Does the Environment Matter for Poverty Reduction? The Role of Soil Fertility and Vegetation Vigor in Poverty Reduction

The debate on the environment-poverty nexus is inconclusive, with past research unable to identify the causal dynamics. This paper uses a unique global panel data set that links (survey and census derived) poverty data to measures of environmental quality at the subnational level. The analysis uses vegetation vigor as a proxy for above-ground environmental quality and soil fertility as proxy for below-ground environmental quality. Rainfall is used to account for endogeneity issues in an instrumental variable approach. This is the first global study using quasi-experimental methods to uncover to what degree environmental quality matters for poverty reduction. The paper draws three main conclusions. (1) The environment matters for poverty reduction. The panel regression suggests that a 10 percent increase in vegetation vigor is associated with a poverty headcount ratio reduction of nearly 0.7 percentage point in rural areas, and 1 percentage point in Sub-Saharan Africa. A 10 percent increase in soil quality leads to a roughly 2 percentage point decrease in poverty rates in rural areas and in Sub-Saharan Africa. (2) The effects of environmental quality on poverty are stronger than its effects on average income, suggesting that the poor benefit disproportionately from environmental quality. (3) In situ environmental quality improvements are pro-poor, in contrast to urbanization. Although urbanization has highly significant and sizable correlations with GDP per capita, it is not significantly correlated with poverty reduction.


Introduction
The world had 1 billion fewer people living in poverty in compared to 1990World Bank, 2016). While poverty remains unacceptably high (using a poverty line of 1.90$/day, 10.7 percent of the world population are still poor 1 ) and challenges remain, these aggregate numbers suggest that significant progress has been made in the past decades. What factors are responsible for the observed decline in poverty headcounts? Human capital formation, economic growth, trade, and institutional strengthening have been suggested as important drivers. Much ink has been spilled on these human development and macro-economic drivers of poverty reduction (see e.g. Gennaioli et al, 2015). Less emphasis has been placed on the role of the environment in poverty reduction.
Most of the world's poor live in rural areas and depend heavily on agricultural or environmental income to make a living (Barbier and Hochard, 2016;Olinto et al., 2013). This includes timber and non-timber forest products; both formally and informally (for an overview, see e.g. Angelsen and Wunder, 2003). Environmental quality is important for both agricultural and environmental income, as healthier ecosystems increase agricultural and environmental yields.
Early empirical studies have identified declining soil fertility and poverty to be related at an aggregate level (Krishna et al., 2006;Barrett & Swallow, 2006). More recently, Barrett and Bevis (2016) find that national GDP per capita is positively correlated with soil nutrient balances in 36 Sub-Saharan African countries for which data are available. Barbier & Hochard (2016) find that around a quarter of people living in low-income countries reside on severely degraded land and that a lower share of people on degraded lands is associated with higher economic growth as well as lower poverty. While prior studies all point towards positive relationships, the evidence cannot be interpreted as causal.
The theoretical channels behind the suggested positive relationship are rather intuitive. The most important productive asset determining land productivity for the rural poor is soil, which grows the crops and delivers the ecosystem services that allow for either selling in the market or consumption. There is extensive literature on the relationship between agricultural productivity and soil fertility. For instance, the water storage capability of soil is an important determinant of plant growth (Wong and Asseng, 2006). Louwagie et al. (2009) find that shallow soils, stoniness or chemical issues such as salinity or acidity are negatively correlated with crop yields. There is also a third channel: The topographical conditions of the soil (elevation, steepness, etc.) affect soil erosion and accessibility by humans and machinery (e.g. Duran Zuazo, 2008). For an overview, see Mueller et al. (2010). Thus, we conclude that locations with good soils are likely to have high agricultural potential and thus have absolute advantage in producing high-value perishable vegetables and other crops.
On the other hand, poor soils may have negative implications for poverty reduction through three mechanisms (Barrett and Bevis, 2016): First, poor and degraded soils have negative effects on agricultural and environmental income. Reduced income on a micro level has not only contemporaneous negative effects on income and poverty on the macro level, but also makes it harder for a household to access nutrients to boost soil productivity in the future. Such links can be self-reinforcing: poor soil constrains capital accumulation and low capital accumulation inhibits investments in improving soils (Eswaran et al., 1997;. 1 Calculated using the World Bank's WDI data set for the year 2016. 4 correlative. Suich et al. (2015) review the evidence of research into the relationship between ecosystem services and poverty alleviation and conclude that "a considerable gap remains in understanding the links between ecosystem services and poverty, how change occurs, and how pathways out of poverty may be achieved based on the sustainable utilisation of ecosystem services." The main contribution of this paper is to provide causal estimates of the impact of environmental quality on poverty reduction. In addition, insights into the importance of environment relative to other factors such as infrastructure development or urbanization will be discussed. In particular, several contributions to existing literature are offered: (a) We use a global subnational data set, (b) we use monetary poverty rates that emerge from survey and census estimations rather than highly modeled measures that may induce multicollinearity with the explanatory variables, (c) we use not one, but three different measures of environmental quality, one for soil fertility and two for vegetation vigor, and (d) we do not draw evidence from cross-sectional evidence only (which likely suffers from omitted variable bias), but also exploit variance over time by implementing a panel fixed effects model. With these methodological refinements, we obtain results that underline the importance of soil and vegetation quality in poverty reduction.

