Dynamic medium term Granger causality between growth and poverty

Transcription

1 Dynamic medium term Granger causality between growth and poverty Salvador Pérez-Moreno Universidad de Málaga and Diana Weinhold LSE Draft, October 2012 Abstract This paper examines the causal relationship between growth and poverty reduction in developing between 1970 and For this purpose, we apply a modified form of traditional Granger causality tests to suit the short times series that are available, and use panel data model evaluation techniques to test the out-of-sample forecasting performance of competing models. We conclude that the evidence supports the hypothesis that growth causes unidirectional poverty reduction measured by the $1/day poverty rate in the period 70s-80s. These results are confirmed by country groups when we split the sample into low- and middle-income and into mid-high- and very-high-inequality. However, in the period 80s-90s economic growth does not causes poverty reduction growth in a Granger-causal fashion, except in low-income for the $1/day poverty rate. JEL classification: I32, O40, C23 Keywords: poverty, growth, causality 1

2 1. Introduction A wide range of empirical analyses on the relationship between aggregate economic growth and poverty have found that the growth is associated with poverty reduction. Ravallion and Chen (1997), Dollar and Kraay (2002), and Bourguignon (2003) all find that aggregate growth significantly reduces poverty, while Quah (2001) makes the important point that in the dynamics of income distributions, first moment effects (i.e. aggregate growth) historically dominate second moment effects (i.e. changes in income distribution) in determining the proportion of the population in poverty. Later work has corroborated these first-order earlier conclusions (Adams (2004), Fusu (2008)) and a number of papers have explored further nuances. For example, Bourguignon (2003) explores the heterogeneity of the growth elasticity of poverty, finding that it is a decreasing function of both income level and degree of income inequality in a country. There is much less consensus surrounding the empirical evidence for poverty having a causal impact on aggregate growth. Nevertheless, there are a number of theoretical arguments that link poverty to growth. Some of the proposed mechanisms include socio-political instability (Alesina and Perotti, 1996; Benhabib and Rustichini, 1996), credit constraints (Galor and Zeira, 1993; Aghion et al., 1999), taxation and redistribution (Alesina and Rodrick, 1994; Persson and Tabellini, 1994), and nutritional and associated cognitive deficits (Dasgupta, 1997, Ravallion, 1997a, Adelman and Hoddinott, 2001). Through such channels poverty may inhibit economic growth and, in turn, poverty reduction may support economic growth. Indeed, Lustig, Arias and Rigolini (2002) argue that the relationship between growth and poverty runs in both directions. They highlight multiple complementarities between growth and poverty reduction and point out that actions to reduce poverty can create virtuous cycles that raise economic growth, in turn reinforcing poverty reduction. However despite the significant amount of research in this area it has been difficult to empirically establish causality using traditional methods. Data on income distribution (poverty, inequality) indicators for developing are generally scarce and sporadic, generating series that are too short and not appropriate for the usual time series econometric approaches. At the same time no convincing instruments for either poverty or growth have been identified that would allow for instrumental variables estimation. We attempt to address this deficit by investigating whether growth and poverty are causally related in a Granger fashion by employing 2

