2.1 Global trends in genetic variation

We used three separate paired t tests to examine whether the mean observed heterozygosity, expected heterozygosity, and allelic diversity differed between time points. Pairs consisted of the modern and historic time-point values of allelic diversity, expected heterozygosity or observed heterozygosity for each species in each independent study that spanned more than one generation. Thus, species that were studied multiple times were initially considered as independent pairs. Paired analyses were necessary because they preserved the vital connection between the time points from the same study. This ensured that we assessed changes in diversity using the unique historical baseline for each modern estimate. A paired analysis is also technically necessary because each marker set will suffer from different ascertainment biases and study-specific limitations. Furthermore, paired data points are not independent and should not be separated. To meet the assumption of normality, we log-transformed the values for each time point (results were stable without transformation). Due to multiple allelic diversity metrics, it was also necessary to standardize the values of allelic diversity by dividing both the historic and modern values by the historic value. The difference in variance that causes between historic and modern values does not violate the assumptions of the paired t test, because it is the difference between the paired values that the test assumptions apply to (McDonald, 2014). The standardization was not necessary for heterozygosity estimates because they were always reported as expected or observed heterozygosity. Five species were present in multiple studies (n = 13). These values were not initially excluded and were treated independently because the studies are independent and the populations were different evolutionary significant units. However, to control for potential bias due to unequal taxonomic representation, we averaged across each taxonomic class and repeated the paired t tests on this reduced data set.

2.2 Factors affecting loss of genetic variation

To examine factors influencing the total change between time points in expected heterozygosity and allelic diversity, we used linear mixed-effects models (lme4, version 1.1.18; Bates, Maechler, Bolker, & Walker, 2015) and backwards stepwise deletion. Observed heterozygosity was not examined because it did not change significantly between time points. To account for unequal taxonomic representation, taxonomic class was included as a random effect. The bottleneck status, age of the oldest sample, if the focal populations were on the island or continental mainland, and number of generations between time points were included as explanatory variables in the full model. The full model contained all listed variables and no interactions because balanced replication was not present across the factors. The response variables were scaled (i.e. the total change in each metric was divided by the historical value), because relative loss is biologically important and this ensured the response variables met the model assumptions. The linear mixed-effects model examining change in allelic diversity was constrained to allelic richness. This is because the different allelic diversity metrics cannot be combined in this analysis, as they are highly variable when combined. Model selection was performed by removing the variable with lowest t value from each nested model, followed by confirming the variable's nonsignificance using a likelihood ratio test (chi-squared). This was repeated until the final model contained only terms that resulted in significant change in the deviance of the two nested models (i.e. p < 0.05 and a t value > 2). Both models only examined studies spanning more than one generation. To ensure the data frame did not change with variable removal, species without a clear Red List status or bottleneck status were removed from the data set. Values of percentage change reported from the linear mixed-effects models are for this reduced data set (n = 64, expected heterozygosity; n = 43, allelic richness).

A chi-squared test was conducted to test whether deviation from Hardy–Weinberg equilibrium proportions was more common in historic or modern time points. This was to ensure the trends that we identified were not driven by inflated historical estimates of diversity due to false alleles from error-prone historic samples (Wandeler, Hoeck, & Keller, 2007). We considered a deviation from Hardy–Weinberg equilibrium to have occurred if any marker proportions were significantly in disequilibrium after corrections for multiple testing. We also required that the tests were separated by time point. If the tests were conducted on data combined across time points, the study was scored as “NA.” The total number of marker proportions not in equilibrium was not considered, as this was typically low (1–2 markers).

All analyses listed above were conducted in R (v.3.5.0; R Core Team, 2018). All tests were two-sided.

