Publc Dsclosure Authorzed Publc Dsclosure Authorzed Publc Dsclosure Authorzed Publc Dsclosure Authorzed THE COMPOSITION OF GROWTH MATTERS FOR POVERTY ALLEVIATION * Norman Loayza The World Bank Abstract Claudo Raddatz The World Bank Ths paper contrbutes to explan the cross-country heterogenety of the poverty response to changes n economc growth. It does so by focusng on the structure of output growth tself. The paper presents a two-sector theoretcal model that clarfes the mechansm through whch the sectoral composton of growth and assocated labor ntensty can affect workers wages and, thus, poverty allevaton. Then, t presents cross-country emprcal evdence that analyzes, frst, the dfferental poverty-reducng mpact of sectoral growth at varous levels of dsaggregaton, and, second, the role of unsklled labor ntensty n such dfferental mpact. The paper fnds evdence that not only the sze of economc growth but also ts composton matters for poverty allevaton, wth the largest contrbutons from labor-ntensve sectors (such as agrculture, constructon, and manufacturng). The results are robust to the nfluence of outlers, alternatve explanatons, and varous poverty measures. Keywords: Poverty, economc growth, producton structure, labor ntensty. World Bank Polcy Research Workng Paper 4077, December 2006 WPS4077 The Polcy Research Workng Paper Seres dssemnates the fndngs of work n progress to encourage the exchange of deas about development ssues. An objectve of the seres s to get the fndngs out quckly, even f the presentatons are less than fully polshed. The papers carry the names of the authors and should be cted accordngly. The fndngs, nterpretatons, and conclusons expressed n ths paper are entrely those of the authors. They do not necessarly represent the vew of the World Bank, ts Executve Drectors, or the countres they represent. Polcy Research Workng Papers are avalable onlne at http://econ.worldbank.org. * We are grateful to Yvonne Chen and Syan Chen for research assstance and to Koch Kume for edtoral assstance. For useful comments, suggestons, and/or data we are ndebted to Maros Ivanc, Francsco Ferrera, Oded Galor, Aart Kraay, Humberto López, Ross Levne, Martn Ravallon, Yona Rubnsten, Lus Servén, and semnar partcpants at Brown Unversty and the World Bank. Support from the Chef Economst Offce of the Latn Amerca and Carbbean Regon of the World Bank s gratefully acknowledged.
THE COMPOSITION OF GROWTH MATTERS FOR POVERTY ALLEVIATION I. Introducton There s lttle doubt that economc growth contrbutes sgnfcantly to poverty allevaton. The evdence s mountng and comng from varous sources: cross-country analyses (Besley and Burgess, 2003; Dollar and Kraay, 2005; Kraay, 2005; and López, 2004), cross-regonal and tme-seres comparsons (Ravallon and Chen, 2004; Ravallon and Datt, 2002), and the evaluaton of poverty evoluton usng household data (Bb, 2005; Contreras, 2001; Menezes-Flho and Vasconcellos, 2004). At the same tme, t s clear that the effect of economc growth on poverty reducton s not always the same. In fact, most studes pont to consderable heterogenety n the poverty-growth relatonshp, and understandng the sources of ths dvergence s a growng area of nvestgaton (Bourgugnon, 2003; Kakwan, Khandker, and Son, 2004; Lucas and Tmmer, 2005, Ravallon, 2004). Most of the receved lterature focuses on soco-economc condtons of the populaton as determnants of the relatonshp between growth and poverty reducton. Thus, wealth and ncome nequalty, lteracy rates, urbanzaton levels, and morbdty and mortalty rates, among others, have been found to nfluence the degree to whch output growth helps reduce poverty. In ths paper we take a dfferent, albet complementary, perspectve on the sources of heterogenety n the poverty-growth relatonshp. We focus on the characterstcs of output growth tself, rather than the demographc, socal, or economc condtons of the populaton. We study how the producton structure of the economy and, specfcally, the sectoral composton of growth affect ts capacty to reduce poverty. Our conjecture s that growth n certan sectors s more poverty reducng than growth n others and that a sector s poverty-reducng capacty s related to ts ntensty n the employment of unsklled labor. There are mportant studes that precede and motvate our work. Thorbecke and Jung (1996) develop a socal-accountng method to estmate the mpact of varous producton actvtes on poverty reducton. The method requres knowledge of complex elastctes connectng the dstrbuton of households wth eght employment and producton sectors. The authors apply the method to Indonesa n the 1980s and fnd that 1
agrcultural and servce sectors contrbute more to poverty reducton that ndustral sectors do. Khan (1999) apples the same methodology to study sectoral growth and poverty allevaton n South Afrca. He fnds that hgher contrbutons are derved from growth n agrculture, servces, and some manufacturng sectors. A dfferent approach conssts of conductng reduced-form analyss on tme-seres data for ndvdual countres. Ths s the approach taken by Ravallon and Datt (1996) to study the evoluton of poverty n Inda durng 1951-91. Lnkng poverty changes to value-added growth rates n the three major sectors of economc actvty, they fnd that growth n agrculture and servces helped reduce poverty n both urban and rural areas whereas ndustral growth dd not reduce poverty n ether. Applyng a smlar methodology for the case of Chna over 1980 2001, Ravallon and Chen (2004) fnd that growth n agrculture emerges as far more mportant than growth n secondary or tertary sectors for the purpose of poverty allevaton. Our work adds to ths lterature along four dmensons. Frst, we present a twosector theoretcal model that clarfes the mechansm through whch the sectoral composton of growth and assocated labor ntensty can affect workers wages --and, thus, poverty allevaton-- even n the absence of market segmentaton. Second, we use cross-country evdence --wth the pros and cons assocated wth ncreasng the underlyng varaton of the data-- allowng us to relate our results to the emprcal macroeconomc lteratures on growth and poverty. Thrd, we employ a level of dsaggregaton that explores the dversty wthn the ndustral sector, hopng to shed lght on why t appears to be less pro poor than agrculture or servces. And, fourth, we explctly consder sectoral employment ntensty as the mechansm through whch the pattern of growth matters for poverty allevaton. The plan of the paper s the followng. Secton II presents a theoretcal model that formalzes our ntal conjecture. It examnes the wage (poverty) effect of output growth n a two-sector economy, where captal and labor are freely moble and the sectors technologes vary accordng to ther labor ntensty. Secton III presents cross-country emprcal evdence that analyzes, frst, the dfferental poverty-reducng mpact of sectoral growth at varous levels of dsaggregaton, and, second, the role of unsklled labor ntensty n such dfferental mpact. Also n ths secton, we subject our basc 2
