariates into sets of dichotomous indic ator variables (that is, to factor the variables). Factoring tends to increase estimation variance, but it also tends to decrease bias and t hus may increase or decrease total accuracy. We conducted a simulation study to examine the impact of factoring on the accuracy of rate esti mation. Factored and unfactored Poisson regression models were fit to follow-up study datasets that were randomly generated from 37,500 popu lation model forms that ranged from subadditive to supramultiplicative . In the situations we examined, factoring sometimes substantially imp roved accuracy relative to fitting the corresponding unfactored model, sometimes substantially decreased accuracy, and sometimes made little difference. The difference in accuracy between factored and unfactore d models depended in a complicated fashion on the difference between t he true and fitted model forms, the strength of exposure and covariate effects in the population, and the study size. It may be difficult in practice to predict when factoring is increasing or decreasing accura cy. We recommend, therefore, that the strategy of factoring variables be supplemented with other strategies for modeling dose-response.