When Is Growth Pro-Poor? Cross-Country Evidence

Transcription

1 WP/04/47 When Is Growh Pro-Poor? Cross-Counry Evidence Aar Kraay

2 2004 Inernaional Moneary Fund WP/04/47 IMF Working Paper Research Deparmen When Is Growh Pro-Poor? Cross-Counry Evidence Prepared by Aar Kraay 1 Auhorized for disribuion by Ashoka Mody March 2004 Absrac This Working Paper should no be repored as represening he views of he IMF. The views expressed in his Working Paper are hose of he auhor(s) and do no necessarily represen hose of he IMF or IMF policy. Working Papers describe research in progress by he auhor(s) and are published o elici commens and o furher debae. Growh is pro-poor if he povery measure of ineres falls. This implies hree poenial sources of pro-poor growh: (a) a high rae of growh of average incomes; (b) a high sensiiviy of povery o growh in average incomes; and (c) a povery-reducing paern of growh in relaive incomes. I empirically decompose changes in povery in a large sample of developing counries ino hese componens. In he medium run, mos of he variaion in changes in povery is due o growh, suggesing ha policies and insiuions ha promoe broad-based growh should be cenral o pro-poor growh. Mos of he remainder is due o povery-reducing paerns of growh in relaive incomes, raher han differences in he sensiiviy of povery o growh in average incomes. Cross-counry evidence provides lile guidance on policies and insiuions ha promoe hese oher sources of pro-poor growh. JEL Classificaion Numbers: I32, O40 Keywords: povery, growh Auhor s Address: akraay@worldbank.org 1 The World Bank, Washingon, DC., and was on secondmen o he IMF s Research Deparmen while pars of his paper were wrien. This paper was prepared in he conex of he pro-poor growh program sponsored by he World Bank s Povery Research and Economic Managemen (PREM) group. The auhor is graeful o Robera Gai, Francisco Ferreira, and Marin Ravallion for helpful discussions, and o Shaohua Chen for providing daa.

3 - 2 - Conens Page I. Inroducion... 3 II. Empirical Framework... 5 III. Daa IV. Resuls V. Relaive Imporance of Sources of Pro-Poor Growh VI. Wha Drives he Sources of Pro-Poor Growh? VII. Conclusions References Tables 1. Correlaions of Povery Measures and Survey Mean Income or Consumpion Decomposing Changes in Povery: All Spells Decomposing Changes in Povery: Long Spells Correlaes of Pro-Poor Growh: All Spells Correlaes of Pro-Poor Growh: Long Spells Mulivariae Growh and Disribuional Change Regressions Figures 1. Relaive Growh Incidence Curves Sensiiviy of Povery o Growh in Percenile p Growh and Povery Reducion: Long Spells Decomposing Changes in he Headcoun Policies and Growh and Disribuion Componens of Changes in Policy... 32

4 - 3 - I. INTRODUCTION In he pas few years, he erm pro-poor growh has become pervasive in discussions of developmen policy. Despie he widespread use of he erm, here appears o be much less consensus as o wha exacly pro-poor growh means, le alone wha is deerminans are. According o one view, growh is pro-poor if he accompanying change in income disribuion by iself reduces povery (Kakwani and Pernia (2000)). However, his definiion is raher resricive, since i implies ha, for example, China s very rapid growh and dramaic povery reducion during he 1980s and 1990s was no pro-poor because he poor gained relaively less han he nonpoor. A broader and more inuiive definiion is ha growh is pro-poor if he povery measure of ineres falls. Ravallion and Chen (2003) propose his definiion and apply i o a paricular povery measure, he Was index. In his paper, I adop he broader definiion and hen apply sandard povery-decomposiion echniques o idenify hree poenial sources of pro-poor growh: (a) a high rae of growh of average incomes; (b) a high sensiiviy of povery o growh in average incomes; and (c) a povery-reducing paern of growh in relaive incomes. I implemen his decomposiion for several povery measures, using household-survey daa for a large sample of counries in he 1980s and he 1990s. I hen use variance decomposiions o summarize he relaive imporance of hese differen sources of pro-poor growh. Finally, I invesigae he correlaes of he sources of pro-poor growh in a large panel of observaions on changes in povery. The main resuls of his paper are he following. Firs, regarding he relaive imporance of he hree poenial sources of pro-poor growh, I find ha roughly half of he variaion in shor-run changes in povery can be explained by growh in average incomes. In he mediumo-long run, beween 66 and 90 percen of he variaion in changes in povery can be accouned for by growh in average incomes. Virually all of he remainder is due o changes in relaive incomes. In conras, cross-counry differences in he sensiiviy of povery o growh in average incomes accoun for very lile of he variaion in changes in povery. Second, I find some evidence ha growh in average household-survey incomes is correlaed wih several of he usual deerminans of growh from he empirical growh lieraure, including insiuional qualiy, openness o inernaional rade, and size of governmen. Alhough he evidence documened here for he correlaes of growh in household-survey incomes is no especially compelling, I argue ha his likely reflecs he limied counry coverage and presence of measuremen error in he household-survey daa on which his paper is based. Third, I find relaively lile evidence ha povery-reducing paerns of growh in relaive incomes are significanly correlaed wih a se of explanaory variables ha he empirical growh lieraure has idenified as significan deerminans of growh in per capia GDP. The same is rue for a number of oher variables ha, while no generally significan for growh, have been suggesed in he lieraure as poenially reducing inequaliy.