Empirical Strategy a. Data
We create a unique geospatial data set linking environment and natural resource measures to poverty and other human development indicators at the subnational level. The geographical unit of analysis of the data set is the administrative unit 1 level, commonly referred to as "province" level, because this is the finest-grained level at which the poverty data set is available. 3 This rich data set allows us to specify a range of different models to examine the relationship of environmental quality (and degradation) and poverty.

Poverty and GDP per capita (outcome measures)
The measurement of poverty we use is the headcount ratio of people falling below $1.90 per day. Even though this is a narrow definition of poverty, 4 it is the official international poverty line and allows us to draw from country poverty maps that the World Bank has helped produce over the last 30 years. We "stitch together" a global map of sub-national poverty measures and pool several years of observations for several countries, creating an unbalanced poverty panel (see table 6 for the detailed descriptive statistics on the poverty observation frequency by country). For the panel specifications, we use the countries in table 6 that have repeated poverty observations. For the cross-sectional specification, we use the latest available observation per country.

5
Gross Domestic Product per capita 5 is our second outcome indicator. We compute this variable using GDP data from Gennaioli et al. (2015) and population data from the Gridded Population of the World (GPW) data set (CIESIN, 2016). 6

Environmental quality -soil fertility and vegetation vigor (explanatory measures)
We use three different measurements to proxy environmental quality. Two measure vegetation vigor, and one measures soil fertility: a) Net Primary Productivity (NPP), as our first measure of vegetation vigor, and b) the Normalized Difference Vegetation Index (NDVI), as our second measure of vegetation vigor for a robustness specification. c) topsoil carbon content (content of carbon in the 30cm of soil layer from the top), as a measure of soil fertility NPP is the rate at which an ecosystem accumulates biomass. It measures how much carbon dioxide plants take in during photosynthesis minus how much carbon dioxide is released during respiration. Hence, it is an indicator for how much of the absorbed carbon becomes part of leaves, roots, stalks or tree trunks. NPP data are captured via the Moderate Resolution Imaging Spectroradiometer (MODIS) aboard NASA's Terra and Aqua satellites. The data are provided in grid format. We use the average value per province in our computations.
NDVI is a satellite image-based measurement of environmental quality. It measures the "greenness", i.e., the relative density and health of vegetation of the planet's surface. This index shows where and how much green leaf vegetation was growing during a certain period. It uses red and near infrared light spectrum data captured by NASA's Terra satellite to compute a monthly indicator of vegetation quality for about 10km² cells. Again, we use the province averages. 7 Generally, it has been found that NPP is a better measure of biomass productivity and biodiversity (see for instance Xu et al., 2012;Phillips et al., 2008) which is why we use NPP for our main specifications and NDVI data in a robustness check.
Soil carbon content is an important measure of plant productivity. It is the percentage of carbon contained in the organic matter in the soil. It is the result of, e.g., decomposing plant and animal residues, root exudates or both living and dead microorganisms. Topsoil carbon content refers to the organic carbon measured in the top 30cm of the soil. It is a major determinant of plant growth and agricultural productivity (see for instance Lal, 2004 These indicators of environmental quality are used together with a set of common control variables to model two dependent variables, namely poverty and GDP per capita.

