The results of logistic analyses when the variables are highly correlated: an empirical example using diet and CHD incidence.

We provide an example which includes highly correlated variables in multivariate logistic analyses relating four nutrients to the incidence of coronary heart disease in 10 years for over 7000 men. Paradoxical results occur for both the inferences to be drawn and variable selection. Different models exist which show that any particular variable both is and is not related significantly to coronary heart disease incidence, and step-up and step-down variable selection algorithms provide drastically different results.