I should have noticed this when I first looked at the pdf, but the problem is that the areas of no prediction are surrounded by points that lie on the same line. For second-order trend removal, you need to have at least three distinct x and three distinct y coordinates in the neighborhood around each prediction location (among other restrictions). So, if all the neighbors fall on the same two lines, there is instability in the predictions, and the prediction can�??t calculate. As Gail discovered, if you increase the maximum neighbors, you�??ll force more points in that don�??t fall on the same two lines.