National leverage points to reduce global pesticide pollution

.


Introduction
Pesticides have vastly improved global agricultural productivity, increased yields, and decreased yield variability and costs (Galt, 2008, Savary et al., 2019).In a context of growing demand for agricultural production, coupled with a growing impact of climatic change on agricultural productivity, pesticides play an important role in the global food system (Deutsch et al., 2018).At the same time, the use of pesticides has severe negative effects on the environment (Beketov et al., 2013, Larsen et al., 2017, Stehle and Schulz, 2015).Global pesticide pollution is not only an existing urgent problem, but might further exacerbate, e.g. because of global production shifts to countries with less stringent environmental policies.Accordingly, the reduction of pesticide pollution risks is a major goal for agricultural policy worldwide (Möhring et al., 2020a).For example, by 2030, the European Union aims to reduce pesticide use and risk by 50 % (Schebesta and Candel, 2020).The environmental risks from pesticide use in agriculture are spatially highly heterogeneously distributed around the world, e.g. between different countries (Tang et al., 2021).Yet, it remains unclear what share of these differences is explained by differences in countries' agricultural systems and policies and how much is explained by other factors, such as environmental differences (e.g.pest pressures) and sub-and supernational socio-economic variables (e.g.regional markets).
In this paper, we use a novel dataset of 21.4 million georeferenced 1 km 2 grid cells of information on pesticide pollution risks globally (Tang, et al., 2021) and employ a spatial regression discontinuity design to identify the role of countries in explaining differences in pesticide pollution risks.The spatial regression discontinuity design allows us to separate the risk of pesticide pollution that is caused by countries' institutions and policies and the risk that is explained by natural differences between the countries (Wuepper et al., 2020a, Wuepper et al., 2020b).More specifically, we quantify the role of national farming systems (e.g. the share of organic farming and of permanent crops grown) and policies (e.g. which pesticide farmers are allowed to use), controlling for other differences between countries (e.g.pest pressure, weather, soils, regional farming traditions and markets).
Previous studies have shown that pesticide use in agriculture and its associated risks on humans and environmental health are high, and are very heterogeneous across farming systems, regions, and countries (Galt, 2008, Kniss, 2017, Kudsk et al., 2018, Lai, 2017, Marcus and De Souza, 2020, Meissle et al., 2010, Schreinemachers and Tipraqsa, 2012, Tang, et al., 2021).Such differences stem from the type of crops that are grown, agricultural practices, pest pressures, and other environmental variables, as well as social, economic, and political contexts.Environmental differences are an obvious explanation why some countries have a higher pesticide pollution risk than others (Oerke, 2006, Savary, et al., 2019).However, environmental differences do not deterministically lead to differences in pesticide use and risks, but there are also a host of important socio-economic explanations.Especially, differences in national farming systems and policies can play a major role in this regard.For example, national pesticide regulations differ considerably between countries all around the world, even neighboring ones.For example, some pesticides that are allowed in Bangladesh are banned in India, whereas some that are allowed in India are banned in China; and some pesticides that are allowed in Colombia and Bolivia are banned in Brazil, and some pesticides that are banned in Mozambique are allowed in Malawi (Watts, 2019).Moreover, on average, the risk of pesticide pollution is reduced in organic farming, compared to conventional farming (Meemken andQaim, 2018, Seufert andRamankutty, 2017).Globally, many countries differ in how much of the national farming area is farmed organically (e.g. in Latin America this is 9 % in Uruguay, it is 2 % in Argentina, and it is below 1 % in Brazil, and in Europe it is 21 % in Austria, 13 % in Czech Republic, and 2 % in Hungary (FIBL, 2021, Willer andLernoud, 2018)).Finally, countries also differ in what they are producing, and some crops (e.g.grapes) require more pest management than others (e.g.sugar cane).Even pest pressures are not entirely controlled by environmental factors, but may also change due to differences in national responses to invasive species (Paini et al., 2016).
While global differences in pesticide use and risk across countries are documented (Tang, et al., 2021, Zhang, 2018) and despite the general presumption that national farming systems and policies must be important in shifting the risk of pesticide pollution up or down, so far, a rigorous econometric estimate of their importance was lacking.Along these lines, while there are several studies focusing on explaining differences in pesticide use and risk across farms and regions (Möhring and Finger, 2022, Möhring et al., 2020b, Staudacher et al., 2020, Tang and Luo, 2021, Waterfield and Zilberman, 2012, Wuepper et al., 2021), global evidence on the determinants of aggregate differences in pesticide use and risk across countries is missing so far.
In the following, we provide such a global analysis.After the quantification of how much countries matter generally for the global risk of pesticide pollution, we investigate the main explanations for this, globally and by world region (Europe, the Americas, Asia, Africa), focusing on various country-differences that could potentially explain why some have a higher pesticide pollution risk than others (e.g.stricter national pesticide regulations).We also examine potential synergies and trade-offs, and quantify the relationships of pesticide pollution with crop yield gaps as well as soil erosion.
The next section (2) provides explanations on our data and empirical approach.Section 3 presents our results, including robustness and sensitivity checks and explorations of mechanisms.Section 4 is a discussion and section 5 concludes.

Data and methods
This section provides a description of our data (a) and empirical framework (b), and which assumptions we are making and how we explore their plausibility (c).

