No biodiversity impact.

Estimates of global biodiversity trends vary depending on their data and mathematical model. The most apocalyptic models gather extensive press coverage, even when based on controversial data (for example, ‘biological annihilation’5, which described trend estimates based largely on expert opinion; or ‘insect Armageddon’, which is based on data disputed by the original collectors6). However, even analyses of the best available data reach conflicting conclusions. An analysis of a global dataset of abundance time series of vertebrates estimated that, on average, vertebrate populations have declined by more than 50% since 1970 (Living Planet Index2 (LPI)); however, other global analyses found that the mean population size7,8 and species richness9,10 have remained stable over similar timeframes. Explanations for the discrepancies have been proposed8,11,12,13, but not resolved.

One crucial consideration is that summary indices may be easily misinterpreted. Calculating the geometric mean across populations is the most common and straightforward approach, but is strongly influenced by extremes. To illustrate, imagine an ecosystem in which one population declined by 99%. Even if a second population increased 50-fold or 393 populations increased by 1% (that is, a large net increase), a geometric mean would show a catastrophic 50% decline. Thus, a geometric mean decline of 50% could arise from substantial, widespread loss that is occurring across many populations (we term this the ‘catastrophic declines’ hypothesis) or from a few extremely declining populations (we term this the ‘clustered declines’ hypothesis). Both scenarios involve important conservation issues, but suggest vastly different underlying problems and require different mitigation strategies14, thus distinguishing between them is of real-world importance.

We derive a Bayesian hierarchical mixture (BHM) model to distinguish between the catastrophic and clustered declines hypotheses. The model statistically separates population trends into extreme declines, typical trends and extreme increases (Fig. 1), while accounting for time-series size, within-population fluctuations, number of populations and among-population variance. We test declines in abundance for more than 14,000 vertebrate populations (from the LPI)15. We chose LPI data because of its large scope, because the data and analytical details were publicly available, and because previous analyses of these data suggested widespread, global declines2.

[Figure omitted]

We first examined whether the previous estimate2 of a mean decline of more than 50% was sensitive to extreme populations: robust declines would support the catastrophic declines hypothesis, whereas high sensitivity to a few populations would support the clustered declines hypothesis (Fig. 1). We then applied our BHM model to assess the evidence for catastrophic or clustered declines globally and by region and taxonomy. Finally, we explore two additional conservation issues. First, we test whether declines occur disproportionately in larger animals (large animals tend to have lower reproductive rates), which might release small animals from predation16. Second, previous analyses often excluded time series with few data points10,12,17, but small time series make up most of the available data. We test the effects of their exclusion18.

Sensitivity of geometric mean to extreme populations

The geometric mean index that underlies the LPI analysis was highly sensitive to extreme populations. Excluding only the 2.4% most-strongly declining populations (354 out of 14,700 populations) reversed the estimate of global vertebrate trends from a loss of more than 50% to a slightly positive growth (Fig. 2). Similarly, excluding 2.4% of the most-strongly increasing populations strengthened the mean decline to 71%. High sensitivity suggests that extreme populations are disproportionately affecting global trend estimates, such that clusters of extreme population decline should be considered explicitly.

[Figure omitted]

Evidence for clustered declines

Among the 57 domain–realm–taxon systems of the LPI, 16 systems contained clusters of extreme decline and 8 contained clusters of extreme growth (of those, 3 systems are repeated, as they had both clusters of extreme decline and growth) (Fig. 3 and Supplementary Table 2). Together, clusters of extreme decline accounted for only 1% of populations across systems (2% of populations in the 16 systems in which they occurred). The mean population trend for extremely declining clusters across the 16 systems was θ2 = −3.94, or approximately 98% loss per year, and deviated substantially from the mean trend of the primary cluster in those systems. Clusters of extreme growth accounted for 0.4% of populations across systems (2.4% in the 8 systems in which they occurred), with θ2 = 3.51, that is, an explosive 33× growth per year (Fig. 3 and Supplementary Table 2).

[Figure omitted]

Extreme clusters showed some taxonomic and geographic patterns. The largest cluster of extreme declines was in Arctic marine mammals, accounting for 7.6% of populations in that system. However, mammalian systems generally had the fewest clusters of extreme decline (19% of 16 systems), followed by reptile–amphibian systems (21% of 14 systems), whereas bird and fish systems had more clusters of extreme declines (31% of 16 and 45% of 11 systems, respectively) (Fig. 3). Clusters of extreme decline occurred throughout the world, half of which occurred in marine realms, whereas extreme increases occurred more in temperate regions or terrestrial realms (Fig. 3).

Extreme population trends occurred predominantly in small time series. Excluding time series with fewer than 10 points not only removed all but two extreme clusters, but also removed 52% of the data (Supplementary Table 3). The higher frequency of extreme trends among small time series was also apparent in the raw data (Fig. 4). Thus the decision of whether to include small time series will have large effects on the resulting estimates of global trends.

[Figure omitted]

Body size was related to population trends. Larger species had three times more extreme declines than increases (15 compared with 5 clusters of extreme declines compared with extreme increases). Comparatively, smaller species had half as many (8) extremely declining and disproportionately more (7) extremely increasing clusters (Supplementary Table 4). Although size-specific models included fewer populations, especially for smaller species, the number of clusters was not uniformly lower (as might be expected given a reduction in power); therefore, the differential occurrence of extremely declining versus increasing clusters suggests that large animals are more vulnerable to extreme declines.