Instrumental variable (measure that delivers exogenous variation)
Mean average annual rainfall is employed as the instrument for annual changes in vegetation vigor (NPP & NDVI) and topsoil carbon. The data are sourced from the Climatic Research Unit in the National Center of Atmospheric Research (NCAR, 2017). The data set contains geographically gridded multiple weather time series from 1901 onwards. We use averages per year and province. A detailed discussion of the validity of using rainfall as an instrument is provided in the research design section.

Control variables
We use a variety of highly relevant control variables to isolate the effect of soil fertility and vegetation vigor on our dependent variables. This is particularly important in the cross-sectional specifications, as they do not have the luxury of province fixed effects. To capture the effect of different terrains, a topographic ruggedness index is employed. This index (see Nunn & Puga, 2012) is given by the square root of the sum of the squared differences in elevation between a given point on the earth surface and the eight points that lie 30 arc-seconds away in the eight major directions of the compass (North, North-East, East, etc.). Hence, the ruggedness index captures small-scale terrain irregularities based on elevation differences.
We also include land use categories (i.e. cropland, forest land, grass land, urban land or other) in the regression. Each LU indicator is measured as a share of the total geographic area. The original data are provided as global raster files with a spatial resolution of 300m by the Land Cover (LC) project of the Climate Change Initiative (CCI) led by the European Space Agency (ESA).
Soil type is a categorical variable that is used as a control variable in the cross-sectional specifications. There are 12 different categories of soil types: Alfisol, Andisol, Ardisol, Emtisol, Gelisol, Histosol, Inceptisol, Mollisol, Oxisol, Spodosol, Ulisol, and Vertisol. Additionally, where there is no soil, the area may be classified as either rock or sand. These soil categories are the result of millennia old geological and ecological processes and are a crucial determinant of soil quality. While a province may have several soil types, we assign the most prevalent soil type to each province. This classification using the aforementioned 12 + 2 soil types follows the soil classification system of the USDA system of soil taxonomy.
We include road density (mean kilometers of roads per province) and population as standard control variables. In contrast to similar works, we do not rely on population data that are modeled using land use or night light data (e.g. Amaral et al., 2005). Such data sets might raise severe endogeneity issues when acting as an explanatory variable for environmental quality. On our level of aggregation, we can rely on census and survey estimates, which greatly minimizes this risk of endogeneity. A detailed description of all variables, including data sources, can be found in table 1. For summary statistics, please refer to table 5 in the appendix. 7

b. Research design
Previous attempts to estimate the effect of environmental quality on poverty suffer from both omitted variables bias and simultaneity (as discussed above). Changes in environmental quality such as degradation of soil or degradation of vegetation vigor (through e.g. application of fertilizers) are endogenous explanatory variables, meaning that OLS would produce biased and inconsistent effect estimates and the coefficients would be correlated with the error term. To overcome these methodological challenges, we use simultaneous equation models with instrumental variables (see e.g. Greene, 2003or Wooldridge, 2010. For vegetation vigor, we specify a panel regression as we can exploit time series of NPP, NDVI, and rainfall. This allows us to create an unbalanced panel 8 as described in specification (1) below. In the case of below ground environmental quality, our measure of soil stored carbon is unfortunately time-invariant. Therefore, we specify a cross-sectional regression. Similarly, we have variation over time regarding the poverty headcount ratio, but we can only make use of our proxy of regional economic activity (GDP) measured at one point in time. To account for omitted variables that may drive the relationship between soil and poverty, we include a battery of control variables.
Starting with the panel specification, the reduced-form empirical models estimated are: 8

