Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Is the Structure of Growth Different in Sub-Saharan Africa?

Is the Structure of Growth Different in Sub-Saharan Africa? Abstract Taking advantage of the 2014 issue of the Povcalnet data and building on earlier works, this paper provides an anatomy of growth in sub-Saharan Africa (SSA) and compares it with that of the developing world (including or excluding SSA) during the three decades between the early 1980s and the early 2010s. We examined both the impact of the pattern of growth on poverty and inequality (the growth–inequality–poverty (G–I–P) nexus); and the less studied reverse causal chain emanating from poverty to subsequent inequality and growth (the poverty–inequality–growth (P–I–G) nexus). For the G–I–P nexus, we found that poverty reduction is becoming more responsive to income growth and improvement in inequality in SSA in the post-2007 years; though the responsiveness remains smaller in SSA than in the developing world (with or without SSA) throughout the three decades examined. For the P–I–G nexus, we found that SSA again differs from the rest of the developing world in that SSA countries with the highest initial poverty incidence appeared to grow subsequently faster—leading to poverty convergence; while the developing world (with or without SSA) experienced a lack of poverty convergence during the same time period. We hypothesise that the main cause of poverty convergence in SSA during 1978–2007 might be due to anti-poverty interventions by governments and foreign public and private aid inversely proportional to the depth of poverty. 1. Introduction The questions at the heart of this paper are, first, whether the structure of growth in sub-Saharan Africa (SSA) is different from that of the developing world (including or excluding SSA) and, second, whether the recent pattern of growth in SSA is different from what it was prior to the beginning of this millennium. We hope to show that both of these questions can be answered affirmatively and we attempt to provide explanations for those differences. These questions are particularly relevant in the light of the extensive literature prior to the new millennium that argued that there were some structural factors such as ethno-linguistic heterogeneity, the peculiar historical origins of African (artificial) states, and endowment constraints such as land-lockedness and tropical climate that contributed to an environment making the growth pattern in SSA different from that in the rest of the world. Even controlling for those obstacles, researchers running international growth regressions almost invariably came up with an ‘African dummy’ to account (in a residual and unexplained way) for the stagnating growth in the region.1 In retrospect, it can be argued that it is the poor quality of governance that was the main culprit for the dismal performance of SSA economies before 2000 (Englebert, 2000; Thorbecke and Ouyang, 2016). By the structure of growth, we mean the dynamic interaction among growth, inequality and poverty (G–I–P) during the process of economic development. A major factor influencing this interaction is the speed and form of the structural transformation. Since around 2000, the pace of growth in SSA has undergone a quantum jump and the pattern of growth has become more inclusive. The structural transformation that was typically flawed in many SSA countries until recently is now more successful. The agricultural sector that in the past was, at best, ignored by the policymakers and, at worst, mercilessly exploited is now starting to get the attention it deserves. Workers released from agriculture move into more productive jobs in other sectors. For thirteen out of fourteen SSA countries for which at least two annual observations were available between 2000 and around 2010, the structural transformation was successful in the sense that the fall in the share of the labour force in agriculture was associated with an increase in per capita GDP (Thorbecke, 2014). Further evidence of progress towards a more inclusive growth pattern is that non-monetary dimensions of well-being improved as well. The Human Development Report (UNDP, 2014) indicates that out of the fourteen best performers as measured by the annual growth rate of the Human Development Index between 2000 and 2013, eleven were from SSA. Starting from a low baseline, many African countries have enjoyed significant improvements in infant mortality, health and school enrolment. The current spell of growth in SSA has been influenced by a number of factors. Among the endogenous factors, at least partially under the control of African states, are (i) the improved treatment of agriculture and, more generally, the overall improvement in the quality of governance; and (ii) the appearance of a middle class. Among the most important exogenous factors are (i) the large jump in the flow of direct foreign investment into the subcontinent and (ii) high global commodity prices and consequent favourable terms of trade (Thorbecke, 2014). Hence our conjectures, that will be subsequently tested, are that the current structure of growth in SSA is significantly different from that prevailing during the period 1960–2000 and also different from the global pattern of growth. In order to test those conjectures, the anatomy of growth, and more particularly the interrelationship among G–I–P, has to be analysed. This is best done by focusing on the G–I–P interrelationship. As these three concepts are not only interrelated but also jointly determined, one has to be careful in any attempt to infer unidirectional causality among them. As will be described below, one can best imagine a G–I–P triangle with influence travelling in both directions. In analysing the transmission of influence in such a socio-economic system, one can start at any one of the three corners. Traditionally, development economics focused on the directional effects emanating from income growth to inequality and poverty. In the remainder of this paper, we refer to this case as the G–I–P nexus. It is only recently that the lens has been placed on the reverse impact of poverty on growth and inequality which we refer to as the poverty–inequality–growth (P–I–G) nexus. The rest of the paper is organised as follows: Section 2 presents and analyses the G–I–P nexus and each of the links in the transmission of influence originating with growth. The impact on inequality and poverty of different structures of growth, such as pro-poor growth, exclusive and inclusive growth, is explored. A case is made that recent SSA growth has become more inclusive. Section 3 focuses on the P–I–G nexus and reviews recent research findings that high initial poverty and inequality levels affect subsequent growth negatively. Section 4 uses two different data sets to estimate the global and SSA structural coefficients of the G–I–P nexus based on Bourguignon's (2003) identity model. Our results confirm our conjectures that the historical growth pattern in SSA is different from that of the developing world (including or excluding SSA); and that the recent SSA growth spell is different from that prevailing up to the mid 2000s. Section 5, in turn, estimates the structural coefficients of the P–I–G nexus and finds that, in contrast with the results, Ravallion (2012) obtained from a globally representative sample of about hundred developing countries, high initial poverty does not appear to slow down subsequent growth in the African subcontinent. Finally, Section 6 concludes. 2. The impact of growth on inequality and poverty The chain of influence linking growth to poverty has been thoroughly researched and is relatively clear and well understood. It is based on the G–I–P interrelationship. The pace of growth and its pattern, in any given country, are determined by the forces of globalisation and the development strategy that is being followed. The process of globalisation is largely exogenous (outside the control of the State), while the development strategy is at least partially endogenous (under the control of the State). Depending on the specific initial conditions prevailing in a country, the combined effects of globalisation and the development strategy will give rise to, and result in different speeds of growth and different structures of growth from exclusive to inclusive. The less unequal the income distribution and the more inclusive the pattern of growth are, the greater the impact of growth will be on poverty reduction. By now there is a rich literature on pro-poor growth. There are essentially two approaches and definitions. The relative definition of pro-poor growth is that the poor benefit proportionately more than the non-poor from the prevailing growth, which implies a fall in income inequality. The absolute definition only requires that the poor benefit from growth.2 Hence, under this approach inequality can increase and still be compatible with pro-poor growth. More recently, inclusive growth has become the new paradigm adopted by the development community. The African Development Bank (2012) defines inclusive growth as ‘economic growth that results in a wider access to sustainable socio-economic opportunities for a broader number of people, regions or countries, while protecting the vulnerable, all being done in an environment of fairness, equal justice, and political plurality’ (p. 2). The inclusive growth approach takes a longer term perspective as the focus is on productive employment as the main instrument, rather than on direct income redistribution, as a means of increasing incomes for excluded groups. Unlike the pro-poor growth agenda that focuses mainly on the welfare of the poor, inclusive growth is concerned with opportunities in the labour force for poor and middle class alike. Figure 1 presents schematically the G–I–P nexus. Each of the links in the causal chain plays a role in the transmission of the combined effects of globalisation (e.g., through trade and foreign direct investment) and the development strategy (e.g., country-specific policies and institutions)—as prime movers. These links have been described in detail in Thorbecke (2015) and are only very briefly reviewed here. Figure 1: View largeDownload slide Globalisation and Development Strategy and Interrelationship Among G–I–P. Figure 1: View largeDownload slide Globalisation and Development Strategy and Interrelationship Among G–I–P. Starting, first, on the left-hand side of Figure 1, the two prime movers affect the pattern of growth in a given economy. In SSA, before around 2000, most countries stagnated and the structure of growth tended to be exclusive and contributing to a more unequal income distribution. In contrast, the present growth spell is characterised by a quantum jump in the pace of growth and some evidence that the pattern had become more inclusive. Hence, the G–P link (the upper left arrow in Figure 1), that in the pre-2000 era, was very weak, in the sense that the growth elasticity of poverty reduction in SSA was only about half that in other developing regions, in the recent, spell has become stronger resulting in a significant reduction in poverty. Income inequality is still relatively very high in the African subcontinent—although it is falling in a number of countries. The I–P link (lower left arrow) reveals the role of inequality as a filter between growth and poverty reduction. A more inclusive pattern of growth can keep inequality in check with favourable consequent poverty outcomes. Next, the I–G link is subject to two conflicting theoretical approaches. The Neo-classical theory argues that an uneven income distribution is a pre-condition to growth as the rich have a higher marginal propensity to save than the poor. Hence, for any given total GDP, an unequal income distribution will generate a larger flow of savings cum investment. On the other hand, the New Political Economy of Development argues convincingly that greater income inequality is likely to dampen growth through a variety of channels, such as the diffusion of political and social instability, unproductive rent-seeking activities and increased insecurity of property rights. We subsequently explore this proposition and find some support for it. More specifically in Section 4, we found that the impact of reducing high initial inequality through income growth has become more important for effective poverty reduction in the current SSA growth period. The final link appearing in Figure 1 is the G–I link, which is still referred to in the literature as ‘Kuznets’ Law’, i.e. that at very low levels of development, growth will be un-equalising up to a threshold per capita income amount and beyond that equalising. The present consensus is that this ‘law’ does not hold as a generalisation and lost its immutability. The specific initial conditions and policies of an economy largely determine the impact of growth on inequality. One of the earliest attempts to formalise the interrelationship among the G–I–P nexus is the identity model built by Bourguignon (2003). The identity model assumes a log-normal income distribution and can be used to explain the heterogeneity of G, I, P outcomes depending on country-specific conditions and more specifically the wide range of growth elasticities of poverty reduction observed among countries. Essentially, the dependent variable is the growth in poverty expressed as a function of initial values and growth rates of income and inequality, the ratio of the initial ratio of the poverty line to average income and some interacting terms. Even though, one cannot make any causal inferences on the basis of an identity model, the Bourguignon model has been used to shed light on the relative roles and importance of income and inequality in reducing poverty in different regions and countries. Fosu (2015) applied the identity model to a sample of some hundred developing countries using observations from growth spells between 1981 and 2007. He concluded that ‘Viewed within a global context … the relatively low levels of income appear to be a major factor in inhibiting the effectiveness of income and inequality improvements in producing poverty reduction in SSA countries generally’ (p. 56). In summary, the impact of growth on poverty is relatively clear and well understood. In contrast the reverse influence from poverty and inequality to subsequent growth has only recently been analysed and identified as a potentially important transmission mechanism. A number of empirical studies have thrown light on the major effect initial poverty (and inequality) can have on future growth. This is the theme of the next section. 3. The impact of poverty and inequality on subsequent growth In the past, this link tended to be dismissed on the ground that any policies directed to reducing poverty detracted from growth. This was based on the firm belief of a trade-off between equity and efficiency. Any measure to reduce poverty today would reduce future growth. A number of recent empirical studies have questioned this view and thrown light on the major effect high initial poverty (and inequality) can have on future growth. An early and path-breaking study (Perry et al., 2006) made a case for a pro-growth poverty reduction strategy on the ground that there are multiple channels through which the existence of poverty acts as a major obstacle to growth. Examples of such channels and poverty traps are that poor people (i) have limited access to credit and financial markets, which seals them off potentially profitable and productive investment opportunities; (ii) often suffer from ill health and malnutrition that affects their productivity; (iii) attend low-quality schools that constrain their human capital. Careful econometric work by Lopez and Servén (2009) provided empirical support for the contention that high poverty can be a major obstacle to growth. They used a standard growth model augmented to include a poverty measure among the explanatory variables and controlled for other factors. Estimating the resulting specification on a large country panel data set using a generalised method of moments (GMM) approach to control for endogeneity of other regressors, the authors found that poverty has a negative impact on growth that is significant both statistically and economically. On average, a 10% increase in poverty reduces annual growth by 1%. The mechanism retarding growth in countries with high initial poverty was found to be the deterrence of investment, especially when the degree of financial development is limited. The next important empirical case that high poverty incidence could reduce subsequent growth was made by Ravallion (2012). In an attempt to explain why poverty convergence is not occurring worldwide, he used a sample of growth spells from almost hundred developing countries covering the period from 1978 to 2007. Among his main findings are that (i) high initial poverty rates have sizeable negative impacts on the growth rate; (ii) it is high poverty, not inequality per se, that retards growth; (iii) to the extent that higher overall inequality comes with higher poverty at a given mean, it yields lower growth rates and (iv) the growth elasticity of poverty reduction tends to be smaller in countries with a higher initial poverty rate. The immediate question that is elicited by these results is that of endogeneity between growth and poverty. Slow growth contributes to (if not causes) poverty. No wonder then that initial high poverty would be associated with low growth. However, Ravallion (2012) rules out this interpretation by making clear that ‘We see that the finding that a higher initial poverty rate implies a lower subsequent growth rate in the mean (at given initial mean) is robust to allowing for the possible endogeneity of the initial mean and initial poverty rate’ (p. 514). The common theme of the above studies is that high poverty today dampens subsequent growth. Yet, surprisingly, in Section 5 where we apply the Ravallion's (2012) model to his SSA sample of 28 countries, we find little evidence of any association between initial poverty levels and subsequent growth rates—contrary to the global pattern showing that high initial poverty was positively and significantly correlated with lower subsequent growth (demonstrating a lack of poverty convergence). Figure 2 captures the comparative static effects over two time periods (t) and (t + 1) and shows graphically the connection between the traditional G–I–P nexus and the P–G–I nexus. The left-hand side of Figure 2 reproduces the G–I–P nexus as it appears in Figure 1, while the right-hand side of Figure 2 indicates the channels through which the initial incidence of poverty at time (t) influences subsequent growth and inequality at time (t + 1). Figure 2: View largeDownload slide The G–I–P Nexus at Time t and the Reverse P–G–I Nexus at Time (t + 1). Figure 2: View largeDownload slide The G–I–P Nexus at Time t and the Reverse P–G–I Nexus at Time (t + 1). In the next section, we attempt to test whether the SSA growth structure is different from the global structure. We use the Bourguignon identity model to estimate the structural coefficients of the G–I–P nexus for the SSA sample and compare the SSA estimates with those derived from the global data sets. 4. Estimating the SSA structural coefficients of the G–I–P nexus 4.1 Econometric model To explore the structural coefficients of the G–I–P nexus in SSA, we used the identity model initially proposed by Bourguignon (2003) and subsequently developed by Fosu (2009, 2011, 2015) into the following log-linear model3:   p=b1+b2y+b3yG′+b4y(Z/Y)+b5g+b6gG′+b7g(Z/Y)+b8G′+b9(Z/Y), (1) where p is the annualised change in log poverty rate; y is the annualised growth rate of log per capita income; g is annualised change in log Gini coefficient; G′ is the log of initial Gini; (Z/Y) is the ratio of the poverty line Z to initial income Y, also expressed in natural logarithm. All changes are between two surveys and are annualised. The coefficients to be estimated are bi(i=1,…,9). The sign of b2 is expected to be negative, as income growth should lead to poverty reduction, ceteris paribus. The signs of b3 and b4 are expected to be positive, as both higher initial inequality (G′) and lower initial income (higher Z/Y) should weaken the effect of income growth (y) on poverty reduction. The sign of b5 is expected to be positive, as rising inequality is often associated with increasing poverty. But a negative b5 is also possible, in which case rising inequality contributes to poverty reduction. The sign of b6 can be either positive or negative: if it is positive, then rising inequality will have a greater (negative or positive depending on the sign of b5) impact on poverty reduction in countries with higher initial inequality; if it is negative, then rising inequality will have a smaller impact on poverty reduction in these countries. Similarly, the sign of b7 can also be either positive or negative; in the latter case rising inequality has a smaller impact on poverty reduction in countries with higher Z/Y (higher initial poverty or lower initial income). Finally, the signs of b8 and b9 are expected to be positive, as both higher initial inequality and higher initial poverty should impede poverty reduction. Using the estimates of b’s from equation (1), we can derive the growth elasticity of poverty reduction ( Ey) and inequality elasticity of poverty reduction ( Egini):   Ey=b2+b3G′+b4Z/Y (2)  Egini=b5+b6G′+b7Z/Y (3) The elasticity equations clearly suggest that the growth effectiveness of poverty reduction is not only related to income growth per se, but also to the initial inequality level and initial income level. Similarly, initial levels of inequality and income also affect how sensitive the change in poverty is to a change in inequality. It should also be noted that the sign of Ey is expected to be negative as growth reduces poverty, while the sign of Egini is expected to be positive as rising inequality increases poverty. To see whether the growth pattern in SSA differs from that in the developing world (including or excluding SSA), we need to estimate b1–b9 and calculate the elasticities for the SSA region, the entire developing world and the non-SSA developing world. We used two data sets for our empirical exploration. The first data set is essentially the same as that used by Fosu (2015) who derived country-level data from the 2009 issue of the World Bank Povcalnet database4. While this data set—hereinafter referred to as the Povcalnet 1 data set—has in total 539 observations from 124 developing countries covering the 1977–2007 period for the $1.25 a day poverty line, our regressions are based on 368 usable observations collected in 94 developing countries during 1982–2007.5 The second data set is an update of the Povcalnet 1 data set, which takes advantage of the 2014 issue of the Povcalnet database. This most recent and enriched data set—hereinafter referred to as the Povcalnet 2 data set—contains 