Bias, incompleteness and the ‘known unknowns’ in the Holocene faunal record

Abstract

Long-term faunal data are needed to track biodiversity change and extinction over wide spatio-temporal scales. The Holocene record is a particularly rich and well-resolved resource for this purpose but nonetheless represents a biased subset of the original faunal composition, both at the site-level assemblage and when data are pooled for wider-scale analysis. We investigated patterns and potential sources of taxonomic, spatial and temporal bias in two Holocene datasets of mammalian occurrence and abundance, one at the global species level and one at the continental population-level. Larger-bodied species are disproportionately abundant in the Holocene fossil record, but this varies according to trophic level, probably owing to past patterns of human subsistence and exploitation. Despite the uneven spatial distribution of mammalian occurrence records, we found no specific source of sampling bias, suggesting that this error type can be avoided by intensive data collection protocols. Faunal assemblages are more abundant and precisely dated nearer to the present as a consequence of taphonomy, past human demography and dating methods. Our study represents one of the first attempts to quantify incompleteness and bias in the Holocene mammal record, and failing to critically assess the quality of long-term faunal datasets has major implications for understanding species decline and extinction risk.

This article is part of a discussion meeting issue ‘The past is a foreign country: how much can the fossil record actually inform conservation?’

1. Introduction

In order to respond to the ongoing biodiversity extinction crisis [1,2], it is imperative to understand patterns and drivers of species and population losses across taxonomy, time and space. For example, the disappearance of large-bodied mammals across the late Quaternary is known to have reshaped current-day patterns of mammalian diversity [3] and extinction risk [4,5]. Additional vertebrate species extinctions, range contractions and population losses have been documented in the fossil and historical record across continental and insular regions throughout the late Quaternary up to the present day as a result of climatic changes and anthropogenic impacts [6,7]. Long-term archival data are therefore needed to provide context for measuring changes in biodiversity over wide temporal and spatial scales.

The faunal record of the Holocene Epoch, from approximately 11 700 yr BP to present, is particularly well suited for this purpose. As the most recent geological epoch, it has a rich and well-resolved record documenting vertebrate occurrences over wide spatio-temporal scales [8] and includes extinct and extant species, permitting analysis of factors that influence both extinction and survival. Furthermore, the relative stability of Holocene climate in comparison with the preceding Late Pleistocene means that anthropogenic drivers of biodiversity loss are generally more straightforward to identify and interpret [7], in comparison to the ongoing debate around human versus climatic drivers of late Quaternary megafaunal extinctions [9,10]. As the time period during which major human demographic, subsistence and technological transitions occurred [11], the Holocene is well suited to track faunal responses to a wide variety of human pressures, from low-density hunter–gatherers to settled farming communities and more recent industrial and urbanized societies.

However, faunal records in long-term environmental archives only ever represent a biased subsample of an ecosystem or community [12] owing to multiple processes that influence patterns of taxonomic composition and abundance during the deposition and excavation of faunal assemblages. First-order (pre-excavation) taphonomic modifications of fossil assemblages can include biotic disturbances from other organisms (e.g. trampling, scavenging and burrowing animals, or plant root growth), and abiotic disturbances such as wind, rain, floods and heat [13]. Intrinsic qualities of the faunal deposition can also affect its survival; for example, the largest, hardest, densest bones and teeth generally survive best in response to weathering, burial and decomposition [14,15], potentially biasing the relative representation of different-sized species or individuals (e.g. ontogenetic stages) in the faunal record. Other ecological traits are less well understood in terms of their role in controlling relative patterns of preservation in bone deposits, but factors such as trophic guild [16] and habitat [17] can also affect the frequency with which different species are represented in faunal archives.

It is also important to distinguish between faunal assemblages from natural fossil (palaeontological) sites and those from archaeological sites. In general, fossil sites have been deposited by ‘natural’ (non-anthropogenic) means, for example, from animals dying, decomposing and being buried and preserved by natural sedimentological processes or through the accumulation of prey species by non-human predators such as raptors. By contrast, faunal remains from archaeological sites (hereafter ‘zooarchaeological’ assemblages) are by definition found in human contexts, such as killing sites, refuse pits or deliberate burials, which reflect prehistoric or historical processes of animal exploitation such as hunting, herding or domestication. A human selection filter reflecting subsistence, social and cultural preferences (both positive and negative) towards certain species has therefore further influenced the composition of zooarchaeological assemblages in addition to non-anthropogenic first-order taphonomic modifications, and it may not be easy to disentangle these biases from an understanding of which species were formerly present and/or abundant in the local environment [15]. As human populations rapidly expanded during the Holocene from the Neolithic agricultural revolution onwards, zooarchaeological assemblages are by far the most abundant source of faunal records for this time period [8,18].