5 - 4 - Taken ogeher, hese resuls underscore he imporance of growh in average incomes for povery reducion. This, in urn, suggess ha a policy package focusing on known deerminans of growh in average incomes, such as he proecion of propery righs, sable macroeconomic policies, and openness o inernaional rade should be a he hear of propoor growh sraegies. Moreover, he absence of compelling evidence ha hese facors are sysemaically correlaed wih he changes in income disribuion ha maer mos for povery reducion suggess ha here are no obvious rade-offs policies ha lead o growh in average incomes are unlikely o sysemaically resul in adverse effecs on povery hrough heir effecs on relaive incomes. This does no mean ha growh in average incomes is sufficien for povery reducion. Raher, he resuls presened here sugges ha cross-counry evidence is unlikely o be very informaive abou he policies and insiuions ha are likely o lead o povery-reducing paerns of growh in relaive incomes. This suggess ha more micro-level and case-sudy research may be useful in shedding ligh on he deerminans of povery-reducing disribuional change. This paper is relaed o a growing empirical lieraure on growh, inequaliy, and povery. Mos immediaely, his paper builds on Dollar and Kraay (2002). In ha paper, he auhors defined he poor as hose in he boom quinile of he income disribuion and empirically invesigaed he deerminans of growh in incomes of he poores quinile. In a large panel of counries, we found ha growh in incomes of he poor racked growh in average incomes roughly one for one. Since he growh rae of average incomes of he poor is jus he sum of he growh rae of average incomes and he growh rae of he firs-quinile share, our paper showed ha neiher average incomes nor a large se of oher conrol variables were significanly correlaed wih changes in he firs-quinile share. Tha paper conribued o a growing lieraure on he deerminans of inequaliy, including Li, Squire, and Zhou (1998); Gallup, Radele, and Warner (1998); Spilimbergo, Londono, and Szekely (1999); Leamer and ohers (1999); Easerly (1999); Barro (2000); Foser and Szekely (2001); and Lundberg and Squire (2003). This paper differs from Dollar and Kraay (2002), as well as much of he exising lieraure on deerminans of inequaliy, in wo key respecs. Firs, insead of looking a relaive povery measures or inequaliy, I focus primarily on changes in absolue povery measures as he dependen variable. 2 As is well undersood, changes in absolue povery measures are complicaed and nonlinear funcions of underlying changes in average income and income disribuions. 3 The second conribuion of his paper is o empirically consruc he exac 2 A noable early excepion is Ravallion and Chen (1997), who esimae regressions of changes in absolue povery on changes in mean incomes using a panel of household surveys from developing counries. 3 See, for example, Bourguignon (1999) for a lognormal example.

6 - 5 - measures of disribuional change ha maer for changes in povery for a large sample of counries, raher han simply looking a common summary saisics of inequaliy such as he Gini coefficien or quinile shares. This means ha I can empirically sudy he conribuions of growh and disribuional change o changes in povery wihou having o make resricive assumpions abou he shape of he underlying income disribuion. 4 Despie hese differences, he main conclusions of his paper are similar o hose in Dollar and Kraay (2002). In paricular, boh papers find ha growh in average incomes maers a grea deal for reducions in boh relaive and absolue povery. Boh papers also find lile evidence ha common deerminans of growh, as well as a number of oher variables, are robusly correlaed wih paerns of disribuional change ha maer for povery reducion. The res of his paper proceeds as follows. Secion II reviews sandard poverydecomposiion echniques and uses hem o illusrae he channels hrough which growh and disribuional change maer for changes in a number of povery measures. Secion III describes he daase of changes in povery in a large sample of developing counries on which he empirical analysis is based. Secion IV presens he resuls. Secions V and VI provide evidence on he relaive imporance of he sources of pro-poor growh, as well as evidence on some of he correlaes of hese sources. Secion VII concludes. II. EMPIRICAL FRAMEWORK In his secion I use sandard echniques o decompose he change in povery ino hree componens: (a) growh in average incomes; (b) he sensiiviy of povery o changes in average incomes; and (c) changes in relaive incomes. Le y (p) denoe he income of he p h percenile of he income disribuion a ime. This can be wrien as a funcion of average dl income, µ, and he Lorenz curve, L (p), i.e. y = µ. Le P denoe he following dp generic addiive povery measure: 1 P = f ( y ) dp (1) 0 4 For example, Lopez (2003) invesigaes he deerminans of growh and change in he Gini coefficien, and hen draws conclusions regarding he likely effecs on povery by assuming ha he disribuion of income is lognormal, so ha here is a one-o-one mapping beween he Gini coefficien and he Lorenz curve.

7 - 6 - This noaion capures a number of differen povery measures. For example, if θ z y f ( y, θ ) = up o he headcoun, H y 1 ( ) z z = where z is he povery line, and zero aferwards, we have he Foser-Greer-Thorbecke class which includes he headcoun z (θ=0), he povery gap (θ=1), and he squared povery gap (θ=2). If ( ) f y = ln y up o he povery line and zero aferwards, we have he Was povery index. Anoher possibiliy is a broader social welfare funcion wihou a disconinuiy a he povery line, such as Akinson s (1970) equally disribued equivalen income (EDEI). In his case f θ ( y ) y θ = for all p, and 1 P is he povery measure of ineres. Nex, we can differeniae his povery measure wih respec o ime o ge: 5 1 dp = η g dp (2) d 0 Equaion (2) ells us ha he rae of change in he povery measure is he average across all perceniles of he income disribuion of he growh rae of each percenile muliplied by he sensiiviy of he povery measure o growh in ha percenile. In paricular, df ( y ) η y is he semi-elasiciy of he povery measure wih respec o he dy income of he p h percenile. This erm capures he effec on povery of a small change in incomes of individuals a he p h percenile of he income disribuion. This sensiiviy is dy 1 muliplied by g, which capures he growh rae of incomes a each d y percenile of he income disribuion. Ravallion and Chen (2003) refer o his funcion as he growh incidence curve. The overall change in povery hen consiss of he average across all perceniles of he produc of hese wo erms. In order o separae ou he effecs of growh in average incomes, we can re-wrie equaion (2) by adding and subracing average growh o ge: 5 Differeniaing under he inegral sign in equaion (1) requires he applicaion of Leibniz s rule. Noe ha he erm involving he derivaive of he upper limi of inegraion is zero, since he povery measures are zero when evaluaed a he incomes of hose a he povery line. For EDEI boh he upper and lower limis of inegraion are consan and so he derivaive simply passes hrough he inegral sign.