result to a comprehensve set of robustness checks that account for the nfluence of outler and extreme observatons, for potental alternatve explanatons, and for varous poverty measures. Secton IV offers some concludng remarks. II. The Model We now present a two-sector model wth asymmetrc technologes to help us understand the relaton between sectoral growth and poverty allevaton. We focus on the two-sector case for smplcty, but the results are analogous for the n-sector case. The economy s populated by two types of ndvduals: poor and rch. Both types are endowed wth n unts of labor, derve utlty from the consumpton of a fnal good, and have the same dscount factor ρ and nstantaneous utlty functon u(c)=log(c). However, only rch ndvduals have access to an asset a that allows them to transfer wealth across perods. Ths settng mples that the ncome and consumpton of poor ndvduals depend only on the real wage rate. Thus we assume that the rate of poverty reducton s related only to the growth rate of real wages. Although ths s an extreme assumpton, t smplfes consderably the analyss and s roughly consstent wth the low savng rates observed both n poor countres and poor households wthn a country. 1 The fnal good, y, s produced by a perfectly compettve frm usng a constantreturns-to-scale technology and two ntermedate goods, y 1 and y 2, as nputs accordng to ( ) 1 β β β 1 2, y = y + y The fnal good can be used not only for consumpton but also as captal n the producton of the ntermedate goods. Each ntermedate good s produced by a perfectly compettve frm accordng to the followng technology wth labor-augmentng technologcal progress, 1 Schmdt-Hebbel and Serven (1999) show that savng rates ncrease wth ncome across countres. In poor countres, the savng rates are below 10%. Attanaso and Székely (1998) provde evdence on households savng rates at dfferent levels of the ncome dstrbuton n Mexco. Ther data show that savng rates ncrease strongly wth ncome and dsplay even negatve values up to the 25th percentle of the household ncome dstrbuton. 3
( ) (1 α ) α y = k An, = 1,2 where k and n are sector s captal and labor, respectvely, and A captures the level of technology, whch evolves exogenously accordng to A = exp( gt). We assume that captal s perfectly moble across sectors and does not deprecate. Intersectoral allocaton In what follows we use ths setup to derve the relaton between the composton of growth and the evoluton of the real wage rate, whch gven our assumptons maps nto the ncome and consumpton of the poor. We wll only focus on the aspects of the model that are relevant for the dervaton of ths expresson and omt several sde aspects of the characterzaton. Under perfect competton, the prce charged by the fnal-good frm, p, equals ts unt cost of producton. Then, ( ) 1 1 1 1 ε ε ε 1 2, p= p + p where ε 1 = (1 β). Solvng the optmzaton problem of the fnal-good frm and settng the prce of the fnal good as a numerare, we obtan the followng frst order condtons, py Y ε 1 ε y = s =, = 1,2 Y (1) whch characterze the share of the fnal good producton value that goes to each ntermedate sector. Gven that the producton of the fnal good exhbts constant returns to scale these shares add up to one. Combnng the frst order condtons, we obtan the followng expresson for the demand of ntermedate goods, y y 1 2 p = p 2 1 ε, (2) 4
whch shows that ε corresponds to the (constant) elastcty of substtuton between the ntermedate goods. Under perfect competton, each ntermedate-good frm determnes ts demand for labor and captal takng factor and output prces as gven. Then, the frstorder condtons correspondng to the ntermedate-good frm are gven by y = ωn rk, 1,2, pα = p (1 α ) = (3) Equatons (2) and (3)--whch correspond to the standard condtons for statc effcency-- plus the condtons of factor market equlbrum -- k + k =, and 1 2 k n1+ n2 = n-- determne the allocaton of labor and captal across sectors at every moment. Although n prncple we could use the prevous equatons to determne the relatve prces of the ntermedate goods p 1 / p 2 as a functon of the aggregate captallabor rato k/ n, technologcal parameters, and sector productvtes A, ths problem cannot be solved n closed form except n some specal cases that restrct the values of ε and the α (see Myagwa and Papageorgu, 2005, for a dscusson). Nevertheless, we can use these equatons to characterze the evoluton of real labor ncome, whch s the object of nterest for our emprcal analyss. The evoluton of real labor ncome Accordng to the frst-order condtons for ntermedate-good frms, and focusng wthout loss of generalty on ntermedate good 1, the rate of change of the real wage can be wrtten as ˆ ω = pˆ + yˆ nˆ (4) 1 1 1 where the hat denotes the rate of change of a varable ( xˆ = dx / x ). The frst two terms of ths expresson correspond to the evoluton of the value of sector 1 output n terms of the fnal good ( p1y 1). From equaton (1) ths corresponds to 1 1 sˆ1 Yˆ ε + = yˆ1+ ( s ˆ ˆ 1y1+ s2y2) (5) ε ε where we have used the fact that Yˆ = s ˆ ˆ 1y1+ s2y2because of constant returns to scale. 5