Data and sources
We here describe our data and sources in three parts: (i) pesticide data, (ii) environmental and agricultural data, and (iii) socio-economic data.The available data comes at different temporal resolutions.Overall, our analysis corresponds to circa 2015, but some variables can date back to year 2000.
Pesticide Data) For this study, we selected 21.4 million 1 km 2 grid cells of agricultural areas that are located no further than 100 km away from at least one international land-border and matched them with the georeferenced global pesticide pollution risk scores from Tang, et al. (2021).The risk score data has an original resolution of 5 arcmin (about 10x10 kilometers at the equator), encompassing 38.54 million km 2 of agricultural land (including pasture), and covering the areas of 168 countries.The risk scores consider the ecological risk posed by 92 active ingredients (refer to Table S2 in the Supplementary Information of Tang, et al. (2021) for the full list of considered active ingredients) on four environmental compartments, namely soil, surface water, groundwater, and atmosphere.The estimation of risk scores is described in detail in Tang et al. and we provide here only a brief overview.To calculate the risk scores, the predicted environmental concentrations (PEC) of the selected active ingredients in each environmental compartments were first estimated using a spatially explicit environmental model (Trevisan et al., 2009).The model describes the fate and transport of pesticides and explicitly accounts for crop-specific pesticide application, soil properties (e.g., porosity, bulk density, organic carbon content, field capacity), groundwater characteristics (e.g., water table depth, groundwater thickness, net recharge rate), hydrometeorological conditions (e.g., rainfall, air temperature, evapotranspiration with respect to year 2015), pesticide physicochemical properties (e.g., adsorption capacity, degradation rate, volatility, ecotoxicity), and topography (refer to Table S1 in the Supplementary Information of Tang et al. (2021) for the details of the input data used).Next, the predicted no effect concentrations (PNEC) of the selected active ingredients in soil, surface water, and atmosphere were determined based on median lethal concentration (LC50) of earthworms, fishes, and rats (inhalation) (Lewis et al., 2016) using an assessment factor approach (Institute for Health and Consumer Protection, 2003).The PNEC for groundwater was taken as 0.1 µg l − 1 for all active ingredients following the European Commission guidelines (European Commission, 2006).Then, the risk quotient of an active ingredient was calculated as the ratio between PEC and PNEC.Because each grid cell has multiple cropping systems that receive applications of multiple active ingredients, the additive effect of pesticide mixtures was considered by calculating the risk point in an environmental compartment as the log-transformed sum of the risk quotients of all active ingredients used in that grid cell.Ultimately, the risk score in a grid cell was calculated as the maximum of the risk points across the four environmental compartments based on the assumption that there is no additive effect across environmental compartments (e.g., pesticide concentration in the atmosphere has no effect on aquatic organisms).A region can be considered at low risk if 0 < risk score ≤ 1, at medium risk if 1 < risk score ≤ 3, and at high risk if risk score >3.Based on the average species sensitivity distribution curve for pesticides, a risk score >0 corresponds to a more than 5 % probability for a random non-target species to be affected by pesticides, whereas a risk score >1 and >3 imply that the probability for a random species to experience an effect by pesticides is greater than 45 % and 90 %, respectively (Fig. 1), noting that an effect does not imply mortality.
The crop-specific pesticide application rates used to determine the risk scores were sourced from PEST-CHEMGRIDSv1.01 (Maggi et al., 2019) which provides the global georeferenced application rates for a total of 95 active ingredients used in six dominant crops (alfalfa, corn, cotton, rice, soybean, and wheat) and four aggregated crops (vegetables and fruit, orchards and grapes, pasture and hay, and other crops).For the estimation of risk scores, three active ingredients (i.e., Bacillus amyloliquefaciens, calcium polysulfide, and petroleum oil) were Fig. 1.The relationship between the pesticide risk score and the probability for a random species to experience an effect, determined based on the species sensitivity distribution curves using parameters in Nagai (2016).The figure is redrawn after Supplementary Fig. 1 in Tang et al. (2021).Details on the estimation are reported in Methods in Tang et al. (2021).
excluded due to insufficient information relative to their physicochemical properties and ecotoxicities.The application rates were estimated based on the crop-specific pesticide use data from the USGS Pesticide National Synthesis Project database (Baker, 2017).Among the 512 active ingredients provided in the USGS database, the top 20 most used (by mass) in each of the crop classes mentioned above were selected.Specifically, in each crop class, the active ingredients were sorted in descending order based on the total mass used and the top 20 with highest mass were selected.This selection leads to 200 active ingredients, but some of the active ingredients are recurring across different crop classes, hence, yielding a total of 95 unique active ingredients.The estimation at global scale was then conducted by first using spatial statistical methods that account for 20 covariates that include soil physical properties, hydroclimatic variables, agricultural quantities, and socio-economic indices (see Table 1 in Maggi, et al. (2019) for the full list of databases).Next, national factors (e.g., agricultural practices, machinery capacity, access to pesticides) were then implicitly accounted for by constraining the estimates against countrylevel pesticide use data reported on FAOSTAT (2019) (i.e., the total pesticide mass in a country cannot exceed that reported by FAOSTAT) and by explicitly accounting for country-specific approval for adopting genetically modified pesticide-resistant crops as reported by the International Service for the Acquisition of Agri-Biotech Applications ISAAA (2018), as well as country-specific pesticide bans (or no approval for use) as reported by European Commission (2016) and Watts (2019).Among the 168 countries included in the estimates, pesticide application rates of 28 countries (Afghanistan, Andorra, United Arab Emirates, Benin, Bosnia and Herzegovina, Democratic Republic of the Congo, Cuba, Djibouti, Western Sahara, Gabon, Georgia, Equatorial Guinea, French Guiana, Cambodia, Liberia, Liechtenstein, Mongolia, Nigeria, Puerto Rico, North Korea, Singapore, Sierra Leone, San Marino, Somalia, Serbia, Swaziland, Uzbekistan, Yemen) were not constrained against FAO as pesticide use data in those countries were not available in FAOSTAT.The estimates were validated and benchmarked against independent and publicly available national active ingredient use data from the United Kingdom, Australia, South Korea, and South Africa (see Technical Validation section and Maggi, et al. (2019) Figure 9).
Environmental and Agricultural Variables) Global temperature data comes from Menne et al. (2018), rainfall data comes from PSD (2019), and topography data comes from Robinson et al. (2014).The data on the global distribution of the potential native vegetation comes from Bastin et al. (2019) and is expressed as the percentage of each 1 km 2 grid cell that would be covered by trees without human impact.This is based on training a random forest machine learning algorithm on environmental variables that explain the observed tree cover in protected areas all around the world and then using these environmental variables to extrapolate this pattern outside protected areas.The data on countries yield gaps was originally produced by Mueller et al. (2012) for the year 2000 and we use an updated version for the year 2015 produced by Ritchie and Roser (2021).The data on soil erosion on croplands (in t/ ha/yr) comes from Borrelli et al. (2017).
Socio-Economic Variables) Data on countries' GDP per capita was obtained from the World Bank (2020).Country-specific data on machinery in use on agricultural land (unit/ha-cropland, on average between 2000 and 2017), fertilizer use (kg/ha, on average between 2000 and 2017), and employment in agricultural sector (on average between 2013 and 2019) were obtained from FAOSTAT (2019).Data on countries' organic farming share (in % farmland) comes from the FIBL (2021).Data on countries' pesticide bans was obtained from Watts (2019) and the European Commission (2016).Pesticide bans are a feasible proxy for pesticide policy: they are easy to measure, and thus this data is globally comprehensively available.However, we acknowledge that bans are by far not the only regulatory tool that policy makers have and use.Other policy instruments regulate amounts, timing, and mandatory safety precautions, and they are often geographically differentiated, such that they are e.g.stricter near sensitive water bodies etc.Thus, our list of pesticide bans is only a very coarse proxy to pesticide policy and it takes into account only the regulation aspect.

