Linear, nonlinear or categorical: How to treat complex associations in regression analyses? Polynomial transformations and fractional polynomials

摘要

Nonlinear approaches to assess exposure-outcome relations are still fairly uncommon in public health research. The predominant reliance on linear associations and categorized continuous predictors is surprising, given the availability of powerful alternatives with sophisticated and user friendly software implementations. This simplicity threatens one of the major aims in regression analyses: to obtain an unbiased mean estimate of the dependent variable conditional on the predictor variables.