A global assessment of drivers and risks associated with pollinator decline

Pollinator decline has attracted global attention, and substantial efforts are underway to respond, through national pollinator strategies and action plans. These policy responses require clarity on what is driving pollinator decline, and what risks it generates for society, in different parts of the world. Using a formal expert elicitation process, we evaluated relative regional and global importance of eight pressures driving pollinator decline, and ten consequent risks to human well-being. Our results indicate that global policy responses should focus on reducing pressure from changes in land cover and conguration, land management, and pesticides, as these were considered very important drivers in most regions. We quantify for the rst time how the importance of drivers, and risks from pollinator decline, differ among regions. For example, losing access to managed pollinators was only considered a serious risk to people in North America, whereas yield instability in pollinator-dependent crops, classed as a serious or high risk in four regions, presented only moderate risk in Europe and North America. Overall, perceived risks were substantially higher in the Global South. Despite extensive, research on pollinator decline, our analysis reveals considerable scientic uncertainty about what this means for human society.

IPBES's more recently published global assessment on biodiversity and ecosystem services 21 , the pollinators assessment did not directly compare the relative importance of major drivers of pollinator decline, or make any integrated assessment of the risks it generates for society, either at global or regional levels. Consequently, although researchers have made broad, global recommendations about how to respond to pollinator decline 2 , addressing speci c drivers and risks at national or regional scales appropriate for policy implementation has been more challenging, resulting in often ineffective policies 22 .
Here, we used a structured expert elicitation technique and a globally representative group of pollinator and pollination experts to evaluate the relative importance of eight major direct drivers (or causes) of observed pollinator decline, and the risks to human well-being associated with ten direct impacts of pollinator decline de ned by the IPBES report 4 (Table 1; Supplementary Table 1). We separately assessed each of six global continental regions, with the exception that, for biogeographic and geopolitical reasons, the Paci c islands were grouped with Asia (Asia-Paci c) and not with Australia and New Zealand (see Methods; Figure S1). Indirect impacts, such as increased land conversion in response to lower crop yields, were not assessed. We did not consider interactions between multiple drivers, although such interactions are likely to in uence pollinator decline 1 , because knowledge about driver interactions remains largely incomplete and insu cient for the scale and scope of analysis here.
Understanding and communicating risks to human well-being associated with biodiversity loss play a central role in raising awareness of our mutual dependence on nature, and in driving the transformative societal change required to conserve and restore global biodiversity worldwide 23 . We take a scienti ctechnical approach, in which a risk is understood as the probability of a speci c hazard or impact taking place. We used a semi-quantitative risk matrix, with risk scores calculated as the product of probability, scale and severity of impacts, and a 'four-box model' (Table 2) established by the IPBES to communicate levels of con dence 4 , thus highlighting the key known "unknowns" in current scienti c understanding.
Our assessment used a modi ed Delphi technique 24 , an approach designed to reduce bias, but particularly suitable for elicitation of expert judgements about complex issues, where the judgement requires a range of different perspectives and areas of expertise not necessarily held by each participant 24 .

Results
What's driving pollinator declines? Figure 1 shows nal scores for the importance of the six drivers de ned in Table 1, following three rounds of scoring. Globally, land cover and con guration, and land management were the most important drivers of pollinator declines (Figure 1; Ext Data Tables 2 & 4). Land cover and con guration was scored 'very important' in all six regions, while land management was the only variable considered to be 'the most important' in any region (Europe) and was 'very important' in all other regions except Africa (Figure 1). These conclusions are supported by considerable evidence from multiple regions [25][26][27] and continuing global trends towards agricultural expansion, conventional intensi cation, and urbanization in regions of the Global South, driven by international trade 28 . Land management was considered less important in Africa, where access to the necessary nancial and technical capital to intensify production is still limited 29 and where there was considerable uncertainty (categorised as 'inconclusive') over the in uence of land cover and con guration (Figure 1).