8 dp dµ 1 dµ 1 = η ( ) + ( ) ( ) p dp η p g p dp (3) d d µ 0 0 d µ Equaion (3) idenifies he hree sources of pro-poor growh ha we have been discussing: (a) growh in average incomes; (b) he sensiiviy of povery o growh in average incomes; and (c) growh in relaive incomes. The firs erm in equaion (3) capures he firs wo dµ 1 sources of pro-poor growh. I consiss of growh in average incomes,, muliplied d µ by a erm summarizing he sensiiviy of he povery measure o changes in average incomes, 1 0 η dp. This sensiiviy is simply he average across all perceniles of he sensiiviy of povery o growh in each percenile of he income disribuion. The second erm in Equaion (3) capures he remaining source of pro-poor growh: changes in relaive incomes. In paricular, his hird source of pro-poor growh is he average across all perceniles of he income disribuion of he produc of (a) he growh rae of income in he p h percenile relaive o average income growh, and (b) he sensiiviy of povery o growh in ha percenile. For example, if he povery measure of ineres is very sensiive o growh among he poores, and if he income of he poores grows faser han average incomes, hen povery will fall. Equaion (3) is useful for hinking abou he various definiions and sources of pro-poor growh. For example, he Kakwani and Pernia (2000) definiion of pro-poor growh saes ha growh is pro-poor if and only if he second erm in equaion (3) is negaive, i.e., he paern of growh in relaive incomes is such ha he povery measure falls. A broader definiion of pro-poor growh suggesed by Ravallion and Chen (2003) is ha growh is propoor if he povery measure of ineres falls. According o his definiion, here are hree poenial sources of pro-poor growh: (a) rapid growh in average incomes; (b) a high sensiiviy of povery o growh in average incomes; and (c) a povery-reducing paern of growh in relaive incomes. In he empirical secion of his paper, I will use daa on income disribuions and average incomes for a large sample of developing counries o consruc hese hree sources of propoor growh, documen heir relaive imporance, and invesigae heir deerminans. Before doing so, however, i is useful o examine he key ingrediens in equaion (3) in more deail: dµ 1 he paern of growh in relaive incomes, g, and he funcion summarizing d µ he sensiiviy of povery o growh in each percenile, η (p). Figure 1 graphs wo examples of he paern of growh in relaive incomes, for China over he period , and for Indonesia over he period In China, according o he household survey average incomes grew a 14 percen per year, and he dollar-a-day headcoun measure of povery fell from 51 percen o 33 percen of he populaion.

9 - 8 - However, here was also a sharp increase in inequaliy during his period, wih he Gini coefficien rising from 34 o 40. The paern of relaive income growh raes shown in he relaive growh incidence curve highlighs his paern of increased inequaliy. Growh in he poores 80 perceniles of he populaion was below average growh, while he riches 20 percen of he populaion saw above-average growh. In Indonesia, survey mean income fell dramaically beween 1996 and 1999 a nearly 9 percen per year as a resul of he Eas Asian financial crisis. Ye during his period, he paern of growh in relaive incomes was povery-reducing. Inequaliy as measured by he Gini coefficien fell from 36.5 o The relaive growh incidence curve is downward sloping, indicaing ha incomes of he richer perceniles of he income disribuion fell faser han incomes of poorer perceniles. In fac, below-average growh was recorded only for he riches 20 percen of he populaion. Despie his pro-poor paern of relaive income growh, he headcoun measure of povery increased from 8 percen o 13 percen of he populaion, driven by he large negaive growh effec. Consider nex he sensiiviy of he povery measure o growh in differen perceniles of he income disribuion. In he case of he Foser-Greer-Thorbecke class, θ 1 η = θ ( y / z) ( 1 y / z) up o he headcoun, and zero aferwards. For he Was index, η = 1 up o he headcoun, and zero aferwards. Finally, for EDEI, we θ have η = θ y. Noe ha hese sensiiviies depend no only on he povery measure of ineres, bu also on he enire disribuion of income as summarized by y (p). Figure 2 graphs hese sensiiviies, using he acual disribuion of income in China in 1990 as an example, o show how differen povery measures are sensiive o growh in differen perceniles of he income disribuion. In he case of he headcoun, his sensiiviy is zero everywhere excep jus below he povery line where i spikes down o minus infiniy. This is because he headcoun simply adds up he number of people below he povery line small increases in income of inframarginal poor people ha do no bring hem above he povery line will no reduce he headcoun. The same is rue for increases in incomes of hose above he povery line, including he near-poor jus above he povery line. The case of he headcoun already illusraes he broader poin of Figure 2: wheher a given paern of growh is pro-poor or no depends crucially on he povery measure of ineres. In paricular, if pro-poor growh in he sense of reducing he headcoun measure of povery is he objecive, hen a pro-poor growh sraegy should focus exclusively on raising he incomes of hose jus a he povery line, and should ignore everyone else. This srong and slighly absurd conclusion is in par driven by he choice of he headcoun as he povery measure of ineres. Consider nex he povery gap and he squared povery gap. The povery gap is mos sensiive o growh in incomes of hose a he povery line, bu is also sensiive o growh in incomes of everyone below he povery line. The inuiion for his is he following: he povery gap reflecs a social welfare funcion which is indifferen o he disribuion of income among poor people. In his case a given rae of average growh resuls in a larger absolue increase in income for a person near he povery line, and so he povery