The last term n equaton (4) s the evoluton of employment n sector 1. The frst order condtons of ntermedate good frms wth respect to labor presented n equaton (3) together wth equaton (2) mply that α 1 n 2 y 1 α n y 2 1 2 ε 1 ε = 1, (6) whch, after log-dfferencng and usng the labor market clearng condton n= n1+ n2, results n the followng expresson for the growth rate of employment n sector 1, where l 2 s the share of employment n sector 2 ( n 2 / n). ε 1 nˆ1 = l2 ( yˆ ˆ 1 y2) + nˆ, (7) ε Fnally, puttng together equatons (5) and (7) and re-arrangng terms we obtan the growth rate of the real wage rate, 2 2 1 ˆ ε ω = syˆ + ( l s) yˆ (8) = 1 ε = 1 where, n a slght abuse of notaton, y ˆ now represent the growth rates n per-capta terms. Ths equaton ndcates that the growth of real labor ncome s drven by two components. The frst one, correspondng to the frst term on the left-hand sde of equaton (8), s the growth of per-capta GDP. An ncrease n per-capta GDP corresponds to a hgher output per worker that maps nto hgher wages. The contrbuton of a sector s growth to ths term depends exclusvely on ts sze, as captured by ts share on fnal-good output, s. The second component captures the reallocaton effects. The mpact of a sector s growth on ths component depends on the elastcty of substtuton across sectors n the producton of the fnal good (ε ) and on a sector s labor ntensty, as captured by the dfference between ts labor share of total employment, l, and ts share n total output s. Startng from equatons (1) and (3), t can be shown that ths dfference corresponds to 6
1 1 l s =, = 1,2 α s s 1+ 1+ α s s (9) whch ndcates that the dfference l -s s hgher for sectors wth a hgher share of labor n total output, α. Ths means that growth n a labor ntensve sector wll have an addtonal effect on wages beyond ts mpact on aggregate growth, as long as the elastcty of substtuton s suffcently hgh (specfcally above 1, accordng to eq. (8)). 2 The elastcty of substtuton s relevant because t determnes whether (and by how much) labor wll move nto or out of a growng sector: the hgher the elastcty of substtuton, the more labor moves nto that sector. If the elastcty s too low (below 1) labor actually moves out of an expandng sector; however, as the elastcty ncreases and surpasses a threshold value (equal to 1), labor starts to flow nto the growng sector. Wth a hgh (low) elastcty of substtuton, the prce adjustment requred by an ncrease n the relatve output of a sector s small (large) so that labor needs to move nto (out of) the expandng sector to acheve wage equalzaton (ths can be clearly seen n eq. (6)). Equaton (8) also shows that there are two cases n whch the growth rate of real labor ncome depends only on GDP growth: () when the technologes of the ntermedate sectors are dentcal ( α1 = α2), and () when the elastcty of substtuton s equal to one (the Cobb-Douglas case). The frst case s trval: f there are no asymmetres across sectors, uneven growth s rrelevant. In the second case, under a Cobb-Douglas producton functon, sectoral labor shares are constant, and any adjustment n relatve quanttes results only n a correspondng change n relatve prces. Uneven sectoral growth, not requrng labor reallocaton across sectors, would not affect real wages. Thus, omttng the composton of growth as a determnant of real wage ncreases and poverty allevaton s equvalent to assumng that ether sectors do not dffer n ther labor ntenstes or ther elastcty of substtuton s equal to one. Although not explctly derved n the model, t should be noted that the presence of technologcal progress s mportant for the long-run mplcatons of the model (that s, 2 Consder the followng example. Suppose that sector 1 s more labor ntensve than sector 2 (α 1 /α 2 >1), so that l 1 -s 1 >0, and that t experences an exogenous ncrease n productvty. If the elastcty of substtuton s suffcently hgh, labor wll move nto sector 1 where t s relatvely more productve, pushng the wage rate up. The opposte wll happen f the elastcty of substtuton s relatvely low (below 1). 7
beyond transtonal dynamcs). If the model exhbts a balanced-growth path, the growth rate of each sector y ˆ and of the economy wll be exclusvely determned by the growth rates of productvty n all the dfferent sectors of the economy (the g s). Characterzng the balanced growth path of the model s beyond the scope of ths paper; nevertheless, equaton (8) s vald both durng transtonal dynamcs and n balanced growth. Our assumpton that poverty changes are only a functon of the growth rate of real labor ncome corresponds to assumng that h ˆ = ψ ( ˆ ω), where ĥ s the growth rate of poverty. In the emprcal secton of the paper, we wll estmate the parameters of the lnearzed verson of ths relaton hˆ = γ ˆ 0 + γω 1 as our benchmark case, but we wll also consder some non-lnear relatons n our robustness analyss. III. Emprcal Evdence Our emprcal analyss conssts of two related sectons. In the frst, we address the connecton between the pattern of growth and poverty allevaton by dsaggregatng growth nto ts sectoral components and examnng ther correspondng effects on poverty. Ths s the tradtonal approach, and, thus, t allows us to place our analyss n the context of the receved lterature. The second emprcal secton modfes the sectoral analyss by ntroducng labor ntensty as the source of the dfferental mpact of sectoral growth on poverty reducton. Ths approach s derved from the theoretcal model and, thus, establshes the lnk between theory and emprcs n the paper. Data and sample Our sample conssts of a cross-secton of developng countres wth comparable measures of poverty changes, dsaggregated value-added growth rates at 3- and 6-sector levels, and unsklled employment at the same levels of dsaggregaton. In practce, our dataset s the result of combnng the Kraay (2005) database on poverty spells, 3 World 3 The Kraay database results from processng ncome dstrbuton data for a large number of developng countres. In turn, ts source s the collecton of household survey data estmated from prmary sources and made comparable across countres by Martn Ravallon and Shaohua Chen at the World Bank. For detals, see Kraay (2005). 8
