Life-history predicts global population responses to the weather in the terrestrial mammals

With the looming threat of abrupt ecological disruption due to a changing climate, predicting which species are most vulnerable to environmental change is critical. The life-history of a species is an evolved response to its environmental context, and therefore a promising candidate for explaining differences in climate-change responses. However, we urgently need broad empirical assessments from across the worlds ecosystems to explore these predictions. Here, we use long-term abundance records from 157 species of terrestrial mammal and a two-step Bayesian meta-regression framework to investigate the link between annual weather anomalies, population growth rates, and species-level life-history. Overall, we found no consistent effect of temperature or precipitation anomalies on annual population growth rates. Furthermore, population responses to weather anomalies were not predicted by phylogenetic covariance, and instead there was variability in weather responses for populations within a species. Crucially, however, long-lived mammals with smaller litter sizes had responses with a reduced absolute magnitude compared to their shorter-living counterparts with larger litters. These results highlight the role of species-level life-history in driving responses to the environment.


Introduction
characterising the life cycle of an organism to quantify their resilience to perturbations (Capdevila et In this study, we investigated annual population responses to temperature and precipitation in populations of terrestrial mammals across the world's ecosystems. Importantly, we tested whether lifehow observed annual population growth rates were influenced by weather anomalies (annual deviation 115 from long-term average weather patterns) using autoregressive additive models that accounted for 116 temporal autocorrelation in abundance records and overall abundance trends. Then, we used a 117 phylogenetically controlled Bayesian meta-regression with weather effect coefficients as the response 118 variable to address three key questions: 1) Are there consistent temperature and precipitation effects on 119 abundance change across the terrestrial mammals? 2) How are these patterns influenced by covariance 120 both within and between species, and are there vulnerable biomes? 3) Can species-level life-history 121 traits predict the magnitude of population responses to the weather? The terrestrial mammals are an 122 ideal study system to explore the predictors of population responses to climate change because they are 123 a well-studied group with a combination of intensive abundance monitoring across the globe (Almond 124 et al., 2020), detailed life-history information for hundreds of species (Conde et al., 2019;Myhrvold et 125 al., 2015) and a highly resolved phylogeny to facilitate phylogenetic comparative analyses (Upham et 126 al., 2019). Furthermore, there is growing evidence from the mammals of the mechanistic links between 127 the climate, demography, and population dynamics (Coulson et al., 2001;Paniw et al., 2019Paniw et al., , 2021;

131
We assessed population responses to weather in 486 long-term abundance time-series records of 157 132 species of terrestrial mammals from across the world's ecosystems (Fig. 1). The time-series records 133 ranged in duration from 10 years to 35 years, with mean and median record lengths across records of 134 15.7 and 14 years, respectively (Fig. 1). The records were distributed across 13 terrestrial biomes, 135 including both tropical and temperate regions, but were generally biased towards north western Europe 136 and North America. We had records from 12 of 27 mammalian orders recognised by the IUCN Red

162
The size of the point gives the record duration in years. The histogram in the bottom left gives the distribution of record lengths across the whole dataset. The 163 bar graph in the bottom right is a frequency distribution of each of the mammal orders analysed in the current study.

No consistent population response to weather
Overall, we found no consistent effect of either temperature or precipitation anomalies on annual population growth rates in the terrestrial mammals (Fig.2). The raw weather effects on population coefficient of -0.32 (±1.76 SD) and mean precipitation coefficient of 0.07 (±0.81 SD). Furthermore,

169
95% of records had temperature and precipitation coefficients between -4.29-3.17 and -1.41-1.88, respectively. Nevertheless, approximately 8% (n = 42) of temperature effects and 1% of precipitation 171 were greater than 3 or less than -3, indicating that small clusters of populations experienced more 172 extreme annual responses to the weather (Fig.2). Our Bayesian meta-regression, controlling for both within species variance, phylogenetic covariance and differences in sample size (number of years) 174 between records, mirrored the lack of consistent weather effects on population growth. The posterior 175 mean global intercept, ̅, for temperature effects was 0.02 [-0.21-0.25

