Prediction Bias Induced by Plot Size in Forest Growth Models
Abstract
Many forest growth models use plot-level variables, such as stem density or basal area, as predictors. Even if these variables are area-based, so that their expectations do not change with plot size when expressed per unit of area, a change in plot size may induce a prediction bias if the growth model is nonlinear with respect to these variables. In this study, we show that the effect of plot size on forest growth models can be seen as an error term that affects the plot-level entries and that error propagation theory can be used to predict such plot size-induced biases. For a density-dependent matrix population model in a tropical rainforest in French Guiana, the prediction bias was up to 23% when the plot size was decreased from 6.25 to 0.01 ha. The bias could be positive or negative, depending on the curvature of the response variable to the plot-level variables. Error propagation theory provided an approximate analytical estimator of bias. This estimator may help in understanding and correcting the prediction bias that is observed in some growth-and-yield models when the plot size used for prediction differs from that used for model fitting.