Spatial regression discontinuity design: Quantifying and explaining border discontinuities
Before we focus on the role of specific country characteristics, such as countries' organic farming share or their banning of certain pesticides, we begin with quantifying how much countries overall influence the risk of pesticide pollution.For this, we zoom into international border areas and analyze whether there are discontinuities in the risk of pesticide pollution across international borders.In section c below, we discuss the necessary assumptions that must be fulfilled such that this approach identifies countries' causal effect (Wuepper and Finger, 2022).The main idea is that countries' influence is sharply delineated by their borders whereas the influence of environmental and geographic confounders changes either smoothly or randomly across them (Hahn et al., 2001, Lee and Lemieux, 2010, Turner et al., 2014).As an example, most countries around the world are environmentally and geographically highly non-comparable and this makes it challenging to understand how much the difference in pesticide pollution between any two countries is explained by the two countries themselves vis-à-vis their locations.Moving closer and closer to the countries' border, they become more comparable, but there is a trade-off between reducing the influence of confounders and shrinking the sample size.Thus, a first step of our approach is to statistically estimate the optimal "bandwidth", i.e. the optimal maximum border distance beyond which we drop observations from our analysis (Calonico et al., 2020, Cattaneo andVazquez-Bare, 2016).We can then estimate a regression discontinuity model that identifies countries' border discontinuities, relying on the existence of a continuously distributed "running variable" (such as border distance in kilometers) and a sharp "cut-off"-value in the running variable that determines treatment (such as international borders that determine to which country an agricultural area belongs) (Wuepper, et al., 2020a, Wuepper, et al., 2020b).Formally, we estimate: Here, τ ij is the discontinuity between countries i and j.Y i (1) and Y i (0) are the two potential outcomes to our effect of interest, i.e. the pesticide pollution risk on the same agriculturally used land (with identical environmental conditions), one belonging to country i and one to country j.X is the distance to the border, which controls for all spatially continuously distributed confounding factors (e.g.rainfall and temperature), and c is the threshold at which the country influence abruptly changed (X = 0).The second part of the equation defines exactly how identification is achieved.We fit two separate regression lines on each border side and their vertical distance close to c identifies τ ij , under the assumption that without τ ij , the pesticide pollution risk would be distributed spatially continuously across the border (Wuepper and Finger, 2022).To estimate the average border discontinuity in pesticide pollution risk globally, we estimate the following regression: Here, Y n is the pesticide pollution risk of 1 km 2 grid cell n as computed by Tang, et al. (2021), T i is a binary indicator function for whether country i has more pesticide pollution than country j, such that τ ij identifies the globally average border discontinuity, d A i and d B i are the border distances on both sides of each border, such that β 1 and β 2 capture all confounders that correlate with border distance, θ b are border fixed effects capturing all general differences between the borders all around the world, ϑ n is each 1 km 2 grid cell's longitude and latitude, and ∊ n is an error term.
To estimate country-specific effects, we re-define the binary indicator T i to equal one if a grid cell belongs to the country of interest and zero if it belongs to a direct neighboring country.To explore the role of specific country characteristics, such as their share of organic farming for example, we re-define the binary indicator T i to equal one if a grid cell belongs to a country with a higher value in that explanatory variable than the country's direct neighbor at each border and zero otherwise.Standard errors were always clustered by border, except for country and border specific estimates, in which case standard errors were specified to be heteroscedasticity robust.