10 - 9 - measure is mos sensiive o hose neares he povery line, bu is non-zero for all poor people. The squared povery gap is also sensiive o growh in he incomes of all hose below he povery line, bu he sensiiviy is now U-shaped. Growh in incomes of he riches and poores of hose below he povery line maers leas, and he squared povery gap is mos sensiive o growh in incomes of poor people somewhere in beween hese wo exremes. The inuiion for his again depends on he underlying social welfare funcion, which now values absolue ransfers from richer o poorer poor people. This however is offse by he fac ha a given average growh rae resuls in a larger absolue increase in income for richer poor people. This is why he sensiiviy of he povery measure o growh is a non-monoonic funcion of he income percenile. The Was index has he propery ha i is equally sensiive o growh in all perceniles below he povery line. This is why Ravallion and Chen (2003) argue ha a good measure of propoor growh is he average (across all perceniles) growh rae of hose below he povery line, i.e., he average growh rae of incomes of he poor. In his paper I go furher and decompose he average growh rae of incomes of he poor ino growh in average incomes and he average growh rae of he poor relaive o growh in average incomes. This allows me o disinguish beween he effecs of growh in average incomes and growh in relaive incomes on he Was measure, and all he oher measures considered here. This disincion is no rivial, as we will see in he empirical secion of he paper ha here is more evidence for he correlaes of growh in average incomes han growh in relaive incomes. Finally, when inequaliy aversion is posiive, i.e., θ<1, he EDEI measure is mos sensiive o growh in incomes of he poores, bu is non-zero for all income perceniles. The key difference wih he oher povery measures is ha here is no longer a disconinuiy a he povery line growh in all pars of he income disribuion maers for povery reducion, wih growh among he poores maering mos. To reierae, he imporan poin of Figure 2 is ha povery measures differ in heir sensiiviy o growh in differen perceniles of he income disribuion. As a resul, a given paern of relaive income growh migh be pro-poor (in he sense ha he povery measure falls) for some povery measures, bu no for ohers. Moreover, if we ake seriously he objecive of pro-poor growh wih respec o a paricular povery measure, hen his requires a growh sraegy focusing on paricular pars of he income disribuion. For example, pro-poor growh wih respec o he headcoun requires an emphasis on hose jus below he headcoun, while pro-poor growh wih respec o EDEI wih srong inequaliy-aversion requires inervenions argeed o reaching he poores of he poor. Finally consider he average across all perceniles of he sensiiviy η (p) of povery o growh in incomes of percenile p. Recall from equaion (3) ha his average sensiiviy measures he effec of growh in average incomes on he povery measure. We have been referring o high values of his average sensiiviy of povery o growh in average incomes as one of he hree poenial sources of pro-poor growh. For he Foser-Greer-Thorbecke class

11 of povery measures, his average sensiiviy can be expressed in erms of he povery measure iself when θ is no equal o zero, η ( p ) dp = ( P ( θ ) P ( θ 1) ) 1 0, where P (θ) denoes he FGT measure wih parameer θ. 6 In he case where θ is zero, he sensiiviy of he 1 L '( H ) headcoun o growh in average incomes is: η dp = which can be µ L ''( H ) 0 expressed as he slope of he densiy of income a he povery line. For he Was measure, he average elasiciy is simply minus one imes he headcoun. For EDEI, he sensiiviy o average growh is EDEI iself, implying ha he elasiciy of EDEI wih respec o growh in average incomes is one. While hese resuls are useful for analyically characerizing he sensiiviy of he differen povery measures o growh in average incomes, we will see shorly ha cross-counry differences in he sensiiviy of povery o growh in average incomes are no empirically very imporan, in he sense ha hey explain lile of he crosscounry variaion in he firs erm in equaion (3). We herefore do no discuss hem furher here. III. DATA The objecive of he res of his paper is o use he analyic framework discussed above o decompose observed changes in povery ino he hree erms discussed above: (a) growh in average incomes; (b) he sensiiviy of povery o growh in average incomes; and (c) changes in relaive incomes. Afer consrucing hese hree erms for a large sample of developing counries, I use hem o idenify he relaive imporance of, and facors correlaed wih, hese various sources of pro-poor growh. I use household survey daa on average incomes and en poins on he Lorenz curve for a large number of surveys, as compiled by Marin Ravallion and Shaohua Chen a he World Bank. Their daa comes direcly from primary sources, has been meiculously cleaned, and is available a hp:// 7 Depending on he counry, he surveys measure eiher he disribuion of income or he disribuion of consumpion. Average income or consumpion is measured in 1993 dollars and is adjused for crosscounry differences in purchasing power pariy. Since I am ineresed in changes in povery over ime, I ake only counries wih a leas wo household surveys. This resuls in a oal of 285 surveys covering 80 developing counries. Mos of he survey daes are in he 1990s, wih some counries exending back o he 1980s. I use he World Bank s dollar-a-day povery line which in 1993 dollars is $1.08 per day, or $393 per year. 6 This resul can be found in Kakwani (1993). 7 I am graeful o Shaohua Chen for kindly providing key daa from all of he household surveys, including some ha is no available on he povery monioring websie.

