WPS4077
THE COMPOSITION OF GROWTH MATTERS
FOR POVERTY ALLEVIATION*
Norman Loayza Claudio Raddatz
The World Bank The World Bank
Abstract
This paper contributes to explain the cross-country heterogeneity of the poverty response
to changes in economic growth. It does so by focusing on the structure of output growth
itself. The paper presents a two-sector theoretical model that clarifies the mechanism
through which the sectoral composition of growth and associated labor intensity can
affect workers' wages and, thus, poverty alleviation. Then, it presents cross-country
empirical evidence that analyzes, first, the differential poverty-reducing impact of
sectoral growth at various levels of disaggregation, and, second, the role of unskilled
labor intensity in such differential impact. The paper finds evidence that not only the size
of economic growth but also its composition matters for poverty alleviation, with the
largest contributions from labor-intensive sectors (such as agriculture, construction, and
manufacturing). The results are robust to the influence of outliers, alternative
explanations, and various poverty measures.
Keywords: Poverty, economic growth, production structure, labor intensity.
World Bank Policy Research Working Paper 4077, December 2006
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the
exchange of ideas about development issues. An objective of the series is to get the findings out quickly,
even if the presentations are less than fully polished. The papers carry the names of the authors and should
be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely
those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors,
or the countries they represent. Policy Research Working Papers are available online at
http://econ.worldbank.org.
*We are grateful to Yvonne Chen and Siyan Chen for research assistance and to Koichi Kume for editorial
assistance. For useful comments, suggestions, and/or data we are indebted to Maros Ivanic, Francisco
Ferreira, Oded Galor, Aart Kraay, Humberto López, Ross Levine, Martin Ravallion, Yona Rubinstein, Luis
Servén, and seminar participants at Brown University and the World Bank. Support from the Chief
Economist Office of the Latin America and Caribbean Region of the World Bank is gratefully
acknowledged.
THE COMPOSITION OF GROWTH MATTERS FOR POVERTY ALLEVIATION
I. Introduction
There is little doubt that economic growth contributes significantly to poverty
alleviation. The evidence is mounting and coming from various sources: cross-country
analyses (Besley and Burgess, 2003; Dollar and Kraay, 2005; Kraay, 2005; and López,
2004), cross-regional and time-series comparisons (Ravallion and Chen, 2004; Ravallion
and Datt, 2002), and the evaluation of poverty evolution using household data (Bibi,
2005; Contreras, 2001; Menezes-Filho and Vasconcellos, 2004). At the same time, it is
clear that the effect of economic growth on poverty reduction is not always the same. In
fact, most studies point to considerable heterogeneity in the poverty-growth relationship,
and understanding the sources of this divergence is a growing area of investigation
(Bourguignon, 2003; Kakwani, Khandker, and Son, 2004; Lucas and Timmer, 2005,
Ravallion, 2004). Most of the received literature focuses on socio-economic conditions
of the population as determinants of the relationship between growth and poverty
reduction. Thus, wealth and income inequality, literacy rates, urbanization levels, and
morbidity and mortality rates, among others, have been found to influence the degree to
which output growth helps reduce poverty.
In this paper we take a different, albeit complementary, perspective on the sources
of heterogeneity in the poverty-growth relationship. We focus on the characteristics of
output growth itself, rather than the demographic, social, or economic conditions of the
population. We study how the production structure of the economy and, specifically, the
sectoral composition of growth affect its capacity to reduce poverty. Our conjecture is
that growth in certain sectors is more poverty reducing than growth in others and that a
sector's poverty-reducing capacity is related to its intensity in the employment of
unskilled labor.
There are important studies that precede and motivate our work. Thorbecke and
Jung (1996) develop a social-accounting method to estimate the impact of various
production activities on poverty reduction. The method requires knowledge of complex
elasticities connecting the distribution of households with eight employment and
production sectors. The authors apply the method to Indonesia in the 1980s and find that
1
agricultural and service sectors contribute more to poverty reduction that industrial
sectors do. Khan (1999) applies the same methodology to study sectoral growth and
poverty alleviation in South Africa. He finds that higher contributions are derived from
growth in agriculture, services, and some manufacturing sectors.
A different approach consists of conducting reduced-form analysis on time-series
data for individual countries. This is the approach taken by Ravallion and Datt (1996) to
study the evolution of poverty in India during 1951-91. Linking poverty changes to
value-added growth rates in the three major sectors of economic activity, they find that
growth in agriculture and services helped reduce poverty in both urban and rural areas
whereas industrial growth did not reduce poverty in either. Applying a similar
methodology for the case of China over 19802001, Ravallion and Chen (2004) find that
growth in agriculture emerges as far more important than growth in secondary or tertiary
sectors for the purpose of poverty alleviation.
Our work adds to this literature along four dimensions. First, we present a two-
sector theoretical model that clarifies the mechanism through which the sectoral
composition of growth and associated labor intensity can affect workers' wages --and,
thus, poverty alleviation-- even in the absence of market segmentation. Second, we use
cross-country evidence --with the pros and cons associated with increasing the underlying
variation of the data-- allowing us to relate our results to the empirical macroeconomic
literatures on growth and poverty. Third, we employ a level of disaggregation that
explores the diversity within the industrial sector, hoping to shed light on why it appears
to be less pro poor than agriculture or services. And, fourth, we explicitly consider
sectoral employment intensity as the mechanism through which the pattern of growth
matters for poverty alleviation.
The plan of the paper is the following. Section II presents a theoretical model that
formalizes our initial conjecture. It examines the wage (poverty) effect of output growth
in a two-sector economy, where capital and labor are freely mobile and the sectors'
technologies vary according to their labor intensity. Section III presents cross-country
empirical evidence that analyzes, first, the differential poverty-reducing impact of
sectoral growth at various levels of disaggregation, and, second, the role of unskilled
labor intensity in such differential impact. Also in this section, we subject our basic
2
result to a comprehensive set of robustness checks that account for the influence of
outlier and extreme observations, for potential alternative explanations, and for various
poverty measures. Section IV offers some concluding remarks.
II. The Model
We now present a two-sector model with asymmetric technologies to help us
understand the relation between sectoral growth and poverty alleviation. We focus on the
two-sector case for simplicity, but the results are analogous for the n-sector case.
The economy is populated by two types of individuals: poor and rich. Both types
are endowed with n units of labor, derive utility from the consumption of a final good,
and have the same discount factor and instantaneous utility function u(c)=log(c).
However, only rich individuals have access to an asset a that allows them to transfer
wealth across periods. This setting implies that the income and consumption of poor
individuals depend only on the real wage rate. Thus we assume that the rate of poverty
reduction is related only to the growth rate of real wages. Although this is an extreme
assumption, it simplifies considerably the analysis and is roughly consistent with the low
saving rates observed both in poor countries and poor households within a country.1
The final good, y, is produced by a perfectly competitive firm using a constant-
returns-to-scale technology and two intermediate goods, y1 and y2, as inputs according to
1
y = y1 + y2
( ) ,
The final good can be used not only for consumption but also as capital in the
production of the intermediate goods. Each intermediate good is produced by a perfectly
competitive firm according to the following technology with labor-augmenting
technological progress,
1Schmidt-Hebbel and Serven (1999) show that saving rates increase with income across countries. In poor
countries, the saving rates are below 10%. Attanasio and Székely (1998) provide evidence on households'
saving rates at different levels of the income distribution in Mexico. Their data show that saving rates
increase strongly with income and display even negative values up to the 25th percentile of the household
income distribution.
3
yi = ki(1-i )(An ) , i =1,2
i
i i
where ki and ni are sector i's capital and labor, respectively, and Ai captures the level of
technology, which evolves exogenously according to Ai = exp(git). We assume that
capital is perfectly mobile across sectors and does not depreciate.
Intersectoral allocation
In what follows we use this setup to derive the relation between the composition
of growth and the evolution of the real wage rate, which given our assumptions maps into
the income and consumption of the poor. We will only focus on the aspects of the model
that are relevant for the derivation of this expression and omit several side aspects of the
characterization.