3 a modified form of Granger causality testing originally suggested in Granger and Huang (1997) that is adapted specifically for the use of short time series panel data. The approach builds on the original insight of Granger (1969) that if a variable X Granger-causes some variable Y, then including information on past values of X in addition to past values of Y (plus any other relevant information) should improve the out-of-sample forecasting ability of Y compared to an otherwise identical model without the information on X. The remaining sections are structured as follows. Section 2 describes the data, section 3 discusses the methodology and the results, and final discussion is presented in section 4. Results and country tables are presented in section Data We use a panel of 52 developing for the years 1970, 1980, 1990 and 1998, with poverty data estimated by Sala-i-Martin (2002) for the conventional poverty lines of $1/day and $2/day in 1985 PPP rates, and data of GDP per capita (constant 1995 US$) from World Development Indicators. The sample consists of all developing for which Sala-i-Martin (2002) provide poverty data and the World Bank supplies GDP per capita data for all years considered. By regions, it consists of 21 from Sub-Saharan Africa, 21 from Latin America and the Caribbean, 6 from South Asia, and 4 from East Asia. In order to test for the relationship between growth and poverty reduction by using groups of with similar key characteristics, we divide economies into income groups according to 1998 GNI per capita, calculated using the World Bank Atlas method. In particular, low-income includes with 1998 GNI per capita of $760 or less, and middle-income includes with 1998 GNI per capita of $761 to $9360. In addition, as many of the theoretical mechanisms that link poverty causally to growth involve political economic relationships that depend on the degree of inequality as well as poverty, we separate the sample into two segments according to the level of inequality. Although most examined present considerable inequality levels, we differentiate between mid-high-inequality and very highinequality. This includes those that present an average Gini index of 0.50 or more for 3

4 the period , according to the data from World Development Indicators 1. The included by region, income level and inequality level are listed in Table Results and Discussion 3.1 Contemporaneous Correlation in the Cross Section Initially we consider the contemporaneous correlation between aggregate growth and poverty growth in our data. This approach allows us to create a benchmark result against which our dynamic analysis can be compared. Furthermore, by examining the effects of changes of poverty while controlling for the initial levels, we mitigate the influence of unobserved timeinvariant characteristics that could confound a pure levels-regression. Thus, to start with, we consider the following models: Gpov Ggdp it it = + β1 gdpit + β 2Lgpd it 1 + β 3Lpovit 1 α + ε = + β1 povit + β 2Lgpd it 1 + β 3Lpovit 1 α + ε it it where Gpov and Lpov are the growth rate and log-level of the respective poverty indicators, and Ggdp and Lgdp are the growth rate and log-level of GDP per capita, with level terms measured at the beginning of the period. In Tables 1A and 1B we consider the cross section correlation between the growth of GDP per capita and growth of $1/day and $2/day poverty rates for the full sample, as well as for the periods , , and separately, because of the existence of significant particularities in economic, political and social terms in developing for each decade. In Table 1A we present the results for the equation in which the growth of poverty is a function of the growth of GDP per capita, whilst the Table 1B shows the results for the equation in which the growth of GDP per capita is a function of the growth of poverty. Consistent with the existing literature, the contemporaneous analysis shows a strong, negative correlation between economic growth and poverty growth. The statistical significance of 1 There are five of the sample for which World Development Indicators do not provide Gini index data between 1970 and Thus, taking into consideration inequality data after 1998 and other additional information, we consider Barbados, China, India and Pakistan as middle high-inequality, and Gabon as a very highinequality country. 4

5 the poverty reducing effect of growth is weaker in the 1990s for $1/day poverty rates, however. These contemporaneous correlations presented in Tables 1A and 1B do not imply causality however; indeed, Lustig, Arias and Rigolini (2002) argue convincingly that growth and poverty are jointly determined, and if one direction of causality runs from poverty to growth then the model may be miss-specified and may suffer from an endogenous bias. We thus turn to the task of attempting to identify the causal channels between these variables Testing for Causality It has become common place in the development literature to test for causality using instrumental variables estimation (for a discussion and critique of this phenomenon see Deaton (2009)). There is no guarantee that good instruments can be found for many important questions, however; and as Deaton (2009) and others have pointed out, the use of poor quality instruments may be of dubious value. For many phenomena the judicious exploration of the dynamics of a time series relationship may provide a good alternative. In particular, in those cases in which it is less likely that expectations of the future may be a first-order determinant of contemporaneous behaviour we can employ concepts of Granger causality. We exploit the fact that the correctly specified version of a causal model will do much better forecasting out of sample than the misspecified version; we use dynamics to identify causality. As noted above, however, we do not have extensive annual time series data on poverty and inequality. Instead we use short panels in which each observation represents a 10-year average. While this reduces the number of observations compared to a (hypothetical) annual series, in fact there are several serious advantages as well. In the first place, it is unlikely that the periodicity of any causal relationship from poverty to growth (other than a tautological identity as both variables are characteristics of the underlying dynamic path of the income distribution) is as short as one year. Annual data is also subject to cyclical influences and noise that make detecting an underlying relationship especially tricky. Nevertheless using a short panel of with fewer observations may make it more difficult to detect any causal linkages. In effect, for us to find that there is a Granger-causal relationship from poverty to growth we must find that poverty rates and changes in poverty from ten years ago still help us to make better forecasts about current (average) economic growth than 5

