A comparison of different ways of including baseline counts in negative binomial models for data from falls prevention trials