702 usable observations collected from 102 developing countries during 1986–2012. If the 1986–2007 sub-set of the Povcalnet 2 data included the same information as the Povcalnet 1 data set for the same timespan, as we initially assumed, then any difference between regression results derived from these two data sets should reveal the change in growth structure within a region in the post-2007 years. Unfortunately, this is not the case as it appears that the World Bank periodically revises and updates its historical data sets.6 Hence, a comparison of regressions run on the two data sets reflects not only the structural changes that might have occurred in the post-2007 period but also the revisions made by the World Bank. One issue that needs to be highlighted at the outset is that of the representativeness of the SSA subsample. The Povcalnet 2 data set for SSA that includes 103 observations collected from 35 SSA countries over the timespan 1986–2012 can be claimed to be representative.7 There is, however, a concern about the representativeness of the Povcalnet 1 SSA subsample. The Povcalnet 1 SSA subsample we used in this analysis contains 66 observations collected from 28 SSA countries during 1987–2006.8 The relatively small number of observations and the fact that some countries with multiple observations (e.g., Côte d'Ivoire) are given unduly high weight raise serious questions about the representativeness of this SSA subsample. Due to this concern, Fosu (2015) decided not to estimate coefficients separately for the SSA subsample, but instead to use structural coefficients of the entire developing world for his analysis of SSA growth structure. While these are valid concerns and should be kept in mind when interpreting our regression results, we decided to estimate SSA's own structural coefficients during 1987–2006 and reported them in Table 1 column (2). For one thing, compared with the SSA subsample Fosu (2015) would have used, the Povcalnet 1 SSA subsample we used in this analysis is slightly more representative in the sense that ours contains more observations from more countries collected during a longer period of time.10 An important result is that we find that the inequality elasticity of poverty reduction ( Egini) for SSA taken as a whole is very different depending on the structural coefficients used: SSA Egini estimated using SSA's own structural coefficients is only about half of that estimated using structural coefficients of the entire developing world.11 This suggests that estimating SSA's own structural coefficients is necessary for our understanding of the G–I–P nexus in SSA; though the region has a less than ideal sample size and composition. Table 1: SSA Versus Global Structural Coefficients, Povcalnet 1 Data Set, 1982–2007   Global  SSA  Non-SSA  1982–2007  1987–2006  1982–2007  (1)  (2)  (3)  Growth of log income (b2)  −12.49***  −9.01**  −15.02***  (−5.52)  (−2.27)  (−6.80)  Growth of log income × log initial Gini (b3)  3.04***  2.00*  4.03***  (4.88)  (1.93)  (6.15)  Growth of log income × log (poverty line/initial income) (b4)  1.33***  0.53***  2.14***  (4.43)  (2.59)  (6.25)  Change in log Gini (b5)  19.90***  −9.46  24.14***  (3.42)  (−1.49)  (3.78)  Change in log Gini × log initial Gini (b6)  −5.00***  2.65  −6.34***  (−3.24)  (1.57)  (−3.35)  Change in log Gini × log (poverty line/initial income) (b7)  −2.46***  −0.60  −3.08***  (−4.03)  (−1.01)  (−3.30)  Log initial Gini (b8)  0.11  0.01  0.19  (0.66)  (0.13)  (0.93)  Log (poverty line/initial income) (b9)  0.01  0.12  0.02  (0.17)  (1.03)  (0.20)  Intercept (b1)  −0.43  −0.0003  −0.72  (−0.71)  (−0.001)  (−1.00)  N  368  66  302  R2  0.69  0.89  0.70    Global  SSA  Non-SSA  1982–2007  1987–2006  1982–2007  (1)  (2)  (3)  Growth of log income (b2)  −12.49***  −9.01**  −15.02***  (−5.52)  (−2.27)  (−6.80)  Growth of log income × log initial Gini (b3)  3.04***  2.00*  4.03***  (4.88)  (1.93)  (6.15)  Growth of log income × log (poverty line/initial income) (b4)  1.33***  0.53***  2.14***  (4.43)  (2.59)  (6.25)  Change in log Gini (b5)  19.90***  −9.46  24.14***  (3.42)  (−1.49)  (3.78)  Change in log Gini × log initial Gini (b6)  −5.00***  2.65  −6.34***  (−3.24)  (1.57)  (−3.35)  Change in log Gini × log (poverty line/initial income) (b7)  −2.46***  −0.60  −3.08***  (−4.03)  (−1.01)  (−3.30)  Log initial Gini (b8)  0.11  0.01  0.19  (0.66)  (0.13)  (0.93)  Log (poverty line/initial income) (b9)  0.01  0.12  0.02  (0.17)  (1.03)  (0.20)  Intercept (b1)  −0.43  −0.0003  −0.72  (−0.71)  (−0.001)  (−1.00)  N  368  66  302  R2  0.69  0.89  0.70  Notes: This table gives b1ˆ–b9ˆ in Bourguignon's Identity Model using the Povcalnet 1 data set and the fixed-effect (FE) regression procedure. Log (poverty line/initial income) measures initial poverty condition, where poverty line is $1.25/day. T-ratios are in parenthesis and based on White standard errors (are heteroscedasticity-consistent). T-ratios are compared against critical values adjusted for sample sizes.9 ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. We applied three regression procedures to the Bourguignon model: Ordinary Least Square (OLS), FE and the GMM. The OLS and FE results are very similar (in both magnitude and significance level) whether they are applied to the Povcalnet 1 or the Povcalnet 2 data sets and their respective SSA subsamples. The GMM procedure, however, generates only significant estimates for the Povcalnet 1 data set12 but insignificant results for the Povcalnet 2 data set, even though the sample size of the latter is almost twice that of the former.13 The poor performance of GMM estimates may be related to weak identification (and even under-identification in some cases) of the GMM model used, which we inherited from Fosu (2011, 2015) and used as instrumental variables (IVs) lagged income, lagged Gini index, and their interaction terms with regional dummies.14 Specifically, the identification tests statistics suggest that these IVs are only weakly related to the endogenous regressor they are supposed to instrument (income growth) no matter which of the two data sets and their SSA subsamples are used;15 further, these IVs are barely relevant to the endogenous regressor when the procedure is run for the Povcalnet 1 SSA subsample.16 In the presence of weak identification, GMM estimates may perform poorly (Stock et al., 2002; Stock and Yogo, 2005); and under-identification of the model when applied to Povcalnet 1 SSA subsample means that little is gained from instrumenting. Following suggestion of Baum et al.(2003), we tried to address the issue of weak identification by reducing the number of IVs; but this effort was in vain. Lagged variables are usually viewed as good instruments that address not only the endogeneity issue but also measurement errors (Lopez and Servén, 2009; Ravallion, 2012). However, if over time some structural change happened to the growth pattern, lagged income and lagged Gini coefficients may no longer be valid instruments for income growth. We suspect this may well be the case especially for the SSA region which went through significant structural changes since around 2000 as discussed in Section 1. Further, we wonder if the GMM regression procedure is really necessary in our analysis. After all, GMM estimator comes with a price: it is less efficient compared with the OLS estimator (which is but less efficient than the FE estimator); and it has poor finite sample performance compared with regular IV estimator (which is a special case of GMM). Econometricians agree that GMM is appropriate if endogeneity and heteroscedasticity exist at the same time. While the presence of heteroscedasticity in our data is confirmed by the Pagan–Hall test statistic,17 we find little evidence that income growth endogeneity is a serious concern in our data. The endogeneity test suggests that for the Povcalnet 1 global sample, the Povcalnet 2 global sample, and its SSA subsample, income growth can all be treated as exogenous.18 Due to the above considerations, we report only b1–b9 estimates from the FE regression procedure.19 Compared with OLS and GMM, the FE approach has three advantages: (i) it addresses endogeneity related to cross-country heterogeneity, which OLS cannot address but is demonstrated to have been significant among SSA countries by Fosu (2015); and (ii) the FE estimator is more efficient than the OLS estimator as it addresses serial correlation of the error terms, yet OLS estimator is more efficient than the GMM estimator; and (iii) the FE method allows slightly larger sample size than the GMM estimator as it has fewer explanatory variables. One major concern of the FE approach is that its attenuation bias may be too strong to reveal true relationship when applied to data with independent measurement errors in each period (Hauk and Wacziarg, 2009; Ravallion, 2012). Our FE results however are significant, suggesting that the true relationship may only be stronger. 4.2 Empirical results Table 1 presents our FE estimates of b1–b9 based on the Povcalnet 1 sample, its SSA subsample, and its non-SSA subsample. Since the comparison between the full sample and the SSA subsample conveys the same message as the comparison between the non-SSA and the SSA subsamples, we shall focus on explaining the former. The estimates for the entire developing world reported in column (1) are similar to those reported in Fosu (2015) following a GMM regression procedure, suggesting that the FE method is a valid substitute of the GMM method at least for the global sample. Table 1 column (2) presents the FE estimates of structural coefficients based on 66 observations from 28 SSA countries collected during 1987–2006. As mentioned at the outset of this section, the Povcalnet 1 SSA subsample is less ideal in sample representativeness. It does, however, allow us to see that the growth structure in SSA is very different from that of the entire developing world. Given the above estimates, we then calculated the growth elasticity of poverty reduction ( Ey) and inequality elasticity of poverty reduction ( Egini) for SSA and the entire developing world using formula provided in Equations (3) and (4). We report our elasticity estimates in Table 2. Note that these estimates are similar to those reported in Fosu (2009, Table 2) using data collected during 1982–2004.20 Table 2: Growth and Inequality Elasticity of Poverty Reduction for Global and SSA Regions, Povcalnet 1 Data Set, 1982–2007 Region  Entire developing world  SSA  Non-SSA developing world  1982–2007  1987–2006  1982–2007  Ey  −2.91  −1.56  −3.26  Egini  4.48  0.62  5.23  Estimates from  368 observations from 94 countries  66 observations from 28 countries  302 observations from 66 countries  Region  Entire developing world  SSA  Non-SSA developing world  1982–2007  1987–2006  1982–2007  Ey  −2.91  −1.56  −3.26  Egini  4.48  0.62  5.23  Estimates from  368 observations from 94 countries  66 observations from 28 countries  302 observations from 66 countries  Comparing b1–b9 in the two columns of Table 1, we see that during 1982–2007, growth (as reflected by b2) had a significant but much smaller impact on poverty reduction in SSA than in the entire developing world (−9.01** versus −12.49***). Further, as reflected by b3 and b4, high initial inequality and low initial income significantly weaken the growth elasticity of poverty reduction ( Ey) in SSA at about the same relative scale and strength as they weaken Ey in the entire developing world. Consequently, the growth elasticity of poverty reduction in SSA was only about a half of that in the developing world (Table 2 row 1: −1.56 versus −2.91). Compared with the entire developing world, the SSA region also had a much smaller inequality elasticity of poverty reduction ( Egini) during 1982–2007 (Table 2 row 2: 0.62 versus 4.48). The main reason seems to be that rising inequality had little impact (as reflected by b5) on poverty reduction in SSA taken as a whole but a very strong poverty-increasing effect elsewhere (−9.46 versus 19.90***).21 This is imaginable if one considers the enclave-type, exclusive growth that many resource-rich SSA economies experienced before the early 2000s, where resource rent had been enjoyed by only the group in power (Ndulu, 2008; Thorbecke and Ouyang, 2016). The above result, however, should not be taken as evidence that inequality plays a minor role in poverty reduction in SSA. Our country-level elasticity estimates suggest large variability among SSA countries.22 In his recent study using also the Povcalnet 1 data but different regression procedure, Fosu (2015) also found considerable country-level heterogeneity with respect to the relative contribution of income growth and inequality improvement in SSA. Next, we perform the same analysis to the enriched and expanded Povcalnet 2 data set and its SSA subsample and report the estimates of b1–b9 in Table 3. Again, we report estimates for the full sample, its SSA subsample, and its non-SSA subsample; but we focus on explaining the comparison between the full sample and the SSA subsample, as its implications are similar to those one would obtain from comparing the SSA and the non-SSA subsamples. Table 3: SSA Versus Global Structural Coefficients, Povcalnet 2 Data Set, 1986–2012   Global  SSA  Non-SSA  1986–2012  1986–2012  1986–2012  (1)  (2)  (3)  Growth of log income (b2)  −9.91**  −9.96  −10.42***  (−2.41)  (−1.42)  (−4.49)  Growth of log income × log initial Gini (b3)  1.96*  2.40  1.83**  (1.82)  (1.37)  (2.98)  Growth of log income × log (poverty line/initial income) (b4)  −0.21  0.91**  −0.79*  (−0.58)  (2.09)  (−2.36)  Change in log Gini (b5)  24.31***  15.84  25.03***  (2.97)  (1.19)  (5.03)  Change in log Gini × log initial Gini (b6)  −5.23**  −4.21  −5.06  (−2.37)  (−1.20)  (−3.72)  Change in log Gini × log (poverty line/initial income) (b7)  −0.51  −4.62***  0.36  (−0.66)  (−3.19)  (0.54)  Log initial Gini (b8)  0.70***  0.02  0.62**  (3.46)  (0.19)  (2.75)  Log (poverty line/initial income) (b9)  0.01  −0.15*  0.01  (0.18)  (−1.44)  (0.17)  Intercept (b1)  −2.63***  −0.13  −2.30**  (−3.46)  (−0.30)  (−2.72)  N  702  103  599  R2  0.54  0.68  0.55    Global  SSA  Non-SSA  1986–2012  1986–2012  1986–2012  (1)  (2)  (3)  Growth of log income (b2)  −9.91**  −9.96  −10.42***  (−2.41)  (−1.42)  (−4.49)  Growth of log income × log initial Gini (b3)  1.96*  2.40  1.83**  (1.82)  (1.37)  (2.98)  Growth of log income × log (poverty line/initial income) (b4)  −0.21  0.91**  −0.79*  (−0.58)  (2.09)  (−2.36)  Change in log Gini (b5)  24.31***  15.84  25.03***  (2.97)  (1.19)  (5.03)  Change in log Gini × log initial Gini (b6)  −5.23**  −4.21  −5.06  (−2.37)  (−1.20)  (−3.72)  Change in log Gini × log (poverty line/initial income) (b7)  −0.51  −4.62***  0.36  (−0.66)  (−3.19)  (0.54)  Log initial Gini (b8)  0.70***  0.02  0.62**  (3.46)  (0.19)  (2.75)  Log (poverty line/initial income) (b9)  0.01  −0.15*  0.01  (0.18)  (−1.44)  (0.17)  Intercept (b1)  −2.63***  −0.13  −2.30**  (−3.46)  (−0.30)  (−2.72)  N  702  103  599  R2  0.54  0.68  0.55  Notes: This table gives b1ˆ–b9ˆ in Bourguignon's Identity Model using the Povcalnet 2 data set and the FE regression procedure. T-ratios are in parenthesis and based on White standard errors (corrected for heteroscedasticity). Log (poverty line/initial income) measures initial poverty condition, where poverty line is $1.25/day. T-ratios are compared against critical values adjusted for sample sizes. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. Based on the above estimates, we calculated the income and growth elasticities of poverty reduction during 1986–2012 for the entire developing world and SSA, respectively; and report these elasticities in Table 4. Table 4: Growth and Inequality Elasticity of Poverty Reduction for Global and SSA Regions, Povcalnet 2 Data Set, 1986–2012 Region  Entire developing world  SSA  Non-SSA  (1986–2012)  (1986–2012)  (1986–2012)  Ey  −2.32  −1.20  −2.34  Egini  5.64  1.62  5.27  Estimates from  702 observations from 102 countries  103 observations from 35 countries  599 observations from 67 countries  Region  Entire developing world  SSA  Non-SSA  (1986–2012)  (1986–2012)  (1986–2012)  Ey  −2.32  −1.20  −2.34  Egini  5.64  1.62  5.27  Estimates from  702 observations from 102 countries  103 observations from 35 countries  599 observations from 67 countries  Results presented in Tables 3 and 4 suggest that during 1986–2012, the impact of income growth on poverty reduction was essentially the same in SSA as in the entire developing world (−9.96** versus −9.91**). At given income growth rate (y), high initial inequality also weakens Ey at about the same strength in SSA as in the entire developing world (2.4 versus 1.96*). But because high initial poverty (high Z/Y or low initial income) weakens Ey significantly more in SSA than in the entire developing world (0.91** versus −0.21), Ey in SSA remains about half of that in the developing world during 1986–2012 (−1.12 versus −2.13), as it did during 1982–2007 (−0.56 versus −2.91 as reported in Table 2). Note that during 1982–2007, the main reason that SSA has a smaller Ey than the rest of the developing world is that poverty reduction is less responsive to income growth in SSA; in contrast, the main reason for SSA to still have a smaller Ey during 1986–2012 is that initial poverty weakens Ey more in SSA, while poverty reduction has become equally responsive to income growth in SSA as it is in the entire developing world. This suggests that directly addressing high initial poverty in SSA has become more important for poverty reduction in the region. The inequality elasticity of poverty reduction ( Egini) in the SSA region, in contrast, has significantly increased compared with the earlier period (Povcalnet 1) in the sense that it was close to a third of its developing world counterpart during 1986–2012 (1.62 versus 5.64), while it was less than a fifth of its developing world counterpart during the earlier 1982–2007 period (0.80 versus 4.48). The significant increase in SSA's Egini suggests that poverty reduction in SSA as a whole has become more responsive to improvement in inequality than before. This is what one would expect as growth has become more inclusive in the SSA region; as inclusiveness means that any socio-economic improvement, including a reduction in inequality, would benefit more poor people and thus have a greater impact on poverty reduction.23 We also notice that at given inequality change rate (g), high initial poverty now weakens Egini significantly more in SSA than in the entire developing world (−4.62*** versus −0.51), while it has little impact on Egini during 1982–2007 (−0.60 versus −2.46*** as reported in Table 1)—a finding that again suggests the increasing importance of directly attacking high initial poverty in SSA. To summarise, our above analysis of the G–I–P nexus in SSA provides some empirical evidence in support of what we conjectured in the beginning of this paper based on existing (mostly) qualitative evidence. First, our empirical results suggest that during the three decades between the early 1980s and the early 2010s, the growth pattern in the SSA region is indeed different from that in the entire developing world in the sense that poverty reduction in SSA is less responsive to income growth and improving inequality. Second, the growth pattern within SSA as a whole appears to have changed over time. More specifically, SSA seems to be slowly catching up as evidenced by the estimates derived from the Povcalnet 2 data set that includes the more recent observations from 2008 to 2012.24 We find that (i) the poverty-reducing effect of income growth in SSA has increased to become slightly greater than that in the entire developing world and (ii) poverty reduction in SSA has become more responsive to improving inequality. Associated with these improvements, however, are the increasingly stronger Ey- and Egini-weakening effect of high initial inequality and high initial poverty, suggesting that addressing high initial inequality and low initial income has become more urgent than it was before for the poor people in SSA countries to benefit more from income growth and improving inequality. If we also consider the social benefits of reducing persistent and high inequality and poverty, the importance of directly attacking initial poverty and inequality is further reinforced. 5. Estimating the African structural coefficients of the P–I–G nexus 5.1 Econometric model Compared with its long recognition and extensive exploration of the impact of growth and inequality on poverty (the G–I–P nexus), the development community—as discussed in Section 3—had paid less attention to the reverse influence from poverty and inequality to subsequent growth (the P–I–G Nexus) until the most recent decade. Specifically, there are two strands of literature on the P–I–G nexus: one focusing on the direct impact of inequality on subsequent income growth; the other focusing on the direct impact of initial poverty on income growth. The former strand of literature has so far reached no unanimous conclusions: Alesina and Rodrik (1994), Perotti (1996) and Ravallion (2012) found a negative relationship between inequality and growth on the basis of cross-section data.25Li and Zou (1998) and Forbes (2000), in contrast, found a positive relationship between the two using aggregate panel data. Barro (2000) found the impact of inequality on growth dependent on the country's level of income. Banerjee and Duflo (2003), in turn, argued that there is an inverted U-shape relationship between income growth and the changes in inequality. Our emphasis and focus here is on the strand of literature, which explores the direct impact of initial poverty on subsequent income growth. While the theoretical literature on links between income or consumption growth and poverty-related variables (such as under-investment in human and/or physical capital, risk aversion and poor institutions) is relatively extensive,26 and several empirical studies attempted to uncover these indirect poverty–growth links;27 there have been very few direct empirical assessments of the impact of initial poverty on subsequent growth. One of the earliest empirical attempts to explore the direct impact of initial poverty on subsequent income growth was made by Lopez and Servén (2009). Using country-level panel data from some hundred developed and developing countries during 1960–2000,28Lopez and Servén (2009) found that a ten percentage-point increase in the headcount poverty rate reduces annual per capita income growth by about one percentage point. Following a similar empirical strategy29 but using cross-section data from ninety developing countries during 1978–2007, Ravallion (2012) found a lack of poverty convergence in the entire developing world in spite of evidence for mean (income and also consumption) convergence and the advantage of growth.30 He then demonstrated that the lack of poverty convergence is because initial poverty directly impedes subsequent income growth, besides weakening the growth elasticity of poverty reduction which has been extensively documented in the G–I–P literature.31 Given our conjecture that the growth pattern in SSA is different from that of the entire developing world, it is essential to analyse how initial poverty affects income growth and poverty reduction in Africa. A first effort to explore the P–I–G nexus in the African continent was made by Shimeles et al.