Methods
Population time series data
We obtain populations monitored to extirpation from two sources: i) the Living Planet database (LPD) (http://www.livingplanetindex.org/data_portal), containing annual population abundance data for over 25000 vertebrate populations between 1950-2019 and ii) from previously published work on the extinction vortex (Fagan & Holmes 2006). A diverse range of methods to monitor population abundance are included in the LPD, with the caveat to inclusion in the dataset being that monitoring should be reputable, appropriate for the species and consistent through time. A detailed outline of inclusion criteria for the populations in the LPD are provided by Loh et al. (2005). In some cases, complete censuses of the population were carried out, whereas in others population abundance was monitored using indirect indicators. In the absence of evidence to the contrary, regardless of the method used in population monitoring, we assume that indices of population abundance are representative of the true population size at any given point in time.
We define extirpation as a population declining to a zero-abundance count at the end of the time series and identify populations from the LPD that showed this. Zero-abundance counts occurring before the end of the time series might indicate a relatively low species detectability and, correspondingly, a high rate of observation error (Brook et al. 2006). To minimize the possibility of including populations that were not actually extinct and to avoid inflating annual variation in population abundance, we omit time series where zero counts occurred and were followed by subsequent observations. In addition, we only consider populations where the time between the penultimate abundance count and the zero-abundance count (signifying extirpation) was no more than one year, so that we can ascertain the exact year in which the population went extinct. Furthermore, to avoid introducing possible bias from short time series, we only include time series with at least 10 counts of population abundance.
Based on these filtering criteria, we produce a dataset of 55 population extirpations of 52 different species, including two elasmobranchs, five actinopterygians, one amphibian, one reptile, nine mammals and 34 birds. Our dataset of time series has a mean length of 15.98 (±6.65) years.
Life history data
We compile life history data for all species in this dataset from various life history databases (Myhrvold et al. 2015; Oliveira et al. 2017; Froese & Pauly 2000), extracting data on log-transformed (base 10) adult body mass in kg. Additionally, where possible, we collate up to six other traits indicative of life history speed: maximum longevity, female maturity, incubation time, fledgling age, number of litters per year and litter/clutch size. We log-transform (base 10) these trait values and investigate their relationship with body mass using linear models.
LMM/GLMMs
We perform all statistical analyses using R version 3.6.1 (R Core Team, 2019). We perform three statistical analyses to investigate how population dynamics change in the region of an extinction event. For each analysis we use linear or generalized linear mixed effects models (LMMs/GLMMs) in the ‘nlme’ (Pinheiro et al. 2019) and ‘glmmTMB’ (Brooks et al. 2017) packages respectively, to account for context-specific factors that could mask the effect size of fixed effects on the response variables. We account for the nested random effects of our data using a mixed modelling framework, with population nested inside species, nested inside units, nested inside data type. This accounts for the site-specific effects on the population dynamics, the potential effects of relatedness at the species level and the potential effects of units of measurement. These are nested within data type, which is consistent across species and populations, but may vary between species and units. We present full descriptions of the fixed and nested random effects used in our LMMs/GLMMs in Table 1. We also normalize our fixed effects to enable easier comparison of the relative importance of each variable in the models. As avian taxa are overwhelmingly represented in the dataset (65.38% of species), we perform our analyses on all populations together, as well as two subsets: i) only avian populations and ii) only non-avian populations. For each analysis, we employ AIC-based model selection to identify the best fitting model using the ‘MuMIn’ package (Bartón 2019).