Binary logistic regression with stratified survey data

Standard inference techniques are only valid if the design is ignorable. Two approaches that take the design into account are compared using binary logistic regression. The modelbased approach includes relevant design variables as independents and the designbased approach use design weights. The approaches are exemplified using a cross-sectional stratified mail survey, where associations between urinary incontinence and health related variables among women are studied. Stratification variable age is strongly associated with the dependent variable while stratification variable geography is primarily administrative. The final models contain the same variables irrespective of which approach is used, and the design seem to be only slightly informative. The modelbased approach gives more efficient estimates since geography is not included in the model. But since point estimates differ slightly there is a tradeoff between bias and efficiency, so a designbased approach which includes geography in the design weights might also be advocated.