Panel specification of the vegetation vigor determinants of poverty
The panel is specified as Pov i,t α β 1 vegetation i,t β 2 LU i,t β 3 pop i,t μ c θ t μ c * θ t ϵ i,t 1 where i denotes province, c denotes country, and t denotes year. The dependent variable pov measures the poverty headcount rate. The explanatory variables are vegetation quality (NPP or NDVI), a vector of the five land use categories, population, country fixed effects ( ), year fixed effects ( ) and countryspecific time trends (μ c * θ t , as controlling for unobserved heterogeneity across countries and time is crucial. We use province-clustered standard errors, which account for within-country clustering of errors. NPP, NDVI, and population enter the model as log-transformed measures. We are interested in the intensive margin of vegetation 9 and control for the extensive margin. An increase in vegetation vigor of a province may be driven by territorially expanding vegetation, or by intensifying the vegetation vigor within bounded areas. By controlling for land use change trends, we attempt to control for changing patterns in land uses (the extensive margin) and isolate the effect of intensifying the plant productivity in each land use (the intensive margin).
Model (1) identifies the potential effect of the environment exploiting variation over time. However, unlike NPP and NDVI data, subnational GDP data as well as soil fertility data are time-invariant, meaning that the effects of soil fertility on poverty and GDP can only be assessed using spatial variation. The same holds true for the effects of vegetation vigor on GDP.

Cross-sectional specifications (soil fertility as determinant of poverty & GDP)
2 3 where GDP measures GDP per capita and soil refers to the measure of soil quality. LU is a vector of shares of the five land use categories. pop refers to population, rug refers to ruggedness and road refers to the road density of the province under consideration. are country fixed effects and are year dummies, capturing that our poverty observations are from different years. 10 refers to the error term. GDP per capita, population, ruggedness, road density and precipitation enter the model as log-transformed measures. Our cross-sectional specifications exploit within-country inter-province variation in 9 Changes in environmental quality, as proxied by changes in NPP or NDVI, can have many origins: they can be caused by climatic, ecological, geochemical, and human influences on the biosphere. In other words, the vegetation changes can be driven by bad land management, overexploitation of renewable resources such as forests, the destruction of biodiversity, droughts, and climate change (Nemani et al., 2003;Ainsworth et al., 2012). Here, we evaluate the effects of changes in environmental quality on the pace of poverty reduction, but without disentangling the origins of these changes, which may have come from any of these sources and are a subject for future scrutiny. Environmental changes driven by human behavior are the reason for endogeneity concerns, which we seek to overcome using an IV approach. 9 environmental quality in estimating potential effects on poverty rates (model 2) and on GDP per capita (model 3).
We use a suite of highly relevant and geospatially heterogenous control variables to reduce the problem of omitted variables bias (which was much less of an issue with the panel specification). For instance, controlling for land use categories is important, as another reason for why our environmental quality measures change is due to changing LU patterns rather than due to changes in vegetation quality driven by rainfall, which we are interested in isolating. To control for different levels of infrastructure across provinces, we use road density (average km of roads per province). Controlling for ruggedness is crucial, as terrain ruggedness may on the one hand be connected to higher levels of poverty due to difficulties regarding agricultural land use and on the other hand may even have positive correlations with poverty reduction due to historical reasons (see Nunn & Puga, 2012).
As a robustness check we also carry out spatial econometrics exercises. To minimize the risk of biased estimates due to neglected spatial dependence of our data and to gain additional robustness, we specify spatial Durbin models to account for both the possibility of spatially lagged dependent and explanatory variables. In addition, we run Moran's I tests on the residuals of the second-stage regressions of selected IV specifications (results not shown). The results are robust to this explicit modeling of spatial dependence, however, the estimated coefficients become quantitatively smaller.