Most second-order (post-excavation) changes arise from sampling strategy and can therefore be partly controlled for by standardized excavation methods on a site-by-site basis, although the extent to which standardized methods are used is not necessarily reported [19]. Indeed, there has been a notable recorded bias in historical excavations, which often focused primarily on the discovery and description of larger fossil specimens and associated skeletal elements, meaning that large-bodied species have been relatively well documented, whereas many smaller-bodied species remain poorly known from environmental archives [20]. Furthermore, when records are pooled into larger datasets, uneven patterns of site distribution across landscapes can be problematic for spatial analysis. While the discovery and location of palaeontological and archaeological sites are always partly opportunistic and random, the spatial bias in the spread of sites can further distort our understanding of underlying natural faunal distributions. For example, sites can be actively searched for in areas of cultural interest, or where archaeological or fossil material is known to survive well. Countries also differ greatly both physically (e.g. topography, underlying geology, extent of urban versus rural areas) and in their political and economic histories, all of which can affect the amount of excavation, i.e. sampling, that has been undertaken.

The variety of dating methods used for zooarchaeological analysis also means that the precision and accuracy of dating may vary across time. Direct dating methods such as radiocarbon dating have been widely used in an effort to establish reliable megafaunal extinction chronologies and modern human arrival dates during important periods of late Quaternary environmental change, in order to disentangle potential extinction drivers [21,22]. However, for the more climatically stable and human-dominated Holocene, research has instead tended to focus on how faunal assemblages reflect past human society and culture [23] rather than on composition and dynamics of the faunal communities themselves. As a consequence, there is often little incentive to date individual bones, and zooarchaeological assemblages are often simply associated with the age of their associated archaeological site. Reported dating of faunal material can thus vary from broad temporal categories such as ‘Late Mesolithic’, a period potentially spanning a couple of thousand years, to a specific cultural period defined to a couple of hundred years. The extent of this disparity in dating methods, and any associated bias that it may place on the precision and resolution of available faunal records over time, should therefore also be carefully considered where records have been combined into large datasets.

In summary, it cannot be assumed that even the recent faunal record faithfully reflects source communities either at the site- or landscape-level, and the potential extent to which pre- and post-excavation biases may actively distort the interpretation of underlying patterns of species distributions, range shifts and extinctions should be critically assessed. To address this substantial but often overlooked concern, we therefore quantified taxonomic, spatial and temporal incompleteness/unevenness in two Holocene mammal datasets and investigated how potential sources of bias influence our understanding of: (i) species-level extinction at the global scale; (ii) population-level change for individual species at the regional scale. These complementary analyses into the quality of the recent faunal record provide important implications for the use of Holocene and older datasets in informing ecological baselines, reconstructing extinction processes and assessing extinction risk.

2. Material and methods

(a). Global faunal record

We compiled a global dataset of 255 mammal species that are known to have become globally extinct during the postglacial period (from the Holocene to the recent historical period), together with their recorded country-level geographical occurrences, description dates (or date of first publication if the taxon has remained formally undescribed beyond initial identification as a new species) and estimated body mass [7] (see the electronic supplementary material, table S1). This comprehensive dataset includes both species that were originally described from extant populations that have subsequently become extinct (‘modern’, e.g. thylacine Thylacinus cynocephalus), and species that have only been recorded from the Holocene fossil and/or zooarchaeological records and never observed alive in recent times by scientists (‘fossil’, e.g. woolly mammoth Mammuthus primigenius). Body mass data for a small number of extinct mammals were taken from the PanTHERIA database [24], while data for remaining species were taken from previously published body mass estimates for Holocene extinct mammals derived from regression equations based on extant congeners or skeletal measurement parameters [5].