12 Using hese surveys, I consruc wo daases of spells of changes in povery. In he firs daase, I consider all possible spells for each counry, discarding only hose few cases where he survey changes from an income o an expendiure survey or vice versa. This resuls in 205 spells of povery changes. The lengh of hese spells is quie shor, averaging 3.5 years and ranging from one o 13 years. In order o be able o look a changes over longer horizons, I also consruc a daase consising of one spell per counry, where he iniial and final years are chosen so as o maximize he lengh of he spell given available daa. This resuls in a se of 80 spells, wih an average lengh of 8.2 years, and ranging from wo o 19 years. Finally I eliminae all spells where he headcoun measure of povery is negligible in eiher he iniial or final period, i.e., below wo percen, and I also drop a number of spells where he average annual growh rae in he survey mean is implausibly large, i.e., more han 15 percen in absolue value. This reduces he firs daase o 128 spells covering 58 counries wih an average lengh of 3.5 years, and he second daase o 42 spells wih an average lengh of 9.6 years. In order o consruc he differen povery measures and heir decomposiions discussed in he previous secion, I need he full Lorenz curve and no jus he 10 poins provided in he Ravallion-Chen daa. To obain his, I assume ha he Lorenz curve has he following funcional form: β γ ( 1 (1 p) ), α 0, 0 < β 1, 1 α L = p γ (4) This paricular parameerizaion is a member of a family of ordered Lorenz curves proposed by Sarabia, Casillo, and Sloje (1999). I esimae he parameers of his Lorenz curve for each survey using an algorihm suggesed by he same auhors. This involves selecing all possible combinaions of hree poins on he Lorenz curve, and hen for each combinaion finding values of α, β, and γ such ha he Lorenz curve passes hrough hese hree poins. The final esimaes of α, β, and γ are hen found by averaging across all he resuling esimaes of hese parameers, discarding hose for which he parameer resricions indicaed in Equaion (4) ha are required for he Lorenz curve o have posiive firs and second derivaives do no hold. I hen obain he quanile funcion by analyically differeniaing he Lorenz curve and muliplying by average income. Using his, I can immediaely consruc η (p) for each povery measure of ineres, as well as he growh incidence curve over he y observed discree inerval, g = 1. y 1 IV. RESULTS I begin by consrucing he povery measures of ineres (he headcoun, he povery gap, he squared povery gap, he Was index) for he iniial and final years of each spell. I hen compue average annual changes in hese measures, normalizing each by is iniial value so as o ge proporionae changes ha are more easily comparable across povery measures. Table 1 repors he simple correlaions of he levels and average annual growh raes in hese

13 povery measures wih he corresponding log-levels and growh raes of survey mean income. These simple correlaions are all negaive, and are large in absolue value, especially hose in levels and hose for he long spells. Figure 3 graphs he proporional change in each povery measure agains he growh rae of average incomes, using he sample of long spells. In each case, here is a srong and highly significan negaive relaionship beween changes in povery and change in average incomes. There is somewha more dispersion around his average relaionship for he more boom-sensiive povery measures such as he povery gap and he squared povery gap, han for he headcoun measure. However, his may simply reflec he greaer sensiiviy of boom-sensiive povery measures o measuremen error in individual incomes, as I argue in more deail below. Table 1 and Figure 3 confirm he widely undersood empirical regulariy ha povery measures end o fall as average incomes increase. In he following secions his paper will documen he relaive imporance of he differen sources of pro-poor growh discussed above, and some evidence on he correlaes of growh in average and relaive incomes. V. RELATIVE IMPORTANCE OF SOURCES OF PRO-POOR GROWTH I now documen he relaive imporance of he hree sources of pro-poor growh ha we have been discussing. I do his in wo seps. I firs decompose he change in povery in each spell ino a growh componen and a disribuion componen using he decomposiion suggesed by Da and Ravallion (1992), which is he discree-ime analog of equaion (3). Le P ( µ, L ) denoe a povery measure based on mean income a ime, µ, and he Lorenz curve a ime, L. I hen wrie he proporional change in he povery measure over he discree inerval beween ime and -1 as: P ( µ, L ) P( µ 1, L 1 ) P( µ, L ) 1 1 = P ( µ, L 1 ) P( µ 1, L 1 ) P( µ, L ) P ( µ 1, L ) P( µ 1, L 1 ) P( µ, L ) ε (5) The firs erm on he righ-had side is he growh componen of he change in povery, and is consruced as he proporional difference beween he iniial povery measure and a hypoheical povery measure compued using he second period mean bu he firs period Lorenz curve. The second erm is he disribuion componen which is compued as he proporional difference beween he iniial povery measure and a hypoheical povery measure consruced using he firs period mean bu he second period Lorenz curve. These wo componens are he discree-ime analogs of he wo erms in equaion (3). Unlike equaion (3), however, here is also a residual erm because he decomposiion is done over a discree and no an infiniesimal inerval. I measure he proporional changes on he lef- and righ-hand side of equaion (5) as log differences and normalize by he lengh of he inerval o ge average annual percen changes in povery and is growh and disribuion componens for each spell.

