Estimating a multilevel dominant height-age model from nested data with generalized errors

Abstract

A dominant height-age model was developed following a multilevel nonlinear mixed-model approach. Random effects were modeled at the plot and the tree levels, with two random parameters at each level. In addition to the random effects, a variance function was used to model heteroscedasticity, and various covariance structures were evaluated to account for residual autocorrelation. A new covariance structure, the modified spatial power structure, is proposed. This new structure was found to provide the best fit to the model. Although residual autocorrelations were still not completely removed by this structure, they were substantially reduced. Model validation results using an independent data set confirmed that the final model with the modified spatial power structure produced more accurate and precise dominant height predictions. This was true for both expansion methods, and the empirical best linear unbiased predictor (EBLUP) expansion showed better results than the ZERO expansion. The model with the modified spatial power structure fitted by the EBLUP expansion method, therefore, was chosen as the preferred model for making tree-specific, dominant height predictions. The advantages and relevance of the multilevel nonlinear mixed-model approach for forest growth and yield modeling are discussed. © 2011 by the Society of American Foresters.