In order to investigate whether pre- and post-excavation modifications to fossil assemblages are potentially biasing the taxonomic composition of the Holocene faunal record, we first investigated whether there was a relationship between species body mass and description date in the overall dataset. We then investigated this relationship in a more nuanced way, by assessing whether the relationship between body mass and description date was modified by whether extinct species were modern or fossil, and whether species inhabited continental or insular regions or both. We conducted analyses using phylogenetic generalized least squares (PGLS), which fits a linear model controlling for the non-independence between species resulting from the phylogenetic structure in the data. Although species description date is not itself a heritable trait, it may correlate with traits such as body mass. Following [25,26], we arbitrarily selected the first phylogenetic tree from 1000 available trees in [27]. We used the pgls function in the R package ‘caper’ [28], with taxonomy standardized between datasets where possible. In total, 89 extinct mammal species in our dataset were absent from the phylogeny (electronic supplementary material, table S2), so this analysis was undertaken with a reduced dataset of 166 species.

(b). Continental zooarchaeological record

We used an extensive dataset of 18 588 zooarchaeological records for 23 large mammal species (greater than 2 kg) in Europe spanning the Holocene, which has previously been used to reconstruct long-term processes of mammalian range decline and extinction [29] (electronic supplementary material, table S3). These data also represent a subset of a larger Holocene vertebrate database that has been widely used to investigate regional faunal turnover, refugia and extinction across various vertebrate taxa [8].

(i). Taxonomic bias

We converted the number of zooarchaeological records for each species into the proportion of sites occupied across its extent of occurrence (defined as a minimum convex polygon enclosing all sites). This was to account for: (i) species' differing range sizes (which might skew the overall abundance of each species), and (ii) areas of Europe with low densities of zooarchaeological sites, which might artificially lower estimated abundance for species’ ranges that fell within these areas (electronic supplementary material, table S3). As a proxy measure for zooarchaeological abundance, ‘proportion of occupied sites' will hereafter be referred to as ‘abundance’. We excluded two species from analyses: (i) wild horse (Equus ferus), because accurate records for this species only cover the first half of the Holocene owing to identification problems with domestic horses in the later Holocene [30]; and (ii) fallow deer (Dama dama), because its abundance and distribution were heavily affected by human-mediated introduction beyond its natural European range [31].

We compared observed zooarchaeological abundance with ecological variables that could potentially influence whether a species might be recorded in the zooarchaeological record. Owing to the small size of the response variable (n = 21) we selected only two predictor variables, body mass and trophic level (herbivore/carnivore), on the basis that body mass represents a useful proxy for a range of other life-history variables (e.g. reproductive output) [32], and that humans are likely to interact differently with herbivores and carnivores, e.g. as subsistence versus competitors for prey. We predominantly used life-history data from the PanTHERIA database [24]; data on the extinct aurochs (Bos primigenius) were taken from a global database of late Quaternary mammals [33]. Body mass was logarithmically transformed for analysis. We analysed determinants of abundance using generalized linear models (GLMs) with quasi-binomial errors to account for the response variable being a proportion and displaying overdispersion. We calculated ‘quasi-Akaike information criterions’ (QAICs) using the R package AICcmodavg [34] for model comparison, with the model with the lowest QAIC value interpreted as having the best explanatory power for explaining the influence of body mass and trophic level on zooarchaeological abundance. We also re-ran the analysis including species as a random effect.

In order to assess whether any observed relationship between zooarchaeological abundance and body mass or trophic level could either be attributed to taphonomic bias or instead reflected natural patterns, we compared our analysis of zooarchaeological abundance with regressions of population density and body mass from an ecological dataset of global mammal populations [35]. We selected species with a mean body mass between 0.5 and 700 kg and within the orders Artiodactyla, Carnivora and Perissodactyla, representing the body size distribution and orders that were also present in the Holocene dataset in order to ensure the two datasets were comparable. This mass range also covered an order of magnitude, considered sufficient to overcome any potential biasing effects on body size–abundance relationships [36]. We compared our Holocene dataset with two different ecological datasets: (i) one from North America only (as the most faunally intact and environmentally analogous present-day ecosystem to compare with the Holocene of northern Eurasia); and (ii) all continents pooled together. We also conducted separate regressions for all three datasets by trophic level (herbivore/carnivore). As an additional method of comparison, we calculated confidence intervals for all model slopes (all data, herbivore only and carnivore only) for the three datasets, with significant differences in abundance–body mass distributions interpreted if 95% confidence intervals did not overlap. All analyses were conducted using the glm and glmer packages ‘MASS’ [37] and ‘lme4’ [38] in R 3.5.0 [39].

