The best-fitting model for risk of CD was selected based on (1) the lowest value of AIC, (2) non-significant differences in deviance from more complex models and (3) significant parameter estimates. Logistic regression was conducted using PROC GENMOD in SAS version 9.1.3 (SAS Institute, Cary, NC, USA). Random residual effects of twin resemblance and repeated measurement were accommodated by using the Generalized Estimating Equation (GEE) algorithm incorporated in the SAS GENMOD procedure on the simplifying assumption of constant correlation between measures within monozygotic and dizygotic twin clusters.