Bank (2005) data on sectoral value added, 4 and Purdue Unversty s Global Trade Analyss Project database (GTAP, 2005) on labor shares. We focus on changes occurrng over long horzons, where the poverty reductoneconomc growth relatonshp s most stable. For ths reason we use only one spell per country, where the duraton of the spell corresponds to the longest perod for whch ntal and fnal poverty data exst for the country. The rest of the varables (e.g., value added growth rates and labor ratos) are calculated over the correspondng perod per country. The dependent varable s the proportonal change n poverty over a perod of tme (spell) per country. Specfcally, ths s the annualzed change n poverty as proporton to average poverty over the perod. 5 Gven ts mportance n the lterature, the benchmark poverty measure n the paper s the headcount poverty ndex, defned as the fracton of the populaton wth ncome below a gven poverty lne. In robustness exercses, however, we use alternatve measures of poverty, comprsng other members of the Foster-Greer-Thorbecke class of measures (the poverty gap and the squared poverty gap) and the Watts ndex. Followng conventon for cross-country comparablty, the poverty lne s set to $1 per person per day, converted nto local currency usng a purchasng-power-party adjusted exchange rate. Regardng the explanatory varables, we work wth growth rates of sectoral valueadded and employment data at two levels of dsaggregaton. The frst s the tradtonal sectoral dvson of agrculture, ndustry, and servces. The second one dsaggregates ndustry further nto mnng, manufacturng, utltes, and constructon. Sectoral growth rates are calculated drectly from data on sectoral value added as annualzed log changes of per capta value added between the end and start of the correspondng spell. Employment data s calculated ndrectly from data on sectoral value added and payments 4 The World Bank (2005) data on sectoral value added s complemented wth statstcs from the Inter- Amercan Development Bank and the Unted Natons. 5 That s, proportonal poverty change = 1 P P F I *, where P represents the poverty measure; T, the T ( P + P F I )/2 length of the spell; and the subscrpts I and F, ntal and fnal, respectvely. Calculatng the proportonal change wth respect to the average measure allows us to avod abnormally large proportonal changes when very low ntal and/or fnal measures are present, as would be the case f log dfferences were used. Kraay (2005) uses the latter procedure and then s forced to drop a consderable number of observatons. Were we to use Kraay s method, we would be workng wth 32 country observatons, rather than 51, the sample sze of our benchmark regresson. 9
to unsklled workers. Under the assumpton of wage equalzaton, the rato of unsklled workers n a sector to total unsklled workers n the country s calculated as the rato of payments to unsklled workers n the sector to total payments to unsklled workers n the economy. Regardng data for ths calculaton, only one observaton per country or per smlar countres s avalable from the orgnal source (GTAP). 6 The resultng sample conssts of 55 countres for 3-sector data and 51 countres for 6-sector data. Appendx 1 provdes the lst of countres ncluded n the sample, as well as the ntal and fnal years of ther correspondng spell. Appendx 2 provdes defntons and sources for all varables used n our emprcal exercses, and Appendx 3 presents basc summary statstcs on the 51-country sample. Poverty reducton and sectoral growth We are nterested n estmatng the effect of sectoral growth on poverty reducton. The regresson equaton can then be wrtten as, I + δ sj yˆ 0 j =1 hˆ j = δ + ε j (10) where ĥ s the annualzed rate of change of the headcount poverty ndex, ŷ s the annualzed rate of change of sectoral value added, s s the sectoral value added share n GDP, and the subscrpt and j represent sector and country, respectvely. All growth rates are expressed n per capta terms, and the sector shares are calculated from constantprce magntudes. 7 The set I conssts of three or sx sectors, dependng on whether ndustry s consdered as a whole or dsaggregated nto ts four major categores. In prncple, t may be possble to estmate the poverty effect of output changes n a levels 6 Gven that, n most cases, the date of ths observaton dffers sgnfcantly from the years of our poverty spell, we frst use the GTAP data to compute the rato of payments to unsklled workers to a sector s value added and assume t to be constant over tme, for a gven sector and country. We then use ths rato and the sector s share n total value added durng our spell (from World Bank, 2005) to compute the correspondng rato of unsklled workers n the sector to total unsklled workers n the country. Under wage equalzaton, the rato of unsklled workers n a sector to total unsklled workers can be wrtten as l l = α s α s, where α s the rato of unsklled labor payments to sector s k value added, and s k k k k k s the share of sector k n total value added. 7 Calculatng the shares from nomnal magntudes would more closely approxmate the theoretcal model. However, we work wth constant-prce shares because, frst, ther resultng country coverage s larger than when usng current-prce shares, and, second, they are very smlar and render bascally the same econometrc results. 10
regresson. However, the lterature advces a regresson n dfferences to control for fxed effects that may be drvng both poverty and output, such as a host of country-specfc development-related varables n our cross-country settng. Our regresson specfcaton weghts sectoral growth by ts relatve sze. As Ravallon and Chen (2004) pont out, ths specfcaton has the advantage that t allows for a smple test of whether the growth composton matters: If the null hypotheses that the coeffcents δ are equal to each other cannot be rejected, then the sectoral regresson collapses to one where GDP growth s the only relevant explanatory varable. In ths case, only sze and not composton of growth would matter for poverty allevaton. Our regresson specfcaton also allows for testng whether these sectors can be grouped n dfferent categores, not accordng to ther output characterstcs but accordng to ther relatonshp wth poverty reducton. Ths wll become mportant when we study the case of sx-sector dsaggregaton. Table 1 presents the results when GDP s decomposed nto agrculture, ndustry, and servces. The regressons are conducted usng both the full sample of 55 countres and the subset of 51 countres for whch sx-sector data are avalable. The latter exercse s conducted wth the purpose of comparson wth the sx-sector analyss. In both samples (columns 1 and 3, respectvely), the sze-adjusted value-added growth rates of all sectors fal to carry statstcally sgnfcant coeffcents. Moreover, the hypothess that the coeffcents are the same cannot be rejected. The lack of ndvdual sgnfcance of sectoral growth rates and the nablty to separate ther effects ndcates that the three major sectors are hghly lnked n ther relatonshp wth poverty reducton. Ths may be nterpreted as evdence aganst the mportance of growth composton for poverty allevaton, but t may also be the result of workng wth nsuffcently dsaggregated output categores. We examne the latter possblty below when we analyze the sx-sector case. Before dong that, however, we can take the falure to reject the equalty of coeffcents at face value and estmate a constraned regresson that assumes equal sectoral effects. Apart from approxmaton errors, ths s equvalent to regressng poverty changes on GDP growth rates. These results are presented n columns 2 and 4 for each of the samples, respectvely. In both 11