(2016), who applied Ravallion (2012)’s empirical models to his African subsample containing thirty-two African countries including four Northern African countries. Their analysis suggests that Africa enjoyed poverty convergence during 1978–2007 as initial poverty did not impede subsequent growth in Africa as it did in the entire developing world; which they suggest could be explained by the anti-poverty programmes some African governments put in place since the mid-1990s. Since our focus in this paper is SSA, we further restrict our analysis here to the SSA region; that is, we excluded the four North African countries in the Shimeles et al.(2016) analysis. Hence, the results we reported in this paper are based on the Ravallion (2012) data set32 covering ninety developing countries that were surveyed at least twice during 1978–2007 and its SSA subsample covering twenty-eight SSA countries surveyed during the same period. Differences in the analysis results based on the global sample and its SSA subsample should reflect the differences between the P–G–I relationship in SSA and that in the entire developing world. Incidentally, more recent data have become available since the Ravallion (2012) data were collected in December 2008; and the extended data set now covers 105 developing countries that were surveyed at least twice during 1978–2012. The SSA subsample of this extended data set contains thirty-two SSA countries surveyed during 1980–2011.33 We have tried to apply the extended data set to Ravallion (2012)’s empirical models and have obtained some preliminary results, but recently we noticed some updates of the Povcalnet data which may change the analysis results. We therefore decide not to report these results at this stage and in this paper.34 Two key questions and valid concerns have to be addressed before reporting and analysing the results we obtained from applying the Ravallion (2012) models to the SSA subsample: (i) is the SSA sample of countries and growth spells representative of the population? And (ii) is there sufficient variability (variance) within the SSA sample to capture the true relationship? Although it would be preferable to have access to an even larger SSA sample, we feel that the present sample is relatively representative of the geographical and economic diversity of the subcontinent as indicated by footnote 31 that gives the list of countries and the first and last years of the surveys used to derive the growth spells. In order to check on the variability of the SSA subsample compared to the full (whole developing world) sample, we derived the kernel distribution of the log initial headcount index (measured against the $2/day poverty line) for the full sample used by Ravallion (2012) and separately for the SSA subsample. As expected and shown in Figure 3, the former distribution has a much wider spread than the latter. Clearly, the small variability of the SSA sample means that our results should be qualified accordingly. With this qualification in mind, our analysis does provide evidence supporting our conjecture that the P–I–G relationship in SSA is very different from that in the developing world as a whole. Figure 3: View largeDownload slide Kernel Distribution of Log Initial Headcount Index by Region (1978–2007). Figure 3: View largeDownload slide Kernel Distribution of Log Initial Headcount Index by Region (1978–2007). We first explored whether convergence in mean income and poverty convergence exist in our two SSA subsamples using, respectively, the following empirical model from Ravallion (2012, Equations (1) and (3), p. 511)35:   gi(μit)=αi+βilnμit−1+εit (4) (Test for income convergence)   giHit=αi⁎+βi⁎lnHit−1+εit⁎ (5) (Test for poverty convergence). The dependent variable gi(xit)(≡(ln(xit)−ln(xit−τ))/τ) denotes the annualised change in log mean income ( μit) and log headcount index ( Hit) between year (t−τ) and year t in Equations (4) and (5), respectively; and the independent variable of interest is, respectively, log mean income and log headcount index in the beginning of the entire period of τ years. The coefficients of the initial values give the rates of unconditional convergence. For conditional convergence rates, we added—following Ravallion (2012)—a set of controls comprising per capita consumption from national accounts, primary school enrolment rate, life expectancy at birth and the relative price index of investment goods as a measure of policy distortion—all for earliest survey date and expressed in natural logarithm. 5.2 Empirical results As in Section 4, we performed analysis for the full sample, its SSA subsample and its non-SSA subsample and reported all these results. But we shall only explain the difference between SSA and the whole developing world, as the comparison between the SSA and non-SSA subsamples conveys essentially the same information. As given in Table 5, we find that both SSA and the entire developing world enjoyed strong mean income convergence during 1978–2007. Further, the former enjoyed stronger mean convergence than the latter, which is expected as the SSA region on average has a higher initial poverty level than the entire developing world. The estimates in Table 5 are robust to the inclusion of controls and also the inclusion of North African countries. Table 5: Convergence in Mean Income (1978–2007)   Unconditional world  Conditional world  Unconditional SSA  Conditional SSA  Unconditional non-SSA  Conditional non-SSA  Initial log mean income  −0.0174***  −0.0465***  −0.0276***  −0.0485***  −0.0207*  −0.0456***  (−3.2017)  (−10.6022)  (−3.4963)  (−5.3002)  (−2.5429)  (−9.1244)  N  97  88  28  28  69  60  R2  0.1351  0.4982  0.1709  0.4676  0.1294  0.5827    Unconditional world  Conditional world  Unconditional SSA  Conditional SSA  Unconditional non-SSA  Conditional non-SSA  Initial log mean income  −0.0174***  −0.0465***  −0.0276***  −0.0485***  −0.0207*  −0.0456***  (−3.2017)  (−10.6022)  (−3.4963)  (−5.3002)  (−2.5429)  (−9.1244)  N  97  88  28  28  69  60  R2  0.1351  0.4982  0.1709  0.4676  0.1294  0.5827  Notes: This table gives βˆ in the regression (4) in the text: gi(μit)=αi+βilnμit−τ+γiXit−τ+εit. T-ratios are in parenthesis and based on White standard errors (corrected for heteroscedasticity). T-ratios are compared against critical values adjusted for sample sizes. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. Estimates in Table 6, for its part, suggest that SSA experienced significant and sizable poverty convergence during 1978–2007; while the entire developing world experienced a lack of poverty convergence during this period as revealed in Ravallion (2012).36 As in Table 5, the findings in Table 6 are robust to the inclusion of control variables and the choice of poverty lines, though not to the inclusion of North African countries.37 Table 6: Convergence in Poverty (1978–2007)   Unconditional world  Conditional world  Unconditional SSA  Conditional SSA  Unconditional SSA  Conditional SSA  Initial log headcount index ($2 a day)  0.0059  −0.0099  −0.0254**  −0.0409***  0.0042  −0.0090  (0.5901)  (−0.6967)  (−2.3736)  (−4.4003)  (0.3632)  (−0.5478)  N  89  84  28  28  61  56  R2  0.0080  0.1202  0.1340  0.4321  0.0029  0.1755  Initial log headcount index ($1.25 a day)  −0.0052  −0.0248  −0.0251**  −0.0372***  −0.0099  −0.0244  (−0.452)  (−1.5746)  (−2.3861)  (−4.4219)  (−0.7013)  (−1.2516)  N  82  78  28  28  54  50  R2  0.0053  0.1451  0.1127  0.3638  0.0128  0.1957    Unconditional world  Conditional world  Unconditional SSA  Conditional SSA  Unconditional SSA  Conditional SSA  Initial log headcount index ($2 a day)  0.0059  −0.0099  −0.0254**  −0.0409***  0.0042  −0.0090  (0.5901)  (−0.6967)  (−2.3736)  (−4.4003)  (0.3632)  (−0.5478)  N  89  84  28  28  61  56  R2  0.0080  0.1202  0.1340  0.4321  0.0029  0.1755  Initial log headcount index ($1.25 a day)  −0.0052  −0.0248  −0.0251**  −0.0372***  −0.0099  −0.0244  (−0.452)  (−1.5746)  (−2.3861)  (−4.4219)  (−0.7013)  (−1.2516)  N  82  78  28  28  54  50  R2  0.0053  0.1451  0.1127  0.3638  0.0128  0.1957  Notes: This table gives βˆ in the regression (5) in the text: gi(Hit)=αi⁎+βi⁎lnHit−τ+γi⁎Xit−τ+εit⁎. T-ratios are in parenthesis and based on White standard errors (corrected for heteroscedasticity). T-ratios are compared against critical values adjusted for sample sizes. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. Hence, it can be inferred from the above evidence that the structure of growth in SSA is different from that of the entire developing world during the period of 1978–2007. To explain the existence of poverty convergence in SSA during 1978–2007, we recall that Ravallion (2012) demonstrated that the rate of poverty reduction is jointly determined by the income growth rate and the initial poverty level. Specifically, countries with lower initial income tend to experience faster growth in mean income and hence faster poverty reduction.38 This is referred to as the mean convergence effect on poverty convergence. The initial poverty level, for its part, affects the rate of poverty reduction through two distinct channels. On the one hand, it directly hinders subsequent income growth and hence lowers the rate of poverty convergence through low income growth; on the other hand, it indirectly impedes poverty reduction through weakening the growth elasticity of poverty reduction (or making poverty reduction less responsive to income growth). Ravallion (2012, Equation (11), p. 520) summarises the above into the following equation that decomposes the rate of poverty convergence into three components:   ∂g(Hit)∂lnHit−τ=ηβ(1−Hit−τ)(∂lnHit−τ∂lnμit−τ)−1+ηγ(1−Hit−τ)+[−ηg(μit)Hit−τ] (6)(Mean convergence effect) (Direct effect (Poverty elasticity of poverty) effect). Equation (6) is not an identity but instead derived from the two empirical models specified below:   giμit=α+βlnμit−τ+γlnHit−τ+εit (7)Ravallion (2012)), Equation (4) and also Equation (10.2))   g(Hit)=η(1−Hit−τ)g(μit)+vit (8)Ravallion (2012), Equation (10.1)). Parameters β and γ in Equation (6), therefore, are from Equation (7), where the annualised change in log mean income ( giμit≡lnμit−lnμit−τ)/τ) is regressed on the initial levels of income and poverty, or income and poverty in the beginning of the growth spell of τ years. Parameter η in Equation (6), for its part, is from Equation (8) and is the regression coefficient of the ‘initial-poverty-adjusted income growth rate’ on the change in poverty. As Ravallion (2012, Table 4, p. 528) convincingly demonstrated, ‘the key proximate determinant of the rate of poverty reduction is η, the “poverty-adjusted growth rate” ( (1−Hit−τ)g(μit); rather than the ordinary growth rate ( g(μit)’ (pp. 519–520). The expression ∂lnHit−τ∂lnμit−τ in Equation (6) is the regression coefficient of g(Hit on g(μit, or the standard growth elasticity of poverty reduction denoted as Ey in Section 4. Finally, Hit−τ and g(μit) are the sample means of initial poverty rate and mean income growth rate. Estimates of Equation (7) based on the Ravallion (2012) data and its SSA subsample are presented in Table 7. During 1982–2007, initial poverty significantly retards income growth in the entire developing world as a whole (−0.0201*** with t-ratio = −5.51) but has little impact on income growth in Africa (−0.0035 with t-ratio = −0.14). Again, these results are robust to the choice of poverty lines (1.25 or 2 international dollars per day) and the inclusion of a set of control variables including the initial Gini index.39 Table 7: Regressing Mean Income Growth on Initial Poverty and Income (1978–2007)   World  SSA  Non-SSA  Log initial income  −0.0395***  −0.0293*  −0.0440***  (−8.4095)  (−1.7913)  (−8.8327)  Log initial headcount index ($2 line)  −0.0201***  −0.0035  −0.0200***  (−5.5135)  (−0.1363)  (−5.7577)  N  90  28  62  R2  0.275  0.1711  0.3080  Log initial income  −0.0348***  −0.0249  −0.0375***  (−5.0507)  (−1.4517)  (−4.6644)  Log initial headcount index ($1.25 line)  −0.01*  0.0035  −0.0094*  (−2.4131)  (0.2306)  (−2.2337)  N  84  28  56  R2  0.2425  0.1714  0 .2563    World  SSA  Non-SSA  Log initial income  −0.0395***  −0.0293*  −0.0440***  (−8.4095)  (−1.7913)  (−8.8327)  Log initial headcount index ($2 line)  −0.0201***  −0.0035  −0.0200***  (−5.5135)  (−0.1363)  (−5.7577)  N  90  28  62  R2  0.275  0.1711  0.3080  Log initial income  −0.0348***  −0.0249  −0.0375***  (−5.0507)  (−1.4517)  (−4.6644)  Log initial headcount index ($1.25 line)  −0.01*  0.0035  −0.0094*  (−2.4131)  (0.2306)  (−2.2337)  N  84  28  56  R2  0.2425  0.1714  0 .2563  Notes: This table gives βˆ in the regression (7) in the text: giμit=α+βlnμit−τ+γlnHit−τ+εit. T-ratios are in parenthesis and based on White standard errors (corrected for heteroscedasticity). T-ratios are compared against critical values adjusted for sample sizes. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. Next we report estimates of Equation (8) in Table 8. The negative sign of η suggests that in both the entire developing world and the SSA region, higher initial poverty weakens the poverty-adjusted growth elasticity of poverty reduction during 1978–2007. Comparing the magnitude of the two η s, we see that initial poverty weakens the (poverty-adjusted) growth effectiveness of poverty reduction less in the SSA region than in the entire developing world (−2.2694*** versus −2.4676***). Table 8: Regressions of Poverty-Adjusted Income Growth Rate on the Change in Poverty (1978–2007)   World  SSA  Non-SSA  Poverty-adjusted income growth (1−Hit−τ)g(μit)($2 line)  −2.4676***  −2.2694***  −2.4653***  (−7.3670)  (−6.5728)  (−7.1981)  N  89  28  61  R2  0.6714  0.8257  0.6662    World  SSA  Non-SSA  Poverty-adjusted income growth (1−Hit−τ)g(μit)($2 line)  −2.4676***  −2.2694***  −2.4653***  (−7.3670)  (−6.5728)  (−7.1981)  N  89  28  61  R2  0.6714  0.8257  0.6662  Notes: This table gives βˆ in the regression (8) in the text: g(Hit)=η(1−Hit−τg(μit)+vit. T-ratios are in parenthesis and based on White standard errors (corrected for heteroscedasticity). T-ratios are compared against critical values adjusted for sample sizes. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. Now we have estimated all parameters needed to compute the three components of poverty convergence rate specified in Equation (6). We present our computation results in Table 9. Numerical figures in this table seem to suggest that while the three components almost perfectly predicted the actual rate of poverty convergence in the entire developing world during 1978–2007 as suggested in Ravallion (2012), they could not fully account for the poverty convergence that SSA experienced during the period: the sum of the three effects predicts a lack of poverty convergence (−0.0035) in SSA, while a regression analysis of the actual data suggest a strong unconditional poverty convergence in the region (−0.0254 with t = −2.37 and R2 = 0.134). Table 9: Decomposition of the Rate of Poverty Convergence (1978–2007) Poverty convergence rate = (1) + (2) + (3)  Full sample (N = 90)  SSA subsample (N = 28)  Non-SSA subsample (N = 61)  Mean convergence effect (1)  −0.0476  −0.0312  −0.0583  Direct effect of poverty (2)  0.0263  −0.0017  0.0347  Poverty elasticity effect (3)  0.0161  0.0364  0.0061  Poverty convergence rate = (1) + (2) + (3)  −0.0052  −0.0035  −0.0174  Unconditional poverty convergence rate from regression (see Table 6 upper panel)  0.0059  −0.0254**  0.0042    (t = 0.59, R2 = 0.008, N = 89)  (t = 2.37, R2 = 0.134, N = 28)  (t = 0.3632, R2 = 0.0029, N = 61)  Poverty convergence rate = (1) + (2) + (3)  Full sample (N = 90)  SSA subsample (N = 28)  Non-SSA subsample (N = 61)  Mean convergence effect (1)  −0.0476  −0.0312  −0.0583  Direct effect of poverty (2)  0.0263  −0.0017  0.0347  Poverty elasticity effect (3)  0.0161  0.0364  0.0061  Poverty convergence rate = (1) + (2) + (3)  −0.0052  −0.0035  −0.0174  Unconditional poverty convergence rate from regression (see Table 6 upper panel)  0.0059  −0.0254**  0.0042    (t = 0.59, R2 = 0.008, N = 89)  (t = 2.37, R2 = 0.134, N = 28)  (t = 0.3632, R2 = 0.0029, N = 61)  Notes: This decomposition is based on Equation (6): ∂g(Hit)∂lnHit−τ=ηβ(1−Hit−τ)(∂lnHit−τ∂lnμit−τ)−1+ηγ(1−Hit−τ)+[−ηg(μit)Hit−τ]. The first term in this equation is the mean convergence effect, the second the direct effect of initial poverty and the third the poverty elasticity effect. Parameters β and γ are from Equation (7): giμit=α+βlnμit−τ+γlnHit−τ+εit. The expression ( ∂lnHit−τ∂lnμit−τ) in Equation (6) is the standard growth elasticity of poverty reduction. Parameter η is from Equation (8): gHit)=η(1−Hit−τgμit)+vit and is referred to as the ‘initial-poverty-adjusted growth elasticity of poverty reduction’. Hit−τ and g(μit) are the sample means of initial poverty rate and mean income growth rate. What, then, could have contributed to the poverty convergence in SSA during 1978–2007? Using data from forty-two Ethiopian administrative districts covering about 80,000 households during 1996–2011 and from thirty-three Rwandan districts covering over 27,000 households during 2000–2010, Shimeles et al.(2016) found strong poverty convergence and a positive association between initial poverty and subsequent growth in mean income. Since both countries have introduced anti-poverty policies favouring lagging regions during the periods examined and have registered impressive economic growth, Shimeles et al. (2016) suggested that deliberate policy interventions by African governments and public and private aid flows might explain why the region achieved poverty convergence. 6. Summary and conclusions The main objective of this paper is to understand better the anatomy of growth in SSA and compare it to that of the developing world as a whole (including or excluding SSA). Key to the analysis of the structure of growth is the interrelationship among G–I–P. These three concepts are jointly determined in any given socio-economic system (such as a country or a region) under the influence of the process of globalisation and the country-specific development strategy that act as prime movers. The forces of globalisation are largely exogenous (outside the control of State), while the development strategy is at least partially endogenous (under the control of the State). One can best imagine a G–I–P triangle with influence travelling in both directions. In analysing the transmission of influence, one can start at any corner of the G–I–P triangle. Traditionally, research in development economics focused on the impact of influence emanating from the pattern of growth on inequality and poverty which we refer to as the G–I–P nexus. This is the theme of Section 2, where the effects of different structures of growth, such as ‘pro-poor growth’ and ‘inclusive growth’, on inequality and poverty are reviewed. The main message of this section is that the pace of growth in SSA underwent a quantum jump around the turn of the century and the structure of growth appears to have become more inclusive. The reverse causal chain emanating from poverty to subsequent inequality and growth, or what we refer to as the P–I–G nexus, is discussed in Section 3. We review recent studies and argue that a strong case can be made—based on data from the entire developing world—that high initial poverty can dampen subsequent growth. We proceeded in Section 4 to test econometrically whether the G–I–P nexus (i.e., the impact of growth on inequality and poverty) was different in SSA than in the developing world and whether the structure of growth in SSA had actually undergone a structural break around the beginning of this millennium. Building on the earlier work of Fosu (2009, 2011, 2015), we used the Bourguignon (2003) identity model to estimate and derive structural coefficients of the G–I–P nexus for the developing world and, separately, for the SSA region. We relied on two data sets, covering two different periods: Povcalnet 1 (1982–2007) and Povcalnet 2 (1986–2012). The regression results provide some empirical support for our initial conjectures. First, the growth pattern in the SSA region, in the three decades between the early 1980s and the early 2010s, was indeed different from that of the whole developing world. The responsiveness of poverty reduction to income growth and to change in inequality was significantly less in the SSA region than in the rest of the developing world. Second, the growth pattern within the SSA region has changed over time as poverty reduction has become more responsive to income growth, improvement in inequality and the level of initial poverty in the post-2007 years. The evidence suggests that SSA is slowly catching up. Next, in Section 5, we tested the P–I–G nexus to determine whether the finding that high initial poverty retarded future growth and preempted poverty convergence obtained from a large global sample of ninety developing countries covering growth spells over an extended period (1978–2007) also applied to the SSA region. Again we found that SSA differs from the rest of the developing world in that initial poverty did not dampen subsequent growth in this region during 1978–2007 while it did in the rest of the developing world during the same time period. In fact, SSA countries with the highest initial poverty incidence appeared to grow subsequently faster—leading to poverty convergence. The same finding was further confirmed at the interregional level in Ethiopia and to a lesser degree in Rwanda by Shimeles et al.(2016). We hypothesise that the main cause of poverty convergence in SSA during 1978–2007, in contrast with the lack of convergence in the global sample, might be due to anti-poverty interventions by governments and foreign public and private aid inversely proportional to the depth of poverty. Our results and findings are subject to two key caveats: (i) whether the various samples of the African countries we used in our regressions are sufficiently representative of the underlying population; and (ii) whether the sample sizes and variances are sufficient to draw valid inferences. While we recognise that larger and more diversified samples would have been preferable, we believe that we can defend the validity of our results as discussed in the text. Finally, we can draw some policy implication from our study. The analysis of the G–I–P nexus implies that an inclusive growth strategy focused on productive employment creation can spread the benefits of growth more widely and reduce extreme inequality and poverty. In turn, the analysis of the P–I–G nexus suggests that a pro-growth poverty reduction strategy focused on alleviating poverty directly through such interventions as social protection schemes can contribute to an acceleration of the pace of growth and to a more inclusive pattern. By removing some major constraints on the behaviour of the poor, the latter may be better able to invest in their education and adopt riskier but more efficient technologies. By now, there are many examples of interventions that by reducing current poverty contribute to raising productivity in the future. Since it is often difficult to draw a clear distinction as to whether a specific measure comes under one or the other strategy, it is perhaps best to consider these two strategies as complementary and mutually re-enforcing. Acknowledgements The authors would like to thank Augustin Fosu and the anonymous referees for helpful comments on earlier versions of this paper. We would also like to thank the African Development Bank for funding this research. Footnotes 1 Easterly and Levine (1997) were the first of many to refer to an ‘African dummy’. 