Growth and Poverty Revisited: Why Don’t We See Poverty Convergence?

1. Keynote presentation for the Annual Conference of the Center for the Study of African Economies, Oxford University, March 2012 Growth and Poverty Revisited:Why Don’t We See Poverty Convergence? Martin Ravallion Development Research Group, World Bank

2. Two “stylized facts” about development • The advantage of backwardness—the mean convergence property whereby diminishing returns to aggregate capital imply higher growth rates in countries starting out with a low mean. • The advantage of growth—economic growth is distribution-neutral on average => economic growth reduces poverty. Strong form: “economic growth is the main driving force for poverty reduction.” Ample empirical support for both stylized facts in the literature

3. Stylized fact 1: The advantage of backwardness • Growth empirics: Conditional convergence has never been rejected to my knowledge, though differing speeds of convergence.

4. Stylized fact 1: The advantage of backwardness • Growth empirics: Conditional convergence never been rejected to my knowledge, though differing speeds of convergence. Annualized growth rate Initial mean Initial country characteristics

5. Stylized fact 2: The advantage of growth Economic growth is typically “pro-poor” Average slope = -2 (lower using NAS) Sizeable variance in impact of growth on poverty: A 1% rate of growth will bring anything from a modest drop in the poverty rate of 0.6% to a more dramatic 3.5% annual decline (95% CI).

6. An implication of these stylized facts has gone unnoticed: We should see poverty convergence • Taken together, these stylized facts imply that poorer countries will grow faster, which will give them higher rates of poverty reduction than less poor countries. • So we should see poverty convergence: a catching up process whereby poorer countries experience a higher (proportionate) rate of progress against poverty.

7. Indeed, we should see common speed of convergence in standard specifications • Poverty rates should converge at the same speed as the means in standard specifications in which the parameters of the dynamic processes for growth and poverty reduction are independent of the initial level of poverty. • Simple expository log-linear model: Growth model for the mean: + Poverty model: = Implied growth model for poverty: • The parameter determining the speed of convergence should be the same for the mean as the poverty measure

8. But we do not find any sign of poverty convergence in the data! • As this paper will show, there is no correlation across countries between the initial levels of poverty and subsequent proportionate rates of poverty reduction. • The overall incidence of poverty is falling in the developing world, but no faster in the poorest countries. Something is missing from our standard model of growth and poverty reduction.

9. What is missing from these two stylized facts of development? The parameters of these models of growth and poverty reduction must vary systematically with the initial level of poverty.

10. Outline • Theories of distribution-dependent growth • Past evidence on growth and initial distribution • Data and descriptive statistics • Testing the relevance of initial distribution to growth • Initial distribution and the effect of growth on poverty • Implications for poverty convergence • Implications for Africa and development policy

11. Theories of distribution-dependent growthbased on credit-market failures • Market failure attributed to information asymmetries, notably that lenders are imperfectly informed about borrowers. • Key analytic feature: a suitably nonlinear relationship between an individual’s initial wealth and her future wealth (the “recursion diagram”). • With diminishing marginal products of capital, the mean future wealth will be a quasi-concave function of the distribution of current wealth. • Thus higher current inequality implies lower future mean wealth at a given value of current mean wealth, i.e., lower growth. • Examples: Galor and Zeira (1993), Benabou (1996), Aghion and Bolton (1997) and Banerjee and Duflo (2003).

12. Nonlinear dynamics: Inequality handicaps growth Multiple equilibria Unique equilibrium 12

13. Dynamic implication of borrowing constraints: poverty also impedes growth • In a growth model with a borrowing constraint: • Those with sufficient wealth will reach their unconstrained optimum, equating the marginal product of capital with the interest rate. • But the “wealth poor,” for whom the borrowing constraint is binding, will not be able to do so. • Higher inequality in such an economy implies lower growth (e.g., Banerjee and Duflo, 2003). • However, essentially the same model also implies that higher current wealth poverty for a given mean also implies lower growth. • An unambiguously higher initial headcount index of poverty holding the initial mean constant implies a lower growth rate.

