Statistical analysis
We used logit survival as the response variable in our models to account for nonlinearassociations with extrinsic and intrinsic predictors. Prior to analysis, we log10 transformed body mass and clutch size due to skewness, and scaled latitude and climate data to z scores by subtracting their mean and dividing by their standard deviation. Most variables were weakly correlated, although both measures of temperature reached Spearman rank correlations >0.75 (Table S1). To estimate adult survival rates along the latitudinal gradient, we used a multi-level meta-analytical framework using the R package metafor(Viechtbauer 2010), which fits random and fixed effects models, weighting effect sizes by the inverse of their squared standard error. For each model developed, we accounted for sources of non-independence in our dataset that can arise when multiple survival estimates are extracted from the same study, are available for the same species, and / or due to common ancestry, by fitting study identity, species identity, and phylogeny as random intercepts. To incorporate phylogeny, we used a majority rules consensus tree derived from a set of 1,000 randomly-selected trees based on the global phylogeny of birds (Jetzet al. 2012), and pruned to our study taxa (Fig. S1) with the R package phytools (Revell 2012). We used the branch length from this consensus tree to specify values for the model variance-covariance matrix.
We first ran a random effects only model on the entire dataset using therma.mv function to estimate a pooled effect size of global avian survival rates. Given potential differences in selection pressures experienced by passerines vs. nonpasserines, species from Old World (Afrotropics, Indomalayan, Palearctic) vs. New World (Neotropics, Nearctic) biogeographic realms, and mainland vs. island bird populations, we also evaluated separate meta-analytic models using effect sizes for these six data subsets. We considered point estimates to be different from one another if their 95% confidence intervals (CI) did not overlap. We quantified total heterogeneity of each dataset by calculating Cochran’s Q and I2statistics (Higgins & Thompson 2002).
To test for publication bias in our global dataset we used three complimentary methods: (1) We visually assessed asymmetry of funnel plots (Fig. S2); they appeared close to symmetrical. (2) We removed studies that reported survival estimates for >10 species, and which accounted for 64% of effect sizes, and reran the analysis. We repeated this procedure for studies conducted for <10 years to examine the effects of study duration on survival estimates. (3) We fit additional models where study method (i.e., live-recapture, dead recovery, or both) was used as a moderator or whether package aukwas used to calculate the geographic coordinates. Results of this sensitivity analysis were all qualitatively similar to the global mean survival rate based on the entire dataset (Fig. S3).
We conducted meta-regressions (meta-analyses incorporating explanatory variables, hereafter referred to as “moderators”) whereby we determined the effects on species-specific adult survival rates of (1) latitude, (2) extrinsic climatic factors, and (3) intrinsic traits in accordance with hypotheses described from the primary literature. We began by comparing fit of a latitude-only model, where regression slopes varied between hemispheres, to single-predictor linear models testing the influence of moderators on adult survival rates (Table S2). We next used AICC values (Burnham & Anderson 2002) to guide selection of a multi-predictor model of extrinsic climatic factors and intrinsic traits separately. Starting with the moderator that had the lowest AICC value, we sequentially added the next strongest moderator until AICC was no longer improved (Table S3). We considered the model that minimized AICCthe most appropriate if it had fewer parameters and was at least 2 AICC less than the next most competitive model (Arnold 2010). All of the intrinsic moderators we assessed improved model fit and were carried forward to the next step of model development. Temperature seasonality (Temp Seasonality) provided the best model fit for extrinsic moderators. We then combined both sets of moderators into a joint extrinsic / intrinsic model and repeated analysis using the global dataset and each of the six data subsets.