6 a model which relies solely on past information about levels and growth of GDP alone (and vice versa). Thus a test for Granger causality using ten-year data creates a much higher hurdle than in-sample correlation for rejection of the null of no causality. If there is no causal relationship and the contemporaneous correlation is spurious for some reason, we will fail to reject the null hypothesis of no causality. If there is a causal relationship that acts only in the short run (for example, a reduction in poverty that increases/decreases growth for only one or two years, or vice versa) then we will also fail to reject the null hypothesis of no causality. The specification of the models is fairly straightforward. We control for initial log-levels of both the dependent and the independent variables in each model, thus controlling for all unobserved variables that determined GDP (or poverty) levels at that initial point in time. In as much as these variables are time-invariant, the initial level will capture their effect and reduce the chances of omitted variable and deep simultaneity biases. We further control for the unobservable time-invariant characteristics that could be driving the levels of both GDP and poverty rates (for example, deep institutional quality) by modelling our variables as growth rates (to eliminate the levels averages). Thus an identifying assumption here (for unbiased estimates of the coefficients) is that there are no omitted unobserved trends in the growth rates. However the overall approach we adopt is fairly robust to deviations from these assumptions as we do not rely on our coefficient estimates to identify causality but instead on the out-of-sample forecasting ability of each model. Finally, rather than relying on exogeneity of the regressors, we exploit on dynamics and the evaluation of out-of-sample predictions to identify causality. The fact that we use ten-year intervals works to our advantage here; in the case of growth and poverty we would need to see future growth rates ten years off exerting a causal effect on current rates of change in poverty (that contains unique informational content to the effect on current economic growth and level of GDP) to create a reverse causality bias. This is not impossible; current economic activity is often driven by expectations of future outcomes. However economic growth is very difficult to predict and we would expect much of the current adjustment to future expectations to be captured in the GDP levels and growth measures. In any case this is much less likely a problem than is faced in any contemporaneous analysis. In particular, let us consider the following four models: 6

7 r Gpov it = α + β 1 Gpov it 1 + β 2 Lpov it 1 + β 3 Ggdp it 1 + β 4 Lgdp it 1 + β r C i + ε it (1) r Gpov it = α + β 1 Gpov it 1 + β 2 Lpov it 1 + β r C i + ε it (2) r r Ggdp it = α + β 1 Ggdp it 1 + β 2 Lgdp it 1 + β 3 Gpov it 1 + β 4 Lpov it 1 + β r C i + ε it (3) r Ggdp it = α + β 1 Ggdp it 1 + β 2 Lgdp it 1 + β r C i + ε it (4) r r r where variables are as defined above, with the addition of the regional dummies r Ci for the regions Sub-Saharan Africa, Caribbean, and South Asia to control for potential regional fixed effects in growth rates of GDP and poverty. Models (1) and (2) are rival models of poverty growth rates, and models (3) and (4) are rival models of GDP per capita growth rates. Models (1) and (3) include information on past growth rates and levels of both variables in question, and models (2) and (4) only include lagged endogenous variables and the regional dummies. Thus, for instance, if model (1) can forecast growth rate of poverty more accurately than model (2) then information about past GDP per capita growth was important and we conclude that there is evidence of Granger causality from GDP to poverty. However, if model (2) forecasts better than model (1), or there is no difference, then we conclude that including information about GDP per capita growth does not help to predict poverty variations and therefore we fail to reject the null hypothesis of no causal relationship from growth to poverty. Following Granger and Huang (1997) and Weinhold and Reis (2001), we have adopted a sum-difference test that takes the following form. Consider the forecast errors η = from 1it η2it models (1) and (2) (or from models (3) and (4)) where i and t denote not-in-sample cross section and time periods. The null hypothesis is H η = η (5) : it 2it It then follows that if H 0 can be rejected, the model with the lowest forecast error variance should be accepted as being significantly superior to the competing model. For this purpose, Granger and Huang suggest constructing the following variables: 7