Under perfect competition, the price charged by the final-good firm, p , equals its
unit cost of production. Then,
1
p = p1
( 1- + p2 1- )1- ,
where = (1- )-1. Solving the optimization problem of the final-good firm and setting
the price of the final good as a numeraire, we obtain the following first order conditions,
-1
pi yi
= si = , i =1,2 (1)
Y yi
Y
which characterize the share of the final good production value that goes to each
intermediate sector. Given that the production of the final good exhibits constant returns
to scale these shares add up to one.
Combining the first order conditions, we obtain the following expression for the
demand of intermediate goods,
y1 = , p2
(2)
y2 p1
4
which shows that corresponds to the (constant) elasticity of substitution between the
intermediate goods. Under perfect competition, each intermediate-good firm determines
its demand for labor and capital taking factor and output prices as given. Then, the first-
order conditions corresponding to the intermediate-good firm are given by
yi = ni pi(1-i) ,i
rki
(3)
pii = =1,2,
Equations (2) and (3)--which correspond to the standard conditions for static
efficiency-- plus the conditions of factor market equilibrium --k1 + k2 = k, and
n1 + n2 = n -- determine the allocation of labor and capital across sectors at every moment.
Although in principle we could use the previous equations to determine the
relative prices of the intermediate goods p1 / p2 as a function of the aggregate capital-
labor ratio k / n, technological parameters, and sector productivities Ai , this problem
cannot be solved in closed form except in some special cases that restrict the values of
and the i (see Miyagiwa and Papageorgiu, 2005, for a discussion). Nevertheless, we
can use these equations to characterize the evolution of real labor income, which is the
object of interest for our empirical analysis.
The evolution of real labor income
According to the first-order conditions for intermediate-good firms, and focusing
without loss of generality on intermediate good 1, the rate of change of the real wage can
be written as
^ = p^1 + y^1 -n^1 (4)
where the hat denotes the rate of change of a variable ( x^ = dx / x ). The first two terms of
this expression correspond to the evolution of the value of sector 1 output in terms of the
final good ( p1y1). From equation (1) this corresponds to
s^1 +Y^ = -1 y^1 + (s1y^1 + s2 y^2 )
1
(5)
where we have used the fact that Y^ = s1y^1 + s2y^2 because of constant returns to scale.
5
The last term in equation (4) is the evolution of employment in sector 1. The first
order conditions of intermediate good firms with respect to labor presented in equation
(3) together with equation (2) imply that
-1
1 nn1 y1
2 2 =1, (6)
y2
which, after log-differencing and using the labor market clearing condition n = n1 + n2 ,
results in the following expression for the growth rate of employment in sector 1,
n^1 = l2 -1
(7)
(y^1 - y^2)+n^,
where l2 is the share of employment in sector 2 ( n2 / n ).
Finally, putting together equations (5) and (7) and re-arranging terms we obtain
the growth rate of the real wage rate,
^ = siy^i +
2
i=1 -1 2
(8)
(l i-si)y^i
i=1
where, in a slight abuse of notation, y^i now represent the growth rates in per-capita
terms.
This equation indicates that the growth of real labor income is driven by two
components. The first one, corresponding to the first term on the left-hand side of
equation (8), is the growth of per-capita GDP. An increase in per-capita GDP
corresponds to a higher output per worker that maps into higher wages. The contribution
of a sector's growth to this term depends exclusively on its size, as captured by its share
on final-good output, si . The second component captures the reallocation effects. The
impact of a sector's growth on this component depends on the elasticity of substitution
across sectors in the production of the final good ( ) and on a sector's labor intensity, as
captured by the difference between its labor share of total employment, li , and its share
in total output si . Starting from equations (1) and (3), it can be shown that this difference
corresponds to
6
li - si = 1 - 1 , i =1,2 (9)
1+ -i si
i
s-i 1+ s-i
si
which indicates that the difference li-si is higher for sectors with a higher share of labor in
total output, i . This means that growth in a labor intensive sector will have an additional
effect on wages beyond its impact on aggregate growth, as long as the elasticity of
substitution is sufficiently high (specifically above 1, according to eq. (8)).2
The elasticity of substitution is relevant because it determines whether (and by
how much) labor will move into or out of a growing sector: the higher the elasticity of
substitution, the more labor moves into that sector. If the elasticity is too low (below 1)
labor actually moves out of an expanding sector; however, as the elasticity increases and
surpasses a threshold value (equal to 1), labor starts to flow into the growing sector. With
a high (low) elasticity of substitution, the price adjustment required by an increase in the
relative output of a sector is small (large) so that labor needs to move into (out of) the
expanding sector to achieve wage equalization (this can be clearly seen in eq. (6)).
Equation (8) also shows that there are two cases in which the growth rate of real
labor income depends only on GDP growth: (i) when the technologies of the intermediate
sectors are identical (1 = 2 ), and (ii) when the elasticity of substitution is equal to one
(the Cobb-Douglas case). The first case is trivial: if there are no asymmetries across
sectors, uneven growth is irrelevant. In the second case, under a Cobb-Douglas
production function, sectoral labor shares are constant, and any adjustment in relative
quantities results only in a corresponding change in relative prices. Uneven sectoral
growth, not requiring labor reallocation across sectors, would not affect real wages.
Thus, omitting the composition of growth as a determinant of real wage increases and
poverty alleviation is equivalent to assuming that either sectors do not differ in their labor
intensities or their elasticity of substitution is equal to one.
Although not explicitly derived in the model, it should be noted that the presence
of technological progress is important for the long-run implications of the model (that is,
2Consider the following example. Suppose that sector 1 is more labor intensive than sector 2 (1/2>1), so
that l1-s1>0, and that it experiences an exogenous increase in productivity. If the elasticity of substitution is
sufficiently high, labor will move into sector 1 where it is relatively more productive, pushing the wage rate
up. The opposite will happen if the elasticity of substitution is relatively low (below 1).
7
beyond transitional dynamics). If the model exhibits a balanced-growth path, the growth
rate of each sector y^i and of the economy will be exclusively determined by the growth
rates of productivity in all the different sectors of the economy (the gi s). Characterizing
the balanced growth path of the model is beyond the scope of this paper; nevertheless,
equation (8) is valid both during transitional dynamics and in balanced growth.
Our assumption that poverty changes are only a function of the growth rate of real
labor income corresponds to assuming that h^ = (^) , where h^ is the growth rate of
poverty. In the empirical section of the paper, we will estimate the parameters of the
linearized version of this relation h^ = 0 +1^ as our benchmark case, but we will also
consider some non-linear relations in our robustness analysis.
III. Empirical Evidence
Our empirical analysis consists of two related sections. In the first, we address
the connection between the pattern of growth and poverty alleviation by disaggregating
growth into its sectoral components and examining their corresponding effects on
poverty. This is the traditional approach, and, thus, it allows us to place our analysis in
the context of the received literature. The second empirical section modifies the sectoral
analysis by introducing labor intensity as the source of the differential impact of sectoral
growth on poverty reduction. This approach is derived from the theoretical model and,
thus, establishes the link between theory and empirics in the paper.
Data and sample
Our sample consists of a cross-section of developing countries with comparable
measures of poverty changes, disaggregated value-added growth rates at 3- and 6-sector
levels, and unskilled employment at the same levels of disaggregation. In practice, our
dataset is the result of combining the Kraay (2005) database on poverty spells,3 World
3 The Kraay database results from processing income distribution data for a large number of developing
countries. In turn, its source is the collection of household survey data estimated from primary sources and
made comparable across countries by Martin Ravallion and Shaohua Chen at the World Bank. For details,
see Kraay (2005).
8
Bank (2005) data on sectoral value added,4 and Purdue University's Global Trade
Analysis Project database (GTAP, 2005) on labor shares.
We focus on changes occurring over long horizons, where the poverty reduction-
economic growth relationship is most stable. For this reason we use only one spell per
country, where the duration of the spell corresponds to the longest period for which initial
and final poverty data exist for the country. The rest of the variables (e.g., value added
growth rates and labor ratios) are calculated over the corresponding period per country.