cases the growth elastcty of poverty s negatve, statstcally sgnfcant, and a lttle over 1 n magntude. Table 2 presents the results when GDP s further dsaggregated nto agrculture, servces, and ndustry s four major categores. We work wth both the full sample of countres and the reduced sample obtaned by applyng the Kraay (2005) crtera for elmnatng extreme observatons (see footnote 4). The results are smlar n both cases, so we dscuss only those usng the full sample. In the unconstraned regresson, only manufacturng growth carres a sgnfcantly negatve coeffcent, although agrculture growth also approaches a level of sgnfcant poverty allevaton effect. The pattern of sgns s dverse across sectors, wth agrculture, manufacturng, and constructon, carryng negatve coeffcents, whle mnng, utltes, and servces presentng postve ones. The relatvely large dsperson across countres makes t dffcult to learn much about dfferences n growth elastctes of poverty across sectors unless we restrct the model to be estmated. We can do ths by pullng together sectors that appear to have smlar effects on poverty. A frst approxmaton s to group together sectors that present negatve coeffcents n the unconstraned regresson, and do lkewse wth those that carry postve coeffcents. Before groupng them, we can test the equalty of ther coeffcents. These tests (shown at the bottom of Table 2, column 1) ndcate that agrculture, manufacturng, and constructon (the sectors carryng negatve coeffcents) can be pulled together, whle mnng, utltes, and servces (all carryng postve coeffcents) can form a sngle category. Applyng these restrctons, we can estmate the correspondng constraned regresson, whose results are presented n column 2. Growth n agrculture, manufacturng, and constructon now appear to have a clear, sgnfcant poverty reducng effect. In contrast, growth n mnng, utltes, and servces do not seem to reduce poverty (or worsen t for that matter), once growth n other sectors s controlled for. The test for the equalty of coeffcents n the constraned regresson confrms that the two groups (agrculture/manufacturng/constructon on one sde and mnng/utltes/servces on the other) have statstcally dfferent mpacts on poverty (see bottom of columns 2 and 4). 12
Poverty reducton and labor-ntensve growth Why would some sectors growth contrbute to poverty allevaton more than growth n others? There are a few potental explanatons. One s the relatonshp between the geographc locaton of a sector s producton and the ncdence of poverty n the area. Accordng to ths argument, agrcultural growth would have a large mpact on poverty allevaton because the poor are concentrated n rural areas. A second explanaton emphaszes market segmentaton, whch would prevent wage gans n one sector to be transmtted to the rest. Our theoretcal model formalzes a thrd explanaton accordng to whch a sector s labor ntensty determnes ts mpact on poverty reducton, even n the presence of free labor moblty. The basc result of our theoretcal model lnks wage ncreases to sectoral growth and s gven n equaton (9). The mult-sector verson of ths equaton can be wrtten as, I I ε 1 ˆ ω = s yˆ + ( l s ) yˆ (11) = 1 ε = 1 That s, wage grows proportonally to aggregate output (frst term) wth a premum (second term) f growng sectors are suffcently labor ntensve. Assumng that wage ncrease and poverty reducton are lnearly related, hˆ = θ0 + θω 1ˆ., then changes n poverty can be expressed as a functon of sectoral growth, Collectng terms, I I hˆ = θ 0 + θ s yˆ 1 + θ2 ( l s ) yˆ (12) = 1 = 1 I l hˆ = θ + + 0 θ1 θ 2 θ 2 s yˆ (13) = 1 s Ths expresson ndcates that a sector s growth effect on poverty reducton depends on ts labor ntensty, l s. To the extent that sectors dffer concernng ther labor ntenstes, ths explans why ther effects on poverty allevaton are not the same. Moreover, snce n prncple labor ntenstes can vary not only across sectors but also across countres for the same sector, then the sectoral growth elastctes of poverty reducton may be country specfc. Ths may explan n part why our sectoral regressons are so lackng n precson. 13
How dfferent s labor ntensty across sectors and across countres? And, s the pattern of sectoral growth elastctes of poverty consstent wth ther labor ntenstes? Fgure 1 presents box-plots for the cross-country dstrbuton of labor ntenstes ( l s ) correspondng to the sx sectors under examnaton. We notce that, frst, wth dfferent degrees, these sectors exhbt a remarkable dsperson across countres; and second, n spte of ths dsperson, t s possble to dentfy a rankng of labor ntenstes across sectors. Agrculture and constructon, followed by manufacturng, seem to be the most labor-ntensve sectors, havng all of them a medan l s rato larger than 1. The constructon sector s notceable for the large dsperson of ts cross-country dstrbuton of labor ntensty. Conversely, manufacturng shows a concentrated dstrbuton, partcularly regardng the nter-quartle range, whch may explan why ts coeffcent s estmated wth suffcent precson to acheve statstcal sgnfcance. Mnng and utltes, followed by servces, are the least labor-ntensve sectors, wth medan ratos below 1 n all cases. Mnng and utltes also show consderable dsperson across countres n ther labor ntensty, whle servces presents the most concentrated dstrbuton of the sx major sectors. The pattern of coeffcents on sectoral growth estmated above s consstent wth the noton that labor ntensty determnes a sector s nfluence on poverty allevaton. The sectors wth medan labor ntenstes greater than 1 --agrculture, constructon, and manufacturng-- carry negatve coeffcents; whle those wth medan labor ntenstes lower than 1 --mnng, utltes, and servces-- have postve coeffcents. Moreover, the rankng of labor ntenstes (n decreasng order) concdes exactly wth the rankng of sectoral coeffcents (from more to less negatve) estmated for the reduced sample and wth those estmated for the full sample wth only one excepton (mnng and servces swtch places). The consstency between labor ntenstes and the pattern of estmated sectoral growth coeffcents s suggestve, but a more formal test can be conducted on the bass of out theoretcal model. Equaton (12) can be wrtten as a regresson equaton of the change n poverty on aggregate and sectoral growth, l s 14