8 SUM i12 = η 1i + η 2i (6) DIFi 12 = η1 i η 2i (7) regression SUM Then a test of the null hypothesis is equivalent to a test of whether δ = 0 from the i12 = α + δ DIFF i 12 + ξi (8) In our case, we regress the model (8) using bootstrapped standard errors, which are more robust than traditional standard errors. An advantage of this approach for data like ours which quite limited in the time series (with only three time observations of the level variables and two observations of the growth rates for each country) is that we do not rely on precise estimates of the in sample parameters but rather on running a model competition in which all models are equally handicapped by the scarcity of time series observations. At the same time the ten-year periodicity of the data in turn implies that each observation embodies a lot of unique information and the subsequent time series variation is less likely to be driven by cyclical and short term shocks. Nevertheless, the short time series does preclude a traditional Granger-causality approach in which we predict out-of-sample across time. With our shorter panel data set we take advantage of the fact that we have variation across space as well as through time. In particular, we estimate the models on N-1 cross section observations and use the resulting coefficient estimates to generate a forecast of the dependent variable for the remaining (not-in-sample) unit. In this way, we systematically generate N different forecast errors for each model. Then, we test whether these forecast errors are statistically different from each other as described above. Tables 2A, 2B, 3A and 3B reflect the results of the modified form of the traditional Granger causality test discussed above. In particular, tables 2A and 3A present summary statistics of the mean squared forecasting error from the two competing models of poverty growth (models 1 and 2) and the two competing models of GDP per capita growth (models 3 and 4) for the periods and , respectively. In addition, tables 2B and 3B include the corresponding results of the sum-difference test using bootstrapped standard errors. 8

9 As can be observed in Tables 2A and 2B, the results obtained for reveal that the poverty growth model that includes information on former GDP per capita growth (model 1) performs better at out-of-sample forecasting than the model that excludes GDP per capita growth (model 2) when poverty is measured using the $1/day poverty rate. In other words, we reject the null of no causality and find that aggregate economic growth Granger-causes poverty growth. This result is consistent across country groups when we split the sample into low- and middleincome and into mid-high- and very-high-inequality. However, for the $2/day poverty rate, the poverty growth model that excludes GDP per capita growth (model 2) performs better at out-of-sample forecasting, thus in this case we cannot reject the null hypothesis of no causality from growth to poverty. Tables 3A and 3B present the results from an analysis of We find that growth Granger-causes poverty reduction measured by the $1/day poverty rate in low-income, while among higher income we fail to find Granger causality. In other words, for higher income the poverty growth models that include information on former GDP per capita growth (model 1) perform worse at out-of-sample forecasting than the model that excludes GDP per capita growth (model 2), or the difference between the better performance of model 1 as compared to that of model 2 is not statistically significant. Some authors have found that higher initial inequality tends to reduce the impact of growth on poverty (see, e. g., Ravallion, 1997b; Fusu, 2008). Our analysis is not focused on parameter estimation per se and thus is not suited to testing for differences in the poverty elasticity of growth. However, we do not find that high inequality breaks the link; indeed our results suggest that growth Granger-causes poverty reduction in both mid-high- and very-high-inequality as well. Finally, none of our results suggest any Granger-causal relationship from poverty reduction to aggregate economic growth. In particular, while tables 2A and 3A show that in some cases the GDP per capita growth model that includes information on former poverty growth (model 3) performs slightly better at out-of-sample forecasting than the model that excludes poverty growth (model 4), the difference is never statistically significant. 9