2 Thus, an extreme case where the average income of the poor increases by only 1% following a GDP growth spell of 8% would be considered pro-poor according to the absolute definition. 3 The assumption that per capita income follows log-normal distribution is empirically tested and confirmed by Lopez and Servén (2005). They also found that this assumption does not hold for per capita consumption. 4 We thank Fosu for kindly sharing with us his data. 5 Fosu's regression is based on 320 usable observations for the $1.25 poverty line from 89 developing countries during 1981–2007 (Fosu 2015, p. 51 and endnote 10). Our usable sample size is slightly larger as we used a different regression method due to the concern of weak identification in Fosu's GMM procedure. See more discussion about this in the text. 6 In fact, the Povcalnet 1’s 1986–2007 subsample has 359 usable observations, while the Povcalnet 2’s 1986–2007 subsample contains 529 usable observations. 7 These thirty-five SSA countries, with the number of years surveyed in parentheses, are as follows: Angola (1), Benin (1), Botswana (3), Burkina Faso (3), Burundi (2), Cape Verde (1), Cameroon (2), Central African Republic (2), Chad (1), Republic of Congo (1), Côte d'Ivoire (8), Ethiopia (3), Gambia (1), Ghana (4), Guinea (4), Guinea-Bissau (1), Kenya (3), Lesotho (4), Madagascar (5), Malawi (2), Mali (3), Mauritania (5), Mauritius (1), Mozambique (2), Niger (4), Nigeria (4), Rwanda (3), Senegal (4), Sierra Leone (1), South Africa (5), Swaziland (2), Tanzania (3), Togo (1), Uganda (7) and Zambia (6). 8 These twenty-eight SSA countries, with the number of years surveyed in parentheses, are as follows: Botswana (1), Burkina Faso (2), Burundi (2), Cameroon (1), Central African Republic (1), Côte d'Ivoire (6), Ethiopia (3), Gambia (1), Ghana (4), Guinea (2), Guinea-Bissau (2), Kenya (3), Lesotho (3), Madagascar (4), Malawi (1), Mali (2), Mauritania (3), Mozambique (1), Niger (2), Nigeria (3), Rwanda (1), Senegal (3), Sierra Leone (1), South Africa (2), Swaziland (1), Tanzania (1), Uganda (5) and Zambia (5). 9 When the sample size is small, a critical value larger than 1.96 would be needed for a 5% significance level. In Table 1 Column 2 where the sample size is only 66 while the number of controls is 8, the critical value is 2.0017 for the 5% significance level and 2.3924 for the 1% level. 10 Fosu's GMM regression would have used fifty-eight observations from twenty-five SSA countries during 1987–2005; while our FE regression used sixty-six observations from twenty-eight SSA countries during 1987–2006. 11 Inequality elasticity based on Fosu (2015)’s global and our SSA structural coefficients are, respectively, 1.47 and 0.8. Income elasticity based on Fosu (2015)’s global and our SSA structural coefficients are, respectively, −1.37 and −1.56. 12 The GMM estimates for the global sample are exactly the same as those reported in Fosu (2015), which should be as exactly the same usable sample and IVs are used. 13 GMM results are available upon request. 14 See Fosu (2011) for a detailed list of instrument variables. 15 The Cragg–Donald (1993)F statistic and their Stock–Yogo critical values for 5% significance level (reported in brackets) for GMM applied to the Povcalnet 1 sample, its SSA subsample, the Povcalnet 2 sample, its SSA subsample are 3.47 (21.36), 4.56 (16.85), 1.59 (21.36) and 3.01 (16.85), respectively. Since the F statistics are smaller than the corresponding critical values, we cannot reject the null that the model is weakly identified. 16 The under-identification test statistic for GMM applied to the Povcalnet 1 SSA subsample is 7.88 with a χ2(4) p-value of 0.0961. Since the p-value is very close to 0.1, we can only barely reject the null that says the model is under-identified at 10% significance level. 17 The Pagan–Hall test statistics (and p-values) for GMM applied to the Povcalnet 1 sample, its SSA subsample, the Povcalnet 2 sample, its SSA subsample are 45.593 (χ2(26) p-value 0.01), 21.303(χ2(11) p-value = 0.0304), 100.2 (χ2(26) p-value = 0.000) and 27.401 (χ2(11) p-value = 0.004), respectively. Since the p-values are smaller than 0.05, we reject the null of no heteroscedasticity. 18 The Hausman test statistics (and p-values) for GMM applied to the Povcalnet 1 sample, its SSA subsample, the Povcalnet 2 sample, its SSA subsample are 0.297 (0.586), 6.187 (0.013), 0.147 (0.702) and 0.102 (0.750), respectively. Except for the p-value for the Povcalnet 1 SSA subsample, the p-values are all larger than 0.05, suggesting that we cannot reject the null that the specified endogenous regressor (annualised growth of log mean income) can actually be treated as exogenous. 19 OLS estimates are similar and therefore not reported for parsimony. They are available upon request. 20 Using World Bank growth and poverty data from effectively 1982 to 2004 and random-effects regression procedures, Fosu (2009) found that growth elasticity is, respectively, −1.1 and −2.9 for the SSA and global sample; and that inequality elasticity is 1.1 and 5.4 for the SSA and global sample, respectively. 21 Also note that the initial levels of poverty and inequality both had little impact on Egini in SSA (as reflected by b6 and b7). 22 Country-level growth and income elasticity estimates are not reported for parsimony, but are available upon request. 23 To explain the increase from a technical perspective, we look at Equation (3) that reveals the three components of inequality elasticity ( Egini=b5+b6G′+b7Z/Y). Comparing results reported in Tables 1 and 3, we see that the change in b5 estimate has a positive net effect on Egini; while the changes in b6 and b7 estimates both have a negative net effect on Egini, given that initial poverty (G′) and initial poverty (Z/Y) from the two data sets are largely the same (which we have confirmed). Since our estimation returns a more positive estimate for Egini from Povcalnet 2, the positive effect dominated the negative effect. That is, the increase of SSA Egini from 0.62 during 1987–2006 to 1.62 during 1986–2012 is mainly driven by a change in b5, which suggests that rising inequality coexisted with poverty reduction in SSA before 2007; but has become an obstacle to further poverty reduction in the region during the post-2007 years. 24 Ideally, we would have liked to truncate the Povcalnet 2 data set into two sub-periods 1986–2000 and 2000–2012 to try to capture the structural break in the growth pattern. Unfortunately, the number of observations is too small for us to obtain significant results. 25 Ravallion (2012), however, finds the negative relationship significant ‘only at the 10% level and when [growth is calculated] using consumption data from the national survey’ (p. 514). 26 Azariadis and Stachurski (2005) provide a good survey of poverty traps, i.e., vicious cycles of poverty and stagnation. On how poverty may retard growth through suppressing investments in health, education and physical capital, see Dasgupta and Ray (1986), Galor and Zeira (1993) and Banerjee and Newman (1994), respectively. On risk aversion as a poverty-perpetuating mechanism, see Banerjee (2000) and Dercon (2003). Note that Stiglitz (1969) first argues that risk aversion leads to under-investment. On poor institutions as potential sources of poverty traps, see Engerman and Sokoloff (2006). 27 Quah (1993), Bloom et al. (2003) and Azariadis and Stachurski (2004) found empirical evidence that cross-country distribution of per capita income exhibits bimodality. At the micro level, Jalan and Ravallion (2002) found a significant link between household consumption growth and aggregate (at the local level) physical and human capital endowments in China; which they argue is consistent with the existence of geographic poverty traps. 28 Lopez and Servén (2009) constructed their own poverty data using a log-normal approximation that dates back to Gibrat (1931). This approximation allows them to use per capita income data which are more available to estimate the less available poverty data for a given poverty line. The assumption that per capita income follows log-normal distribution is empirically tested and confirmed in Lopez and Servén (2005). 29 Both papers regress income growth rate on initial poverty and initial per capita income/consumption. The main difference is that Ravallion (2012) exploited only the cross-section dimension of the data; while Lopez and Servén (2009) exploited both the time-series and the cross-section dimensions of the data. More specifically, in Ravallion (2012)’s main analysis sample of ninety countries, each country is surveyed only twice. He then generated for each country an annualised growth rate of log per capita income or consumption for the period between the two surveys, and regressed this growth rate on the log headcount index in the first round of survey. Clearly, his main analysis sample is a cross-sectional data set. In Lopez and Servén (2009)’s analysis sample of some hundred countries, in contrast, each country can be surveyed multiple times; for every two surveys, there is a growth rate of log per capita income. Their analysis sample is therefore an unbalanced panel data set. They regressed this growth rate on log poverty ratio of the earlier round of survey between the two. The advantage of Ravallion (2012)’s strategy—which Lopez and Servén (2009) also used and reported results in their Table 8—is that it eliminates any concern about reverse causality. The potential loss is that it does not control for cross-country heterogeneity. Ravallion (2012), however, argues that FE regression has attenuation bias that is too strong to reveal any true relationship given the nature of his data (p. 517). 30 Mean convergence refers to the situation that countries with lower initial income enjoy faster mean income or consumption growth; sometimes, it also refers to the advantage of backwardness. The advantage of growth refers to that a higher mean income or consumption tends to come with a lower incidence of absolute poverty. Poverty convergence, where countries with higher initial poverty rate experience faster poverty reduction, is expected to follow as a consequence of mean convergence and the advantage of growth. 31 Ravallion (2012, p. 520) found that the direct effect of poverty is 0.026, the poverty elasticity effect is 0.016 and the sum of the two negative impact of initial poverty on poverty reduction almost exactly offset the mean convergence effect, or the advantage of backwardness that is −0.047. 32 We thank Ravallion (2012) for making his data set publicly available on the American Economic Review website. 33 The countries (first survey year, last survey year) in the 1978–2012 SSA subsample are as follows: Angola (2000, 2009), Botswana (1986, 1994), Burkino Faso (1994, 2009), Burundi (1992, 2006), Cote d'Ivoire (1985, 2008), Cameroon (1996, 2007), Central African Republic (1993, 2008), Ethiopia (1982, 2011), Gamiba (1998, 2003), Ghana (1988, 2006), Guinea (1991, 2007), Guinea-Bissau (1991, 2002), Kenya (1992, 2006), Lesotho (1987, 2003), Madagascar (1980, 2010), Malawi (1998, 2010), Mali (1994, 2010), Mauritania (1987, 2008), Mozambique (1997, 2008), Namibia (1993, 2004), Niger (1992, 2008), Nigeria (1986, 2011), Rwanda (1985, 2011), Senegal (1991, 2011), Seychelles (1995, 2007), Sierra Leone (1990, 2011), South Africa (1993, 2009), Swaziland (1995, 2010), Tanzania (1992, 2007), Togo (2006, 2011), Uganda (1989, 2010) and Zambia (1991, 2010). The 1978–2012 data set also contains five North African countries: Algeria (1988, 1995), Egypt (1991, 2009), Morocco (1985, 2007), Tunisia (1985, 2010) and Djibouti which however is only surveyed once (2000). 34 It is important to note that the P–I–G analysis and the earlier G–I–P analysis are not based on exactly the same data. The former uses only information collected from the first and the last survey in a certain country to explore the impact of initial poverty on subsequent growth. The latter, in contrast, used information collected from all surveys in a certain country to elicit the impact of growth and change in inequality on the pace of poverty reduction. Therefore, our decision to not to report the P–I–G results for the longer period in this paper does not suggest that our G–I–P results for the longer period should be subject to further validation. 35 Note that while Ravallion (2012) denotes the dependent variable in his empirical model for mean convergence (Equation (1), p. 511) as ∆lnμit≡ln(μit)−ln(μit−1), he did this for simplicity by assuming evenly spaced data. He actually regressed the annualised change in log mean income (gi(μit)≡(lnμit−lnμit−τ)/τ). Similarly, he regressed the annualised change in log poverty (gi(Hit)≡(lnHit−lnHit−τ)/τ) instead of ∆lnHit specified in his empirical model for poverty convergence (Equation (3), p. 511). 36 Ravallion (2012, Figure 1, p. 505) suggests a lack of unconditional poverty convergence for poverty measured against the $2/day line. He mentions that poverty convergence is still absent when he adds controls and/or uses different poverty lines (p. 512). 37 When we add North African countries to the 1978–2007 SSA subsample, we found only conditional poverty convergence in Africa during 1978–2007 for poverty measured against the $2 per day poverty line (−0.0595 with t = −2.4199 and R2 = 0.5497). 38 The strong link between income growth and poverty reduction can be quickly confirmed. For the SSA subsample, the regression coefficient of poverty reduction rate (annualised difference in log headcount index at $2 line) on income growth rate (annualised difference in log mean income) is −0.465 with t = −6.13 and R2 = 0.660. For the full sample, Ravallion (2012, p. 512) reports that the regression coefficient of annualised difference in log headcount index at $2 line on annualised difference in log mean income is −1.085 with a t-ratio of −4.128 and R2 of 0.377. 39 It might be conjectured that initial poverty is picking up some other aspects of initial distribution, especially inequality which some literature has identified as having a direct negative impact on income growth. To investigate this, we ran Equation (7) with a set of controls including log of initial Gini index, initial per capita consumption from national account, primary school enrolment rate, life expectancy at birth and the relative price index of investment goods. It turns out the findings presented in Tables 9 and 10 still hold. Regression results are available upon request. References African Development Bank ( 2012) African Economic Outlook 2012 . Tunis, Tunisia: Abidjan, Côte d'Ivoire. PubMed PubMed  Alesina A., Rodrik D. ( 1994) ‘ Distributive Policies and Economic Growth’, Quarterly Journal of Economics , 109 ( 2): 465– 90. Google Scholar CrossRef Search ADS   Azariadis C., Stachurski J. ( 2004) A Forward Projection of the Cross-country Income Distribution . Belgium: Mimeo, Universite Catholique of Louvain. Azariadis C., Stachurski J. ( 2005) ‘Poverty traps’, in P. Aghion, S. Durlauf (eds) Handbook of Economic Growth , Vol. 1. Amsterdam, Netherlands: Elsevier (North Holland Publishing Co.), pp. 295– 384. Banergee A. ( 2000) ‘ The Two Poverties’, Working Paper 01-16, Department of Economics, M.I.T. Boston, Massachusetts, USA. Banerjee A. V., Duflo E. ( 2003) ‘ Inequality and Growth: What Can the Data Say’, Journal of Economic Growth , 8 ( 3): 267– 99. Google Scholar CrossRef Search ADS   Banerjee A. V., Newman A. F. ( 1994) ‘ Poverty, Incentives, and Development’, American Economic Review , 84 ( 2): 211– 5. Barro R. J. ( 2000) ‘ Inequality and Growth in a Panel of Countries’, Journal of Economic Growth , 5 ( 1): 5– 32. Google Scholar CrossRef Search ADS   Baum C. F., Schaffer M. E., Stillman S. ( 2003) ‘ Instrumental Variables and GMM: Estimation and Testing’, Stata Journal , 3 ( 1): 1– 31. Bloom E., Canning D., Sevilla J. ( 2003) ‘ Geography and Poverty Traps’, Journal of Economic Growth , 8 ( 4): 355– 378. Google Scholar CrossRef Search ADS   Bourguignon F. ( 2003) ‘The growth elasticity of poverty reduction: explaining heterogeneity across countries and time periods’, in Inequality and Growth: Theory and Policy Implications . Washington, DC: World Bank, pp. 3– 26. Cragg J. G., Donald S. G. ( 1993) ‘ Testing Identifiability and Specification in Instrumental Variable Models’, Econometric Theory , 9 ( 02): 222– 40. Google Scholar CrossRef Search ADS   Dasgupta P., Ray D. ( 1986) ‘ Inequality as a Determinant of Malnutrition and Unemployment: Theory’, The Economic Journal , 96 ( 384): 1011– 34. Google Scholar CrossRef Search ADS   Dercon S. ( 2003) Risk and Poverty: A Selective Review of the Issues . Oxford, UK: Mimeo, Oxford University. Easterly W., Levine R. ( 1997) ‘ Africa's Growth Tragedy: Policies and Ethnic Divisions’, The Quarterly Journal of Economics , 112 ( 4): 1203– 50. Google Scholar CrossRef Search ADS   Engerman S. L., Sokoloff K. L. ( 2006) ‘The persistence of poverty in the Americas: The role of institutions’, in S. Bowless, S. Durlauf, K. Hoff (eds) Poverty Traps . New Jersey: Princeton University Press, pp. 43– 78. Englebert P. ( 2000) ‘ Solving the Mystery of the AFRICA Dummy’, World Development , 28 ( 10): 1821– 35. Google Scholar CrossRef Search ADS   Fosu A. K. ( 2009) ‘ Inequality and the Impact of Growth on Poverty: Comparative Evidence for Sub-Saharan Africa’, Journal of Development Studies , 45 ( 5): 726– 45. Google Scholar CrossRef Search ADS   Fosu, A. K. ( 2011) Growth, Inequality, and Poverty Reduction in Developing Countries: Recent Global Evidence . UNU-WIDER Working Paper 2011-001. Helsinki, Finland: UNU-WIDER. Fosu A. K. ( 2015) ‘ Growth, Inequality and Poverty in Sub-Saharan Africa: Recent Progress in a Global Context’, Oxford Development Studies , 43 ( 1): 44– 59. Google Scholar CrossRef Search ADS   Forbes K. ( 2000) ‘ A Reassessment of the Relationship Between Inequality and Growth’, American Economic Review , 90 ( 4): 869– 87. Google Scholar CrossRef Search ADS   Galor O., Zeira J. ( 1993) ‘ Income Distribution and Macroeconomics’, Review of Economic Studies , 60 ( 1): 35– 52. Google Scholar CrossRef Search ADS   Gibrat R. ( 1931) Les Inégalités Économiques. Paris: Sirey. Hauk W. R. Jr, Wacziarg R. ( 2009) ‘ ‘A Monte Carlo Study of Growth Regressions’’, Journal of Economic Growth , 14 ( 2): 103– 47. Google Scholar CrossRef Search ADS   Jalan J., Ravallion M. ( 2002) ‘ Geographic Poverty Traps? A Micro Model of Consumption Growth in Rural China’, Journal of Applied Econometrics , 17 ( 4): 329– 46. Google Scholar CrossRef Search ADS   Li H., Zou H. F. ( 1998) ‘ Income Inequality is Not Harmful for Growth: Theory and Evidence’, Review of Development Economics , 2 ( 3): 318– 34. Google Scholar CrossRef Search ADS   Lopez H., Servén L. ( 2005) ‘ A normal relationship? Poverty, growth, and inequality’, World Bank Policy Research Working Paper 3814. Washington DC, USA. Lopez, H., Servén, L. ( 2009). Too Poor to Grow. World Bank Policy Research Working Paper Series 5012, http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.493.8341&rep=rep1&type=pdf. Ndulu B. J. (ed.) ( 2008) The Political Economy of Economic Growth in Africa, 1960–2000 , Vol. 2. Cambridge, UK: Cambridge University Press. Perotti R. ( 1996) ‘ Growth, Income Distribution, and Democracy: What the Data Say’, Journal of Economic growth , 1 ( 2): 149– 87. Google Scholar CrossRef Search ADS   Perry G., Arias O., López H., Maloney W., Servén L. ( 2006) Poverty Reduction and Growth: Virtuous and Vicious Circles . World Bank Publications. Google Scholar CrossRef Search ADS   Quah D. ( 1993) ‘ Empirical Cross-section Dynamics in Economic Growth’, European Economic Review , 37 ( 2): 426– 34. Google Scholar CrossRef Search ADS   Ravallion M. ( 2012) ‘ Why Don't We See Poverty Convergence?’, American Economic Review , 102 ( 1): 504– 23. Google Scholar CrossRef Search ADS   Shimeles, A., E. Thorbecke, and Y. Ouyang ( 2016) Why Don't We See Poverty Convergence in Sub-Saharan Africa?  Paper submitted for possible publication. Stiglitz J. ( 1969) ‘ The Effects of Income, Wealth and Capital Gains Taxation on Risk-Taking’, The Quarterly Journal of Economics , 83 ( 2): 263– 83. Google Scholar CrossRef Search ADS   Stock J. H., Wright J. H., Yogo M. ( 2002) ‘ A Survey of Weak Instruments and Weak Identification in Generalized Method of Moments’, Journal of Business & Economic Statistics , 20 ( 4): 518– 29. Google Scholar CrossRef Search ADS   Stock J. H., Yogo M. ( 2005) ‘Testing for weak instruments in linear IV regression’, in D. W. K. Andrews, J. H. Stocks (eds) Identification and Inference for Econometric Models: Essays in Honor of Thomas Rothenberg . Cambridge, UK: Cambridge University Press. Google Scholar CrossRef Search ADS   Thorbecke E. ( 2014) ‘How inclusive is the present growth pattern in sub-Saharan Africa?’, in E. Zedillo, O. Cattaneo, H. Wheeler (eds), Africa at a Fork in the Road. Taking Off or Disappointment Once Again?  Connecticut, USA: Yale Center for the Study of Globalization. Thorbecke E. ( 2015) ‘The anatomy of growth and development in sub-Saharan Africa’, in A. McKay, E. Thorbecke (eds) Economic Growth and Poverty Reduction in Sub-Saharan Africa: Current and Emerging Issues . Oxford, UK: Oxford University Press. Google Scholar CrossRef Search ADS   Thorbecke E., Ouyang Y. ( 2016) ‘Is sub-Saharan Africa finally catching up?’, in M. Andersson, T. Axelsson (eds) Diverse Development Paths and Structural Transformation in the Escape from Poverty . Oxford, UK: Oxford University Press, pp. 236– 65. Google Scholar CrossRef Search ADS   UNDP ( 2014) Human Development Report 2014 . Washington, DC: United Nations Development Program. © The Author 2017. Published by Oxford University Press on behalf of the Centre for the Study of African Economies. All rights reserved. For permissions, please email: journals.permissions@oup.com http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of African Economies Oxford University Press

Is the Structure of Growth Different in Sub-Saharan Africa?

Journal of African Economies , Volume 27 (1) – Jan 1, 2018

Loading next page...