where, I lj h ˆ j = θ + θ yˆ j θ sj y 0 1 + 2 j + ε j 1 ˆ s (14) =1 j I yˆ s yˆ s (per capta) GDP growth. The coeffcent θ 1 ndcates the sze = 1 effect of growth on poverty reducton, whle θ 2 reveals ts composton effect. Negatve sgns are expected for both coeffcents f growth helps reduce poverty and f the labor ntensty of growng sectors has an addtonal mpact on poverty allevaton. In order to estmate equaton (14), t s crucal to obtan data on labor ntenstes by sector and country. As explaned above, we derve these data from nformaton on sectoral value added from World Bank (2005) and payments to unsklled workers from the Global Trade Analyss Project (GTAP). We focus on unsklled workers as they are lkely to best represent the poor n each country. Equaton (14) provdes a drect test of the model, and ths s our basc and preferred specfcaton. However, there are other possbltes. Frst, f we beleve that labor ntenstes are technologcal drven and common across countres, then we can use a sngle l s rato for each sector for all countres. Ths may be a good strategy f we are uncertan as to the qualty of the data on labor ntenstes per country. We mplement ths specfcaton by replacng the country-specfc labor ntenstes n equaton (16) by ther correspondng sample medan per sector. Second, a dscrete or categorcal verson of the test can be derved by assumng that sectoral growth can have ether a hgh or a low mpact on poverty reducton dependng on whether ts labor ntensty l s s, respectvely, above or below a certan threshold, whch we set equal to 1. Ths approach s useful f we are stll uncertan as to the precse measure of labor ntenstes but don t beleve that they are common across countres. We mplement ths specfcaton by allocatng sectors nto two groups accordng to ther labor ntensty, regressng poverty changes on the growth rates of hgh and low labor-ntensty groups, and then testng for the dfference between ther respectve coeffcents. Notce that the composton of these groups can vary from country to country. Table 3 presents the estmaton results for the drect regresson mpled by the model (column 2), the two alternatve specfcatons (columns 3 and 4), and a benchmark 15
regresson wth (per capta) GDP growth as sole explanatory varable (column 1). coeffcents on aggregate growth ( θ 1 ) are always sgnfcantly negatve, wth larger magntudes when labor ntensty s controlled for. Most relevant for our purposes, the coeffcent on labor-ntensty-weghted sectoral growth --or labor-ntensve growth, for short-- ( θ 2 ) s also negatve and hghly statstcally sgnfcant n our preferred specfcaton (column 2). Interestngly, the regresson ft ncreases consderably (from 15 to 28%) once nformaton on labor ntensty s added to that on aggregate growth. Fgure 2 shows a partal-regresson plot lnkng the change n poverty and labor-ntensve growth; t confrms a negatve pattern that s well establshed by most observatons n the sample (we consder the ssue of outlers below.) Thus, t appears that n addton to the sze of growth, the composton of growth regardng ts labor ntensty s statstcally and economcally relevant for explanng poverty reducton. The coeffcent on labor-ntensve growth s also negatve and statstcally sgnfcant when we use medans per sector across countres to measure labor ntensty (column 3). However, the ft of the regresson declnes somewhat, revealng that country-specfc data on labor ntenstes contrbute useful nformaton for growth composton to explan poverty changes. A smlar message s obtaned from the alternatve specfcaton based on groupng sectors by labor ntensty (column 4). The coeffcent on growth n hgh labor-ntensty sectors s negatve and statstcally sgnfcant, whle that on growth n low labor-ntensty sectors s much smaller and not sgnfcant. In fact, the null hypothess that these two coeffcents are the same s rejected wth a p-value of 0.07. The R-squared n ths case falls consderably wth respect to the preferred case, confrmng that the precse numercal values on country-specfc labor ntenstes provde relevant nformaton that cannot be captured by categorcal ndcators. The Robustness to outlers and extreme observatons Table 4 presents the results related to the analyss of robustness to oulers, wth our basc regresson repeated n column 1 for comparson purposes. The data on labor ntensty, l s, present a few extreme values that are lkely to represent ether measurement error or rare crcumstances; n order to avod ther undue nfluence, n our basc specfcaton, we truncated the cross-country dstrbuton of labor ntensty per 16
sector to values rangng from 5 to 95 percentle of the orgnal dstrbuton. Column 2 presents the regresson results when these extreme values are not truncated. The coeffcent on labor-ntensve growth contnues to be statstcally negatve, although ts level of sgnfcance and the regresson ft dmnsh a lttle. Inspecton of Fgure 2 may rase questons as to the nfluence of some countres n our basc results. To dspel these doubts, we run the regresson usng a procedure that weghs observatons accordng to how they ft the pattern establshed by the rest. Ths s the robust regresson presented n column 3. We also run the regresson completely excludng possble outlers, dentfed as the countres that receve weghts below 0.7 of a maxmum of 1 n the robust procedure. These countres are Argentna, Estona, Latva, and Senegal; and the correspondng results are shown n column 4. In both cases, the coeffcent of nterest remans negatve and hghly statstcally sgnfcant, wth a magntude that s almost the same as that n the benchmark case. It s reassurng that the regresson ft ncreases consderably when the outlers are excluded. As mentoned above, the way we calculate poverty changes (that s, proportonal wth respect to the average poverty level n the spell) produces fewer extreme values than the standard way of takng log dfferences. Ths allows