See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/337698100

Countries and the global rate of soil erosion

Article in Nature Sustainability · January 2020

DOI: 10.1038/s41893-019-0438-4

CITATIONS READS

246 13,095

3 authors:

David Wuepper Pasquale Borrelli

University of Bonn Università Degli Studi Roma Tre
54 PUBLICATIONS 1,879 CITATIONS 185 PUBLICATIONS 15,071 CITATIONS

SEE PROFILE SEE PROFILE

Robert Finger
ETH Zurich
279 PUBLICATIONS 7,788 CITATIONS

SEE PROFILE

All content following this page was uploaded by David Wuepper on 25 October 2020.

The user has requested enhancement of the downloaded file.

 Analysis
https://doi.org/10.1038/s41893-019-0438-4

Countries and the global rate of soil erosion

David Wuepper 1
*, Pasquale Borrelli 2
and Robert Finger1

Soil erosion is a major threat to food security and ecosystem viability, as current rates are orders of magnitude higher than
natural soil formation. Governments around the world are trying to address the issue of soil erosion. However, we do not know
whether countries have much actual control over their soil erosion. Here, we use a high-resolution, global dataset with over
35 million observations and a spatial regression discontinuity design to identify how much of the global rate of soil erosion
is actually affected by countries and which country characteristics, including their policies, are associated with this. Overall,
moving just across the border from one country to the next, the rate of soil erosion changes on average by ~1.4 t ha−1 yr−1, which
reveals a surprisingly large country effect. The best explanation we find is countries’ agricultural characteristics.

R
ising populations and incomes make it increasingly challeng-
ing to meet the world’s demand for agricultural products and
ecosystem services. We must produce more food and energy
as well as increase the provision of ecosystem services1,2. In this con-
text, the protection of the world’s soils and the prevention of soil
erosion is a high priority3–5. However, global soil erosion is currently
increasing6, despite many countries already attempting to address
this issue. It could of course be that countries play a relatively
unimportant role in the global rate of soil erosion. However, there
is ample evidence pointing in the opposite direction7–11. In the fol-
lowing, we provide a quantitative assessment of countries’ influence
on the global rate of soil erosion and we examine potential explana-
tions for the differences we see. The main empirical challenges here
are that countries’ territories also naturally exhibit distinct erosion
rates and both supra- and sub-national characteristics tend to cor-
relate with country characteristics. Thus, we cannot simply compare
countries’ average erosion rates to understand countries’ impact.
We are interested in the effect of an area belonging to one country
and not another—that is, the causal effect of countries on the global
rate of soil erosion.
An illustrative example
To make the empirical challenge clear, take the example of Haiti,
which has a much higher rate of soil erosion than Finland6. The
difference is unlikely to be solely the causal effect of the countries.
Specifically, the relative role of the countries versus other socioeco-
nomic effects and geographic differences is unclear. To find this
out and determine the effect of all the other countries around the
world we use a spatial regression discontinuity design12,13 with high-
resolution, globally gridded datasets6 (resolution: 1 km2; reference
year: 2012; number of observations: >35 million). The main idea is
that most international borders are predominantly ‘political’ (that
is, they have little or no overlap with a natural discontinuity in the
rate of soil erosion). Thus, actual border discontinuities in the rate
of soil erosion can be distinguished from natural border disconti-
nuities. An illustrative example is shown in Fig. 1.
Haiti and the Dominican Republic are located on the same
island (Hispaniola), sharing a comparable environment, especially
at their border, but they differ in their land-use trajectories. The
data of Bastin et al.14 show that without human presence, the whole
island of Hispaniola would be covered by the same dense natural
soil cover (Fig. 1a). Combining the data from Bastin et al.14 with
those of Borrelli et al.6, we are even able to model the natural rate
of soil erosion that we would observe if there were no humans on
the island. When we plot this natural rate of soil erosion in equal-
sized bins as a function of border distance, it can be seen that the
natural rate of soil erosion would show continuity across the border
(and so this border would be categorized as purely political) (Fig. 1b).
However, on satellite images15 it can be seen that the actual soil
cover is discontinuous exactly at the border, with denser vegetation
in the Dominican Republic (more forest), than in Haiti (more bare
land) (Fig. 1c). This translates into a clear border discontinuity in
the actual soil erosion rate6 (Fig. 1d), which is entirely human made.
Distinguishing political borders from environmental borders
To formally test the assumption that border discontinuities in soil
erosion do not simply reflect natural discontinuities, we globally
model the natural rate of soil erosion for each of our 1 km2 pixels,
combining data on the erosion effects of rainfall, topography and
soil characteristics6 and on the natural vegetation that we would
observe in the absence of human activity14 (Methods). We esti-
mate that about 40% of all borders are ‘natural’ in the sense that
there are statistically significant discontinuities in the natural rate
of soil erosion. Yet, these natural erosion discontinuities are small
(mean ≈ 0.03 t ha−1 yr−1). In contrast, the country discontinuities in
the actual rate of soil erosion are large: around 1.4 t ha−1 yr−1 on all
land and 1.8 t ha−1 yr−1 on cropland. These values reflect the average
impact of an area being in one country and not in another. In some
regions, the discontinuities are considerably larger. For example,
Brazil is on average causing 4 t ha−1 yr−1 more soil erosion than its
neighbours, whereas Germany is causing 0.2 t ha−1 yr−1 less. The dif-
ference in erosion rates between Haiti and the Dominican Republic
is 50 t ha−1 yr−1.
Quantifying soil erosion discontinuities at borders
For the estimation of the border discontinuities, we only use obser-
vations that are within a statistically optimal distance of the border.
The closer we restrict our data to the border, the more comparable
the observations become (for example, right at the border, obser-
vations share the same topography, rainfall, soil and naturally
occurring vegetation), but the estimator loses precision16. Thus,
there exists a maximum border distance at which we have a statisti-
cal optimum. In our study, this is on average 17 km on each side
of a border. In our baseline specification, we control for border