 
/lp/oxford-university-press/is-the-structure-of-growth-different-in-sub-saharan-africa-8HrkNFDoHq

References (37)

Publisher
Oxford University Press
Copyright
© The Author 2017. Published by Oxford University Press on behalf of the Centre for the Study of African Economies. All rights reserved. For permissions, please email: journals.permissions@oup.com
ISSN
0963-8024
eISSN
1464-3723
DOI
10.1093/jae/ejw032
Publisher site
See Article on Publisher Site

Abstract

Abstract Taking advantage of the 2014 issue of the Povcalnet data and building on earlier works, this paper provides an anatomy of growth in sub-Saharan Africa (SSA) and compares it with that of the developing world (including or excluding SSA) during the three decades between the early 1980s and the early 2010s. We examined both the impact of the pattern of growth on poverty and inequality (the growth–inequality–poverty (G–I–P) nexus); and the less studied reverse causal chain emanating from poverty to subsequent inequality and growth (the poverty–inequality–growth (P–I–G) nexus). For the G–I–P nexus, we found that poverty reduction is becoming more responsive to income growth and improvement in inequality in SSA in the post-2007 years; though the responsiveness remains smaller in SSA than in the developing world (with or without SSA) throughout the three decades examined. For the P–I–G nexus, we found that SSA again differs from the rest of the developing world in that SSA countries with the highest initial poverty incidence appeared to grow subsequently faster—leading to poverty convergence; while the developing world (with or without SSA) experienced a lack of poverty convergence during the same time period. We hypothesise that the main cause of poverty convergence in SSA during 1978–2007 might be due to anti-poverty interventions by governments and foreign public and private aid inversely proportional to the depth of poverty. 1. Introduction The questions at the heart of this paper are, first, whether the structure of growth in sub-Saharan Africa (SSA) is different from that of the developing world (including or excluding SSA) and, second, whether the recent pattern of growth in SSA is different from what it was prior to the beginning of this millennium. We hope to show that both of these questions can be answered affirmatively and we attempt to provide explanations for those differences. These questions are particularly relevant in the light of the extensive literature prior to the new millennium that argued that there were some structural factors such as ethno-linguistic heterogeneity, the peculiar historical origins of African (artificial) states, and endowment constraints such as land-lockedness and tropical climate that contributed to an environment making the growth pattern in SSA different from that in the rest of the world. Even controlling for those obstacles, researchers running international growth regressions almost invariably came up with an ‘African dummy’ to account (in a residual and unexplained way) for the stagnating growth in the region.1 In retrospect, it can be argued that it is the poor quality of governance that was the main culprit for the dismal performance of SSA economies before 2000 (Englebert, 2000; Thorbecke and Ouyang, 2016). By the structure of growth, we mean the dynamic interaction among growth, inequality and poverty (G–I–P) during the process of economic development. A major factor influencing this interaction is the speed and form of the structural transformation. Since around 2000, the pace of growth in SSA has undergone a quantum jump and the pattern of growth has become more inclusive. The structural transformation that was typically flawed in many SSA countries until recently is now more successful. The agricultural sector that in the past was, at best, ignored by the policymakers and, at worst, mercilessly exploited is now starting to get the attention it deserves. Workers released from agriculture move into more productive jobs in other sectors. For thirteen out of fourteen SSA countries for which at least two annual observations were available between 2000 and around 2010, the structural transformation was successful in the sense that the fall in the share of the labour force in agriculture was associated with an increase in per capita GDP (Thorbecke, 2014). Further evidence of progress towards a more inclusive growth pattern is that non-monetary dimensions of well-being improved as well. The Human Development Report (UNDP, 2014) indicates that out of the fourteen best performers as measured by the annual growth rate of the Human Development Index between 2000 and 2013, eleven were from SSA. Starting from a low baseline, many African countries have enjoyed significant improvements in infant mortality, health and school enrolment. The current spell of growth in SSA has been influenced by a number of factors. Among the endogenous factors, at least partially under the control of African states, are (i) the improved treatment of agriculture and, more generally, the overall improvement in the quality of governance; and (ii) the appearance of a middle class. Among the most important exogenous factors are (i) the large jump in the flow of direct foreign investment into the subcontinent and (ii) high global commodity prices and consequent favourable terms of trade (Thorbecke, 2014). Hence our conjectures, that will be subsequently tested, are that the current structure of growth in SSA is significantly different from that prevailing during the period 1960–2000 and also different from the global pattern of growth. In order to test those conjectures, the anatomy of growth, and more particularly the interrelationship among G–I–P, has to be analysed. This is best done by focusing on the G–I–P interrelationship. As these three concepts are not only interrelated but also jointly determined, one has to be careful in any attempt to infer unidirectional causality among them. As will be described below, one can best imagine a G–I–P triangle with influence travelling in both directions. In analysing the transmission of influence in such a socio-economic system, one can start at any one of the three corners. Traditionally, development economics focused on the directional effects emanating from income growth to inequality and poverty. In the remainder of this paper, we refer to this case as the G–I–P nexus. It is only recently that the lens has been placed on the reverse impact of poverty on growth and inequality which we refer to as the poverty–inequality–growth (P–I–G) nexus. The rest of the paper is organised as follows: Section 2 presents and analyses the G–I–P nexus and each of the links in the transmission of influence originating with growth. The impact on inequality and poverty of different structures of growth, such as pro-poor growth, exclusive and inclusive growth, is explored. A case is made that recent SSA growth has become more inclusive. Section 3 focuses on the P–I–G nexus and reviews recent research findings that high initial poverty and inequality levels affect subsequent growth negatively. Section 4 uses two different data sets to estimate the global and SSA structural coefficients of the G–I–P nexus based on Bourguignon's (2003) identity model. Our results confirm our conjectures that the historical growth pattern in SSA is different from that of the developing world (including or excluding SSA); and that the recent SSA growth spell is different from that prevailing up to the mid 2000s. Section 5, in turn, estimates the structural coefficients of the P–I–G nexus and finds that, in contrast with the results, Ravallion (2012) obtained from a globally representative sample of about hundred developing countries, high initial poverty does not appear to slow down subsequent growth in the African subcontinent. Finally, Section 6 concludes. 2. The impact of growth on inequality and poverty The chain of influence linking growth to poverty has been thoroughly researched and is relatively clear and well understood. It is based on the G–I–P interrelationship. The pace of growth and its pattern, in any given country, are determined by the forces of globalisation and the development strategy that is being followed. The process of globalisation is largely exogenous (outside the control of the State), while the development strategy is at least partially endogenous (under the control of the State). Depending on the specific initial conditions prevailing in a country, the combined effects of globalisation and the development strategy will give rise to, and result in different speeds of growth and different structures of growth from exclusive to inclusive. The less unequal the income distribution and the more inclusive the pattern of growth are, the greater the impact of growth will be on poverty reduction. By now there is a rich literature on pro-poor growth. There are essentially two approaches and definitions. The relative definition of pro-poor growth is that the poor benefit proportionately more than the non-poor from the prevailing growth, which implies a fall in income inequality. The absolute definition only requires that the poor benefit from growth.2 Hence, under this approach inequality can increase and still be compatible with pro-poor growth. More recently, inclusive growth has become the new paradigm adopted by the development community. The African Development Bank (2012) defines inclusive growth as ‘economic growth that results in a wider access to sustainable socio-economic opportunities for a broader number of people, regions or countries, while protecting the vulnerable, all being done in an environment of fairness, equal justice, and political plurality’ (p. 2). The inclusive growth approach takes a longer term perspective as the focus is on productive employment as the main instrument, rather than on direct income redistribution, as a means of increasing incomes for excluded groups. Unlike the pro-poor growth agenda that focuses mainly on the welfare of the poor, inclusive growth is concerned with opportunities in the labour force for poor and middle class alike. Figure 1 presents schematically the G–I–P nexus. Each of the links in the causal chain plays a role in the transmission of the combined effects of globalisation (e.g., through trade and foreign direct investment) and the development strategy (e.g., country-specific policies and institutions)—as prime movers. These links have been described in detail in Thorbecke (2015) and are only very briefly reviewed here. Figure 1: View largeDownload slide Globalisation and Development Strategy and Interrelationship Among G–I–P. Figure 1: View largeDownload slide Globalisation and Development Strategy and Interrelationship Among G–I–P. Starting, first, on the left-hand side of Figure 1, the two prime movers affect the pattern of growth in a given economy. In SSA, before around 2000, most countries stagnated and the structure of growth tended to be exclusive and contributing to a more unequal income distribution. In contrast, the present growth spell is characterised by a quantum jump in the pace of growth and some evidence that the pattern had become more inclusive. Hence, the G–P link (the upper left arrow in Figure 1), that in the pre-2000 era, was very weak, in the sense that the growth elasticity of poverty reduction in SSA was only about half that in other developing regions, in the recent, spell has become stronger resulting in a significant reduction in poverty. Income inequality is still relatively very high in the African subcontinent—although it is falling in a number of countries. The I–P link (lower left arrow) reveals the role of inequality as a filter between growth and poverty reduction. A more inclusive pattern of growth can keep inequality in check with favourable consequent poverty outcomes. Next, the I–G link is subject to two conflicting theoretical approaches. The Neo-classical theory argues that an uneven income distribution is a pre-condition to growth as the rich have a higher marginal propensity to save than the poor. Hence, for any given total GDP, an unequal income distribution will generate a larger flow of savings cum investment. On the other hand, the New Political Economy of Development argues convincingly that greater income inequality is likely to dampen growth through a variety of channels, such as the diffusion of political and social instability, unproductive rent-seeking activities and increased insecurity of property rights. We subsequently explore this proposition and find some support for it. More specifically in Section 4, we found that the impact of reducing high initial inequality through income growth has become more important for effective poverty reduction in the current SSA growth period. The final link appearing in Figure 1 is the G–I link, which is still referred to in the literature as ‘Kuznets’ Law’, i.e. that at very low levels of development, growth will be un-equalising up to a threshold per capita income amount and beyond that equalising. The present consensus is that this ‘law’ does not hold as a generalisation and lost its immutability. The specific initial conditions and policies of an economy largely determine the impact of growth on inequality. One of the earliest attempts to formalise the interrelationship among the G–I–P nexus is the identity model built by Bourguignon (2003). The identity model assumes a log-normal income distribution and can be used to explain the heterogeneity of G, I, P outcomes depending on country-specific conditions and more specifically the wide range of growth elasticities of poverty reduction observed among countries. Essentially, the dependent variable is the growth in poverty expressed as a function of initial values and growth rates of income and inequality, the ratio of the initial ratio of the poverty line to average income and some interacting terms. Even though, one cannot make any causal inferences on the basis of an identity model, the Bourguignon model has been used to shed light on the relative roles and importance of income and inequality in reducing poverty in different regions and countries. Fosu (2015) applied the identity model to a sample of some hundred developing countries using observations from growth spells between 1981 and 2007. He concluded that ‘Viewed within a global context … the relatively low levels of income appear to be a major factor in inhibiting the effectiveness of income and inequality improvements in producing poverty reduction in SSA countries generally’ (p. 56). In summary, the impact of growth on poverty is relatively clear and well understood. In contrast the reverse influence from poverty and inequality to subsequent growth has only recently been analysed and identified as a potentially important transmission mechanism. A number of empirical studies have thrown light on the major effect initial poverty (and inequality) can have on future growth. This is the theme of the next section. 3. The impact of poverty and inequality on subsequent growth In the past, this link tended to be dismissed on the ground that any policies directed to reducing poverty detracted from growth. This was based on the firm belief of a trade-off between equity and efficiency. Any measure to reduce poverty today would reduce future growth. A number of recent empirical studies have questioned this view and thrown light on the major effect high initial poverty (and inequality) can have on future growth. An early and path-breaking study (Perry et al., 2006) made a case for a pro-growth poverty reduction strategy on the ground that there are multiple channels through which the existence of poverty acts as a major obstacle to growth. Examples of such channels and poverty traps are that poor people (i) have limited access to credit and financial markets, which seals them off potentially profitable and productive investment opportunities; (ii) often suffer from ill health and malnutrition that affects their productivity; (iii) attend low-quality schools that constrain their human capital. Careful econometric work by Lopez and Servén (2009) provided empirical support for the contention that high poverty can be a major obstacle to growth. They used a standard growth model augmented to include a poverty measure among the explanatory variables and controlled for other factors. Estimating the resulting specification on a large country panel data set using a generalised method of moments (GMM) approach to control for endogeneity of other regressors, the authors found that poverty has a negative impact on growth that is significant both statistically and economically. On average, a 10% increase in poverty reduces annual growth by 1%. The mechanism retarding growth in countries with high initial poverty was found to be the deterrence of investment, especially when the degree of financial development is limited. The next important empirical case that high poverty incidence could reduce subsequent growth was made by Ravallion (2012). In an attempt to explain why poverty convergence is not occurring worldwide, he used a sample of growth spells from almost hundred developing countries covering the period from 1978 to 2007. Among his main findings are that (i) high initial poverty rates have sizeable negative impacts on the growth rate; (ii) it is high poverty, not inequality per se, that retards growth; (iii) to the extent that higher overall inequality comes with higher poverty at a given mean, it yields lower growth rates and (iv) the growth elasticity of poverty reduction tends to be smaller in countries with a higher initial poverty rate. The immediate question that is elicited by these results is that of endogeneity between growth and poverty. Slow growth contributes to (if not causes) poverty. No wonder then that initial high poverty would be associated with low growth. However, Ravallion (2012) rules out this interpretation by making clear that ‘We see that the finding that a higher initial poverty rate implies a lower subsequent growth rate in the mean (at given initial mean) is robust to allowing for the possible endogeneity of the initial mean and initial poverty rate’ (p. 514). The common theme of the above studies is that high poverty today dampens subsequent growth. Yet, surprisingly, in Section 5 where we apply the Ravallion's (2012) model to his SSA sample of 28 countries, we find little evidence of any association between initial poverty levels and subsequent growth rates—contrary to the global pattern showing that high initial poverty was positively and significantly correlated with lower subsequent growth (demonstrating a lack of poverty convergence). Figure 2 captures the comparative static effects over two time periods (t) and (t + 1) and shows graphically the connection between the traditional G–I–P nexus and the P–G–I nexus. The left-hand side of Figure 2 reproduces the G–I–P nexus as it appears in Figure 1, while the right-hand side of Figure 2 indicates the channels through which the initial incidence of poverty at time (t) influences subsequent growth and inequality at time (t + 1). Figure 2: View largeDownload slide The G–I–P Nexus at Time t and the Reverse P–G–I Nexus at Time (t + 1). Figure 2: View largeDownload slide The G–I–P Nexus at Time t and the Reverse P–G–I Nexus at Time (t + 1). In the next section, we attempt to test whether the SSA growth structure is different from the global structure. We use the Bourguignon identity model to estimate the structural coefficients of the G–I–P nexus for the SSA sample and compare the SSA estimates with those derived from the global data sets. 4. Estimating the SSA structural coefficients of the G–I–P nexus 4.1 Econometric model To explore the structural coefficients of the G–I–P nexus in SSA, we used the identity model initially proposed by Bourguignon (2003) and subsequently developed by Fosu (2009, 2011, 2015) into the following log-linear model3:   p=b1+b2y+b3yG′+b4y(Z/Y)+b5g+b6gG′+b7g(Z/Y)+b8G′+b9(Z/Y), (1) where p is the annualised change in log poverty rate; y is the annualised growth rate of log per capita income; g is annualised change in log Gini coefficient; G′ is the log of initial Gini; (Z/Y) is the ratio of the poverty line Z to initial income Y, also expressed in natural logarithm. All changes are between two surveys and are annualised. The coefficients to be estimated are bi(i=1,…,9). The sign of b2 is expected to be negative, as income growth should lead to poverty reduction, ceteris paribus. The signs of b3 and b4 are expected to be positive, as both higher initial inequality (G′) and lower initial income (higher Z/Y) should weaken the effect of income growth (y) on poverty reduction. The sign of b5 is expected to be positive, as rising inequality is often associated with increasing poverty. But a negative b5 is also possible, in which case rising inequality contributes to poverty reduction. The sign of b6 can be either positive or negative: if it is positive, then rising inequality will have a greater (negative or positive depending on the sign of b5) impact on poverty reduction in countries with higher initial inequality; if it is negative, then rising inequality will have a smaller impact on poverty reduction in these countries. Similarly, the sign of b7 can also be either positive or negative; in the latter case rising inequality has a smaller impact on poverty reduction in countries with higher Z/Y (higher initial poverty or lower initial income). Finally, the signs of b8 and b9 are expected to be positive, as both higher initial inequality and higher initial poverty should impede poverty reduction. Using the estimates of b’s from equation (1), we can derive the growth elasticity of poverty reduction ( Ey) and inequality elasticity of poverty reduction ( Egini):   Ey=b2+b3G′+b4Z/Y (2)  Egini=b5+b6G′+b7Z/Y (3) The elasticity equations clearly suggest that the growth effectiveness of poverty reduction is not only related to income growth per se, but also to the initial inequality level and initial income level. Similarly, initial levels of inequality and income also affect how sensitive the change in poverty is to a change in inequality. It should also be noted that the sign of Ey is expected to be negative as growth reduces poverty, while the sign of Egini is expected to be positive as rising inequality increases poverty. To see whether the growth pattern in SSA differs from that in the developing world (including or excluding SSA), we need to estimate b1–b9 and calculate the elasticities for the SSA region, the entire developing world and the non-SSA developing world. We used two data sets for our empirical exploration. The first data set is essentially the same as that used by Fosu (2015) who derived country-level data from the 2009 issue of the World Bank Povcalnet database4. While this data set—hereinafter referred to as the Povcalnet 1 data set—has in total 539 observations from 124 developing countries covering the 1977–2007 period for the $1.25 a day poverty line, our regressions are based on 368 usable observations collected in 94 developing countries