Predicting species composition when environmental drivers are missing

Species distribution models provide valuable tools for marking species’ range limits and predicting individual species’ responses to environmental change.

However, multi-species patterns like richness and co-occurrence involve correlations among species, and the processes driving these correlations are not always observable.

In other words, some sites will be less suitable or more suitable for a group of species than expected by these models, depending on variables that have not even been measured.

Current (nonlinear) models cannot account for these variables, which limits their utility for community ecology.

How can we predict species composition without knowing what these unmeasured factors are or how they affect community assembly?

I designed a stochastic neural network[4] (to be made available as an R package called mistnet). The model uses latent random variables (Z) for unobserved environmental factors. Since these variables can take on a range of values, the model can make a range of distinct predictions for a given set of environmental data.