I'm interested in examining which of a suite of environmental variables is most important to predicting the distribution of a given species A, given a set of presence/absence points. I would like to include in my list of environmental variables a species distribution model of a different species (species B), as I think it may be an important interacting species that can influence the presence of species A. The distribution model of species B is one that I also would have made using presence/absence points that I also have. I haven't found any literature saying I can't do this, but I also haven't found any saying I can. Is this "allowed," or has this been already rejected as a viable option for distribution modeling?

  • $\begingroup$ Seems reasonable to me. $\endgroup$
    – mkt
    Oct 3, 2019 at 19:07
  • $\begingroup$ When you say you want to include a species distribution model as a variable, do you mean the you want to include the predictions from that model as a predictor in your new model? That sounds ok to me, although bear in mind that there will be uncertainty around those predictions. Maybe you could run additional models using the upper and lower prediction intervals as your predictor, to get an idea of how much that uncertainty affects your results. $\endgroup$
    – rw2
    Oct 8, 2019 at 11:53
  • 1
    $\begingroup$ That is what I mean! Thank you very much for the advice, that is exceptionally helpful. $\endgroup$
    – aeiche01
    Oct 9, 2019 at 12:02
  • $\begingroup$ Have you looked into join species distribution modelling? It is more complex, but if your goal is to look at the interaction between two species and how that affects their distribution, it could be the way to go. $\endgroup$
    – Liam G
    Jan 21, 2020 at 22:02

1 Answer 1


There is no statistical reason this is not allowed, in my view this decision should be based on the ecology.

However, you should consider whether the species distribution model is actually any good - do your model evaluation. Also - with the prediction surface you will also have some uncertainty. Lots of two-stage approaches ignore this uncertainty. If you have a full Bayesian analysis then you can use your posterior surface as a prior in the other model. If you are being frequentist and your model is quick to fit you might want to consider bootstrapping as a way to propagate uncertainty.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.