We assessed 42 theodolite tracks containing ship transits to find natural experiments

that could be used to model the probability of a whale responding. Of the 42 tracks considered, 35 could be considered in a before-during natural experimental framework, with sufficient information to quantify changes in whale behavior before and during a ship transit. The 7 tracks that had to be dropped contained insufficient information about whale behavior before and/or during the ship’s transit to evaluate response; sparse information RO4929097 on the ship’s track was not the limiting factor. Scoring each experiment as either a response or a non-response required using all values greater than or equal to some severity score cutoff as a somewhat arbitrary threshold. To account for the subjective nature of this step, analyses were run using severity scores of both 2 and 3 as cutoffs. There was insufficient coverage and resolution in the data to consider other levels of the Southall score as cutoffs. We modeled the probability that a whale did (1) or did not (0) show a behavioral response to a ship transit, in a GLM framework. Candidate click here covariates included

natural (WhaleID, Year, Month, TimeOfDay, Age, and Sex) and anthropogenic (CAR, TUG and COL; Ship_Speed; PCA1; N_other_boats; RL_rms and RL_weighted) variables. With a binomial response, one has the choice of several link functions, including logit,

probit or complementary log–log. The logit link is the default for most logistic regressions. We used a probit link, because this imposes the classic sigmoidal shape thought to underlie conventional dose–response curves (Miller et al., 2012). We did not have sufficient data to be able to test of alternative relationships; instead, we are assuming that killer whales will not respond to noise below some unknown, but low, received level, and that all whales would respond to noise at some unknown high level (even if that level is beyond the range of our data). In other words, the model structure assumes that if there is a dose–response relationship, it will follow a classic sigmoidal shape common to all toxicology studies, and the data are used to estimate parameters describing the curve we suspect is there. If there is no support from the data for fitting the curve, then each term will have a coefficient of zero and we will be left with an intercept-only model. We used a stepwise procedure to consider all possible combinations of candidate independent variables to choose the lowest Akaike Information Criterion (AIC; (Burnham and Anderson, 2002)). We used function stepwise in the “Rcmdr” library ( Fox, 2005) to select the combination of terms that provided the best fit to the data, with AIC score penalizing the addition of unnecessary terms.