during 1982–2007.5 The second data set is an update of the Povcalnet 1 data set, which takes advantage of the 2014 issue of the Povcalnet database. This most recent and enriched data set—hereinafter referred to as the Povcalnet 2 data set—contains 702 usable observations collected from 102 developing countries during 1986–2012. If the 1986–2007 sub-set of the Povcalnet 2 data included the same information as the Povcalnet 1 data set for the same timespan, as we initially assumed, then any difference between regression results derived from these two data sets should reveal the change in growth structure within a region in the post-2007 years. Unfortunately, this is not the case as it appears that the World Bank periodically revises and updates its historical data sets.6 Hence, a comparison of regressions run on the two data sets reflects not only the structural changes that might have occurred in the post-2007 period but also the revisions made by the World Bank. One issue that needs to be highlighted at the outset is that of the representativeness of the SSA subsample. The Povcalnet 2 data set for SSA that includes 103 observations collected from 35 SSA countries over the timespan 1986–2012 can be claimed to be representative.7 There is, however, a concern about the representativeness of the Povcalnet 1 SSA subsample. The Povcalnet 1 SSA subsample we used in this analysis contains 66 observations collected from 28 SSA countries during 1987–2006.8 The relatively small number of observations and the fact that some countries with multiple observations (e.g., Côte d'Ivoire) are given unduly high weight raise serious questions about the representativeness of this SSA subsample. Due to this concern, Fosu (2015) decided not to estimate coefficients separately for the SSA subsample, but instead to use structural coefficients of the entire developing world for his analysis of SSA growth structure. While these are valid concerns and should be kept in mind when interpreting our regression results, we decided to estimate SSA's own structural coefficients during 1987–2006 and reported them in Table 1 column (2). For one thing, compared with the SSA subsample Fosu (2015) would have used, the Povcalnet 1 SSA subsample we used in this analysis is slightly more representative in the sense that ours contains more observations from more countries collected during a longer period of time.10 An important result is that we find that the inequality elasticity of poverty reduction ( Egini) for SSA taken as a whole is very different depending on the structural coefficients used: SSA Egini estimated using SSA's own structural coefficients is only about half of that estimated using structural coefficients of the entire developing world.11 This suggests that estimating SSA's own structural coefficients is necessary for our understanding of the G–I–P nexus in SSA; though the region has a less than ideal sample size and composition. Table 1: SSA Versus Global Structural Coefficients, Povcalnet 1 Data Set, 1982–2007   Global  SSA  Non-SSA  1982–2007  1987–2006  1982–2007  (1)  (2)  (3)  Growth of log income (b2)  −12.49***  −9.01**  −15.02***  (−5.52)  (−2.27)  (−6.80)  Growth of log income × log initial Gini (b3)  3.04***  2.00*  4.03***  (4.88)  (1.93)  (6.15)  Growth of log income × log (poverty line/initial income) (b4)  1.33***  0.53***  2.14***  (4.43)  (2.59)  (6.25)  Change in log Gini (b5)  19.90***  −9.46  24.14***  (3.42)  (−1.49)  (3.78)  Change in log Gini × log initial Gini (b6)  −5.00***  2.65  −6.34***  (−3.24)  (1.57)  (−3.35)  Change in log Gini × log (poverty line/initial income) (b7)  −2.46***  −0.60  −3.08***  (−4.03)  (−1.01)  (−3.30)  Log initial Gini (b8)  0.11  0.01  0.19  (0.66)  (0.13)  (0.93)  Log (poverty line/initial income) (b9)  0.01  0.12  0.02  (0.17)  (1.03)  (0.20)  Intercept (b1)  −0.43  −0.0003  −0.72  (−0.71)  (−0.001)  (−1.00)  N  368  66  302  R2  0.69  0.89  0.70    Global  SSA  Non-SSA  1982–2007  1987–2006  1982–2007  (1)  (2)  (3)  Growth of log income (b2)  −12.49***  −9.01**  −15.02***  (−5.52)  (−2.27)  (−6.80)  Growth of log income × log initial Gini (b3)  3.04***  2.00*  4.03***  (4.88)  (1.93)  (6.15)  Growth of log income × log (poverty line/initial income) (b4)  1.33***  0.53***  2.14***  (4.43)  (2.59)  (6.25)  Change in log Gini (b5)  19.90***  −9.46  24.14***  (3.42)  (−1.49)  (3.78)  Change in log Gini × log initial Gini (b6)  −5.00***  2.65  −6.34***  (−3.24)  (1.57)  (−3.35)  Change in log Gini × log (poverty line/initial income) (b7)  −2.46***  −0.60  −3.08***  (−4.03)  (−1.01)  (−3.30)  Log initial Gini (b8)  0.11  0.01  0.19  (0.66)  (0.13)  (0.93)  Log (poverty line/initial income) (b9)  0.01  0.12  0.02  (0.17)  (1.03)  (0.20)  Intercept (b1)  −0.43  −0.0003  −0.72  (−0.71)  (−0.001)  (−1.00)  N  368  66  302  R2  0.69  0.89  0.70  Notes: This table gives b1ˆ–b9ˆ in Bourguignon's Identity Model using the Povcalnet 1 data set and the fixed-effect (FE) regression procedure. Log (poverty line/initial income) measures initial poverty condition, where poverty line is $1.25/day. T-ratios are in parenthesis and based on White standard errors (are heteroscedasticity-consistent). T-ratios are compared against critical values adjusted for sample sizes.9 ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. We applied three regression procedures to the Bourguignon model: Ordinary Least Square (OLS), FE and the GMM. The OLS and FE results are very similar (in both magnitude and significance level) whether they are applied to the Povcalnet 1 or the Povcalnet 2 data sets and their respective SSA subsamples. The GMM procedure, however, generates only significant estimates for the Povcalnet 1 data set12 but insignificant results for the Povcalnet 2 data set, even though the sample size of the latter is almost twice that of the former.13 The poor performance of GMM estimates may be related to weak identification (and even under-identification in some cases) of the GMM model used, which we inherited from Fosu (2011, 2015) and used as instrumental variables (IVs) lagged income, lagged Gini index, and their interaction terms with regional dummies.14 Specifically, the identification tests statistics suggest that these IVs are only weakly related to the endogenous regressor they are supposed to instrument (income growth) no matter which of the two data sets and their SSA subsamples are used;15 further, these IVs are barely relevant to the endogenous regressor when the procedure is run for the Povcalnet 1 SSA subsample.16 In the presence of weak identification, GMM estimates may perform poorly (Stock et al., 2002; Stock and Yogo, 2005); and under-identification of the model when applied to Povcalnet 1 SSA subsample means that little is gained from instrumenting. Following suggestion of Baum et al.(2003), we tried to address the issue of weak identification by reducing the number of IVs; but this effort was in vain. Lagged variables are usually viewed as good instruments that address not only the endogeneity issue but also measurement errors (Lopez and Servén, 2009; Ravallion, 2012). However, if over time some structural change happened to the growth pattern, lagged income and lagged Gini coefficients may no longer be valid instruments for income growth. We suspect this may well be the case especially for the SSA region which went through significant structural changes since around 2000 as discussed in Section 1. Further, we wonder if the GMM regression procedure is really necessary in our analysis. After all, GMM estimator comes with a price: it is less efficient compared with the OLS estimator (which is but less efficient than the FE estimator); and it has poor finite sample performance compared with regular IV estimator (which is a special case of GMM). Econometricians agree that GMM is appropriate if endogeneity and heteroscedasticity exist at the same time. While the presence of heteroscedasticity in our data is confirmed by the Pagan–Hall test statistic,17 we find little evidence that income growth endogeneity is a serious concern in our data. The endogeneity test suggests that for the Povcalnet 1 global sample, the Povcalnet 2 global sample, and its SSA subsample, income growth can all be treated as exogenous.18 Due to the above considerations, we report only b1–b9 estimates from the FE regression procedure.19 Compared with OLS and GMM, the FE approach has three advantages: (i) it addresses endogeneity related to cross-country heterogeneity, which OLS cannot address but is demonstrated to have been significant among SSA countries by Fosu (2015); and (ii) the FE estimator is more efficient than the OLS estimator as it addresses serial correlation of the error terms, yet OLS estimator is more efficient than the GMM estimator; and (iii) the FE method allows slightly larger sample size than the GMM estimator as it has fewer explanatory variables. One major concern of the FE approach is that its attenuation bias may be too strong to reveal true relationship when applied to data with independent measurement errors in each period (Hauk and Wacziarg, 2009; Ravallion, 2012). Our FE results however are significant, suggesting that the true relationship may only be stronger. 4.2 Empirical results Table 1 presents our FE estimates of b1–b9 based on the Povcalnet 1 sample, its SSA subsample, and its non-SSA subsample. Since the comparison between the full sample and the SSA subsample conveys the same message as the comparison between the non-SSA and the SSA subsamples, we shall focus on explaining the former. The estimates for the entire developing world reported in column (1) are similar to those reported in Fosu (2015) following a GMM regression procedure, suggesting that the FE method is a valid substitute of the GMM method at least for the global sample. Table 1 column (2) presents the FE estimates of structural coefficients based on 66 observations from 28 SSA countries collected during 1987–2006. As mentioned at the outset of this section, the Povcalnet 1 SSA subsample is less ideal in sample representativeness. It does, however, allow us to see that the growth structure in SSA is very different from that of the entire developing world. Given the above estimates, we then calculated the growth elasticity of poverty reduction ( Ey) and inequality elasticity of poverty reduction ( Egini) for SSA and the entire developing world using formula provided in Equations (3) and (4). We report our elasticity estimates in Table 2. Note that these estimates are similar to those reported in Fosu (2009, Table 2) using data collected during 1982–2004.20 Table 2: Growth and Inequality Elasticity of Poverty Reduction for Global and SSA Regions, Povcalnet 1 Data Set, 1982–2007 Region  Entire developing world  SSA  Non-SSA developing world  1982–2007  1987–2006  1982–2007  Ey  −2.91  −1.56  −3.26  Egini  4.48  0.62  5.23  Estimates from  368 observations from 94 countries  66 observations from 28 countries  302 observations from 66 countries  Region  Entire developing world  SSA  Non-SSA developing world  1982–2007  1987–2006  1982–2007  Ey  −2.91  −1.56  −3.26  Egini  4.48  0.62  5.23  Estimates from  368 observations from 94 countries  66 observations from 28 countries  302 observations from 66 countries  Comparing b1–b9 in the two columns of Table 1, we see that during 1982–2007, growth (as reflected by b2) had a significant but much smaller impact on poverty reduction in SSA than in the entire developing world (−9.01** versus −12.49***). Further, as reflected by b3 and b4, high initial inequality and low initial income significantly weaken the growth elasticity of poverty reduction ( Ey) in SSA at about the same relative scale and strength as they weaken Ey in the entire developing world. Consequently, the growth elasticity of poverty reduction in SSA was only about a half of that in the developing world (Table 2 row 1: −1.56 versus −2.91). Compared with the entire developing world, the SSA region also had a much smaller inequality elasticity of poverty reduction ( Egini) during 1982–2007 (Table 2 row 2: 0.62 versus 4.48). The main reason seems to be that rising inequality had little impact (as reflected by b5) on poverty reduction in SSA taken as a whole but a very strong poverty-increasing effect elsewhere (−9.46 versus 19.90***).21 This is imaginable if one considers the enclave-type, exclusive growth that many resource-rich SSA economies experienced before the early 2000s, where resource rent had been enjoyed by only the group in power (Ndulu, 2008; Thorbecke and Ouyang, 2016). The above result, however, should not be taken as evidence that inequality plays a minor role in poverty reduction in SSA. Our country-level elasticity estimates suggest large variability among SSA countries.22 In his recent study using also the Povcalnet 1 data but different regression procedure, Fosu (2015) also found considerable country-level heterogeneity with respect to the relative contribution of income growth and inequality improvement in SSA. Next, we perform the same analysis to the enriched and expanded Povcalnet 2 data set and its SSA subsample and report the estimates of b1–b9 in Table 3. Again, we report estimates for the full sample, its SSA subsample, and its non-SSA subsample; but we focus on explaining the comparison between the full sample and the SSA subsample, as its implications are similar to those one would obtain from comparing the SSA and the non-SSA subsamples. Table 3: SSA Versus Global Structural Coefficients, Povcalnet 2 Data Set, 1986–2012   Global  SSA  Non-SSA  1986–2012  1986–2012  1986–2012  (1)  (2)  (3)  Growth of log income (b2)  −9.91**  −9.96  −10.42***  (−2.41)  (−1.42)  (−4.49)  Growth of log income × log initial Gini (b3)  1.96*  2.40  1.83**  (1.82)  (1.37)  (2.98)  Growth of log income × log (poverty line/initial income) (b4)  −0.21  0.91**  −0.79*  (−0.58)  (2.09)  (−2.36)  Change in log Gini (b5)  24.31***  15.84  25.03***  (2.97)  (1.19)  (5.03)  Change in log Gini × log initial Gini (b6)  −5.23**  −4.21  −5.06  (−2.37)  (−1.20)  (−3.72)  Change in log Gini × log (poverty line/initial income) (b7)  −0.51  −4.62***  0.36  (−0.66)  (−3.19)  (0.54)  Log initial Gini (b8)  0.70***  0.02  0.62**  (3.46)  (0.19)  (2.75)  Log (poverty line/initial income) (b9)  0.01  −0.15*  0.01  (0.18)  (−1.44)  (0.17)  Intercept (b1)  −2.63***  −0.13  −2.30**  (−3.46)  (−0.30)  (−2.72)  N  702  103  599  R2  0.54  0.68  0.55    Global  SSA  Non-SSA  1986–2012  1986–2012  1986–2012  (1)  (2)  (3)  Growth of log income (b2)  −9.91**  −9.96  −10.42***  (−2.41)  (−1.42)  (−4.49)  Growth of log income × log initial Gini (b3)  1.96*  2.40  1.83**  (1.82)  (1.37)  (2.98)  Growth of log income × log (poverty line/initial income) (b4)  −0.21  0.91**  −0.79*  (−0.58)  (2.09)  (−2.36)  Change in log Gini (b5)  24.31***  15.84  25.03***  (2.97)  (1.19)  (5.03)  Change in log Gini × log initial Gini (b6)  −5.23**  −4.21  −5.06  (−2.37)  (−1.20)  (−3.72)  Change in log Gini × log (poverty line/initial income) (b7)  −0.51  −4.62***  0.36  (−0.66)  (−3.19)  (0.54)  Log initial Gini (b8)  0.70***  0.02  0.62**  (3.46)  (0.19)  (2.75)  Log (poverty line/initial income) (b9)  0.01  −0.15*  0.01  (0.18)  (−1.44)  (0.17)  Intercept (b1)  −2.63***  −0.13  −2.30**  (−3.46)  (−0.30)  (−2.72)  N  702  103  599  R2  0.54  0.68  0.55  Notes: This table gives b1ˆ–b9ˆ in Bourguignon's Identity Model using the Povcalnet 2 data set and the FE regression procedure. T-ratios are in parenthesis and based on White standard errors (corrected for heteroscedasticity). Log (poverty line/initial income) measures initial poverty condition, where poverty line is $1.25/day. T-ratios are compared against critical values adjusted for sample sizes. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. Based on the above estimates, we calculated the income and growth elasticities of poverty reduction during 1986–2012 for the entire developing world and SSA, respectively; and report these elasticities in Table 4. Table 4: Growth and Inequality Elasticity of Poverty Reduction for Global and SSA Regions, Povcalnet 2 Data Set, 1986–2012 Region  Entire developing world  SSA  Non-SSA  (1986–2012)  (1986–2012)  (1986–2012)  Ey  −2.32  −1.20  −2.34  Egini  5.64  1.62  5.27  Estimates from  702 observations from 102 countries  103 observations from 35 countries  599 observations from 67 countries  Region  Entire developing world  SSA  Non-SSA  (1986–2012)  (1986–2012)  (1986–2012)  Ey  −2.32  −1.20  −2.34  Egini  5.64  1.62  5.27  Estimates from  702 observations from 102 countries  103 observations from 35 countries  599 observations from 67 countries  Results presented in Tables 3 and 4 suggest that during 1986–2012, the impact of income growth on poverty reduction was essentially the same in SSA as in the entire developing world (−9.96** versus −9.91**). At given income growth rate (y), high initial inequality also weakens Ey at about the same strength in SSA as in the entire developing world (2.4 versus 1.96*). But because high initial poverty (high Z/Y or low initial income) weakens Ey significantly more in SSA than in the entire developing world (0.91** versus −0.21), Ey in SSA remains about half of that in the developing world during 1986–2012 (−1.12 versus −2.13), as it did during 1982–2007 (−0.56 versus −2.91 as reported in Table 2). Note that during 1982–2007, the main reason that SSA has a smaller Ey than the rest of the developing world is that poverty reduction is less responsive to income growth in SSA; in contrast, the main reason for SSA to still have a smaller Ey during 1986–2012 is that initial poverty weakens Ey more in SSA, while poverty reduction has become equally responsive to income growth in SSA as it is in the entire developing world. This suggests that directly addressing high initial poverty in SSA has become more important for poverty reduction in the region. The inequality elasticity of poverty reduction ( Egini) in the SSA region, in contrast, has significantly increased compared with the earlier period (Povcalnet 1) in the sense that it was close to a third of its developing world counterpart during 1986–2012 (1.62 versus 5.64), while it was less than a fifth of its developing world counterpart during the earlier 1982–2007 period (0.80 versus 4.48). The significant increase in SSA's Egini suggests that poverty reduction in SSA as a whole has become more responsive to improvement in inequality than before. This is what one would expect as growth has become more inclusive in the SSA region; as inclusiveness means that any socio-economic improvement, including a reduction in inequality, would benefit more poor people and thus have a greater impact on poverty reduction.23 We also notice that at given inequality change rate (g), high initial poverty now weakens Egini significantly more in SSA than in the entire developing world (−4.62*** versus −0.51), while it has little impact on Egini during 1982–2007 (−0.60 versus −2.46*** as reported in Table 1)—a finding that again suggests the increasing importance of directly attacking high initial poverty in SSA. To summarise, our above analysis of the G–I–P nexus in SSA provides some empirical evidence in support of what we conjectured in the beginning of this paper based on existing (mostly) qualitative evidence. First, our empirical results suggest that during the three decades between the early 1980s and the early 2010s, the growth pattern in the SSA region is indeed different from that in the entire developing world in the sense that poverty reduction in SSA is less responsive to income growth and improving inequality. Second, the growth pattern within SSA as a whole appears to have changed over time. More specifically, SSA seems to be slowly catching up as evidenced by the estimates derived from the Povcalnet 2 data set that includes the more recent observations from 2008 to 2012.24 We find that (i) the poverty-reducing effect of income growth in SSA has increased to become slightly greater than that in the entire developing world and (ii) poverty reduction in SSA has become more responsive to improving inequality. Associated with these improvements, however, are the increasingly stronger Ey- and Egini-weakening effect of high initial inequality and high initial poverty, suggesting that addressing high initial inequality and low initial income has become more urgent than it was before for the poor people in SSA countries to benefit more from income growth and improving inequality. If we also consider the social benefits of reducing persistent and high inequality and poverty, the importance of directly attacking initial poverty and inequality is further reinforced. 5. Estimating the African structural coefficients of the P–I–G nexus 5.1 Econometric model Compared with its long recognition and extensive exploration of the impact of growth and inequality on poverty (the G–I–P nexus), the development community—as discussed in Section 3—had paid less attention to the reverse influence from poverty and inequality to subsequent growth (the P–I–G Nexus) until the most recent decade. Specifically, there are two strands of literature on the P–I–G nexus: one focusing on the direct impact of inequality on subsequent income growth; the other focusing on the direct impact of initial poverty on income growth. The former strand of literature has so far reached no unanimous conclusions: Alesina and Rodrik (1994), Perotti (1996) and Ravallion (2012) found a negative relationship between inequality and growth on the basis of cross-section data.25Li and Zou (1998) and Forbes (2000), in contrast, found a positive relationship between the two using aggregate panel data. Barro (2000) found the impact of inequality on growth dependent on the country's level of income. Banerjee and Duflo (2003), in turn, argued that there is an inverted U-shape relationship between income growth and the changes in inequality. Our emphasis and focus here is on the strand of literature, which explores the direct impact of initial poverty on subsequent income growth. While the theoretical literature on links between income or consumption growth and poverty-related variables (such as under-investment in human and/or physical capital, risk aversion and poor institutions) is relatively extensive,26 and several empirical studies attempted to uncover these indirect poverty–growth links;27 there have been very few direct empirical assessments of the impact of initial poverty on subsequent growth. One of the earliest empirical attempts to explore the direct impact of initial poverty on subsequent income growth was made by Lopez and Servén (2009). Using country-level panel data from some hundred developed and developing countries during 1960–2000,28Lopez and Servén (2009) found that a ten percentage-point increase in the headcount poverty rate reduces annual per capita income growth by about one percentage point. Following a similar empirical strategy29 but using cross-section data from ninety developing countries during 1978–2007, Ravallion (2012) found a lack of poverty convergence in the entire developing world in spite of evidence for mean (income and also consumption) convergence and the advantage of growth.30 He then demonstrated that the lack of poverty convergence is because initial poverty directly impedes subsequent income growth, besides weakening the growth elasticity of poverty reduction which has been extensively documented in the G–I–P literature.31 Given our conjecture that the growth pattern in SSA is different from that of the entire developing world, it is essential to analyse how initial poverty affects income growth and poverty reduction in Africa. A first effort to explore the P–I–G nexus in the African continent was made by Shimeles et al.