10 4. Conclusions Although the relationship between economic growth and poverty has been widely examined in recent literature, most empirical studies generally focus on the study of an in-sample contemporaneous correlation, often assuming that the only direction of causality is from growth to poverty. In this paper we carry out a bi-directional causal analysis of the relationship between growth and poverty reduction in developing and exploit a combination of dynamics and out-of-sample forecasting to identify causality. To this end, we apply a modified form of traditional Granger causality tests to suit the short times series that are available and use panel data model evaluation techniques to test the out-of-sample forecasting performance of competing models. The results for show that forecasts of future poverty growth in all country groups (measured by the $1/day poverty rate) are significantly improved when information on past economic growth is included. Moreover, information on poverty growth does not improve forecasts of future economic growth. Thus our analysis suggests, consistent with existing literature, that during this period growth causes unidirectional poverty reduction measured by the $1/day poverty rate in a Granger-causal fashion. However our results are not consistent for poverty rates measured by the $2/day criteria, suggesting that the estimated growth-poverty relationships in the literature may be sensitive to alternative measures of poverty. For the period we observe that the causal relationship in the direction from economic growth to poverty reduction, measured by the $1/day poverty rate, is only statistically significant in low-income, a result that is consistent with Bourguignon's (2003) observation that the growth elasticity of poverty is a function of income level. Again, we find that when poverty is measured by the $2/day poverty rate, we cannot reject the null hypothesis of no causality, nor can we reject no causality from changes in poverty to growth. Overall our results are mostly consistent with the conclusions reached by most previous empirical studies in that we find that aggregate economic growth seems to be effective in contributing to longer-range extreme poverty reduction, especially in the first stages of economic development. While our econometric 'hurdle' for finding causality is quite high, making it much more difficult for us to reject the null of no causality than in most traditional econometric approaches, the failure to find causality from growth to poverty reduction measured at the $2/day 10

11 level, as well as the failure to find this relationship in higher-income developing in the later sample period is indicative that the growth elasticity of poverty is heterogeneous across income levels and time periods. 11

12 5. Tables and Results TABLE 1A: Contemporaneous Cross-Section Analysis. Sala-i-Martin s Poverty Data and GDP per capita (constant 1995 US$) a Dependent variable: Gpov it Full sample ( ) Poverty rate $1/day t=1980 ( ) t=1990 ( ) t=1998 ( ) Full sample ( ) (1.60) (-6.46)** (-2.12)* (-0.64) Poverty rate $2/day t=1980 ( ) t=1990 ( ) t=1998 ( ) Constant (3.08)** (2.40)* (4.40)** (0.86) (1.08) (1.17) (-2.96)** Ggdp it (-4.99)** (-3.50)** (-2.80)** (-1.96) (-2.98)** (-3.58)** (-3.13)** Lgdp it 1 (-3.22)** (-2.86)** (-4.05)** (-0.96) (-1.68) (-1.33) (1.89) Lpov it 1 (-2.69)** (-0.23) (-5.11)** (-1.05) (-0.38) (-0.79) (4.43)** 2 R N a Estimates were obtained using ordinary least squares, with the dependent variable being difference in log of the respective poverty measure. Heteroskedasticity-consistent t-statistics are shown in parentheses. * Significant at the 0.05 level. ** Significant at the 0.01 level. 12

