Correlation-Regression Analysis for Understanding Dominant Height Projection Accuracy

2023

e1-e10

Artikel aus einer Zeitschrift

200210395

Dominant height predictions tend to be less accurate with increasing projection intervals, a common occurrence not well understood statistically although intuitively self-evident and taken for granted. In this study, we used linear prediction theory (essentially correlation-regression) for height projection to better understand this
phenomenon as related to three mean-covariance models, including an unstructured (essentially sample means, variances, and correlations) model and two structured models with means approximated by using the Chapman-Richards function and variances-covariances directly or indirectly modeled (i.e., mixed-effects modeling). An empirical evaluation of dominant height projection by the three models was performed using a second-rotation loblolly pine (Pinus taeda L.) plantation data set. The results showed that weakening correlations inherent in distant remeasurements largely determine projection accuracy to decline and that the squared coefficient of correlation (R2) can serve as an upper-limit measure of regression-type linear prediction of heights given one single prior measurement. The unstructured model directly estimated means and covariances without involving any function approximation; therefore, it provided best fits in terms of likelihood-based measures and better (smaller root mean squared error) prediction when applied to new individual plots, becoming a nearly top-limit standard model for comparison with other linear predictors of heights.