1
Agricultural Economics and Policy Group, ETH Zurich, Zurich, Switzerland. 2Environmental Geosciences, University of Basel, Basel, Switzerland.
*e-mail: dwuepper@ethz.ch

Nature Sustainability | VOL 3 | January 2020 | 51–55 | www.nature.com/natsustain 51

Analysis Nature Sustainability
a c

b Haiti Dominican Republic d Haiti Dominican Republic

Natural soil erosion rate (t ha–1 yr–1)

Soil erosion rate (t ha–1 yr–1)

150 150

100 100

50 50

0 0
50 0 50 50 0 50
Border distance (km) Border distance (km)

Fig. 1 | The border between Haiti and the Dominican Republic. a, Without humans, the island of Hispaniola would have a homogeneous natural soil
cover mostly made up of forest. b, Combining data on weather, geography and natural vegetation, we can model the potential natural rate of soil erosion,
which would be low and very continuously distributed across the border. c, In contrast, because of historically distinct deforestation rates, there is a sharp
discontinuity in actual soil cover between the two countries, as can be seen from space. A segment of the border is shown as a yellow line, with Haiti to
the left and the Dominican Republic to the right. Credit: NASA/Goddard Space Flight Center Scientific Visualization Studio. d, Compared with the natural
soil erosion rate, the actual soil erosion rate is much higher and shows a sharp discontinuity right at the border, which mirrors the soil cover discontinuity.
Because there is no natural erosion discontinuity, the actual erosion discontinuity reveals the impact of the countries.

distance (separately on each side of a border), include border fixed
effects and cluster the standard errors at the borders. Among our
multiple robustness checks, we also estimate the border discontinu-
ity separately at the 40% ‘natural’ and 60% ‘political’ borders, try a
very wide bandwidth of 70 km maximum distance on each border
side and include a quadratic polynomial of longitude and latitude
as well as the predicted natural rate of soil erosion. Finally, we also
estimate ‘placebo’ discontinuities by artificially shifting all borders
8 km away from their real location. We find no discontinuity at
these ‘placebo’ borders, reflecting the fact that only actual political
borders matter17.
The prominent role of agriculture
Here, we analyse country characteristics that are associated with
the identified border discontinuities (Fig. 2). It should be noted
that these are descriptive statistical associations rather than
cleanly identified causal explanations, as many country charac-
teristics co-vary.
Our exploration of mechanisms is based on the soil erosion
that is explained entirely by land use on cropland. This setting
has the lowest measurement error and is thus most suitable to
explore mechamisms. On average, countries differ by two per-
centage points in how much of their maximum erosion potential
they ‘achieve’. The corresponding estimate in absolute erosion is
the 1.8 t ha−1 yr−1 mentioned previously. The two percentage points
are the benchmark magnitude to which we can compare the mag-
nitude of all potential explanations. We sort countries according
to each potential mechanism and estimate a separate regression
discontinuity for each mechanism (Methods). For example, we
sort countries with a higher agricultural gross domestic product
(GDP) share to the ‘left’ side of the border (as represented on a
graph) and their neighbours to the ‘right’, and then we determine
whether we find a significant discontinuity between these coun-
tries. In this way, we estimate that about half (that is, ~1 percent-
age point) of the average discontinuity is explainable by differences
in agricultural characteristics (for example, the agricultural GDP,
the agricultural employment shares and the agricultural intensity).
In contrast, countries’ general environmental policy performance,
their population density or their general GDP cannot explain why
there are discontinuities in the erosion rate at international bor-
ders. This does not imply that these factors are irrelevant for soil
erosion at large. Rather, it suggests that they might operate at other
levels. Other factors that could potentially explain the remaining
~50% of the country impact include specific land-use targeted
policies18,19 or cultural differences20.
A map of the global heterogeneity in country impacts
Figure 3 illustrates the global distribution of country impacts. For
this map, all estimated border discontinuities have been aggregated
by country.
Figure 3 indicates a considerable global potential to mitigate
soil erosion, especially in some large countries in Asia and Latin
America (such as Brazil and China). As the world’s soils are a
globally valuable resource that is increasingly degraded and many
countries have considerable potential to mitigate this problem,
this underlines the value of projects for international cooperation,
such as the Global Soil Partnership of the Food and Agriculture
Organization (FAO; http://www.fao.org/global-soil-partnership),
which might increase incentives and support for countries that
can make a difference.
Methods
Our main data source is the data of Borrelli et al.6, who use a combination of
remote-sensing and statistical data in the framework of the revised universal
soil loss equation (RUSLE) to globally model the magnitude and distribution of
soil erosion. The two main strengths of the dataset are that it has been produced
globally with a consistent method and that it has a high resolution. To model soil
erosion, individual determinants are first quantified and then brought together in
the following multiplicative equation:
Erosioni ¼ Ri ´ Ki ´ LSi ´ Ci ð1Þ

