Is OBIS Sufficient?

First, any attempt to describe modern biodiversity, including the OBIS database, is plagued by similar sampling biases, as in the fossil record (Hughes et al. Reference Hughes, Orr, Ma, Costello, Waller, Provoost, Yang, Zhu and Qiao2021). Inevitably, some regions will be undersampled (e.g., the Southern Ocean; Bonnet-Lebrun et al. Reference Bonnet-Lebrun, Sweetlove, Griffiths, Sumner, Provoost, Raymond, Ropert-Coudert and Van de Putte2023), some taxa will be underrepresented (e.g., bryozoans; Kopperud et al. Reference Kopperud, Lidgard and Liow2022), and some environments will be overrepresented (e.g., shallow coastal habitats; Klein et al. Reference Klein, Appeltans, Provoost, Saeedi, Benson, Bajona, Peralta and Bristol2019; Shaw et al. Reference Shaw, Briggs and Hull2020). Klein et al. (Reference Klein, Appeltans, Provoost, Saeedi, Benson, Bajona, Peralta and Bristol2019) note that half of described marine species are not found in OBIS (see also Webb and Vanhoorne Reference Webb and Vanhoorne2020), and those that are tend to be common and well known. However, a closer look at the OBIS dataset in comparison with PBDB data shows that our results are unlikely to arise from spatial or other biases within OBIS. For example, while more records come from Europe and North America in the OBIS, mirroring the socioeconomic and access issues seen in the PBDB (Fig. 5A), the bias is much less pervasive or impactful. Even where raw sampling numbers are relatively lower, biodiversity hotspots in places like the Indo-Pacific, Gulf of Mexico, and the African south coast still stand out within our subset of the OBIS dataset (Fig. 5B). By comparing data from the most recent PBDB time bin (the late Miocene–Quaternary) with the OBIS dataset, it becomes clear that OBIS is better sampled across latitude (Fig. 6) and taxonomy (Fig. 7). Southern Hemisphere sampling in the PBDB is comparatively incomplete, and latitudinal sampling deficiencies are only worse in deep time. The most recent PBDB time bin is also overwhelmingly dominated by mollusks, especially bivalves (69% and 41%, respectively), while the OBIS contains a taxonomically broader suite of organisms (only 37% mollusk and 18% bivalve; see also Webb and Vanhoorne Reference Webb and Vanhoorne2020). The median OBIS cell contains five of the six phyla examined in this study, while the median recent PBDB cell only contains two. Finally, comparing cells that are occupied by occurrences in both datasets shows that well-represented cells in the PBDB are typically well sampled in the OBIS, but the reverse is not true (Fig. 8). In sum, there are many more cells sampled in the OBIS and those cells are consistently better sampled than those of the PBDB. Even if sampling of the OBIS is not complete—biological sampling can never be considered 100% comprehensive—sampling within the OBIS is certainly comprehensive enough for comparisons to the PBDB and for global spatial diversity statistics.

Figure 5.

Ocean Biodiversity Information System (OBIS) occurrences and richness in our subset of the data. A, OBIS occurrences within equal-area cells. B, OBIS generic richness within equal-area cells. Patterns in A are largely driven by spatial sampling biases (e.g., a greater number of scientific institutions in the Northern Hemisphere), but important marine invertebrate diversity hotspots can be seen in B, implying that imperfect spatial sampling has not greatly distorted spatial diversity patterns in the OBIS.

Figure 6.

Latitudinal sampling in the Ocean Biodiversity Information System (OBIS) dataset (blue) versus the most recent (mid-Miocene to Recent) time bin of the Paleobiology Database (PBDB) (red). OBIS sampling is more complete in the Southern Hemisphere (A), but both exhibit the latitudinal diversity gradient (B).

Figure 7.

The number of phyla per sampled cell in the Ocean Biodiversity Information System (OBIS) dataset (A) versus the most recent time bin of the Paleobiology Database (PBDB) (B).

Figure 8.

The numbers of families present in cells occupied in both the Ocean Biodiversity Information System (OBIS) and most recent Paleobiology Database (PBDB) data. All diverse PBDB cells are also diverse in OBIS, but the reverse is not the case.

