A completely automated model selection procedure resulted in two quite different models, depending on the severity score cutoff that

was used to define response. Assuming that a response is given by a score of 2 or greater on the Southall scale, the model selected by an automated stepwise procedure was (Model 1): equation(Model 1) Response2∼Year+CAR+COL+TUG+Month+Age+RL_rms,Response2∼Year+CAR+COL+TUG+Month+Age+RL_rms, Estimate Std. error z Value Pr(>|z|) (Intercept) 699.74410 324.52124 2.156 0.0311* Year −0.34602 0.15989 −2.164 0.0305* CAR −10.30153 5.23157 −1.969 0.0489* COL −6.09617 3.02291 −2.017 0.0437* TUG −9.54309 MEK pathway 4.89167 −1.951 0.0511. Month −3.04004 1.62113 −1.875 0.0608. Age 0.06393 0.02682 2.383 0.0172* RL_rms 0.18178 0.11832 1.536 0.1244 Signif. codes: 0‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1

‘ ’ 1 Binomial models are somewhat difficult to interpret with respect to explanatory power, and the usual R summaries for binomial GLMs do not contain the kind of R-squared summary statistics one normally expects in a regression. There is a tool1 (“binomTools”) to extract information from binomial models to give an idea about their explanatory power. We used function Rsq in package binomTools to illustrate, roughly, how much explanatory power each model had, and to assess Raf inhibitor how much additional explanatory power the various models had when including or excluding information on received level. We found that Model 1 had an R-squared value of approximately 0.58. We reran all models with the cutoff for scoring a response set this time to ⩾3 on the Southall scale. In this case, both forward and backward stepwise model selection indicated that the preferred model was [Model 2]: equation(Model 2) Response3∼Sex+N-other-boats,Response3∼Sex+N-other-boats,which means that a killer whale’s response to the passage

of a ship (using a severity score of ⩾3 as a cutoff), on average, was best explained by the number of small vessels in the area and the sex of the whale. Using strictly automated procedures, Model 2 did not include information on received noise level at the whale. Because a central focus of this study is to understand TCL whether noise was a better predictor of behavior than other variables, we compared the selected model (Model 2) to one that also contained information on received noise level. We found that equation(Model 3) Response3∼Sex+N-other-boats+RL-rms,Response3∼Sex+N-other-boats+RL-rms,had similar support from the data as Model 2. The difference between Model 2 and Model 3 was ΔAIC = 1.41, which means that there is no strong statistical support for dropping noise level from the model. On the contrary, explanatory power of the model increased from R-squared = 0.23–0.25 when we included a term for RL. We therefore proceeded on the grounds of management interest, and used Model 3 for interpretation. Estimate Std. error z Value Pr(>|z|) (Intercept) −8.54322 465.47010 −0.018 0.9854 SexM −1.54243 0.62471 −2.469 0.