The effects of mixed-mode protocols on attrition in a household panel survey: why do some effects emerge only at later waves?