2. The spatial distribution of rural populations in developing countries

One of the first studies to determine the spatial distribution of populations on less favored lands globally was CGIAR (1999), which concluded that nearly two-thirds of the rural population of developing countries–almost 1.8 billion people–lives on less-favored lands, including marginal agricultural lands, forest and woodland areas, and arid zones. By applying national rural poverty percentages, CGIAR (1999) determined that 633 million poor people lived on less favored lands in developing countries, or around two-thirds of the total rural poor (see also CAWMA, 2008).

A subsequent analysis by the World Bank (2003) sought to identify the percentage of total population in a selection of low and middle-income economies located on ‘fragile lands’ in 2000. This classification comprised four categories of land: terrain greater than 8 per cent median slope, soil unsuitable for rainfed agriculture, arid and dry semi-arid land without access to irrigation, and forests (deciduous, evergreen and mixed). The study estimated that nearly 1.3 billion people in 2000–almost a fifth of the world's population–lived in such areas in developing regions, and concluded that since 1950, the estimated population in developing economies on ‘fragile lands’ may have doubled (World Bank, 2003).

The World Bank (2008) employed the definition proposed by Pender and Hazell (Reference Pender, Hazell, Pender and Hazell2000) for less favored areas to determine the spatial distribution of rural populations in 2000. However, the analysis was able to determine only the distribution of rural population on lands limited by rainfall (arid and semi-arid lands) and in remote areas with poor market access. Around 430 million people in developing countries in 2000 lived in such distant rural areas, and nearly half (49 per cent) of these populations were located in arid and semi-arid regions characterized by frequent moisture stress that limits agricultural production (World Bank, 2008).

Using a variety of global spatially referenced data sets, we analyzed the spatial distribution of rural population across developing countries in 2000, following the classification of less favored agricultural land and areas of figure 1 (see ‘Technical notes, data sources and mapping methods’ in online appendix B). Less favored agricultural land consists of irrigated land on terrain greater than 8 per cent median slope; rainfed land with a length of growing period (LGP) of more than 120 days but either on terrain greater than 8 per cent median slope or with poor soil quality; semi-arid land (land with LGP 60–119 days); and arid land (land with LGP<60–119 days). These various land areas were identified by using FAO Global Agro-Ecological Zones (GAEZ) Data Portal version 3 datasets (available online at http://gaez.fao.org/), combined with national boundaries from the Gridded Population of the World, version 3 (GPWv3) of the Center for International Earth Science Information Network (CIESIN) and Centro Internacional de Agricultura Tropical (CIAT). Agricultural land extent was obtained from the Pilot Analysis of Global Ecosystems (PAGE) (http://www.ifpri.org/dataset/pilot-analysis-global-ecosystems-page), and rural populations determined from the rural-urban extent dataset that was published as part of CIESIN Global Rural Urban Mapping Project (GRUMPv1). Market accessibility was used to identify remote areas using Nelson (Reference Nelson2008) as released by the Global Environment Monitoring Unit of the Joint Research Centre of the European Commission. Following Nelson (Reference Nelson2008), we identify market access as less than five hours of travel to a market city with a population of 50,000 or more.

The results of this analysis for the main regions comprising 124 of the 139 low and middle-income economies as classified by the World Bank (2014) are depicted in table 1.Footnote ³ Almost 36 per cent of the 2000 rural population in these developing countries was located on less favored agricultural land, although this share ranged from 23 per cent in the Middle East and North Africa to 56 per cent in Europe and Central Asia. Over 37 per cent of the rural population in 2000 was in less favored agricultural areas, and about 8 per cent on remote less favored agricultural lands. The latter also amounted to 22 per cent of all the rural population on less favored agricultural land, with this share varying from 12 per cent in Europe and Central Asia to 30 per cent in Sub-Saharan Africa.

Table 1. Spatial distribution of rural population by major developing region, 2000

Notes: Less favored agricultural land (LFAL) consists of irrigated land on terrain greater than 8% median slope; rainfed land with a length of growing period (LGP) of more than 120 days but either on terrain greater than 8% median slope or with poor soil quality; semi-arid land (land with LGP 60–119 days); and arid land (land with LGP<60 days). These various land areas were determined by employing in Arc GIS 10.1 the datasets from the FAO Global Agro-Ecological Zones (GAEZ) Data Portal version 3 (http://gaez.fao.org/) combined with national boundaries from the Gridded Population of the World, Version 3 (GPWv3) of the Center for International Earth Science Information Network (CIESIN) and Centro Internacional de Agricultura Tropical (CIAT). Agricultural land extent was obtained from the Pilot Analysis of Global Ecosystems (PAGE) (http://www.ifpri.org/dataset/pilot-analysis-global-ecosystems-page), and rural populations determined from the rural-urban extent dataset that was published as part of CIESIN Global Rural Urban Mapping Project (GRUMPv1). Market accessibility was used to identify remote areas using Nelson, (Reference Nelson2008) as released by the Global Environment Monitoring Unit of the Joint Research Centre of the European Commission. Market access is identified as less than five hours of travel to a market city with a population of 50,000 or more.

