World Development Vol. 36, No. 8, pp.

1317–1341, 2008
 2008 Elsevier Ltd. All rights reserved
0305-750X/$ - see front matter
www.elsevier.com/locate/worlddev
doi:10.1016/j.worlddev.2007.08.003

Social Spending, Human Capital, and Growth

in Developing Countries
EMANUELE BALDACCI, BENEDICT CLEMENTS,
SANJEEV GUPTA and QIANG CUI *
International Monetary Fund, Washington, DC, USA
Summary. — Using panel data from 118 developing countries in 1971–2000, this paper explores the
channels linking social spending, human capital, and growth and compares the effects of alternative
economic policy interventions. With separate modeling for education and health capital, explicit
control for governance, and incorporation of nonlinearity, the paper finds that both education
and health spending have a positive and significant impact on education and health capital, and
thus support higher growth. Also, other policy interventions, such as improving governance and
taming inflation, can achieve similar results. Hence, higher spending alone is likely insufficient to
achieve the Millennium Development Goals.
 2008 Elsevier Ltd. All rights reserved.

Key words — economic growth, human capital, social spending, MDGs

1. INTRODUCTION help countries meet the MDGs (via their salu-

tary effects on economic growth) are of great
The role of human capital in fostering eco- interest. While positive externalities or market
nomic development is well recognized in the failures may justify the involvement of the pub-
growth literature. Following Romer (1986) lic sector in these areas, this does not, in itself,
and Lucas (1988), human capital has been iden- indicate that higher spending per se is the most
tified not only as a key contributor to growth effective or the only policy intervention for
and poverty alleviation, but also as an ultimate helping meet the MDGs. 2
development objective to expand human free- This paper seeks to contribute to this debate
dom more generally (Ravallion & Chen, 1997; by providing an integrated assessment of the
Schultz, 1999; Sen, 1999; Squire, 1993). The role of social spending and other policy inter-
growing global focus on the Millennium Devel- ventions on human capital, economic growth,
opment Goals (MDGs) 1 has further high- and social indicators. Building upon earlier
lighted the importance of making tangible studies, we analyze the dynamic direct and indi-
progress in key education and health indica- rect effects of social spending on human capital
tors. and growth, while taking into account the inter-
A crucial issue in this regard is the role of action between education and health interven-
public policy in helping countries foster human tions. The empirical estimates are based on a
capital and meet the MDGs. Education and panel dataset covering 118 developing countries
health are two essential dimensions of human from 1975 to 2000. The paper also examines the
capital and core elements of the MDGs. In impact of different policy interventions for fos-
most countries, the public sector plays a domi- tering human capital and growth.
nant role in providing the educational and
health services necessary to build human capi-
tal. As such, the impact of this spending on so-
cial indicators, and the impact of higher * We would like to thank Alex Segura-Ubiergo, David
spending versus other policy interventions (such Coady, David Newhouse, Simon Johnson, and three
as improvements in fiscal sustainability or anonymous referees for useful comments on an earlier
improvements in governance) that might also draft. Final revision accepted: August 10, 2007.
1317
1318 WORLD DEVELOPMENT

The remainder of the paper is structured as education capital (Boozer, Ranis, Stewart, &
follows. In Section 2, a review of the existing lit- Suri, 2003; Ranis & Ramirez, 2000).
erature is provided. In Section 3, an explana- The empirical literature on the effects of
tion of the data and model is given. Section 4 health capital on growth is relatively thin. Con-
provides the empirical results, including robust- ceptually, a healthy person cannot only work
ness tests. Section 5 summarizes the simulated more effectively and efficiently, but also devote
effects of different policy interventions on more time to productive activities. Based on
growth and social indicators. Section 6 dis- microeconomic evidence, Strauss and Thomas
cusses policy implications and concludes the (1998) argue that health status explains varia-
paper. tion in wages at least as much as education lev-
els. Recent experimental or quasi-experimental
studies, such as Thomas and Frankeberg
2. LITERATURE REVIEW (2002) and Thomas et al. (2003), have found
that specific health sector interventions help
Most studies in the literature have focused recipients raise earnings significantly, and gen-
on only one segment of the social spending— eral indicators of health and nutrition status
social indicators—economic growth nexus. are significant predictors of economic success.
That is, studies either assess the growth effects Research at the macro level can better capture
of improving education or health indicators, or the potential externalities of health sector inter-
the impact of public spending on these indica- ventions, and several recent studies support the
tors. Furthermore, research on the first stream positive contribution of health capital to
has concentrated primarily on education capi- growth. Barro (1996b), Bloom and Canning
tal, and has often focused on the impact of (2003), Bloom, Canning, and Sevilla (2004)
the initial stock of education capital on and Gyimah-Brempong and Wilson (2004) find
growth. These studies generally confirm that that health capital indicators positively influ-
enrollment and/or schooling boosts growth ence aggregate output. For the countries in
(Barro, 1996a, 1996b; Barro & Sala-i-Martin, their sample, about one-fourth of economic
1995; Levine & Renelt, 1992; Mankiw, Romer, growth was attributable to improvements in
& Weil, 1992; Sala-i-Martin, 1997). Using a health capital, and improvements in health con-
more refined measure on skills available for ditions equivalent to one more year of life
14 OECD countries, Coulombe, Tremblay, expectancy are associated with higher growth
and Marchand (2004) find that a country with of up to 4 percentage points per year.
literacy scores 1% higher than the average Studies examining the impact of social spend-
experiences an increase in per capita GDP ing on social indicators have produced mixed
growth of 1.5 percentage points. Some studies, results both for industrial and developing coun-
however, find the macroeconomic evidence to tries. For example, based on cross-sectional
be unconvincing (Benhabib & Spiegel, 1994; data for developing countries, Baldacci, Guin-
Pritchett, 1996) and inconsistent with the find- Sui, and de Mello (2003) and Gupta, Verho-
ings at the microeconomic level on the returns even, and Tiongson (2002b) find that social
to education. spending is an important determinant of educa-
These findings on the link between education tion and health outcomes. 3 They also find that
capital and economic growth raise a number of education spending has a greater effect on social
important issues. First, country heterogeneity indicators than health outlays. The positive ef-
matters. For example, papers utilizing samples fect of social spending on social indicators is
that include developed countries tend to find also supported by Anand and Martin (1993),
weaker results, which is consistent with dimin- Psacharopoulos (1994), Hojman (1996), Bidani
ishing returns in education. In light of this het- and Ravallion (1997) and Psacharopoulos and
erogeneity, Jones and Olken (2005) argue that Patrinos (2002). At the same time, a number
the within-country dimension is critical for of studies have found insignificant or very weak
explaining the determinants of growth. Second, linkages between public education outlays and
the way in which education capital is measured education indicators (Flug, Spilimbergo, &
and modeled can affect the empirical results Wachtenheim, 1998; Mingat & Tan, 1992,
(Krueger & Lindahl, 2001). Finally, it is impor- 1998; Noss, 1991). Similarly, a number of stud-
tant to incorporate feedback effects between ies find that the contribution of health spending
education capital and growth relationship to to health status—as measured by infant mortal-
correctly gauge the growth effects of enhancing ity or child mortality—is either small or statisti-
 SOCIAL SPENDING, HUMAN CAPITAL, AND GROWTH IN DEVELOPING COUNTRIES 1319

cally insignificant (Filmer & Pritchett, 1997; Fil- In the few studies where these interactions have
mer, Hammer, & Pritchett, 1998; Musgrove, been incorporated, the salutary effects have
1996; Pritchett, 1996). In contrast, Gupta, been found to be significant. Ranis and
Verhoeven, and Tiongson (2003) find a positive Ramirez (2000) argue that explicitly modeling
relationship between public spending on health the two-way relationship between human capi-
care and the health status of the poor. tal and growth leads to the finding of a statisti-
The mixed results from the literature on the cally significant relationship in both directions
effectiveness of education and health spending in developing countries. A rare exception that
in improving social indicators can be traced examines the effects of spending on growth is
to several factors. First, existing studies have Gyimah-Brempong and Wilson (2004), who
been hampered by data availability and mea- find a positive and robust link between invest-
surement problems. Statistics on social spend- ment in health and growth in both sub-Saharan
ing and the relevant social outcomes are African and OECD countries.
relatively scarce compared to macroeconomic The present study builds on this existing liter-
indicators. The use of large panel data series ature and attempts to correct some of the weak-
with better consistency in compilation can help nesses described above. In particular, the
mitigate these problems. Second, there is the is- interaction between the education and health
sue of model design. Many studies have used sectors are incorporated in the sectoral
traditional linear models to capture the spend- equations. The impact of governance on the
ing-outcomes relationship, while in many cases productivity of public spending is also fully
a nonlinear specification yields more robust re- incorporated into the analysis. In addition,
sults. Third, the literature has often failed to the system allows for feedback between the
capture the interaction between the education higher growth attributable to better human
and health sector and the significant spillovers capital and country income levels.
between them. In effect, this leads to an under-
statement of the effect of spending and social
indicators (Mayer-Foulkes, 2003; Miguel & 3. DATA AND MODEL
Kremer, 2003). Higher levels of education, for
example, increase public awareness and the (a) Methodology
capacity of families to address their own health
needs. At the same time, better health enhances The econometric approach in this paper aims
the effective and sustained use of the knowl- to capture the potential impact of social spend-
edge and skills that individuals acquire through ing on social indicators and growth in the con-
education (Schultz, 1999). Barro (1996b) fur- text of an endogenous growth model. For this
ther argues that better health can reduce the purpose, we use general specifications for real
depreciation of education capital, and thus in- per capita income growth, total investment,
creases the favorable effect of education on education capital, and health status equations,
growth. drawing on a set of robust explanatory vari-
Another weakness in much of the research is ables used in the existing literature. We also in-
the failure to control for variables that may clude controls for the quality of governance in
have an impact on the effectiveness of social all four equations. In addition, in the growth
spending. In particular, recent research has also equation, we explicitly model both the level
highlighted the important role of institutions and changes in education and health capital
and governance in mediating the nexus between as a key determinant of economic performance.
social spending, indicators, and growth. Poor In the education and health equations, we allow
governance has been identified as a key cause interactions between education and health cap-
of ineffective social spending (e.g., Abed & ital indicators and the presence of nonlinear ef-
Gupta (2002); Gupta, Davoodi, & Alonso- fects at different levels of spending. We also
Termé, 2002a; Mauro, 1998; Rajkumar & include, as appropriate, lagged values of the
Swaroop, 2002). Rodrik, Subramanian, and spending variables to take into account the
Trebbi (2004) and Hausmann, Pritchett, and appropriate dynamic structure of the links
Rodrik (2005) further highlighted the importance between social indicators and spending vari-
of governance in sustaining economic growth. ables. 4 Finally, we conduct a set of simulations
Finally, few studies have examined the feed- based on the estimate results to assess the im-
back effects between social spending, social pact of higher social spending and other key
indicators, and growth in an integrated system. economic interventions on social indicators,
1320 WORLD DEVELOPMENT

