I have never heard a formal lecture answering this question even after many years in epidemiology. It should be parsimonious of course to avoid too many missings but seems largely a subjective approach to keep or drop a variable. It was therefore quite helpful to find now an online lecture that exemplifies a sound approach – check out unc.edu/courses/2006spring. I already used anova to compare models (at least since my move from SAS to R) while using AIC is something that I am adding now to my toolbox.
> sapply(list(model1,model2,model3),AIC)
I was using loess smoother in the past to decide how to enter a variable in the model, but will give now also rcspline.plot a chance
> library(Design)
> rcspline.plot(y=frogs$pres.abs, x=frogs$meanmin, nk=5, m=20)