%0 Research article
%T Comparing regression estimation techniques when predicting diameter distributions of Scots pine on drained peatlands
%A Siipilehto, Jouni
%A Sarkkola, Sakari
%A Mehtätalo, Lauri
%D 2007
%J Silva Fennica
%V 41
%N 2
%R doi:10.14214/sf.300
%U https://silvafennica.fi/article/300
%X We compared different statistical methods for fitting linear regression  models to a longitudinal data of breast height diameter (dbh)  distributions of Scots pine dominated stands on drained peatlands. The  parameter prediction methods for two parameters of Johnson’s S<sub>B</sub> distribution, fitted to basal-area dbh distributions, were: 1) a linear  model estimated by ordinary least squares (OLS), 2) a multivariate  linear model estimated using the seemingly unrelated regression approach  (SUR), 3) a linear mixed-effects model with random intercept (MIX), and  4) a multivariate mixed-effects model (MSUR). The aim was to clarify  the effect of taking into account the hierarchy of the data, as well as  simultaneous estimation of the correlated dependent variables on the  model fit and predictions. Instead of the reliability of the predicted  parameters, we focused on the reliability of the models in predicting  stand conditions. Predicted distributions were validated in terms of  bias, RMSE, and error deviation in the generated quantities of the  growing stock. The study material consisted of 112 successively measured  stands from 12 experimental areas covering the whole of Finland (total  of 608 observations). Two independent test data sets were used for model  validation. All the advanced regression techniques were superior to  OLS, when exactly the same independent stand variables were included.  SUR and MSUR were ranked the overall best and second best, respectively.  Their ranking was the same in the modeling data, whereas MSUR was  superior in the peatland test data and SUR in the mineral soil test  data. The ranking of the models was logical, but may not be widely  generalized. The SUR and MSUR models were considered to be relevant  tools for practical forest management planning purposes over a variety  of site types and stand structures.