A method to quantify confounding in regression analyses applied to data on diet and CHD incidence