Assumptions and test
Our empirical framework is based on several assumptions that can be empirically assessed.We show the results of these empirical assessments in the section 3 below (Figs. 3 and 4).These assumptions are: 1) Our first and most basic assumption is that we have a continuously distributed "running variable" that has a clear and sharp "threshold" value at which our "treatment" of interest is switched on and off.This is given in our setting, as the distance of each grid cell to an international border gives us this "running variable", the threshold value is border distance zero, and the treatment that we are interested in is whether a grid cell belongs to one country or another, which changes exactly at the borders, by definition.
2) A second assumption is that there are no "confounding treatment" happening at the borders.For example, if the international borders between countries with more and with less pesticide pollution where systematically placed on top of environmental or geographic borders (e.g.biomes, mountains, deserts), then we could not identify how much of the estimated discontinuity in pesticide pollution risk comes from natural versus anthropogenic differences between countries.To test this, we used a summary measure for potentially relevant environmental differences that pre-date the creation of the international borders, which is the native vegetation potential absent human impact (Bastin, et al., 2019).We substituted this native vegetation potential for the pesticide pollution risk as a placebo-test and checked that we do not estimate a border discontinuity for this outcome.It should be noted that this does not guarantee the absence of natural discontinuities at individual borders, and in fact we know from the literature that some international borders do coincide with natural discontinuities (Wuepper, et al., 2020a, Wuepper, et al., 2020b).To probe whether this affects our results, we used two ways to divide the borders around the world into "natural" and "nonnatural".For this, we estimated the discontinuity in the native  vegetation potential separately at each border and then categorized the borders either according to the statistical significance of the discontinuity (at the 5 % level) or according to whether it is larger or smaller than 2 %, independent of statistical significance and confirmed that all estimates are similar to each other.3) A third assumption is that generally, the pesticide pollution risk is spatially continuously distributed and we do not estimate random discontinuities wherever we choose to look for one.To probe this, we arbitrarily moved all borders 20 km away from their real location and indeed did not estimate a discontinuity in those locations.4) A fourth assumption is that differences in the quality and reliability of the underlying data have no systematic influence on our results.
The pesticide pollution risk of Tang, et al. ( 2021) was generally estimated with application rates constrained against pesticide use data from the FAO but this was not possible for 28 countries for which data was not available.We confirmed that only focusing on "FAO verified data" does not meaningfully impact our results.
Together, these tests increase our confidence into the internal validity of our analysis.5) We also confirm that the border discontinuities in pesticide pollution risk are not driven by differences in dominant crops across borders.
The dominant crop in a location was determined using the harvested area maps estimated in Monfreda et al. ( 2008) which includes 175 crops.A crop was considered as the dominant if it has the largest harvested area across all crops.Whether we include crop fixed effects or focus only on border where the same crop is dominant on both sides, our results stay robust.More specifically, the reduction in the estimated discontinuity is barely visible and statistically insignificant.This means, the country-effect is not driven by which crops are grown, but by how they are grown, except for the general exception of the share of permanent crops, which are associated with an increased pesticide pollution risk.6) Finally, an important question regards the external validity of our results.For this we use two additional tests, which suggest that we can extrapolate our findings from the border regions.First, we find that the risk of pesticide pollution only slightly changes as a function of border distance and the pesticide pollution risk in border regions generally correlates strongly with the countries' average pesticide pollution risk (correlation above 90 %)

Limitations and uncertainty
There are a couple of limitations and uncertainties underlying this study.First, PEST-CHEMGRIDS does not include all registered active ingredients used in agriculture and may miss to account for some active ingredients common for use in other countries.In addition, it does not account for the illegal use of banned substances.The selection of top 20 active ingredients for each crop may also miss to account for those used at high dose but over a small area.Information about the use of active ingredients is currently scarce and sparse (Mesnage et al., 2021), and therefore benchmarking the estimates for all countries is currently not possible.However, PEST-CHEMGRIDS is currently the only publicly available, data-driven inventory that provides crop-specific georeferenced application rates for different active ingredients.Secondly, there are uncertainties in the data and assumptions taken to model the pollution risk.The estimation of the pesticide pollution risk assumed a single pesticide application at an annual rate, all agricultural fields are adjacent to surface water bodies, and maximum exposure of non-target organisms in time and in space.Thirdly, the estimation of the risk scores does not consider pesticide risk on human health, legacy pollution from pesticides that were banned before 2015, and the effect of the accumulation of pesticides and their degradation products over time and thus may not fully capture the pervasiveness of certain pesticides.
To account for the uncertainties in application rates and other environmental data, an uncertainty analysis was conducted in Tang, et al. (2021).Specifically, the sensitivities of 11 input variables (application rates, soil bulk density, porosity, water content, organic carbon content, water table depth, groundwater thickness, net recharge rate, topography, rainfall, and atmospheric temperature) were quantified based on Monte-Carlo approach, i.e., model simulations were repeated multiple times with values of variables randomly extracted over the parameter space.A total of 50,000 model realization per grid cell (i.e., about 60 billion realizations in total) were conducted with variables randomly extracted over ± 50 % of the reference values using a uniform distribution.The analysis shows that approximately 22 % of the risk score estimates are highly certain with less than 9 % having low certainty (see details in Methods and Supplementary Fig. 3 in Tang et al, 2021).