The dependent variable is the proportional change in poverty over a period of
time (spell) per country. Specifically, this is the annualized change in poverty as
proportion to average poverty over the period.5 Given its importance in the literature, the
benchmark poverty measure in the paper is the headcount poverty index, defined as the
fraction of the population with income below a given poverty line. In robustness
exercises, however, we use alternative measures of poverty, comprising other members of
the Foster-Greer-Thorbecke class of measures (the poverty gap and the squared poverty
gap) and the Watts index. Following convention for cross-country comparability, the
poverty line is set to $1 per person per day, converted into local currency using a
purchasing-power-parity adjusted exchange rate.
Regarding the explanatory variables, we work with growth rates of sectoral value-
added and employment data at two levels of disaggregation. The first is the traditional
sectoral division of agriculture, industry, and services. The second one disaggregates
industry further into mining, manufacturing, utilities, and construction. Sectoral growth
rates are calculated directly from data on sectoral value added as annualized log changes
of per capita value added between the end and start of the corresponding spell.
Employment data is calculated indirectly from data on sectoral value added and payments
4 The World Bank (2005) data on sectoral value added is complemented with statistics from the Inter-
American Development Bank and the United Nations.
1
5That is, proportional poverty change = * PF - PI , where P represents the poverty measure; T, the
T (PF + PI ) / 2
length of the spell; and the subscripts I and F, initial and final, respectively. Calculating the proportional
change with respect to the average measure allows us to avoid abnormally large proportional changes when
very low initial and/or final measures are present, as would be the case if log differences were used. Kraay
(2005) uses the latter procedure and then is forced to drop a considerable number of observations. Were we
to use Kraay's method, we would be working with 32 country observations, rather than 51, the sample size
of our benchmark regression.
9
to unskilled workers. Under the assumption of wage equalization, the ratio of unskilled
workers in a sector to total unskilled workers in the country is calculated as the ratio of
payments to unskilled workers in the sector to total payments to unskilled workers in the
economy. Regarding data for this calculation, only one observation per country or per
similar countries is available from the original source (GTAP).6
The resulting sample consists of 55 countries for 3-sector data and 51 countries
for 6-sector data. Appendix 1 provides the list of countries included in the sample, as
well as the initial and final years of their corresponding spell. Appendix 2 provides
definitions and sources for all variables used in our empirical exercises, and Appendix 3
presents basic summary statistics on the 51-country sample.
Poverty reduction and sectoral growth
We are interested in estimating the effect of sectoral growth on poverty reduction.
The regression equation can then be written as,
h^j = 0 + i sij y^ij +
I
(10)
j
i=1
where h^ is the annualized rate of change of the headcount poverty index, y^ is the
annualized rate of change of sectoral value added, s is the sectoral value added share in
GDP, and the subscript i and j represent sector and country, respectively. All growth
rates are expressed in per capita terms, and the sector shares are calculated from constant-
price magnitudes.7 The set I consists of three or six sectors, depending on whether
industry is considered as a whole or disaggregated into its four major categories. In
principle, it may be possible to estimate the poverty effect of output changes in a levels
6Given that, in most cases, the date of this observation differs significantly from the years of our poverty
spell, we first use the GTAP data to compute the ratio of payments to unskilled workers to a sector's value
added and assume it to be constant over time, for a given sector and country. We then use this ratio and the
sector's share in total value added during our spell (from World Bank, 2005) to compute the corresponding
ratio of unskilled workers in the sector to total unskilled workers in the country. Under wage equalization,
the ratio of unskilled workers in a sector to total unskilled workers can be written as
lk l = ksk s i i , where is the ratio of unskilled labor payments to sector's k value added, and s
k k
i
is the share of sector k in total value added.
7Calculating the shares from nominal magnitudes would more closely approximate the theoretical model.
However, we work with constant-price shares because, first, their resulting country coverage is larger than
when using current-price shares, and, second, they are very similar and render basically the same
econometric results.
10
regression. However, the literature advices a regression in differences to control for fixed
effects that may be driving both poverty and output, such as a host of country-specific
development-related variables in our cross-country setting.
Our regression specification weights sectoral growth by its relative size. As
Ravallion and Chen (2004) point out, this specification has the advantage that it allows
for a simple test of whether the growth composition matters: If the null hypotheses that
the coefficients i are equal to each other cannot be rejected, then the sectoral regression
collapses to one where GDP growth is the only relevant explanatory variable. In this
case, only size and not composition of growth would matter for poverty alleviation. Our
regression specification also allows for testing whether these sectors can be grouped in
different categories, not according to their output characteristics but according to their
relationship with poverty reduction. This will become important when we study the case
of six-sector disaggregation.
Table 1 presents the results when GDP is decomposed into agriculture, industry,
and services. The regressions are conducted using both the full sample of 55 countries
and the subset of 51 countries for which six-sector data are available. The latter exercise
is conducted with the purpose of comparison with the six-sector analysis. In both
samples (columns 1 and 3, respectively), the size-adjusted value-added growth rates of all
sectors fail to carry statistically significant coefficients. Moreover, the hypothesis that
the coefficients are the same cannot be rejected.
The lack of individual significance of sectoral growth rates and the inability to
separate their effects indicates that the three major sectors are highly linked in their
relationship with poverty reduction. This may be interpreted as evidence against the
importance of growth composition for poverty alleviation, but it may also be the result of
working with insufficiently disaggregated output categories. We examine the latter
possibility below when we analyze the six-sector case. Before doing that, however, we
can take the failure to reject the equality of coefficients at face value and estimate a
constrained regression that assumes equal sectoral effects. Apart from approximation
errors, this is equivalent to regressing poverty changes on GDP growth rates. These
results are presented in columns 2 and 4 for each of the samples, respectively. In both
11
cases the growth elasticity of poverty is negative, statistically significant, and a little over
1 in magnitude.
Table 2 presents the results when GDP is further disaggregated into agriculture,
services, and industry's four major categories. We work with both the full sample of
countries and the reduced sample obtained by applying the Kraay (2005) criteria for
eliminating extreme observations (see footnote 4). The results are similar in both cases,
so we discuss only those using the full sample. In the unconstrained regression, only
manufacturing growth carries a significantly negative coefficient, although agriculture
growth also approaches a level of significant poverty alleviation effect. The pattern of
signs is diverse across sectors, with agriculture, manufacturing, and construction,
carrying negative coefficients, while mining, utilities, and services presenting positive
ones.
The relatively large dispersion across countries makes it difficult to learn much
about differences in growth elasticities of poverty across sectors unless we restrict the
model to be estimated. We can do this by pulling together sectors that appear to have
similar effects on poverty. A first approximation is to group together sectors that present
negative coefficients in the unconstrained regression, and do likewise with those that
carry positive coefficients. Before grouping them, we can test the equality of their
coefficients. These tests (shown at the bottom of Table 2, column 1) indicate that
agriculture, manufacturing, and construction (the sectors carrying negative coefficients)
can be pulled together, while mining, utilities, and services (all carrying positive
coefficients) can form a single category.
Applying these restrictions, we can estimate the corresponding constrained
regression, whose results are presented in column 2. Growth in agriculture,
manufacturing, and construction now appear to have a clear, significant poverty reducing
effect. In contrast, growth in mining, utilities, and services do not seem to reduce poverty
(or worsen it for that matter), once growth in other sectors is controlled for. The test for
the equality of coefficients in the constrained regression confirms that the two groups
(agriculture/manufacturing/construction on one side and mining/utilities/services on the
other) have statistically different impacts on poverty (see bottom of columns 2 and 4).
12
Poverty reduction and labor-intensive growth
Why would some sectors' growth contribute to poverty alleviation more than
growth in others? There are a few potential explanations. One is the relationship
between the geographic location of a sector's production and the incidence of poverty in
the area. According to this argument, agricultural growth would have a large impact on
poverty alleviation because the poor are concentrated in rural areas. A second
explanation emphasizes market segmentation, which would prevent wage gains in one
sector to be transmitted to the rest. Our theoretical model formalizes a third explanation
according to which a sector's labor intensity determines its impact on poverty reduction,
even in the presence of free labor mobility.