(2016), who applied Ravallion (2012)’s empirical models to his African subsample containing thirty-two African countries including four Northern African countries. Their analysis suggests that Africa enjoyed poverty convergence during 1978–2007 as initial poverty did not impede subsequent growth in Africa as it did in the entire developing world; which they suggest could be explained by the anti-poverty programmes some African governments put in place since the mid-1990s. Since our focus in this paper is SSA, we further restrict our analysis here to the SSA region; that is, we excluded the four North African countries in the Shimeles et al.(2016) analysis. Hence, the results we reported in this paper are based on the Ravallion (2012) data set32 covering ninety developing countries that were surveyed at least twice during 1978–2007 and its SSA subsample covering twenty-eight SSA countries surveyed during the same period. Differences in the analysis results based on the global sample and its SSA subsample should reflect the differences between the P–G–I relationship in SSA and that in the entire developing world. Incidentally, more recent data have become available since the Ravallion (2012) data were collected in December 2008; and the extended data set now covers 105 developing countries that were surveyed at least twice during 1978–2012. The SSA subsample of this extended data set contains thirty-two SSA countries surveyed during 1980–2011.33 We have tried to apply the extended data set to Ravallion (2012)’s empirical models and have obtained some preliminary results, but recently we noticed some updates of the Povcalnet data which may change the analysis results. We therefore decide not to report these results at this stage and in this paper.34 Two key questions and valid concerns have to be addressed before reporting and analysing the results we obtained from applying the Ravallion (2012) models to the SSA subsample: (i) is the SSA sample of countries and growth spells representative of the population? And (ii) is there sufficient variability (variance) within the SSA sample to capture the true relationship? Although it would be preferable to have access to an even larger SSA sample, we feel that the present sample is relatively representative of the geographical and economic diversity of the subcontinent as indicated by footnote 31 that gives the list of countries and the first and last years of the surveys used to derive the growth spells. In order to check on the variability of the SSA subsample compared to the full (whole developing world) sample, we derived the kernel distribution of the log initial headcount index (measured against the $2/day poverty line) for the full sample used by Ravallion (2012) and separately for the SSA subsample. As expected and shown in Figure 3, the former distribution has a much wider spread than the latter. Clearly, the small variability of the SSA sample means that our results should be qualified accordingly. With this qualification in mind, our analysis does provide evidence supporting our conjecture that the P–I–G relationship in SSA is very different from that in the developing world as a whole. Figure 3: View largeDownload slide Kernel Distribution of Log Initial Headcount Index by Region (1978–2007). Figure 3: View largeDownload slide Kernel Distribution of Log Initial Headcount Index by Region (1978–2007). We first explored whether convergence in mean income and poverty convergence exist in our two SSA subsamples using, respectively, the following empirical model from Ravallion (2012, Equations (1) and (3), p. 511)35:   gi(μit)=αi+βilnμit−1+εit (4) (Test for income convergence)   giHit=αi⁎+βi⁎lnHit−1+εit⁎ (5) (Test for poverty convergence). The dependent variable gi(xit)(≡(ln(xit)−ln(xit−τ))/τ) denotes the annualised change in log mean income ( μit) and log headcount index ( Hit) between year (t−τ) and year t in Equations (4) and (5), respectively; and the independent variable of interest is, respectively, log mean income and log headcount index in the beginning of the entire period of τ years. The coefficients of the initial values give the rates of unconditional convergence. For conditional convergence rates, we added—following Ravallion (2012)—a set of controls comprising per capita consumption from national accounts, primary school enrolment rate, life expectancy at birth and the relative price index of investment goods as a measure of policy distortion—all for earliest survey date and expressed in natural logarithm. 5.2 Empirical results As in Section 4, we performed analysis for the full sample, its SSA subsample and its non-SSA subsample and reported all these results. But we shall only explain the difference between SSA and the whole developing world, as the comparison between the SSA and non-SSA subsamples conveys essentially the same information. As given in Table 5, we find that both SSA and the entire developing world enjoyed strong mean income convergence during 1978–2007. Further, the former enjoyed stronger mean convergence than the latter, which is expected as the SSA region on average has a higher initial poverty level than the entire developing world. The estimates in Table 5 are robust to the inclusion of controls and also the inclusion of North African countries. Table 5: Convergence in Mean Income (1978–2007)   Unconditional world  Conditional world  Unconditional SSA  Conditional SSA  Unconditional non-SSA  Conditional non-SSA  Initial log mean income  −0.0174***  −0.0465***  −0.0276***  −0.0485***  −0.0207*  −0.0456***  (−3.2017)  (−10.6022)  (−3.4963)  (−5.3002)  (−2.5429)  (−9.1244)  N  97  88  28  28  69  60  R2  0.1351  0.4982  0.1709  0.4676  0.1294  0.5827    Unconditional world  Conditional world  Unconditional SSA  Conditional SSA  Unconditional non-SSA  Conditional non-SSA  Initial log mean income  −0.0174***  −0.0465***  −0.0276***  −0.0485***  −0.0207*  −0.0456***  (−3.2017)  (−10.6022)  (−3.4963)  (−5.3002)  (−2.5429)  (−9.1244)  N  97  88  28  28  69  60  R2  0.1351  0.4982  0.1709  0.4676  0.1294  0.5827  Notes: This table gives βˆ in the regression (4) in the text: gi(μit)=αi+βilnμit−τ+γiXit−τ+εit. T-ratios are in parenthesis and based on White standard errors (corrected for heteroscedasticity). T-ratios are compared against critical values adjusted for sample sizes. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. Estimates in Table 6, for its part, suggest that SSA experienced significant and sizable poverty convergence during 1978–2007; while the entire developing world experienced a lack of poverty convergence during this period as revealed in Ravallion (2012).36 As in Table 5, the findings in Table 6 are robust to the inclusion of control variables and the choice of poverty lines, though not to the inclusion of North African countries.37 Table 6: Convergence in Poverty (1978–2007)   Unconditional world  Conditional world  Unconditional SSA  Conditional SSA  Unconditional SSA  Conditional SSA  Initial log headcount index ($2 a day)  0.0059  −0.0099  −0.0254**  −0.0409***  0.0042  −0.0090  (0.5901)  (−0.6967)  (−2.3736)  (−4.4003)  (0.3632)  (−0.5478)  N  89  84  28  28  61  56  R2  0.0080  0.1202  0.1340  0.4321  0.0029  0.1755  Initial log headcount index ($1.25 a day)  −0.0052  −0.0248  −0.0251**  −0.0372***  −0.0099  −0.0244  (−0.452)  (−1.5746)  (−2.3861)  (−4.4219)  (−0.7013)  (−1.2516)  N  82  78  28  28  54  50  R2  0.0053  0.1451  0.1127  0.3638  0.0128  0.1957    Unconditional world  Conditional world  Unconditional SSA  Conditional SSA  Unconditional SSA  Conditional SSA  Initial log headcount index ($2 a day)  0.0059  −0.0099  −0.0254**  −0.0409***  0.0042  −0.0090  (0.5901)  (−0.6967)  (−2.3736)  (−4.4003)  (0.3632)  (−0.5478)  N  89  84  28  28  61  56  R2  0.0080  0.1202  0.1340  0.4321  0.0029  0.1755  Initial log headcount index ($1.25 a day)  −0.0052  −0.0248  −0.0251**  −0.0372***  −0.0099  −0.0244  (−0.452)  (−1.5746)  (−2.3861)  (−4.4219)  (−0.7013)  (−1.2516)  N  82  78  28  28  54  50  R2  0.0053  0.1451  0.1127  0.3638  0.0128  0.1957  Notes: This table gives βˆ in the regression (5) in the text: gi(Hit)=αi⁎+βi⁎lnHit−τ+γi⁎Xit−τ+εit⁎. T-ratios are in parenthesis and based on White standard errors (corrected for heteroscedasticity). T-ratios are compared against critical values adjusted for sample sizes. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. Hence, it can be inferred from the above evidence that the structure of growth in SSA is different from that of the entire developing world during the period of 1978–2007. To explain the existence of poverty convergence in SSA during 1978–2007, we recall that Ravallion (2012) demonstrated that the rate of poverty reduction is jointly determined by the income growth rate and the initial poverty level. Specifically, countries with lower initial income tend to experience faster growth in mean income and hence faster poverty reduction.38 This is referred to as the mean convergence effect on poverty convergence. The initial poverty level, for its part, affects the rate of poverty reduction through two distinct channels. On the one hand, it directly hinders subsequent income growth and hence lowers the rate of poverty convergence through low income growth; on the other hand, it indirectly impedes poverty reduction through weakening the growth elasticity of poverty reduction (or making poverty reduction less responsive to income growth). Ravallion (2012, Equation (11), p. 520) summarises the above into the following equation that decomposes the rate of poverty convergence into three components:   ∂g(Hit)∂lnHit−τ=ηβ(1−Hit−τ)(∂lnHit−τ∂lnμit−τ)−1+ηγ(1−Hit−τ)+[−ηg(μit)Hit−τ] (6)(Mean convergence effect) (Direct effect (Poverty elasticity of poverty) effect). Equation (6) is not an identity but instead derived from the two empirical models specified below:   giμit=α+βlnμit−τ+γlnHit−τ+εit (7)Ravallion (2012)), Equation (4) and also Equation (10.2))   g(Hit)=η(1−Hit−τ)g(μit)+vit (8)Ravallion (2012), Equation (10.1)). Parameters β and γ in Equation (6), therefore, are from Equation (7), where the annualised change in log mean income ( giμit≡lnμit−lnμit−τ)/τ) is regressed on the initial levels of income and poverty, or income and poverty in the beginning of the growth spell of τ years. Parameter η in Equation (6), for its part, is from Equation (8) and is the regression coefficient of the ‘initial-poverty-adjusted income growth rate’ on the change in poverty. As Ravallion (2012, Table 4, p. 528) convincingly demonstrated, ‘the key proximate determinant of the rate of poverty reduction is η, the “poverty-adjusted growth rate” ( (1−Hit−τ)g(μit); rather than the ordinary growth rate ( g(μit)’ (pp. 519–520). The expression ∂lnHit−τ∂lnμit−τ in Equation (6) is the regression coefficient of g(Hit on g(μit, or the standard growth elasticity of poverty reduction denoted as Ey in Section 4. Finally, Hit−τ and g(μit) are the sample means of initial poverty rate and mean income growth rate. Estimates of Equation (7) based on the Ravallion (2012) data and its SSA subsample are presented in Table 7. During 1982–2007, initial poverty significantly retards income growth in the entire developing world as a whole (−0.0201*** with t-ratio = −5.51) but has little impact on income growth in Africa (−0.0035 with t-ratio = −0.14). Again, these results are robust to the choice of poverty lines (1.25 or 2 international dollars per day) and the inclusion of a set of control variables including the initial Gini index.39 Table 7: Regressing Mean Income Growth on Initial Poverty and Income (1978–2007)   World  SSA  Non-SSA  Log initial income  −0.0395***  −0.0293*  −0.0440***  (−8.4095)  (−1.7913)  (−8.8327)  Log initial headcount index ($2 line)  −0.0201***  −0.0035  −0.0200***  (−5.5135)  (−0.1363)  (−5.7577)  N  90  28  62  R2  0.275  0.1711  0.3080  Log initial income  −0.0348***  −0.0249  −0.0375***  (−5.0507)  (−1.4517)  (−4.6644)  Log initial headcount index ($1.25 line)  −0.01*  0.0035  −0.0094*  (−2.4131)  (0.2306)  (−2.2337)  N  84  28  56  R2  0.2425  0.1714  0 .2563    World  SSA  Non-SSA  Log initial income  −0.0395***  −0.0293*  −0.0440***  (−8.4095)  (−1.7913)  (−8.8327)  Log initial headcount index ($2 line)  −0.0201***  −0.0035  −0.0200***  (−5.5135)  (−0.1363)  (−5.7577)  N  90  28  62  R2  0.275  0.1711  0.3080  Log initial income  −0.0348***  −0.0249  −0.0375***  (−5.0507)  (−1.4517)  (−4.6644)  Log initial headcount index ($1.25 line)  −0.01*  0.0035  −0.0094*  (−2.4131)  (0.2306)  (−2.2337)  N  84  28  56  R2  0.2425  0.1714  0 .2563  Notes: This table gives βˆ in the regression (7) in the text: giμit=α+βlnμit−τ+γlnHit−τ+εit. T-ratios are in parenthesis and based on White standard errors (corrected for heteroscedasticity). T-ratios are compared against critical values adjusted for sample sizes. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. Next we report estimates of Equation (8) in Table 8. The negative sign of η suggests that in both the entire developing world and the SSA region, higher initial poverty weakens the poverty-adjusted growth elasticity of poverty reduction during 1978–2007. Comparing the magnitude of the two η s, we see that initial poverty weakens the (poverty-adjusted) growth effectiveness of poverty reduction less in the SSA region than in the entire developing world (−2.2694*** versus −2.4676***). Table 8: Regressions of Poverty-Adjusted Income Growth Rate on the Change in Poverty (1978–2007)   World  SSA  Non-SSA  Poverty-adjusted income growth (1−Hit−τ)g(μit)($2 line)  −2.4676***  −2.2694***  −2.4653***  (−7.3670)  (−6.5728)  (−7.1981)  N  89  28  61  R2  0.6714  0.8257  0.6662    World  SSA  Non-SSA  Poverty-adjusted income growth (1−Hit−τ)g(μit)($2 line)  −2.4676***  −2.2694***  −2.4653***  (−7.3670)  (−6.5728)  (−7.1981)  N  89  28  61  R2  0.6714  0.8257  0.6662  Notes: This table gives βˆ in the regression (8) in the text: g(Hit)=η(1−Hit−τg(μit)+vit. T-ratios are in parenthesis and based on White standard errors (corrected for heteroscedasticity). T-ratios are compared against critical values adjusted for sample sizes. ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. Now we have estimated all parameters needed to compute the three components of poverty convergence rate specified in Equation (6). We present our computation results in Table 9. Numerical figures in this table seem to suggest that while the three components almost perfectly predicted the actual rate of poverty convergence in the entire developing world during 1978–2007 as suggested in Ravallion (2012), they could not fully account for the poverty convergence that SSA experienced during the period: the sum of the three effects predicts a lack of poverty convergence (−0.0035) in SSA, while a regression analysis of the actual data suggest a strong unconditional poverty convergence in the region (−0.0254 with t = −2.37 and R2 = 0.134). Table 9: Decomposition of the Rate of Poverty Convergence (1978–2007) Poverty convergence rate = (1) + (2) + (3)  Full sample (N = 90)  SSA subsample (N = 28)  Non-SSA subsample (N = 61)  Mean convergence effect (1)  −0.0476  −0.0312  −0.0583  Direct effect of poverty (2)  0.0263  −0.0017  0.0347  Poverty elasticity effect (3)  0.0161  0.0364  0.0061  Poverty convergence rate = (1) + (2) + (3)  −0.0052  −0.0035  −0.0174  Unconditional poverty convergence rate from regression (see Table 6 upper panel)  0.0059  −0.0254**  0.0042    (t = 0.59, R2 = 0.008, N = 89)  (t = 2.37, R2 = 0.134, N = 28)  (t = 0.3632, R2 = 0.0029, N = 61)  Poverty convergence rate = (1) + (2) + (3)  Full sample (N = 90)  SSA subsample (N = 28)  Non-SSA subsample (N = 61)  Mean convergence effect (1)  −0.0476  −0.0312  −0.0583  Direct effect of poverty (2)  0.0263  −0.0017  0.0347  Poverty elasticity effect (3)  0.0161  0.0364  0.0061  Poverty convergence rate = (1) + (2) + (3)  −0.0052  −0.0035  −0.0174  Unconditional poverty convergence rate from regression (see Table 6 upper panel)  0.0059  −0.0254**  0.0042    (t = 0.59, R2 = 0.008, N = 89)  (t = 2.37, R2 = 0.134, N = 28)  (t = 0.3632, R2 = 0.0029, N = 61)  Notes: This decomposition is based on Equation (6): ∂g(Hit)∂lnHit−τ=ηβ(1−Hit−τ)(∂lnHit−τ∂lnμit−τ)−1+ηγ(1−Hit−τ)+[−ηg(μit)Hit−τ]. The first term in this equation is the mean convergence effect, the second the direct effect of initial poverty and the third the poverty elasticity effect. Parameters β and γ are from Equation (7): giμit=α+βlnμit−τ+γlnHit−τ+εit. The expression ( ∂lnHit−τ∂lnμit−τ) in Equation (6) is the standard growth elasticity of poverty reduction. Parameter η is from Equation (8): gHit)=η(1−Hit−τgμit)+vit and is referred to as the ‘initial-poverty-adjusted growth elasticity of poverty reduction’. Hit−τ and g(μit) are the sample means of initial poverty rate and mean income growth rate. What, then, could have contributed to the poverty convergence in SSA during 1978–2007? Using data from forty-two Ethiopian administrative districts covering about 80,000 households during 1996–2011 and from thirty-three Rwandan districts covering over 27,000 households during 2000–2010, Shimeles et al.(2016) found strong poverty convergence and a positive association between initial poverty and subsequent growth in mean income. Since both countries have introduced anti-poverty policies favouring lagging regions during the periods examined and have registered impressive economic growth, Shimeles et al. (2016) suggested that deliberate policy interventions by African governments and public and private aid flows might explain why the region achieved poverty convergence. 6. Summary and conclusions The main objective of this paper is to understand better the anatomy of growth in SSA and compare it to that of the developing world as a whole (including or excluding SSA). Key to the analysis of the structure of growth is the interrelationship among G–I–P. These three concepts are jointly determined in any given socio-economic system (such as a country or a region) under the influence of the process of globalisation and the country-specific development strategy that act as prime movers. The forces of globalisation are largely exogenous (outside the control of State), while the development strategy is at least partially endogenous (under the control of the State). One can best imagine a G–I–P triangle with influence travelling in both directions. In analysing the transmission of influence, one can start at any corner of the G–I–P triangle. Traditionally, research in development economics focused on the impact of influence emanating from the pattern of growth on inequality and poverty which we refer to as the G–I–P nexus. This is the theme of Section 2, where the effects of different structures of growth, such as ‘pro-poor growth’ and ‘inclusive growth’, on inequality and poverty are reviewed. The main message of this section is that the pace of growth in SSA underwent a quantum jump around the turn of the century and the structure of growth appears to have become more inclusive. The reverse causal chain emanating from poverty to subsequent inequality and growth, or what we refer to as the P–I–G nexus, is discussed in Section 3. We review recent studies and argue that a strong case can be made—based on data from the entire developing world—that high initial poverty can dampen subsequent growth. We proceeded in Section 4 to test econometrically whether the G–I–P nexus (i.e., the impact of growth on inequality and poverty) was different in SSA than in the developing world and whether the structure of growth in SSA had actually undergone a structural break around the beginning of this millennium. Building on the earlier work of Fosu (2009, 2011, 2015), we used the Bourguignon (2003) identity model to estimate and derive structural coefficients of the G–I–P nexus for the developing world and, separately, for the SSA region. We relied on two data sets, covering two different periods: Povcalnet 1 (1982–2007) and Povcalnet 2 (1986–2012). The regression results provide some empirical support for our initial conjectures. First, the growth pattern in the SSA region, in the three decades between the early 1980s and the early 2010s, was indeed different from that of the whole developing world. The responsiveness of poverty reduction to income growth and to change in inequality was significantly less in the SSA region than in the rest of the developing world. Second, the growth pattern within the SSA region has changed over time as poverty reduction has become more responsive to income growth, improvement in inequality and the level of initial poverty in the post-2007 years. The evidence suggests that SSA is slowly catching up. Next, in Section 5, we tested the P–I–G nexus to determine whether the finding that high initial poverty retarded future growth and preempted poverty convergence obtained from a large global sample of ninety developing countries covering growth spells over an extended period (1978–2007) also applied to the SSA region. Again we found that SSA differs from the rest of the developing world in that initial poverty did not dampen subsequent growth in this region during 1978–2007 while it did in the rest of the developing world during the same time period. In fact, SSA countries with the highest initial poverty incidence appeared to grow subsequently faster—leading to poverty convergence. The same finding was further confirmed at the interregional level in Ethiopia and to a lesser degree in Rwanda by Shimeles et al.