2.3.3. Modeling analyses

We fitted generalized additive mixed models (GAMM) to the data, with the gam function mgcv R package. We used a binomial family argument with a logit‐link function to estimate the parameters of an inverse‐logit selection model based on seal foraging dives and random points (Johnson et al., 2006). Foraging dives and pseudo‐absences were the response variable, taking the values 1 and 0, respectively. To consider intra‐individual autocorrelation, we included individual as a random effect. Environmental variables were treated as fixed effects. Bathymetry, tidal current, distance from shore, and distance from the last haulout were included as continuous variables; sediments were treated as categorical variable. When one sediment type was over‐represented, the model was forced to consider this sediment type as reference level (otherwise reference sediment type was included alphabetically). The multicollinearity between covariates was assessed using the VIF value (Kutner et al., 2004). The best model was selected using the AIC (Akaike, 1973). Firstly, we fitted one model per site for each species to focus at the local scale. Secondly, to highlight the importance of modeling habitat selection around colonies, particularly when local habitat characteristics differ, we also fitted a global model for each species using pooled data from all colonies and included “site” as a random factor. We did not include Kenmare for harbor seals’ global model, as tidal current dataset was not available for this area. However, as we had problem of convergence when running global model, we only included one diving point on three. For models fitted for each colony, we calculated the importance of each covariate using the prediction function of the GAMM, providing an index of the relative importance of each covariate in the chosen model. Maps of habitat selection predicted by the model were created with ArcGIS for all sites.