13 TABLE 1B: Contemporaneous Cross-Section Analysis. Sala-i-Martin s Poverty Data and GDP per capita (constant 1995 US$) b Dependent variable: Ggdp it Full sample ( ) Poverty rate $1/day t=1980 ( ) t=1990 ( ) t=1998 ( ) Full sample ( ) (1.33) (-7.80)** (-1.31) (0.06) Poverty rate $2/day t=1980 ( ) t=1990 ( ) t=1998 ( ) Constant (1.14) (-0.44) (2.27)* (0.38) (0.76) (2.01)* (-2.13)* Gpov it (-7.21)** (-1.35) (-5.83)** (-2.52)* (-2.00) (-3.01)** (-5.17)** Lgdp , it 1 (-0.66) (0.76) (-2.22)* (0.09) (-0.84) (-2.23)* (1.97) Lpov it 1 (-0.51) (1.67) (-1.75) (-5.38)** (0.14) (-0.85) (2.53)* 2 R N b Estimates were obtained using ordinary least squares, with the dependent variable being difference in log of GDP per capita. Heteroskedasticity-consistent t-statistics are shown in parentheses. * Significant at the 0.05 level. ** Significant at the 0.01 level. 13

14 Table 2A: Mean-squared forecast error (1970s-1980s) c Model 1 Model 2 Model 3 Model 4 Poverty headcount ratio at $1 (PPP) a day Full sample N=52 Lowincome Middleincome Mid-highinequality N=31 Very highinequality N=21 Poverty headcount ratio at $2 a day (PPP) Full sample N=52 Lowincome (2.7037) (3.6665) (4.0296) (3.4171) (1.6091) (0.7623) (3.4006) (4.1941) (4.6994) (4.6016) (2.8090) (0.6360) (0.2008) (0.1783) (0.2978) (0.2348) (0.2240) (0.2345) (0.2019) (0.1666) (0.3004) (0.2268) (0.2005) (0.2019) c Out-of-sample causality test. Standard deviations of forecast error are shown in parentheses. Model 1: a model of poverty growth including GDP per capita growth Model 2: a model of poverty growth excluding GDP per capita growth Model 3: a model of GDP per capita growth including poverty growth Model 4: a model of GDP per capita growth excluding poverty growth (0.4184) (0.2998) (0.1986) (0.1666) Middleincome (1.3989) (1.1165) (0.2852) (0.3004) Midhighinequality N= (0.7197) (0.5684) (0.3380) (0.2268) Very highinequality N= (0.3843) (0.6439) (0.2049) (0.2005) 14

15 Table 2B: Sum-difference test results for poverty growth (models 1 and 2) and GDP per capita growth (models 3 and 4) (1970s-1980s) d Models 1 and 2 Poverty headcount ratio at $1 (PPP) a day Full sample N=52 Lowincome Middleincome Midhighinequality N= (-0.04) (-2.17)** (0.12) (0.19) Very highinequality N= (0.78) (-2.53)** (0.05) (0.87) Poverty headcount ratio at $2 a day (PPP) Full sample N=52 Lowincome Middleincome Constant (0.07) (-0.02) (-0.15) (-0.16) (0.12) (-0.56) DIFFij (-2.21)** (-1.99)** (-1.78) (1.53) (0.66) (1.82) Constant (0.14) (0.14) (0.8300) (0.09) (0.14) (0.29) DIFFij (-0.16) (1.63) (-0.17) (0.71) (0.54) (-0.76) d T-statistics shown in parentheses based on bootstrapped standard errors obtained from 500 bootstrap replications. ** Significant at the 0.05 level Models 3 and 4 Midhighinequality N= (0.06) (2.35) (-0.10) (1.35) Very highinequality N= (0.23) (-1.93) (-0.02) (0.28) 15