183
Spatial effects and variation between species

184
We tested whether there were differences in weather responses among ecological biomes because biome 185 effects may be indicative of more extreme responses to weather in some habitats. Using leave-one-out 186 cross-validation, we compared the predictive performance of the model including the effect of biome 187 relative to the base model, and we found no evidence for an influence of biome on either temperature 188 (Δelpd = -0.67 relative to base model) or precipitation (Δelpd = -0.73) effects (see Fig. S16-17 for more 189 information). Furthermore, we explored the role of spatial autocorrelation at driving differences in 190 weather coefficients across records using Morans I tests and spatially explicit meta-regressions but did 191 not find evidence for spatial autocorrelation in weather effects (Figs. S19-S21). We also incorporated 192 both phylogenetic covariance ( 2 ) and species-level variance ( ) to capture both amongand within-species variation. Interestingly, we found far greater levels of within-species variation in 194 temperature responses compared to among-species variance (Fig. 2c) Table S1 & S1). However, univariate 221 models including litter size also had a higher predictive performance than the base model (Δelpd = 3.98

222
and Δelpd = 0.8 for temperature and precipitation, respectively). For temperature, the second-best 223 predictive model was the one that included univariate effects for longevity, bodymass and litter size 224 (Δelpd = 4.54; Table S1), and this model was also competitive for precipitation (Δelpd = 0.69; Table   225 S2). Therefore, in both cases we selected the models including all univariate life-history effects.

226
For both temperature and precipitation, our results highlight that shorter-living mammals with 227 greater litter sizes experienced weather effects of a greater magnitude than longer-living, slowly 228 reproducing mammals (Fig. 3). The magnitude of weather responses was negatively associated with

320
Two points are omitted from the plotting panel due to large mean coefficient values and high standard 321 errors, which are visible on the plot.