52 Nature Sustainability | VOL 3 | January 2020 | 51–55 | www.nature.com/natsustain

Nature Sustainability
Overall country effect
Agricultural GDP share
Agricultural intensity
Agricultural employment share
Primary production
GDP per capita
Population density
Environmental performance index
Environmental policy stringency
Environmental policy enforcement
Sustainable Development Goals
–2 –1 0 1 2
Percentage points
Fig. 2 | Agriculture is the best explanation. This coefficient plot is based on
individual regression discontinuity estimates, according to which countries
are sorted by a different characteristic in each specification. The left-side
(dependent) variable is the soil cover effect on cropland, which reflects
how much of the maximum potential soil erosion actually occurs in each
grid cell, as a percentage. Diamonds are point estimates, and bars show
the 90%, 95% and 99% confidence intervals. As a benchmark, we sort
countries according to their estimated individual coefficients. The average
discontinuity here is two percentage points. We can compare the effects
of all other country characteristics to this benchmark. Countries’ economic
reliance on agriculture (that is, the agricultural GDP share) explains half the
overall discontinuity (agriculture being more important = 1 percentage point
more soil erosion), and so do a country’s agricultural intensity as measured
by fertilizer use per hectare (more intensity = 1 percentage point less soil
erosion) and a country’s agricultural employment share (again, agriculture
being more important = 1 percentage point more soil erosion). A country’s
primary production (agriculture, forestry and fishery) explains almost as
much, but this is estimated to be less significant, and it is only statistically
significant at the 10% level. In contrast, differences in GDP per capita do
not produce a significant discontinuity and neither do various measures of
general environmental policy performance (including the environmental
performance index, the reported environmental policy stringency and
enforcement indices and the Sustainable Development Goals index).
where Erosioni is the estimated annual soil erosion per hectare (Mg ha−1 yr−1) of
pixel i, Ri is a parameter capturing rainfall erosivity (MJ mm h−1 ha−1 yr−1), Ki is a
parameter capturing soil erodibility (Mg h MJ−1 mm−1), LSi is a parameter capturing
the effect of topography (dimensionless) and Ci is a parameter capturing the effect
of land cover (percentage change of erosion rate compared with the maximum
possible from all the natural factors).
Since the erosion data is modelled, cross-validation is important
(Supplementary Information and Supplementary Fig. 1). Borrelli et al.6 show that
their global map is generally consistent with field measurements, but for highlands,
for example, modelled erosion is lower than measured erosion, and for cratons,
modelled erosion is at the higher end of the field measurements. In the case of the
highlands, this discrepancy could relate to the small number of field measurements
(n = 44), which are not representative. Either way, in the analysis, general
differences between borders are absorbed by including border fixed effects.
The rainfall erosivity factor is computed from pluviograph data of
3,625 weather stations that are part of the global rainfall erosivity database
(GloREDa)21. For processing, the US Department of Agriculture Rainfall Intensity
Summarization Tool (the RIST software) is used in a first step, and then a Gaussian
process regression is trained on data from the WorldClim database (http://www.
worldclim.org), including average monthly precipitation, average minimum and
maximum monthly precipitation, average monthly temperature, precipitation of
the wettest month, precipitation of the driest month and precipitation seasonality,
to predict the rainfall erosivity factor globally for each 1 km2 pixel.
Nature Sustainability | VOL 3 | January 2020 | 51–55 | www.nature.com/natsustain
Analysis
The soil erodibility factor is estimated based on intrinsic topsoil properties,
such as organic matter, texture, permeability and structure. This information is
provided by the International Soil Reference and Information Centre (ISRIC)
SoilGrids database22. For processing, the equations of Wischmeier and Smith23
are used.
The topography factor, based on slope length and steepness, is computed using
the approach of Desmet and Govers24 and data from the hole-filled Shuttle Radar
Topography Mission (SRTM) digital elevation model and Advanced Spaceborne
Thermal Emission and Reflection Radiometer (ASTER) Global Digital Elevation
Model (GDEM) data products25,26. The computational operations are carried out
using System for Automated Geoscientific Analyses (SAGA)27.
The main data sources for the soil cover factor are NASA’s Moderate Resolution
Imaging Spectroradiometer (MODIS) Land Cover Type product (MCD12Q1) and
the MOD44B Vegetation Continuous Fields (VCF) product, classified according to
the International Geosphere Biosphere Programme (IGBP) system28. The data were
obtained from NASA’s EarthData facility (https://reverb.echo.nasa.gov) and pre-
processed using the MODIS Reprojection Tool to obtain baseline land-cover maps.
These maps were subsequently improved with data on forests from Hansen et al.29
and data on croplands by Monfreda et al.30, Ramankutty et al.31, the FAOSTAT
database (http://www.fao.org/faostat/en/#data) of the FAO and the Land Processes
Distributed Active Archive Center32. To move from soil cover to its effect on soil
erosion, literature values are assigned to groups of land covers with similar erosion
effects (see Borrelli et al.6 for details and sources). Then, the erosion effects of
different land covers are weighted by the percentage shares of each land cover on
each pixel.
The final product is a globally gridded dataset of spatially explicit soil erosion
estimates. The nominal time period of our data is 2012, but it should be noted that
the input data come from many periods around this date, in a few cases going as far
back as the year 200030,31.
To produce a comparable map for the hypothetical natural rate of soil erosion,
we combine the rainfall erosivity (R), the soil erodibility (K) and the topography
(LS) factors from above but use a soil cover factor (C) that reflects how natural
vegetation in each world region would look without human impact. To compute
this, we use the data of Bastin et al.14, who trained a random forest machine-
learning algorithm on protected areas all round the world (areas which arguably
showcase the natural vegetation for each region in the world). The assumption
is that the random forest algorithm can learn in these protected areas how
climatic, edaphic and topographic variables relate to variation in the natural
vegetation (for example, the share of forests and grasslands). The advantage over,
for example, traditional regression-based techniques is that which explanatory
variables are used for prediction and how they are used is completely data driven,
and the outcomes usually far outperform specifications assembled by humans.
K-fold cross-validation has demonstrated the quality of the model: 71% of
the global variation in the natural vegetation observed in protected areas was
correctly predicted14.
From a large initial number of explanatory variables, ten cluster variables
are created and used. Five describing a grid cell’s climate, three its soil and two
its topography: annual mean temperature, annual precipitation, precipitation
seasonality, mean temperature of the wettest quarter, precipitation of the driest
quarter, depth to bedrock, organic carbon stock in the top layer, sand content in
the top layer, elevation and hill-shade. Using the optimal combination and form of
these ten explanatory factor variables, the trained random forest machine-learning
algorithm is able to accurately predict for each of our 1 km2 grid cells how its
natural vegetation would look in the absence of humans.
The next step is to integrate the predicted natural vegetation into our estimate
of the natural soil erosion rate. For this, a natural soil cover factor is created that
is a weighted average per pixel of the soil cover effect of grass- and shrubland
on the one hand and forest on the other. Using this new soil cover factor, we can
re-estimate equation (1) for the hypothetical, natural rate of soil erosion. Except for
the soil cover factor, all other factors remain as before, reflecting the effects of soil
characteristics, rainfall and topography. Since our estimates of the natural rate of
soil erosion and the actual rate of soil erosion have the same spatial resolution and
share most of their input data, they are comparable, and we can investigate how
differently they are distributed.
For our analysis, we use a spatial regression discontinuity design12,13,17,33,34.
Ideally we would be able to conduct a controlled experiment in which we would
randomly assign areas of land to different countries and compare how soil erosion
rates change. Without such randomization, it is a challenge to learn what role
countries as political entities play compared with other differences between
countries (for example, rainfall intensity). With the spatial regression discontinuity
design, we try to approximate this ideal. We focus exclusively on areas that are
close to international borders, assuming that observations just to the left and
right of a border are more comparable than observations that are farther apart34.
We make the (falsifiable) assumptions that potential confounding factors are
continuously distributed across the borders and that the only relevant discontinuity
right at the border is that the country changes. The main threat is borders that
have been endogenously placed—that is, borders where there are natural features
that affect soil erosion and that coincide with the country borders. For example, a
mountain range might naturally divide two countries, and the natural rate of soil
53
Analysis Nature Sustainability