Developing countries are all low and middle-income economies with 2012 per capita income of US$12,615 or less (World Bank, 2014).

Given the evidence that a sizable proportion of the rural population was located on less favored lands and in remote areas in 2000, we explore next how this spatial population distribution might influence overall poverty, as measured by a poverty headcount indicator. In subsequent sections, we examine empirically how the spatial distribution of rural population in 2000 may have affected poverty changes across developing countries in the subsequent 2000–2012 period.

3. Spatial distribution and poverty

To illustrate the potential influence of the shares of rural population on less favored agricultural land and remote areas on poverty, we begin by defining a standard poverty relationship, which depends both on the average income of the population and the properties of the Lorenz curve that determine income distribution. If people living in these disadvantaged regions have lower incomes, then the share of the rural population with poorer access to markets and located on low quality agricultural land should affect average income of the entire population and thus also the properties of the Lorenz curve. We adjust the corresponding measure of poverty, the poverty headcount index, accordingly.

Following Gastwirth (Reference Gastwirth1971), the inverse of the continuously differentiable cumulative distribution function F(y) defines the quantile function for p; i.e., the income level y below which we find a proportion p of the population. That is, y = F ⁻¹(p) = y(p).Footnote ⁴ This leads directly to derivation of the Lorenz curve, the graphical representation of the fraction of total income that the holders of the lowest pth fraction of incomes possess,

(1)$$L(p)={1 \over \mu }\int_0^p F^{-1}(t)dt, \quad L_p ={\partial L \over \partial p}={y(p) \over \mu }>0, \quad L_{pp} >0,\quad 0\le p\le 1,$$

where $\mu =\int_{0}^{\infty} {ydF(y)} =\int_{0}^{\infty} {yf(y)dy}$ is the mean income of the population. Thus the derivative of the Lorenz curve with respect to p gives the ratio of the income of that share of the population to the average income of the entire population n. As y′ (p) > 0, the Lorenz curve is an increasing and convex function of p.

Defining H as the poverty headcount index, i.e., the share of the population with income no higher than a defined poverty line z, it follows that H = F(z) and thus z = F ⁻¹(H). Inverting the derivative of the Lorenz curve in (1), and evaluating it at p = H, yields

(2)$$H=L_p^{-1} \left( {{z \over \mu }} \right).$$

Expression (2) indicates that the level of poverty may change due to a change in the mean income μ relative to the poverty line or due to a change in the properties of the Lorenz curve that determine relative income inequalities.Footnote ⁵

Our conjecture is that the spatial distribution of population, especially the proportion of rural populations on less favored agricultural land and remote areas, may also influence the incidence of poverty, as measured by (2). Denote s as the share of the rural population in these marginal agricultural locations. Because of the poor agricultural productivity and/or returns to these lands, households living in these locations are likely to experience stagnant and low standards of living compared to identical households with access to better land, markets and infrastructure. It follows that a larger share of rural population on less favored agricultural land and remote areas s adversely impacts average income, i.e., μ = μ (s), μ′ < 0. However, if s influences average income, then from (1) it also influences the Lorenz curve L(p).

This suggests that (2) can be expressed as

(3)$$H=L_p^{-1} \left( {{z \over \mu }(s); \, s} \right).$$

As before, the incidence of poverty is inversely related to the mean income of the population, but now the share of the rural population on less favorable agricultural land and remote areas s also affects poverty. This direct influence of the spatial distribution of the rural population on poverty changes is an empirically testable hypothesis. In addition, if s influences average income μ, then it may also affect the poverty-reducing impacts of income growth, which has been the focus of extensive cross-country poverty analysis.Footnote ⁶ This indirect influence of the spatial distribution of the rural population on poverty changes via income growth is a second hypothesis worth exploring empirically. As (3) indicates, the key variables for such an empirical analysis of both hypotheses are the spatial distribution of rural population s, the poverty headcount index H, a given poverty line z, and mean income μ.