while accounting for the lagged and feedback use lagged levels of the endogenous variables
effects among variables. and, where appropriate, some exogenous vari-
As Temple (1999) and Baltagi (2001) argue, ables as instruments for the 2SLS. To assess
panel estimators are the most appropriate whether these instruments are valid—in the
choices for growth regressions, as they do not sense that they only affect the dependent vari-
constrain technology to be the same across able through their correlation with the endoge-
countries and better account for dynamic ef- nous variables for which they instrument—we
fects that are not captured in purely cross-sec- conduct Sargan overidentification tests and
tional regressions. Nonetheless, the use of joint-F tests for the first-stage regressions. 6
panel estimates also involves several important The system GMM estimator addresses endoge-
difficulties that need to be addressed: small neity by simultaneously solving level and differ-
sample bias, measurement error, endogeneity, ence equations and does not use exogenous
and heterogeneity. instruments. 7 Given the limitations of an IV
There are a number of different estimators approach, and the significant drop in sample
that can be used to overcome these problems, size (due to the use of lagged variables for the
each with its own strengths and weaknesses. endogenous variables and the limited coverage
In this paper, we utilize the fixed-effect model of some of the exogenous instruments), these
(LSDV) as our baseline model, but also use results should be interpreted with caution.
alternative estimators to assess the robustness All estimators are vulnerable to measurement
of the results. The LSDV is to be preferred error. However, to help minimize this problem,
for three main reasons. First, LSDV minimizes we compare our results with those from a ro-
distortions from heterogeneity and fixed-ef- bust estimator, 8 which places lower weight on
fects-related endogeneity. Second, the tech- outliers. We also use five-year averages for all
nique is a good choice in light of the short variables to reduce measurement error. 9 Such
time series available for each country. As indi- averaging is also useful to smooth possible
cated by Islam (2000), Monte Carlo results sug- business cycle effects. The robustness of results
gest that LSDV estimators provide excellent is further checked by using alternative indica-
small sample properties and perform well rela- tors of health and education capital. Moreover,
tive to General Method of Moments (GMM) we compare the results with those obtained
and instrumental variable (IV) panel estima- from a latent variable approach, which corrects
tors. 5 Third, LSDV provides a solid basis for for measurement error (Baldacci et al., 2003).
the simulations of the system of equations; as Finally, a FGLS estimator, with specific con-
shown in Duncan (1975) and Baldacci et al. trols for error term autocorrelation and heter-
(2004), LSDV produces unbiased and efficient oskedasticity, is also utilized.
estimates for a group of equations under the The modeling strategy is broadly as follows
assumption that the disturbances are uncorre- (see section below for equation-by-equation
lated across equations, which is consistent with specification). The growth equation follows a
the data as subsequent test based on seemingly typical neoclassical growth formulation, with
unrelated regression fails to reject this assump- separate variables for education and health
tion. capital and an additional variable to capture
LSDV cannot, however, address all prob- governance. The investment equation is stan-
lems, including those related to endogeneity. dard for this literature. The education and
For example, endogeneity may arise in our health capital equations also follow the existing
equations due to reverse causality between the literature but allow for the dynamic effects of
dependent and independent variables, such as lagged spending, nonlinear effects, and interac-
growth, investment, and human capital indica- tion between education and health capital. Fi-
tors. In this light, we estimate two IV-type esti- nally, governance is explicitly modeled in all
mators, 2SLS and system GMM, to further equations.
correct for endogeneity, including the kind aris-
ing from the correlation of error terms in the (i) Growth equation
presence of lagged explanatory variables. As Drawing upon Mankiw et al. (1992), Barro
shown in Temple (1999), finding good instru- (1996a, 1996b), Bassanini and Scarpetta
ments for IV estimators is a considerable chal- (2001), Bloom et al. (2004) and Gyimah-Brem-
lenge in growth models that incorporate pong and Wilson (2004), the growth equation is
dynamic effects. Following Baldacci, Hillman, based on a neoclassical growth framework aug-
and Kojo (2004) and Bloom et al. (2004), we mented by the separate inclusion of education
 SOCIAL SPENDING, HUMAN CAPITAL, AND GROWTH IN DEVELOPING COUNTRIES 1321

capital and health capital and governance. The primary enrollment. For these countries,
per capita output equation is assumed to take increases in education spending are unlikely
the following form: to be allocated to primary education and
thus higher education spending may not be
y ¼ f ðsk ; he; ed; XÞ;
associated with higher primary enrollment.
where y is real per capita GDP; sk is the invest- In low-income countries, where primary
ment ratio; he denotes health capital; ed repre- enrollment rates are lower, greater priority
sents education capital; and X denotes the set may be given to primary education. More
of macro and institutional control variables, generally, the use of an enrollment rate
such as the fiscal balance, inflation rate, trade enjoys several advantages: (i) the enrollment
openness, and governance that augment the rate is considered as a good proxy for other
baseline specification of the model. measures of the stock of human capital, such
In addition, we assume that there is a rela- as years of schooling (Woszmann, 2003),
tionship between both the initial stock and which is also confirmed in our sample; 11
increment in human capital with per capita (ii) the enrollment rate is closely linked with
GDP growth. 10 As such, the growth of real the targets embedded in the MDGs; and (iii)
per capita GDP g is given by enrollment data are largely based on actual
country data, rather than projected or
g ¼ f ðsk ; he; Dhe; ed; Ded; XÞ: estimated data often used for other indica-
Therefore, we consider the baseline growth tors. The use of the enrollment rate as a
equation as the following: proxy is consistent with much of the litera-
ture examining the role of human capital
git ¼ a1i þ g1t þ b11  lnðy it1 Þ þ b12  ni;t on growth (e.g., Bils & Klenow, 2000; Man-
kiw et al., 1992; Ranis & Ramirez, 2000). 12
þ b13  skit þ b14  Ed it1 þ b15  Heit1 Nevertheless, given that the combined
þ b16  DEd it þ b17  DHeit enrollment indicator is an imperfect mea-
sure, the results should be interpreted with
Xn
caution.
þ b1m ðXmit Þ þ uit ; ð1Þ
m¼8
• DEd refers to changes in education capital.
While the stock of human capital affects pro-
where ductivity growth, changes in education capi-
• git is real per capita GDP growth; tal reflect an adjustment in the level of a
• ai and gt denote the country-specific effect productive input (e.g., educated labor). On
and period-specific effect, respectively; the relationship between human capital
• logðy t1 Þ is the lagged logarithm of per and growth, Lucas (1988) emphasized that
capita income to control for the expected changes in human capital promote growth,
reduction in growth rates as per capita whereas Romer (1986, 1990) argued that
incomes rise and there is convergence to the stock of human capital drives growth
steady growth rates; mainly via innovation. The derivation of
• skit denotes the investment ratio, measured the growth equation also leads to the inclu-
in terms of gross capital formation (both pri- sion of both the stocks and changes in
vate and public) in percent of GDP. As an human capital variables. 13 Therefore, both
increase in the investment ratio captures an variables are included.
increase in the stock of physical capital, its • Heit refers to the stock of health capital,
coefficient is expected to be positive; and Dhe refers to changes in health capital.
• Edit refers to the stock of education capi- The health condition of the population
tal, which is proxied by the sum of the gross affects output growth. While it is desirable
primary and secondary enrollment rate. A to account for both mortality and morbidity
combined enrollment rate is used, given the as a full measure of health capital, such data
desire to use a single measure that could cap- are not available. Instead, the logarithm of
ture progress in building human capital for under-5 child mortality is used to proxy the
countries at different stages of development, stock of health capital. To facilitate interpre-
and the fact that data on education spending tation, the signs of the coefficients on mortal-
are only available at an aggregated level. ity rates are reversed so that the positive
Many middle-income countries, for exam- coefficients correspond to improvements in
ple, have already achieved high levels of health status. 14 Mortality data are more
1322 WORLD DEVELOPMENT

readily available for a larger set of countries linear effect of fiscal balances on investment.
over a longer time period than other poten- The equation is as follows:
tial proxies. They are also directly related
to indicators chosen to monitor progress I it ¼ a2i þ g2t þ b21 logðy t1 Þ þ b22 nit
toward the MDGs. Following the derivation
of the growth equation, as well as in Bloom þ b23 Eduit þ b24 Heait þ b25 Balanceit
et al. (2004) and Gyimah-Brempong and þ b26 LowdefBalanceit þ b27 pit
Wilson (2004), both stock and changes in
health capital are included. þ b28 Highpit þ b29 Lowgovit þ uit ; ð2Þ
• nit is the rate of growth of the total
population. where all macro variables are defined as in the
• Xmit consists of control variables. Trade growth equation but only the levels of the edu-
openness (open), changes in terms of trade cation and health indicators are included.
(dtot), fiscal balance (Balance), and inflation
(inf) were included, as they have been fre- (iii) Education equation
quently identified as key determinants of Specification of the education equation draws
growth (Baldacci et al., 2004; Barro, 1996a; upon Baldacci et al. (2003) and Gupta et al.
Levine & Renelt, 1992). Trade openness is (2002b), and other papers as described under
expressed as the sum of total imports and each explanatory variable. In addition to key
exports in relation to GDP. Inflation rate determinants used in the earlier literature, we
is the logarithm of 1 plus the change in included a health capital indicator and an inter-
CPI. Following Fischer (1993) and Baldacci action term to capture the potential effect gov-
et al. (2004), the high-inflation dummy refers ernance may have on the effectiveness of
to cases where the inflation rate exceeds 20% education spending. Finally, we tested for the
per year. An interaction variable of low-def- potential lagged effect of education spending. 16
icit fiscal balance (countries with deficits Only one period of lagged spending was found
below 3% of GDP), LowdefBalance, is also to be significant and hence included as an addi-
created by multiplying fiscal balance with a tional explanatory variable. The equation is
low-deficit dummy 15 to account for possible therefore as follows:
nonlinear effect of fiscal policy. A dummy
for poor governance (Lowgov) was also Eduit ¼ a3i þ g3t þ b31 lnðy it Þ þ b32  Popu15it
included, based on the anticorruption and þ b33  Heai;t þ b34  Urbanit
democratic accountability index from Inter-
national Country Risk Guide (ICRG) com- þ b35 Quality it þ b36 EduSpendingit
piled by the Political Risk Services group. þ b37 EduSpendingit1 þ b38 EduSpending
Including both indexes gives a more com-
plete measure of governance. The dummy  Lowgovit þ b39 EduSpendingi;t1
takes the value of one (zero otherwise) when  Lowgovit1 þ b310 Femaleit þ uit : ð3Þ
the index value is lower than the mean. Also,
we include both the level of the annual infla- This equation examines the direct impact of
tion rate (p) and the high-inflation dummy education spending on education capital, as
(Highp) to measure the possible nonlinear proxied by the composite primary and second-
effects of inflation on growth. ary school enrollment rate.
• uit is the error term. • Income level (y). Higher per capita income
raises the demand for education. This vari-
(ii) Investment equation able is measured as the logarithm of real
The investment equation captures the effect per capita GDP in PPP dollar terms.
of human capital on investment, following • School age population (Popu15). The age
Fischer (1993), Benhabib and Spiegel (1994) structure of the population can affect school
and Mauro (1998). We also add variables that enrollment (Mingat & Tan, 1992). Age
address the investment climate (as proxied by structure is captured by the share of popula-
our governance indicator) and macroeconomic tion below 15 years old.
environment (fiscal balance and inflation). Fol- • Health capital (Hea). A healthier popula-
lowing Baldacci et al. (2004), we include an tion is more likely to be able to afford to
interactive term between the low-deficit dummy invest in education. As in the growth and
and fiscal balance to address the potential non- investment equations, the under-5 child
 SOCIAL SPENDING, HUMAN CAPITAL, AND GROWTH IN DEVELOPING COUNTRIES 1323