14. Other theories Other theoretical models => Poverty can impede growth and (hence) poverty reduction: • Inequality or poverty restricts efficiency-enhancing cooperation amongst people, such that public goods needed for growth are underprovided or efficiency-enhancing policy reforms are blocked (Bardhan et al., 2000). • Lasting (adverse) productivity effects of poor nutrition, esp., in childhood (Dasgupta and Ray, 1986; Cunha and Heckman, 2007). • Lopez and Servén (2009) introduce a subsistence consumption requirement into the utility function in the model of Aghion et al. (1999) and show that higher poverty incidence (failure to meet the subsistence requirement) implies lower growth. 14

15. High poverty can also be welfare inefficient through negative externalities of poverty • In marked contrast to relative deprivation theory, some evidence indicates that poor people benefit (objectively and subjectively) from having better off friends and neighbors. • Evidence for Malawi (Lokshin and Ravallion, 2010) • The dominance of the positive externality for the rural poor implies that there will be a socially sub-optimal incentive to take actions that increase one’s own income, given the spillover effect on others. • In other words, there will be too much poverty from an efficiency point of view.

16. Past evidence on growth and inequality • Empirical support for the view that a higher Gini index of inequality impedes growth; Alesina and Rodrik (1994), Persson and Tabellini (1994), Birdsall et al., (1995), Clarke (1995), Perotti (1996), Deininger and Squire (1998) and Knowles (2005) • However, not all the evidence has been supportive; also see Li and Zou (1999), Barro (2000) and Forbes (2000). • The main reason why these studies have been less supportive appears to be that they have allowed for country-level fixed effects. • Simulation studies: coefficients are heavily biased toward zero in fixed-effects growth regressions (Hauk and Wacziarg, 2009). • The significance of the Gini index in past studies may be an omitted variable bias. If poverty is the relevant variable, inequality could still be significant given that one also controls for the initial mean.

17. Mysterious control variables • The specification choices in past empirical work have lacked a clear theoretical justification. • Consider three popular predictors of growth, namely human development, the investment share, and financial development. • Basic schooling and health attainments (often significant in growth regressions) are potential channels linking initial distribution to growth; see original Galor and Zeria (1993) model. • Share of investment in GDP (robust predictors of growth rates) is one of the main channels through which distribution affects growth from the theoretical literature. • Private credit (as a share of GDP) has been used as a measure of “financial sector development” in explaining growth and poverty reduction (Beck et al., 2000, 2007). But theories based on borrowing constraints imply that the aggregate flow of credit in the economy depends on the initial distribution.

18. Data for this study • 92 developing countries with at least two national surveys since 1980 and earliest survey finds some households lived in poverty. • Transition economies of EECA mostly excluded. Spurious “convergence.” • Longest spell between two surveys for each country. • Both surveys used the same welfare indicator, either consumption or income per person, following standard measurement practices. • When both were available, consumption was generally preferred. Three-quarters of the spells use consumption. • Comparability problems between surveys remain, such as differences in recall periods and imputation/valuation methods. • Median year of first survey is 1991; median for second is 2004. Median interval is 13 years and varies from three to 27 years. • All changes between the surveys are annualized.

19. Data cont., • National accounts data were mapped as closely as possible to the survey dates, interpolating as need be. • All monetary measures are in constant 2005 prices (using country-specific Consumer Price Indices). • All international comparisons are at PPP using the results of the 2005 International Comparisons Program.

20. Measures of distribution • Poverty is mainly measured by the headcount index for $2.00 per day at 2005 PPP, which is the median poverty line amongst developing countries. • Also lower line of $1.25 a day (mean of poorest 15 countries) and a much higher line of $13 a day in 2005 (US line) • Inequality is measured by the usual Gini index. • Middle class: four measures: • The proportion of the population living in the interval $2 to $13 a day at 2005 purchasing power parity (PPP) (Ravallion, 2010). • Western middle class: Those living above US poverty line can be thought of as the “middle class” (and “rich”) by Western standards. • Share of middle three quintiles (Easterly) • “Miser index” (Lind and Moene): polarization between rich and poor (inverse measure of middle class).

21. Convergence? • The survey means exhibit convergence. But the poverty measures do not.

22. N Estimated convergence parameters Note: Controls included initial consumption per capita from the NAS, primary school enrollment rate, life expectancy at birth, and the price index of investment goods from Penn World Tables (6.2), which is a widely-used measure of market distortions; all three variables are matched as closely as possible to the date of the earliest survey.

23. Tests of robustness to measurement errors • Subsample of 70 countries with 3+ surveys. • Three further tests more robust to measurement errors: • Test 1: Estimated OLS trend regressed on log initial mean. • Test 2: Means from first two surveys, with growth rates from last two surveys. • Test 3: Initial values from first survey, with growth rates from last two. • In each case: convergence in means but not poverty measures. The two stylized facts are confirmed in these data, but no sign of poverty convergence. Why?