(ii). Spatio-temporal bias

The spatial spread of Holocene faunal data across Europe (electronic supplementary material, figure S1) indicated that numbers of zooarchaeological records differed substantially between countries, with higher representation of records in central and western Europe. We investigated alternative possible explanations for this pattern by calculating the total number of zooarchaeological records (interpreted as a proxy for research output) for each country. To determine whether number of zooarchaeological records per country reflected high output from a few sites rather than wider research effort across multiple sites, we tested the relationship between numbers of zooarchaeological records and zooarchaeological sites per country using Pearson's correlation coefficient. Numbers of records and sites were extremely highly positively correlated (r₃₈ = 0.91, p < 0.05), so we retained number of zooarchaeological records as the chosen metric for further analysis.

One possible explanation for variation in the number of zooarchaeological records across Europe is variation in country wealth and resources available for zooarchaeological research. Gross domestic product (GDP), a country's total economic activity based on the market value of all goods and services, is a widely used appropriate proxy measurement of a country's wealth [40]. However, it has been demonstrated that a country's land area, GDP and population size are all positively correlated [40], meaning that raw data on numbers of zooarchaeological records and GDP might not be appropriate for analysis as both values might show collinearity with land area. Regression analysis confirmed that country land area and GDP across Europe were positively correlated (r₃₈ = 0.35, p < 0.05). Therefore: (i) we corrected a total number of zooarchaeological records by land area into a measure of density of records for each country; and (ii) we used GDP per capita rather than GDP, with data obtained from the World Bank website [41]. We analysed the relationship between the density of zooarchaeological records and GDP per capita for each country using a GLM with quasipoisson errors to account for overdispersion. GDP per capita was logarithmically transformed for analysis.

We also investigated whether the spatially uneven spread of zooarchaeological data could instead be related to variable topography across Europe, which might, for example, affect landscape accessibility for research. We calculated the average elevation for each country using a high-resolution 30-arc seconds (approx. 1 km) elevation map from the WorldClim database [42]. We analysed the relationship between the number of zooarchaeological records and average elevation for each country using a negative binomial GLM to account for count data with considerable overdispersion. Average elevation was logarithmically transformed for analysis. We also compared the elevation profile for the zooarchaeological dataset to that of Europe to assess whether there were differences in the overall range of values and to check that the average elevation was not being biased by widely outlying values. We calculated the elevation of each zooarchaeological data point at a 30-arc raster square resolution and then randomly sampled the same number of points from an ArcMap layer of European elevation and plotted both datasets as histograms.

In order to examine any changes or patterns in dating precision over time, we plotted the lower and upper date range by midpoint for each zooarchaeological record. We also plotted all dated Holocene records against a theoretical linear increase in the number of zooarchaeological records, in order to identify changes in the rate of accumulation of records through time.

3. Results

(a). Global faunal record

The number of extinct species known from Holocene and historical contexts has increased over time since the mid-eighteenth century with peaks of new species descriptions in the early twentieth century and again close to the present (figure 1). This pattern has primarily been driven by a description of ‘fossil’ species (n = 199), with description of ‘modern’ (historically extant) species (n = 56) largely concluded by the 1950s. Nearly four times as many recently extinct mammal species have been described from islands (n = 196) compared to continental regions (n = 55), with four species having past geographical ranges that included both continents and islands. We found a significant negative relationship between description date and body mass across the ‘fossil’ data subset, with larger-bodied species generally described earlier than smaller-bodied species (est = −0.0030, s.e. = 0.0013, t₁₀₉ = −2.000, p < 0.05; R² = 0.048, λ = 1.00), but not for the ‘modern’ data subset (est = −0.00025, s.e. = 0.00057, t₅₃ = −0.43, p > 0.05; R² = 0.0035, λ = 1.00) (figure 2). For the combined fossil and modern dataset, we found a significant negative relationship between description date and body mass on continental regions, with larger-bodied species again generally described earlier (est = −0.0031, s.e. = 0.0014, t₄₁ = −2.141, p ≤ 0.05; R² = 0.10, λ = 1.00) and a non-significant relationship on islands (est = −0.00021, s.e. = 0.00050, t₁₀₃ = −0.410, p ≥ 0.05; R² = 0.00163, λ = 0.973) (figure 2). Re-analysis following removal of major outliers did not change the results (electronic supplementary material, table S5).

Figure 1.

