No specific permits were required for the described field studies. Generalized linear models were used to analyze the relationship between the tree attributes and (1) the total number of lichen species on each tree, (2) the number of species of conservation concern on each tree (which in this study included red-listed species (Gärdenfors 2010) and indicator species, the latter used to indicate forests of high conservation value in conservation assessments; (Nitare learn more 2000), (3) presence or absence on each tree of the four most frequently occurring

lichen species of conservation concern (Collema furfuraceum (Arnold) Du Rietz, Lecanora impudens Degel., Leptogium saturninum (Dicks.) Nyl., and Lobaria pulmonaria L. Hoffm. Species number (1 and 2 above) was modeled with a Poisson distribution and with an identity link function

to the explanatory variables (tree attributes), while presence or absence of individual species was modeled with a binomial distribution and a logit link function (i.e. logistic regression). The choice of distributions and link functions was based on their fit with the data. Prior to analysis, all explanatory variables were first checked for strong correlations (here >0.6 in a bivariate plot). Where correlations were present, we excluded those variables from further analysis that we judged were of least practical use for identifying retention SCH772984 datasheet trees in the field. Tree age, size of branches, and size and width of tree crown were thus excluded due to their strong correlation with bark crevices (tree age) and tree diameter (size of branches and size and width DNA Synthesis inhibitor of crown). We detected no overdispersion in the Poisson-modeled data. We used model-averaging to derive parameter estimates for each explanatory variable (see tree attributes in Table 2), to overcome the problem with model selection uncertainty. All possible subsets of models were thus constructed (i.e. 256 models) and we used the second-order Akaike information criterion AICC (which penalizes models with many explanatory variables) to calculate relative likelihoods and Akaike weights for all models (Burnham and Anderson 2002). Akaike weights can be interpreted as the probability that each model

is the best model, given the data and set of considered candidate models. Model-averaged parameter estimates and associated standard errors and confidence intervals were calculated for all parameters across the models with a ΔAICC ⩽ 2 (on average 12 models), which are models that can be said to have “substantial support” (Burnham and Anderson, 2002 and Grueber et al., 2011). To reduce bias in parameter estimates, we denoted the estimate of parameters not included in any given model within the candidate set to zero and thus averaged parameter estimates over all models, not just those containing the parameter (Burnham and Anderson, 2002 and Lukacs et al., 2010). The statistical software package Statistica was used for all modeling (StatSoft 2011).