Someone once said that anyone who talks about 'controlling for a variable' probably doesn't understand statistics. I'm one of those people, alas.

I've been using the R package Hmsc to build a spatial joint species distribution model of plant collection records in New Guinea - so the response variable is a matrix of species presence/absence across sites (regions) and there are several covariates, mostly environmental and soil variables. One covariate comes from a cost-distance raster that estimates a region's difficulty of access. I added this because it seemed obvious from the map that most collections came from easy-to-reach areas.

What I want to estimate is the real species richness of a region, adjusting for the fact that it may be hard to get to and, as a result, have few collection records. How can I distinguish the signal coming from the environmental variables from the noise of collection bias?

In a normal regression I might set the cost-distance value to zero and make predictions for all my regions from the other covariates, as if every region were perfectly easy to get to (would this make sense?) - should I do this here? And since variance partitioning shows the cost-distance element to be a small part of the variance (the random effect of the latent spatial term accounts for nearly 50%), do I even need to? I'm worried that the skewed collection efforts due to accessibility have got swallowed up and lost in the spatial term, although there are plenty of other potential causes of spatial autocorrelation.


Your Answer

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

Browse other questions tagged or ask your own question.