Cumulative and decadal descriptions of extinct Holocene mammal species, AD 1750–AD 2012.

Figure 2.

The relationship between log body mass of extinct Holocene mammal species and description date, modified by whether they were described from modern or fossil specimens (top row) or whether they originated from continental or insular regions (bottom row). Slopes taken from linear regressions; see main text for regression information.

(b). Continental zooarchaeological record

The number of zooarchaeological records varied hugely between species in our dataset, from nearly 4000 records for red deer (Cervus elaphus) to fewer than 20 records for wolverine (Gulo gulo) (electronic supplementary material, table S3). We found a positive but non-significant relationship between zooarchaeological abundance and body mass (0.21 ± 0.11, p = 0.07) and a significant relationship between abundance and trophic level (1.45 ± 0.39, p < 0.01), with herbivores having a higher number of records than carnivores. Models with trophic level as the single explanatory variable had the highest explanatory power, both in our original models and in models that included species as a random factor (figure 3a; electronic supplementary material, table S4).

Figure 3.

Comparison of the abundance-body mass relationship between zooarchaeological and ecological datasets. (a) Relationship between the proportion of occupied sites in which a species occurs in the zooarchaeological record and log body mass. (b,c) Relationship between population density and body mass in the orders Artiodactyla, Carnivora and Perissodactyla, within the body size range 0.5–700 kg from the ecological dataset published by Damuth [35]; data for North America (b) and all continents pooled together (c). Population density and body mass are log transformed. Closed circles represent herbivores, open circles represent carnivores. Dashed lines represent a simple linear regression for herbivore subsets, solid lines represent a simple linear regression for carnivore subsets; see main text for regression information.

Modern mammalian population density was significantly negatively associated with body mass in the global dataset (−0.27 ± 0.09, p < 0.01) (figure 3c) and was negatively but not quite significantly associated with body mass in the North American data subset (−0.32 ± 0.16, p = 0.054) (figure 3b). Confidence intervals of slopes for the Holocene zooarchaeological and North American datasets overlapped but did not overlap between the Holocene zooarchaeological and global datasets (electronic supplementary material, table S4), indicating that these two datasets were significantly different. When the data were subdivided by the trophic guild, we found a significant negative body mass–abundance relationship across all modern mammal population density datasets (figure 3b,c; electronic supplementary material, table S4).

We found no significant body mass–abundance relationships across either trophic level subset of Holocene data. However, for herbivores, the slope was negative and 95% confidence intervals overlapped with modern herbivore datasets, indicating that it was not significantly different. By contrast, for carnivores, the slope was positive and did not overlap with modern carnivore datasets, indicating that it had a significantly different relationship (figure 3a–c; electronic supplementary material, table S4).

The number of zooarchaeological records varied considerably between countries (figure 4). However, when country size was controlled for, we found no association between the density of zooarchaeological records and a country's GDP per capita (0.37 ± 0.31, p = 0.24) or average elevation (−0.20 ± 0.16, p = 0.21). Elevation profiles showed that zooarchaeological records were present up to 2000 m above sea level (m.a.s.l.) but were generally absent at higher elevations (electronic supplementary material, figure S2). While the overall elevation profile for Europe reaches 3000 m.a.s.l., over 95% of the continent is below 2000 m.a.s.l., indicating that the two datasets were not significantly different.

Figure 4.

Map of Europe showing the density of mammalian Holocene zooarchaeological records (Azimuthal Equidistant projection, cell size: 1° × 1°).

The precision of dating for records across the Holocene was also temporally uneven. Overall, the average date range for each zooarchaeological record became narrower closer to the present, with this shift becoming particularly marked around 0 BC/AD (figure 5). Accumulation of successive records across the Holocene was not linear over time, showing a low rate of accumulation during the early Holocene, relatively steady accumulation for much of the mid-Holocene and accelerating in the later Holocene close to the present.

Figure 5.

Plot showing (i) ranges of estimated dates for all zooarchaeological records through the Holocene (black bars), and (ii) the accumulation of records in the zooarchaeological record through time (red line). The grey slope denotes a theoretical linear rate of accumulation.