Instrumental variables approach for both the panel and the cross-sectional specifications
An instrumental variable design is employed to control for the possible endogeneity of soil fertility and vegetation vigor on poverty. We use rainfall (as sources of exogenous variation for soil fertility), NPP, and NDVI (for a similar research design, see for instance Kiuri, 2016).
To qualify as a good instrument, the candidate variable must be (a) strongly associated to the endogenous variable (in our case environmental quality as proxied by soil fertility and vegetation vigor), (b) not affect the outcome variable (in our case poverty and GDP) through any channel other than through environmental quality, and (c) it must be (as-if) randomly assigned (see e.g. Angrist & Pieschke, 2015).
Condition (a) is easily met by the instrumental variable rainfall. The annual variations in the rainfall measure are one of the most crucial determinants of vegetation vigor and biomass productivity (see several papers in the agronomic literature such as Vlam et al., 2014 andSchippers et al., 2015). This is confirmed by the strength of the first-stage regressions that are provided in the appendix. Precipitation influences above ground biomass by affecting inter alia seed germination, seedling growth, and plant phenology (see e.g. Kang et al., 2013;Liu et al., 2014;Yan et al., 2014). Many studies have used rainfall as a predictor of biomass productivity. Furthermore, precipitation is also the main input factor in Revised Universal Soil Loss Equation (RUSLE) models (see e.g. Angulo-Martínez and Beguería, 2009;Hernando and Romana, 2015) and in the GIS-based Universal Soil Loss model (Angima et al., 2003;Fu et al., 2005;Lufafa et al., 2003). 11 There have been several debates on the validity of rainfall as an instrument, particularly on criterion (b), which we reviewed carefully (see below). We argue, recognizing that the perfect instrument does not exist, that our paper is probably the one among the reviewed lot that is best positioned to use rainfall as an instrument. This is mostly because in our case rainfall is much more closely linked to our treatment variables vegetation growth and soil stored carbon than any other treatment variable used in the literature, but also because the exclusion criterion (b) is unlikely violated. Many studies have recently shown the relationship of this measure with many other variables, which in turn might be related to poverty and welfare. Sarsons (2015) for example finds evidence from India that rainfall might affect welfare through other channels, namely transportation and the ability to organize, particularly during extreme events, calling into question the findings of the seminal IV papers from e.g. , Miguel et al. (2004), Miguel (2005), Yang & Choi (2007). Concretely, the concern with using rainfall as an IV for the case of this paper is that extreme rainfall events (such as flooding) can lead to the destruction of property and affect poverty, outside the channels of soil fertility and vegetation changes. The same rationale applies to droughts, that they for example kill livestock due to heat stress in addition due to reduced vegetation vigor. We overcome the flooding and drought identification threat by excluding extreme rainfall events from our main IV specifications (titled 'no outlier' specifications in the results section). Furthermore, by including road density and ruggedness we control for the transportation identification threat. An empirical indication that this exclusion restriction holds is that the OLS specification indicates that rainfall is not a statistically significant predictor in its own right (in addition to NPP, NDVI and soil fertility being in the specification) of poverty rates or GDP per capita.
There is no way to fully avail the identification strategy of possible violations using rainfall, as there may yet even other channels which we have not pondered and which have not yet been discussed in the literature, through which rainfall may affect welfare, although that may be said of any IV. We argue that if there is a fitting case for using rainfall as an IV, using it for isolating the exogenous variation in environmental quality is probably the best candidate. Rainfall is perhaps the most important determinant of plant growth, particularly so in areas with little irrigation.
In low-income and middle-income countries, economies are particularly dependent on the primary sector such as agriculture and forestry. We argue that increased quantities of rainfall increase crop yields and the environmental services from surrounding ecosystems, a mechanism which is especially strong in Sub-Saharan Africa, where only 4% of area cultivated is equipped for irrigation as compared to for instance 28% in North Africa (see You et al., 2010). We rely on the same logic, but instead instrument plant growth directly. Therefore, we posit that annual mean rainfall only affects poverty through biomass productivity, and that outside this channel it has no effect on poverty reduction. To emphasize the theoretical efforts made to ensure the restriction holds, we again stress the additional control variables and excluded extreme rainfall events mentioned above.

