Interaction and model selection
Manered Wildner, Andreas Markuzzi · Journal of Internal Medicine · 1997
DEAR SIR, we appreciate the recent review of Hallqvist et al. [l] regarding the evaluation of interaction between causes in cardiovascular epidemiology.We agree that the selection of the analytical model has implications for the interpretation of hdings.Hallqvist and colleagues criticize the evaluation of interaction based on any model besides the additive.These authors follow in their deductive logic Rothman's concept of causality [2, 31, which is well established in the epidemiologic community.This paradigm commonly assumes the appropriateness of the additive model for the description and explanation of observed effects of independent causes, and hence assumes an inherent interaction between causes where other models have been fitted, such as the model underlying logistic regression.Rothman justifies the additivity of independent causes by a small risk assumption [4], which is essential for the truncated probabilistic relationship Pr(A+B) Pr,,, + Pr(B)