4. Discussion

(a). Bias in Holocene mammal baselines and extinction risk

Our analysis of description dates for globally extinct mammals demonstrates that we still have an incomplete and biased understanding of mammalian extinctions and past levels of biodiversity even for the Holocene, the most recent interval of geological time, with numbers of ongoing new species descriptions continuing to increase over time. Indeed, the discovery curve of recently extinct mammals shows a nearly exponential increase from the 1950s onwards, rather than any signs of levelling off close to the present. This pattern contrasts markedly with the global trajectory of species descriptions for extant mammals over recent decades, which continue to increase but at a relatively reduced rate compared to the total number of extant species already known [43]. Estimating the number of recently extinct mammal species that remain to be discovered and described is therefore an important research goal. However, extrapolating the number of undescribed species from a temporal pattern of past species descriptions is associated with large margins of error unless the inventory of a group is largely complete and beginning to level off over time, which is demonstrably not the case for recently extinct mammals [44,45]. Rapid advancements in ancient DNA and genomics have also facilitated the identification of new species based on fossil material and museum specimens in recent years, and these techniques will doubtless only increase our ability to describe new species into the future [46]. Overall, it therefore remains surprisingly challenging to make even basic assumptions about past patterns of diversity, biogeography, community composition and severity and dynamics of past human impacts for one of the most-studied and ecologically significant animal groups.

Our analysis also highlights that we have a poorer understanding of past diversity of recently extinct small-bodied mammals, notably rodents and bats, compared to larger-bodied species (e.g. carnivores, artiodactyls, perissodactyls, proboscideans). Small mammals continue to be described regularly from the Holocene palaeontological and zooarchaeological records, while the largest species were mostly described by AD 1900. The earlier discovery of larger-bodied mammals in the Holocene record is associated in part with taphonomic processes that preferentially select for the deposition and survival of larger skeletal elements [15]. It also probably reflects the common historical preference for taxonomists to describe larger, ‘charismatic’ species first for scientific kudos [7,20]. Even without explicit size-based biases in fossil preservation or description, specific targeted sampling strategies such as fine-mesh sieving are needed to recover remains of smaller vertebrates and invertebrates, and these were not widely employed until the latter half of the twentieth century [47]. Once recovered, smaller bones and teeth can also be more difficult to identify to species level [48], so that the quality of described data can easily be biased towards larger species in Holocene faunal assemblages.

The extinct Holocene mammal fauna currently consists largely of insular species, with geographical hotspots of known global Holocene species extinctions including the Caribbean, Madagascar and insular Mediterranean, and with the vast majority of insular species (nearly 82%) only described during the last 100 years. However, we only found a significant relationship between body mass and description date for continental regions, and not for insular regions. On islands, this lack of significance is at least partly attributable to the general pattern of reduced body masses observed in extinct and extant insular mammals, associated with ecological resource limitation driving dwarfing of large-bodied lineages under the ‘island rule’ [49]. By contrast, continental regions harboured a wider range of body masses, with the largest generally being found and described first. Many species may also have been described later from insular regions owing to their increased geographical remoteness and inaccessibility, and because many major island systems are located in the tropics where preservation of long-term environmental archives is generally poorer and more recently developed techniques can be required to identify and date regional faunal remains [50,51]. Tropical species also tend to exhibit smaller geographical range sizes, which has also been shown to be a negative correlate of description date in mammals [52].

Small body size has been associated with an overall lower risk of extinction for mammals in both the past [5] and the present [53,54], principally because smaller-bodied species have higher reproductive rates that enable faster population growth and recovery [53]. However, this pattern may be more complicated within specific geographical areas and ecoregions, with higher levels of recent extinction and current risk observed in the smallest-bodied species for some regional faunas (e.g. Australia, Caribbean) [53–55], and it has been recognized that the very smallest species in some vertebrate groups may be at higher risk of extinction owing to small geographical range sizes [56]. Our analyses indicate that our baseline understanding of mammalian diversity remains incomplete at both global and regional levels, with our knowledge of the extent of recent small mammal extinctions in particular likely to be underestimated owing to incomplete and biased sampling. This bias in our baseline hinders our ability to compare recent levels of species losses and assess relative patterns of extinction risk across fundamental gradients of mammalian diversity and ecology.

(b). Bias in zooarchaeological data through space and time

Bias in species occurrence data has been studied extensively in ecological [57], historical [58,59] and palaeontological [60,61] datasets; however, our analysis represents, to our knowledge, the first attempt to directly quantify multiple sources of bias in a large zooarchaeological dataset. Owing to the fact that patterns of human influence may differ between zooarchaeological deposits, we note that our analysis relates specifically to large mammals for the European Holocene and may not necessarily reflect patterns of bias in other zooarchaeological datasets.