Results
We detail three main findings: (1) Environmental quality -both above-ground as well as below-groundhas significant and sizeable negative effects on poverty, particularly in rural areas and in Sub-Saharan Africa.
(2) The relationship between environmental quality and GDP per capita is less strong, and only significant in certain specifications, suggesting that environmental quality is particularly important for the poor, compared to the average income earner. (3) In-situ environmental quality improvements are propoor, in contrast to urbanization. While urbanization has highly significant and sizeable correlations with GDP per capita, it is not significantly correlated with poverty reduction, unlike environmental quality, which is significantly correlated with both.

a. Panel Fixed Effects evidence
The correlation between concurrent environmental quality changes and poverty reduction are strong (see figure 1). The graph shows the conditional relationship of vegetation vigor (as measured by Net Primary Productivity) and poverty after controlling for other possible predictors of poverty in the panel specification. It can be clearly seen that increasing vegetation vigor is associated with accelerated poverty reduction.

Figure 1: The relationship between changes in vegetation & poverty reduction
Note: Each dot represents an equally sized bin of observations (grouped over the x-axis). Within these bins, the average of the xand y-variable is computed and used in a scatterplot. The plot gives the conditional effect of log(NPP) on the residualized poverty headcount ratio after controlling for several covariates. They are created by running OLS regressions equivalent to table 2, specification 1.
However, it does not appear as if environmental quality causally influences poverty reduction for the average province in our global sample, as shown in table 2 (despite the OLS specification (1) being significant, the IV specifications (4) and (5) are not). Environmental quality however is much more important for poorer and more rural areas, as shown by specification (2) and (3). In fact, among rural areas and Sub-Saharan Africa, poverty is reduced significantly by vegetation vigor increasing, as shown by specifications (6)  Conditional NPP/Poverty 12 rates. Similarly, increases in vegetation vigor as measured by NDVI resulted in poverty rate reductions in rural areas (see robustness check in appendix). The reasons for such significant and sizeable effects in Sub-Saharan Africa and rural areas has likely to do with livelihoods there being more dependent on the environment, among other reasons that are discussed below.
Using the data variation available and assuming a linear relationship, we find that the effects of NPP range between a reduction of the poverty headcount ratio by 4.8 percentage points (in times of extremely large increases in vegetation) and an increase in the poverty headcount ratio of around 3 percentage points (in times of a large decrease in vegetation vigor). In that sense, the effects we find are significantly larger than the effects suggested in Okwi et al. (2007). The interquartile range of the effects is -0.28 percentage points and +0.39 percentage points for the rural areas. The coefficients of the other control variables are in line with the effects suggested by the literature. Note however that all effects other than those of environmental quality are not 'exogenized,' meaning they may not be interpreted as causal. That said, important lessons can be drawn from correlations as well. For example, road density, our proxy for infrastructural development, is an important and strong predictor of poverty reduction. This is a long established and well-known effect in development economics. For an overview of possible theoretical channels see for instance Brenneman & Kerf (2002). Urbanization (as measured by the growing area of cities), another variable that is consistently correlated with poverty reduction in the development economics literature (Rakodi, 2014), does not show consistently significant correlations with poverty reduction. Ruggedness in turn is statistically positively related to poverty rates, suggesting that rougher terrains make it harder to escape poverty. The direction of control variable coefficients is generally in line with the literature, which gives confidence in the quality of the data and empirical approach. 12

b. Cross-sectional findings
There is a negative relationship between poverty and soil quality. This means the better the soil quality, the lower the poverty rates (see figure 2). The correlations after controlling for fixed effects point in the direction of poverty being a problem in regions with low soil quality. This relationship is even stronger when we look at the rural subset of our data. Due to the unavailability of time-variant soil fertility and subnational GDP measures, we use a cross-sectional model specification.