The basic result of our theoretical model links wage increases to sectoral growth
and is given in equation (9). The multi-sector version of this equation can be written as,
^ = si y^i +
I (lI
-si) y^i
(11)
i=1 -1 i
i=1
That is, wage grows proportionally to aggregate output (first term) with a premium
(second term) if growing sectors are sufficiently labor intensive. Assuming that wage
increase and poverty reduction are linearly related, h^ =0 +^., then changes in poverty
1
can be expressed as a function of sectoral growth,
h^ =0 +1 si y^i +2 (li - si ) y^i
I I
(12)
i=1 i=1
Collecting terms,
h^ = 0 + 1 -2 +2
I li si y^i (13)
i=1 si
This expression indicates that a sector's growth effect on poverty reduction depends on
its labor intensity, li si . To the extent that sectors differ concerning their labor
intensities, this explains why their effects on poverty alleviation are not the same.
Moreover, since in principle labor intensities can vary not only across sectors but also
across countries for the same sector, then the sectoral growth elasticities of poverty
reduction may be country specific. This may explain in part why our sectoral regressions
are so lacking in precision.
13
How different is labor intensity across sectors and across countries? And, is the
pattern of sectoral growth elasticities of poverty consistent with their labor intensities?
Figure 1 presents box-plots for the cross-country distribution of labor intensities (li si )
corresponding to the six sectors under examination. We notice that, first, with different
degrees, these sectors exhibit a remarkable dispersion across countries; and second, in
spite of this dispersion, it is possible to identify a ranking of labor intensities across
sectors. Agriculture and construction, followed by manufacturing, seem to be the most
labor-intensive sectors, having all of them a median li si ratio larger than 1. The
construction sector is noticeable for the large dispersion of its cross-country distribution
of labor intensity. Conversely, manufacturing shows a concentrated distribution,
particularly regarding the inter-quartile range, which may explain why its coefficient is
estimated with sufficient precision to achieve statistical significance. Mining and
utilities, followed by services, are the least labor-intensive sectors, with median li si
ratios below 1 in all cases. Mining and utilities also show considerable dispersion across
countries in their labor intensity, while services presents the most concentrated
distribution of the six major sectors.
The pattern of coefficients on sectoral growth estimated above is consistent with
the notion that labor intensity determines a sector's influence on poverty alleviation. The
sectors with median labor intensities greater than 1 --agriculture, construction, and
manufacturing-- carry negative coefficients; while those with median labor intensities
lower than 1 --mining, utilities, and services-- have positive coefficients. Moreover, the
ranking of labor intensities (in decreasing order) coincides exactly with the ranking of
sectoral coefficients (from more to less negative) estimated for the reduced sample and
with those estimated for the full sample with only one exception (mining and services
switch places).
The consistency between labor intensities and the pattern of estimated sectoral
growth coefficients is suggestive, but a more formal test can be conducted on the basis of
out theoretical model. Equation (12) can be written as a regression equation of the
change in poverty on aggregate and sectoral growth,
14
h^j =0 +1y^ j +2 sI lij-1sij y^ij +
(14)
j
i=1 ij
where, y^ si y^i is (per capita) GDP growth. The coefficient 1 indicates the size
I
i=1
effect of growth on poverty reduction, while 2 reveals its composition effect. Negative
signs are expected for both coefficients if growth helps reduce poverty and if the labor
intensity of growing sectors has an additional impact on poverty alleviation.
In order to estimate equation (14), it is crucial to obtain data on labor intensities
by sector and country. As explained above, we derive these data from information on
sectoral value added from World Bank (2005) and payments to unskilled workers from
the Global Trade Analysis Project (GTAP). We focus on unskilled workers as they are
likely to best represent the poor in each country.
Equation (14) provides a direct test of the model, and this is our basic and
preferred specification. However, there are other possibilities. First, if we believe that
labor intensities are technological driven and common across countries, then we can use a
single li si ratio for each sector for all countries. This may be a good strategy if we are
uncertain as to the quality of the data on labor intensities per country. We implement this
specification by replacing the country-specific labor intensities in equation (16) by their
corresponding sample median per sector. Second, a discrete or categorical version of the
test can be derived by assuming that sectoral growth can have either a high or a low
impact on poverty reduction depending on whether its labor intensity li si is,
respectively, above or below a certain threshold, which we set equal to 1. This approach
is useful if we are still uncertain as to the precise measure of labor intensities but don't
believe that they are common across countries. We implement this specification by
allocating sectors into two groups according to their labor intensity, regressing poverty
changes on the growth rates of high and low labor-intensity groups, and then testing for
the difference between their respective coefficients. Notice that the composition of these
groups can vary from country to country.
Table 3 presents the estimation results for the direct regression implied by the
model (column 2), the two alternative specifications (columns 3 and 4), and a benchmark
15
regression with (per capita) GDP growth as sole explanatory variable (column 1). The
coefficients on aggregate growth (1) are always significantly negative, with larger
magnitudes when labor intensity is controlled for. Most relevant for our purposes, the
coefficient on labor-intensity-weighted sectoral growth --or labor-intensive growth, for
short-- (2) is also negative and highly statistically significant in our preferred
specification (column 2). Interestingly, the regression fit increases considerably (from 15
to 28%) once information on labor intensity is added to that on aggregate growth. Figure
2 shows a partial-regression plot linking the change in poverty and labor-intensive
growth; it confirms a negative pattern that is well established by most observations in the
sample (we consider the issue of outliers below.) Thus, it appears that in addition to the
size of growth, the composition of growth regarding its labor intensity is statistically and
economically relevant for explaining poverty reduction.
The coefficient on labor-intensive growth is also negative and statistically
significant when we use medians per sector across countries to measure labor intensity
(column 3). However, the fit of the regression declines somewhat, revealing that
country-specific data on labor intensities contribute useful information for growth
composition to explain poverty changes. A similar message is obtained from the
alternative specification based on grouping sectors by labor intensity (column 4). The
coefficient on growth in high labor-intensity sectors is negative and statistically
significant, while that on growth in low labor-intensity sectors is much smaller and not
significant. In fact, the null hypothesis that these two coefficients are the same is rejected
with a p-value of 0.07. The R-squared in this case falls considerably with respect to the
preferred case, confirming that the precise numerical values on country-specific labor
intensities provide relevant information that cannot be captured by categorical indicators.
Robustness to outliers and extreme observations
Table 4 presents the results related to the analysis of robustness to ouliers, with
our basic regression repeated in column 1 for comparison purposes. The data on labor
intensity, li si , present a few extreme values that are likely to represent either
measurement error or rare circumstances; in order to avoid their undue influence, in our
basic specification, we truncated the cross-country distribution of labor intensity per
16
sector to values ranging from 5 to 95 percentile of the original distribution. Column 2
presents the regression results when these extreme values are not truncated. The
coefficient on labor-intensive growth continues to be statistically negative, although its
level of significance and the regression fit diminish a little.
Inspection of Figure 2 may raise questions as to the influence of some countries in
our basic results. To dispel these doubts, we run the regression using a procedure that
weighs observations according to how they fit the pattern established by the rest. This is
the robust regression presented in column 3. We also run the regression completely
excluding possible outliers, identified as the countries that receive weights below 0.7 of a
maximum of 1 in the robust procedure. These countries are Argentina, Estonia, Latvia,
and Senegal; and the corresponding results are shown in column 4. In both cases, the
coefficient of interest remains negative and highly statistically significant, with a
magnitude that is almost the same as that in the benchmark case. It is reassuring that the
regression fit increases considerably when the outliers are excluded.
As mentioned above, the way we calculate poverty changes (that is, proportional
with respect to the average poverty level in the spell) produces fewer extreme values than
the standard way of taking log differences. This allows us to keep a larger number of
observations in the sample than would be the case if we applied the criteria in Kraay
(2005). We check whether our basic results still hold in this reduced sample (of 32
countries), and the results are presented in column 5. The sign, significance, and even
magnitude of the coefficient on labor-intensive growth are remarkably similar as those
using the full sample, with a slight gain in regression fit. All in all, the results in Table 4
allow us to conclude that our basic results are robust to the possible presence of outlier or
extreme observations.