mortality rate is used to proxy the stock of significant impact and was dropped. 17 Fur-
health capital. thermore, in contrast to the education equa-
• Urbanization (Urban). As shown in Gupta tion, lagged spending was found to have no
et al. (2002b), households in urban areas are significant effect, and therefore the equation
more likely to send their children to school takes the following form:
because access to education is typically bet-
ter in urban areas. Heait ¼ a4i þ g4t þ b41  lnðy i;t Þ þ b42 Fertility it
• Quality indicator (Quality). Following
þ b43 Eduit þ b44 Urbanit
Gupta et al. (2003), the quality of education
is proxied by the school repetition rate. A þ b45 HeaSpendingit þ b46 HeaSpending
higher repetition rate generally corresponds  Lowgovit þ b47 Femaleit þ u4it : ð4Þ
to a lower quality of education.
• Education spending (EduSpending). Both
the current five-year average and lagged • Income level. Past studies indicate that this
five-year average of education spending, in variable is a crucial determinant of health
percent of GDP, were used. A logarithmic status (Carrin & Politi, 1996; Pritchett &
functional form was used to achieve the best Summers, 1996).
fit and control for diminishing returns to • Gender equality. Similar to the education
spending. Ideally, more cross-country time equation, we used this indicator to capture
series data, disaggregated by level of expen- the institutional factors that may affect
diture, could be utilized, but data for a wide health outcome via female education. A
set of countries are not available. number of studies have identified female edu-
• Poor governance (Lowgov). As shown in cation as an important determinant of the
Abed and Gupta (2002), Baldacci et al. health status of infants and children, as well
(2003) and Gupta et al. (2002b), poor gover- as the population in general (e.g., Schultz,
nance has a strong adverse impact on the 1993; World Bank, 1993). 18 For example,
effect of public spending on social indicators. in developing countries, women play a more
The effects are captured by the interaction important role in family health and sanita-
terms with the spending variables. tion. Furthermore, female education is posi-
• Gender equality (Female) in education tively associated with infants’ health and
also affects aggregate education capital, negatively associated with fertility rates.
given that female enrollment rates tend to While it would be desirable to measure gen-
be lower than that of males. Furthermore, der equality for society as a whole, such data
women are often the primary educators are not available. Instead, the five-year aver-
and nurturers of children, and therefore age of share of female students in primary
investing in girls’ education can reap signifi- and secondary schools is used as a proxy.
cantly greater returns than investing in • Urbanization. Schultz (1993) finds that
boys’ education (Summers, 1992). There- the mortality rate is higher for rural,
fore, higher overall enrollment can be low-income, agricultural households, and
expected in societies with better gender therefore urbanization can be an important
equality, given that in these societies, the determinant of the health status of the pop-
rate of return on education is highest. This ulation. This effect is captured by including
effect is proxied by the share of female stu- the share of the urban population in the
dents in primary and secondary schools. regression.
• Poor governance. This is specified in the
(iv) Health equation same way as in the education equation.
Specification of the health equation draws • Current health spending. The five-year
from Baldacci et al. (2003) and Filmer et al. average of health spending as a percent of
(1998). As with the education equation, an GDP was used. A logarithmic functional
interactive term between poor governance and form was adopted to achieve the best fit
health spending is included to address potential and control for decreasing returns to spend-
effect of governance on the effectiveness of these ing. This is consistent with Deaton (2004),
outlays. Similar to the education equation, we who finds that the benefits from investment
started with a general specification to allow in health are stronger in countries at the ini-
education capital to affect health outcome di- tial stages of development. Similar to educa-
rectly, but this effect was found to have no tion spending, data availability precludes a
1324 WORLD DEVELOPMENT

further breakdown within health spending. The availability of the data for different vari-
The significance of current health spend- ables varies significantly. Macroeconomic data,
ing—but not lagged outlays—can be attrib- such as GDP growth, investment, and inflation,
uted to a number of reasons: First, higher cover all countries for 1971–2000. The coverage
health spending fully translates into better of social indicators and spending data is often
outcomes immediately (that is, 5-year time more limited. This makes the total numbers of
period) 19; and second, health conditions observations in each regression lower than
are not cumulative and must be maintained those in Table 1. 23 Available observations
by regular health care. Thus, higher spend- may also differ between regressions, depending
ing today does not guarantee better health on the variables used, and the availability of
five years from now. instruments may lead to further differences in
sample sizes for the IV estimates.
(v) Decomposition of short-run effects
Following equation-by-equation estimates,
we conducted a decomposition of the direct 4. EMPIRICAL RESULTS
and indirect effects of social spending and other
key policy-related variables on education and (a) Main results
health capital, and also on growth. Given that
education capital was found not to have any ef- The results for the growth, investment, edu-
fect on health, the system of equations can be cation, and health equations are presented in
estimated in a recursive fashion: Tables 2–5. 24 In most cases the coefficients
y ¼ By þ Cx þ n are statistically significant, and all equations
have a good fit.
or in a reduced form y ¼ Px þ n; ð5Þ
where y is a vector of endogenous variables; x is (i) Growth equation
a vector of exogenous variables; B is a lower tri- Changes in education capital are found to
angular matrix containing coefficients for direct positively contribute to economic growth, and
effects of each pair of endogenous variables i the level of education capital is also found to
and j with elements and bij 5 0 and bji = 0; C have positive effects. In the baseline, an increase
is a matrix of direct effects of the exogenous in the growth rate of education capital raises
variables on the endogenous variables, and n the growth rate by 0.1 percentage point, and a
is a vector of the equations’ error terms. The to- similar increase in the level of education capital
tal effect of x on y can be calculated by affects the growth rate positively. Most results
P = (I  B)1C, and the indirect effect of x are similar across the different model estimates.
on y is then given by H = (I  B)1  I. 20 The similarity of the LSDV results with those
provided by the robust estimator indicates that
(b) Data the effect of outliers is minimal. The coefficient
estimates are also similar to those from the
A panel dataset for 118 developing countries 2SLS and system GMM estimators. This sug-
from 1971 to 2000 was compiled for the pur- gests that endogeneity bias does not have a sig-
poses of the paper (see Table 1 for a description nificant effect on our results. The education
of the data and Appendix I for the list of coun- capital coefficients in 2SLS, however, have
tries). Five-year averages are used to smooth much reduced significance, which is likely due
short-term fluctuations and minimize the mea- to the sharp rise in standards errors of the
surement errors that can be present in annual 2SLS estimates and the decline in sample size
observations. All macroeconomic variables are by about 30% when instruments are used.
from the IMF’s World Economic Outlook The impact of health capital on growth dif-
2003 database, while social indicators and fers from that of education capital. Changes
social spending data are from the latest in the health capital indicator positively affect
Barro–Lee dataset, 21 the World Bank (2003) growth with a coefficient of about 1, but the
database, and a social spending database com- coefficient on the level of health capital is insig-
piled from IMF staff reports and the IMF’s nificant. This seems to be consistent with a
Government Financial Statistics. Data on gover- much faster rate of depreciation of health cap-
nance are taken from ICRG index of corrup- ital, consistent with the findings in Barro
tion and democratic accountability. 22 (1996b).
 SOCIAL SPENDING, HUMAN CAPITAL, AND GROWTH IN DEVELOPING COUNTRIES 1325

Table 1. Summary descriptive statisticsa

Variable Observation Mean Standard deviation Minimum Maximum
Real per capita GDP growth (in percent) 652 0.9 3.7 17.8 11.0
Income level (log of real per capita GDP)b 652 9.6 3.2 10.7 16.9
Investment ratio (in percent of GDP) 652 22.2 8.1 1.4 50.2
Primary and secondary gross enrollment 652 134.9 47.9 16.3 238.5
rate (in percent)
Under-5 child mortality (per 1000 live births) 652 105.7 76.4 7.0 349.5
Education spending (in percent of GDP) 538 3.5 1.6 0.6 9.7
Health spending (in percent of GDP) 313 2.0 1.4 0.3 8.7
Population growth (in percent) 652 2.2 1.3 3.3 8.2
Fertility rate (per 1000 people per year) 651 4.7 1.9 1.1 10.0
Immunization rate against measles 470 63.8 26.2 0.0 99.0
(per 1000 people)
Inflation rate (log of changes in CPI) 652 0.2 0.4 0.1 3.4
High-inflation dummy (inflation >20%) 652 0.3 0.4 0.0 1.0
Trade openness (in percent of GDP) 609 68.2 39.2 0.1 247.7
Changes in terms of tradec 646 2.2 17.0 21.2 365.9
Fiscal balance (in percent of GDP) 523 4.9 5.4 29.8 13.9
Governance index (1–12; higher value 347 5.9 1.9 1.0 10.5
equals better governance)
Anti-corruption index (1–6; higher value 347 3.1 1.2 0.0 5.9
equals lower corruption)
Democratic accountability index (1–6; 347 2.8 1.0 0.0 5.1
higher value equals better accountability)
Poor governance dummy (1 for countries 652 0.2 0.4 0.0 1.0
below the average governance index)
Share of under-15 population (in percent) 651 39.3 8.1 16.6 50.2
Urbanization (share of urban population) 647 41.2 21.1 2.9 91.4
School repetition rate (in percent) 411 12.4 9.5 0.1 38.6
Share of female students in primary and 448 44.3 6.9 14.9 59.6
secondary schools (in percent)
Sources: Various databases from the IMF (2003), the Political Risk Service Group (2004), and the World Bank
(2003) (see Appendix A for details).
a
The summary data cover 118 countries with available 5-year averages for GDP, investment, inflation, population
growth, and education and health capital indicators for at least two periods. The coverage of macroeconomic data,
such as growth, investment, and inflation, is significantly better than data on social spending. For example, the
coverage of education and health spending is in the range of 36–112 countries in 1971–2000. The observations
available in each regression may differ and are usually lower than the total observations, depending on the avail-
ability of the relevant variables in each regression. In addition, the decline in sample size is often significant in the IV
models, due to additional restrictions imposed by the availability of exogenous instruments and the use of lagged
variables as instruments.
b
Negative values occur in 10 observations because small values of real per capita GDP becomes negative in loga-
rithm terms for Angola and Democratic Republic of Congo in periods with conflicts.
c
The terms of trade variable compares the relative price of a country’s exports and imports. A positive change
denotes an improvement in the terms of trade.

Other key policy-related variables also affect fiscal balance by 1 percentage point of GDP
growth. Inflation reduces growth, while stron- is associated with higher real per capita growth
ger fiscal balances bolster economic perfor- of 0.01 percentage point, but such salutary ef-
mance. An increase in the rate of inflation by fects disappear when the deficit is already lower
1 percentage point is found to reduce growth than 1% of GDP. Investment also positively af-
by about 0.1 percentage point, but inflation fects growth, while openness to trade has a po-
higher than 10% lowers growth by 0.1 percent- sitive though small impact on growth as well.
age point. In addition, an improvement in the Governance does not appear to affect growth
1326 WORLD DEVELOPMENT