4. Data and descriptive statistics

As indicated in table 1, we have estimated four spatial distribution variables for the rural population in 2000 for 124 low and middle-income economies. These variables, which represent s, are:

• • the share (per cent) of the rural population on less favored agricultural land (henceforth s ₁),

• • the share (per cent) of the rural population in less favored agricultural areas (s ₂),

• • the share (per cent) of the rural population on remote less favored agricultural land (s ₃), and

• • the share (per cent) of the rural population on less favored agricultural lands located on remote land (s ₄).

We use these variables to test the two hypotheses concerning the direct and indirect influences that the spatial distribution of the rural population has on poverty changes for developing countries from 2000 to 2012.

Following the recent poverty analysis literature,Footnote ⁷ we obtain our cross-country measures of a given poverty line z, the poverty headcount index H, and mean income μ from PovcalNet, the online tool for poverty measurement developed by the Development Research Group of the World Bank (available online at http://iresearch.worldbank.org/PovcalNet/). PovcalNet produces internationally comparable country level poverty and income distribution estimates based on standardized household surveys across 127 developing countries. From this database, we identify 83 low and middle-income economies with at least two suitable household surveys from 2000 to 2012. The longest available spell between surveys is used for each country, and both surveys use the same welfare indicator, either consumption or income per person. The median interval between surveys is eight years, and it varies from two to eleven years.Footnote ⁸ All monetary measures are in constant 2005 prices and are at Purchasing Power Parity (PPP).

The poverty headcount index H is the percentage of the population living in households with consumption per capita (or income when consumption is not available) below the poverty line. We follow Ravallion (Reference Ravallion2012) and choose a poverty line z of US$2 per person per day at 2005 PPP. In the initial survey year, the median poverty headcount index across all 83 countries was 42.85 per cent, but ranged widely from 0.29 to 95.44 per cent. By the final survey year, the median poverty headcount was 27.86 per cent, and it varied from 0.08 to 93.49 per cent.

Mean income μ is the average monthly (2005 PPP US$) per capita income or consumption expenditure from the household surveys for each country in the relevant year. In the initial survey year, the median per capita monthly income was US$100 across all 83 countries, and ranged from US$24 to 2,003. In the final survey year, median income was US$115, and varied from US$28 to 2,012. Finally, inequality is measured by the usual Gini index, which was also obtained from the PovcalNet cross-country household surveys for the relevant years. Table 2 summarizes the descriptive statistics for the key variables used in the poverty analysis for our sample of 83 developing countries.

Table 2. Descriptive statistics of key poverty analysis variables

Notes: Based on a sample of 83 developing countries. See online appendix tables A1–A3.

We also employ a number of control variables in our analysis, following the approach of similar poverty analyses.Footnote ⁹ The controls are inflation, government consumption as a share of GDP, arable land per capita, agricultural value added as a share of GDP and per worker, investment as a share of GDP, trade openness, primary school enrollment, and life expectancy. These variables were obtained from the World Development Indicators (World Bank, 2014), and as far as possible, for 2000 and our sample of 83 countries. Other controls include a dummy for landlocked country as defined by UNDP (http://unctad.org/en/pages/aldc/Landlocked%20Developing%20Countries/List-of-land-locked-developing-countries.aspx), for small island developing states as defined by UNESCO (http://www.unesco.org/new/en/natural-sciences/priority-areas/sids/about-unesco-and-sids/sids-list/), and distance from equator for each country. We also employ rule of law and democracy (voice and accountability) indices, from the Worldwide Governance Indicators (http://data.worldbank.org/data-catalog/worldwide-governance-indicators), which were averaged over 1996–2000 for each country. Finally, we use regional dummies for the six main developing country regions (see table 1).

5. Estimation strategy and results

Our estimation strategy follows previous literature that estimates the growth elasticity of poverty with respect to mean income empirically, where the change in poverty between two periods of time is explained by the growth of income (Ravallion, Reference Ravallion1997 Reference Ravallion2001 Reference Ravallion2012; Ravallion and Chen, Reference Ravallion and Chen1997; Dollar and Kraay, Reference Dollar and Kraay2002; Bourguignon, Reference Bourguignon, TS and Turnovsky2003; Adams and Page, Reference Adams and Page2005). As noted in the Introduction, this literature generally confirms that the estimated growth elasticity of poverty is positive, although initial income inequality can also mitigate the poverty-reducing impact of income growth. In addition, Ravallion (Reference Ravallion2012) also examines whether initial level of poverty affects directly the change in poverty over two periods of time, or whether initial poverty exerts a significant indirect impact via the elasticity of poverty reduction with respect to growth.