Alternative explanations
It may be argued that the importance of the growth composition term is due to its
correlation with other variables that affect poverty changes. The results in Table 5 check
for this possibility by allowing for alternative explanations in turn. First is the issue of
agricultural growth. Given that agriculture is the sector with the highest labor intensity in
most countries, it may be argued that our growth composition variable is just capturing
17
the presence of agriculture, which may affect poverty reduction for reasons unrelated to
labor intensity. We examine this possibility by adding (size-adjusted) agriculture value
added growth as an independent explanatory variable to our basic specification (see
column 1). While the coefficient on agricultural growth is negative but not statistically
significant, the coefficient on labor-intensive growth retains its sign, significance, and
magnitude with respect to our basic specification. This suggests that the importance of
agricultural growth in poverty reduction that has been recognized in the literature is
mostly due to its intensive use of unskilled labor. Most significantly, the importance of
labor intensity in growth's ability to reduce poverty appears to be relevant across all
sectors.
Second is the connection with inequality. A prominent explanation in the
literature as to the differing effect of income growth on poverty reduction is that higher
inequality dampens the beneficial impact of growth (see Ravallion 2004 for references).
If more unequal countries have growth biased against labor-intensive sectors --because,
for instance, inequality induces policies that make labor markets more rigid--, then
excluding inequality from our analysis could be biasing the results in our favor. To
account for this possibility, we control for inequality by adding the Gini coefficient as an
independent explanatory variable (column 2) and by interacting it with both the growth
size and composition terms (column 3). This also captures possible non-linearities in the
relation between wage growth and poverty reduction. In both cases, the growth
composition term remains negative and significant, as in our preferred specification. The
significance of GDP growth per se suffers when the interactions with the Gini coefficient
are added given its high collinearity with the interaction term.
Third is the issue of measurement error due to the discrepancy between national
accounts and household surveys on data for mean income growth. As mentioned above,
poverty measures are constructed from household survey information; and in most
studies connecting poverty and mean income growth, the same source is used for both
variables. We, however, do it otherwise: since our focus is on production and its
composition, we had to use data from national accounts for the explanatory variables. It
is well-known that income growth derived from household survey data shows large and
sometimes systematic differences with that obtained from national accounts (see Deaton,
18
2003). If these differences are correlated with labor intensity in the country, the
coefficient on growth composition may be biased. Moreover, the bias could be in our
favor if national accounts underreported production from unskilled workers. To account
for this possibility, we include mean income growth from household surveys as an
additional explanatory variable (see column 4). As expected, this variable carries a
negative and significant coefficient, and its inclusion produces both an improvement in
the regression fit and a decline in the magnitude of the coefficients on the size and
composition of growth. However, both coefficients remain negative and statically
significant, confirming our hypothesis.
Fourth is the issue of growth endogeneity. Our analysis has been conducted in
differences in order to control for country-specific structural factors that affect poverty
and production jointly. Still, it can be argued that improvements in poverty drive
production growth --possibly through higher rates of accumulation of human capital and
savings-- thus making the analysis in differences also subject to the endogeneity critique.
Although this does not apply to the variable on the composition of growth, its coefficient
may still be biased if composition and size of growth are correlated. To control for the
potential endogeneity of growth, we instrument for it using the average GDP growth of
the country's trading partners as the source of exogenous variation. The instrumental
variable procedure (whose results are presented in column 5) renders coefficients on the
size and labor-intensity of growth that remain negative and highly significant. Moreover,
their magnitudes are even larger that in the benchmark case, indicating that the
endogeneity of growth, if any, was playing against the hypothesis advocated in the paper.
Alternative poverty measures
Our analysis has used the headcount poverty index as the benchmark measure of
poverty given its prominence in both the empirical literature and policy circles.
However, our simple theoretical model builds the case for the importance of the
composition of growth by focusing on its relationship not with poverty directly but with
labor wages. The connection with poverty is made by assuming that wages affect
poverty according to a linear function, which combined with the basic result of the model
brings about the paper's main regression equation. Yet, the linearity of the relationship
19
between wages and the headcount poverty index may be called into question by
considering that the elasticity of this measure to marginal changes in income is nil except
around the poverty line. (Naturally, the justification for the linear assumption in the
paper is that we are dealing with more than marginal changes in income/production.) To
dispel these doubts, we use other poverty measures that are more closely related to wages
and that respond to changes in income over a wider range of the income distribution.
Table 6 shows the results when alternative poverty measures are used to construct
the dependent variable; these measures are the average poverty gap, the average squared
poverty gap, and the Watt's poverty index (columns 1-3, respectively). In all cases, the
size of GDP growth and --most importantly for our purposes-- its labor intensity carry
negative and quite significant coefficients. The regression fit does not improve when we
use these alternative poverty measures instead of the standard headcount index; actually,
the R2 is one-third lower when using the simple poverty gap. Finally, in column 4 we
examine to what extent our benchmark result comes from the connection between labor-
intensive GDP growth and improvements in the incomes of the poor. For this purpose,
we add the growth of the average poverty gap as an additional explanatory variable. We
find that the growth composition term retains its sign and significance, but its size shrinks
to less than one third than in the benchmark. This indicates that, at least partially, labor-
intensive growth affects the headcount poverty index through the incomes of the poor.
The mechanism: distribution or mean component of poverty changes?
The last issue we examine is the mechanism through which the labor intensity of
GDP growth matters for poverty alleviation. In particular, does it affect the distribution
or the mean component of poverty changes? To answer this question we implement the
decomposition introduced by Datt and Ravallion (1992), according to which changes in
poverty can be broken down into the portion due to changes in mean income holding
income distribution constant (i.e., unchanged Lorenz curve), the portion due to changes in
the distribution of income holding constant its mean, and an approximation residual.
Then, we estimate the respective effects of the size and the labor-intensity of GDP
growth on each of these components, applying the restriction that the combined effect of
each explanatory variable must be the same as its corresponding effect on the overall
20
poverty change (which is given by the benchmark regression). We implement this
estimation through a constrained Seemingly-Unrelated-Regression-Equation procedure
(SURE).
The results are presented in Table 7. When SURE estimation ignores the
influence of outliers (columns 1 and 2), we find that labor-intensive growth affects
poverty changes exclusively through their mean component. When we control for the
influence of outliers (Cols. 3 and 4), labor-intensive growth still affects significantly the
mean component, but now it appears to also affect the distribution component though less
strongly and significantly. The strength of the mean-income channel relative to the
distribution channel indicates that labor-intensive growth should not be associated with
zero-sum income changes across households. It's not that labor-intensive growth is
poverty reducing mainly because it implies redistribution from rich to poor. Although
labor-intensive growth improves the relative standing of the poor, its main effect on
poverty is through its beneficial impact on their absolute income.
IV. Concluding Remarks
The first concern that developing countries face in their objective to reduce
poverty is the lack of sufficient economic growth. This is justifiably so given that no
lasting poverty alleviation has occurred in the absence of sustained production growth.
However, growth's sheer size does not appear to be a sufficient condition for profound
poverty reduction. In fact, a complaint often heard in countries around the world is that
the poverty response to growth is sometimes disappointing.
A general argument for the resilience of poverty relies on either the lack of
opportunities presented to the poor or their inability to take advantage of them. If the
poor are malnourished, are uneducated, live in remote areas, or are discriminated against,
the gains of economic growth are likely to escape them. This paper offers a
complementary perspective supporting the general argument on the lack of opportunities.
In a nutshell, the paper argues that not only the size of economic growth matters for
poverty alleviation but also its composition in terms of intensive use of unskilled labor,
the kind of input that the poor can offer to the production process.
21
The paper first illustrates the connection between wage expansion (poverty
reduction), labor intensity, and sectoral growth through a multi-sector theoretical model.
Then, considering the model's insights, it conducts a set of cross-country empirical
exercises on poverty changes as the dependent variable. The paper finds that the impact
of growth on poverty reduction varies from sector to sector and that there is a systematic
pattern to this variation. Sectors that are more labor intensive (in relation to their size)
tend to have stronger effects on poverty alleviation. Thus, agriculture is the most
poverty-reducing sector, followed by construction, and manufacturing; while mining,
utilities, and services by themselves do not seem to help poverty reduction.