3.1 Global trends in genetic variation

We documented an average 5.4% (±18.8% SD) global decrease in mean expected heterozygosity (t = 2.74, mean of historical time point (log) = −0.590, mean of modern time point (log) = −0.672, mean of differences = 0.081, 95% confidence interval (CI) = 0.022–0.141, df = 69, p = 0.008, paired t test). A slightly higher decrease of 6.5% (±17.8% SD) was seen for mean allelic diversity (t = 3.46, mean of historical time point (log) = 0, mean of modern time point (log) = −0.094, mean of differences = 0.094, CI = 0.040–0.148, df = 72, p = 0.0009, paired t test). In contrast to expected heterozygosity and allelic richness, observed heterozygosity did not change significantly (nonsignificant increase of 1% ±19% SD; t = 0.42, mean of historical time point (log) = −0.7, mean of modern time point (log) = −0.713, mean of differences = 0.013, CI = −0.047–0.072, df = 60, p = 0.67, paired t test). Across taxonomic averages, we also identified a significant decline in expected heterozygosity of 3.8% (±4.9% SD; t = 2.47, mean of historical time point (log) = −0.560, mean of modern time point (log) −0.620, mean of differences = 0.04, CI = 0.0033–0.0771, df = 9, p = 0.0358, paired t test) and decline in allelic diversity of 3.7% (±10.1% SD). However, this was nonsignificant (t = 1.36, mean of historical time point (log) = 0, mean of modern time point (log) = −0.05, mean of differences = 0.049, CI = −0.0342–0.1340, df = 8, p = 0.2084, paired t test). Trends in observed heterozygosity remained nonsignificant (t = −1.51, mean of historical time point (log) = −0.664, mean of modern time point (log) = −0.649, mean of differences = −0.015, CI = −0.0375–0.0078, df = 8, p = 0.1686).

3.2 Factors affecting loss of genetic variation

The magnitude of change in expected heterozygosity was significantly associated with the age of the oldest sample in a study (ê = −0.0011 ± 0.0005 SE, t = −2.307, GLMM) and whether the populations occurred on an island or mainland (ê = 0.1908 ± 0.0671 SE, t = −2.846, GLMM). The final model contained only these two explanatory variables and taxonomic class as a random factor. All other variables in the model were nonsignificant, including the IUCN Red List of Threatened Species status (Endangered: ê = −0.0573 ± 0.1141 SE, t = −0.502; Vulnerable: ê = −0.0513 ± 0.1197 SE, t = −0.429; Near Threatened: ê = −0.0227 ± 0.1194 SE, t = −0.190; Least Concern: ê = −0.1036 ± 0.1086 SE, t = −0.945; Invasive: ê = −0.0899 ± 0.148 SE, t = −0.607, GLMM), bottleneck status (ê = 0.0195 ± 0.0569 SE, t = 0.343, GLMM) and number of generations a study spanned (ê = 0.0004 ± 0.0014 SE, t = 0.266, GLMM). There was a negative relationship between the total expected heterozygosity lost and age of the oldest sample (Figure 1), meaning that studies from closer to the industrial revolution showed significantly greater loss of expected heterozygosity. Island species also showed a significantly larger loss of expected heterozygosity, declining by an average of 27.6% (±28.1% SD) relative to an average loss of 2.1% (±15.8 SD) in mainland populations.

The analysis of allelic diversity (measured by allelic richness) yielded similar results to those for expected heterozygosity. No significant effect was found for IUCN Red List of Threatened Species status (Endangered: ê = −0.1752 ± 0.0866 SE, t = −2.023; Vulnerable: ê = −0.0351 ± 0.0855 SE, t = −0.410; Near Threatened: ê = −0.1075 ± 0.0832 SE, t = −1.291; Least Concern: ê = −0.1814 ± 0.0824 SE, t = −2.201; Invasive: ê = −0.1975 ± 0.1038 SE, t = −1.903, GLMM) or bottleneck status (ê = 0.0007 ± 0.0044 SE, t = 0.161, GLMM). In contrast to expected heterozygosity, there was also no significant effect of the age of the oldest sample (ê = −3^e−05 ±7^e−05 SE, t = −0.436, GLMM). Instead, the number of generations a study spanned had a significant association with loss of allelic diversity (ê = −0.001 ± 0.0004 SE, t = −2.307, GLMM). However, the trends across the two metrics were similar, and studies spanning more generations showed greater loss of allelic diversity (Figure 2). Island species (30.9% ±26.1 SD) also showed substantially greater loss of allelic diversity relative to those on a mainland (4.6% ±12.3 SD) (ê = −0.1908 ± 0.0671 SE, t = −2.846). Finally, we found that deviations from Hardy–Weinberg equilibrium proportions were not more common in either time point (χ² = 1.4104, df = 1, p = 0.235, chi-squared test).