Discussion
Our results provide an important empirical link between a species' life-history and its population responses to environmental change. While we found no consistent patterns of responses to temperature and precipitation anomalies across the mammals, life-history traits relating to the pace of life were associated with responses to weather. Namely, shorter-living species with increased litters sizes, or species characterised with 'fast' life-history traits, responded with a greater magnitude compared to those with 'slow' life-history traits. While it has long been theorised that an organism's life-history traits evolve in response to, and as an adaptation to, environmental conditions (Stearns, 1992), rarely has this theory been tested at a global scale. We find strong support for the hypothesis that longevity, and 'slow' life-history characteristics more generally, buffer organisms against short-term variability in the environment (Morris et al., 2008) and add to a small number of studies linking population demography and the climate (Compagnoni et al., 2021;Paniw et al., 2021). We predict that in the short term abrupt ecological disruption from climate change will have a disproportionate impact the abundance of shorter-lived species with higher reproductive output. We do not argue that long-lived species are less vulnerable to climate change. Over longer time-scales, species with slow life-history are also slower to recover from perturbations (Gamelon et al., 2014). However, critically our results highlight the potential utility of life-history traits for predicting species vulnerability to climate change.
Demography has a vital role to play in predicting population declines in the Anthropocene and in highlighting targets for conservation management (Conde et al., 2019;Richards et al., 2021). Our study emphasises this role, demonstrating the predictive power of life-history traits when investigating responses to environmental change. However, there are limitations and barriers to the utility of demography in conservation. Only 1.3% of tetrapods globally have sufficient demographic information with which to estimate population dynamics (Conde et al., 2019). Here, we used summary traits that are available for many species (maximum longevity and mean litter size), but in particular maximum recorded longevity, while sufficient as a broad indicator, is strongly influenced by sampling variance and a flawed measure of longevity differences between taxa (Moorad et al., 2012). Ideally, lifetables with mortality and reproduction trajectories across the lifecycle can be combined with data on external drivers to investigate detailed patterns in population dynamics, rather than relying on abundance trends (Desforges et al., 2018;Jackson et al., 2019). The recent development of the demographic resilience framework, which uses demographic data across the lifecycle to simulate how a population may respond to perturbations (Capdevila et al., 2020), has excellent potential in extending these findings to explore demographic relationships with climate responses in detail. Unfortunately, however, detailed (st)agespecific demographic information is not currently available for a majority of species, but growing in availability rapidly (Salguero-Gómez et al., 2016). Therefore, there is a need to continue to increase the collection of demographic data (and other traits) for many more species than are currently available (Conde et al., 2019), so that we may predict population changes with respect to environmental change.
Achieving this target may revolutionise the way we quantify species vulnerability to climate change (Antão et al., 2020;Dornelas et al., 2019;Leung et al., 2020;Paniw et al., 2021), helping to prevent extinctions before they occur.
In line with recent global assessments of biodiversity in the face of climatic change (Paniw et al., 2021), we did not find an overall consistent effect of weather anomalies on population growth rates.
This may in part reflect the fact that abundance changes are a higher-order process determined by complex interactions between demographic processes that counteract each other (Leung et al., 2020;Paniw et al., 2021). However, our results contrast with findings of linear associations between mammal abundance and temperature change (Spooner et al., 2018). These differences may reflect our approach to investigate annual changes, rather than long-term trends. Significant population trends from long time-series are detectable from smaller component time-series even when sampling is incomplete (Wauchope et al., 2019), and thus responses detected in trends may reflect broader changes in response to the climate that are not detected in models of annual change. Furthermore, we estimated linear, annual effects of weather on population growth rates, where population responses may actually be more complex non-linear patterns or lagged effects. However, the detection of climate effects on average trends may also be confounded by effects of other (sometimes more dominant) drivers (e.g. habitat loss) (Daskalova, Myers-Smith, Bjorkman, et al., 2020). Nevertheless, our findings can be explained in light of recent studies from the Living Planet Database that have found that the large majority of records do not exhibit population declines (Leung et al., 2020).
Interestingly, we did not find evidence for phylogenetic covariance in weather responses between species. Recent evidence from birds indicated strong phylogenetic covariance in vital rates, particularly in adult survival, and the incorporation of phylogenetic information greatly improved predictive performance when imputing vital rates (James et al., 2020). Therefore, as with overall patterns, our findings may reflect the trade-offs between vital rates, which cancel one another out when scaling up to population-level processes such as population growth rates in response to the weather . Furthermore, for long-term time-series, there may also be temporal trade-offs in vital rates, where for example investing heavily into survival in one year (in response to climate) may impact subsequent reproduction for several years, decreasing the magnitude of population growth rates.
The extent of phylogenetic covariance in vital rate responses and trade-offs remains unknown, understanding how the climate impacts demographic rates across species may provide a useful tool for imputing population responses to the climate across the tree of life (James et al., 2020).
We highlight the importance of variation in population responses to climate within a species range. Sampling heterogeneity has recently been shown to have broad implications for metrics of population dynamics, where demographic rates are poorly correlated among sampling sites for the same species (Engbo et al., 2020;Römer et al., 2021). Therefore, inferences obtained from monitoring single populations or studies may not accurately portray species-level variability. This has broad implications for macroecology, particularly for population viability assessments (PVA) and species-distribution modelling. First, as well as suffering from data quality issues in their parameterisation (Chaudhary & Oli, 2020), our findings suggest that PVAs based on data from a single population may not accurately reflect population viability across a species' geographic range. Therefore, incorporating detailed demographic data, and investigating differences in population responses across a range, could greatly improve our perspective on population viability (Desforges et al., 2018). Second, presence-only models of species distributions that do not account for the fact that responses to the environment within a species range do not accurately represent species distributions (Benito Garzón et al., 2019). Moving towards trait-based monitoring and explicitly including demographic processes with mechanistic links to appropriate drivers into species distribution models could greatly improve predictions of climate change impacts on the biosphere (Trisos et al., 2020).
Ultimately, improving our predictions of how humans are influencing the natural world is paramount to prevent rapid declines to global biodiversity (Kissling et al., 2018). This however requires a large shift towards both broad and detailed monitoring of global biodiversity. We show that linking species traits such as life-history to changes in the environment may equip us with tools to predict and prevent future losses.

