Comparison of different non-parametric growth imputation methods in the presence of correlated observations

Two different nearest neighbour methods and generalized additive models were applied to impute individual tree 5-year diameter increment for Scots pine (Pinus sylvestris L.) and Norway spruce (Picea abies (L.) H. Karst.). The aim was to compare the performance of different types of non-parametric methods when observations in the data are correlated. The study was implemented by defining different restrictions to the pool of possible reference trees. Observations from the same stand as the target tree are usually excluded while applying non-parametric imputation and cross-validation methods for evaluation. However, the nearest neighbours may all be selected from one particular stand, if the stand-level variables contain much weight in the distance function, and the neighbours are selected based on these stand-level variables. This may affect the stand-level results - for example, if the neighbours of a tree growing in a damp site are all selected from a dry site. The results showed that in general the differences in accuracy among the different methods and restriction alternatives were not remarkable under the assumption that the trees from the same stand as the target tree were excluded. However, the stand-level and regional results were slightly improved by not including many neighbours from one stand, implying that using many similar neighbours is an inefficient procedure. Hence, restricting the number of mutually correlated neighbours would be appropriate when considering the accuracy of stand-level or regional growth estimates.