LAUSR

There are now several examples in the ecological literature where, for spatial prediction like kriging, non-Euclidean distances were used in autocorrelation models developed under a Euclidean distance assumption. This leads to a problem where prediction variances may be negative, and generally leads to unreliable standard errors for prediction. My objective is to help ecologists understand the problem and avoid this mistake. I introduce a class of models that are easy to construct, based on linear mixed models, that perform well and guarantee that prediction standard errors will be positive. Received: 13 November 2017 | Accepted: 30 December 2017 DOI: 10.1111/2041-210X.12979