After this sectoral-driven empirical analysis, the paper conducts a more direct test
of the model by considering poverty reduction a function of not only aggregate growth
(which would represent growth's size effect) but also a measure of labor-intensive growth
(which would represent its composition effect). The results confirm that poverty
alleviation indeed depends on the size of growth. Moreover, they also indicate that
poverty reduction is stronger when growth has a labor-intensive inclination. This central
result of the paper is robust to the influence of outlier and extreme observations, holds
true for various poverty measures (such as the headcount index, the average poverty gap,
and the Watt's index), and is not driven away by alternative explanations --such as the
importance of agricultural growth in reducing rural poverty, the role of inequality in
dampening the beneficial impact of growth, and the statistical discrepancy between
household surveys and national accounts. Finally, analysis on the mechanisms through
which labor-intensive growth reduces poverty allows us to conclude that this positive
effect does not require or imply redistribution from rich to poor. Although labor-
intensive growth improves the relative standing of the poor, its main effect on poverty is
given by its beneficial impact on their absolute income.
From a positive perspective, these results may help understand the considerable
disparity in the poverty reaction to economic growth and, in particular, why in some
circumstances poverty is irresponsive to production improvements. This would be the
case of, for instance, a country experiencing a mining or oil boom that is unaccompanied
by growth in other sectors. From a normative perspective, this study does not provide
grounds for "industrial" (or selective) policies as it does not deal with the sources of
22
sectoral growth, the complex links across sectors, or the political economy of government
intervention. Instead, the results of the paper suggest that policy distortions that
discourage labor employment or induce capital-biased technological innovation are ill-
advised to reduce poverty. Removing biases against labor, whether policy-induced or
not, can effectively create opportunities for the poor in growing economic activities and,
thus, help them break away from their condition.
23
References
Attanasio, Orazio P., and Miguel Szekely, 1998, "Household Savings and Income
Distribution in Mexico," Working Paper No.390, Inter-American Development Bank,
Research Department.
Besley, Timothy, and Robin Burgess, 2003. "Having Global Poverty," Journal of
Economic Perspectives, 17(3), pp. 3-22.
Bibi, Sami, 2005. "When is Economic Growth Pro-Poor? Evidence from Tunisia," Cahier
de recherche/Working Paper 05-22, CIRPEE.
Bourguignon, Francois, 2003. "The Growth Elasticity of Poverty Reduction: Explaining
Heterogeneity Across Countries and Time Periods," in Inequality and Growth: Theory
and Policy Implications, Theo S. Eicher, and Stephen J. Turnovsky eds., MIT Press;
Cambridge, MA.
Contreras, Dante, 2001. "Economic Growth and Poverty Reduction by Region: Chile
1990-96," Development Policy Review, 19(3), pp.291-302.
Deaton, Angus, 2003. "Measuring Poverty in a Growing World (or Measuring Growth in
a Poor World)," NBER Working Paper No. 9822.
Dollar, David, and Aart Kraay, 2002. "Growth is Good for the Poor," Journal of
Economic Growth, 7(3), pp. 195-225.
Kakwani, Nanak, Shahid Khandker, and Hyun H. Son, 2004. "Pro-Poor Growth:
Concepts and Measurement with Country Case Studies," UNDP International Poverty
Center, Working Paper No.1, Brazil.
Khan, Haider, 1999. "Sectoral Growth and Poverty Alleviation: A Multiplier
Decomposition Technique Applied to South Africa," World Development, 27(3), pp. 521-
530.
Kraay, Aart, 2005. "When is Growth Pro-Poor? Evidence from a Panel of Countries,"
Forthcoming, Journal of Development Economics.
Lopez, J. Humberto, 2004. "Pro-Growth, Pro-Poor: Is There a Tradeoff?" Policy
Research paper No. 3378, World Bank, Washington, D.C.
Lucas, Sarah, and Peter Timmer, 2005. "Connecting the Poor to Economic Growth: Eight
Key Questions," Working Paper, Center for Global Development, Washington, D.C.
Menezes-Filho, Naercio, and Ligia Vasconcellos, 2004, "Operationalising Pro-Poor
Growth: A Country Case Study on Brazil," Paper in the Operationalising Pro-Poor
Growth Work Programme Series by AFD, BMZ, DFID and the World Bank.
24
Miyagiwa, K. and Papageorgiou, C. (2005). "The elasticity of substitution, Hick's
conjecture and economic growth", Working Paper, Department of Economics, Lousiana
State University.
Ravallion, Martin, 2004. "Pro-Poor Growth: A Primer," Policy Research paper No.3242,
World Bank, Washington, D.C.
Ravallion, Martin, and Gaurav Datt, 1996. "How Important to India's Poor Is the Sectoral
Composition of Economic Growth?" The World Bank Economic Review, 10(1), pp.1-25.
Ravallion, Martin, and Gaurav Datt, 2002. "Why Has Economic Growth Been More Pro-
Poor in Some States of India than Others?" Journal of Development Economics, 68(2),
pp. 381-400.
Ravallion, Martin, and Shaohua Chen, 2004. "China's (Uneven) Progress Against
Poverty," Policy Research paper No. 3408, World Bank, Washington, D.C.
Schmidt-Hebbel K., and Serven, L., 1999. "Saving in the world: the stylized facts?"
in The Economics of Savings and Growth: Theory, Evidence, and Implications for Policy,
K. Schmidt-Hebbel, and L.Serven eds. Cambridge University Press; Cambridge, New
York and Melbourne, pp. 6-32.
Thorbecke, Eric, and Hong-Sang Jung, 1996. "A Multiplier Decomposition Method to
Analyze Poverty Alleviation," Journal of Development Economics, 48(2), pp. 279-300.
25
Table 1. Poverty Reduction and Sectoral Growth: 3-Sector Disaggregation
In all regressions, the dependent variable is the annualized growth rate of the headcount poverty index during
the longest spell available for each country. The independent variables are individual sector's per capita value
added growth weighted by the share of this sector's value added in total GDP. In the fully constrained
regression, all sectors are forced to have a common coefficient. The test presented at the bottom of the
unconstrained regression support this restriction.
Full Sample with 3-sector data Full sample with 6-sector Data
Unconstrained Fully constrained Unconstrained Fully constrained
(1) (2) (3) (4)
Agriculture growth -3.718 -1.470** -5.351 -1.359**
(per capita, share-weighted) (3.647) (0.582) (6.119) (0.638)
Industry growth -2.343 -1.470** -2.420 -1.359**
(per capita, share-weighted) (1.636) (0.582) (1.738) (0.638)
Services growth 0.167 -1.470** 0.591 -1.359**
(per capita, share-weighted) (2.041) (0.582) (2.135) (0.638)
Constant 0.006 0.015 -0.000 0.010
(0.020) (0.017) (0.022) (0.018)
Test AG=IND=SER AG=IND=SER
Test p-value 0.62 0.51
Observations 55 55 51 51
R-squared 0.13 -- 0.12 --
Numbers in parentheses are robust standard errors.
* significant at 10%; ** significant at 5%; *** significant at 1%
26
Table 2. Poverty Reduction and Sectoral Growth: 6-Sector Disaggregation
In all regressions, the dependent variable is the annualized growth rate of the headcount poverty index during the longest
spell available for each country. The independent variables are individual sector's per capita value added growth weighted
by the share of this sector's value added in total GDP. In the partially constrained regressions, the sectors carrying
coefficients of the same sign (in the unconstrained regression) are forced to have a common coefficient. The tests
presented at the bottom of the unconstrained regressions support these restrictions. The reduced sample results from
applying the criteria in Kraay (2005).
Full sample Reduced sample
Unconstrained Partially constrained Unconstrained Partially constrained
(1) (2) (3) (4)
Agriculture growth -11.204 -4.119** -11.695 -3.416**
(per capita, share-weighted) (7.269) (1.629) (7.010) (1.353)
Mining growth 1.120 1.373 4.661 2.386
(per capita, share-weighted) (4.628) (1.459) (4.171) (1.507)
Manufacturing growth -3.829* -4.119** -3.624** -3.416**
(per capita, share-weighted) (2.175) (1.629) (1.422) (1.353)
Utilities growth 10.726 1.373 12.329** 2.386
(per capita, share-weighted) (9.605) (1.459) (5.651) (1.507)
Construction growth -6.314 -4.119** -4.518 -3.416**
(per capita, share-weighted) (5.699) (1.629) (4.614) (1.353)
Services growth 1.491 1.373 2.123 2.386
(per capita, share-weighted) (2.308) (1.459) (2.645) (1.507)
Constant -0.005 0.001 -0.036 -0.031*
(0.026) (0.019) (0.022) (0.018)
Test 1 MIN=U=SER MIN=U=SER
Test p-value 0.65 0.43
Test 2 AG=MA=C AG=MA=C
Test p-value 0.58 0.16
Test 3 AG=MIN AG=MIN
Test p-value 0.06 0.04
Observations 51 51 31 31
R-squared 0.17 -- 0.29 --
Numbers in parentheses are robust standard errors.