Evidence for catastrophic declines

In contrast to the extreme clusters, the primary clusters accounted for the vast majority (98.6%) of populations across the 57 LPI systems. The overall growth rate of primary clusters was close to zero: θ1 = −0.00035, corresponding to around 1.7% loss over 50 years, given a constant rate across populations and time (Fig. 5). In addition, in contrast to extreme clusters, primary cluster trends were robust to time-series size, as excluding series with fewer than 10 data points yielded a similar overall global trend (θ1 = 0.0043) (Extended Data Fig. 3).

[Figure omitted]

Although the global BHM model reveals considerably more nuance than a geometric mean index, analysing across systems still masked important patterns. When systems were analysed separately (Supplementary Table 2), primary population clusters were strongly declining (θ1 < −0.015) with high certainty (95% credible intervals not overlapping zero) in three systems, all of which occurred in the Indo-Pacific realm (freshwater mammals, freshwater birds and terrestrial birds) (Fig. 3). This suggests that this region has the highest risk of system-wide declines and should be a conservation priority. By contrast, the primary cluster was increasing with high certainty in seven systems, six of which were in temperate regions. In addition, seven additional systems had strongly declining primary population clusters but with less certainty (95% credible intervals overlapped zero), four of which were amphibian or reptile groups. Finally, 14 systems showed strong but low-certainty increases, with no obvious taxonomic nor geographic patterns (Fig. 3).

Each primary cluster also contained variation among populations. In the 10 systems with significant or non-significant mean declines where θ1 < −0.015, 87% of the individual populations showed strong declines (Fig. 5). These 10 systems accounted for around 20% of the total global vertebrate populations, but for around 61% of strong declines. The multimodality observed in Fig. 5 was an outcome of aggregating unimodal primary clusters across systems, and suggests that there are heterogeneous stressor levels among systems (that is, similar principles to those that cause extreme clusters within systems). The remaining approximately 11% of strongly declining populations were distributed across 47 out of 57 systems; it is unclear whether they represent a deviation from the natural dynamics that are expected to occur in any naturally variable system.

Primary cluster trends were related to body size, but not as predicted. In comparison to the overall patterns for larger animals, the same systems showed significant declines and increases, but two additional temperate systems showed significant increases (Extended Data Fig. 4 and Supplementary Table 4). Smaller species also appeared to decline more than larger species; there were 27 systems in which smaller species had more-negative growth rates than larger species, compared with 18 systems in which the reverse was true. However, analyses of the smaller species were based on substantially fewer populations, and trends were generally not significant (Supplementary Table 4), so patterns remain tentative.

Discussion

By re-analysing a comprehensive dataset of global wildlife population trends, we show that previously estimated global declines are driven by a few extremely declining populations. Removing only 2.4% of declining populations reversed the estimated global trends from more than 50% mean decline since 1970 to a slightly positive growth. Our BHM model revealed that clusters of extreme decline are widespread and occur disproportionately in larger species, and that a few clusters of extreme increase also exist and occur disproportionately in smaller species. This is consistent with previous arguments of ‘trophic downgrading’16.

Clusters of extreme declines were largely due to small time-series datasets. However, neither random sampling error nor ‘saw tooth’ population dynamics (in which ultimately stable populations experience sudden declines followed by gradual increases) can fully explain this association (see Supplementary Information for a full discussion). Additional explanations are needed. Extreme trends could reflect transient populations that naturally leave or enter a survey area19, which could represent natural dynamics. Alternatively, researchers may stop sampling after populations become (close to) extirpated, although the converse has also been suggested20. A third possibility is that some regions experience both lower sampling effort and greater declines, such that poorly sampled datasets correlate with factors linked to vulnerability, such as lower national wealth or conservation investment. Understanding why small time series contain so many extreme declines is particularly important given that studies that did not find widespread declines often excluded short time series7,10,12, potentially reconciling divergent findings among studies.

Once extreme clusters were statistically separated, no global trend remained across typical populations (that is, primary clusters; 98.6% of populations). However, aggregating systems into one global trend hid important variation. Three systems, all of which occurred in the Indo-Pacific realm, showed widespread vertebrate declines across typical populations. Moreover, among typical populations smaller species may be faring worse than larger ones. Although these results were tentative given lower sample sizes and high uncertainty, this trend is contrary to common conservation assumptions and so merits additional research.

Our results emphasize an important point: biodiversity trends within and across regions and taxa are highly disparate. This probably reflects differences in both susceptibility and exposure to anthropogenic environmental change21,22,23. Unravelling this variation is imperative to understand in which regions biodiversity is threatened the most24 and which conservation actions promote stability or recovery. A productive global conversation about conservation requires that both scientists and media pay more attention to variation and resist the temptation of simple summary indices.

Shifting the message from ubiquitous catastrophe to foci of concern, also touches on human psychology. Continual negative and guilt-ridden messaging can cause despair, denial and inaction25,26. If everything is declining everywhere, despite the expansion of conservation measures in recent decades, it would be easy to lose hope. Our results identify not only regions that need urgent action to ameliorate widespread biodiversity declines, but also many systems that appear to be generally stable or improving, and thus provide a reason to hope that our actions can make a difference.