Pesticides were scored as 'important' or 'very important' drivers of pollinator decline in all regions, with the greatest con dence in Europe and Asia/Paci c (Figure 1). Pesticides were considered less important than land use and land management in Europe and Australia/New Zealand, but much more important in Africa ( Figure 1). The adverse effects of pesticides on pollinators have received considerable attention in recent years, following studies demonstrating widespread exposure 30 and detrimental effects on populations 31,32 or diversity 27 . There is far less evidence available to quantify the exposure in regions beyond Europe and North America. Also, pesticide regulations are weaker in the Global South, adding considerably to the risk 4 .
Climate change was considered an 'important' or 'very important' driver in every region. There was, however, unanimous lack of con dence over its importance relative to other drivers. In every region except Africa median con dence scores were 'medium' and in Africa, seven of the ten scorers responded that climate change effects are 'unknown' (Figure S2 and Supplementary Table 2). Long-term data scarcity limit and confound the demonstration of current climate change effects on pollinators, and available studies are restricted to few taxa such as bumblebees 13 and butter ies 33 .
Genetically modi ed organisms (GMOs) were considered the least important driver overall, except in South America (Figure 1), which is the second largest producer of GM crops among our regions, after North America 34 . Emerging evidence of potential impacts of herbicide-tolerant crops and associated glyphosate use on honey bees was discussed in the South American context (now reviewed 35 ). Levels of con dence and agreement were lower overall for GMOs and invasive alien species as drivers of pollinator decline, due to very limited available evidence. In the case of GMOs, impacts are di cult to separate from the effects of land cover and con guration, because such crops are often produced in large monocultures.
What are the risks to human well-being? Figure 2 shows the nal risk scores following three rounds of scoring, partitioned into probability and magnitude (scale × severity), for each of the direct impacts listed in Table 1, in each major global region. Overall, loss of wild pollinator diversity and crop pollination de cit were the highest and most widespread risks, scoring as serious or high risks in every region (see Figure 2, Supplementary Tables 3 & 7). Although much of the published evidence for pollinator declines is from Europe and North America (where the evidence was considered 'well established') 1 , there is growing evidence of pollinator declines in other regions 19 , including vertebrate pollinators 36 , along with global evidence of general biodiversity decline 23 . Evidence for pollination de cits is also growing across several regions 8, [37][38][39] (Figure 2), although for Australia/NZ and Africa, the degree of con dence was 'inconclusive', indicating low amounts of evidence and low agreement among our experts (see Table 2 for de nitions). This is a particular concern in Africa and Asia-Paci c, where pollinated crops are both nutritionally 5 and economically 40 valuable to livelihoods and well-being. Yield instability in pollinator-dependent crops, which is higher than that for non-dependent crops at global scale 41 , was classed as a serious or high risk in four of the six regions but moderate in Europe and North America, where highly pollinator dependent crops tend to be less widely grown and less important to total agricultural output. Direct impacts of wild fruit production losses had very low risk scores in economically developed regions of North America, Europe and, Australia/New Zealand (median scores <6), but, classed as a serious risk in Africa, Asia-Paci c and, South America ( Figure 2). These regions are dominated by low-to middle-income countries, where at least for Africa and Asia-Paci c, large portions of the population live in rural communities 42 .
Risks were greatest in South America compared to other regions (Supplementary Table 3: mean risk score across all ten impacts = 48.2), with four 'high' risks (pollination de cits, yield instability, food system resilience and wild pollinator diversity) and ve 'serious' risks (all others except managed pollinators). This re ects the high diversity of insect pollinated crops grown and exported throughout the region, often by smallholder farmers in and around areas of natural habitats that contain a high diversity of pollinating insects 43 . Continuing losses of pollinators are therefore likely to destabilise both regional food production and international trade, affecting livelihoods across the region. Like other regions of the Global South, South America is also home to a high diversity of extant indigenous cultures and people, many of whom rely on subsistence agriculture and natural resources such as non-timber forest products 44 , increasing the risks from a decline in honey, wild fruits, and cultural values.
In contrast to South America, Africa had very low risk scores for honey production and managed pollinators (both 'low' risk; see Figure 2 and Supplementary Table 3). Beekeeping is unique in Africa since it is the only global region that has large, genetically diverse populations of native honey bees (Apis mellifera) still thriving in the wild 45 . In fact, numbers of managed hives are increasing in many African countries due to limited colony losses and managed honey bee populations relatively resilient to Varroa mite 46 .