24. Poverty and growth • Benchmark regression: where Note: The regression is consistent with a derivative of current mean with respect to lagged mean that is less than unity, but fades toward zero at sufficiently long gaps between survey rounds.

25. Poverty and growth • Benchmark regression: where Growth rate Initial mean Initial poverty rate ($2/day) Note: The regression is consistent with a derivative of current mean with respect to lagged mean that is less than unity, but fades toward zero at sufficiently long gaps between survey rounds.

26. Alternative specifications

27. Using countries with 3+ surveys IVs for “initial” mean and poverty drawn from earlier surveys

28. Encompassing regressions

29. Initial distribution and growth elasticity of poverty reduction • In general, the growth elasticity of poverty reduction will depend on the initial distribution. • Past work has focused on inequality as the relevant aspect of initial distribution. • Mechanically, however, a high (low) initial poverty rate might be expected to imply a lower (higher) absolute elasticity. • If consumption is log-normally distributed and inequality is unchanged with growth in the mean then the (absolute) elasticity of the poverty rate to mean consumption will be strictly decreasing in the poverty rate as well as inequality.

30. Regressions for rate of poverty reduction Homogeneity tests passes => the relevant growth rate is the poverty-adjusted rate, as given by the growth rate times one minus the poverty rate.

31. Rate of poverty reduction is proportional to the distribution-corrected rate of growth

32. Poverty makes growth less pro-poor • The (absolute) growth elasticity of poverty reduction tends to be lower in countries with a higher initial poverty rate. • Poorer countries tend to experience lower proportionate effects on their poverty measures from any given rate of growth. • At an initial poverty rate of 10% (about one standard deviation below the mean) the elasticity is about -3 (using the IVE) while it falls to about -0.7 at a poverty rate of 80% (about one standard deviation above the mean). • The interaction effect with the poverty rate is stronger than that with the partial elasticity of poverty reduction derived analytically.

33. So why don’t we see poverty convergence? Preferred model: Poverty convergence elasticity:

34. To recap on the findings • There is an adverse effect on growth of high initial poverty at a given mean. • This is consistent with theoretical models of economic growth incorporating borrowing constraints. • Although, there are alternative explanations (e.g., nutrition and productivity) • A high initial incidence of poverty also entails a lower subsequent rate of progress against poverty at a given growth rate. • Thus we do not see poverty convergence, despite the fact that growth reduces poverty and mean living standards are converging amongst developing countries.

35. Implications

36. Implications for Africa? • In 2008, 48% of the population of Sub-Saharan Africa (SSA) lived below $1.25 a day at 2005 PPP; 69% below $2. • Highest $1.25 a day poverty rate in developing world; near tie with South Asia for highest $2 a day poverty rate. • Dummy variables for SSA are insignificant in these models. Africa is explained by the models. • Implication: High current poverty rates in SSA will make it harder to reduce poverty in the future: • Growth rates will be lower than otherwise expected • And the growth will have less impact on poverty; SSA would need an even higher growth rate than (say) China to attain the same pace of progress against poverty.

37. However, look at the pattern in 30 years agoPoverty incidence ($1.25 a day), 1981-2008 East and South Asia were able to overcome the disadvantage of high poverty. Encouraging signs that SSA is also doing so.

38. Implications for development policy? • For many poor countries, the growth advantage of starting out with a low mean ("conditional convergence") is lost due to high poverty. • Inequality can be a constraint on growth and (hence) poverty reduction. And high inequality can make growth less poverty reducing. • However, not all “inequalities” matter: The key aspect of initial distribution in poor countries is poverty itself, at a given initial mean. • High current inequality is only a handicap if it entails a high incidence of poverty at given mean consumption.

39. Policy implications cont., • Policies that act directly to reduce poverty can be seen as important elements of longer-term poverty reduction through economic growth. • Two types of policies: 1. Pro-poor redistributive policies(incl. social protection) can have an efficiency role as well their traditional equity role. 2. Policies that make markets work better for the poor, esp. credit, but also land. • In both cases, careful micro evaluation and monitoring will be crucial. • As will flexibility in responding to evidence on what works and what does not in specific contexts.

40. Thank you for your attention!