Logistic regression models for wind and snow damage in northern Finland based on the National Forest Inventory data

The susceptibility of forest stands to wind and snow damage was predicted with basic logistic regression models from a large systematic sample and with conditional logistic regression from a smaller matched study. Data was from the 8th National Forest Inventory, recorded in northern Finland during 1992-1994. The current investigation adds to previous studies by considering the role of individual explanatory variables more closely instead of only predicting the absolute risk. From both, the basic logistic and the matched models, we obtain odds ratios for explanatory variables, which estimate the so called relative risk; i.e. increase in the damage risk by a particular factor. In our study, matching by cluster or municipality was used to control the possible effects of stochastic local factors that were not quantified, such as wind and snow conditions, on the odds ratios. According to the basic logistic regression model, the susceptibility of a stand to wind damage was increased by: large mean diameter, high stand age, seed-tree cutting, special cutting (cutting for ditches, roads or power lines, or sanitation cutting after damage), and decreasing temperature sum. Significant exposure factors for snow damage were: decreasing temperature sums, elevation >200 m a.s.l., conifer-dominance, mineral soil, undrained, unthinned, development stage pole stand and no proximity to stand edge. The distribution of damaged stands at the landscape level could not be predicted well with the basic logistic regression model, since it is only concerned with the susceptibility component of the probability, and not the occurrence of the damaging agents; i.e. unfavourable weather. Odds ratios can be used as an alternative to probabilities for comparing the damage risks related to combinations of exposure factors from site, stand and forest management. The matched approach showed promise as an alternative method for modelling the odds ratios. The results suggest that matching was useful for wind damage, but less so for snow damage. If the incidence of winds can be taken into account in some way, as with the matched model, the effect of exposure factors on damage risk can be estimated more precisely. Wind damage is a more stochastic phenomenon than snow damage: at least in northern Finland, snow damage may be more tied to more regularly occurring climatic factors at the site than wind damage. Consequently, the basic logistic regression model where these can be included as explanatory variables was better for snow damage than the matched model. (C) 2000 Elsevier Science B.V. All rights reserved.