The risk of loss of aesthetic values, happiness, or well-being associated with wild pollinators or wild plants dependent on pollinators was perhaps the most di cult to score in all regions. In some contexts, one can make an argument that aesthetic values associated with pollinators are increasing, as people become more aware of their roles, beauty, and diversity. Discussions focused on what constitutes aesthetic values and how they might be changing in response to pollinator decline. This risk varied regionally, with South America and Africa scored highest (42) and lowest (4) risk, respectively (Fig. 2,  Supplementary Table 3). While clear links exist between people and pollinators or pollinator-dependent plants in both regions, for South America, these links are often related to speci c threatened taxa, such as hummingbirds and orchids. In Africa, connections with pollinator-dependent plants are frequently associated with entire landscapes, such as the ower-rich shrubland of Namaqualand, southern Africa, making potential impacts of pollinator decline on aesthetic values less clear.
Europe was the region where human well-being was considered at the lowest risk from pollinator declines overall (mean risk score = 19.6), with no 'high' risks, and only two 'serious' risks (pollination de cit and wild pollinator diversity). Unlike South America, many European countries grow few crops that are highly pollinator dependent and food systems, particularly within the European Union, are highly industrialised and globalised, greatly reducing the importance of wild fruits and buffering against the impacts of global change on food system resilience (both 'low' risk). Despite evidence that habitats containing pollinatordependent plants are aesthetically valued in Europe 47 , their cultural importance may be lower than elsewhere in the world, although this was highly uncertain, with our risk score for 'cultural values' in Europe categorised as 'inconclusive' due to low con dence and low agreement among scorers.
Loss of access to managed pollinators was only considered a serious risk to people in North America, where honey bees Apis mellifera represent a key input to large scale, industrialised cropping systems such as almond 48 , and have suffered serious declines in the past due to outbreaks of disease, pests and 'colony collapse disorder' 49 . The probability of the same occurring in say, South America or Asia-Paci c, was considered far lower, even if the severity of the impact would be similar ( Figure 2). Experts were divided (low agreement) on the risk from losing managed pollinators in Europe (Figure 2), where markets for pollination services are less well developed 50 , and South America, where the number of managed honeybee colonies has expanded substantially but pressures on their populations remain high 9 .
Across both risks and drivers, there was high agreement but low con dence for most factors, placing them in the 'established but incomplete' con dence category. Our con dence in several direct impacts was low because of numerous gaps in knowledge about the ecology and status of all but the most common pollinator species, and the relationships between pollinators, human economies, and culture 20 . Furthermore, while statistical information on crop production, managed pollinators, and honey production is often collected at a national scale, the quality of these data varies considerably within a region and over time, and does not capture global subsistence agriculture, particularly in the Global South.

Discussion
Worldwide, the order of importance of drivers of pollinator decline in our analysis ( Figure 1) differs from the order of relative impact of direct causes of biodiversity loss (or 'changes in the fabric of life') presented by Diaz et al, based on the IPBES global assessment 23 . In both cases, land use change (here, land cover and con guration) for terrestrial realms is the most important, but for the whole of nature, 'direct exploitation' is the next most important driver, followed by climate change, pollution and invasive alien species. For pollinators, direct exploitation is broadly equivalent to 'Pollinator management' (not including direct harvesting of pollinators or pollinator products, which is not suggested as a major driver of pollinator decline). This was ranked with lower importance than climate change, pesticides, and pests and pathogens in our assessment. For pollinators, climate change was ranked below pesticides as a driver, perhaps re ecting more complete evidence that current pesticide use negatively impacts pollinator populations 14,31 , through a range of sublethal effects. Climate change impacts on pollinators are likely to be longer term. Much of the current evidence shows shifting ranges, which only sometimes translate into population declines 13 , or highly uncertain projected future distributions under climate change. Although these two analyses used different methods for ranking drivers (Diaz et al 23 quanti ed the relative impact of each driver, based on rankings in published studies comparing two or more drivers), it is not surprising that the relative importance of drivers differs, when focusing on a functionally de ned subset of organisms (pollinators) that are almost all relatively small in size.