Figure 2: Poverty and Soil Quality
Note: Each dot represents an equally sized bin of observations (grouped over the x-axis). Within these bins, the average of the xand y-variable is computed and used in a scatterplot. The residuals from a regression of poverty headcount ratio on country fixed effects are on the y-axis. Top soil carbon content is on the x-axis. Left panel is overall sample, right panel is areas with high dependence on agriculture.
This inverse correlation of soil fertility and poverty (as shown in figure 2 and the OLS specification in table 3) is also significant in the specifications that isolate the effects of poverty (see specification 3). An increase in top soil carbon content of 10 percent reduces the poverty headcount ratio by around 2 percentage points in the global and the rural samples. Rural Sub-Saharan African areas are particularly povertyreducing as a consequence of soil quality improving: a 10 percent increase is soil fertility is linked to a 12 It is also worth noting that the results of the first-stage regressions in the appendix are broadly in line with what we would expect from a theoretical point of view. For instance, a higher forest share is significantly correlated with higher NPP/NDVI levels, whereas a higher grassland share is connected to a lower biomass level. Similar observations can be made with respect to the cross-sectional first-stage regressions. roughly 9 percentage points reduction in poverty rates. The positive soil fertility effects on poverty reduction are in line with the agricultural literature that connects farm income and soil erosion, see for example Hazarika and Honda (2001).
Aside from land use, we also control for soil type, which is a very important control measure in the cross section, as it allows us, along with land use, to isolate the intensive margin changes, and control for poverty reduction effects due to changes in land use. The soil-GDP elasticities are much less consistently significant (see table 4) than the soil-poverty elasticities. Higher soil quality does not contribute to higher GDP per capita. However, caution is warranted in interpreting these findings, because while the existence of a positive effect, as shown in the poverty example, is a good indication that there are indeed significant relationships, the absence of such may also be due to poor data quality 13 and a small sample size and resulting low powered analysis. Aside from the positive effects of environmental quality on poverty reduction, and effect of the intensive margin, there are also interesting associations displayed by measures of the extensive margin such as land use changes. Infrastructure (road density) development shows significant positive relationships with GDP per capita and significant negative relationships with poverty ratios in the panel estimation and the crosssectional specifications.
Not in line with our expectations, crop land expansion is not positively associated with poverty reduction, quite to the contrary (in both the panel and the cross-sectional OLS specifications). The direction of the effects may be an indication that it has not been the poor who have been benefitting from land conversion (such as forest-to-crop conversion) in the overall sample. In fact, poverty rates might increase after areas are converted into cropland according to our results. However, much more research into the reasons for this unexpected positive correlation between poverty rates and cropland is necessary.
Urbanization (urban expansion) is significantly linked to increasing GDP per capita, and has in fact the largest coefficient among any land use change measures, by a large margin, attesting to the powerful links between urbanization and economic development. However, the effects on poverty rates are not significant. This is a puzzling observation. As we had mentioned before, only environmental quality is 'exogenized' and therefore all other relationships can be interpreted merely as correlative. That said, the absence of a correlation between urbanization and poverty, especially despite such strong and significant relationships with GDP per capita, suggests that urbanization is not an inclusive demographic process. On the other hand, the Stable Unit Treatment Variable Assumption (SUTVA) may be violated, meaning that it could be that those in rural areas with the most talent moved to the cities (in other provinces), contributing to economic growth there, but being absent from contributing to the economic development process in the rural origin provinces, which would have helped to reduce poverty (and not sent back home enough remittances to counteract that demographic effect).
The relative strength of the effects of the different explanatory variables on poverty need to be assessed with caution, as it is hard to grasp at what expense each of these changes happen. For example, looking at the original OLS specification reveals that although a 10 percent change in NPP is associated with a 0.4 percentage point change in poverty rates, a 1 percent change in road density only leads to a 0.3 percentage point change in poverty rates. This could indicate that improving environmental quality has larger ties to poverty reduction than improvements of road density. However, the relative costs for pursuing one over the other are not part of the modeling, and this specification does not provide causal identification. Additional research efforts to compare the relative effects of urbanization, infrastructure development, and in-situ development of environmental quality in rural areas are needed.