Average erosion effect per country (t ha–1 yr–1)

>3 (+) 0 >3 (–)

Fig. 3 | A global map of countries’ soil erosion performance. The map is based on soil erosion discontinuities between each country and all of its
neighbours (unweighted). Darker green indicates that a country has a more positive impact on the global rate of soil erosion than its neighbours (that is, it
has a dampening effect), and darker red indicates that a country has a more negative impact on the global rate of soil erosion (that is, it has higher erosion
rates). For all countries that perform more poorly than their neighbours, the map shows how much soil erosion they could mitigate (per hectare and year)
just by catching up to their neighbours. Even the better performing countries could probably further improve their impact, so this map probably presents a
lower bound for countries’ erosion mitigation potential—especially in regions where soil erosion is high. For reference, the global average soil erosion rate
is 2.4 t ha−1 yr−1, and we estimate the average country discontinuity at 1.4 t ha−1 yr−1.

erosion could differ on each side of the border because rainfall, soils, topography or
natural vegetation is different.
Our first test is thus to estimate at which borders we find a discontinuity
in the hypothetical natural rate of soil erosion. We do so just like we test for
discontinuities in the actual rate of soil erosion. The model, from Calonico
et al.35, is a non-parametric local polynomial regression with robust confidence
intervals. The first step in the estimation process is the choice of the optimal
bandwidth—that is, the optimal maximum distance to the border. The closer
we restrict our data to the borders, the more we mitigate omitted variable
bias; but we also increase the variance of our estimates, because we reduce our
sample size36. On average, the optimal distance in our dataset is 17 km on
each side.
We then estimate the following for each country pair in our data:
Yi ¼ α þ β1 Di þ β2 distancei þ β3 distancei ´ Di þ εi ð2Þ
where Yi is the potential natural rate of soil erosion of pixel i and α is a constant.
Indicator Di takes the value 1 if a pixel i is located in adjacent country A and 0 if
it is located in adjacent country B in each pair. Our ‘running variable’ is distancei,
which is the distance from each pixel to the border, and this is fitted separately
on each side. Here, β1 indicates whether there would be a natural soil erosion
discontinuity at a given location of a current border, even if the countries and
their borders did not exist. This is how we classify borders as ‘political’ and
‘natural’. If we find a border discontinuity in the hypothetical natural soil erosion
rate, the border is ‘natural’, and if we find no discontinuity, the
border is ‘political’.
To estimate the actual soil erosion discontinuities at individual borders, we
substitute Yi in equation (2) for either the (actual) soil erosion rate in t ha−1 yr−1 or
the (actual) soil cover effect on the rate of erosion (the C factor), depending on the
specification. In the baseline specifications, we only control for border distance
on each side of the border. For robustness tests, we also include lower-order
polynomials (linear or quadratic) of longitude and latitude, and the natural rate of
soil erosion.
When estimating the global average border discontinuity, we need to decide
which country of each country pair to sort to the ‘left’ side of the border and
which to the ‘right’. We could, for example, sort all countries with a smaller
average soil erosion rate to the left and all with a larger rate to the right. This,
however, could result in a downward bias of the actual discontinuity if some
countries’ impact differences have the opposite sign from their difference in
average soil erosion. The most precise global aggregation of individual border
discontinuities is achieved by basing the sorting on estimated individual border
discontinuities. An important control variable is the border indicator θij, which
absorbs all idiosyncratic particularities of each border. Standard errors, εi, are also
clustered at the borders, for the same reason. When using all data at once, we use
a parametric regression (but allow the border distance to have separate effects
on each side of a border). As bandwidth, we use the average optimal bandwidth
estimated from all individual border specifications, which is 17 km on each border
side. However, we also estimate specifications using alternative bandwidths, up
to 70 km on each side, to probe the sensitivity of the estimates. Our baseline
specification is then
Yi ¼ γ þ δ1 Di þ δ2 distancei þ δ3 distancei ´ Di þ θij þ εi ð3Þ
where Yi continues to be either the soil erosion rate or the soil cover effect (the C
factor), depending on the specification. The term γ is a constant and δ1–3 are model
parameters to be estimated. What changes in comparison with the individual
border specification is that now Di equals 1 for each country that has a potentially
positive impact and 0 otherwise. As robustness checks, we again include low-order
polynomials (linear or quadratic) of longitude and latitude, and the natural rate of
soil erosion. The interpretation of β1 is the global average border discontinuity in
the rate of soil erosion.
To understand which country characteristics are associated with the estimated
border discontinuities, we re-sort the countries according to each. Specifically, we
define Di to equal 1 if the country has a higher value in an explanatory variable
(for example, a higher income, more stringent environmental policies, and so on)
and 0 otherwise. Because all else remains the same, we can directly compare the
discontinuity in specific country characteristics to the magnitude of the overall
discontinuity.
Data availability
Data can be retrieved from Wuepper et al.37 and from the corresponding author
upon reasonable request.
Code availability
Code and programs can be retrieved from Wuepper et al.37 and from the
corresponding author upon reasonable request.
Received: 22 March 2019; Accepted: 30 October 2019;
Published online: 2 December 2019
References
1. Foley, J. A. et al. Solutions for a cultivated planet. Nature 478, 337–342
(2011).
2. Mueller, N. D. et al. Closing yield gaps through nutrient and water
management. Nature 490, 254–257 (2012).
3. Amundson, R. et al. Soil and human security in the 21st century. Science 348,
1261071 (2015).
4. Kaiser, J. Wounding Earth’s fragile skin. Science 304, 1616–1618 (2004).
5. Montanarella, L. Agricultural policy: govern our soils. Nature 528, 32–33
(2015).
6. Borrelli, P. et al. An assessment of the global impact of 21st century land use
change on soil erosion. Nat. Commun. 8, 2013 (2017).