Materials and Methods
To assess the effects of weather on population growth rates we collated information on global weather and the abundance, life-history and phylogeny of the terrestrial mammals. All analyses were carried out using R version 4.0.5 (R Core Team, 2021). For all data on the terrestrial mammals, taxonomies were resolved using the taxize package version 0.9.98 (Chamberlain et al., 2020) and matched using the Global Biodiversity Information Facility database (https://www.gbif.org/). All code used in the current study and full descriptions of the analyses are archived in the Zenodo repository (doi:10.5281/zenodo.4707232), which was created from the following GitHub repository https://github.com/jjackson-eco/mammal_weather_lifehistory.

Data selection
For full descriptions of the data selection process please refer to S1. Long-term annual time-series abundance data from across the terrestrial mammals were obtained from the Living Planet Database found at https://livingplanetindex.org/data_portal. Abundance is measured in several ways (e.g. population counts and density, which does not impact population responses, see Fig. S19), and so we natural-log-transformed population growth rates to ensure that weather effects were comparable across records. Our final dataset contained 486 geo-referenced records from 157 terrestrial mammal species, which was used in all subsequent analyses (Fig. 1).
Global weather data was obtained from version 1.2.1 of the CHELSA monthly gridded temperature and precipitation dataset at a spatial resolution of 30 arc seconds (~ 1km 2 ) for all months between 1979-2013 across the globe's land surface (Karger et al., 2017). Generally, we expect that organisms will respond to deviations in the weather compared to the average values, as opposed to raw weather variables. Furthermore, across the globes surface the variance in weather variables changes substantially, which may influence population responses. Thus, we explored population responses for the key weather variable of standardised annual anomalies, and then validated our approach using annual weather variance. These weather anomalies are the average deviation of the temperature and precipitation from expected values in a given year.
We used three key species-level life-history traits that are available for a large number of species: maximum longevity, mean litter size and mean adult body mass. Life-history data were collected from

Weather effects on annual population growth rates
To assess comparative population responses to weather in the terrestrial mammals we used a two-step meta-regression approach. First, for each record we estimated the effect of annual weather anomalies (and weather variance) on population growth rates. We calculated the standardised proportional population growth rate in year as where is the abundance in year , transformed to prevent observations of 0.
Then, with as the response variable, we estimated the effect of temperature and precipitation anomalies on population growth using generalised additive mixed models (GAMMs) fit using the gamm function of the mgcv package (Wood, 2017). We opted to use a general linear-modelling framework as opposed to a state-space approach, which is often employed for time-series to account for measurement error and estimate trends (see Daskalova, Myers-Smith, & Godlee, 2020). The primary reason for this choice was that we aimed to assess broad comparative patterns in population change, and did not expect systematic errors in model parameters due to measurement error. Furthermore, Daskalova, Myers-Smith, & Godlee (2020) found that abundance trend terms were highly correlated between linear and state-space approaches across the LPD, which would be expected if there are not systematic errors in measurement across the database. We did however test the implication of this choice by employing a state-space approach (see S2).
Changes in abundance are influenced by several drivers of population dynamics including habitat loss (Daskalova, Myers-Smith, Bjorkman, et al., 2020) and population processes such as density dependence (Brook & Bradshaw, 2006), which may confound any influence of the weather on abundance. Therefore, because we aimed to assess the isolated impact of weather anomalies, accounting for these trends in abundance and temporal autocorrelation was crucial. We initially explored the extent of autocorrelation in abundance patterns using timeseries analysis and found evidence for lag 1 autocorrelation in abundance, but not for greater lags (Fig. S3&S4). Furthermore, we tested the potential impact of density dependence on estimating environmental effects using an autoregressive timeseries simulation and found that environmental effects were robust to density dependence even for short timeseries (Fig. S5). Thus, for each record, we model population growth rate in each year as where 0 is the intercept and is a linear parametric term with coefficient for the weather (temperature or precipitation anomaly) in year . Here, positive coefficients indicate that positive weather anomalies i.e. hotter/wetter years, were associated with population increases, and vice versa.
Identical additive regression models were run using weather variances as the weather variable . The term ( ) captures the effect of year as a non-linear trend, where the function is a thin plate regression spline with a basis dimension of five (Wood, 2003). The function was also fitted with an order 1 autoregressive (AR(1)) correlation structure, as specified in the nlme package (Pinheiro et al., 2014). Thus, the term ( ) incorporates both the non-linear trend in abundance and temporal autocorrelation.
Finally, we validated our additive model approach by testing other models to calculate weather effects, including linear regressions both including and excluding temporal trends or density dependence, state-space models, and a temporally autocorrelated model fit using the glmmTMB package (Brooks et al., 2017) (S2; Fig. S7-S11). Weather coefficients generated using linear year effects were positively correlated to those from additive models (Fig. S9), and additive model coefficients were highly correlated with those from state-space models (Fig. S10 & S11).