Table 2. Growth equation-dependent variable: average real per capita GDP growth
LSDVa Robust regression FGLS 2SLS Sys. GMM
Investment 0.159 0.145 0.167 0.106 0.164
(7.50)*** (7.66)*** (14.18)*** (1.35) (30.93)***
Population growth 0.041 0.043 0.012 0.031 0.025
(3.09)*** (3.57)*** (2.46)** (1.46) (5.64)***
Catch-up variable 0.150 0.237 0.002 0.227 0.008
(4.87)*** (8.59)*** (1.23) (5.21)*** (5.79)***
Education capital (t  1) 0.093 0.121 0.027 0.082 0.011
(2.44)** (3.58)*** (2.80)*** (0.45) (1.10)
Health capital (t  1) 0.019 0.004 0.010 0.051 0.027
(0.80) (0.21) (1.35) (0.76) (5.94)***
Changes in education capital 0.098 0.135 0.062 0.120 0.009
(2.13)** (3.31)*** (2.86)*** (0.39) (0.67)
Changes in health capital 0.130 0.107 0.110 0.307 0.084
(2.21)** (2.06)** (3.74)*** (1.52) (10.07)***
Trade openness 0.001 0.001 0.001 0.002 0.001
(2.40)** (2.12)** (7.34)*** (3.30)*** (11.96)***
Changes in terms of trade 0.001 0.000 0.000 0.000 0.002
(1.43) (0.51) (0.72) (0.00) (7.70)***
Fiscal balance 0.005 0.004 0.005 0.003 0.005
(3.46)*** (3.53)*** (6.40)*** (1.14) (8.00)***
Low-deficit fiscal balance 0.002 0.003 0.001 0.001 0.011
(0.62) (0.76) (0.31) (0.23) (6.85)***
Inflation rate 0.058 0.051 0.071 0.024 0.165
(2.95)*** (2.87)*** (4.99)*** (0.89) (21.76)***
High-inflation dummy 0.054 0.045 0.037 0.039 0.052
(3.16)*** (2.95)*** (3.97)*** (1.71)* (10.12)***
Low-governance dummy 0.022 0.018 0.025 0.004 0.077
(1.58) (1.39) (3.70)*** (0.19) (16.14)***
Constant 0.704 1.645 0.528 1.664 0.058
(1.64) (4.29)*** (6.82)*** (1.27) (0.97)
Observations 432 432 426 283 318
R-squared 0.74 0.80
P-value (joint F-test)b 0.001
P-value (overidentification) 0.17 0.80
Absolute value of t statistics in parentheses.
a
Baseline regression.
b
Following Baldacci et al. (2004), Bloom et al. (2004) and Temple (1999), we use lags of the endogenous variables as
instruments. The endogenous variables are investment and human capital indicators. We also use lagged government
expenditure as an instrument for the investment rate and current changes in education and health capital indicators.
In this context, we assume that government expenditure directly affects investment and human capital indicators
because government is the primary provider of education and health services in the countries covered in the paper.
Overall government expenditure is unlikely to affect growth, other than through investment and human capital. The
joint F-test in the first stage and the overidentification test fail to reject the validity of these instruments.
*
Significant at 10%.
**
Significant at 5%.
***
Significant at 1%.

directly, although it has indirect effects via not. An increase in health capital of 1 percent-
physical and human capital, as discussed be- age point is associated with an increase in the
low. investment-to-GDP ratio of 0.03 percentage
point. LSDV, the robust estimator, and the
(ii) Investment equation FGLS estimator produce a positive and signif-
The results indicate that health capital affects icant coefficient on health capital, but the IV
investment, although education capital does estimates do not show significant results for this
 SOCIAL SPENDING, HUMAN CAPITAL, AND GROWTH IN DEVELOPING COUNTRIES 1327

Table 3. Investment equation-dependent variable average annual investment ratio

LSDVa Robust regression FGLS 2SLS Sys. GMM
Population growth 0.496 0.834 0.024 0.247 0.548
(1.26) (2.49)** (0.10) (0.49) (1.29)
Education capital 0.025 0.000 0.027 0.237 0.024
(1.29) (0.01) (3.04)*** (1.16) (1.04)
Health capital 0.028 0.045 0.028 0.119 0.019
(1.69)* (3.23)*** (4.63)*** (1.26) (1.39)
Catch-up variable 0.939 1.688 1.193 0.075 2.872
(1.58) (3.34)*** (4.33)*** (0.07) (5.89)***
Changes in terms of trade 0.057 0.001 0.024 0.081 0.090
(1.96)* (0.03) (1.10) (2.08)** (2.49)**
Trade openness 0.070 0.123 0.064 0.042 0.087
(5.34)*** (11.11)*** (8.35)*** (1.83)* (4.91)***
Inflation 0.961 0.268 0.139 1.826 0.263
(1.00) (0.33) (0.24) (1.05) (0.24)
High-inflation dummy 2.198 1.127 1.082 2.636 5.364
(2.55)** (1.54) (2.22)** (2.38)** (5.07)***
Fiscal balance 0.079 0.209 0.094 0.102 0.241
(1.14) (3.53)*** (2.01)** (1.21) (1.93)*
Low-deficit fiscal balance 0.056 0.078 0.263 0.318 0.285
(0.27) (0.45) (2.00)** (1.05) (0.99)
Low-governance dummy 1.607 2.038 0.897 2.132 2.428
(2.09)** (3.13)*** (1.95)* (1.66)* (2.66)***
Constant 32.397 38.690 24.553 63.588 33.354
(5.42)*** (7.62)*** (9.07)*** (1.89)* (6.47)***
Observations 494 494 487 455 357
R-squared 0.67 0.79
P-value for joint F-testb 0.000
P-value for overidentification 0.19 0.13
Absolute value of t statistics in parentheses.
a
Baseline regression.
b
Education and health capital indicators are instrumented by their lagged values, lagged under-15 population share,
and lagged female illiteracy. These variables are highly correlated with human capital indicators but are not likely to
affect investment, except through their effects on human capital. The joint F-test in the first stage and overidentifi-
cation test fail to reject validity of these instruments.
*
Significant at 10%.
**
Significant at 5%.
***
Significant at 1%.

indicator. In contrast, the inflation and gover- effect and a lagged effect. The coefficient on cur-
nance variables are more robust, with similar rent education spending is about 22 and that on
results under different estimators. High infla- the lagged education spending is about 10.
tion is associated with a reduction in the invest- Based on the average education spending and
ment-to-GDP ratios of 2 percentage points. enrollment rates in the sample, these coeffi-
Similarly, poor governance, which is measured cients imply that an increase in education
as a governance rating lower than the develop- spending of 1 percentage point of GDP would
ing country average, is associated with a reduc- raise enrollment rates by 6 percentage points
tion in the investment ratio of 2 percentage in the current five-year period and another 3
points. Finally, similar to the literature, trade percentage points in the following period. 25
openness stimulates investment. However, the impact of governance turns out
to be decisive in the spending effects. The salu-
(iii) Education capital equation tary effects of education spending are offset
Education spending positively affects educa- significantly in a poor-governance environ-
tion capital through both a contemporaneous ment. The positive impact of higher education
1328 WORLD DEVELOPMENT

Table 4. Education equation-dependent variable: average school enrollment rate

LSDVa Robust regression FGLS 2SLS Sys. GMM
Health capital 0.188 0.130 0.362 0.313 0.678
(3.13)*** (3.03)*** (32.88)*** (1.50) (5.81)***
Current per capita income 8.228 6.866 0.465 2.681 1.030
(3.01)*** (3.52)*** (0.44) (0.29) (0.29)
Share of under-15 population 0.815 0.205 0.977 0.707 0.545
(1.79)* (0.63) (16.59)*** (0.61) (1.17)
Urbanization 0.109 0.100 0.028 0.275 0.304
(0.45) (0.57) (1.24) (0.52) (1.13)
Repetition rate 0.024 0.341 0.189 1.080 0.977
(0.07) (1.50) (3.87)*** (1.48) (2.27)**
Share of female students 1.417 2.361 0.969 0.293 1.189
(2.84)*** (6.61)*** (15.96)*** (0.28) (1.50)
Current education spending (in logs) 21.598 17.455 6.906 45.679 18.618
(4.00)*** (4.52)*** (3.99)*** (1.75)* (3.58)***
Lagged education spending (in logs) 9.776 7.520 5.053 14.641 29.702
(1.92)* (2.06)** (3.38)*** (1.11) (4.84)***
Low-governanceb * current education spending 5.262 1.364 0.420 3.207 5.681
(2.32)** (0.84) (0.46) (1.07) (2.49)**
Low-governanceb * lagged education spending 3.529 4.926 4.735 3.650 6.090
(1.43) (2.79)*** (12.55)*** (1.26) (2.79)***
43.389 4.778 144.346 129.215 195.110
Constant (1.14) (0.18) (15.73)*** (1.29) (3.33)***
Observations 300 299 254 168 148
R-squared 0.95 0.98 0.55
P-value for joint F-testc 0.001
P-value for overidentification 0.31 0.69
Absolute value of t statistics in parentheses.
a
Baseline regression.
b
The low-governance dummy is interacted with the spending term to compare the effects of spending in different
governance environments, and thus no governance level variable is included in the regression.
c
The health capital indicator and education spending are instrumented by the immunization rate, fertility rate, and a
democracy index. These instruments are either highly correlated with health status or likely to affect education
spending decisions, but none is likely to contribute directly to education outcomes. The joint F-test in the first stage
and overidentification tests fail to reject validity of these instruments.
*
Significant at 10%.
**
Significant at 5%.
***
Significant at 1%.

spending on education capital is reduced in half this result are further explored in the impact
in countries with poor governance. decomposition in Table 6.
In addition, the results confirm a strong link
from health capital to education capital. The (iv) Health capital equation
elasticity of education capital in response to Health spending has a positive and signifi-
an improvement in health capital is about 1.3, cant impact on health capital, but its effects
which indicates that a drop in the child mortal- are only contemporaneous. An increase in
ity rate by 1 percentage point is associated with health spending of 1 percentage point of GDP
an increase in the enrollment rate of 1.3 per- is associated with a rise in the under-5 child sur-
centage points. vival rate of 0.2 percentage point. But unlike
Income is found to have a positive and signif- education spending, lagged health spending
icant effect, consistent with notion that improv- has no further effect on health indicators.
ing human capital can catalyze a virtuous cycle Consistent with the literature, income is nega-
between growth and human development tively associated with child mortality rates,
(Ranis & Ramirez, 2000). The implications of whereas fertility rates raise them. In contrast, a
 SOCIAL SPENDING, HUMAN CAPITAL, AND GROWTH IN DEVELOPING COUNTRIES 1329

Table 5. Health equation-dependent variable: average annual under-5 child mortality rate
LSDVa Robust regression FGLS 2SLS
Current per capita income 9.425 5.805 16.422 8.615
(3.47)*** (4.83)*** (9.28)*** (1.97)*
Urbanization 0.457 0.326 0.116 0.972
(1.38) (2.24)** (1.55) (1.54)
Share of female students 3.287 1.255 3.054 4.164
(7.13)*** (6.15)*** (11.59)*** (4.59)***
Current health spending (in logs) 5.182 7.105 0.853 28.553
(1.68)* (5.20)*** (0.50) (1.51)
Low-governanceb * current health spending 5.157 3.803 0.092 13.034
(1.69)* (2.82)*** (0.05) (1.77)*
Fertility rate 12.160 8.358 17.024 11.750
(5.49)*** (8.53)*** (17.07)*** (3.00)***
Constant 208.100 128.591 275.175 254.196
(6.13)*** (8.57)*** (16.70)*** (4.03)***
Observations 267 255 230 188
R-squared 0.98 1.00
P-value for joint F-testc 0.000
P-value for overidentification 0.07
Absolute value of t statistics in parentheses.
a
Baseline regression.
b
The low-governance dummy is interacted with the spending term to compare the effects of spending in different
governance environment, and no additional governance level is included in the regression.
c
Health spending is instrumented by lagged spending and the democracy index to control for endogeneity. Similar to
previous regressions, democracy is likely to affect spending decisions but does not directly contribute to health status.
The validity of lagged spending, as an instrument, is also confirmed by the SIC tests, which fail to confirm a lagged
effect of health spending on health indicators. The joint F-test in the first stage fails to reject validity of these
instruments, but the overidentification test indicates that the instruments are only valid at the 90% confidence level.
*
Significant at 10%.
**
Significant at 5%.
***
Significant at 1%.

higher share of female students in school is asso- rect effects. The results are summarized in Ta-
ciated with lower child mortality rates. Although ble 6.
education capital has no direct effect on health The impact of education and health capital
capital, it may nevertheless affect health status on growth varies for different income and re-
through these indirect channels. As argued by gional country groups. The impact of educa-
Barro and Sala-i-Martin (1995), Ranis and tion capital on growth is more pronounced
Ramirez (2000) and Boozer et al. (2003), better in low-income countries, where an increase
education indicators are associated with lower of 1 percentage point in the primary and sec-
fertility rates. Gupta et al. (2003) also find a ondary enrollment rates is associated with 0.1
robust relationship between a mother’s educa- percentage point increase in per capita GDP
tion level and child mortality rates. growth. This effect is 1.5 times that in mid-
Health spending is only effective, however, in dle-income countries. Geographically, the im-
an environment where governance is adequate. pact is highest in sub-Saharan Africa and
In a poor-governance environment, higher pub- lowest in Eastern Europe and Central Asia.
lic outlays for health have no effect on child As for the impact of health capital on growth,
mortality rates. the impact of a unit reduction in the under-5
child mortality rate is smaller in low-income
(v) Short-run effect decomposition and country countries than middle-income countries. How-
group differences ever, this masks the fact that in terms of the
Following single equation estimations, we elasticities, the impact of reducing child mor-
carried out a decomposition of direct and indi- tality is about 4 times as large when compared
1330 WORLD DEVELOPMENT

