Nonlinear models of height growth for Douglas-fir in southwestern Oregon

Graduation date: 1986A review is made of methods which assess the bias\ud and non-normality of parameter estimates and predictions\ud obtained with nonlinear regression. Particular emphasis\ud is placed upon curvature measures of nonlinearity,\ud related measures of parameter and prediction bias, and\ud the effects of reparameterizations. Alternate models of\ud individual tree height growth are compared on the basis\ud of mean square error, intrinsic nonlinearity, parameter\ud effects nonlinearity, and estimated bias. While results\ud are specific to the data examined, some general\ud conclusions are made concerning appropriate models for\ud individual tree height growth. Both the Richards and a\ud Weibull-type growth model are found to adequately\ud describe individual tree height growth, with low levels\ud of intrinsic nonlinearity, and acceptable parameter\ud effects nonlinearity following repararneterization. Some\ud evidence is found for a modification of either the\ud Richards or Weibull model to include an asymptotic\ud linear growth rate when modeling the height growth of\ud some western conifers past the age of 200.\ud Stem analysis data on Douglas-fir height growth in\ud mixed confier stands located in southwestern Oregon are\ud used to develop a system of dominant height growth and\ud site index prediction. The Weibull model is used\ud successfully to develop a polymorphic height growth\ud prediction equation. A linear model, estimated with\ud site index as the dependent variable, is used to predict\ud site index. A comparision is made of pooled least\ud squares and random coefficient estimation methods. The\ud random coefficient method is found to more closely model\ud the shape of early height growth, but appears to result\ud in more biased predictions and performs very poorly on\ud older height growth, with both the estimation and\ud validation data. Alternative error assumptions are\ud examined with the pooled data method. The best\ud performance in validation is obtained with assumption of\ud independent errors, heteroscedastic across trees