Bayesian meta-regression
Second, with the weather effects from each record as the response variable, we explored comparative patterns in population responses to weather using a Bayesian meta-regression framework implemented in the brms package (Bürkner, 2017). Separate models were fit for temperature and precipitation.
Bayesian meta-regression was used to address three key questions: 1) Were there consistent population responses to weather across the terrestrial mammals? 2) How did population responses vary within and between species and were there spatial patterns across biomes? 3) Does life-history predict the magnitude of population responses? To address questions 1 and 2, we used Gaussian models controlling for both phylogenetic and species-level covariance. The full model for record and species is given by equation 3 below where the weather effect (z-transformed for analyses), is given by a multivariate normal distribution with mean and phylogenetic covariance matrix . The global intercept is given by ̅, which estimates overall patterns in weather effects across records, addressing question 1. We incorporated phylogenetic covariance using a Brownian motion model, with the correlation matrix given by (calculated from the maximum clade-credibility tree) and variance factor 2 , from which between-species variance was estimated. We incorporated an intercept-only varying effect for species with the term [ ] , from which within-species variance was estimated with . The term gives the spatial effect of biome on weather responses. Thus, estimating within ( ) species variance, between ( 2 ) species variance and the spatial effect of biome ( ), we explored question 2. All metaregression models also included the linear effect of record length (scaled number of years in the record) on weather effects, which was estimated using . For all meta-regression models, we used regularising priors obtained from prior predictive simulations of the slope, intercept and exponential variance terms (McElreath, 2020a(McElreath, , 2020b, to reflect the constraints in the raw data across species (see S3 and Fig. S12-S15 for details). Gaussian meta-regression models were also fit for weather effects calculated using the annual weather variance, and the results obtained were largely identical to those obtained for weather anomalies (Fig. S22).
For question 3, although on average we expect that species life-history influences the magnitude of responses to the environment, we have little evidence to suggest that life-history per se influences the directionality of responses (Morris et al., 2008). Thus, to address this question we explored how maximum longevity, litter size and adult body mass influenced the absolute magnitude of weather responses, | |, using Gamma regression models with a log link. The full model for record and species is given by equation 4 below where is a shape parameter that was fit with a Gamma prior, and refers to a set of linear lifehistory terms ( 1 1 + ⋯ ) that were explored using model selection. Specifically, for the three lifehistory traits, we explored a set of models incorporating univariate, multivariate and 2-way interaction terms, as well as a base model excluding all life-history effects. For the full set of ten candidate models please refer to the supplementary information (table S1 & S2). All life-history effects were fit with the same Normal prior, with mean 0 and standard deviation 0.3 (S2; Fig. S14). We assessed the predictive performance of candidate models using leave-one-out cross-validation implemented in the loo package (Vehtari et al., 2017). Models were compared using the Bayesian LOO estimate of out-of-sample predictive performance, or the expected log pointwise predictive density (elpd) (Vehtari et al., 2017).
All final meta-regression models were run over 3 Markov chains, with 4000 total iterations and 2000 warmup iterations per chain. Model convergence was assessed by inspecting Markov chains, and the degree of mixing between chains using ̂.