Results
Below, we begin with a quantification of the effect of national farming systems, based on revealing border discontinuities in the global risk of pesticide pollution, which we first plot and then estimate using a regression discontinuity design (Subsection A).We then continue with explorations of more specific explanations, as well as synergies and trade-offs (Subsection B).

Quantifying the effect of national farming systems
The basis for our analyses is a new global map of the environmental risk of pesticide pollution (Tang, et al., 2021).The risk of pesticide pollution is quantified as a risk score, representing the probability for pesticides to pose an effect on non-target organisms due to their use in agricultural settings, but does not include the risks to human health (see discussion in the data section).A risk score greater than 0 signifies that there is a more than 5 % probability for a random species to experience an effect due to pesticide use, whereas a risk score of 3 signifies that the probability is equal to 90 %.A major strength of the data is that it is consistently modelled at the global scale, guaranteeing that country differences are real and not just country-specific measurement artifacts.
Fig. 2a shows a lot of variation in the environmental risk of pesticide pollution -within and especially between countries.The latter is suggestive for the role of national farming systems, but it is not yet causally identified, as there exist alternative explanations for why countries might differ in their environmental risk of pesticide pollution, including simply natural environmental differences (Wuepper, et al., 2020a, Wuepper, et al., 2020b).To control for such confounding factors, we use a regression discontinuity design, which consists of two main steps.First, we limit our analysis to all areas statistically and optimally close to the countries' borders, to minimize the confounding influence of natural differences between any country pair.Secondly, in our actual econometric model, we implicitly control for all continuously distributed confounding factors, such as natural pest pressure, soil permeability, and rainfall, by controlling for the distance of each grid cell to the border.We then estimate whether there is a statistically significant discontinuity in the risk of pesticide pollution right at the border, which can only be explained by national farming systems and not by confounding factors.A similar empirical framework has been previously used to identify countries' role in the global rate of soil erosion (Wuepper, et al., 2020a), and how countries affect the relationship of nitrogen pollution and crop yield gaps (Wuepper, et al., 2020b).Its fundamental idea can be illustrated visually (Fig. 2b-f) by plotting the risk of pesticide pollution of grid cells close to a border and sorting all grid cells from countries with more pesticide pollution on one side and all grid cells from countries with less pesticide pollution on the other.A smooth transition of the risk of pesticide pollution across country borders would signify that national farming systems have no significant effect in influencing the pesticide pollution, whereas a sharp discontinuity, which is what we see here, suggests that differences in national farming systems do matter for the distribution of the risk of pesticide pollution (Fig. 2b).An important first check, which can also be visualized, is whether country borders might be endogenous on average, in the sense that perhaps the natural confounders we try to avoid also exhibit such a border discontinuity.We show that environmental variables such as topography, rainfall, temperature, and native tree covers do not exhibit a border discontinuity (Fig. 2 c-f) and thus, natural confounders cannot explain the apparent border discontinuity in the environmental risk of pesticide pollution (Fig. 2 b).
Plot shows ten point estimates and respective 95 % confidence intervals.The sample size within the optimal bandwidth of 24 km is 7,159,124.The main estimate is a discontinuity of 0.5 (1).As a first placebo estimate, the estimated discontinuity in the native vegetation potential is 0 (2) and as a second placebo test, we arbitrarily move all borders 20 km away from their real location, and there is also no discontinuity (3).When crop fixed effects are included (4) or we limit the analysis to places where the same crop is dominant on both side of the border (5), the discontinuity is basically unchanged.When we categorize border as natural (6) and non-natural (7), according to whether we find any statistically significant discontinuity in environmental conditions, both estimates are very close.When we categorize border as natural (8) and non-natural (9), according to whether we find any discontinuity in environmental conditions of more than 2 %, there is a slight difference but it is not statistically significant.Finally, when we use only FAO verified data, the estimate remains similar (10).
The observed patterns in the raw data already foreshadow our following rigorous econometric results (Fig. 3).Our first main result is that the average border discontinuity is a shift in the risk score by 0.5 points, which is considerable given that Tang, et al. (2021) classify the risk score as low when between 0 and 1, as medium when between 1 and 3, and as high when above 3.Thus, simulating a change in the risk score by 0.5 often means moving countries from one risk category (low, medium, high) to another.When shifting e.g. from medium-risk to lowrisk (i.e., a reduction of risk score from 1.5 to 1), the probability for a random species to experience an effect decreases from about 62 % to 45 %.However, the relationship between risk score and the probability for non-target organisms to experience an effect is not linear, that is, a reduction of 0.5 point in high-risk regions (e.g., risk score > 3) will have a smaller environmental gain than a reduction in medium or low-risk regions.Overall, our estimates imply that we could roughly reduce the global risk of pesticide pollution by about 33 % 1 if the countries causing more pesticide pollution would get to the level of their less polluting neighbors.This would require that at all the international border around the world, the pesticide pollution risk in the countries with higher values would be brought down to the level of the neighboring country with lower values.At some borders, this would mean only modest regulatory or market changesat other borders this would require vast and fundamental changes, including a change in policies, technology, and production focus.Fig. 3 shows that our global estimate is robust to a variety of tests (Fig. 3).For example, the estimated effect is unaffected by the inclusion of crop fixed effects and when we only analyze borders where the same crop is dominant on both sides, and we estimate the same effect at natural and non-natural borders.When we restrict our analysis only to data that could independently be verified with FAO data, the estimate also remains the same.In contrast, we do not estimate a discontinuity at 'placebo borders' that we randomly moved 20 km away from their real location (placebo test 1) and we also do not estimate a discontinuity in the native tree cover potential at the real border locations (placebo test 2).The next question is then, what exactly explains this important influence of countries.An important further consideration regards the external validity of our findings.How representative are pesticide dynamics near political borders compared to other agricultural areas?Fig. 4 shows two tests.First, we compare the average risk of pesticide pollution at different distance to the border (a), and second, we compare the pesticide pollution risk near borders with the countries' averages (b).Both comparisons suggest that our study focused on borders are able to reveal more general patterns.
Fig. 5a maps how each countries' national farming system affects the global risk of pesticide pollution.It should, however, be noted that these estimates are only directly comparable between neighboring countries because that is how they are estimated (see methods section).For example, Spain's farming system causes more pesticide pollution than that of its neighbors Portugal and France do, but Spain's farming system cannot be directly compared to Germany's farming system, with which it does not share a border.Fig. 5b maps the uncertainty of these estimates (i.e. the range of the 95 % confidence interval for each estimate).Finally, Fig. 6 shows the average discontinuity per continent.This reveals that despite the existence of the European Union and its Common Agricultural Policy, the largest differences between national farming systems are actually found in Europe.The second largest are found in Asia.
In the next subsection, we identify more specific and concrete explanations why exactly national farming systems are so important for the global environmental risk of pesticice pollution.Then, we also examine some potential trade-offs and synergies. 1The countries causing more pesticide pollution than their neighbors have an average risk score of 1.5.The estimated discontinuity is 0.5, which implies that half of all countries globally could reduce their risk score from 1.5 to 1 if they would close the gap to their neighbors, so the global pesticide pollution risk could be reduced by 66% x 50%.This number decreases if those neighbors increase their pesticide pollution in response.However, also the countries with lower pesticide pollution risk likely have potential to reduce their pesticide pollution, in which case an even higher global reduction in the risk of pesticide pollution is possible.

