Forward variable selection based on models fitted to harbour porpoise data, based on the cross-validation log-Likelihood (CVLL).

ΔCVLL is the change in CVLL by adding a (smooth of the) covariate. Sea state and Site entered the model as factor variables. “te(X,Y)” represents a tensor product smooth of X and Y coordinates (Bardsey projection). “s” represents a thin plate regression spline smoother (or cubic regression spline for cyclic smoothers, i.e. for the covariates Lunar cycle, aspect and tidal state). The best model contained all variables up to slope.