16 Table 3A: Mean-squared forecast error (1980s-1990s) e Poverty headcount ratio at $1 (PPP) a day Full sample N=52 Lowincome Middleincome Mid-highinequality N=31 Very highinequality N=21 Poverty headcount ratio at $2 a day (PPP) Full sample N=52 Lowincome Model (2.1181) (0.9638) (3.7575) (2.3239) (3.0380) (0.4019) (0.1545) Model (2.1377) (1.3334) (3.6605) (2.2825) (2.4158) (0.4167) (0.1594) Model (0.1906) (0.3870) (0.1333) (0.2236) (0.2209) (0.1798) (0.3233) Model (0.1785) (0.2365) (0.1522) (0.1901) (0.2116) (0.1785) (0.2365) e Out-of-sample causality test. Standard deviations of forecast error are shown in parentheses. Model 1: a model of poverty growth including GDP per capita growth Model 2: a model of poverty growth excluding GDP per capita growth Model 3: a model of GDP per capita growth including poverty growth Model 4: a model of GDP per capita growth excluding poverty growth Middleincome (0.5278) (0.5586) (0.1553) (0.1522) Mid-highinequality N= (0.4087) (0.3843) (0.1962) (0.1901) Very highinequality N= (0.6265) (0.4521) (0.2294) (0.2116) 16

17 Table 3B: Sum-difference test results for poverty growth (models 1 and 2) and GDP per capita growth (models 3 and 4) (1980s-1990s) f Poverty headcount ratio at $1 (PPP) a day Full sample N=52 Lowincome Middleincome Midhighinequality N= (0.13) (0.63) (0.04) (3.83) Very highinequality N= (-0.20) (2.31) (0.16) (0.32) Poverty headcount ratio at $2 a day (PPP) Full sample N=52 Lowincome Middleincome Mid-highinequality N= Constant Models (0.11) (-0.03) (0.04) (-0.20) (0.01) (-0.37) (-0.18) 1 and DIFFij (-0.46) (-3.52)** (0.65) (-0.60) (-0.60) (-0.60) (2.57) Constant Models (0.12) (0.03) (0.16) (0.02) (0.02) (0.10) (-0.09) 3 and DIFFij (3.44) (1.35) (-1.51) (0.42) (1.07) (0.45) (0.81) f T-statistics shown in parentheses based on bootstrapped standard errors obtained from 500 bootstrap replications. ** Significant at the 0.05 level Very highinequality N= (0.05) (2.80) (0.04) (0.42) 17

18 Table 4. Countries by region, income level and inequality level Country Region Income Level Inequality Level Bangladesh South Asia Low Middle high Barbados Caribbean Middle Middle high Bolivia Caribbean Middle Very high Botswana Sub-Saharan Africa Middle Very high Brazil Caribbean Middle Very high Burkina Faso Sub-Saharan Africa Low Middle high Burundi Sub-Saharan Africa Low Middle high Central African Rep. Sub-Saharan Africa Low Very high Chile Caribbean Middle Very high China East Asia Low Middle high Colombia Caribbean Middle Very high Costa Rica Caribbean Middle Middle high Cote d'ivoire Sub-Saharan Africa Low Middle high Dominican Republic Caribbean Middle Middle high Very high Ecuador Caribbean Middle Very high El Salvador Caribbean Middle Gabon Sub-Saharan Africa Middle Very high Gambia, The Sub-Saharan Africa Low Very high Ghana Sub-Saharan Africa Low Middle high Guatemala Caribbean Middle Very high Guinea-Bissau Sub-Saharan Africa Low Middle high Guyana Caribbean Middle Very high Honduras Caribbean Low Very high India South Asia Low Middle high Indonesia East Asia Low Middle high 18