Consequently, to analyze the possible direct and indirect influences of our spatial distribution variables s _k in 2000 on poverty changes from 2000 to 2012 in our 83 sample countries, we follow a similar estimation strategy to that of Ravallion (Reference Ravallion2012). Thus, our basic regression is

(4)$$\gamma _i (H_{it} )=\alpha _0 +\alpha _1 \ln (v_{it-\tau } )+\lpar {\beta_0 +\beta_1 v_{it-\tau } } \rpar \gamma _i (\mu _{it} )+ w_{it},$$

where i is each country observation, t is the final survey date, τ is the length of time between surveys, and ω_it is the error term. In equation (1), the dependent variable of the regression is

$$\gamma _i (H_{it} )\equiv \ln \left( {{H_{it} \over H_{it-\tau}}} \right)/\tau,$$

which is the annualized change in log headcount poverty rate for US$2 a day between surveys, and thus represents growth in poverty.Footnote ¹⁰ As noted above, across our sample of 83 countries, the median survey spell is eight years, and it varies from two to eleven years (i.e., the full 2000–2012 period). A standard hypothesis in the literature is that changes in poverty over time will be influenced by growth in income (Ravallion, Reference Ravallion2001, Reference Ravallion2012; Dollar and Kraay, Reference Dollar and Kraay2002; Adams and Page, Reference Adams and Page2005; Kraay, Reference Kraay2006). That is, as the mean per capita income across surveyed households rises, one would expect their average poverty rate to fall. Thus, a key explanatory variable in determining changes in poverty between surveys in equation (1) is the annual growth in income per person, which is represented by the annualized change in log survey mean income between surveys $\gamma_{i} (\mu _{it})\equiv \ln (\mu _{it}/\mu _{it-\tau})/\tau $.

Equation (4) also specifies that the change in poverty over time could be influenced by the initial level of another variable of interest v _it−τ. In Ravallion (Reference Ravallion2012), this variable of interest is the initial poverty level H _it−τ. In our analysis, we are also interested in how each of the four spatial distribution variables in 2000, i.e., s _kit−τ, might influence the change in poverty. As indicated in (4), any such variable of interest could have a direct impact on changes in poverty over time, $\alpha_{1} \ln (v_{it-\tau }) \gamma_{i} (\mu_{it})$, or it could affect the poverty-reducing impact of income growth, $(\beta_{0} +\beta_{1} v_{it-\tau})\gamma_{i} (\mu_{it})$.

It follows that two tests of restrictions on the various parameters estimated by (4) determine the direct and indirect influence of any variable of interest v _it−τ on the annualized change in poverty (Ravallion, Reference Ravallion2012). For example, rejection of the null hypothesis α₁ = 0 for H _it−τ or s _kit−τ indicates that initial poverty or spatial distribution levels have a direct influence on changes in poverty over time, and subsequently, the magnitude of α₁ determines whether this influence is positive or negative. Failure to reject the null hypothesis of homogeneity, i.e., β₀ + β₁ = 0, confirms that initial poverty or spatial distribution levels have an indirect influence through ‘adjusting’ the growth elasticity of poverty reduction. That is, the restriction implies β₀ = −β₁ and the correct regressor is $(1-v_{it-\tau } )\gamma _{i} (\mu _{it} )$. From (4), the expected sign of the coefficient of this regressor is negative. Thus, in the case that both restrictions α₁ = 0 and β₀ + β₁ = 0 hold, then the regression becomes

$$\gamma _i (H_{it}) = \alpha_0 +\beta_1 (1-v_{it-\tau})\gamma _i (\mu _{it}) + w_{it}, \quad \beta _1 < 0.$$

Our strategy for estimating (4) involves four sets of regressions. First, we replicate the analysis by Ravallion (Reference Ravallion2012), using the initial poverty level H _it−τ for v _it−τ in (4), for our sample of 83 countries over 2000 to 2012. Second, we repeat the analysis but include separately (in log form) each of our four spatial distribution variables s _kit−τ as independent regressors. Third, we re-estimate (4) for our sample of countries, but this time using each of our four spatial distribution variables s _kit−τ for v _it−τ. Finally, allowing for the possibility that the ‘poverty-adjusted’ growth rate is influenced by the ‘spatial distribution-adjusted’ growth rate, we estimate IV, SUR and 3SLS regressions of (4) taking this into account. That is, if H _it−τ is affected by any s _kit−τ, which is a possibility implied by equation (3), then $(1-H_{it-\tau})\gamma _{i} (\mu_{it})$ is determined by $(1-s_{kit-\tau})\gamma _{i} (\mu _{it})$. For all four steps, we estimate the regressions both with and without additional control variables.

