Outliers and/or measurement errors on the permanent sample plot data

Estonian Agricultural University, Institute of Forestry and Rural Engineering, Kreutzwaldi 5, 51014 Tartu, Estonia. Email: [email protected] __________________________________________________________________________________ Abstract Tree-wise re-measured permanent sample plot data are needed to create and maintain modern forest growth models for Estonia. For that purpose an all-Estonian network of forest sample plots was established by the Department of Forest Management at Estonian Agricultural University in 1995. Tree species, storey, azimuth, distance from plot centre, two breast height diameters and faults were determined for each tree, and total height, crown base and height of dry branches were measured for sample trees. Routine inspection of forest plot measurement data for outliers was necessary because it provided information about the causes and consequences of measurement errors. At the moment, 715 forest sample plots have been established. A total of 749 outliers have been checked for errors on 112 plots in 2004. In this study several statistical methods were tested for the detection of outliers. Empirical distributions of most tree variables by species and storey were analyzed using Grubb’s test, Dixon’s test and the 2-sigma rule. Multivariate methods (residual diagnostics) were used for detecting outliers in the interaction of several variables. Analyzing empirical diameter distributions, Dixon’s test for small samples and 2-sigma rule for large samples were more effective than Grubbs’ test. The results of regression diagnostics of influential observations showed that Covratio method was the most sensitive one for outlier detection in nonlinear model and Dffits method in linear models. The most effective method for checking outliers from height-diameter data was the studentized residual. Outlier detection plays important role in modeling, inference and even data processing because outliers can lead to model misspecification, biased parameter estimation and poor forecast. __________________________________________________________________________________