14 Tables 2 and 3 repor he resuls of applying his decomposiion o he wo daases of spells. Throughou hese wo ables, I use he following variance decomposiion o summarize he relaive imporance of he various componens. If X and Y are wo correlaed random variables, hen I define he share of he variance of X+Y due o variaion in X as VAR( X ) + COV ( X, Y). 8 The op panel of each able documens he imporance VAR( X ) + VAR( Y) + 2 COV ( X, Y ) of he residual relaive o he sum of he growh and disribuion componens of he change in povery. The firs column shows he variance of he sum of he growh and disribuion componens, he second column he variance of he residual, and he hird he covariance beween he wo. The final column repors he share of he variance of changes in povery due o he growh and disribuion componens, which is virually one for all povery measures. This simply reflecs he fac ha he variance of he residual erm is iny relaive o he variance in measured changes in povery. This can also be verified visually from he op panel of Figure 4, which graphs he change in povery on he horizonal axis, and he sum of he growh and disribuion componens on he verical axis, using he daase of long spells. The slope of he OLS regression line is he share of he variance in povery changes due o he growh and disribuion componens, and one minus he slope is he share due o he residual erm. I is clear from his graph ha changes in povery are largely accouned for by he sum of he growh and disribuion componens, wih very lile of he variaion due o he residual. The middle panel of Tables 2 and 3 does he same variance decomposiion, bu now o assess he imporance of he growh componen relaive o he disribuion componen of changes in povery. For he sample of all spells, beween one-hird and one-half of he variaion in changes in povery is due o he growh componen, wih he remainder due o changes in disribuion. The sory is quie differen for he long spells, where he growh componen of changes in povery dominaes, accouning for beween 65 and 90 percen of changes in povery. In boh ables, he growh componen is relaively less imporan for boomsensiive povery measures such as he povery gap and he squared povery gap. The middle panel in Figure 4 graphically summarizes his second decomposiion for he long spells sample, ploing he growh componen of changes in povery on he verical axis, and he sum of he growh and disribuion componens on he horizonal axis. Again, he slope of he OLS regression line can be inerpreed as he share of he variaion on he horizonal axis due o he growh componen. Visually inspecing his graph, i is clear ha if povery reducion is large, i is mosly because he growh componen of povery reducion is large. The boom panel of Tables 2 and 3 furher disenangles he growh componen ino growh in average incomes, and he sensiiviy of povery o growh in average incomes, i.e., i 8 When X and Y are normally disribued, his variance decomposiion has a very naural inerpreaion. I ells us how much he condiional expecaion of X increases for each uni ha we observe he sum (X+Y) o be above is mean value.

15 separaes he firs erm in equaion (3) ino is wo componens. Since he decomposiion we have been using applies o sums of random variables, I ake he logarihm of he absolue value of he growh componen, which hen becomes he sum of he logarihm of he absolue value of growh, and he logarihm of he absolue value of he average sensiiviy of povery o growh, and apply he decomposiion o his sum. Tables 2 and 3 show ha over 80 percen of he cross-counry variaion in he growh componen of changes in povery is due o crosscounry differences in average income growh, and very lile is due o cross-counry differences in he sensiiviy of povery o average income growh. The boom panel of Figure 4 illusraes his, bu wihou he log ransform required o do he variance decomposiion. On he horizonal axis I graph he growh componen of he change in povery, while on he verical axis I graph growh in average incomes. While he slope of his regression canno be inerpreed as a variance share, i neverheless is very clear ha crosscounry differences in he growh componen of povery are overwhelmingly accouned for by cross-counry differences in growh. Pu differenly, i is clear from his graph ha if he growh componen of povery reducion is large, i is mos likely ha growh iself was large, raher han ha he sensiiviy of povery o growh was large. 9 Two sriking feaures of Tables 2 and 3 meri furher discussion: (a) he share of he variance due o growh is smaller over he shor horizons represened in he daase of all spells, and is larger in he daase of long spells; and (b) in boh daases, he share of he variaion in povery measures due o growh declines as he povery measures become more boomsensiive, for example when we move from he headcoun o he povery gap o he squared povery gap. We can undersand hese properies beer wih he help of a simple example using he EDEI povery measure. We can wrie he discree proporional change in EDEI as: ln EDEI ( ln EDEI ( θ ) ln EDEI (1) ) ( θ ) = ln EDEI (1) + (6) Recall ha EDEI(1) is jus average income, and ha he sensiiviy of EDEI o growh in average incomes is one. As a resul, Equaion (6) is a way of wriing he Da Ravallion decomposiion for his measure, wih he firs erm corresponding o he growh componen 9 A firs glance his resul seems inconsisen wih Ravallion (1997), who documens ha he sensiiviy of povery o growh varies significanly wih iniial inequaliy. However, using eiher sample of spells I can replicae he resul ha he ineracion of growh wih he iniial Gini coefficien is significanly correlaed wih he change in headcoun measures of povery. Inuiively, he difference beween he resuls here and hose in Ravallion (1997) can be undersood as follows: alhough he ineracion of growh wih iniial inequaliy is significan in explaining changes in povery, i does no add much o he explanaory power of he regression in my samples. Pu differenly, alhough here are cross-counry differences in he sensiiviy of povery o growh which are significanly correlaed wih iniial inequaliy, in he daa hese differences are dominaed by he much larger cross-counry differences in growh iself.

16 and he second o he disribuion componen. Moreover, he disribuional change componen of he change in his povery measure corresponds o he change in a paricular inequaliy measure: i is simply he proporional change in one minus he Akinson inequaliy measure. 10 For he purpose of his example, assume ha he logarihm of household incomes is disribued normally wih mean µ and sandard deviaion σ. As he number of households in each counry becomes large, i is sraighforward o see ha ln EDEI(θ) converges in θ 2 probabiliy o µ + 2 σ. 11 Using his resul, we can wrie he Da Ravallion decomposiion for EDEI as: p θ 1 2 ln EDEI ( θ ) µ + σ (7) 2 Equaion (7) is helpful for undersanding he wo key feaures of Tables 2 and 3 menioned 2 above. Suppose ha µ and σ are independen across counries. Then he share of 2 variance of changes in povery due o growh wih be [ ] θ 1 2 V µ [ ] [ ] / V µ + V σ. 2 The furher θ is from one, i.e., he more EDEI weighs incomes of he poor (for θ<1) or he rich (for θ>1), he smaller is he share of he variance of change in povery due o growh and he larger is he share due o he disribuion componen. In oher words, he more boomsensiive (or for ha maer, op-sensiive), he povery measure, he larger will be he conribuion of changes in relaive incomes o changes in he povery measure. This suggess an explanaion why he share of he variance of changes in povery due o growh declines as he povery measures become more boom-sensiive. Equaion (7) is also helpful for hinking abou why he share of he variance of changes in all povery measures due o growh is smaller in he shor run han in he long run. One possible explanaion is ha measuremen error in changes in inequaliy is relaively more imporan han measuremen error in changes in average incomes when he period under consideraion is shor. I is no clear how one migh direcly documen ha his is he case. However, i is 10 The Akinson class of inequaliy measures is 1-EDEI(θ)/EDEI(1). 11 This is because EDEI(θ) 1/θ is he sample average of incomes raised o he power θ. As he number of households becomes large, his converges o he expecaion of income raised o he power θ. If incomes are lognormally disribued, we can use he momen generaing funcion of he lognormal disribuion o evaluae his expecaion o obain he resul in he ex.