Our results show that body mass scales inversely to population density in natural populations of large mammals, a relationship that has also been demonstrated elsewhere [62,63]. We would therefore expect fewer large-bodied species to be present in the zooarchaeological record if zooarchaeological abundance reflected underlying patterns of ecological abundance and rarity in sampled source communities. By contrast, abundance increases with higher body mass in our zooarchaeological dataset; although this relationship was not significant, its slope did differ significantly from that of the global ecological mammal dataset. The power of body mass–abundance relationships in the North American and zooarchaeological datasets may have been reduced by the small sample sizes and smaller range of body masses available for analysis, with paucity of data points leading to a type 2 error, a problem that has also been noted in previous large-scale analyses of fossil mammal assemblages [64]. The disproportionately higher abundance of large-bodied mammals in the zooarchaeological record probably reflects preferential human hunting of these species. The likelihood that larger-bodied vertebrates have been a primary focus of prehistoric human hunting effort is supported by the pattern of terrestrial and insular extinctions following human arrival or technological change across the globe during the late Quaternary [7,65], and wild megafaunal vertebrates continue to be overharvested for consumption in many regions today [66]. While the increased vulnerability of large-bodied mammals to humans is also strongly related to their slower reproductive life histories, these species are highly detectable and non-arboreal and tend to be diurnal, and so are likely to have come into contact with humans relatively frequently [17].

However, the abundance–body mass relationship observed in our zooarchaeological dataset was also confounded by trophic level, which was the only consistently significant predictor of zooarchaeological abundance across all models, with herbivores more abundant than carnivores. This almost certainly reflects the increased likelihood of prehistoric humans to hunt herbivores for subsistence, as well as the higher available biomass of large herbivores in ecosystems compared to carnivore biomass. Interestingly, the scaling of abundance and body mass in the zooarchaeological record also differs between trophic groups; small herbivores were more abundant than larger herbivores, a pattern shown in natural populations, whereas large carnivores were more abundant than smaller carnivores, which is significantly different from the pattern seen in natural populations [35]. This unexpected finding may reflect the fact that because herbivores were more heavily exploited overall for subsistence, humans were less discriminate and hunted all body size classes opportunistically, leading natural patterns of relative abundance to be reflected in the zooarchaeological record. The particular dominance of species such as red deer and wild boar (Sus scrofa) in Europe's zooarchaeological record may also be partly owing to forest laws that afforded protection to these ‘noble game’ species for recreational hunting, e.g. in mediaeval Britain [67]. By contrast, humans were less likely to come into contact with smaller carnivores and more likely to see large carnivores as competitors for both wild prey and domestic livestock, and so might have disproportionately targeted these species. Indeed, large carnivores are more likely to be involved in human-wildlife conflict than smaller carnivores for these reasons today [68]. Large carnivores such as wolf (Canis lupus) and brown bear (Ursus arctos) were also the focus of recreational hunting in mediaeval Europe [67].

Although the spatial spread of zooarchaeological data was uneven across Europe, we found that the relative wealth of a country did not influence research output across European countries once geographical size was taken into account. This pattern may partly reflect the fact that archaeology is a fairly international endeavour and therefore academics will often fund, carry out or collaborate on excavations in countries other than their own. For example, within our study area, there is a strong international academic presence in Anatolian archaeology [69] and parts of the Caucasus [70]. Furthermore, even where research funds are reduced, countries can nonetheless have strong traditions of academic research and achieve relatively high research outputs through the efforts of only a few researchers. A final reason may be that the literature searches carried out to compile the database included grey literature in numerous European languages as well as research published in international journals, and therefore reached a range of research repositories beyond those dependent on or linked to higher levels of funding [29,71]. This finding is therefore a strong argument for employing exhaustive and thorough data collection protocols when compiling species occurrence datasets based on zooarchaeological assemblages in order to minimize sources of researcher bias. There is also increasing availability of large, open-access databases of zooarchaeological and palaeontological records that will continue to facilitate the incorporation of long-term archives into biodiversity assessments [72]. However, as these datasets are usually derived from multiple and often secondary sources, great care is needed in auditing and curating in order to minimize the use of erroneous or poor-quality data [73]. We also found no elevational bias in the spatial spread of zooarchaeological data across Europe. This may be owing to the fact that many prehistoric human populations are known to have lived at high altitudes in Europe (and elsewhere globally), and that archaeological research is increasingly conducted in remote and hard-to-access landscapes including at higher elevations [74].