* significant at 10%; ** significant at 5%; *** significant at 1%
27
Table 3. Poverty Reduction and Labor-Intensive Growth
In all regressions, the dependent variable is the annualized growth rate of the poverty headcount during the longest spell
available for each country. GDP growth is the average growth rate of real GDP per capita during the corresponding spell.
Labor intensive growth is, for each country, the sum across sectors of the product of a sector's per capita value-added
growth, its share on total GDP, and its surplus use of unskilled labor. For each sector and country, the surplus use of
unskilled labor is the difference between its labor intensity (the ratio of a sector's share of total unskilled labor employment to
its share of total value added) and one. In the calculation of the median-weighted labor intensive growth, country-specific
sectoral labor intensity is replaced by the cross-country median sectoral labor intensity. The growth of high (low) labor
intensity sectors is the share-weighted growth of the sectors with labor intensity greater (lower) than 1.
Only volume of growth Adding composition of growth
Country-specific lj/sj Median lj/sj High and low lj/sj
(1) (2) (3) (4)
GDP growth -1.597** -1.936*** -2.640***
(0.657) (0.569) (0.653)
Labor intensive growth -13.622***
(4.607)
Median-weighed labor -21.419**
intensive growth (9.429)
Growth of high labor -3.429**
intensity sectors (H) (1.521)
Growth of low labor -0.194
intensity sectors (L) (0.650)
Constant 0.014 0.007 0.006 0.010
(0.018) (0.016) (0.017) (0.019)
Test H=L
Test p-value 0.07
Observations 51 51 51 51
R-squared 0.15 0.28 0.25 0.13
Numbers in parentheses are robust standard errors.
* significant at 10%; ** significant at 5%; *** significant at 1%
28
Table 4. Robustness to Outliers and Different Samples
In all regressions, the dependent variable is the annualized growth rate of the poverty headcount during the longest spell
available for each country. GDP growth is the average growth rate of real GDP per capita during the corresponding spell.
Labor intensive growth is, for each country, the sum across sectors of the product of a sector's per capita value-added
growth, its share on total GDP, and its surplus use of unskilled labor. For each sector and country, the surplus use of
unskilled labor is the difference between its labor intensity (the ratio of a sector's share of total unskilled labor employment
to its share of total value added) and one. Column (1) reproduces the benchmark regression for reference. In Column (2),
the measure of unskilled labor intensity did not trim the outliers. Column (3) shows the results obtained a procedure that is
robust to outliers. Column (4) reports the results obtained after dropping Argentina, Estonia, Latvia, and Senegal, the
largest outliers, from the sample. Column (5) shows the results obtained using the restricted sample that results from
applying the criteria in Kraay (2005).
Including Outliers Robust to Excluding Reduced
Benchmark
for lj/sj Outliers Outliers Sample
(1) (2) (3) (4) (5)
GDP growth -1.936*** -1.872*** -2.134*** -2.483*** -1.754***
(0.569) (0.582) (0.531) (0.485) (0.625)
Labor intensive growth -13.622*** -11.069** -13.174*** -13.147*** -11.350**
(4.607) (4.954) (4.772) (4.274) (4.861)
Constant 0.007 0.008 0.008 0.009 -0.006
(0.016) (0.016) (0.016) (0.013) (0.014)
Observations 51 51 51 47 32
R-squared 0.28 0.25 0.29 0.48 0.31
Numbers in parentheses are robust standard errors.
* significant at 10%; ** significant at 5%; *** significant at 1%
29
Table 5. Allowing for Alternative Explanations
In all regressions, the dependent variable is the annualized growth rate of the poverty headcount during the longest
spell available for each country. GDP growth is the average growth rate of real GDP per capita during the
corresponding spell. Labor intensive growth is, for each country, the sum across sectors of the product of a sector's
per capita value-added growth, its share on total GDP, and its surplus use of unskilled labor. For each sector and
country, the surplus use of unskilled labor is the difference between its labor intensity (the ratio of a sector's share of
total unskilled labor employment to its share of total value added) and one. Column (1) controls for the (share-
weighed) growth in agricultural value added. Column (2) controls for a direct effect of the Gini coefficient. Column
(3) controls for potential interactions between the Gini inequality coefficient and, respectively, the volume and
composition of growth. Column (4) controls for the growth in mean income (or expenditure) from household
surveys. Column (5) accounts for the potential endogeneity of GDP growth using the weighted average GDP growth o
Controlling for:
Agricultural Inequality Interactions Survey mean Endogeneity of
growth with inequality growth GDP growth
(1) (2) (3) (4) (5)
GDP growth -1.706*** -1.999*** -1.309 -1.239* -3.656***
(0.635) (0.546) (2.040) (0.627) (0.772)
Labor intensive growth -13.500*** -13.483*** -25.776* -9.194** -17.557***
(4.692) (4.718) (13.734) (4.218) (5.042)
Agricultural growth -6.523
(share-weighted) (6.123)
Gini*GDP growth -1.568
(4.797)
27.726
Gini*Labor intensitive growth (28.883)
Gini -0.001
(0.001)
Survey mean growth -0.515***
(0.162)
Constant 0.006 0.059 0.007 0.005 0.035*
(0.017) (0.047) (0.017) (0.015) (0.021)
Observations 51 51 51 51 50
R-squared 0.30 0.30 0.30 0.44
Numbers in parentheses are robust standard errors.
* significant at 10%; ** significant at 5%; *** significant at 1%
30
Table 6. Using Alternative Poverty Measures
In all regressions, the dependent variable is the annualized growth rate of the
corresponding poverty measure indicated below during the longest spell available for
each country. GDP growth is the average growth rate of real GDP per capita during the
corresponding spell. Labor intensive growth is, for each country, the sum across sectors
of the product of a sector's per capita value-added growth, its share on total GDP, and its
surplus use of unskilled labor. For each sector and country, the surplus use of unskilled
labor is the difference between its labor intensity (the ratio of a sector's share of total
unskilled labor employment to its share of total value added) and one. In columns (1)-(3),
the poverty measure is the (average) poverty gap, the (average) squared poverty gap,
and the Watt's index, respectively. In column (4) the poverty measure is the headcount
poverty index, as in the benchmark regression; however, in this column, the growth of the
poverty gap is used to control for the mean income of the poor.
Alternative poverty measure:
Squared
Poverty Poverty Watt's Poverty Controlling for
Gap Gap Index Poverty Gap
(1) (2) (3) (4)
GDP growth -2.068*** -5.283** -4.228** -0.513***
(0.725) (2.430) (1.658) (0.175)
Labor intensitve growth -14.220** -29.159** -24.260** -3.830**
(6.940) (12.588) (10.478) (1.782)
Growth of poverty gap 0.689***
(0.048)
Constant 0.017 0.079 0.058 -0.004
(0.025) (0.062) (0.046) (0.005)
Observations 51 51 51 51
R-squared 0.19 0.28 0.27 0.94
Numbers in parentheses are robust standard errors.