(2016). We hypothesise that the main cause of poverty convergence in SSA during 1978–2007, in contrast with the lack of convergence in the global sample, might be due to anti-poverty interventions by governments and foreign public and private aid inversely proportional to the depth of poverty. Our results and findings are subject to two key caveats: (i) whether the various samples of the African countries we used in our regressions are sufficiently representative of the underlying population; and (ii) whether the sample sizes and variances are sufficient to draw valid inferences. While we recognise that larger and more diversified samples would have been preferable, we believe that we can defend the validity of our results as discussed in the text. Finally, we can draw some policy implication from our study. The analysis of the G–I–P nexus implies that an inclusive growth strategy focused on productive employment creation can spread the benefits of growth more widely and reduce extreme inequality and poverty. In turn, the analysis of the P–I–G nexus suggests that a pro-growth poverty reduction strategy focused on alleviating poverty directly through such interventions as social protection schemes can contribute to an acceleration of the pace of growth and to a more inclusive pattern. By removing some major constraints on the behaviour of the poor, the latter may be better able to invest in their education and adopt riskier but more efficient technologies. By now, there are many examples of interventions that by reducing current poverty contribute to raising productivity in the future. Since it is often difficult to draw a clear distinction as to whether a specific measure comes under one or the other strategy, it is perhaps best to consider these two strategies as complementary and mutually re-enforcing. Acknowledgements The authors would like to thank Augustin Fosu and the anonymous referees for helpful comments on earlier versions of this paper. We would also like to thank the African Development Bank for funding this research. Footnotes 1 Easterly and Levine (1997) were the first of many to refer to an ‘African dummy’. 2 Thus, an extreme case where the average income of the poor increases by only 1% following a GDP growth spell of 8% would be considered pro-poor according to the absolute definition. 3 The assumption that per capita income follows log-normal distribution is empirically tested and confirmed by Lopez and Servén (2005). They also found that this assumption does not hold for per capita consumption. 4 We thank Fosu for kindly sharing with us his data. 5 Fosu's regression is based on 320 usable observations for the $1.25 poverty line from 89 developing countries during 1981–2007 (Fosu 2015, p. 51 and endnote 10). Our usable sample size is slightly larger as we used a different regression method due to the concern of weak identification in Fosu's GMM procedure. See more discussion about this in the text. 6 In fact, the Povcalnet 1’s 1986–2007 subsample has 359 usable observations, while the Povcalnet 2’s 1986–2007 subsample contains 529 usable observations. 7 These thirty-five SSA countries, with the number of years surveyed in parentheses, are as follows: Angola (1), Benin (1), Botswana (3), Burkina Faso (3), Burundi (2), Cape Verde (1), Cameroon (2), Central African Republic (2), Chad (1), Republic of Congo (1), Côte d'Ivoire (8), Ethiopia (3), Gambia (1), Ghana (4), Guinea (4), Guinea-Bissau (1), Kenya (3), Lesotho (4), Madagascar (5), Malawi (2), Mali (3), Mauritania (5), Mauritius (1), Mozambique (2), Niger (4), Nigeria (4), Rwanda (3), Senegal (4), Sierra Leone (1), South Africa (5), Swaziland (2), Tanzania (3), Togo (1), Uganda (7) and Zambia (6). 8 These twenty-eight SSA countries, with the number of years surveyed in parentheses, are as follows: Botswana (1), Burkina Faso (2), Burundi (2), Cameroon (1), Central African Republic (1), Côte d'Ivoire (6), Ethiopia (3), Gambia (1), Ghana (4), Guinea (2), Guinea-Bissau (2), Kenya (3), Lesotho (3), Madagascar (4), Malawi (1), Mali (2), Mauritania (3), Mozambique (1), Niger (2), Nigeria (3), Rwanda (1), Senegal (3), Sierra Leone (1), South Africa (2), Swaziland (1), Tanzania (1), Uganda (5) and Zambia (5). 9 When the sample size is small, a critical value larger than 1.96 would be needed for a 5% significance level. In Table 1 Column 2 where the sample size is only 66 while the number of controls is 8, the critical value is 2.0017 for the 5% significance level and 2.3924 for the 1% level. 10 Fosu's GMM regression would have used fifty-eight observations from twenty-five SSA countries during 1987–2005; while our FE regression used sixty-six observations from twenty-eight SSA countries during 1987–2006. 11 Inequality elasticity based on Fosu (2015)’s global and our SSA structural coefficients are, respectively, 1.47 and 0.8. Income elasticity based on Fosu (2015)’s global and our SSA structural coefficients are, respectively, −1.37 and −1.56. 12 The GMM estimates for the global sample are exactly the same as those reported in Fosu (2015), which should be as exactly the same usable sample and IVs are used. 13 GMM results are available upon request. 14 See Fosu (2011) for a detailed list of instrument variables. 15 The Cragg–Donald (1993)F statistic and their Stock–Yogo critical values for 5% significance level (reported in brackets) for GMM applied to the Povcalnet 1 sample, its SSA subsample, the Povcalnet 2 sample, its SSA subsample are 3.47 (21.36), 4.56 (16.85), 1.59 (21.36) and 3.01 (16.85), respectively. Since the F statistics are smaller than the corresponding critical values, we cannot reject the null that the model is weakly identified. 16 The under-identification test statistic for GMM applied to the Povcalnet 1 SSA subsample is 7.88 with a χ2(4) p-value of 0.0961. Since the p-value is very close to 0.1, we can only barely reject the null that says the model is under-identified at 10% significance level. 17 The Pagan–Hall test statistics (and p-values) for GMM applied to the Povcalnet 1 sample, its SSA subsample, the Povcalnet 2 sample, its SSA subsample are 45.593 (χ2(26) p-value 0.01), 21.303(χ2(11) p-value = 0.0304), 100.2 (χ2(26) p-value = 0.000) and 27.401 (χ2(11) p-value = 0.004), respectively. Since the p-values are smaller than 0.05, we reject the null of no heteroscedasticity. 18 The Hausman test statistics (and p-values) for GMM applied to the Povcalnet 1 sample, its SSA subsample, the Povcalnet 2 sample, its SSA subsample are 0.297 (0.586), 6.187 (0.013), 0.147 (0.702) and 0.102 (0.750), respectively. Except for the p-value for the Povcalnet 1 SSA subsample, the p-values are all larger than 0.05, suggesting that we cannot reject the null that the specified endogenous regressor (annualised growth of log mean income) can actually be treated as exogenous. 19 OLS estimates are similar and therefore not reported for parsimony. They are available upon request. 20 Using World Bank growth and poverty data from effectively 1982 to 2004 and random-effects regression procedures, Fosu (2009) found that growth elasticity is, respectively, −1.1 and −2.9 for the SSA and global sample; and that inequality elasticity is 1.1 and 5.4 for the SSA and global sample, respectively. 21 Also note that the initial levels of poverty and inequality both had little impact on Egini in SSA (as reflected by b6 and b7). 22 Country-level growth and income elasticity estimates are not reported for parsimony, but are available upon request. 23 To explain the increase from a technical perspective, we look at Equation (3) that reveals the three components of inequality elasticity ( Egini=b5+b6G′+b7Z/Y). Comparing results reported in Tables 1 and 3, we see that the change in b5 estimate has a positive net effect on Egini; while the changes in b6 and b7 estimates both have a negative net effect on Egini, given that initial poverty (G′) and initial poverty (Z/Y) from the two data sets are largely the same (which we have confirmed). Since our estimation returns a more positive estimate for Egini from Povcalnet 2, the positive effect dominated the negative effect. That is, the increase of SSA Egini from 0.62 during 1987–2006 to 1.62 during 1986–2012 is mainly driven by a change in b5, which suggests that rising inequality coexisted with poverty reduction in SSA before 2007; but has become an obstacle to further poverty reduction in the region during the post-2007 years. 24 Ideally, we would have liked to truncate the Povcalnet 2 data set into two sub-periods 1986–2000 and 2000–2012 to try to capture the structural break in the growth pattern. Unfortunately, the number of observations is too small for us to obtain significant results. 25 Ravallion (2012), however, finds the negative relationship significant ‘only at the 10% level and when [growth is calculated] using consumption data from the national survey’ (p. 514). 26 Azariadis and Stachurski (2005) provide a good survey of poverty traps, i.e., vicious cycles of poverty and stagnation. On how poverty may retard growth through suppressing investments in health, education and physical capital, see Dasgupta and Ray (1986), Galor and Zeira (1993) and Banerjee and Newman (1994), respectively. On risk aversion as a poverty-perpetuating mechanism, see Banerjee (2000) and Dercon (2003). Note that Stiglitz (1969) first argues that risk aversion leads to under-investment. On poor institutions as potential sources of poverty traps, see Engerman and Sokoloff (2006). 27 Quah (1993), Bloom et al. (2003) and Azariadis and Stachurski (2004) found empirical evidence that cross-country distribution of per capita income exhibits bimodality. At the micro level, Jalan and Ravallion (2002) found a significant link between household consumption growth and aggregate (at the local level) physical and human capital endowments in China; which they argue is consistent with the existence of geographic poverty traps. 28 Lopez and Servén (2009) constructed their own poverty data using a log-normal approximation that dates back to Gibrat (1931). This approximation allows them to use per capita income data which are more available to estimate the less available poverty data for a given poverty line. The assumption that per capita income follows log-normal distribution is empirically tested and confirmed in Lopez and Servén (2005). 29 Both papers regress income growth rate on initial poverty and initial per capita income/consumption. The main difference is that Ravallion (2012) exploited only the cross-section dimension of the data; while Lopez and Servén (2009) exploited both the time-series and the cross-section dimensions of the data. More specifically, in Ravallion (2012)’s main analysis sample of ninety countries, each country is surveyed only twice. He then generated for each country an annualised growth rate of log per capita income or consumption for the period between the two surveys, and regressed this growth rate on the log headcount index in the first round of survey. Clearly, his main analysis sample is a cross-sectional data set. In Lopez and Servén (2009)’s analysis sample of some hundred countries, in contrast, each country can be surveyed multiple times; for every two surveys, there is a growth rate of log per capita income. Their analysis sample is therefore an unbalanced panel data set. They regressed this growth rate on log poverty ratio of the earlier round of survey between the two. The advantage of Ravallion (2012)’s strategy—which Lopez and Servén (2009) also used and reported results in their Table 8—is that it eliminates any concern about reverse causality. The potential loss is that it does not control for cross-country heterogeneity. Ravallion (2012), however, argues that FE regression has attenuation bias that is too strong to reveal any true relationship given the nature of his data (p. 517). 30 Mean convergence refers to the situation that countries with lower initial income enjoy faster mean income or consumption growth; sometimes, it also refers to the advantage of backwardness. The advantage of growth refers to that a higher mean income or consumption tends to come with a lower incidence of absolute poverty. Poverty convergence, where countries with higher initial poverty rate experience faster poverty reduction, is expected to follow as a consequence of mean convergence and the advantage of growth. 31 Ravallion (2012, p. 520) found that the direct effect of poverty is 0.026, the poverty elasticity effect is 0.016 and the sum of the two negative impact of initial poverty on poverty reduction almost exactly offset the mean convergence effect, or the advantage of backwardness that is −0.047. 32 We thank Ravallion (2012) for making his data set publicly available on the American Economic Review website. 33 The countries (first survey year, last survey year) in the 1978–2012 SSA subsample are as follows: Angola (2000, 2009), Botswana (1986, 1994), Burkino Faso (1994, 2009), Burundi (1992, 2006), Cote d'Ivoire (1985, 2008), Cameroon (1996, 2007), Central African Republic (1993, 2008), Ethiopia (1982, 2011), Gamiba (1998, 2003), Ghana (1988, 2006), Guinea (1991, 2007), Guinea-Bissau (1991, 2002), Kenya (1992, 2006), Lesotho (1987, 2003), Madagascar (1980, 2010), Malawi (1998, 2010), Mali (1994, 2010), Mauritania (1987, 2008), Mozambique (1997, 2008), Namibia (1993, 2004), Niger (1992, 2008), Nigeria (1986, 2011), Rwanda (1985, 2011), Senegal (1991, 2011), Seychelles (1995, 2007), Sierra Leone (1990, 2011), South Africa (1993, 2009), Swaziland (1995, 2010), Tanzania (1992, 2007), Togo (2006, 2011), Uganda (1989, 2010) and Zambia (1991, 2010). The 1978–2012 data set also contains five North African countries: Algeria (1988, 1995), Egypt (1991, 2009), Morocco (1985, 2007), Tunisia (1985, 2010) and Djibouti which however is only surveyed once (2000). 34 It is important to note that the P–I–G analysis and the earlier G–I–P analysis are not based on exactly the same data. The former uses only information collected from the first and the last survey in a certain country to explore the impact of initial poverty on subsequent growth. The latter, in contrast, used information collected from all surveys in a certain country to elicit the impact of growth and change in inequality on the pace of poverty reduction. Therefore, our decision to not to report the P–I–G results for the longer period in this paper does not suggest that our G–I–P results for the longer period should be subject to further validation. 35 Note that while Ravallion (2012) denotes the dependent variable in his empirical model for mean convergence (Equation (1), p. 511) as ∆lnμit≡ln(μit)−ln(μit−1), he did this for simplicity by assuming evenly spaced data. He actually regressed the annualised change in log mean income (gi(μit)≡(lnμit−lnμit−τ)/τ). Similarly, he regressed the annualised change in log poverty (gi(Hit)≡(lnHit−lnHit−τ)/τ) instead of ∆lnHit specified in his empirical model for poverty convergence (Equation (3), p. 511). 36 Ravallion (2012, Figure 1, p. 505) suggests a lack of unconditional poverty convergence for poverty measured against the $2/day line. He mentions that poverty convergence is still absent when he adds controls and/or uses different poverty lines (p. 512). 37 When we add North African countries to the 1978–2007 SSA subsample, we found only conditional poverty convergence in Africa during 1978–2007 for poverty measured against the $2 per day poverty line (−0.0595 with t = −2.4199 and R2 = 0.5497). 38 The strong link between income growth and poverty reduction can be quickly confirmed. For the SSA subsample, the regression coefficient of poverty reduction rate (annualised difference in log headcount index at $2 line) on income growth rate (annualised difference in log mean income) is −0.465 with t = −6.13 and R2 = 0.660. For the full sample, Ravallion (2012, p. 512) reports that the regression coefficient of annualised difference in log headcount index at $2 line on annualised difference in log mean income is −1.085 with a t-ratio of −4.128 and R2 of 0.377. 39 It might be conjectured that initial poverty is picking up some other aspects of initial distribution, especially inequality which some literature has identified as having a direct negative impact on income growth. To investigate this, we ran Equation (7) with a set of controls including log of initial Gini index, initial per capita consumption from national account, primary school enrolment rate, life expectancy at birth and the relative price index of investment goods. It turns out the findings presented in Tables 9 and 10 still hold. Regression results are available upon request. References African Development Bank ( 2012) African Economic Outlook 2012 . Tunis, Tunisia: Abidjan, Côte d'Ivoire. PubMed PubMed  Alesina A., Rodrik D. ( 1994) ‘ Distributive Policies and Economic Growth’, Quarterly Journal of Economics , 109 ( 2): 465– 90. Google Scholar CrossRef Search ADS   Azariadis C., Stachurski J. ( 2004) A Forward Projection of the Cross-country Income Distribution . Belgium: Mimeo, Universite Catholique of Louvain. Azariadis C., Stachurski J. ( 2005) ‘Poverty traps’, in P. Aghion, S. Durlauf (eds) Handbook of Economic Growth , Vol. 1. Amsterdam, Netherlands: Elsevier (North Holland Publishing Co.), pp. 295– 384. Banergee A. ( 2000) ‘ The Two Poverties’, Working Paper 01-16, Department of Economics, M.I.T. Boston, Massachusetts, USA. Banerjee A. V., Duflo E. ( 2003) ‘ Inequality and Growth: What Can the Data Say’, Journal of Economic Growth , 8 ( 3): 267– 99. Google Scholar CrossRef Search ADS   Banerjee A. V., Newman A. F. ( 1994) ‘ Poverty, Incentives, and Development’, American Economic Review , 84 ( 2): 211– 5. Barro R. J. ( 2000) ‘ Inequality and Growth in a Panel of Countries’, Journal of Economic Growth , 5 ( 1): 5– 32. Google Scholar CrossRef Search ADS   Baum C. F., Schaffer M. E., Stillman S. ( 2003) ‘ Instrumental Variables and GMM: Estimation and Testing’, Stata Journal , 3 ( 1): 1– 31. Bloom E., Canning D., Sevilla J. ( 2003) ‘ Geography and Poverty Traps’, Journal of Economic Growth , 8 ( 4): 355– 378. Google Scholar CrossRef Search ADS   Bourguignon F. ( 2003) ‘The growth elasticity of poverty reduction: explaining heterogeneity across countries and time periods’, in Inequality and Growth: Theory and Policy Implications . Washington, DC: World Bank, pp. 3– 26. Cragg J. G., Donald S. G. ( 1993) ‘ Testing Identifiability and Specification in Instrumental Variable Models’, Econometric Theory , 9 ( 02): 222– 40. Google Scholar CrossRef Search ADS   Dasgupta P., Ray D. ( 1986) ‘ Inequality as a Determinant of Malnutrition and Unemployment: Theory’, The Economic Journal , 96 ( 384): 1011– 34. Google Scholar CrossRef Search ADS   Dercon S. ( 2003) Risk and Poverty: A Selective Review of the Issues . Oxford, UK: Mimeo, Oxford University. Easterly W., Levine R. ( 1997) ‘ Africa's Growth Tragedy: Policies and Ethnic Divisions’, The Quarterly Journal of Economics , 112 ( 4): 1203– 50. Google Scholar CrossRef Search ADS   Engerman S. L., Sokoloff K. L. ( 2006) ‘The persistence of poverty in the Americas: The role of institutions’, in S. Bowless, S. Durlauf, K. Hoff (eds) Poverty Traps . New Jersey: Princeton University Press, pp. 43– 78. Englebert P. ( 2000) ‘ Solving the Mystery of the AFRICA Dummy’, World Development , 28 ( 10): 1821– 35. Google Scholar CrossRef Search ADS   Fosu A. K. ( 2009) ‘ Inequality and the Impact of Growth on Poverty: Comparative Evidence for Sub-Saharan Africa’, Journal of Development Studies , 45 ( 5): 726– 45. Google Scholar CrossRef Search ADS   Fosu, A. K. ( 2011) Growth, Inequality, and Poverty Reduction in Developing Countries: Recent Global Evidence . UNU-WIDER Working Paper 2011-001. Helsinki, Finland: UNU-WIDER. Fosu A. K. ( 2015) ‘ Growth, Inequality and Poverty in Sub-Saharan Africa: Recent Progress in a Global Context’, Oxford Development Studies , 43 ( 1): 44– 59. Google Scholar CrossRef Search ADS   Forbes K. ( 2000) ‘ A Reassessment of the Relationship Between Inequality and Growth’, American Economic Review , 90 ( 4): 869– 87. Google Scholar CrossRef Search ADS   Galor O., Zeira J. ( 1993) ‘ Income Distribution and Macroeconomics’, Review of Economic Studies , 60 ( 1): 35– 52. Google Scholar CrossRef Search ADS   Gibrat R. ( 1931) Les Inégalités Économiques. Paris: Sirey. Hauk W. R. Jr, Wacziarg R. ( 2009) ‘ ‘A Monte Carlo Study of Growth Regressions’’, Journal of Economic Growth , 14 ( 2): 103– 47. Google Scholar CrossRef Search ADS   Jalan J., Ravallion M. ( 2002) ‘ Geographic Poverty Traps? A Micro Model of Consumption Growth in Rural China’, Journal of Applied Econometrics , 17 ( 4): 329– 46. Google Scholar CrossRef Search ADS   Li H., Zou H. F. ( 1998) ‘ Income Inequality is Not Harmful for Growth: Theory and Evidence’, Review of Development Economics , 2 ( 3): 318– 34. Google Scholar CrossRef Search ADS   Lopez H., Servén L. ( 2005) ‘ A normal relationship? Poverty, growth, and inequality’, World Bank Policy Research Working Paper 3814. Washington DC, USA. Lopez, H., Servén, L. ( 2009). Too Poor to Grow. World Bank Policy Research Working Paper Series 5012, http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.493.8341&rep=rep1&type=pdf. Ndulu B. J. (ed.) ( 2008) The Political Economy of Economic Growth in Africa, 1960–2000 , Vol. 2. Cambridge, UK: Cambridge University Press. Perotti R. ( 1996) ‘ Growth, Income Distribution, and Democracy: What the Data Say’, Journal of Economic growth , 1 ( 2): 149– 87. Google Scholar CrossRef Search ADS   Perry G., Arias O., López H., Maloney W., Servén L. ( 2006) Poverty Reduction and Growth: Virtuous and Vicious Circles . World Bank Publications. Google Scholar CrossRef Search ADS   Quah D. ( 1993) ‘ Empirical Cross-section Dynamics in Economic Growth’, European Economic Review , 37 ( 2): 426– 34. Google Scholar CrossRef Search ADS   Ravallion M. ( 2012) ‘ Why Don't We See Poverty Convergence?’, American Economic Review , 102 ( 1): 504– 23. Google Scholar CrossRef Search ADS   Shimeles, A., E. Thorbecke, and Y. Ouyang ( 2016) Why Don't We See Poverty Convergence in Sub-Saharan Africa?  Paper submitted for possible publication. Stiglitz J. ( 1969) ‘ The Effects of Income, Wealth and Capital Gains Taxation on Risk-Taking’, The Quarterly Journal of Economics , 83 ( 2): 263– 83. Google Scholar CrossRef Search ADS   Stock J. H., Wright J. H., Yogo M. ( 2002) ‘ A Survey of Weak Instruments and Weak Identification in Generalized Method of Moments’, Journal of Business & Economic Statistics , 20 ( 4): 518– 29. Google Scholar CrossRef Search ADS   Stock J. H., Yogo M. ( 2005) ‘Testing for weak instruments in linear IV regression’, in D. W. K. Andrews, J. H. Stocks (eds) Identification and Inference for Econometric Models: Essays in Honor of Thomas Rothenberg . Cambridge, UK: Cambridge University Press. Google Scholar CrossRef Search ADS   Thorbecke E. ( 2014) ‘How inclusive is the present growth pattern in sub-Saharan Africa?’, in E. Zedillo, O. Cattaneo, H. Wheeler (eds), Africa at a Fork in the Road. Taking Off or Disappointment Once Again?  Connecticut, USA: Yale Center for the Study of Globalization. Thorbecke E. ( 2015) ‘The anatomy of growth and development in sub-Saharan Africa’, in A. McKay, E. Thorbecke (eds) Economic Growth and Poverty Reduction in Sub-Saharan Africa: Current and Emerging Issues . Oxford, UK: Oxford University Press. Google Scholar CrossRef Search ADS   Thorbecke E., Ouyang Y. ( 2016) ‘Is sub-Saharan Africa finally catching up?’, in M. Andersson, T. Axelsson (eds) Diverse Development Paths and Structural Transformation in the Escape from Poverty . Oxford, UK: Oxford University Press, pp. 236– 65. Google Scholar CrossRef Search ADS   UNDP ( 2014) Human Development Report 2014 . Washington, DC: United Nations Development Program. © The Author 2017. Published by Oxford University Press on behalf of the Centre for the Study of African Economies. All rights reserved. For permissions, please email: journals.permissions@oup.com

Journal

Journal of African EconomiesOxford University Press

Published: Jan 1, 2018

There are no references for this article.