Second, our results cannot be dismissed as noise. If spatial sampling factors are important for driving the similarity between SQS diversity and resampled modern diversity, then removing the spatial component of the dataset through randomization should eliminate the correlation, and this is exactly what happens. When the taxonomic column of the OBIS dataset is randomized, effectively removing all relationships between space and the distribution of organisms on the Earth, the model fails to accurately predict bin-to-bin change in the SQS value of the fossil record (adj. R ² = 0.023, p = 0.156; Table 1). Furthermore, all versions of the model, including the spatially randomized one, predict bin-to-bin change in SQS value better than does the number of fossil occurrences through time. If SQS reduced the influence of spatial sampling bias as much as it does numerical sampling bias, we would expect the correlation between SQS and the number of occurrences through time to be similar to the correlation between SQS and resampled modern diversity, and this is not the case (Table 1).

Finally, our model is methodologically complex but theoretically simple. From a theoretical standpoint, the model can be conceived as sampling the modern landscape of biodiversity with the same or similar spatial fidelity as that in the fossil record. If the spatial distribution of fossil sampling were sufficient, there should be no appreciable change in predicted richness, as the same data are being sampled each time. There is no particularly good reason to believe that marine biodiversity in deep time was spatially arrayed in a similar manner to modern biodiversity—in fact there are many reasons to believe this not to be the case. While some manifestation of the LDG seems to be more or less persistent, the shape of that relationship has varied significantly over time, from unimodal to bimodal to flat (Crame Reference Crame2001; Alroy et al. Reference Alroy, Aberhan, Bottjer, Foote, Fursich, Harries, Hendy, Holland, Ivany, Kiessling, Kosnik, Marshall, McGowan, Miller, Olszewski, Patzkowsky, Peters, Villier, Wagner, Bonuso, Borkow, Brenneis, Clapham, Fall, Ferguson, Hanson, Krug, Layou, Leckey, Nürnberg, Powers, Sessa, Simpson, Tomasovych and Visaggi2008; Mannion Reference Mannion2020), and the latitudinal and longitudinal distribution of taxa have clearly changed with the changing configuration of continents and ocean basins (Valentine Reference Valentine1973). Nevertheless, the modern landscape can suffice as a null hypothesis for our spatial sampling model, and if the spatial distribution of organisms on the Earth has changed dramatically through deep time, it should be relatively easy to reject our null hypothesis. Again this is not the case. Our results suggest that, when estimating trends or changes in global-scale diversity, even SQS-standardized fossil diversity remains substantially impacted by spatial sampling factors.

Insights from Model Residuals

If the resampled OBIS model can be considered a null model of expected SQS fossil diversity, a closer look at time bins when the raw model and subsequent adjustments perform poorly may be informative.

First, a fundamental question arising from paleobiodiversity studies is whether the dramatic apparent increase in biodiversity into the Cenozoic and toward the present is real or an artifact (Alroy et al. Reference Alroy, Marshall, Bambach, Bezusko, Foote, Fürsich, Hansen, Holland, Ivany, Jablonski, Jacobs, Jones, Kosnik, Lidgard, Low, Miller, Novack-Gottshall, Olszewski, Patzkowsky, Raup, Roy, Sepkoski, Sommers, Wagner and Webber2001). Consistent with Alroy (Reference Alroy2010), our analysis of SQS richness shows an approximately twofold increase in the average richness of Cenozoic time bins compared with those of the Paleo-Mesozoic, even when unconsolidated sediments are excluded from the analysis. To test the hypothesis of a step change in marine biodiversity following the Cretaceous–Paleogene mass extinction, we compare models that include no time-dependent variable, a “Cenozoic or not” factor, and time as a continuous variable (i.e., bin age in millions of years). Results (Table 1) clearly favor the “Cenozoic or not” model as the model with the most explanatory power for both SIB and SQS fossil diversity estimates, indicating the likelihood of a real stepwise increase in marine biodiversity starting near the Cretaceous/Paleogene boundary. However, geographic sampling proxies like the resampled OBIS diversity estimate and the summed MST length of PBDB-sampled cells accurately explain the observed increase in SQS and SIB richness during the Cenozoic, indicating little support for a continuous increase in richness among marine invertebrates over the last 66 Myr. Close and colleagues likewise found a stepwise increase in the biodiversity of marine animals (Close et al. Reference Close, Benson, Saupe, Clapham and Butler2020b) and terrestrial vertebrates (Close et al. Reference Close, Benson, Alroy, Carrano, Cleary, Dunne, Mannion, Uhen and Butler2020a) in the early Cenozoic that they could not attribute to a spatial sampling artifact. Whether this observed stepwise increase is real or the result of additional bias (e.g., more splitting of Cenozoic genera; Hendy Reference Hendy, Allison and Bottjer2010) remains an open question and deserves closer investigation.

