What can cetacean stranding records tell us? A study of UK and Irish cetacean diversity over the past 100 years

There are many factors that may explain why cetaceans (whales, dolphins, and porpoises) strand. Around the UK and Ireland, over 20,000 stranding records have been collected since 1913, resulting in one of the longest, continuous, systematic stranding data sets in the world. We use this data set to investigate temporal and spatial trends in cetacean strandings and use generalized additive models (GAMs) to investigate correlates of strandings. We find a dramatic increase in strandings since the 1980s, most likely due to increases Corresponding author (e-mail: ellen.coombs.14@ucl.ac.uk). MARINE MAMMAL SCIENCE, 000(000): 1–29 (2019) © 2019 Society for Marine Mammalogy DOI: 10.1111/mms.12610

. It is 71 therefore important to monitor cetaceans to determine the impacts of these pressures on 72 their abundance and behavior (Bejder et al. 2006). As with other marine species, cetaceans 73 can prove difficult to study as they are often wide-ranging and spend most of their lives 74 submerged under water (Evans and Hammond, 2004). Frequently employed monitoring 75 techniques, such as surveying from boats, are not only expensive and time consuming, but 76 are often biased towards conspicuous species or those that respond positively to boat 77 presence, such as bottlenose dolphins (Tursiops truncatus) and short-beaked common 78 dolphins (Delphinus delphis; Evans and Hammond, 2004). One approach to these 79 constraints is to use strandings data, i.e., records of cetaceans that have washed ashore.  (Hurrell, 1995) that can alter cetacean distributions and lead to strandings 128 (Simmonds andEliott, 2009, Schumann et al. 2013). Anthropogenic impacts such as military 129 sonar can cause cetaceans to surface quickly resulting in fatal decompression sickness 130 (Jepson et al. 2003 (Table 1) as an offset in the model. To further investigate the impacts of 274 sampling effort, we ran two case study models that look at differences in population between 275 the populated southern UK, and the less populated northern UK (Supplemental information;

276
Regional study 1 & 2). Smooths were modelled using a thin plate spline basis with shrinkage 277 (Marra and Wood, 2011), which allowed terms to be removed from the model (i.e., their 278 effect size shrunk to zero) during fitting, thus terms were selected during model fitting. As we 279 wanted to model species-specific effects, we included a factor-smooth interaction between 280 11 year of stranding and species; this term fitted a smooth of time for each species but allowed 281 common smooths to be fitted for the other covariates. An advantage of this approach is that 282 the per-species smooths are estimated as deviations from a base-level smooth, so some 283 information is shared between species. We fitted models with the following candidate 284 response count distributions: Poisson, quasi-Poisson, negative binomial, and Tweedie. We

415
Temporal and spatial patterns in strandings varied across and within species (Fig. 2, 3).

428
Overall, cetacean strandings records have increased over the past century, with a rapid rise 429 from the late 1980s to the present (Fig. 4). There were several prominent spikes in stranding 430 numbers before the 1990s (Fig. 4)  increase in stranding records for both mysticetes and odontocetes (Fig. 3, 4). Most strandings were of odontocetes, therefore the plot for odontocetes and all species 458 combined show a similar pattern (Fig. 4). Most strandings occurred around the south coast 459 of England and the west coasts of Ireland and Scotland (Fig. 5, S6). This pattern was 460 particularly evident in common dolphin and harbor porpoise strandings (Fig. S4, S5).

461
Stranding hotspots in southern and southwest England were first documented from 1926-462 1950 (Fig. 5). Over the next 25 yr

472
We found significant effects for NAO, SST, and fish catch (P < 0.05, P < 0.001, P = 0.02, 473 respectively) suggesting the smooth of these variables were significantly different from "no 474 effect" (Table 2). However, the estimated degrees of freedom (EDF) were very low (i.e., less 475 than, or not much greater than 1) indicating that the number of individuals that strand was 476 not strongly influenced by any of our predictor variables apart from year of stranding (Table   477 2, Fig. 6

487
We found significant P-values for some variables in our sensitivity analyses (see below for 488 details) suggesting the smooth of these variables were significantly different from "no effect".  (Tables S4, S5, S7-S12, S15 and Fig. S15, S16, S18-494 S23); therefore, we only report the differences below. All results are compiled in Table   495 S15.

497
Species identification models

527
We found a significant effect for maximum SST, and fishing catch (P < 0.005, P < 0.001, 528 respectively) (Table S11, Fig. S22) but otherwise the models for odontocetes and mysticetes 529 were qualitatively similar to those for the full dataset.

531
Habitat models 532 We found significant effects for maximum SST, and fishing catch (P = 0.001, P < 0.001, 533 respectively) but overall the results were the same as in the models without a habitat smooth 534 (Table S12, Fig. S23). 535 536 Regional models 537 The two regional models had different EDFs, with higher EDFs found in the southwest (region 1) 538 model (Table 3). We found significant P-values for all of the variables except for maximum k-index 539 21 and maximum SST in both models (Table S13, S14). The region 1 model had an EDF of 6.62 for 540 NAO but the relationship was not particularly "wiggly". We therefore interpret this as having little effect 541 on the number of stranded individuals (Wood, 2017). The EDFs for the other variables were still too 542 low to be fully conclusive (Table 3, Fig. S24, S25).  (Vanselow et al. 2017) and that these regional and species-specific definitions were not 658 investigated in our macroecological study. We did not find a correlation between 659 geomagnetic fluctuations and strandings in our regional models, perhaps because these 660 effects are population, or season-specific.

676
We found only a slight correlation between SST and stranding records in our main model.

677
The EDF was so low, that this is not a conclusive correlate of strandings.

682
We also found no effects of SST on strandings in our regional models (southwest UK and 683 northwest UK

741
We also highlight that the spatio-temporal difference between the death of the animal and its