us to keep a larger number of observatons n the sample than would be the case f we appled the crtera n Kraay (2005). We check whether our basc results stll hold n ths reduced sample (of 32 countres), and the results are presented n column 5. The sgn, sgnfcance, and even magntude of the coeffcent on labor-ntensve growth are remarkably smlar as those usng the full sample, wth a slght gan n regresson ft. All n all, the results n Table 4 allow us to conclude that our basc results are robust to the possble presence of outler or extreme observatons. Alternatve explanatons It may be argued that the mportance of the growth composton term s due to ts correlaton wth other varables that affect poverty changes. The results n Table 5 check for ths possblty by allowng for alternatve explanatons n turn. Frst s the ssue of agrcultural growth. Gven that agrculture s the sector wth the hghest labor ntensty n most countres, t may be argued that our growth composton varable s just capturng 17
the presence of agrculture, whch may affect poverty reducton for reasons unrelated to labor ntensty. We examne ths possblty by addng (sze-adjusted) agrculture value added growth as an ndependent explanatory varable to our basc specfcaton (see column 1). Whle the coeffcent on agrcultural growth s negatve but not statstcally sgnfcant, the coeffcent on labor-ntensve growth retans ts sgn, sgnfcance, and magntude wth respect to our basc specfcaton. Ths suggests that the mportance of agrcultural growth n poverty reducton that has been recognzed n the lterature s mostly due to ts ntensve use of unsklled labor. Most sgnfcantly, the mportance of labor ntensty n growth s ablty to reduce poverty appears to be relevant across all sectors. Second s the connecton wth nequalty. A promnent explanaton n the lterature as to the dfferng effect of ncome growth on poverty reducton s that hgher nequalty dampens the benefcal mpact of growth (see Ravallon 2004 for references). If more unequal countres have growth based aganst labor-ntensve sectors --because, for nstance, nequalty nduces polces that make labor markets more rgd--, then excludng nequalty from our analyss could be basng the results n our favor. To account for ths possblty, we control for nequalty by addng the Gn coeffcent as an ndependent explanatory varable (column 2) and by nteractng t wth both the growth sze and composton terms (column 3). Ths also captures possble non-lneartes n the relaton between wage growth and poverty reducton. In both cases, the growth composton term remans negatve and sgnfcant, as n our preferred specfcaton. The sgnfcance of GDP growth per se suffers when the nteractons wth the Gn coeffcent are added gven ts hgh collnearty wth the nteracton term. Thrd s the ssue of measurement error due to the dscrepancy between natonal accounts and household surveys on data for mean ncome growth. As mentoned above, poverty measures are constructed from household survey nformaton; and n most studes connectng poverty and mean ncome growth, the same source s used for both varables. We, however, do t otherwse: snce our focus s on producton and ts composton, we had to use data from natonal accounts for the explanatory varables. It s well-known that ncome growth derved from household survey data shows large and sometmes systematc dfferences wth that obtaned from natonal accounts (see Deaton, 18
2003). If these dfferences are correlated wth labor ntensty n the country, the coeffcent on growth composton may be based. Moreover, the bas could be n our favor f natonal accounts underreported producton from unsklled workers. To account for ths possblty, we nclude mean ncome growth from household surveys as an addtonal explanatory varable (see column 4). As expected, ths varable carres a negatve and sgnfcant coeffcent, and ts ncluson produces both an mprovement n the regresson ft and a declne n the magntude of the coeffcents on the sze and composton of growth. However, both coeffcents reman negatve and statcally sgnfcant, confrmng our hypothess. Fourth s the ssue of growth endogenety. Our analyss has been conducted n dfferences n order to control for country-specfc structural factors that affect poverty and producton jontly. Stll, t can be argued that mprovements n poverty drve producton growth --possbly through hgher rates of accumulaton of human captal and savngs-- thus makng the analyss n dfferences also subject to the endogenety crtque. Although ths does not apply to the varable on the composton of growth, ts coeffcent may stll be based f composton and sze of growth are correlated. To control for the potental endogenety of growth, we nstrument for t usng the average GDP growth of the country s tradng partners as the source of exogenous varaton. The nstrumental varable procedure (whose results are presented n column 5) renders coeffcents on the sze and labor-ntensty of growth that reman negatve and hghly sgnfcant. Moreover, ther magntudes are even larger that n the benchmark case, ndcatng that the endogenety of growth, f any, was playng aganst the hypothess advocated n the paper. Alternatve poverty measures Our analyss has used the headcount poverty ndex as the benchmark measure of poverty gven ts promnence n both the emprcal lterature and polcy crcles. However, our smple theoretcal model bulds the case for the mportance of the composton of growth by focusng on ts relatonshp not wth poverty drectly but wth labor wages. The connecton wth poverty s made by assumng that wages affect poverty accordng to a lnear functon, whch combned wth the basc result of the model brngs about the paper s man regresson equaton. Yet, the lnearty of the relatonshp 19