After correcting for the step up into the Cenozoic, residuals correlate weakly but significantly with two variables drawn from the PBDB dataset (Fig. 4A). First, the percentage of fossil occurrences from relatively shallow environments (roughly, shallower than benthic assemblage zone 4, exclusive) correlates positively with the residuals (adj. R ² = 0.193, p = 0.001). Residual correlation with an environmental variable is unsurprising, because the OBIS model does not take any paleoenvironmental information into account, and biological diversity certainly varies across environments. However, to the degree that times with proportionately more fossils from shallow environments also reflect times of greater continental flooding and expansion of shallow seas, this correlation could reflect a real signal of biotic response to expansion and contraction of habitat area (Peters and Husson Reference Peters and Husson2017). The extent of shallow-marine sediment is also seen to correlate positively with spatially subsampled diversity by Close et al. (Reference Close, Benson, Saupe, Clapham and Butler2020b), suggesting species–area effects (Rosenzweig Reference Rosenzweig1995) not predicted by sampling alone. The significant middle Cambrian overestimation of PBDB diversity, for example, disappears in the four-variable OBIS model that incorporates percent shallow occurrences (Figs. 4A, 9), perhaps reflecting a real increase in diversity during the expansion of shallow-water habitats associated with continental flooding during the Sauk sequence. Given the nature of Paleo-Mesozoic fossil data, a relative increase in the sampling of inferred deep occurrences (or a reduction in the proportion of shallow occurrences) might additionally reflect times of increased ocean anoxia. Deep environments are often inferred from the presence of black shales and dark-colored micrites, which are typically depauperate and more directly indicative of anoxia than they are of deep water (Smith et al. Reference Smith, Schieber and Wilson2019). Anoxia might therefore contribute to the forces producing relatively lower than expected fossil SQS diversity (e.g., Peters Reference Peters2007) in intervals with proportionally less shallow sampling.

Several other environmental variables were tested, including the percentage of tropical and reefal occurrences, both of which are often associated with higher origination and standing diversity in the literature (Aberhan and Kiessling Reference Aberhan, Kiessling and Talent2012), but none explained any significant proportion of variation in the residuals. Rather than showing that these variables are not important for standing diversity, this indicates that modern and paleoenvironmental trends diverge in this respect—in the modern dataset, the tropics and reefal locations have not changed from where they were in the past, but the location of the shallow ocean certainly has. In other words, the modern dataset already has a spatial concentration of diversity at the tropics (and at reefs, which are mostly tropical), so no additional information from the fossil dataset is necessary to explain these spatial patterns. The only other variable to significantly correlate with model residuals was a simple measure of the absolute latitudinal range of fossil occurrences in each time bin (adj. R ² = 0.098, p = 0.017), an unsurprisingly weak correlation, given that the model already attempts to account for latitudinal variation in sampling. Its incorporation as an additional explanatory variable implies that the original OBIS model may have treated latitude with too little granularity.

The OBIS Cenozoic-corrected and four-variable models tend to overestimate richness in time bins at or immediately after the major mass extinctions (Fig. 4A), but model residuals for most of these time bins (Fig. 9) do not stand out, reinforcing that rebound from extinctions largely happens within the succeeding time bin at this scale, as suggested by other studies of regional and global biodiversity (Alroy et al. Reference Alroy, Aberhan, Bottjer, Foote, Fursich, Harries, Hendy, Holland, Ivany, Kiessling, Kosnik, Marshall, McGowan, Miller, Olszewski, Patzkowsky, Peters, Villier, Wagner, Bonuso, Borkow, Brenneis, Clapham, Fall, Ferguson, Hanson, Krug, Layou, Leckey, Nürnberg, Powers, Sessa, Simpson, Tomasovych and Visaggi2008; Close et al. Reference Close, Benson, Saupe, Clapham and Butler2020b). The Late Devonian biodiversity crisis is an exception, perhaps owing to its duration over a longer interval of time. PBDB richness is much lower compared with the OBIS model predictions at the Jurassic/Cretaceous (J/K) boundary (Fig. 4A). The latter is especially interesting given that the J/K boundary is generally poorly understood and may be an interval especially affected by sampling biases (Tennant et al. Reference Tennant, Mannion and Upchurch2016). On the other hand, the interval from the end-Silurian to the Middle Devonian stands out as a period with uncharacteristically high residuals—that is, fossil diversity is higher than expected by the OBIS model. Close et al. (Reference Close, Benson, Saupe, Clapham and Butler2020b) also find SQS diversity to be elevated at this time relative to their spatially standardized analysis. One might infer this to reflect an expansion in marine biodiversity (Klug et al. Reference Klug, Kröger, Kiessling, Mullins, Servais, Frýda, Korn and Turner2010; Ruban Reference Ruban2013), yet the interval encompassing the well-known Ordovician radiation does not stand out as a time of elevated richness compared with model estimates. The Early–Middle Devonian, however, is a particularly spatially well-sampled interval, specifically with an atypically high latitudinal range of sampling, even though the number of occurrences collected from that interval is unremarkable.