Despite high pro le, extensive research on the drivers and impacts of pollinator decline, our analysis reveals considerable scienti c uncertainty about what this means for human society, regionally and globally. There are clear risks of wild pollinator diversity loss and pollination de cits globally yet less is understood about the broader implications for human well-being. The case for action to address pollinator decline is most clearly made for South America. Our process reveals several major knowledge gaps. There is an urgent need for research in Africa, to address the substantial uncertainties around the risks to people from pollination de cits, and the importance of changes in land cover and con guration, as a driver of pollinator decline. In more developed regions, especially North America, we lack understanding of the scale and severity of impacts of pollinator decline on human well-being. Globally, the consequences of climate change for pollinators and pollination remain poorly understood, but its impacts will clearly increase in prominence in the coming decades. As climate change is very likely to interact with other drivers of pollinator decline, a focus on how to mitigate and adapt to it should be central to pollinator research and conservation strategies.

Methods
We assessed drivers and risks using a modi ed version of a formal consensus method known as the Delphi technique 24 , in which the second and third rounds of anonymous, independent scoring took place following detailed discussions at a face-to-face workshop in November 2017. This modi cation of the Delphi technique is frequently used in environmental research, where issues are multi-disciplinary and interpretations of the same phrase can differ strongly among individuals 51 . All but one of the authors of this paper (hereafter 'experts') took part in all rounds of the Delphi process (D.S. facilitated only and did not score). This set of 20 pollination experts was carefully selected to cover the range of necessary expertise, including biodiversity science, economics, social science and indigenous and local knowledge, and to ensure that the main global regions were each represented by at least two scorers either originating from or mainly working in that region. Thirteen of the 21 authors (59%) were also authors of the IPBES global pollinators assessment 4 , mostly nominated by their respective national governments, and the team had a balanced gender ratio of 11 men : 10 women.

De nitions of regions, parameters and scores
We divided the world into six global regions, largely representing geographic continents of North America, South America, Asia, Europe, Africa and Oceania, with one key difference: we included the Paci c islands in a region known as 'Asia-Paci c', rather than combining them with Australia and New Zealand in the geographic continent 'Oceania'. Our 'Asia-Paci c' region is equivalent to most of the Asia-Paci c as de ned by IPBES, but excludes Australia and New Zealand. We named 'Australia/New Zealand' as a separate region, because they are very different from mainland Asia and the Paci c islands, both biogeographically and geopolitically (see Figure S1).
For each region, experts individually assigned probability, scale and severity scores for each of ten impacts of pollinator decline, and importance scores to each of eight drivers of pollinator declines de ned by the IPBES 4 (Table 1), using the ve-point Likert scales described in Table S1. All scores were accompanied by a con dence score of low, medium or high, enabling experts to qualify their judgements with a level of con dence, based on the amount of evidence they were aware of, and its quality.
The following de nitions of probability, scale and severity were available for authors to consult throughout the process: Probability: A high probability of impact suggests that the impact is already taking place or is very likely, at least in some circumstances. Low probability implies that the impact is not taking place or is unlikely. Unknown means there is not enough evidence to make a judgement on whether or not the impact is happening or likely to happen.
Scale of impact either refers to the numbers of people or area affected. Large means there is evidence for impacts on people and livelihoods, either over a large area or affecting many people. Moderate means there is evidence for impacts on people and livelihoods, either over a moderate area or affecting a moderate proportion of people, and small means there is evidence for impacts on people and livelihoods, either in a small localised area, or only affecting a small number of people. Unknown means there is not enough evidence on the scale of this impact to make a judgement.
Severity of impact refers to the nature of the impact on individual people or families. Large means there is evidence for a substantial or severe impact on people and livelihoods. Moderate means there is evidence for a moderate impact on people and livelihoods, and small means a small impact. Unknown means there is not enough evidence on the severity of this impact to make a judgement.
Experts rated the importance of each driver in affecting pollinators, at the present time, in each speci c region, on a 1-5 scale from 'not important' to 'the most important' (Tables 1 and S1).