between wages and the headcount poverty ndex may be called nto queston by consderng that the elastcty of ths measure to margnal changes n ncome s nl except around the poverty lne. (Naturally, the justfcaton for the lnear assumpton n the paper s that we are dealng wth more than margnal changes n ncome/producton.) To dspel these doubts, we use other poverty measures that are more closely related to wages and that respond to changes n ncome over a wder range of the ncome dstrbuton. Table 6 shows the results when alternatve poverty measures are used to construct the dependent varable; these measures are the average poverty gap, the average squared poverty gap, and the Watt s poverty ndex (columns 1-3, respectvely). In all cases, the sze of GDP growth and --most mportantly for our purposes-- ts labor ntensty carry negatve and qute sgnfcant coeffcents. The regresson ft does not mprove when we use these alternatve poverty measures nstead of the standard headcount ndex; actually, the R 2 s one-thrd lower when usng the smple poverty gap. Fnally, n column 4 we examne to what extent our benchmark result comes from the connecton between laborntensve GDP growth and mprovements n the ncomes of the poor. For ths purpose, we add the growth of the average poverty gap as an addtonal explanatory varable. We fnd that the growth composton term retans ts sgn and sgnfcance, but ts sze shrnks to less than one thrd than n the benchmark. Ths ndcates that, at least partally, laborntensve growth affects the headcount poverty ndex through the ncomes of the poor. The mechansm: dstrbuton or mean component of poverty changes? The last ssue we examne s the mechansm through whch the labor ntensty of GDP growth matters for poverty allevaton. In partcular, does t affect the dstrbuton or the mean component of poverty changes? To answer ths queston we mplement the decomposton ntroduced by Datt and Ravallon (1992), accordng to whch changes n poverty can be broken down nto the porton due to changes n mean ncome holdng ncome dstrbuton constant (.e., unchanged Lorenz curve), the porton due to changes n the dstrbuton of ncome holdng constant ts mean, and an approxmaton resdual. Then, we estmate the respectve effects of the sze and the labor-ntensty of GDP growth on each of these components, applyng the restrcton that the combned effect of each explanatory varable must be the same as ts correspondng effect on the overall 20
poverty change (whch s gven by the benchmark regresson). We mplement ths estmaton through a constraned Seemngly-Unrelated-Regresson-Equaton procedure (SURE). The results are presented n Table 7. When SURE estmaton gnores the nfluence of outlers (columns 1 and 2), we fnd that labor-ntensve growth affects poverty changes exclusvely through ther mean component. When we control for the nfluence of outlers (Cols. 3 and 4), labor-ntensve growth stll affects sgnfcantly the mean component, but now t appears to also affect the dstrbuton component though less strongly and sgnfcantly. The strength of the mean-ncome channel relatve to the dstrbuton channel ndcates that labor-ntensve growth should not be assocated wth zero-sum ncome changes across households. It s not that labor-ntensve growth s poverty reducng manly because t mples redstrbuton from rch to poor. Although labor-ntensve growth mproves the relatve standng of the poor, ts man effect on poverty s through ts benefcal mpact on ther absolute ncome. IV. Concludng Remarks The frst concern that developng countres face n ther objectve to reduce poverty s the lack of suffcent economc growth. Ths s justfably so gven that no lastng poverty allevaton has occurred n the absence of sustaned producton growth. However, growth s sheer sze does not appear to be a suffcent condton for profound poverty reducton. In fact, a complant often heard n countres around the world s that the poverty response to growth s sometmes dsappontng. A general argument for the reslence of poverty reles on ether the lack of opportuntes presented to the poor or ther nablty to take advantage of them. If the poor are malnourshed, are uneducated, lve n remote areas, or are dscrmnated aganst, the gans of economc growth are lkely to escape them. Ths paper offers a complementary perspectve supportng the general argument on the lack of opportuntes. In a nutshell, the paper argues that not only the sze of economc growth matters for poverty allevaton but also ts composton n terms of ntensve use of unsklled labor, the knd of nput that the poor can offer to the producton process. 21
The paper frst llustrates the connecton between wage expanson (poverty reducton), labor ntensty, and sectoral growth through a mult-sector theoretcal model. Then, consderng the model s nsghts, t conducts a set of cross-country emprcal exercses on poverty changes as the dependent varable. The paper fnds that the mpact of growth on poverty reducton vares from sector to sector and that there s a systematc pattern to ths varaton. Sectors that are more labor ntensve (n relaton to ther sze) tend to have stronger effects on poverty allevaton. Thus, agrculture s the most poverty-reducng sector, followed by constructon, and manufacturng; whle mnng, utltes, and servces by themselves do not seem to help poverty reducton. After ths sectoral-drven emprcal analyss, the paper conducts a more drect test of the model by consderng poverty reducton a functon of not only aggregate growth (whch would represent growth s sze effect) but also a measure of labor-ntensve growth (whch would represent ts composton effect). The results confrm that poverty allevaton ndeed depends on the sze of growth. Moreover, they also ndcate that poverty reducton s stronger when growth has a labor-ntensve nclnaton. Ths central result of the paper s robust to the nfluence of outler and extreme observatons, holds true for varous poverty measures (such as the headcount ndex, the average poverty gap, and the Watt s ndex), and s not drven away by alternatve explanatons --such as the mportance of agrcultural growth n reducng rural poverty, the role of nequalty n dampenng the benefcal mpact of growth, and the statstcal dscrepancy between household surveys and natonal accounts. Fnally, analyss on the mechansms through whch labor-ntensve growth reduces poverty allows us to conclude that ths postve effect does not requre or mply redstrbuton from rch to poor. Although laborntensve growth mproves the relatve standng of the poor, ts man effect on poverty s gven by ts benefcal mpact on ther absolute ncome. From a postve perspectve, these results may help understand the consderable dsparty n the poverty reacton to economc growth and, n partcular, why n some crcumstances poverty s rresponsve to producton mprovements. Ths would be the case of, for nstance, a country experencng a mnng or ol boom that s unaccompaned by growth n other sectors. From a normatve perspectve, ths study does not provde grounds for ndustral (or selectve) polces as t does not deal wth the sources of 22