Our replication of the estimations in Ravallion (Reference Ravallion2012, table 4) for our sample of 83 countries produces similar results. We conduct both OLS and IV regressions of (4), with the latter estimations using the growth rate in private consumption per capita from the national accounts from World Bank (2014) as the instrument for the growth in mean survey income.Footnote ¹¹ For all OLS and IV specifications, the null α₁ = 0 cannot be rejected, and thus the initial level of poverty H _it−τ has no direct influence on poverty changes over time. However, the homogeneity restriction β₀ + β₁ = 0 also cannot be rejected, confirming that the poverty-adjusted growth rate $(1-H_{it-\tau})\gamma _{i} (\mu _{it})$ is the relevant regressor. When homogeneity is imposed on (4), we obtain a significant and negative poverty-adjusted growth elasticity, which is −2.83 for OLS and −4.85 for IV. The corresponding estimates in (Ravallion, Reference Ravallion2012, table 4) are −2.47 and −3.09, respectively.

In our second set of estimations, we find that including each additional ln s _kit−τ variable in the previous poverty analysis regressions has no effect on the results. None of the coefficients on these spatial distribution measures is significant, and their inclusion did not change, or improve, the poverty-adjusted growth regression results. For example, for the OLS regressions, the estimated poverty-adjusted growth elasticity remained significant and ranged from −2.89 to −3.00, and for IV, from −4.85 to −4.96.

Table 3 depicts the results of our third set of regressions.Footnote ¹² Only the estimations corresponding to the share (per cent) of the rural population in 2000 on less favored agricultural lands on remote land (s ₄) are shown; however, similar results are obtained for the other three s _k spatial distribution measures (see supplementary statistical tables A1–A3 in online appendix A). In all regressions in table 3, the null α₁ = 0 cannot be rejected, and thus the 2000 spatial distribution measure s _4it−τ has no direct influence on poverty changes over time. The homogeneity restriction β₀ + β₁ = 0 also cannot be rejected, confirming that the spatial distribution-adjusted growth rate $(1-s_{kit-\tau})\gamma _{i} (\mu _{it})$ is the relevant regressor. In the model imposing homogeneity, the elasticity of growth adjusted for s _4it−τ is significant in both OLS and IV. However, there is considerable difference in the estimated elasticity, which ranges from −2.10 in the OLS regression (column 5) to −7.19 in the IV estimation (column 6). These results do not change when additional control variables are added to the regressions that impose homogeneity on (4).

Table 3. Estimates of the effect of the spatial distribution of the rural population s _{4it − τ} and income growth $\gamma_{i}(\mu_{it})$ on the proportionate change in poverty $\gamma _{i} (H_{it})$

Notes: The dependent variable is the annualized change in log poverty rate for US$2 a day $\gamma _{i}(H_{it})$; t-ratios based on robust standard errors in parentheses; the IV estimations employ the growth rate in private consumption per capita from the national accounts from World Bank (2014) as the instrument for the growth in mean survey income; **significant at the 1% level; *significant at the 5% level; †significant at the 10% level.

Repeating the regressions of table 3 but using the other three s _k spatial distribution measures instead of s ₄ produces a similar outcome (see online appendix tables A1–A3).Footnote ¹³ When homogeneity is imposed, the coefficient on the spatial distribution-adjusted growth rate $(1-s_{kit-\tau }) \gamma _{i} (\mu_{it})$ is significant and negative, but only in the OLS and not the IV estimations. For the regressions using the share (per cent) of the rural population on less favored agricultural land s _1it−τ to adjust growth, the estimated elasticity is −1.40, for the share (per cent) of the rural population on less favored agricultural areas s _2it−τ the elasticity is −1.59, and for the share (per cent) of the rural population on remote less favored agricultural land s _3it−τ the elasticity is −1.67.

