Statistical habitat models give quantitative descriptions of species-habitat relationships by analysing records of environmental conditions and species incidence by means of logistic regression or other techniques. The most useful model in a given situation is usually unknown a priori. Several data-driven, computer-intensive methods exist to automatically fit a large number of candidate models to the data and to select the "best" one of these. Recent research in medical statistics suggests that these selection methods lead to models with reduced predictive performance. In addition, the assessment of predictive performance is overly optimistic if it is not validated with new data or by using resampling techniques. We investigate two procedures for constraining model complexity - backward stepwise variable selection and penalized maximum likelihood - and analyse internal validation via bootstrap resampling. We use an ecologically motivated data set comprising 21 real environmental variables and artificial incidence data of a hypothetical plant species. Backward stepwise variable selection improved predictive performance of the models. Constraining the model complexity via penalized maximum likelihood further improved model performance. The differences between the model strategies decreased with increasing size of the data set used. Internal validation via bootstrap resampling improved the estimates of model performance.