Figure 9.

Residuals between the resampled four-variable Ocean Biodiversity Information System (OBIS) model and fossil shareholder quorum subsampling (SQS) diversity. Dashed lines show the ages of the “big five” mass extinctions. CM, Cambrian; O, Ordovician; S, Silurian; D, Devonian; C, Carboniferous; P, Permian; Tr, Triassic; J, Jurassic; K, Cretaceous; Pg, Paleogene; Ng, Neogene.

Moving Forward

What can be done to remedy or address the effects of spatial bias in the fossil record? One line of thought is to simply increase sampling efforts in previously undersampled locations, including across the Global South. Given that the density of sampling is much higher in North America and Europe than in the rest of the world (Ye and Peters Reference Ye and Peters2023), much could be learned from this process, and many gaps in knowledge might be filled. However, it is likely that biases originating from the spatially nonrandom distribution of exposed sedimentary rock of appropriate age (rather than from sociopolitical phenomena) would still influence interpretations of the fossil record, even after substantial sampling efforts, because the samples necessary to fill the gaps simply do not exist. Another approach is to focus on numerically standardized paleobiological diversity within multiple areas of comparable spatial extent drawn from each time interval (Close et al. Reference Close, Benson, Alroy, Carrano, Cleary, Dunne, Mannion, Uhen and Butler2020a,Reference Close, Benson, Saupe, Clapham and Butlerb), allowing for direct comparisons across geological time at spatially consistent scales. While both increased sampling and improved understanding of diversity variance within and across equal spatial regions are important for addressing the spatial structure inherent in the fossil record, spatially informed quantitative and statistical methods are critical as well. Paleobiologists will benefit from new focus on the development of novel spatial resampling techniques (e.g., recent work by Antell et al. [Reference Antell, Benson and Saupe2024]) and from interdisciplinary studies with ecologists and neontologists experienced in dealing with spatial sampling questions.

Additionally, future efforts to address spatial bias in the sampling-standardized record of fossil diversity should focus on differentiating the real biogeographic effects imposed by plate tectonics, continental flooding, and global climate change from the artificial variation introduced by the distribution of sampling. The spatial distribution of marine taxa today is not likely to approximate that in, for example, the Middle Devonian, because in the latter, the Northern Hemisphere was mostly devoid of land or shallow water (Scotese Reference Scotese2016) and the continents were less dispersed than today (Zaffos et al. Reference Zaffos, Finnegan and Peters2017). Likewise for the Cretaceous hothouse, with much warmer temperatures and no ice at the poles. As well, given the preponderance of data from North America and Europe and their northward shift out of the tropics over the last ~200 Myr, if anything like the modern LDG is typical of the Phanerozoic, numerically standardized estimates alone are still likely to underestimate mid–late Mesozoic diversity in comparison to the Paleozoic. To what degree are recovered patterns in paleobiodiversity in space and time real and due to such factors, as opposed to driven purely by spatial (and numerical) sampling limitations?

In conclusion, we show that a subset of the OBIS dataset of modern global biodiversity with taxonomic membership comparable to the skeletonized marine invertebrate fossil record, when sampled spatially as dictated by the fossil record, produces predicted richness values and first differences that reasonably correlate with fossil richness estimates, even after state-of-the-art numerical subsampling procedures have been applied. As also deduced by Close et al. (Reference Close, Benson, Saupe, Clapham and Butler2020b) using entirely different methods, these results strongly caution the overinterpretation of biodiversity trends inferred from the fossil record and are consistent with the argument that global marine biodiversity has been largely static over the Phanerozoic, with apparent variation resulting almost entirely from changes in the spatial distribution of sampling. Changes in spatial sampling parameters, like the number and distribution of geographic cells sampled, their summed MST distances, and their latitudinal range, are at least as important for predicting fossil richness values as factors associated with numerical sampling intensity. This analysis highlights the importance of the spatial distribution of sampling in analyses of large-scale fossil biodiversity and echo the call from Antell et al. (Reference Antell, Benson and Saupe2024) for the development of new methodologies that explicitly account for spatial biases in the fossil record.