Table 6. Summary of total-effect coefficients by country groupa

Dependent variable (y) Independent variable (x)
Primary and secondary Under-5 child Current Current Lagged
enrollment survival rateb education health education
spending spending spending
(in percent of the (per thousand (one percent
school age group) live births) of GDP)
All developing countries
Real per capita GDP growth 0.09 0.17 0.54 0.39 0.25
Primary and secondary 0.00 0.19 6.00 0.44 2.72
enrollment rate
Under-5 child survivalb 0.00 0.00 0.00 2.36 0.00
Low-income countries
Real per capita GDP growth 0.12 0.11 0.59 0.48 0.27
Primary and secondary 0.00 0.19 6.54 0.54 2.96
enrollment rate
Under-5 child survivalb 0.00 0.00 0.00 2.88 0.00
Middle-income countries
Real per capita GDP growth 0.08 0.29 0.50 0.34 0.23
Primary and secondary 0.00 0.19 5.54 0.39 2.51
enrollment rate
Under-5 child survivalb 0.00 0.00 0.00 2.07 0.00
Asia
Real per capita GDP growth 0.08 0.24 0.53 0.48 0.24
Primary and secondary 0.00 0.19 5.84 0.54 2.64
enrollment rate
Under-5 child survivalb 0.00 0.00 0.00 2.88 0.00
Eastern Europe & Central Asia
Real per capita GDP growth 0.06 0.46 0.44 0.24 0.20
Primary and secondary 0.00 0.19 4.80 0.27 2.17
enrollment rate
Under-5 child survival 0.00 0.00 0.00 1.44 0.00
Latin America & The Caribbean
Real per capita GDP growth 0.08 0.27 0.56 0.34 0.25
Primary and secondary 0.00 0.19 6.17 0.39 2.79
enrollment rate
Under-5 child survivalb 0.00 0.00 0.00 2.07 0.00
Middle East & North Africa
Real per capita GDP growth 0.08 0.22 0.47 0.48 0.21
Primary and secondary 0.00 0.19 5.17 0.54 2.34
enrollment rate
Under-5 child survivalb 0.00 0.00 0.00 2.88 0.00
Sub-Saharan Africa
Real per capita GDP growth 0.13 0.10 0.55 0.48 0.25
Primary and secondary 0.00 0.19 6.02 0.54 2.72
enrollment rate
Under-5 child survivalb 0.00 0.00 0.00 2.88 0.00
Source: Authors’ calculation.
a
The coefficient measures dy/dx, i.e., the impact of one percentage point change in column variables on row
variables. These coefficients are calculated using estimates from the baseline model. If the original coefficients in the
regressions are in y/ln x form, they are transformed into linear coefficients of dy/dx derived from the equation dy/
d ln x = dy/(dx/x) = (dy/dx)/x.
b
Under-5 child survival rate is used to facilitate the interpretation of a positive effect, and the increase in the under-5
child survival rate matches exactly the reduction in the under-5 child mortality rate.
 SOCIAL SPENDING, HUMAN CAPITAL, AND GROWTH IN DEVELOPING COUNTRIES 1331

Table 7. Simulation on the impact on growth and social indicators from various policy interventionsa
Baseline T+1 T+2 T+3 T+1 T+2 T+3
End of period value Absolute change
Increase education spending by 1% of GDPb
Primary and secondary enrollment 154.0 160.0 167.2 172.5 6.0 7.2 5.4
Implied net primary enrollmentc 89.7 93.2 97.5 100 3.5 4.3 1.0
Per capita GDP growth (in percent) 1.3 1.8 2.0 2.7 0.5 0.2 0.7
Implied headcount ratiod 100.0 95.2 89.9 82.9 4.8 5.3 7.0
Under-5 child mortality (per 1000 live births) 76.0 76.0 70.9 64.7 0.0 5.1 6.2
Increase health spending by 1% of GDPb
Under-5 child mortality (per 1000 live births) 76.0 73.6 69.9 69.9 2.4 3.7 0.0
Primary and secondary enrollment 154.0 154.4 157.7 157.7 0.4 3.3 0.0
Implied net primary enrollmentc 89.7 90.0 91.8 91.9 0.3 1.9 0.0
Per capita GDP growth (in percent) 1.3 1.7 1.7 1.7 0.4 0.0 0.1
Implied headcount ratiod 100.0 95.6 91.2 86.7 4.4 4.4 4.5
Improving governance to be above the world average
Under-5 child mortality (per 1000 live births) 76.0 70.8 70.8 69.7 5.2 0.0 1.2
Primary and secondary enrollment 154.0 159.3 172.4 173.5 5.3 13.2 1.0
Implied net primary enrollmentc 89.7 92.8 100 100 3.1 7.7 0.6
Per capita GDP growth (in percent) 1.3 2.9 2.7 2.8 1.6 0.2 0.1
Implied headcount ratiod 100.0 92.5 85.5 78.2 7.5 7.0 7.3
Reduce inflation by 10 percentage points (e.g., from 40% to 30%)
Per capita GDP growth (in percent) 1.3 1.8 1.8 1.8 0.5 0.0 0.0
Implied headcount ratiod 100.0 95.2 90.5 85.7 4.8 4.8 4.8
Marginal change from high inflation to below 20%
Per capita GDP growth (in percent) 1.3 1.4 1.4 1.4 0.1 0.0 0.0
Implied headcount ratiod 100.0 96.4 92.9 89.3 3.6 3.6 3.6
Reduce fiscal deficit by 1% of GDPe
Per capita GDP growth (in percent) 1.3 1.4 1.4 1.4 0.1 0 0
Implied headcount ratiod 100.0 96.4 92.9 89.3 3.6 3.6 3.6
Sources: World Development Indicators database 2003, World Economic Outlook database 2003, country authorities
and authors’ calculations.
a
Simulations are based on coefficients obtained from the baseline model. The initial baseline values are averages for
developing countries in the sample, and the policy interventions are assumed to take place at the end of baseline year.
In addition, the increased social spending is assumed to be maintained over the simulation period. Furthermore, the
elasticity estimates from the recursive model are assumed to hold for the period covered by the simulation while other
variables stay at the same level. The effects of higher incomes on education and health capital are also captured.
b
The value of 1% of GDP of all developing countries was about $60 billion in 2000. The rise in spending is assumed
to be deficit-neutral, and may be achieved through grants, increased revenue collection, and/or shifting of funding
from unproductive outlays.
c
The implied net primary enrollment rate is listed to relate to the explicit targets on net enrollment rate in MDGs
and calculated under the assumption that the ratio of the net primary enrollment rate over the primary and secondary
enrollment rate remains what it was in 2000.
d
The initial value of the headcount ratio is standardized to be 100, and the implied headcount ratios are the
projected end-of-period values after incorporating the incremental poverty reduction due to growth acceleration. The
impact on the poverty headcount ratio is calculated by the growth elasticity of poverty of 2.6, an average based on
estimates from Ravallion and Chen (1997), Bruno et al. (1998) and Adams (2003).
e
The impact on growth is estimated at the sample mean for deficit at 4% of GDP, and the reduction is assumed to be
achieved through grants, increased revenue collection or cut in unproductive spending without affecting social
spending.

to middle-income countries, owing to high In sum, the most salient results regarding the
initial levels of mortality in the poorest interface between social spending, human capi-
countries. tal, and growth are as follows:
1332 WORLD DEVELOPMENT

Table 8. Short-run total effects decomposition: selected endogenous variablesa

Real per capita Investment Gross Under-5 child
GDP growth ratio enrollment rate survival rate
Total effects
Real per capita GDP growth 0.00 0.70 0.15 0.17
Investment ratio 0.00 0.00 0.00 0.04
Primary and secondary enrollment rate 0.00 0.00 0.00 0.19
Under-5 Child survival rate 0.00 0.00 0.00 0.00
Direct effects
Real per capita GDP growth 0.00 0.70 0.15 0.12
Investment ratio 0.00 0.00 0.00 0.04
Primary and secondary enrollment rate 0.00 0.00 0.00 0.19
Under-5 Child survival rate 0.00 0.00 0.00 0.00
Indirect effects
Real per capita GDP growth 0.00 0.00 0.00 0.05
Investment ratio 0.00 0.00 0.00 0.00
Primary and secondary enrollment rate 0.00 0.00 0.00 0.00
Under-5 Child survival rate 0.00 0.00 0.00 0.00
a
The coefficient measures dy/dx, i.e., the impact of one percentage point change in column variables on row
variables. These coefficients are calculated using estimates from the baseline model. If the original coefficients in the
regressions are in y/ln x form, they are transformed into linear coefficients of dy/dx derived from the equation dy/
d ln x = dy/(dx/x) = (dy/dx)/x.

• Both education capital and health capital is further analyzed in Section 5 to illustrate
positively contribute to output growth, but the role of lagged effects and interactions
through slightly different routes. Unlike between the key variables.
health capital, both the stock and flow of • Education and health capital have strong
education capital affect growth. The effects interlinkages. Health capital contributes to
from the two channels are of similar magni- the accumulation of education capital, with
tude, indicating that education capital pro- an elasticity of about 1.3. Moreover, gender
motes growth not simply as an input but equality in the education system has a bear-
also through level effects. In contrast, the ing on health capital.
only direct effect of health capital on • Governance has significant direct effects
growth is through changes to the stock of on the nexus between social spending and
this capital. The stock of health capital social indicators, and poor governance
does, however, affect growth indirectly via reduces growth mainly through its impact
investment. Furthermore, education and on human capital and investment. Countries
health spending have a positive effect on with poor governance, other things being
indicators of education and health capital. equal, have growth that is about 1.6 percent-
Tables 8 and 9. Thus, social spending also age points lower per year than other coun-
helps boost growth via its effects on human tries. 27 Similarly, poor governance is
capital. 26 associated with an investment rate that is 2
• Education spending has both an immedi- percentage points lower. As shown in Table
ate and a lagged effect on education capital, 9, the effect of governance on growth is
while health spending has only a contempo- transmitted through indirect channels via
raneous effect on health capital. About two- its effects on human capital and investment.
thirds of the direct effect of education spend- As shown in Tables 4 and 5, health spending
ing are realized in the first five years, with has no effect on health indicators in coun-
the remainder realized over the next five- tries suffering from poor governance. Simi-
year span. Furthermore, the higher income larly, the impact of education spending is
brought about by growth can generate addi- also much reduced in such countries. The
tional improvements in education and important role of governance—which had
health capital. The aggregate growth effect not always been incorporated in previous
 SOCIAL SPENDING, HUMAN CAPITAL, AND GROWTH IN DEVELOPING COUNTRIES
Table 9. Short-run total effects decomposition: selected exogenous variables
Population Catch-up Trade Changes in Fiscal Low- High- Inflation Low- Current Gender Current Lagged Health
growth variable openness terms of balance deficit inflation rate governance income equality education education spending
trade fiscal dummy dummy spending spending
balance
Total effects
Real per capita 1.93 15.90 0.15 0.04 0.50 0.00 0.07 0.05 1.65 2.30 0.67 0.54 0.25 0.39
GDP growth
Investment ratio 0.00 0.00 0.07 0.06 0.00 0.00 2.20 0.00 1.69 0.33 0.12 0.00 0.00 0.08
Primary and 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.90 10.00 2.03 6.00 2.72 0.44
secondary
enrollment rate
Under-5 child 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 2.34 9.43 3.29 0.00 0.00 2.36
survival rate
Direct effects
Real per capita 1.93 15.90 0.10 0.00 0.50 0.00 0.06 0.05 0.00 0.00 0.00 0.00 0.00 0.00
GDP growth
Investment ratio 0.00 0.00 0.07 0.06 0.00 0.00 2.20 0.00 1.61 0.00 0.00 0.00 0.00 0.00
Primary and 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.46 8.23 1.42 6.00 2.72 0.00
secondary
enrollment rate
Under-5 child 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 2.34 9.43 3.29 0.00 0.00 2.36
survival rate
Indirect effects
Real per capita 0.00 0.00 0.05 0.04 0.00 0.00 0.02 0.00 1.65 2.30 0.67 0.54 0.25 0.39
GDP growth
Investment ratio 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.08 0.33 0.12 0.00 0.00 0.08
Primary and 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.44 1.77 0.62 0.00 0.00 0.44
secondary
enrollment rate
Under-5 child 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
survival rate

