Prediction intervals: placing real bounds on regression-based allometric estimates of biomass

Abstract

Biomass allometry studies routinely assume that regression models can be applied across species and sites, and that goodness of fit of a regression model to its derivation dataset indicates both the relevance of the model to a new dataset and the likely error. Assuming that a model is relevant for a new sample, a prediction interval is a useful error measure for stand mass. Prediction coverage tests whether the model and hence the interval are appropriate in the new sample. Data for three similar shrubby species from four similar sites were combined in various ways to test the impact of varying levels of biodiverse heterogeneity on the performance of the four models most commonly used in published biomass studies. No one model performed consistently well predicting new data, and validation checks were not good indicators of prediction coverage. The highly variable results suggest that the common models might contain insufficient variables. Euclidean distance was used to quantify the relative similarity of samples as a possible means of estimating prediction coverage; it proved unsuccessful with these data.