In the final set of regressions, we estimate IV, SUR and 3SLS poverty-adjusted growth regressions, allowing for the possibility that the poverty adjusted growth rate is determined by the spatial distribution adjusted growth rate with respect to each of our four s _kit−τ measures. We use both the growth in mean survey income as our measure of $\gamma _{i} (\mu_{it})$ and also instrument for this variable with the growth rate in private consumption per capita. Thus, for these regressions, the relevant system of equations is

(4a)$$\gamma _i (H_{it}) =\alpha _0 +\beta _1 (1-H_{it-\tau})\gamma _i (\mu_{it})+ w_{it},$$

(4b)$$(1-H_{it-\tau})\gamma _i (\mu _{it}) =\delta _0 +\delta _1 (1-s_{kit-\tau}) \gamma_i (\mu_{it} )+ \varepsilon _{it}, \quad k=1,\ldots ,4.$$

In all regressions of (4a) and (4b), and for all spatial distribution measures of s _kit−τ the results are robust, and the coefficients significant and with the expected signs (see online appendix tables A1–A3). That is, the estimated poverty-adjusted elasticity for $(1-H_{it-\tau}) \gamma _{i} (\mu_{it})$ is significant and negative, and in the first stage of the SUR and 3SLS regressions, this variable is positively and significantly influenced by $(1-s_{kit-\tau})\gamma _{i} (\mu_{it})$. These results are also robust when additional control variables are included in the regressions, although almost all the controls are individually and jointly insignificant in all IV, SUR and 3SLS regressions, with the exception of agricultural value added per worker, investment as a share of GDP and the Europe and Central Asia dummy.

Table 4 depicts the 3SLS estimations corresponding to the share (per cent) of the rural population in 2000 on less favored agricultural lands on remote land (s ₄). The elasticity of the poverty-adjusted growth rate with respect to mean survey income is negative and significant (see columns (1) and (5)). Moreover, its value (−2.37 to −3.05) corresponds closely to the estimates for this elasticity in our first set of regressions, when we replicate Ravallion (Reference Ravallion2012, table 4) for our sample of countries. Similarly, the elasticity of poverty-adjusted growth using private consumption as an instrument is significant, negative and close to the estimates from our first set of regressions (−5.07 to −5.80). In addition, we find a positive and significant influence of spatial distribution-adjusted growth $(1-s_{4it-\tau})\gamma _{i} (\mu _{it})$ on poverty-adjusted growth $(1-H_{it-\tau }) \gamma _{i} (\mu_{it})$. This elasticity is 0.63 to 0.69 for growth in mean survey income, and 0.22 to 0.46 when the latter variable is instrumented by growth in per capita consumption.

Table 4. 3SLS estimates of the effect of spatial distribution-adjusted growth $(1-s_{4\iota t-\tau }) \gamma_{i}(\mu_{it})$ on the elasticity of growth-adjusted poverty reduction

Notes: The dependent variable in the first equation is the annualized change in log poverty rate for US$ 2 a day $\gamma_{i}(H_{it})$, and the dependent variable in the second equation is the poverty-adjusted growth rate, $\gamma_{i}(\mu_{it})\cdot (1-H_{it-\tau })$; mean survey income refers to regressions that employ the growth in mean survey income $\gamma_{i}(\mu_{it})$; private consumption refers to regressions that employ the growth rate in private consumption per capita from the national accounts from World Bank, 2014 as the instrument for growth in mean survey income; t-ratios based on robust standard errors in parentheses; **significant at the 1% level; *significant at the 5% level; †significant at the 10% level.

Table 4 also indicates that, when they are significant, the control variables have the expected signs. For example, investment share of GDP reduces overall poverty, and agricultural productivity increases poverty-adjusted growth. Poverty is generally lower, whereas poverty-adjusted growth generally higher, in Europe and Central Asia compared to other developing regions. However, as also found by Ravallion (Reference Ravallion2012), the initial Gini index appears not to be significant.

6. Effects on poverty-reducing impacts of income growth

Table 5 summarizes the implications of our empirical results of each of the four s _k spatial distribution variables for the poverty-reducing impacts of growth in income per capita. For comparison, the table also shows the impacts on changes in poverty from growth in average income only. For example, in the absence of any change in the spatial distribution of rural populations or in initial poverty levels, a one-standard-deviation increase of 3.52 per cent in average income growth in our sample of developing countries, from 3.36 to 6.88 per cent, would reduce poverty by an additional 4.97 per cent each year.

Table 5. Effects of the spatial distribution of rural population on the poverty-reducing impact of growth in income per capita

Notes: The estimates of the poverty-reducing impact of growth in income per capita are $\beta_{1}\delta_{1}(1-s_{kit-\tau}/100)\gamma(\mu_{it})$ for each spatial population distribution variable s _k, where the annualized growth rate in survey income per capita γ(μ) is evaluated at the mean for the sample of 83 countries, which is 3.36% (see table 2). Parameter estimates for β₁ and δ₁ are from three-stage least squares (3SL3) estimations, with and without controls. t-ratios are in parentheses; **significant at the 1% level; N = 79 or 80. See also online appendix tables A1–A3.