54 Nature Sustainability | VOL 3 | January 2020 | 51–55 | www.nature.com/natsustain

Nature Sustainability
7. Mirzabaev, A., Nkonya, E., Goedecke, J., Johnson, T. & Anderson, W. in
Economics of Land Degradation and Improvement – A Global Assessment for
Sustainable Development (eds Nkonya, E., Mirzabaev, A. & von Braun, J.)
167–195 (Springer, 2016).
8. Barbier, E. B. & Hochard, J. P. Does land degradation increase poverty in
developing countries? PLoS ONE 11, e0152973 (2016).
9. Nkonya, E. & Anderson, W. Exploiting provisions of land economic
productivity without degrading its natural capital. J. Arid Environ. 112,
33–43 (2015).
10. Lambin, E. F. et al. The causes of land-use and land-cover change: moving
beyond the myths. Glob. Environ. Change 11, 261–269 (2001).
11. García-Ruiz, J. M. et al. A meta-analysis of soil erosion rates across the world.
Geomorphology 239, 160–173 (2015).
12. Lee, D. S. & Lemieux, T. Regression discontinuity designs in economics.
J. Econ. Lit. 48, 281–355 (2010).
13. Imbens, G. & Lemieux, T. Regression discontinuity designs: a guide to
practice. J. Econom. 142, 615–635 (2008).
14. Bastin, J.-F. et al. The global tree restoration potential. Science 365,
76–79 (2019).
15. NASA/Goddard Space Flight Center Scientific Visualization Studio Haitian
Deforestation. Scientific Visualization Studio https://svs.gsfc.nasa.gov/2640
(2019).
16. Imbens, G. & Kalyanaraman, K. Optimal bandwidth choice for the regression
discontinuity estimator. Rev. Econ. Stud. 79, 933–959 (2012).
17. Cattaneo, M. D., Idrobo, N. & Titiunik, R. A Practical Introduction to
Regression Discontinuity Designs (Cambridge Univ. Press, 2018).
18. Panagos, P., Borrelli, P. & Robinson, D. A. Common agricultural policy:
tackling soil loss across Europe. Nature 526, 195 (2015).
19. Deng, L., Shangguan, Z.-P. & Li, R. Effects of the grain-for-green program on
soil erosion in china. Int. J. Sediment Res. 27, 120–127 (2012).
20. Wuepper, D. Does culture affect soil erosion? Empirical evidence from
Europe. Eur. Rev. Agric. Econ. https://doi.org/10.1093/erae/jbz029 (2019).
21. Panagos, P. et al. Global rainfall erosivity assessment based on high-temporal
resolution rainfall records. Sci. Rep. 7, 4175 (2017).
22. Hengl, T. et al. SoilGrids1km — global soil information based on automated
mapping. PLoS ONE 9, e105992 (2014).
23. Wischmeier, W. H. & Smith, D. D. Predicting Rainfall Erosion Losses: A Guide
to Conservation Planning (US Department of Agriculture, 1978).
24. Desmet, P. & Govers, G. A GIS procedure for automatically calculating the
USLE LS factor on topographically complex landscape units. J. Soil Water
Conserv. 51, 427–433 (1996).
25. Robinson, N., Regetz, J. & Guralnick, R. P. EarthEnv-DEM90: a nearly-global,
void-free, multi-scale smoothed, 90m digital elevation model from fused
ASTER and SRTM data. ISPRS J. Photogramm. Remote Sens. 87,
57–67 (2014).
26. Reuter, H. I., Nelson, A. & Jarvis, A. An evaluation of void-filling
interpolation methods for SRTM data. Int. J. Geogr. Inf. Sci. 21,
983–1008 (2007).
Nature Sustainability | VOL 3 | January 2020 | 51–55 | www.nature.com/natsustain
Analysis
27. Brenning, A., Bangs, D., Becker, M., Schratz, P. & Polakowski, F. Package
‘RSAGA’. The Comprehensive R Archive Network https://CRAN.R-project.org/
package=RSAGA (2018).
28. Loveland, T. R. & Belward, A. The IGBP-DIS global 1km land cover data set,
DISCover: first results. Int. J. Remote Sens. 18, 3289–3295 (1997).
29. Hansen, M. C. et al. High-resolution global maps of 21st-century forest cover
change. Science 342, 850–853 (2013).
30. Monfreda, C., Ramankutty, N. & Foley, J. A. Farming the planet:
2. Geographic distribution of crop areas, yields, physiological types,
and net primary production in the year 2000. Glob. Biogeochem. Cycles 22,
GB1022 (2008).
31. Ramankutty, N., Evan, A. T., Monfreda, C. & Foley, J. A. Farming the planet:
1. Geographic distribution of global agricultural lands in the year 2000. Glob.
Biogeochem. Cycles 22, GB1003 (2008).
32. Land Processes Distributed Active Archive Center (US Geological Survey
and National Aeronautics and Space Administration, 2019);
https://lpdaac.usgs.gov/
33. Cattaneo, M. & Escanciano, J. Regression Discontinuity Designs: Theory and
Applications (Emerald Group Publishing, 2017).
34. Keele, L. J. & Titiunik, R. Geographic boundaries as regression
discontinuities. Polit. Anal. 23, 127–155 (2014).
35. Calonico, S., Cattaneo, M. D. & Titiunik, R. Robust nonparametric
confidence intervals for regression-discontinuity designs. Econometrica 82,
2295–2326 (2014).
36. Cattaneo, M. D. & Vazquez-Bare, G. The choice of neighborhood in
regression discontinuity designs. Obs. Stud. 2, 134–136 (2016).
37. Wuepper, D., Borrelli, P. & Finger, R. Dataset: countries and the global rate of
soil erosion. ETH Zurich Research Collection http://hdl.handle.
net/20.500.11850/361244 (2019).
Author contributions
D.W. contributed mainly to analysis and writing, P.B. contributed mainly to data
preparation and R.F. contributed mainly to writing.
Competing interests
The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/
s41893-019-0438-4.
Correspondence and requests for materials should be addressed to D.W.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2019
55
 Analysis
SUPPLEMENTARY INFORMATION
https://doi.org/10.1038/s41893-019-0438-4

