Self-thinningruleandmortalityfunctionsYuqingYang.pptVIP

Western Mensurationists’ Meeting, June 19–21, 2011, Banff, Alberta Assessing model fitting and prediction approaches in developing nonlinear mixed forest models Yuqing Yang and Shongming Huang Alberta Sustainable Resource Development Why nonlinear mixed models? Nonlinear mixed models Advantages: model heteroskedasticity and correlation account for un-known predictors Produce population-averaged (PA) and subject-specific (SS) predictions Disadvantages: too flexible sometimes PA predictions are typically biased Nonlinear mixed models General model form: Parameter estimation ML and REML to maximize a likelihood function: Linearization approaches: First-order (FO) First-order conditional expectation (FOCE) Direct integral approximation: Adaptive Gaussian quadrature (AGQ) Objectives Compare FO, FOCE AGQ approximation methods Compare ML REML parameter estimation techniques HT-DBH data for AW BA-Age data for SB HT-Age data for PL Base models Final models Residual autocorrelation HT-DBH model: BA-Age HT-Age models: Model fitting methods HT-DBH model: FO FOCE, ML REML AGQ with ML only BA-Age HT-Age models: FO FOCE ML REML Model predictions Model predictions PA predictions Using fixed parameters only: Biased mean response of a population SS predictions Using both fixed and random parameters random parameters need to be predicted first Predicting random parameters Using all obs. per subject Using a sub-sample of obs. per subject HT-DBH data: 1-5 trees per plot BA-Age data: 1-3 obs. per plot HT-Age data: 1-5 obs. per tree Random selection, 100 repetitions SS predictions FO method: FOCE method: AGQ method: q=1, follow FOCE method Comparison statistics mean prediction error: ē, ēi percent mean prediction error: ē%, ēi% an overall measure: ?, ?i, ē and ? - AW ē and ? - SB ē and ? - PL ēi% – AW ēi% – SB ēi% – PL HT predictions - AW BA predictions - SB HT predictions - PL Summary ML REML produced close results