1333
1334 WORLD DEVELOPMENT

research—could help explain why some ear- tional primary and secondary schooling is asso-
lier studies have found a generally weak rela- ciated with an increase in growth ranging
tionship between social spending and social between 0.4 and 0.8 percentage points per year,
indicators. depending on the country group. 30 Low-in-
• Gender equality influences both education come countries and sub-Saharan African coun-
and health indicators. The share of female tries are on the high end, while Eastern Europe
students in primary and secondary schools and Central Asia countries are on the low end.
is positively and significantly associated with The results are broadly consistent under a
education capital and negatively and signifi- number of different estimators controlling for
cantly associated with child mortality. A 1 outliers (robust regression) and endogeneity
percentage point increase (e.g., from 40% to (two-stage least squares and system GMM). 31
41%) in the female share of enrolled students However, there are a few results that suggest
is associated with an increase of 2 percentage caution in interpreting the baseline coefficients.
points in the enrollment rate, and a reduction In particular, in the heath indicator equation,
of 0.3 percentage point in the child mortality the FGLS and 2SLS results show that the coef-
rate. This strongly supports the view that in ficient of spending on health outcomes is not
developing countries, gender equality can significant, even though the 2SLS also yields a
improve education and health outcomes. negative coefficient on the interaction term. In
• Income levels matter for social indicators. addition, in the growth equation, the 2SLS
GDP per capita is robustly and positively coefficients for education capital and changes
correlated with both education and health in education capital are insignificant. This
capital. This indicates that higher income could be attributed to the sharp rise in standard
levels and greater human capital reinforce errors with 2SLS, the weakness of the instru-
each other and contribute to a virtuous cir- ments, and the significant drop in the number
cle of growth and higher human capital. of observations (around 20–30%), which occurs
with the use of lagged variables as instruments.
(b) Robustness checks Solving the growth, investment, education,
and health equations simultaneously in a Seem-
Our estimates generate elasticities that are ingly Unrelated Regression framework does
similar to those found in earlier research. The not lead to qualitatively different results, as
econometric results imply that the elasticity of the Breusch-Pagan test fails to reject the inde-
education indicators with respect to education pendence of the residuals from the four equa-
spending is 0.25. 28 This is close to the upper tions. We also compared the recursive system
range of similar estimates on the returns to pri- with other specifications, including a full-
mary and secondary education in the literature fledged system of equations, and find that our
(Baldacci et al., 2003; Psacharopoulos, 1994; model dominates other specifications based on
Psacharopoulos & Patrinos, 2002). The elastic- the Schwartz Information Criteria.
ity of health indicators with respect to changes The estimates on governance and inflation are
in public health outlays is about 0.05, which is broadly consistent with the literature. For exam-
also consistent with the literature (Gupta et al., ple, Mauro (1998) finds that reducing corruption
2002b; Schultz, 1999). by one standard deviation is associated with an
The coefficients linking human capital indica- increase in growth of 0.5 percentage point a year,
tors and growth are also consistent with the lit- while our results indicate that a discrete jump
erature. On health capital and growth, Bloom from a low-governance to a high-governance
and Canning (2003) and Bloom et al. (2004) country is associated with an increase in the
find elasticities of 1.3 and 1.9, respectively, growth rate of 1.1 percentage points. 32 Fischer
compared with our estimates ranging from 1.4 (1993) and Burdekin et al. (2004) find that lower-
to 3.2 for different country groups. 29 On edu- ing inflation by 10 percentage points raises
cation capital and growth, Levine and Renelt growth by 0.4 and 0.8 33 percentage points,
(1992), Barro (1996a), Bassanini and Scarpetta respectively, while our results indicates that the
(2001), Bils and Klenow (2000) and Sianesi and same inflation reduction is associated with high-
Van Reenen (2003) estimate that an additional er growth of 0.5 percentage point. For the bud-
year of schooling raises the growth rate by 0.3– get deficit, we find a weaker result than most of
3 percentage points per year. Our estimates, the literature. Gupta, Clements, Baldacci, and
based on the sum of primary and secondary Mulas-Granados (2005) find that lowering the
enrollment rate, indicate that one year of addi- deficit by 1 percentage point of GDP is associ-
 SOCIAL SPENDING, HUMAN CAPITAL, AND GROWTH IN DEVELOPING COUNTRIES 1335

ated with an increase in per capita GDP growth deaths per thousand from 2000 to 2015, which
of 0.3–0.5 percentage point, while this paper is still short of the MDG target of 62 deaths per
shows a contemporaneous increase in per capita 1,000 live births (one-third of the 1990 level). 38
growth of 0.01 percentage point. In the simula- Moreover, net school enrollment would rise by
tion of policy interventions, however (see be- about 2 percentage points (i.e., from 90 to 92),
low), we find a much stronger effect of deficit and growth rate would increase by a total of 0.5
reduction than implied by the direct effect in percentage points. Overall, such spending in-
the growth equation alone. creases would greatly help countries move to-
ward the MDGs, but would likely be
insufficient to achieve it in all regions.
5. SIMULATION 34 OF THE IMPACT OF Improve governance to a level above the sam-
POLICY INTERVENTIONS ple average: A change in the governance index
from lower- to higher-than-average is associ-
As shown in previous sections, social indica- ated with an immediate reduction of 0.5 per-
tors and growth are affected by many interre- centage point in the child mortality rate, an
lated factors. On the basis of the empirical increase of 6 percentage points in the primary
estimates presented above, simulations can be and secondary enrollment rate, and a rise of
conducted to assess the impact of different pol- 1.6 percentage points in per capita GDP
icy interventions for improving social indica- growth. Therefore, the positive impact is com-
tors, economic growth, and the poverty parable to an increase in education spending
headcount. As such, these simulations can pro- of 1% of GDP.
vide input into the debate over the effectiveness Reduce inflation: The growth effects of lower
of different policies in helping countries reach inflation (and hence its effects on poverty) are
the MDGs. The simulations assess the impact substantial. Cutting the rate of inflation by 10
of the following policy interventions: (1) an in- percentage points (e.g., from 40% to 30%) is
crease in education spending of 1% of GDP; (2) associated with a 0.5 percentage point increase
an increase in health spending of 1% of GDP; in annual growth. Moreover, countries that re-
(3) an improvement in governance; (4) a reduc- duce their rate of inflation to below 20% would
tion in the budget deficit of 1% of GDP; and (5) also experience an additional fillip to growth of
a reduction in inflation of 10 percentage points. 0.1 percentage point per year.
In measuring the outcome, the gross enrollment Reduce fiscal deficit: Improving the fiscal bal-
rate for primary and secondary schools is con- ance by 1 percentage point of GDP is associ-
verted to a net enrollment rate 35 in the simula- ated with an increase in per capita GDP
tion, as the latter is an explicit target in MDGs. growth by 0.01 percentage point. 39 However,
Each of the simulations assumes that the pol- it does not bring additional lagged positive ef-
icy environment remains unchanged. This im- fects as in the case of social spending. Further-
plies, for example, that increases in public more, as shown in Table 2 and consistent with
spending do not result in an increase in the Gupta et al. (2005), these effects disappear in a
budget deficit or a deterioration in governance. low-deficit environment.
Finally, the simulation is undertaken for an
average developing country in order to estab-
lish a benchmark to inform applied policy anal- 6. CONCLUSION AND POLICY
ysis (Table 7). IMPLICATIONS
Increase education and health spending by 1%
of GDP 36: Raising average education spending The paper finds that a number of policy inter-
by 1% of developing country GDP (and main- ventions could be effective in moving countries
taining it at that higher level) can raise the net toward the MDGs. Both education and health
enrollment rate from 90% to 99% and lower spending have a positive and significant direct
child mortality rate from 76 to 65 per thousand impact on the accumulation of education and
from 2000 to 2015. In addition, the per capita health capital, and a positive and significant
growth rate would rise by about 0.5 percentage indirect impact on growth. An increase in educa-
point per year on average, which could reduce tion spending of 1 percentage point of GDP is
the initial poverty headcount by about 17% associated with 3 more years of schooling on
over the 15-year period. 37 Raising average average and raises the annual growth of GDP
health spending by 1% of GDP would reduce per capita by 1.4 percentage points in 15 years.
the under-5 child mortality rate from 76 to 70 Similarly, an increase in health spending of 1
1336 WORLD DEVELOPMENT

percentage point of GDP is associated with an needed to assess the effects of different compo-
increase of 0.6 percentage point in the under-5 nents of social spending (e.g., primary educa-
child survival rate and a rise of 0.5 percentage tion, secondary education spending).
point in annual per capita GDP growth. Other policy interventions may also achieve
There is a significant time lag between in- improvement of a similar size in social indica-
creases in education spending and the realiza- tors and growth. In particular, strengthening
tion of their full effects on social indicators governance can have a strong payoff for social
and growth. Two-thirds of the direct impact indicators as well as for growth. Therefore,
of education spending is felt within five years, reducing corruption and increasing the
but the full impact materializes with a signifi- accountability for public spending are no less
cant time lag of 10–15 years. Such a lag needs important than increasing spending. In addi-
to be kept in mind when designing policy inter- tion, macroeconomic policies, such as reducing
ventions. The impact of health spending, how- inflation and improving fiscal balances, also
ever, is immediate. The positive effects of both have a positive effect on growth and, in turn,
education and health spending are strongly on the poverty headcount.
influenced by the quality of governance. In The results have a number of implications for
countries suffering from poor governance, the poverty reduction strategies aimed at meeting
positive effects of increased spending on educa- the MDGs. Given the importance of different
tion is reduced, and those of higher health policy interventions, efforts to meet the MDGs
spending can be completely negated. will need to be wide-ranging and include
There are substantial differences in the effects strengthening the macroeconomic environment
of social spending on social indicators and and governance. Relative to the significant cost
growth among different country groups. The po- of raising spending, the moderate effects of so-
sitive effects are the highest in low-income coun- cial spending on indicators also confirm the
tries and sub-Saharan Africa. This supports the important role of reforms aimed at improving
view that social spending can be more effective in the efficiency and targeting of these outlays.
such countries in achieving MDGs, as the mar- Furthermore, while improving human capital
ginal returns to social spending tend to decline will have a salutary effect on growth, it will be
for countries with high levels of social outlays. far from a panacea for unlocking the more ro-
However, given the wide variation of country bust expansion in economic activity needed to
circumstances and the nonlinearity in the effects achieve the MDGs. As such, additional re-
as shown in this paper, the results should be ta- search is needed to address the key policy inter-
ken with caution when applied to policy analysis ventions needed to achieve rapid economic
at the country level. Additional research is also growth.