In the format provided by the authors and unedited.

Countries and the global rate of soil erosion

David Wuepper 1
*, Pasquale Borrelli 2
and Robert Finger1

1
Agricultural Economics and Policy Group, ETH Zurich, Zurich, Switzerland. 2Environmental Geosciences, University of Basel, Basel, Switzerland.
*e-mail: dwuepper@ethz.ch

Nature Sustainability | www.nature.com/natsustain

 Supplementary Materials
Countries and the global rate of soil erosion

Content:

Page 2: A Comparison of measured and modelled erosion rates (Supplementary Section 1)

A visual illustration of the cross validation results from Borrelli, et al. 1

Page 3: Comparing Border Areas to the Rest of the Countries (Supplementary Section 2)

We establish that soil erosion in border areas is highly correlated with countries’ average soil erosion

Page 4: Global Estimates of Countries’ Impact (Supplementary Section 3)

This is a coefficient plot showing our main estimates and robustness checks

Page 5. The US-Mexico Border (Supplementary Section 4)

A visual illustration of our analysis on cropland from a specific border

Page 6: Global Borders Show a Continuity in Natural Erosion and a Discontinuity in Actual Erosion

(Supplementary Section 5)

A visual illustration of our global analysis of the whole sample

Page 7: A Graphical Illustration of Testing for Mechanisms (Supplementary Section 6)

Some visual examples for how we investigate the mechanisms behind the country impact

Page 8: A List of Potential Explanations for a Country Impact (Supplementary Section 7)

A table with variables and sources for the testing of mechanisms

Page 10: Global Distribution of Erosion Within and Between Countries (Supplementary Section 8)

Here we quantify the inequality in soil erosion within and between countries using Gini indices

1
 Supplementary Section 1

Among the extensive tests and cross-validation exercises of Borrelli, et al. 1 is a comparison of the
RUSLE estimates with field measurements (Fig. 6 in the original article). There is generally a high
degree of consistency, but e.g. for highlands, model estimates are lower than what has been measured
in the field. Erosion modelling in highlands tends to be more uncertain than elsewhere, but an alternative
explanation for the deviation seen here is that there are only 44 field measurements available, and these
are unlikely to be representative, as they are not randomly sampled (see supplementary materials of
Borrelli, et al. 1).

Supplementary Figure 1. A Comparison of measured and modelled erosion rates. The figure shows soil
erosion rates measured on agricultural fields under conventional agriculture (n = 779), geologic erosion rates
measured on alpine terrain (n= 44), soil-mantled landscapes (n = 1456), low gradient continental cratons (n = 218),
grassland and scrublands (n = 63), native forests (n = 46) and averages of RUSLE predictions indicated by an
asterisk. Reproduced from ref. 1

2
 Supplementary Section 2

In the regression discontinuity design, we only use data-points that are within a rather narrow bandwidth
around international borders. This reduces unobserved heterogeneity but motivates an investigation into
how much the rate of soil erosion differs between areas close to international borders and areas further
inland. Supplementary Figure 2 shows that there is a strong correlation between countries’ average
soil erosion rate and their average soil erosion rate close to international borders. This suggests that we
can interpret our findings as findings for countries, not just border areas.

Supplementary Figure 2. Comparing Border Areas to the Rest of the Country. The soil erosion rate in border
areas is strongly correlated with country’s overall average soil erosion rate. Thus, our analysis that focuses
exclusively on border areas is plausibly representative for the countries at large.

3
 Supplementary Section 3

The coefficient plot of Supplementary Figure 3 below graphs the estimated global border
discontinuities in the rate of soil erosion. The estimates on all types of land vary between 1.1 and 1.9
tons per hectare and year. A placebo exercise shows that these discontinuities can only be found at really
existing borders. Not shown for brevity, the estimate on cropland alone is slightly larger, at 1.8 tons per
hectare and year.

Supplementary Figure 3. Estimating the Global Average Discontinuity. We sort countries according to their
individually estimated discontinuities, to estimate a global average. We include border fixed effects and border
distance on each side of the border. The bandwidth is 17 km, standard errors are clustered by border. The global
average country effect is 1.4 t/ha/yr. To test the robustness of our estimate, we also estimated separate
specifications for “natural” and “political” borders, widened the bandwith to 70 km border distance on each side,
included a quadratic polynomial of longitude and latitude, controled for the natural erosion rate, and estimated a
placebo estimation for non-existent borders that we placed 8 km parallel to the real borders. For the real borders,
we consistently estimate a discotninuity between 1.1 and 1.9 t/ha/, whereas we estimate no discotninuity at the
placebo borders.