17 worh noing ha his paern of relaive imporance of measuremen error seems quie plausible. Suppose for example ha in every period, log household income is measured wih an addiive zero-mean measuremen error, which is independen of rue incomes and is i.i.d. normal across households. If he number of households is large, his zero-mean measuremen error will no be refleced in average income. However, he variance of measured log incomes will now be σ + ξ, where ξ is he variance of measuremen error. Suppose furher ha he variance of measuremen error flucuaes randomly over ime. As long as he variance of measuremen error does no rend up or down oo fas, he average annual change θ 1 σ σ k ξ ξ k in he disribuion componen of changes in povery, + 2 will be k k smaller he longer is he ime inerval, k. While his is no conclusive, i does sugges ha par of he reason for he relaively smaller imporance of he growh componen of changes in povery over shorer horizons migh simply be measuremen error in household incomes. In summary, he resuls in his subsecion ell us ha, over longer horizons, beween 65 and 90 percen of cross-counry differences in povery changes can be accouned for by growh in average incomes. Over shorer horizons he share of he variance of changes in povery due o changes in growh is somewha smaller, and changes in income disribuion are relaively more imporan. However, his may in par be an arifac of measuremen error in individual incomes. While here are of course cross-counry differences in he sensiiviy of povery changes o average income growh, reflecing cross-counry differences in he iniial disribuion of income, empirically hese are relaively unimporan in undersanding changes in povery. Finally, alhough hese calculaions are done based on a discree-ime decomposiion wih unavoidable residuals, empirically hese residuals are also small and do no derac from he main conclusions. VI. WHAT DRIVES THE SOURCES OF PRO-POOR GROWTH? I now urn o he quesion of wha drives he various sources of pro-poor growh. In ligh of he resuls of he previous secion ha cross-counry differences in he sensiiviy of povery o growh in average incomes are relaively unimporan, I focus primarily on he firs and hird sources of pro-poor growh: growh in average incomes, and changes in relaive incomes. I measure growh in average incomes as he average annual growh rae over he spell of household average income or consumpion. I use five differen measures of changes in relaive incomes. The firs is simply he average annual change in he Gini index, for comparabiliy wih exising resuls on he deerminans of changes in inequaliy. The nex four measures are he discree-ime disribuion componens of he change in each of he four povery measures I have been considering. Recall ha, for infiniesimal changes, he disribuion componen of he change in he headcoun measures he growh rae of incomes of hose a he povery line relaive o average growh. For he povery gap and he squared povery gap, he disribuion componen measures a weighed average of relaive growh raes of hose below he povery line, wih he povery gap giving mos weigh o hose a he povery line. For he Was index, he disribuion componen measures he average growh rae of hose below he povery line relaive o overall growh.

18 There are many limiaions o his daase which make i very difficul o use i o idenify causal deerminans of growh or change in relaive incomes. The sample of counries is quie small, especially when we consider he long spells daase where he deerminans of longererm growh and disribuional change are more likely o be apparen. There is also subsanial measuremen error in he daa on growh in survey means, and for measures of disribuional change. While classical measuremen error in hese dependen variables will no necessarily lead o biases in coefficien esimaes, i will inflae sandard errors and reduce he significance of esimaed coefficiens. Because we have relaively few spells per counry in he daase consising of all spells, and only one per counry in he long spells daase, we canno meaningfully base idenificaion on he wihin-counry variaion in he daa. This raises he possibiliy ha any parial correlaions we uncover may be driven by unobserved counry-specific characerisics excluded from he regressions. The small number of spells per counry also means ha we will no be able o rely on inernal insrumens o achieve idenificaion. 12 In ligh of hese difficulies, my more modes objecive here is o simply documen he parial correlaions beween hese sources of pro-poor growh and a number of righ-hand-side variables of ineres, and o inerpre hem wih an appropriae abundance of cauion. I consider he same lis of righ-hand-side variables as in Dollar and Kraay (2002). In ha paper, we considered a small number of variables ha are frequenly found o be robusly correlaed wih real GDP growh in he cross-counry growh lieraure: insiuional qualiy as proxied by a measure of propery righs proecion (he rule of law indicaor from Kaufmann, Kraay, and Masruzzi (2003)), as well as he World Bank s Counry Policy and Insiuional Assessmen (CPIA) indicaor; openness o inernaional rade (he consan-price local currency raio of expors plus impors o GDP); inflaion as a proxy for sable moneary policy (measured as he logarihm of one plus he CPI inflaion rae); he size of governmen (measured as he share of governmen consumpion in GDP in local currency unis); and a measure of financial developmen (he raio of M2 o GDP in local currency unis). We also considered a number of variables ha are generally less robusly correlaed wih growh, bu ha some sudies have found o be correlaed wih inequaliy, eiher in levels or in differences. These include a measure of democracy (he voice and accounabiliy indicaor from Kaufmann, Kraay, and Masruzzi (2003)); relaive produciviy in agriculure (measured as he raio of value added per worker in agriculure relaive o overall value added per worker, boh in curren local currency unis); and primary educaional aainmen. 12 This is of course especially problemaic for he regressions below ha involve a lagged dependen variable, which, ogeher wih unobserved counry-specific effecs, will make esimaes of he coefficien on he lagged dependen variable inconsisen, and can bias he coefficiens on he oher variables in differen direcions depending on heir correlaion wih he lagged dependen variable.

