Generalized Error Structure for Forestry Yield Models

Abstract

The combined time-series cross-sectional nature of remeasurement data from permanent forest plots is examined with an aim toward improving the precision of yield models fitted with these data. The linear model error term is regarded as an aggregation of plot, time period, and residual random effects with possibly distinct variances and correlations. Four alternative error covariance structures are posited that differ in the manner in which serial correlation, plot variance heterogeneity, and cross-plot correlations are prescribed. Yield models with the presumed error covariance specifications were fitted to a panel of 65 pure, even-aged Douglas-fir plot remeasurements, using two-stage generalized least squares and, in one case, a full maximum likelihood estimation. Ordinary least squares results were used as a basis for comparison. Comparison of the fitted models by prediction error and likelihood criteria indicate ordinary least squares nearly always performs better by the former measure, whereas one or more of the alternate specifications always have higher likelihood. For. Sci. 33(2):423-444.