Policy levers, synergies, and trade-offs
To identify the main national policy levers to reduce the global environmental risk of pesticide pollution, we rely on similar analysis as we have done before, but sorting the observations differently.Instead of sorting by the pesticide pollution risk itself, we now sort by various country characteristics and again estimate whether we find a sharp discontinuity, once globally (Fig. 7 a) and once by world region (Fig. 7 b-e).The separate analyses by world regions are done for two reasons: First, between country-differences in potential explanatory variables have different magnitudes in different world regions, e.g.there is more variation in what pesticides regulations are in place in Europe than in the Americas.Secondly, many relationships can be expected to interact with contextual characteristics in a way that even the sign of a relationship can differ between world regions.
Our analysis leads to the following findings: The share of permanent crops (FAO, 2020) (e.g.grapevines, fruit trees) is associated with a higher risk of pesticide pollution globally and in all world regions except the Americas.On the contrary, the share of organic farmland (FIBL, 2021) is associated with a lower risk of pesticide pollution, globally and in all world regions.Yet, this is only statistically significant in Europe, with its large inter-country differences in organic farming (e.g. less than 1 % in Bosnia and Herzegovina, Ukraine, or Albania, and about 20 % in Austria (FIBL, 2021)).On other continents, the share of organic farming is more homogeneous at low levels, not allowing to identify similar effects as in Europe.Pesticide regulation (Watts, 2019) (measured with the number of banned pesticides per country) is globally the strongest predictor of countries' pesticide pollution risk.The more pesticides a country bans, the lower its pesticide risk.This, however, is again entirely driven by Europe and not statistically significant in other world regions.Indicators of agricultural intensity (FAOSTAT, 2019) (fertilizer, labor, and machines per hectare) are globally associated with a lower risk of pesticide pollution, but this is not statistically significant and differs widely across world regions.
In Africa, higher agricultural intensity is generally associated with a lower pesticide pollution risk, whereas in Asia, this holds for labor and machines but a higher fertilizer intensity is associated with a higher pesticide pollution risk.Neither globally nor in any world region have yield gaps significantly associated with the risk of pesticide pollution.Thus, higher pesticide risk is not necessarily associated with lower yield gaps, i.e. higher yields.However, the risk of pesticide pollution is negatively associated with the risk of soil erosion (Borrelli, et al., 2017) on cropland in the Americas and in Asia.There, soil conservation practices are highly relevant and often comprise chemical pesticides  such as glyphosate. 2 In contrast, there is a positive association between pesticide risk and the risk of soil erosion on cropland in Africa, where a higher risk of soil erosion and pesticide pollution is likely jointly caused by the same socio-economic and political constraints.Finally, we find no clear pattern globally for the countries' per-capita GDP to play an important role (World Bank, 2020) and a heterogeneous pattern in different world regions.In Europe, a higher GDP in a comparably poorer country (for European standards) is associated with a lower pesticide pollution risk and the same holds in Africa (in countries that are relatively poor for African standards).In the Americas, it is the opposite: in relatively poor countries (for American standards), a higher GDP is associated with a higher risk of pesticide pollution.Globally, there is no association between GDP and the risk of pesticide pollution for countries with relatively high GDP, whereas in Africa and Asia there is a tendency towards a positive association, but it is not statistically significant.
Overall, this analysis further suggests that with regulatory convergence (towards the more regulated side), the global risk of pesticide pollution could be reduced by 11 %, an increase in organic agriculture (such that the side with less organic agriculture reaches the level of the side with more organic agriculture), another 6 % could be achieved, and by changing the kinds of crops that are grown (reducing the share of permanent crops on the side with more of them to the level of the side with less), another 15 % would be possible.These estimates are based on the pesticide pollution risk discontinuities we are estimating between countries with more and less pesticide bans, higher and lower share of organic agriculture, and higher and lower share of permanent crops.