19 Table 4. (cont.) Jamaica Caribbean Middle Middle high Kenya Sub-Saharan Africa Low Middle high Lesotho Sub-Saharan Africa Low Very high Madagascar Sub-Saharan Africa Low Middle high Malaysia South Asia Middle Middle high Mali Sub-Saharan Africa Low Very high Mexico Caribbean Middle Middle high Nepal South Asia Low Middle high Caribbean Low Middle high Nicaragua Niger Sub-Saharan Africa Low Middle high Nigeria Sub-Saharan Africa Low Middle high Pakistan South Asia Low Middle high Panama Paraguay Caribbean Middle Very high Caribbean Middle Very high Caribbean Middle Middle high Peru Philippines East Asia Middle Middle high Rwanda Sub-Saharan Africa Low Middle high Senegal Sub-Saharan Africa Low Middle high Sierra Leone Sub-Saharan Africa Low Very high South Africa Sub-Saharan Africa Middle Very high Sri Lanka South Asia Middle Middle high Thailand East Asia Middle Middle high Trinidad and Tobago Uruguay Caribbean Middle Middle high Caribbean Middle Middle high Caribbean Middle Middle high Venezuela Zambia Sub-Saharan Africa Low Very high Zimbabwe Sub-Saharan Africa Low Very high 19

20 References Adams, R. H. (2004): Economic Growth, Inequality and Poverty: Estimating the Growth Elasticity of Poverty. World Development, 32:12, Adelman, H. and J. Hoddinott (2001): Assessing the Impact of Early Childhood Malnutrition on Academic Achievement and Adolescent Height in Rural Zimbabwe. World Bank, mimeo. Aghion, P., Caroli, E. and C. Garcia-Penalosa (1999): Inequality and Economic Growth: The Perspective of the New Growth Theories, Journal of Economic Literature, 37, Alesina, A. and D. Rodrick (1994): Distributive Policies and Economic Growth, Quaternary Journal of Economics, 109, Alesina, A. and R. Perotti (1996): Income Distribution, Political Instability, and Investment, European Economic Review, 40, Benhabib, J. and A. Rustichini (1996): Social conflict and growth. Journal of Economic Growth, 1:1, Bourguignon, F. (2003): The growth elasticity of poverty reduction: Explaining heterogeneity across and time periods. In R. Eicher and S. Turnovsky (eds.): Inequality and growth. MIT Press, Cambridge, MA. Dasgupta, P. (1997): Nutritional Status, the Capacity for Work, and Poverty Traps. Journal of Econometrics, 77:1, Deaton, Angus, Instruments of development: Randomization in the tropics, and the search for the elusive keys to economic development, NBER Working Paper No , January 2009 Dollar D. and A. Kray (2002): Growth is Good for the Poor, Journal of Economic Growth, 7, Fosu, A.K. (2008): Inequality and the growth-poverty nexus: Specification empirics using African data. Applied Economics Letters, 15:7, Galor O. and J. Zeira (1993): Income Distribution and Macroeconomics, Review of Economic Studies, 60:1, Granger, C. W. J. (1969): Investigating Causal Relations by Econometric Models and Cross-Spectral Methods, Econometrica, 37,

21 Granger, CliveW.J.and Ling LingHuang Evaluation of PanelDataModels:Some Suggestions from Time Series, Department of Economics, University of California-San Diego. Lustig, N., O. Arias. and J. Rigolini (2002): Poverty Reduction and Economic Growth: a Two-Way Causality. Sustainable Development Department Technical Paper Series. Inter-American Development Bank. Persson, T. and G. Tabellini (1994): Is Inequality Harmful for Growth?, American Economic Review, 84, Quah, Danny (2001), Some Simple Arithmetic on How Income Inequality and Economic Growth Matter, Draft, London School of Economics Ravallion, M. (1997a): Famines and Economics. Journal of Economic Literature, 35:3, Ravallion, M. (1997b): Can high-inequality developing escape absolute poverty?. Economic Letters, 56:1, Ravallion, M. and S. Chen (1997): What Can New Survey Data Tell Us about Recent Changes in Distribution and Poverty?, World Bank Economic Review, 11:2, Sala-i-Martin, X. (2002): The World Distribution of Income (Estimated from Individual Country Distributions)', Working Paper no. 8933, National Bureau of Economic Research. Weinhold, D. and E. Reis (2001): Model Evaluation and Testing for Causality in Short Panels: The Case of Infrastructure Provision and Population Growth in the Brazilian Amazon, Journal of Regional Science, 41:4,