4.1 Global trends in genetic variation

We identified a significant loss in allelic diversity (6.5%) and expected heterozygosity (5.4%) of modern populations based on data from a comprehensive literature review. The overall trends of decline in expected heterozygosity and allelic diversity are consistent with previous single time-point studies showing a putative global decline in genetic variation (DiBattista, 2008; Miraldo et al., 2016). However, the loss of allelic diversity identified here (6.5%) is smaller in magnitude than a difference previously reported between fragmented and unfragmented natural populations (29%) (DiBattista, 2008). The discrepancy between the two measures can be attributed to the difference in study designs: here, we considered all cross-generational studies, even those where environmental conditions remained essentially constant, whereas the previous analysis exclusively examined single time-point studies and only compared populations with a known environmental difference. The decline in genetic variation identified in this study is also less severe than many estimates of ongoing population declines. Wild vertebrate populations, for example, are estimated to have declined by 58% since 1970, at a rate of 2% per year (World Wildlife Fund, 2016), while common farmland bird populations have declined by 49% (see Pereira, Navarro, and Martins (2012) figure 4b for a clear summary across different metrics). We observed a baseline decline of ~6% in variation across an average of 110 years. This equates to a decline of ~0.05% per year. The slower loss of genetic variation relative to population size is anticipated because loss of genetic variation is broadly linked to genetically effective population size and not census size (Allendorf et al., 2013). It should be noted that this annual decline of ~0.05% should only be used comparatively because, as discussed below, loss of genetic variation is unlikely to occur evenly across time.

In contrast to the other metrics, little to no change was identified in observed heterozygosity. The unequal trends in expected and observed heterozygosity cannot be attributed to either a failure to identify alleles or increased error in historical samples, because deviations from Hardy–Weinberg equilibrium proportions were not more likely in the historic time points. Instead, the mismatch in trends is likely fuelled by an excess of observed heterozygotes in modern populations. This can arise in small or declining populations because the strong genetic drift they experience not only leads to unequal allele frequencies across populations, but can lead to unequal frequencies across the sexes (Rasmussen, 1979; Robertson, 1965). As a result of the frequency mismatch across the sexes, the observed heterozygosity in offspring cohorts will often be greater than that expected under Hardy–Weinberg equilibrium, potentially leading to the trends observed. Many historical time points were also represented by a small number of samples. While expected heterozygosity and measures of allelic diversity (notably allelic richness) are robust to small sample sizes, observed heterozygosity will be underestimated in such scenarios (Pruett & Winker, 2008). This may have lead to underestimated historical value of observed heterozygosity, preventing us from identifying any change in heterozygosity.

The studies that met our criteria were taxonomically biased. For instance, commercially important fish from the class Actinopterygii comprised 31% of our studies. Unsurprisingly, mammals and birds were also well represented, encompassing 14% and 24% of studies, respectively. Worryingly, few studies on insects (n = 12), amphibians (n = 3) or reptiles (n = 2) were found. Plant-based studies were also not in this data set because they often failed to meet our inclusion criteria. A strong taxonomic skew towards commercially important fish, where stocking has been used over the last century, could be biasing the observed trends in genetic diversity and decreasing the loss we observe. Thus, we also compared variation trends averaged across taxonomic classes. The previously identified trends remained reasonably stable: expected heterozygosity declined by an average of 3.8% (significantly) and allelic diversity declined by 3.7% (nonsignificantly). We recommend that the taxonomic gaps be addressed, given the importance of understanding diversity trends in different taxa, for instance, in crop pollinators and nonagricultural plants.

4.2 Factors affecting loss of genetic variation

An approximate decline of 5.4%–6.5% in genetic variation represents a global average, and much larger declines could be occurring in some species, geographical regions or time periods. We therefore examined what factors were correlated with the loss of genetic variation using a general linear mixed-effects model with taxonomic class as a random variable. The loss of neither allelic diversity nor expected heterozygosity was related to species’ IUCN Red List of Threatened Species status (International Union for Conservation of Nature, 2018). Each category of IUCN Red List of Threatened Species was well represented across the studies. For example, 37% of studies in our data set were on species of Least Concern, indicating that a decline in genetic diversity is occurring even for species of low conservation concern and that the trend of decline is not due to a bias towards Critically Endangered or Endangered species (21% of the data set). Recent research has shown that there is similarly little correlation between Red List ranking and contemporary genome-wide heterozygosity, where historical bottlenecks and species-specific life history traits were speculated to be more important (Díez-del-Molino et al., 2018). However, in our analysis bottleneck status also had a nonsignificant effect on the magnitude of variation lost, though metrics of genetic variation were lower in bottlenecked species. Due to the opportunistic nature of historical samples, many studies on bottlenecked populations may have been unable to sample before the declines began. If the historical baseline for genetic variation metrics was taken during or after a bottleneck, the largest effects on variation may have already occurred, preventing us from observing an effect.