Discussion
The results presented above suggest that if national governments around the world would implement effective national and international action-plans, the global risk of pesticide pollution could be considerably reduced, even without necessarily reducing the yield of the main staple crops.We are estimating that around a third of the global risk of pesticide pollution is determined at the national level.An important caveat is that telecoupling of pesticide use must be dealt with (Liu et al., 2013).For example, our analysis shows that the promotion of organic farming helps countries to reduce their risk of pesticide pollution.But such step also reduces yields on average, incentivizing high-yield systems with high inputs and possibly pesticide use in other countries.
Similarly, differences in growing conditions aside, national agricultural systems of neighboring countries can be quite different and face quite different challenges.One example is the specialization of the agricultural sector, not only how much permanent crops, but also e.g.how much high value fruits and vegetables vis a vis the main staple crops.There are of course also major technological differences between To identify these explanatory variables, the regression discontinuity design was estimated after the observation have been sorted according to various country characteristics.For example, to explore the role of organic farming, observations from countries with a higher area-share of organic farming were given a negative border distance and the observations from the neighboring country were given a positive border distance.In Europe, we find a sharp border discontinuity between these countries, but not in the Americas or Africa.Based on 1,582,199 grid cells in Africa, 1,863,364 in the Americas, 2,481,233 in Asia, and 1,707,617 in Europe.
2 There are options to practice no till without heavy reliance on pesticides, but these are rarely used so far.countries, as well as differences in state capacity and budgets.All such factors mean that some countries can achieve a low pesticide risk more easily than others.This is why not only national but also international efforts are important.From a global perspective, there is a common interest in mitigating the global risk of pesticide pollution, e.g. to mitigate biodiversity losses, and this might require international knowledge, technology, and financial exchanges, as well as private sector regulations and initiatives.
Another important caveat are the possible trade-offs that countries face when trying to reduce adverse environmental effects from agriculture (Wuepper, et al., 2020b).We find e.g. that both in the Americas and in Asia, there is an inverse relationship between the risk of pesticide pollution and cropland soil erosion.The reason is that here, soil conservation with reduced tillage or no-till farming is widely practiced, which however currently often requires more reliance on chemical pest management (Margulies, 2012).These trade-offs need to be mitigated or eliminated, e.g. by strengthening the use of low-or no-pesticide input soil conservation practices (Jacquet et al., 2022, Knapp andvan der Heijden, 2018).This can be achieved by increasing agricultural R&D spending, which offers many other benefits and has a high average return on investment (Alston et al., 2021), as well as optimizing extension services (Wuepper, et al., 2021).Also agri-environmental schemes might play a much more prominent role in the future to improve the environmental performance of the agricultural sector, even though so far, experiences have been mixed (Hasler et al., 2022).Importantly, by combining these components, synergies can be unlocked, and it should be noted that especially agri-environmental payments schemes are still mostly in a learning phase, in which governments learn what works best (Wuepper and Huber, 2021).
Policy action shall be especially in three areas: First, increase pesticide use efficiency (e.g. using precision farming techniques), secondly, substitute the most risky pesticides with less risky alternatives (e.g. using biocontrol agents), and third, redesign agricultural systems to reduce the need for pesticides (e.g. using agroecological practices, breeding new varieties) (Finger, 2021).More generally, countries' efforts to reduce pesticide pollution risks optimally target the entire food value chain (input suppliers, farmers, the food industry, and consumers) (Möhring et al., 2020a).
Moreover, it is necessary to address current data gaps.For example, when and how which products were applied to which crop is only coherently documented in a few places on earth (Mesnage, et al., 2021) but this is crucial information for further research and to guide policies.
Our analysis suggest potential for several lines of future research.First of all, there are some factors that likely contribute to pesticide pollution risks that we could not study at the global level, but that are relevant for policy and literature.For example, there are differences between countries and regions how pesticides are handled, ranging from regulation to training, extension and information to technical capabilities.Thus, the same chemicals can pose quite different risks, depending on their handling.Then, so far we are only able to examine the environmental risk at the global scale, but it would be important to also look into the human health risk and differences across crop systems, regions and countries.Moreover, more evidence is still needed on the economics of pesticide use, their profitability and economic risk effects, and complementary explanations for farmer pest management decisions, such as behavioral factors that drive pesticide use (e.g.culture and noncognitive skills).Moreover, moving beyond a strict focus only on farmers, a very promising research avenue is a broadening of focus towards entire food systems and study e.g. the (potential) synergies between actors, and how coordinated actions can move such systems towards higher sustainability.Finally, our global analysis is necessarily broad and general, so smaller scale studies that focus on specific contexts are now highly complementary.
Our analyses are subject to the following limitations: First of all, our preferred data would be precise measurements of pesticide pollution.This is not available at the global level and we have to rely on modelled pesticide pollution risk data (Tang et al., 2021).This is necessarily based on input data of varying resolution and reliability, as well as modelling assumptions.It also only captures environmental effects and not effects on human health, and there are no interaction effects between different active ingredients by assumption.The data has been carefully tested but measurement error is unavoidable at this point.Also our main explanatory variables are not precisely measured.Our proxy for pesticide policies is the number of banned substances, which is but one dimension of several one could imagine.The crop yield gap data currently available at the global level is similarly coarse, based on data of varying precision and quality as well as modelling assumptions.Permanent crops are just one type of crop that requires above-average pest management, and even the share of organic agriculture only captures what is certified as organic and thus misses what is organic but not officially registered as such.Our analysis then focuses on international border areas, so that we can answer our research questions using a spatial regression discontinuity design.We have explored the external validity of this approach and find a generally high external validitybut this is tested at the global level and this surely differs among all the borders we have in our data.Finally, our analysis of the explanations for why there are such striking border discontinuities in the global risk of pesticide pollution is less clear than our analysis of the overall between country-differences.This is because the explanations are not independent from each other.Our final estimates imply e.g. that the global risk of pesticide pollution could be reduced by 11 % via stricter pesticide regulation, by 6 % with more organic agriculture, and by changing the kinds of crops that are grown, another 15 % would be possible.Because empirically, choice of crops, organic farming, and pesticide regulations all affect each other, these estimates should be interpreted as rough indications and not as precise quantifications.