NOTES

1. These MDGs, endorsed by heads of states from 3. These studies find that the effect of social spending
189 member countries of the United Nations, on human development indicators is stronger in cross-
encompass measurable targets in seven areas for sectional samples than when the time dimension is also
the developing world: (i) eradicate extreme poverty added.
and hunger; (ii) achieve universal primary education;
(iii) promote gender equality and empower women; 4. The preferred model results in a system of four
(iv) reduce child mortality; (v) improve maternal equations where the coefficient matrix is triangular (see
health; (vi) combat HIV/AIDS, malaria, and other footnote 14). See also Duncan (1975) on the advantage
diseases; and (vii) ensure environmental sustainability. of using a recursive system of equations.
Therefore, progress on the majority of these goals is
to be measured on the basis of education and health 5. Islam (2000) notes that GMM and IV estimators
indicators. may fail because their theoretical performance depends
on the use of optimal weighting matrix, which in practice
is estimated and contains noises present in the data.
2. On market failures and the rationale for public Amemiya (1967) demonstrates that LSDV is consistent
sector involvement in education and health, see World and asymptotically equivalent to the Maximum Likeli-
Bank (1993, 1995), Psacharopoulos (1994) and Sachs hood Estimator in long time series, although in theory,
et al. (2003). the presence of endogenous variables on the right-hand
 SOCIAL SPENDING, HUMAN CAPITAL, AND GROWTH IN DEVELOPING COUNTRIES 1337

side may lead to some inconsistencies as N approaches 15. The dummy takes the value of one if the deficit
infinity. (fiscal balance) is lower (higher) than the sample mean,
or zero if otherwise.
6. The instruments used in each regression and the
statistics from the Sargan tests on overidentifying 16. The appropriate lag structure was selected by
restrictions and the joint-F tests of the first-stage comparing Schwartz Information Criteria (SIC) values
regressions are listed in Tables 2–5. The justification with different lags in a distributed lag model, and one lag
for the choice of instruments in each regression is was chosen. This implies that the impact of education
described in the notes of these tables. Temple (1999) and spending on social indicators, during a given five-year
Bloom et al. (2004) explain that lagged variables may be period, will take additional time (that is, another five
used as instruments when better ones are not available, years) to be fully realized.
but the choice should be supported by empirical tests,
which we provided in the regression tables.
17. Neither current nor lagged education capital indi-
7. See Blundell and Bond (1998) for further details on cator was found to have a significant effect. We assume
the system GMM estimator for dynamic panel regres- that education spending affects health capital only via
sions. The estimation was implemented using the education capital. As such, education spending was not
xtabond2 routine in STATA. included in either health equation.

8. This estimator reduces weights given to outliers in 18. The primary and secondary enrollment rate, as a
estimation. See Hamilton (1991) for details on the proxy of the aggregate education level of the population,
weighting method. were not found to have any significant impact on health.
This variable was thus dropped from the health equa-
9. Temple (1999) and Krueger and Lindahl (2001) tion. This allows the four equations on growth, invest-
highlight the prevalence of such errors in human capital ment, education, and health to be estimated in a
indicators and the significant impact they can have on recursive system, since the coefficient matrix is triangu-
empirical estimates of the human capital-growth rela- lar. Further support for this specification is given in the
tionship. Our choice is also dictated by the limited section on robustness checks.
availability of enrollment rates at 5-year intervals for
most developing countries. 19. In the case of education spending, on the other
hand, the empirical results indicate that the full effects
10. The derivation of the growth model is based on are not captured within a five-year time period. This
Bassanini and Scarpetta (2001), but human capital is could be attributed to the fact that household demand
further decomposed into education and health capital to for education reacts with a lag as spending (and quality)
address potential interactions of the two sectors. increase and that human capital obtained from educa-
tion is cumulative over time.
11. For countries in our sample with data on years of
schooling (drawn from Barro & Lee (2000)), the 20. The indirect effects can be calculated by subtracting
correlation coefficient is 0.7. the direct-effect matrices B and C from the respective
total-effect matrix P. This decomposition is based on the
12. While the enrollment rate mixes together stocks assumption that current and lagged spending increases
and flows, it can nevertheless capture the overall effects are independent and lagged education and health capital
of education capital on growth (Gemmel, 1996). Also, are predetermined. Also, the decomposition focuses on
following Barro (1996a, 1996b), we adopt a logarithmic the short-run relationship between the current stock and
functional form for education capital, which allows for current growth through the recursive system. Long-run
the declining impact of higher enrollment rates on effects that fully incorporate lagged effects and interac-
growth. tions of the key variables are analyzed in the simulations
discussed in Section 5.
13. See Bassanini and Scarpetta (2001) for details.
21. See Barro and Lee (2000).
14. Gyimah-Brempong and Wilson (2004) use health
spending instead of the actual changes in health capital.
22. For a fuller description of the data, see Appendix A.
This approach is problematic, however, as it assumes a
positive relationship between health spending and health
outcomes, without empirically testing for it. 23. See detailed description in the notes to Table 1.
1338 WORLD DEVELOPMENT

24. The results from alternative specifications (used for country correlation we found (0.6–0.8) between the
the robustness tests) are also reported in the tables. ICRG index and other corruption indices (i.e., the index
These are discussed in Section 4(b). of Transparency International and the World Bank
governance indicator (see Kaufmann, Kraay, & Mast-
25. Following Bils and Klenow (2000), an increase of 6 ruzzi, 2003)).
percentage points in the primary and secondary enroll-
ment is equivalent to one more year of schooling. 33. This is a point estimate for the impact of inflation
on growth in a developing country when the inflation
26. An assessment of the direct and indirect effects is rate is between 3% and 50%.
given in Tables 8 and 9.
34. The simulation is a partial equilibrium analysis
27. The effect captures the impact of a discrete change focusing on the interaction of key variables. It assumes
in governance from lower-than-average to higher-than- that elasticities obtained by the recursive system persist
average. over time. See the notes in Table 7 for details.

28. This figure accounts only for the direct impact of a 35. The implied net primary enrollment rate is calcu-
rise in education spending in two periods. If the lated under the assumption that the ratio of the net
intertemporal effects are fully taken into account, the primary enrollment rate over the primary and secondary
estimated elasticity is 45%. enrollment rate remains what it was in 2000.

29. Assuming the average real per capita growth rate to

be 1.3% per year and the average life expectancy to be 63 36. We assume that the rise in spending is achieved in
in calculating elasticities e = (dg/g)/(dx/x). time t and then maintained at the new level. The increase
in spending is assumed to come from sources that do not
increase the budget deficit (e.g., external grants, domes-
30. Following Bils and Klenow (2000), we consider six
tic revenues, or reductions in other public spending).
times the effect on the average enrollment ratio as
equivalent to one more year of average schooling to
obtain this estimate. 37. See notes in Table 7 for details on the projection
method.
31. The model is also tested by using alternative
education and health indicators, such as infant mortality 38. Based on the sample average of the dataset used in
and primary schooling, and the results are found to be this paper.
qualitatively consistent.

39. We assume that the cut is achieved through grants,

32. We did not find alternative governance indicators
increased revenue collection or reduction of unproduc-
that had sufficient time-series data for use in our model.
tive spending, with social spending on education and
Using alternative corruption measures would most likely
health unchanged.
generate similar results, given the high degree of cross-

REFERENCES

Abed, G. T., & Gupta, S. (Eds.) (2002). Governance, Baldacci, E., Guin-Sui, M. T., & de Mello, L. (2003).
corruption, and economic performance. Washington, More on the effectiveness of public spending on
DC: International Monetary Fund. health care and education: A covariance structure
Adams, R. (2003). Economic growth, inequality, and model. Journal of International Development, 15(6),
poverty: Findings from a new data set. World Bank 709–725.
Policy Research Working Paper No. 2922. Washing- Baldacci, E., Hillman, A., & Kojo, N. (2004). Growth,
ton, DC: World Bank. governance, and fiscal policy transmission channels
Amemiya, T. (1967). A note on the estimation of in low-income countries. European Journal of Polit-
Balestra-Nerlove models. Technical Report No. 4, ical Economy, 20(3), 517–549.
Institute for Mathematical Studies in Social Sciences Baltagi, B. (2001). The econometrics of panel data (2nd
of Stanford University. ed.). New York: John Wiley & Sons.
Anand, S., & Martin, R. (1993). Human development in Barro, R. (1996a). Determinants of economic growth: A
poor countries: On the role of private incomes and cross-country empirical study. NBER Working
public services. Journal of Economic Perspective, 7(1), Paper No. 5968. Cambridge, MA: National Bureau
133–150. of Economic Research.
 SOCIAL SPENDING, HUMAN CAPITAL, AND GROWTH IN DEVELOPING COUNTRIES 1339

Barro, R. (1996b). Health, human, capital and economic Duncan, O. D. (1975). Introduction to structural equa-
growth. Paper for the Program on Public Policy and tions models. New York: Academic Press.
Health, Pan American Health Organization and Filmer, D., & Pritchett, L. (1997). Child mortality and
World Health Organization. Washington, DC: Pan public spending on health: How much does money
American Health Organization. matter? World Bank Policy Research Working Paper
Barro, R., & Lee, J. (2000). International data on No. 1864. Washington, DC: World Bank.
educational attainment: Updates and implications. Filmer, D., Hammer, J., & Pritchett, L. (1998). Health
Center for International Development Working policy in poor countries: Weak links in the chain.
Paper No. 42. Cambridge, MA: Harvard University. World Bank Policy Research Working Paper No.
Barro, R., & Sala-i-Martin, X. (1995). Economic growth. 1874. Washington, DC: World Bank.
New York: McGraw-Hill. Fischer, S. (1993). Role of macroeconomic factors in
Bassanini, A., & Scarpetta, S. (2001). Does human growth. Journal of Monetary Economics, 32(3),
capital matter for growth in OECD countries? 485–512.
Evidence from pooled mean-group estimates. Eco- Flug, K., Spilimbergo, A., & Wachtenheim, E. (1998).
nomics Department Working Paper No. 282. Paris, Investment in education: Do economic volatility and
France: OECD. credit constraints matter? Journal of Development
Benhabib, J., & Spiegel, M. (1994). The role of human Economics, 55(2), 465–481.
capital in economic development: Evidence from Gemmel, N. (1996). Evaluating the impacts of human
aggregate cross-country data. Journal of Monetary capital stocks and accumulation on economic
Economics, 34(2), 143–173. growth: Some new evidence. Oxford Bulletin of
Bidani, B., & Ravallion, M. (1997). Decomposing social Economics and Statistics, 58, 9–28.
indicators using distributional data. Journal of Gupta, S., Clements, B., Baldacci, E., & Mulas-Grana-
Econometrics, 77(1), 125–139. dos, C. (2005). Fiscal policy, expenditure composi-
Bils, M., & Klenow, P. J. (2000). Does schooling cause tion, and growth. Journal of International Money and
growth? American Economic Review, 90(5), 1160– Finance, 24(3), 441–463.
1183. Gupta, S., Davoodi, H., & Alonso-Termé, R. (2002a).
Bloom, D., & Canning, D. (2003). The health and Does corruption affect income inequality and pov-
poverty of nations: From theory to practice. Journal erty? Economics of Governance, 3(1), 23–45.
of Human Development, 4(1), 47–71. Gupta, S., Verhoeven, M., & Tiongson, E. (2002b). The
Bloom, D., Canning, D., & Sevilla, J. (2004). The effect effectiveness of government spending on education
of health on economic growth: A production func- and health care in developing and transition econo-
tion approach. World Development, 32(1), 1–13. mies. European Journal of Political Economy, 18(4),
Blundell, R., & Bond, S. (1998). Initial conditions and 717–737, Also reprinted in Gupta, S., Clements, B.,
moment restrictions in dynamic panel data models. & Inchauste G. (Eds.) (2004). Helping countries
Journal of Econometrics, 87(1), 115–143. develop: The role of fiscal policy (pp. 184–212).
Boozer, M. A., Ranis, G., Stewart, F., & Suri, T. (2003). Washington, DC: International Monetary Fund.
Paths to success: The relationship between human Gupta, S., Verhoeven, M., & Tiongson, E. (2003). Public
development and economic growth. Economic spending on health care and the poor. Health
Growth Center Discussion Paper No. 874. New Economics, 12(8), 685–696.
Haven: Yale University. Gyimah-Brempong, K., & Wilson, M. (2004). Health
Bruno, M., Ravallion, M., & Squire, L. (1998). Equity human capital and economic growth in sub-Saharan
and growth in developing countries: Old and new African and OECD countries. The Quarterly Review
perspectives on the policy issues. In V. Tanzi, & K. of Economics and Finance, 44(2), 296–320.
Chu (Eds.), Income distribution and high growth Hamilton, L. C. (1991). How robust is robust regres-
(pp. 117–146). Cambridge, MA: MIT Press. sion? Stata Technical Bulletin, 2, 21–26, July.
Burdekin, R., Denzau, A. T., Keil, M. W., Sithiyot, T., Hausmann, R., Pritchett, L., & Rodrik, D. (2005).
& Willett, T. D. (2004). When does inflation hurt Growth accelerations. Journal of Economic Growth,
economic growth? Different nonlinearities for differ- 10(4), 303–329.
ent economies. Journal of Macroeconomics, 262(3), Hojman, D. E. (1996). Economic and other determi-
519–532. nants of infant and child mortality in small
Carrin, G., & Politi, C. (1996). Exploring the health developing countries: The case of Central America
impact of economic growth, poverty reduction, and and the Caribbean. Applied Economics, 28(3), 281–
public health expenditure. Macroeconomics, Health, 290.
and Development Series, No. 18. Geneva, Switzer- Internationnal Monetary Fund (2003). Word Economic
land: World Health Organization. Outlook database.
Coulombe, S., Tremblay, J., & Marchand, S. (2004). Islam, N. (2000). Small sample performance of dynamic
Literacy scores, human capital and growth across panel data estimators in estimating the growth-
fourteen OECD countries. International Adult Lit- convergence equation: A Monte Carlo study.
eracy Survey Monograph Series. Ottawa: Statistics Advances in Econometrics, 15, 317–339.
Canada. Jones, B., & Olken, B. (2005). The anatomy of start-stop
Deaton, A. (2004). Health in an age of globalization. growth. NBER Working Paper No. W11528. Cam-
Paper prepared for the Brookings Trade Forum. bridge, MA: National Bureau of Economic
Washington, DC: Brookings Institution. Research.
1340 WORLD DEVELOPMENT