The precision of dating for zooarchaeological records, represented by the length of available date ranges, also varied across the Holocene, providing a further source of bias in our zooarchaeological dataset and influencing its ability to make inferences about past biodiversity baselines and human impacts. This temporal variation partly reflects the mix of absolute and relative methods used to date records; whereas the majority of our zooarchaeological records were indirectly dated, direct radiocarbon dates are generally associated with shorter ranges. However, in general, we found that zooarchaeological records were assigned increasingly specific dates nearer to the present. This finding probably reflects the fact that archaeologically defined time periods (e.g. Mesolithic, Neolithic, Bronze Age) tend to denote changes in human technology and subsistence, and so become narrower towards the present owing to the general acceleration of technological change through time, with archaeologists increasingly confident in constructing chronologies and assigning dates to archaeological material nearer the present. Ideally, zooarchaeological records would be directly radiocarbon dated to ensure that chronologies for investigating faunal turnover and extinction were comparably accurate and consistent across datasets. However, given the sheer quantity of data generated from archaeological sites and the costs involved in absolute dating, this is unfortunately still rarely a realistic option.

Our results also reveal periods of increased accumulation of Holocene zooarchaeological records over time. This pattern might reflect periods of prehistoric human population increase. For example, there is a notable increase in our accumulation curve from approximately AD 500 to AD 1200, coincident with the Mediaeval period, during which there was an estimated six-fold increase in human population in Europe and evidence for associated large-scale forest clearance [75]. Increased pressure on wildlife populations could therefore potentially have led to increased numbers of zooarchaeological records found within these sites. Conversely, it has been demonstrated that younger components will always be more abundant than older components in the archaeological record owing to increased taphonomic destruction of older faunal remains and greater detectability of stratigraphically higher samples [76]. These temporal trends therefore indicate that we have more abundant and precise records of faunal occurrence nearer the present, with important implications for using zooarchaeological datasets of faunal records through time to reconstruct biodiversity change.

5. Disentangling bias from reality in the Holocene faunal record

Our combined analyses demonstrate that bias is widespread, with sources and patterns of bias varying across taxonomic, spatial and temporal scales in the Holocene faunal record. Size bias is a particularly complex issue to resolve, especially when using the recent faunal record to investigate species population dynamics and extinctions at both local and global scales. For example, large mammals tend to be well identified and over-represented in the Holocene fossil and zooarchaeological records relative to their status in natural source populations, whereas our knowledge of many small mammals is generally more incomplete or even altogether unknown, providing a biased understanding of extinction risk across different taxa. By contrast, modern-day abundance has been shown to be a strong predictor of abundance in the Holocene faunal record for other vertebrate groups [18]. Sources and patterns of bias in Holocene faunal datasets therefore need to be carefully identified and quantified on a case-by-case basis, particularly accounting for potential variation in bias between different taxonomic or ecological groups, and considering appropriate spatial extents, timescales and types of faunal data that are most relevant to the specific parameters of different research questions.

However, while this study has addressed potential limitations associated with the use of recent fossil and zooarchaeological data ‘at face value’, we also recognize that Holocene faunal data can be a reliable and important indicator of biodiversity change when bias is accounted for. Indeed, we have previously used the zooarchaeological dataset studied above to reconstruct mammalian range change across the Holocene in Europe, and by controlling for sample size variation across time and between different species we were able to reconstruct the dynamics of range declines and identify taxonomic variation in vulnerability or resilience to human impacts over time [4]. These data provide the only available insights into a wide range of important questions surrounding past environmental and ecological conditions and how they have changed through time in response to different stressors or drivers, and so must remain an essential component of the toolkit available for biodiversity research.

Data accessibility

The datasets supporting this paper are available in the electronic supplementary material.

Authors' contributions

J.J.C. and S.T.T. designed the research and coordinated data collection; J.J.C. and B.C. analysed the data; and J.J.C. and S.T.T. wrote the manuscript.

Competing interests

We have no competing interests.

Funding

J.J.C. was supported by a Natural Environment Research Council grant (grant no. NE/G011745/1). S.T.T. was supported by a Royal Society University Research Fellowship (grant no. UF080320/130573).