* significant at 10%; ** significant at 5%; *** significant at 1%
31
Table 7. The Effect on Distribution and Mean Components of Poverty Changes
In these regressions, changes in poverty are decomposed into the portion due to changes in mean income
holding income distribution constant (i.e., unchanged Lorenz curve), the portion due to changes in the
distribution of income holding constant its mean, and an approximation residual. This is the decomposition
introduced by Datt and Ravallion (1992) and implemented by Kraay (2003) for the cross-section of countries
we use in this paper. Estimation is obtained through a seemingly-unrelated-regression-equation (SURE)
system, where the sum of the corresponding coefficients on the growth, distribution, and residual components
is restricted to be the same as the respective coefficient in the benchmark regression. (The coefficients on the
residual component are not presented.) SURE estimation is conducted ignoring (Cols. 1 and 2) and
controlling for (Cols. 3 and 4) the influence of outliers. Regarding the explanatory variables, GDP growth is the
average growth rate of real GDP per capita during the corresponding spell, and labor intensive growth is, for
each country, the sum across sectors of the product of a sector's per capita value-added growth, its share on tot
SURE SURE - Robust to Outliers
Mean Distribution Mean Distribution
Component Component Component Component
(1) (2) (3) (4)
GDP growth -2.142*** 0.005 -2.231*** 0.041
(0.425) (0.360) (0.253) (0.265)
Labor intensive growth -12.510*** -0.862 -8.586*** -4.318*
(3.792) (3.216) (2.371) (2.482)
Constant -0.001 0.003 0.008 -0.004
(0.015) (0.012) (0.010) (0.010)
Observations 49 49 48 48
Numbers in parentheses are robust standard errors.
* significant at 10%; ** significant at 5%; *** significant at 1%
32
Figure 1. Cross-Country Distribution of Labor Intensity (li/si) per Sector
0 .5 1 1.5 2
From top to bottom:
1. Mining 2. Utilities 3. Services
4. Manufacturing 5. Construction 6. Agriculture
33
Figure 2. Poverty Change and Labor-Intensive Growth
Partial-regression observations, controlling for per capita GDP growth
.3
LVA
.2
index ETH ARG
BLR ZWE
YEM EST
.1 PER NER
CHN KOR
headcount COL
LSO BGD TTO
GHA PRY LTUIND
MYS
0 TUR
LKA NGA
VEN
MARTHA SLV EGY
poverty UGA CRIPAN JOR ZAF
MDG IDN TUN ZMB
of MEX ECU
PHL JAM PAKBRA CHL
HNDGTM
-.1 KEN
VNM
Growth
DOM
-.2 POL
SEN
-.005 0 .005 .01
Labor intensive growth
coef = -13.621678, se = 4.6001684, t = -2.96
34
Appendix 1. Samples of Countries
3-sector data 6-sector data Non-outliers Kraay's criteria
WB Code Country Name Spell (55-countries) (51-countries) (47-countries) (32-countries)
ARG Argentina 1992 - 1998
BDI Burundi 1992 - 1998
BGD Bangladesh 1984 - 1992
BLR Belarus 1988 - 1995
BRA Brazil 1985 - 1998
CHL Chile 1987 - 1998
CHN China 1990 - 1998
COL Colombia 1988 - 1998
CRI Costa Rica 1993 - 1996
DOM Dominican Republic 1989 - 1998
DZA Algeria 1988 - 1995
ECU Ecuador 1988 - 1995
EGY Egypt 1991 - 1999
EST Estonia 1993 - 1995
ETH Ethiopia 1995 - 2000
GHA Ghana 1987 - 1999
GTM Guatemala 1987 - 1989
HND Honduras 1989 - 1998
IDN Indonesia 1987 - 2000
IND India 1983 - 1997
JAM Jamaica 1988 - 2000
JOR Jordan 1987 - 1997
KEN Kenya 1992 - 1997
KOR Korea, Rep. 1988 - 1993
LKA Sri Lanka 1985 - 1995
LSO Lesotho 1986 - 1995
LTU Lithuania 1996 - 2000
LVA Lativa 1993 - 1998
MAR Morocco 1985 - 1999
MDG Madagascar 1993 - 1999
MEX Mexico 1989 - 1998
MLI Mali 1989 - 1994
MRT Mauritania 1988 - 1995
MYS Malaysia 1984 - 1997
NER Niger 1992 - 1995
NGA Nigeria 1985 - 1997
PAK Pakistan 1987 - 1998
PAN Panama 1991 - 1996
PER Peru 1985 - 1994
PHL Philippines 1985 - 2000
POL Poland 1993 - 1998
PRY Paraguay 1990 - 1998
SEN Senegal 1991 - 1994
SLV El Salvador 1989 - 1998
THA Thailand 1988 - 2000
TTO Trinidad and Tobago 1988 - 1992
TUN Tunisia 1985 - 1990
TUR Turkey 1987 - 2000
UGA Uganda 1989 - 1996
VEN Venezuela 1981 - 1998
VNM Vietnam 1993 - 1998
YEM Yemen, Rep. 1992 - 1998
ZAF South Africa 1993 - 1995
ZMB Zambia 1991 - 1998
ZWE Zimbabwe 1990 - 1995
35
Appendix 2. Descriptive Statistics for 51-Country Sample
Poverty Reduction and Sectoral Growth
(a) Univariate
variable mean median min max sd
Growth in headcount poverty index -0.014 -0.028 -0.279 0.267 0.106
Industry growth 0.008 0.004 -0.018 0.070 0.014
Agriculture growth* 0.000 0.000 -0.004 0.005 0.002
Mining growth* 0.000 0.000 -0.014 0.016 0.004
Manufacturing growth* 0.005 0.003 -0.006 0.046 0.009
Utilities growth* 0.001 0.001 -0.002 0.009 0.002
Construction growth* 0.002 0.001 -0.003 0.016 0.003
Services growth* 0.010 0.009 -0.018 0.033 0.011
(b) Bivariate Correlation
Growth in
Agricul- Manu-
head-count Industry Mining Utilities Construc- Services
variable ture facturing
poverty growth growth* growth* tion growth* growth*
growth* growth*
index
Growth in headcount poverty index 1.00
Industry growth -0.32 1.00
Agriculture growth* -0.28 0.42 1.00
Mining growth* -0.12 0.58 0.41 1.00
Manufacturing growth* -0.31 0.91 0.31 0.32 1.00
Utilities growth* -0.11 0.61 0.23 0.17 0.56 1.00
Construction growth* -0.20 0.52 0.05 0.15 0.41 0.40 1.00
Services growth* -0.13 0.60 0.14 0.32 0.57 0.30 0.57 1.00
Note: * All sectoral growths are per capita value added in that sector (weighted by the share of its value added in total GDP).
36
Appendix 3. Descriptive Statistics for 51-Country Sample
Poverty Reduction and Labor Intensive Growth
(a) Univariate
variable mean median min max sd
Growth in headcount poverty index -0.014 -0.028 -0.279 0.267 0.106
Proportional change in poverty gap -0.006 -0.027 -0.316 0.310 0.138
Proportional change in squared poverty gap 0.013 -0.020 -0.513 1.291 0.269
Proportional change in Watt's poverty index 0.005 -0.025 -0.463 0.888 0.222
GDP per capita growth 0.018 0.016 -0.061 0.091 0.026
Labor intensive growth -0.001 -0.001 -0.007 0.006 0.003
Gini coefficient 43.61 43.19 21.78 66.25 10.49
Survey mean growth 0.011 0.016 -0.416 0.301 0.090
Mean component of Pov. change -0.020 -0.028 -0.279 0.266 0.103
Distribution component of Pov. change -0.024 -0.038 -0.252 0.262 0.098
(b) Bivariate Correlation
Propor- Propor-
Propor-
Growth in tional tional Mean Distribution
tional GDP per Labor Gini Survey
head-count change in change in component component
variable change in capita intensive coeffi- mean
poverty squared Watt's of Pov. of Pov.
poverty growth growth cient growth
index poverty poverty growth growth
gap
gap gap
Growth in headcount poverty index 1.00
Proportional change in poverty gap 0.96 1.00
Proportional change in squared poverty gap 0.90 0.90 1.00
Proportional change in Watt's poverty index 0.93 0.95 0.99 1.00
GDP per capita growth -0.45 -0.37 -0.48 -0.46 1.00
Labor intensive growth -0.25 -0.20 -0.19 -0.19 -0.21 1.00
Gini coefficient -0.01 0.07 -0.03 0.01 -0.13 0.06 1.00
Survey mean growth -0.62 -0.53 -0.46 -0.49 0.33 0.19 0.24 1.00
Mean component of Pov. change 1.00 0.96 0.90 0.93 -0.46 -0.25 -0.01 -0.62 1.00
Distribution component of Pov. change 0.73 0.61 0.50 0.54 -0.46 -0.25 -0.17 -0.87 0.73 1.00
37