Non Parametric Procedures to Explore Interactions in Epidemiological Studies

In epidemiologic regression studies it could be interesting to analyze a non linear functional relationship between a response variable and a continuous variable with linear splines. An important and crucial aspect regards the location and number of the cut points that should be estimated. This problem could be even more complicated when the relationship differs among subgroups and robust analytical methods that can disentangle complex effects are required. In this paper we propose a novel procedure that extends MARS (Friedman, 1991) in a conditional logistic regression context. A recursive partitioning scheme is adopted in order to estimate knots location and select the most important predictor variables. We apply the proposed procedure to investigate the short term relationship between temperature and all-cause non-injury mortality.