4
 Supplementary Section 4

In addition to our first example in the main text where we show the soil erosion discontinuity between
Haiti and the Dominican Republic (Fig.1), we like to present here a second example (Supplementary
Figure 4), which illustrates an alternative way to estimate countries’ impact on soil erosion. The
example is the border between Mexico and the US and here, we exclusively focus on cropland. The
maximum erosion potential1 in the border area of Mexico and the United States changes continuously
across the border (i.e. the rate of soil erosion that would occur without any soil cover, based on the
effects of rainfall, soils, and topography)(a). In contrast, the share of the maximum erosion potential
that is actually realized (in %)1 shows a clear discontinuity right at the border (b). When we look at the
hypothetical natural vegetation2 in the area, there is no difference between the side of the US and the
side of Mexico (c). However, this is currently largely irrelevant, because the area is now predominantly
used for agriculture (on both sides). When we look at the area from space3, we see that actual soil cover
is quite dense on the side of the US (vegetation is colored red) and less so on the side of Mexico (d).

Supplementary Figure 4. The Border between Mexico and the United States. The border is purely political in
the sense that there is no discontinuity in the natural soil erosion potential 1 (based on rainfall, topography and soil
characteristics combined with no soil cover, which is the configuration that leads to the maximum soil erosion
rate)(a). How much of this continuously distributed soil erosion potential actually occurs 1 changes discontinuously
at the border (b). Importantly, this is not because naturally the vegetation potential2 is distinct in the two countries
(c). The reason is that the fields in the US are more intensively farmed, leading to a dense soil cover, whereas most
fields on the Mexican side are less intensively farmed, leading to a barer soil cover3 (d). Image credit:
NASA/GSFC/METI/ERSDAC/JAROS, and U.S./Japan ASTER Science Team.

5
 Supplementary Section 5

To visually illustrate the global magnitude of our estimated difference between the country effect and
the natural rate of soil erosion, consider Supplementary Figure 5. Naturally, soil cover would mitigate
almost all soil erosion (~99%), continuously distributed in space (a). In contrast, the actual soil cover
mitigates significantly less soil erosion (~93-94%), and this differs discontinuously by country (b).

Supplementary Figure 5. Globally, The Natural Land Cover Effect on Soil Erosion is Continuously
Distributed around International Borders but the Actual Land Cover Effect shows a Discontinuity. It is
apparent that naturally, there is no difference in the natural land cover effect, which is large. Without human
impact, countries would have erosion rates well below 1% of their natural potential, because an established
vegetation would effectively protect the soil (a). In reality, however, there are important land-use differences
between countries and this leads to large erosion differences (b).

6
 Supplementary Section 6

An approach to test for mechanisms is to sort countries according to potential mechanisms. This can be
graphically illustrated, as seen in Supplementary Figure 6. When we sort countries with a higher
agricultural GDP share to the left and countries with a lower agricultural GDP share to the right, we see
a clear border discontinuity (a). The magnitude of the discontinuity looks very similar when we sort
countries with more intensive agriculture to the left and more extensive agriculture to the right (b). In
contrast, both for population density (c), and for the environmental performance index (d), there is no
obvious discontinuity at the border.

Supplementary Figure 6. Graphical illustrations for the Explanatory Power of Different Mechanisms. We
find that between country differences in agriculture explain about 50% of the country impact on soil erosion,
whereas many other candidate mechanisms seem to operate at other (often more local) levels. As an example, there
is a clear discontinuity discernable when we sort countries that rely economically more on agriculture to the left,
and countries that rely economically less on agriculture to the right (a). As another example, the graph looks very
similar, but basically reversed when we sort countries that farm more intensively to the left and those that farm
less intensively to the right (b). We do not see clear discontinuities when we sort countries e.g. by population
density or environmental performance index (c and d, respectively).

7
 Supplementary Section 7
This table lists our explanatory variables and their sources
Supplementary Table 1. List of Potential Explanations for a Country Effect
Variable Description
gdp Gross domestic product per
capita
education Years of schooling
long-term Opposite to short term
orientation orientation. Societies with a
long-term orientation value e.g.
thrift and education.
savings rate gross disposable income less
final consumption expenditure
ethnic fract Probability that two randomly
sampled individuals in a country
are from different ethnicities
interpersonal Percentage of the population
trust that states others can generally
be trusted
institutions World Governance Indicator,
based on several hundred
individual variables measuring
perceptions of governance, such
as control of corruption and
political stability
government Strength of constraint on
constraints government powers
state fragility The fragility of a state is related
to its government’s capacity to
manage conflict and implements
its policies
bayesian Bayesian Corruption index, with
corruption index a higher score reflecting that a
country is perceived to be less
affected by corruption
environmental Multi-dimensional
performance Environmental Performance
index (epi) score Index. Indicators include
environmental risk exposure, air
quality and pollution, sanitation,
drinking water quality, nitrogen
balance and use efficiency,
change in forest cover, species
protection, trends in carbon
intensity, and several others
Source Prior Results / Hypotheses
data.worldbank.org Mirzabaev, et al. 4 find no
relationship with land
degradation but Borrelli, et al.
1
observe a positive
relationship with soil erosion.
data.worldbank.org Aldy, et al. 5 reports a positive
relationship between
education and environmental
preferences
worldvaluessurvey.or Wuepper 6 reports a positive
g relationship between long-
term orientation on soil
erosion mitigation
data.worldbank.org Possibly an alternative
indicator for a long-term
orientation
anderson.ucla.edu/fac Alesina, et al. 7 find a negative
ulty_pages/romain.w relationship between ethnic
acziarg fractionalization and forest
conservation
worldvaluessurvey.or Dietz, et al. 8 emphasize the
g positive effect of inter-
personal trust for the
management of natural
resources
info.worldbank.org/g Mirzabaev, et al. 4 find a
overnance/wgi positive relationship with land
degradation, whereas Nkonya
and Anderson 9 find a negative
one.
worldjusticeproject.or Alternative measure for the
g/our-work/research- role of institutions
and-data
systemicpeace.org/ins Alternative measure for the
crdata role of institutions
users.ugent.be/~sasta Alternative measure for the
nda/BCI/BCI.html role of institutions
epi.envirocenter.yale. One would expect that
edu countries with a better general
environmental performance
also cause less soil erosion,
but there is no prior research
on this
8
sustainable Sustainable Development Goals sdgindex.org Similar to above
development Index. Can be interpreted
goals index roughly as progress of a country
towards achieving the SDGs, in
percentage
Stringency of Business learders’ perception of Browne, et al. 10 Similar to above
Environmental environmental policy stringency
Policies
Enforcement of Business learders’ perception of Browne, et al. 10 Similar to above
Environmental environmental policy
Policies enforcement
infrastructure The quality of a country’s roads reports.weforum.org For infrastructure and market
access both positive and
negative relationships with
land degradation have been
found, see literature
discussions and own findings
in Mirzabaev, et al. 4 and
Nkonya and Anderson 9
population Population per unit area sedac.ciesin.columbia Same as for infrastructure. See
density .edu/ also Lambin, et al. 11
agric intensity Fertilizer use in kg per hectare fao.org/faostat/en/#da Positive relationship with land
ta degradation in Sub-Africa and
America and a negative one in
Europe and Asia9
agric va In percentage, value added in data.worldbank.org Hypothesized to be associated
the agricultural sector with an erosion reduction
because it increases the value
of the soil
agric the difference in value added Gollin, et al. 12 Hypothesized to be associated
productivity gap between the agricultural and the with an increase in erosion
non-agricultural sector because it indicates
management issues and lack
of resources
agric empl share Percentage of the workforce of a fao.org/faostat/en/#da Relationship not clear.
country that works in ta Possibly associated with more
agriculture soil erosion if mostly indicator
for low agricultural
productivity
forest harvest Amount of wood harvested in fao.org/faostat/en/#da Hypothesized to be associated
m3 ta with an increase in soil
erosion because it reduces soil
cover