The values for each spatial distribution variable correspond, respectively, to one-standard-deviation below the mean, the mean, and one-standard-deviation above the mean for the sample of 83 developing countries in 2000 (see table 2).

As table 5 indicates, this poverty-reducing effect of per capita income growth is altered significantly when we adjust income growth for each of the spatial population distribution variables, s ₁, s ₂, s ₃ and s ₄. A higher share of the rural population on less favored agricultural lands and areas diminishes the poverty-reducing impact of per capita income growth, whereas a lower share enhances the effect of income growth on lowering poverty.

For example, consider an annual per capita income rate of 3.4 per cent, which is the mean for the sample of 83 countries (see table 2). A country with a high share of rural population on less favored agricultural land s ₁, such as 59 per cent (one-standard-deviation above the mean for all 83 countries), would expect to see a rate of poverty reduction of only 1.8 to 1.9 per cent per year. However, if 38 per cent of a country's population is on less favored agricultural land (the mean for all 83 countries), then poverty would be reduced by 2.7 to 2.9 per cent annually. If a country has only 17 of its rural population on less favored agricultural land (one-standard-deviation below the mean), then poverty would decline 3.6 to 3.9 per cent each year. As table 5 indicates, a similar pattern emerges for the other spatial distribution variables; as s ₂, s ₃ and s ₄ increase, the poverty-reducing impact of income growth is diminished. A one-standard-deviation change (also 21 per cent) in the share of rural population located in less favored agricultural areas (s ₂) would mean that poverty would decline by only 1.8 to 1.9 per cent per year. A one-standard-deviation change in the share of rural population located on remote less favored agricultural land (s ₃), which is 8.4 per cent, would lead to a rate of poverty reduction of 3.5 to 4.7 per cent annually. Finally, a one-standard-deviation change in the share of rural population on less favored agricultural land located on remote land (s ₄) by 19 per cent causes the poverty-reducing impact of average income growth to be only 2.9 to 4 per cent per year.Footnote ¹⁴

In sum, our estimation results confirm that the spatial distribution of rural population in developing countries on less favored and remote agricultural land impacts overall poverty. However, the hypothesis that this spatial distribution has a direct influence on poverty is rejected. Instead, the location of rural population on less favored agricultural land and more remote areas impacts poverty by lowering the poverty-reducing impact of income growth. This attenuating effect of a higher initial spatial distribution can be sizeable, as the results in table 5 indicate.

7. Implications for climate change impacts on poverty

There is growing evidence that climate change and associated risks are likely to impact adversely the livelihoods of populations living in marginal and remote agricultural areas, both directly through causing declining agricultural productivity and income, or indirectly through affecting land and natural resource use (McSweeney, Reference McSweeney2005; Carter et al., Reference Carter, Little, Mogues and Negatu2007; Debela et al., Reference Debela, Shively, Angelsen and Wik2012; López-Feldman, Reference López-Feldman2014; Wunder et al., Reference Wunder, Börner, Shively and Wyman2014; Angelsen and Dokken, Reference Angelsen and Dokken2015; Hallegatte et al., Reference Hallegatte, Bangalore, Bonanigo, Fay, Kane, Narloch, Rozenberg, Treguer and Vogt-Schilb2015; Narloch and Bangalore, Reference Narloch and Bangalore2016; Robinson, Reference Robinson2016). An analysis of environmental reliance, poverty and climate vulnerability among more than 7,300 households in forest-adjacent communities in 24 developing countries found that the poor tend to live in the less favorable areas, generate 29 per cent of their income from environmental resources, and are more exposed to extreme and variable climate conditions (Angelsen and Dokken, Reference Angelsen and Dokken2015). The poor rural households from Sub-Saharan Africa located in less-favored arid areas already experience declining incomes directly from extreme climate conditions and from high forest loss, and face further loss of future forest benefits from climate change (Angelsen and Dokken, Reference Angelsen and Dokken2015). In rural Uganda, poorer households attempt to diversify their income sources from use of forests and outside employment; however, large negative shocks, such as droughts and other climate-related disasters force more reliance on forest resources, which leads to more land and natural resource degradation especially among those households with below-average and poor quality land holdings (Debela et al., Reference Debela, Shively, Angelsen and Wik2012). In Vietnam, poorer rural households are much more exposed to multiple environment risks, including climate change, and such risk induce not only lower consumption levels but also lower consumption growth over time (Narloch and Bangalore, Reference Narloch and Bangalore2016). Finally, the probability of participation in resource extraction increases considerably when these households experience climate-related impacts and other shocks to their agricultural systems (López-Feldman, Reference López-Feldman2014). Overall, cross-country evidence appears to confirm that for poor households in LFAA, ‘forests and other wildlands are ‘options of last resort’, which people only select as their primary safety net response when shocks are particularly severe and when, due to adverse household and village conditioning factors, they do not have any easier way out’ (Wunder et al., Reference Wunder, Börner, Shively and Wyman2014: S39).

