Table 3. Results of the hurdle boosted regression tree simplification procedure.

Model simplification	Binomial		Poisson		Hurdle
	Dcv	SE Dcv	Dcv	SE Dcv	Dnull	Dresid	mse	rmpe
saturated	0.696	0.021	3.512	0.272	2.429	0.787	2.417	0.240
binomial	0.944	0.014	3.442	0.435	2.661	0.870	3.040	0.302
Poisson	0.695	0.016	3.822	0.348	2.361	0.896	3.087	0.307
binomial and Poisson	0.946	0.012	3.880	0.304	2.668	1.106	3.866	0.384

Hurdle models were fitted as a two-step process: a binomial and Poisson part. These results show that the saturated model using all spatial predictors for both binomial and Poisson parts had lower prediction deviance (e.g., Dcv) and explanatory deviance (e.g., mse) compared to hurdle models for which the binomial, Poisson or both parts were built using only the most influential spatial predictors.

Abbreviations: Dcv and SE Dcv are the mean and standard error of the 10-fold cross-validation residual deviances, Dnull and Dresid are the mean null and residual deviances, mse is the mean square error and rmpe is the relative mean prediction error.