9
 Supplementary Section 8

To understand the global distribution of soil erosion within and between countries, we conduct analyses
of variance (ANOVA) and compute Gini indices13, the latter using the program of Zeileis and Kleiber
14
. In the analyses of variance, we investigate the soil cover effect and the overall soil erosion, and only
include country fixed effects as explanation. We estimate that countries explain ~28% of the global
variance in land cover related soil erosion (the C-factor) and ~7% of overall soil erosion (in t/ha/yr). On
croplands, these numbers are considerably higher, because the human influence is more direct. Turning
to our measures of soil erosion inequality, we use Gini indices, which are commonly used to measure
income inequality13 but can just as well be used to understand the distribution of soil erosion. The Gini
index ranges from zero (most equal distribution) to one (most unequal distribution). Part a of Supplementary
Figure 7 shows the erosion Gini per country. The average is 0.7. part b shows the average erosion rate
per country. When computing the erosion Gini based on the country averages, it is 0.6. Thus, the
distribution of soil erosion follows globally a roughly similar pattern as the distribution within countries.

Supplementary Figure 7. Distribution of Soil Erosion Within and Between Countries. Part a shows the Gini
index for the soil erosion distribution within countries. The index goes from zero (most equal distribution) to one
(most unequal distribution). Within countries, the Gini has a mean of 0.7. part b of shows countries’ average rate
of soil erosion. Between countries the Gini has a mean of 0.6. This suggests that the distribution of soil erosion is
similar within and between countries, with a little bit more inequality within than between countries.

10
 Supplementary References

1 Borrelli, P. et al. An assessment of the global impact of 21st century land use change on soil
erosion. Nature Communications 8, 2013 (2017).
2 Bastin, J.-F. et al. The global tree restoration potential. Science 365, 76-79 (2019).
3 NASA. NASA/GSFC/METI/ERSDAC/JAROS, and U.S./Japan ASTER Science Team: ASTER
Mexicali, <https://www.jpl.nasa.gov/spaceimages/details.php?id=PIA02659> (2019).
4 Mirzabaev, A., Nkonya, E., Goedecke, J., Johnson, T. & Anderson, W. in Economics of Land
Degradation and Improvement–A Global Assessment for Sustainable Development 167-195
(Springer, Cham, 2016).
5 Aldy, J. E., Kotchen, M. J. & Leiserowitz, A. A. Willingness to pay and political support for a
US national clean energy standard. Nature Climate Change 2, 596 (2012).
6 Wuepper, D. Does Culture Affect Soil Erosion? Empirical Evidence from Europe. European
Review of Agircultural Economics forthcoming (2019).
7 Alesina, A., Gennaioli, C. & Lovo, S. Public goods and ethnic diversity: Evidence from
deforestation in Indonesia. Economica 86, 32-66 (2019).
8 Dietz, T. et al. The impact of climate change on Drylands, with a focus on West Africa. Vol.
null (2004).
9 Nkonya, E. & Anderson, W. Exploiting provisions of land economic productivity without
degrading its natural capital. Journal of Arid Environments 112, 33-43 (2015).
10 Browne, C., Di Battista, A., Geiger, T. & Gutknecht, T. The executive opinion survey: The
voice of the business community. The Global Competitiveness Report 2014–2015, 69-78
(2014).
11 Lambin, E. F. et al. The causes of land-use and land-cover change: moving beyond the myths.
Global environmental change 11, 261-269 (2001).
12 Gollin, D., Lagakos, D. & Waugh, M. E. The Agricultural Productivity Gap. The Quarterly
Journal of Economics 129, 939-993, doi:10.1093/qje/qjt056 (2014).
13 Cowell, F. A. in Handbook of Income Distribution Vol. 1 87-166 (Elsevier, 2000).
14 Zeileis, A. & Kleiber, C. R Package ‘ineq’: Measuring Inequality, Concentration, and Poverty
V 0.2-13, <https://cran.r-project.org/web/packages/ineq/index.html> (2019).

11

View publication stats