Thus, the livelihoods of populations living in marginal and remote agricultural areas already highly dependent on agricultural land and resource commons, and their economic activities display low and even declining labor productivity (Barbier, Reference Barbier2010). Geographical isolation further raises substantially the costs of agricultural commerce and crop production in remote markets, distorts or insulates these markets from economy-wide policy changes, and thus discourages smallholder market participation and investment in improved farming systems and land management (Coxhead et al., Reference Coxhead, Shively and Shuai2002; González-Vega et al., Reference González-Vega, Rodríguez-Meza, Southgate and Maldonado2004; Holden et al., Reference Holden, Shiferaw and Pender2004; Shively and Fisher, Reference Shively and Fisher2004; Jansen et al., Reference Jansen, Rodriguez, Damon, Pender, Chenier and Schipper2006; Barrett, Reference Barrett2008; Narain et al., Reference Narain, Gupta and van't Veld2008; Ansoms and McKay, Reference Ansoms and McKay2010). Climate change impacts, such as drought, erosion, increases in salinity, and changes in precipitation, temperature and hydrology, may impact directly such households through causing declining agricultural productivity and income, or indirectly through causing more overuse and degradation of land and natural resources. Both impacts can lead to larger numbers of rural populations concentrated on less favorable agricultural lands and areas which, as our analysis indicates, will diminish the poverty-reducing impacts of growth in average incomes (see table 5).

The regional scale of climate change impacts may also influence the spatial distribution of rural populations in some developing countries. If changes in precipitation, hydrology, sea level rise, and the frequency of drought occur over significantly large areas, then the likely result is an increase in the unfavorable regions for agricultural production in developing countries (i.e., areas A and B in figure 1). Thus, the share of rural populations on less-favored agricultural lands could increase (Hallegatte et al., Reference Hallegatte, Bangalore, Bonanigo, Fay, Kane, Narloch, Rozenberg, Treguer and Vogt-Schilb2015, Reference Hallegatte, Vogt-Schilb, Bangalore and Rozenberg2017). A number of agricultural regions in the developing world have been identified as climate change ‘hotspots’ that are vulnerable to biophysical decline from climate-induced impacts, including North Africa, the Sahel, the Horn of Africa, parts of southern Africa, Central America, the Andes, western China, Central Asia, and India and South Asia (de Sherbinin, Reference de Sherbinin2014). From 2000 to 2030, agricultural land with poor biophysical conditions and productivity is expected to increase worldwide by 1–2.9 million ha annually, with climate change a major factor (Lambin and Meyfroidt, Reference Lambin and Meyfroidt2011). Case studies in China, Ethiopia, Mexico, Uganda, Rwanda, Chile, and Indonesia indicate declines in overall agricultural productivity due to climate change of around 3 to 7 per cent per year, an order of magnitude larger than the estimated cost of remediation (GCEC, 2014). Around 4 to 9 per cent of coastal agricultural areas in developing countries are likely to be exposed to salt water intrusion and storm surge damages caused by the accelerated sea-level rise accompanying climate change (Dasgupta et al., Reference Dasgupta, Laplante, Murray and Wheeler2011).

Already there is evidence that the concentration of rural population in less favorable agricultural regions has increased since 2000. For example, Barbier et al. (Reference Barbier, López and Hochard2016) find that around 1.5 billion rural people in 2010 lived on less favored agricultural land, which is a 14 per cent increase in the nearly 1.32 billion located on this land in 2000 (see table 1). If this trend continues, and is accelerated by the impacts of future climate change, it could possibly increase the share of rural populations located in unfavorable agricultural areas. As our analysis suggests, the poverty-reducing impacts of per capita income growth in developing countries will be further diminished.