Kaufmann, D., Kraay, A., & Mastruzzi, M. (2003). Ravallion, M., & Chen, S. (1997). What can new survey
Governance matters III: Governance indicators for data tell us about recent changes in distribution and
1996–2002. World Bank Policy Research Working poverty? World Bank Economic Review, 11(2),
Paper 3106. Washington, DC: World Bank. 357–382.
Krueger, A. B., & Lindahl, M. (2001). Education for Rodrik, D., Subramanian, A., & Trebbi, F. (2004).
growth: Why and for whom? Journal of Economic Institutions rule: The primacy of institutions over
Literature, 39(4), 1101–1136. geography and integration in economic development.
Levine, R., & Renelt, D. (1992). A sensitivity analysis of Journal of Economic Growth, 9(2), 131–165.
cross-country growth regressions. American Eco- Romer, P. (1986). Increasing returns and long-run
nomic Review, 82(4), 942–963. growth. Journal of Political Economy, 94(5),
Lucas, R. E. Jr., (1988). On the mechanic of economic 1002–1037.
development. Journal of Monetary Economics, 22(1), Romer, P. (1990). Endogenous technological change.
3–42. Journal of Political Economy, 98(5), S71–S102.
Mankiw, G., Romer, D., & Weil, D. (1992). A contri- Sachs, J., & Members of the Commission on Macro-
bution to the empirics of economic growth. Quarterly economics and Health (2003). Macroeconomics and
Journal of Economics, 107, 407–437. health: Investing in health for economic develop-
Mauro, P. (1998). Corruption and the composition of ment. Geneva, Switzerland: World Health Organiza-
government expenditure. Journal of Public Econom- tion.
ics, 69(2), 263–279. Sala-i-Martin, X. (1997). I just ran two million regres-
Mayer-Foulkes, D. (2003). Market failures in health and sions. American Economic Review, 87(2), 178–183.
education investment for Mexico’s youth. Human Schultz, T. P. (1993). Mortality decline in the low-
Capital and Economic Growth Working Papers. income world: Causes and consequence. Economic
Washington DC: Pan American Health Organiza- Growth Center Discussion Paper No. 681. New
tion. Haven, CT: Yale University.
Miguel, E., & Kremer, M. (2003). Worms: Identifying Schultz, T. P. (1999). Health and schooling investments
impacts on education and health in the presence of in Africa. Journal of Economic Perspectives, 13(3),
treatment externalities. Econometrica, 72(1), 67–88.
159–217. Sen, A. (1999). Development as freedom. New York:
Mingat, A., & Tan, J. (1992). Education in Asia: A Alfred A. Knopf Inc.
comparative study of cost and financing. Washington, Sianesi, B., & Van Reenen, J. (2003). The returns to
DC: World Bank. education: Macroeconomics. Journal of Economic
Mingat, A., & Tan, J. (1998). The mechanics of progress Surveys, 17(2), 157–200.
in education: Evidence from cross-country data. Squire, L. (1993). Fighting poverty. American Economic
World Bank Policy Research Working Paper No. Review, Papers and Proceedings, 83(2), 377–382.
2015. Washington, DC: World Bank. Strauss, J., & Thomas, D. (1998). Health, nutrition, and
Musgrove, P. (1996). Public and private roles in health: economic development. Journal of Economic Litera-
Theory and financing patterns. World Bank Discus- ture, 36(2), 766–817.
sion Paper No. 339. Washington, DC: World Bank. Summers, L. H. (1992). Investing in all the people:
Noss, A. (1991). Education and adjustment: A review of Educating women in developing countries. EDI
the literature. PREM Working Paper WPS 701. Seminar Paper No. 45. Washington, DC: World
Washington, DC: World Bank. Bank.
Political Risk Services (2004). International Country Temple, J. R. W. (1999). The new growth evidence.
Risk Guide database. Journal of Economic Literature, 37(1), 112–156.
Pritchett, L. (1996). Where has all the education gone? Thomas, D., & Frankeberg, E. (2002). Health, nutrition
World Bank Policy Research Working Paper No. and economic prosperity: A micro-economic per-
1581. Washington, DC: World Bank. spective. Bulletin of the World Health Organization,
Pritchett, L., & Summers, L. H. (1996). Wealthier is 80, 106–113.
healthier. Journal of Human Resources, 31(4), Thomas, D., Frankenberg, E., Friedman, J., Habicht, J.,
841–868. Hakimi, M., Ingwersen, N., et al. (2003). Causal
Psacharopoulos, G. (1994). Returns to investment in effect of health on labor market outcomes: Evidence
education: A global update. World Development, from a random-assignment iron supplementation
22(9), 1325–1343. intervention. Manuscript. Los Angeles, CA:
Psacharopoulos, G., & Patrinos, H. (2002). Returns to UCLA/California Center for Population Research.
investment in education: A further update. World World Bank (1993). World development report 1993:
Bank Policy Research Paper No. 2881. Washington, Investing in health. New York: Oxford University
DC: World Bank. Press.
Rajkumar, A., & Swaroop, V. (2002). Public spending World Bank (1995). Priorities and strategies for educa-
and outcomes: Does governance matter? World tion: A World Bank review. Washington, DC: World
Bank Working Papers No. 2840. Washington, DC: Bank.
World Bank. World Bank (2003). World development indicators.
Ranis, G., & Ramirez, A. (2000). Economic growth and Washington, DC: World Bank.
human development. World Development, 28(2), Woszmann, L. (2003). Specifying human capital. Journal
197–219. of Economic Surveys, 17(3), 239–270.
 SOCIAL SPENDING, HUMAN CAPITAL, AND GROWTH IN DEVELOPING COUNTRIES 1341

APPENDIX A. LIST OF COUNTRIES All data series are five-year averages of an-
INCLUDED IN THE SAMPLE AND nual data. Data on real per capita GDP, infla-
DATA SOURCES tion, investment, terms of trade, trade
openness, total government expenditure, and
fiscal balance are based on the World Economic
Outlook (WEO) database. In particular, real
The countries included in the study are Alba- GDP growth rates are calculated by the change
nia, Algeria, Angola, Argentina, Armenia, of GDP in constant prices; the rate of inflation
Azerbaijan, Bangladesh, Barbados, Belarus, is calculated by the rate of increase in the CPI
Belize, Benin, Bolivia, Bosnia and Herzegovina, index; investment, government expenditure,
Botswana, Brazil, Bulgaria, Burkina Faso, Bur- and fiscal balances are calculated as a ratio to
undi, Cambodia, Cameroon, Cape Verde, Cen- GDP at current prices; and trade openness is
tral African Republic, Chad, Chile, China, calculated by the value of total exports and im-
Colombia, Comoros, Democratic Republic of ports of goods and services as a share of GDP.
Congo, Republic of Congo, Costa Rica, Côte Data on child mortality, the school repetition
d’Ivoire, Croatia, Czech Republic, Djibouti, rate, ratio of female students in school, popula-
Dominican Republic, Ecuador, Egypt, El Sal- tion growth, the shares of the under-15 popula-
vador, Eritrea, Estonia, Ethiopia, Fiji, The tion and urban population, and fertility rates
Gambia, Ghana, Guatemala, Guinea, Guinea- are taken from the World Development Indica-
Bissau, Guyana, Haiti, Honduras, Hungary, tors (WDI) database; data on the enrollment
India, Indonesia, Islamic Republic of Iran, Ja- rate is taken from the 2000 Barro–Lee dataset;
maica, Jordan, Kazakhstan, Kenya, Kyrgyz and the enrollment rate is the sum of primary
Republic, Lao People’s Democratic Republic, and secondary gross enrollment rates. Child
Latvia, Lebanon, Lesotho, Libya, Lithuania, mortality and enrollment rates are not available
FYR Macedonia, Madagascar, Malawi, for all years, but five-year averages allow the
Malaysia, Maldives, Mali, Mauritania, Mauri- construction of a consistent panel dataset none-
tius, Mexico, Moldova, Mongolia, Morocco, theless. Data on education spending as a share
Mozambique, Myanmar, Namibia, Nepal, Nic- of GDP is taken from the WDI database, but
aragua, Niger, Nigeria, Oman, Pakistan, health spending as a share of GDP is taken
Panama, Papua New Guinea, Paraguay, from an IMF database that has better country
Peru, Philippines, Poland, Romania, Russia, coverage. Data in the IMF database is com-
Rwanda, Saudi Arabia, Senegal, Sierra Leone, piled from IMF staff reports and the IMF’s
Slovak Republic, Solomon Islands, South Afri- Government Financial Statistics. Data on gover-
ca, Sri Lanka, Sudan, Swaziland, Syrian Arab nance are calculated as the sum of the simple
Republic, Tajikistan, Tanzania, Thailand, annual averages of two indices on corruption
Togo, Trinidad and Tobago, Tunisia, Turkey, and democratic accountability, which are two
Uganda, Ukraine, Uruguay, Venezuela, Viet- components of the ICRG rating produced by
nam, Yemen, Zambia, Zimbabwe. the Political Risk Service Group.

Available online at www.sciencedirect.com