Conclusion and Discussion
As expected, we find evidence that building roads is associated with reductions in poverty rates, as many studies have shown before (see e.g. Jacoby, 2000;Gibson and Rozelle 2003;Mu and van de Walle 2007;Jacoby and Minten, 2009;Khandker et al., 2009). What has not been shown conclusively so far however is whether in-situ improvements of environmental quality significantly reduce poverty. Several authors have suggested that they do not. Okwi et al. (2007), for example conclude that if all of Kenya's soil was raised to its highest quality, an only 1 percentage point reduction in poverty rates would ensue. Wantchekon and Stanig (2015) go even farther and conclude that in Sub-Sharan Africa good soil may be a hindrance for poverty reduction.
We find that environmental quality is indeed important for poverty reduction. Using the unbalanced panel data from the Hidden Dimensions data set, we find that as environmental quality increases, the speed of poverty reduction accelerates. Using instrumental variable regressions and exploiting the exogenous time variations in annual rainfall, we find that increasing environmental quality causes significant accelerations in poverty reduction rates, particularly in rural areas and in Sub-Saharan Africa.
Our analysis does not specifically address any mechanism through which environmental quality helps reduce poverty, which is a major research question that has yet to be tackled (see Andam et al., 2010). Agricultural productivity and higher environmental services are most likely to be the channels through which the poverty reduction effect operates. More research on this issue is required in the future, such as disentangling the different channels of irrigation schemes, 14 the importance of the primary sector, and the effect of high initial poverty rates.
We find a similar pattern in a cross-sectional setting. Soil quality has significant and sizeable effects on poverty rates and on GDP per capita. We think that the underlying mechanism is that as environmental quality increases, agricultural and environmental output increase as well. This might be especially relevant for low income households that draw a larger share of their income from natural resources and the environment (Wunder, 2015).
Landscape restoration can reduce degradation and positively contribute to increasing environmental quality (and therefore also poverty reduction). In the Inner Mongolia region of China, Mu et al. (2013) examined NPP changes from 2001 to 2009 and found NPP increased by 21 percent. Of this increase, 80 percent of the NPP change was due to human activity -notably landscape restoration programs (returning cropland to grassland, converting desert to grassland, and better grazing management). These results 14 The coefficients for Sub-Saharan Africa are usually larger than for the full sample specifications. An interesting idea to pursue for further research would therefore be to explore the mechanisms that lead to this result. It is hard to establish whether the effects are larger in SSA because of the extreme poverty in the region or lower irrigation. The irrigation channel could be instrumentalized using available subnational irrigation data sets, see for instance the AQUASTAT data set (Siebert et al., 2005). However, data quality is low especially in Southern and East Africa. For example, the highest indicated data quality in AQUASTAT is measured in North America with 1.03, whereas it is just 3.85 in Southern Africa (the worst measured quality is 4.00). As this means that the geospatial data in these regions rely on highly modeled estimates, we did not pursue this idea in this paper.
suggest that human activities -and landscape restoration specifically -can have a larger impact on NPP changes than climate. Similar results have been found in the Qinghai-Tibet Plateau (Chen et al., 2014;Cai et al., 2015), and elsewhere in China (Xiao et al., 2015). A cross-country comparison of China, Mongolia, Uzbekistan, and Pakistan finds climate influences to be stronger than human influences in explaining NPP changes, but nonetheless concludes that improvements were largely determined by landscape restoration programs (Yang et al., 2016).
In conclusion, we find that the environment does matter for poverty reduction. The greener the vegetation and the richer the soil, the faster the poverty reduction, particularly so in Sub-Saharan Africa and rural areas. Moreover, we find that the effects of environmental quality on poverty are stronger than its effects on average income, suggesting that the poor benefit disproportionately more from environmental quality. Lastly, we find that although urbanization has highly significant and sizeable correlations with GDP per capita, it is not significantly correlated with poverty reduction.