Conclusion
We identify national leverage points to reduce global pesticide pollution risk, using a novel dataset of georeferenced grid cells and a spatial regression discontinuity design as empirical framework.This allows us to separate the risk of pesticide pollution that is explained by countries' farming systems and policies from the risk of pesticide pollution that is explained e.g. by natural environmental factors, or even socio-economic factors that are not determined at the national level (e.g.idiosyncratic risk preferences).The main leverage points are countries' pesticide regulations, organic farming, and the types of crops the countries grow.We find a trade-off between pesticide risks and soil erosion only in the Americas and in Asia, but not elsewhere, and we do not find a significant trade-off between pesticide risk and crop yield gaps in any part of the world.We estimate that a third of the global pesticide pollution risk is determined at the national level, mostly driven by the policy levers we identify (regulations, organic agriculture, crop types).This implies a large potential for national governments to contribute to a global reduction in the risk of pesticide pollution.Especially important in this regard are, besides domestic improvements, international cooperation and a streamlining of policy efforts to reduce the global risk of pesticide pollution.
Future research might benefit from the availability of new and improved data, allowing to go deeper into the details.It would also be highly interesting and policy-relevant to analyze changes over time, both in general, and to understand how the pesticide pollution risk changes after the introduction of policy changes, new trade deals, or other political and economic changes.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Fig. 2 .
Fig. 2. Border Discontinuities in the Risk of Pesticide Pollution.a, The global distribution of the risk of pesticide pollution(Tang et al., 2021).b, a plot of the spatial distribution of the risk of pesticide pollution as a function of border distance, with observations sorted according to which side of each border has a higher pesticide pollution (on the left with negative border distances) and which has a lower pesticide pollution (on the right with positive border distance).This reveals a sharp border discontinuity at the "global average border".c-f, comparison plots with topography (average slope in %), rainfall (in 100 s of Millimeters), temperature (in degrees Celsius), and the natural tree cover potential (in %)(Bastin et al., 2019), which show that natural differences are spatially continuously distributed across this "global average border" and thus cannot explain the pattern in b.

Fig. 4 .
Fig. 4. External Validity.The risk of pesticide pollution at different border distances (a) and the relationship of the risk of pesticide pollution in border regions and the rest of the countries (b).

D
.Wuepper et al.

Fig. 5 .
Fig. 5.Estimated Border Discontinuity by Country.a, the estimated average border discontinuity in the environmental risk of pesicide pollution aggregated by country.b, the 95% confidence intervals for each country's estimate.Both maps are based on extrapolations from the country-effect estimated at borders, using a spatial regression discontinuity design and the data from Tang et al. (2021).Estimates are based on 7,159,124 individual grid cells.

Fig. 6 .
Fig. 6.The average border discontinuity in the risk of pesticide pollution per continent, based on 1,582,199 grid cells in Africa, 1,863,364 in the Americas, 2,481,233 in Asia, and 1,707,617 in Europe.

Fig. 7 .
Fig. 7. Explanations.Globally (a) and by continent (b -e).To identify these explanatory variables, the regression discontinuity design was estimated after the observation have been sorted according to various country characteristics.For example, to explore the role of organic farming, observations from countries with a higher area-share of organic farming were given a negative border distance and the observations from the neighboring country were given a positive border distance.In Europe, we find a sharp border discontinuity between these countries, but not in the Americas or Africa.Based on 1,582,199 grid cells in Africa, 1,863,364 in the Americas, 2,481,233 in Asia, and 1,707,617 in Europe.

, Ziwei; Felicia Wu and David A Hennessy. 2021
. "Environmental and Economic Concerns Surrounding Restrictions on Glyphosate Use in Corn."Proceedings of the National Academy of Sciences, 118(18).