The change in expected heterozygosity was significantly correlated with the date of the oldest sample in a study. Studies using archival samples from close to the end of the industrial revolution often showed a much greater decline relative to those with time series starting in, or after, the 1950s (Figure 1). The number of generations between time points was also correlated with the total loss of allelic diversity, with studies spanning more generations typically showing a greater decline (Figure 2). Habitat fragmentation and agricultural transformation intensified dramatically after the industrial revolution (Ellis et al., 2010). Consequently, many population declines may have either begun in the 1800s or intensified at this point. Estimates of variation for earlier historical time points, particularly those close to the industrial revolution, are therefore likely predecline or at least more representative of historical levels. Estimates are likely mid-decline for historical time points in the 1950s, where rare alleles in particular may have already been lost. These effects are likely driving the trend of reduced genetic diversity loss with younger historical samples (Figure 1). Multi-time point comparisons of the same populations, ideally using ancient DNA to obtain estimates from before the 1700s, are now needed to confirm when declines in genetic variation began.

Postindustrial revolution human disturbances, such as land use change and habitat fragmentation, are not distributed equally across the globe and are particularly severe within western Europe and South-East Asia (Ellis et al., 2010). Furthermore, population declines since the 1970s are 30% higher in Neotropical regions relative to the global average (World Wildlife Fund, 2018). Regions of extreme change are likely suffering more severe variation loss, and it is essential to identify such locations. Unfortunately, the availability of temporal studies meeting our criteria was globally unequal, making a fine-scale investigation into geographical trends impossible. Over 62.6% (n = 62) of species in the data set were from North America and Europe. This regional bias is similar to, though more extreme than, that previously found for mitochondrial haplotype diversity estimates (Miraldo et al., 2016). This is because of the rarity of historical samples, as well as the cost and specialist knowledge needed to sequence them (Wandeler et al., 2007). Though regional comparisons were impossible, we were able to compare trends in variation loss between island and continental mainland populations. Loss of genetic variation was substantially, and significantly, greater on islands relative to mainland populations. Island species lost 27.6% and 30.9% of expected heterozygosity and allelic diversity (respectively), relative to only 2.1% and 4.6% in mainland species. This matches with our expectations because island species often consist of small and isolated populations and, as a result, are more vulnerable to extreme population declines (Purvis, Gittleman, Cowlishaw, & Mace, 2000). Furthermore, human impacts measured by population density, land transformation, transport infrastructure, invasive species and electrical power infrastructure are higher on islands (Purvis et al., 2000; Sanderson et al., 2002). Together, these human influences are likely fuelling higher rates of fragmentation, population declines and genetic drift that are ultimately leading to the higher loss in expected heterozygosity on islands. Island species have previously been shown to harbour less genetic variation on average than mainland species due to founder effects and small population size (Frankham, 1997; Stuessy, Takayama, López-Sepúlveda, & Crawford, 2014), but to our knowledge, a trend of elevated variation loss has not been previously shown. Islands are hot spots of endemism, and this severe genetic decline will disproportionally elevate the risk of global species loss (Kier et al., 2009), exacerbating the global biodiversity extinction crisis if left unmitigated.

Historical samples are known to have a higher rate of genotyping error and smaller sample sizes, both of which can decrease estimates of heterozygosity and allelic diversity (Wandeler et al., 2007). Within this study, we found no differences in probability of deviation from Hardy–Weinberg equilibrium proportions between time points, suggesting that genotyping errors were not substantially more common in historic time points. We also found that sample sizes were fairly large across both time points; thus, the trends observed here are not likely to be driven by consistently poor historical baseline estimates within data sets. A notable exception to these findings is an extreme outlier in Figure 1, where a gain in diversity was attributed to a poor historical baseline in the original study (Atlapetes pallidiceps, Hartmann, Schaefer, & Segelbacher, 2014). A second outlier is also present, but this was attributed by the original authors to immigration introducing new genetic variation during the study (Capreolus capreolus, Wang, Lang, & Schreiber, 2002). The remaining more minor variation increases are often seen in fish and are likely due to stocking efforts (where populations are supplemented with captive-reared individuals).

An estimated average loss in genetic variation of 5.4%–6.5% may seem surprisingly low. However, our requirement for modern samples (to compare diversity change through time) forced the exclusion of extinct species. Consequently, the species that have undergone the most severe population declines are not in this analysis. Conservative estimates suggest extinction rates are now 100 times the historical baseline rate (Ceballos et al., 2015). Hence, the true total loss of genetic variation is likely to be much greater than estimated, and a ~ 6% decline should be viewed as conservative. Importantly, our study highlights that it is essential to address the taxonomic and geographical gaps in temporal studies and confirm these trends. To ensure cross-compatibility, future studies should ensure that the classical metrics of genetic variation (expected and observed heterozygosity, allelic diversity) continue to be reported, alongside generation times and population size estimates.