We set an a priori expectation of consensus as an interquartile distance of < 2 between scores for a particular element (not including con dence). This still allowed us to distinguish between high and low agreement following criteria in Table 3, in which high agreement is denoted by mean IQR £1 (where half of all scores are the same or an adjacent score) ( Table 2).

Three iterative rounds of scoring
In an initial scoping phase, all experts were invited to comment on the proposed scoring structure described above. Following this, the rst round of scoring was conducted online in October 2017. Each expert was asked to score for all regions, considering the evidence in the IPBES report 4 alongside their own expertise. Experts could add comments to support their scores, and were encouraged to cite parts of the IPBES report 4 and other speci c literature. Scores and comments were compiled, anonymously, and summaries sent to all experts, detailing the median and interquartile range of scores for each element, and the proportions of 'unknown' responses.
Each expert was then assigned a region (always one they were familiar with) and a driver, and asked to play a cynic role, doing focused background research to challenge, refute or support the scores from the rst round, with evidence. Cynic roles were not made known during later discussions but cynics were invited to comment appropriately and to actively introduce new evidence to the discussions.
In November 2017, all experts attended a workshop in Reading, UK. Experts were divided into two groups, which each discussed the results from the rst round, and the evidence that supports them, for three regions. Group 1 discussed and scored in rounds 2 and 3 for Europe, North America and Africa; Group 2 discussed and scored South America, Asia Paci c and Australia/New Zealand. Discussions were facilitated and notes taken throughout. Facilitators kept in contact and discussed any speci c issues arising about how to score, to ensure that both groups responded in the same way. At the end of each part of the discussion, participants scored again for each element of risk, and each driver, for each region in turn. Scoring was conducted independently and anonymously, using Excel spreadsheets on personal laptops. All members of a group were encouraged to score for each region discussed in their group, with the following guidance: "Score if you can (but you don't have to). If you feel con dent to score for a region outside your own personal knowledge, please do so. These issues are complex and open to interpretation. This is why we employ a subjective scoring process, with anonymous scoring. Listen to the discussion, and then score as you understand it." These round 2 results were compiled as before, and any scores with interquartile range (IQR) ³2 (our a priori criterion for consensus), progressed to round 3 for rescoring.
Round 3 scoring took placed on the second day of the workshop in a plenary discussion. This allowed a further opportunity for any consistent differences in scoring or approach across groups to be revealed, but none were evident. Second round scores were presented and made the subject of debate and discussion. Experts scored again anonymously and independently, using laptops, for the regions they scored for in round 2, although the discussion was open to both groups. In total, 19 variables (3 drivers, 16 impacts) were rescored, along with associated con dence levels. ) with IQR ³2 were not agged for rescoring during the workshop and were later rescored during a teleconference. Only ve of the ten scorers from group 2 were able to attend the teleconference, due to time differences, so these four variables have only n=5 scorers in the nal dataset

Analysis
Median scores following the third round of scoring were used to derive risk scores (the product of probability, scale and severity scores) and associated risk categories (boundaries visualised in Figure 2), importance scores for drivers, and con dence categories for all nal scores, following criteria given in Table 2. In assigning con dence categories, the quantity and quality of evidence was based on assigned con dence scores for each risk or driver. The con dence score is the percentage of the maximum possible con dence score (9 for risks, 3 for drivers), represented by the median con dence scores from the nal round, with the three medians summed in the case of impacts (con dence score for risk = (å Con dence scores for probability, scale and severity/9) * 100)).
Overall global scores for the importance of drivers were calculated as a median of the six region-level scores and con dence scores, to ensure equal weight was given to each region (although the numbers were unchanged if individual scores across all six regions were used). We did not calculate overall global risk scores for different impacts of pollinator decline, because these scores were based on assessments of probability, scale and severity for different global regions and it does not make sense to average these across regions. All gures were drawn using the ggplot2 package 52 , in R version 4.0.0 53 .