19 This lis of variables is clearly no an exhausive lis of he poenial deerminans of growh in average incomes or changes in relaive incomes. However, i does provide us wih a useful place o begin looking for he correlaes of growh and disribuional change ha maer for povery reducion. I begin by esimaing a number of very parsimonious regressions for each of he dependen variables of ineres. I regress growh in average incomes on he log-level of iniial period income (o pick up convergence effecs) plus each of he conrol variables described above, one a a ime. I do he same for he change in he Gini coefficien, insead including he iniial level of he Gini coefficien o pick up convergence in his variable. For he remaining four disribuion componens of changes in povery, I simply esimae univariae regressions of each one on each of he righ-hand-side variables. 13 Table 4 shows he resuls using he sample of all spells, and Table 5 shows he same informaion bu using only he smaller sample of long spells. Each enry in hese wo ables corresponds o a differen regression. The rows correspond o each of he indicaed righhand-side variables. The columns correspond o he differen dependen variables. The firs wo columns repor regressions for growh and for he change in he Gini. Boh hese regressions also include eiher iniial log income or he iniial Gini. I do no repor he coefficiens on hese variables o save space, bu hey generally ener negaively and usually significanly in all specificaions, consisen wih available evidence on convergence in boh of hese variables. The remaining columns repor resuls for he disribuion componen of he change in each of he four povery measures. Recall ha hese measures are oriened such ha a reducion corresponds o a reducion in povery. A firs glance a Tables 4 and 5 shows ha very few of he explanaory variables of ineres are significanly correlaed wih he dependen variable of ineres a convenional significance levels. In fac, in he 108 regressions in hese wo ables, here is only one coefficien ha is significan a he 5 percen level, and only hree ha are significan a he 10 percen level. One possible explanaion for he lack of significan resuls is ha he measures of growh and disribuional change on he righ-hand-side are conaminaed by subsanial measuremen error. I is difficul o judge however by how much sandard errors should be adjused o reflec his measuremen error. 14 Raher han ry o assess he saisical significance of he parial correlaions documened in Tables 4 and 5, I simply describe some of he qualiaive paerns ha emerge. 13 Ravallion (2001) documens he empirical imporance of inequaliy convergence using he Gini coefficien. I have experimened wih alernaive iniial inequaliy measures in he regressions involving he disribuional change componens of he various povery measures, bu I find ha none are robusly significan. 14 In Table 4, here is an addiional facor which likely biases sandard errors upward. For counries wih muliple spells of growh or disribuional change, here is likely o be by consrucion a negaive correlaion beween he errors of successive spells. Correcing for his will likely reduce sandard errors somewha.

20 Consider firs insiuional qualiy, as proxied by he rule of law indicaor. This ends o be posiively correlaed wih growh, bu also posiively correlaed wih each of he measures of disribuional change, suggesing ha disribuional change ends o raise povery in counries wih good insiuional qualiy. However, he srengh of he correlaion wih growh is much larger han he correlaions wih disribuional change: he -saisic from he growh regression is abou wice he average -saisic for he differen measures of disribuional change. The voice and accounabiliy measure follows he same paern, likely because i is quie highly correlaed wih rule of law in his sample. In he case of openness o inernaional rade, he correlaion wih growh is generally sronger han he correlaions wih disribuional change. Moreover, he sign of he correlaion wih each of he measures of disribuional change is negaive, indicaing ha disribuional change ends o be povery-reducing in counries ha rade more. Inflaion ends o be exremely weakly correlaed wih growh in his sample, and ends o be posiively correlaed wih disribuional change, bu again he correlaion is very weak. Governmen consumpion is negaively correlaed wih growh, bu ineresingly is also negaively correlaed wih each of he measures of disribuional change, suggesing ha disribuional change ends o be pro-poor in counries wih larger governmens. Financial developmen also appears o be very weakly correlaed wih eiher growh or disribuional change in hese regressions. Relaive produciviy in agriculure is essenially uncorrelaed wih growh, bu ends o be posiively correlaed wih disribuional change measures. Somewha surprisingly he sign of he correlaion suggess ha counries wih higher relaive produciviy in agriculure are more likely o experience povery-increasing changes in relaive incomes. Finally, primary educaion is also virually uncorrelaed wih growh, and also is essenially uncorrelaed wih mos of he disribuional change measures, wih he excepion of he Gini in he long spells regression. Overall, while mos of he parial correlaions documened in Tables 4 and 5 are no saisically significan, he qualiaive paern suggess ha here may be some radeoffs. Rule of law is posiively correlaed wih growh bu also wih povery-increasing shifs in relaive incomes. The opposie is rue for governmen consumpion. In conras rade is posiively correlaed wih growh and wih povery-reducing shifs in relaive incomes. In Table 6 we examine hese possible radeoffs in a slighly richer empirical specificaion, using he daase of long spells. We begin by esimaing a more fully-specified growh regression wih iniial income, and iniial values of insiuional qualiy, rade openness, and size of governmen as righ-hand-side variables. Despie he likely noisiness of he daa, i is possible o find plausible specificaions in which some of he deerminans of growh from he growh lieraure are also significanly correlaed wih growh in he household survey mean. The firs column of Table 6 illusraes one such regression, which includes iniial income, insiuional qualiy, rade openness, and governmen consumpion on he righhand-side. Each of hese variables eners wih signs consisen wih he broader growh lieraure. Iniial income eners negaively, picking up convergence effecs. Insiuional