Using expert knowledge with control of false discovery rate to select regressors for prediction of soil properties

Soil scientists often have many covariates that they can use to predict soil properties by regression. They are ill-advised to use all available covariates uncritically, but methods for selection (whether informal or formal) that depend on data for both the predictors and the predictand are subject to selection bias. In this paper we propose an approach that uses automated methods for selecting variables, but which controls the rate of false rejection of true null hypotheses about the various predictive regression models that are considered. This approach reduces the effects of selection bias. Expert judgement is used both to determine the size of the pool of models that is searched (matching it to the strength of evidence for the existence of good models) and to ensure that the searched subset of possible models includes those that make sense, given our knowledge of the soil. The method is described, and a case study is presented on the prediction of soil properties in a large field in northern New South Wales, Australia.