We hypothesized that the scores participants gave for each component of the risk, or driver importance, were dependent on the impact, or driver, being scored, and on the region being scored, rather than re ecting individual scorer differences. We tested this hypothesis using Cumulative Link Models and Cumulative Link Mixed Models with logit link functions (also called proportional odds or ordinal logistic regression models), with the ordinal package 54 , in R version 4.0.0 53 . The top and bottom two score categories (scores 1 and 2, and 4 and 5 respectively) were collapsed to create three-point scales for probability, scale and severity of impacts, and importance of drivers.
We considered the effect of Region and Impact, or Region and Driver, on score, for each of four dependent variables: probability, scale, severity and importance. 'Unknown' responses were treated as 'na' for this analysis. The dataset was not large enough to examine the interaction between Region and Impact or Driver with this type of model (n£10 scorers for each combination of factors).
For each model, we tested the proportional odds assumption, that the effects of region or impact group were the same, regardless of where the cut-off points were placed across the three score categories, using the nominal test and scale test functions, which use likelihood ratio tests. When this assumption was violated, we used partial proportion odds models where possible, given our data structure. Independent variables that failed the tests were examined, with scale (dispersion of latent variable) allowed to vary among levels of the dependent variable (failure of the scale test) or effects of the relevant factor assumed to be nominal rather than ordinal (failure of the nominal test).
These models do not account for the random effects of scorer or group, because the scorers were divided among two separate groups, each of which only scored half of the regions. We ran Cumulative Link Mixed Models separately for each group, including scorer as a random effect to account for differences between individual scorers. The effects of group cannot be analysed as a random factor with this study design, because there are only two levels. The effect of Group cannot be separated from the effect of Region in a single model.
We used McFadden's pseudo R 2 value (r 2 ) to provide an indication of goodness of t for all models, as recommended by Menard (2002) 55 . This is calculated relative to a null model using the following equation: where LL mod is the log likelihood value for the tted model and LL 0 is the log likelihood for the null model which includes only an intercept as predictor (so that every score is predicted the same probability).

Pests and Pathogens
Parasites, pathogens and disease of all pollinating animals are included, both naturally circulating in populations and those associated with human management. "Bee diseases by definition have some negative impacts at the individual bee, colony or population level. Parasites and pathogens can be widespread in nature but may only become problematic when bees are domesticated and crowded." (Section 2.4.1) Pesticide use "Pesticides (fungicides, herbicides, insecticides, acaricides, etc.) are primarily used in crop and plant protection against a range of pests and diseases and include synthetic chemicals, biologicals, e.g., Bacillus thuringiensis (Bt) or other chemicals of biological origin such as spider venom peptides." (Section 2.3.1.) Veterinary medicines are also included. Land management "[…] Arrangements activities and inputs people undertake in a certain land cover type […]" (Section 2.2.1) This includes mowing, cultivating, grazing, burning and cropping regimes and non-pesticide inputs, particularly fertilizers. Pesticides were considered separately, as there are large amounts of evidence specific to them. Land cover and configuration "Land cover has been defined by the UN FAO as the observed (bio)physical cover on the earth's surface". (Section 2.2.1.) This includes the extent of different habitat and land use types, and their spatial configuration at landscape scale. Invasive alien species "Alien species' are defined as a (non-native, non-indigenous, foreign, exotic) species, subspecies, or lower taxon occurring outside of its natural range (past or present) and dispersal potential (i.e. outside the range it occupies naturally or could occupy without direct or indirect introduction or care by humans) and includes any part, gametes or propagule of such species that might survive and subsequently reproduce. 'Alien invasive species' are alien species that become established in natural or seminatural ecosystems, and are an agent of change, threatening native biological diversity" (Section 2.5.1) GMOs "Genetically modified (GM) organisms (GMOs) are organisms that have been modified in a way that does not occur naturally by mating and/or natural recombination. One of the most common methods to do this is by bioengineering transgene(s) into the new organism. The most common plant transgenes confer herbicide tolerance (HT), or toxicity towards herbivores (insect resistance, IR), although other characteristics have been also engineered (e.g., drought resistance in wheat, nutritional values in sorghum)." (Section 2.3.2.) Climate change "a change in the state of the climate that can be identified … by changes in the mean and/or the variability of its properties, and that persists for an extended period, typically decades or longer." (Section 2.6)