# How to compare the results of the random forest algorithm with the other SDMs?

I'm using the random forests algorithm in species distribution context. I have presence-absence as a dependent variable, so I decided to work with classification. At this point, I recovered the probability of presence this way :

``````predict(RF_MODEL, test_, type='prob')[,2]
``````

If I want to compare the results of this model with other models "glm", "gam" and "brt", can we do this or we have to work with regression to have the estimation of occurrence probabilities as the other models?

• Well, you have to evaluate different measures including model fit, performance as well as estimated spatial uncertainty. In Random Forests the fit is indicated by the internal Bootstrap validation and is in the reported statistics. However, performance is an evaluation of prediction against data not in the model. You can do this via a cross-validation withhold or an approach like `rfUtilities::rf.crossValidation` which produces an error distribution based on a permutation. The spatial uncertainty can be evaluated using an infinitesimal Jackknife or U-statistic of the estimate variance. – Jeffrey Evans Jan 25 at 18:30
• @JeffreyEvans, thank you very much for your very interesting explanations. I want to know if I can consider the probability estimated by RF above as occurrence probability as the for example the one extracted by logistic regression for example? – user2300 Jan 27 at 0:04
• The output of random forest is based off a vote of the tree ensemble. So the predicted probability is the proportion of trees that vote 'Yes'to an occurrence (if using classification). So yes you can treat them as telling you the same thing, but how they reach those conclusions are very different. Note, you can't technically describe either as giving you the occurrence probability without accounting for detectability, so it's more like the probability of detection. – Liam G Jul 20 at 2:08
• @LiamG, thank you very muck for your response. Could you please explain more your last point "Note, you can't technically describe either as giving you the occurrence probability without accounting for detectability, so it's more like the probability of detection.". Why we can't use the occurrence probabilities in this context? – user2300 Jul 21 at 15:21
• You can use the model generated probabilities, but bear in mind that the probability of you observing a species is a function of the probability a site is occupied and the probability of detecting that species given the site is occupied. So if the probability of detection is very high and does not vary with covariates then yes you can interpret the probabilities in relation to occurrence. For cryptic species, the probability of detection may be very low and/or vary with covariates. More of a point to consider when interpreting your predictions relative to the ecology of the species. – Liam G Jul 21 at 23:00