Climate variables are not the dominant predictor of Arctic shorebird distributions

R package semEff [38]. The path coefficients we present are semi-partial correlations, which are the correlation between the unique components of predictors (residualised on other pre- dictors) and the response variable. Note that this approach controls for multicollinearity between predictors, whereas many path analyses present standardized partial regression coeffi- cients which do not control for multicollinearity, which often creates an upward bias in per- ceived effect sizes [39]. Model fit was assessed with Fisher’s C statistic, for which a significant value indicates that no potentially significant missing paths were excluded, as well as visually using plots of residuals calculated from the median coefficient values. To assess the relative strength of individual predictors, we compared the absolute values of their total effects on shorebird species occupancy. The total effect of a predictor is the sum of its direct effects (the path coefficients between the predictor and the shorebird response) and its indirect effects (the product of the path coefficients between the predictor, intermediate predictors, and the shore- bird response). Effects were considered to be significant when 95% confidence intervals did not contain 0. To assess the overall relative contribution of climate-only model predictors and additional predictors to shorebird occupancy, we used a deviance partitioning approach. We used Gener- alized Linear Models (GLMs) with a binomial distribution to determine the deviance

explained by climate and additional predictors. Species were analysed separately. Deviance was partitioned, using the ratio of the null deviance and the deviance explained by one of the following factors: climate predictors, additional predictors, deviance shared between climate and additional predictors, and unexplained deviance. Results We observed 11,636 shorebirds in surveys of 2336 plots from 1994–2018. At least one shore- bird was observed on 54% of plots. By species, occupancy varied from 1% (Red Knot) to 14% (Semipalmated Sandpiper); the average (±SD) species occupancy was 6 ± 4%. Descriptive sta- tistics of the environmental predictor values are in S2 Table. The mean scale of effect for land cover classes was 4400m (median = 3000m). For virtually all land cover classes, the scale of effect varied from the smallest possible scale (200m, just covering the plot) to the largest possi- ble scale (10km) depending on the species. Similarly, each species responded to the different land cover classes at very different spatial scales (S3 Table). The median scale of effect for snowmelt timing was 5000m (median = 3500m). Direct effects of climate-only model predictors accounted for on average 17% of the total explained deviance in shorebird occupancy (Fig 4). The total effect of climate-model predictors (direct effect + shared/indirect effect) accounted for on average 43% of the total explained deviance in shorebird occupancy. The unique effects of additional predictors included in cli- mate+additional models, beyond the indirect effects of climate through these predictors,

accounted for on average 57% of the total explained deviance in shorebird occupancy (Fig 4). Depending on the species, the ratio of deviance explained by climate-only predictors vs addi- tional predictors ranged from 80% climate/20% additional to 10% climate/90% additional. The mean explained deviance for our models of shorebird occupancy was 27% (ranging from 9%- 42%). The mean AUC of our models of shorebird occupancy was 0.87 (ranging from 0.72– 0.93). The full results of the GLMs used for deviance partitioning can be found in S4 Table. Overall, the total effect sizes of climate predictors (direct + indirect sizes; Fig 2) and addi- tional predictors on shorebird occupancy were similar (Fig 5). Our path model included the paths illustrated in Fig 2 (Fisher’s C statistic = 3.14, P = 0.21). The 5 predictors with the largest significant absolute effect on shorebird occupancy were additional predictors, specifically the predictors that described the proportion land cover at the species-specific scale of effect of PLOS ONE Climate variables are not the dominant predictor of Arctic shorebird distributions PLOS ONE | https://doi.org/10.1371/journal.pone.0285115 May 17, 2023 7 / 14