Journal of Development Economics 104 (2013) 107–122

Contents lists available at SciVerse ScienceDirect

Journal of Development Economics

journal homepage: www.elsevier.com/locate/devec

Inequality of opportunity and growth☆

Gustavo A. Marrero a, b,⁎, Juan G. Rodríguez c,1
a
Departamento de Análisis Económico, IUDR, Universidad de La Laguna, Spain
b
CAERP, Spain
c
Departamento de Fundamentos del Análisis Económico I, Universidad Complutense de Madrid, Spain

a r t i c l e i n f o a b s t r a c t

Article history: Theoretical and empirical studies exploring the effects of income inequality upon growth reach a disappointing
Received 8 May 2011 inconclusive result. This paper postulates that one reason for this ambiguity is that income inequality is actually a
Received in revised form 8 March 2013 composite measure of inequality of opportunity and inequality of effort. They may affect growth through oppo-
Accepted 15 May 2013
site channels, thus the relationship between inequality and growth depends on which component is larger. Using
the PSID database for the U.S. in 1970, 1980 and 1990 we find robust support for a negative relationship between
JEL classification:
D63
inequality of opportunity and growth, and a positive relationship between inequality of effort and growth.
E24 © 2013 Elsevier B.V. All rights reserved.
O15
O40

Keywords:
Income inequality
Inequality of opportunity
Economic growth

1. Introduction negative impact on investment in human capital accumulation (Galor

A surge of literature on the effect of income inequality on growth
suggests many channels through which inequality can affect growth.2
On the one hand, the classical approach emphasizes the accumulation
of savings (Bourguignon, 1981; Stiglitz, 1969), unobservable effort
(Mirrlees, 1971; Rebelo, 1991) or the investment project size (Barro,
2000) as some of the main routes through which inequality may enhance
growth. On the other hand, the modern approach has shown that in the
presence of credit market imperfections, inequality has a long-lasting
☆ We are grateful for the assistance with the data base from, Marta Huerta, Nidia García
and Elijah Baley. We also acknowledge the helpful comments and suggestions of two
anonymous referees, David Weil (Editor), François Bourguignon, John Roemer, Francisco
Ferreira, Branko Milanovic, Shlomo Yitzhaki, Rafael Salas, the audiences at the World
Bank (Washington DC), ECINEQ (Buenos Aires), XXXIV Simposio de la Asociación
Española de Economía (Valencia), XII Encuentro de Economía Aplicada (Madrid) and
FEDEA (Madrid). This paper has benefited from the financial support of the Fundación
Ramón Areces. The authors also acknowledge funding from the Ministerio de Ciencia e
Innovación under project reference ECO2010-17590. Responsibility for any error is the
authors' alone.
⁎ Corresponding author. Tel.: +34 922 317123.
E-mail addresses: gmarrero@ull.es (G.A. Marrero), juangabr@pdi.ucm.es
(J.G. Rodríguez).
1
Tel.: +34 91 3942515.
2 4
Surveys on this issue can be found in Bénabou (1996), Aghion et al. (1999), Bertola
et al. (2005) and Ehrhart (2009).
and Zeira, 1993) or entrepreneurial activity (Banerjee and Newman,
1993), and hence on economic growth.3 Thus, overall inequality would
affect growth positively or negatively depending on the channels that
dominate.
However, the vast empirical literature does not indicate that any
of these channels has a predominant influence. As a result, the empir-
ical relationship between inequality and growth is also ambiguous.4
This ambiguity tends to be justified through the quality and type of
data (Deininger and Squire, 1998; Panizza, 2002), the inconsistent
nature of inequality measures (Knowles, 2005), the type of inequality
index (Székeli, 2003), the econometric method (Forbes, 2000), the
model specification (Panizza, 2002) or the set of countries considered
and their degree of development (Barro, 2000). Moreover, as pointed out
by Partridge (1997 and 2005), Barro (2000), Bleaney and Nishizama
(2004) and Voitchovsky (2005), this ambiguity can be due to the fact
3
Alternative channels through which inequality can have a negative effect on
growth are: unproductive investments (Mason, 1988), demand patterns (Marshall,
1988), fertility (Galor and Zang, 1997; Kremer and Chen, 2002), domestic market size
(Murphy et al., 1989), political economy (Alesina and Perotti, 1994; Alesina and
Rodrik, 1994; Persson and Tabellini, 1994), and political instability (Gupta, 1990).
See Banerjee and Duflo (2003), among others, on the inconclusiveness of the cross-
country empirical literature on inequality and growth.

0304-3878/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.jdeveco.2013.05.004
108 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122
that income inequality has distinct offsetting avenues affecting subse-
quent growth in different ways. For instance, the variation in the rela-
tionship between inequality and growth with income level and the
presence of non-linearities could reflect these alternative offsetting
ways by altering economic incentives.5
Following this line of inquiry, we claim that an important reason
for this ambiguous relationship is that income inequality is actually
a composite measure of inequality of opportunity (IO) and inequality
of effort (IE).6 Following Roemer (1993) and Van de Gaer (1993), IO
refers to that inequality stemming from factors, called circumstances,
beyond the scope of individual responsibility, like race and socio-
economic background (i.e., proxy by parental education or wealth),
while IE defines the income inequality caused by individual responsi-
ble choices, like the number of hours worked or the occupational
choice. Roughly speaking, total inequality can be viewed as the result
of heterogeneity in circumstances, which involves individual initial
conditions, and the exerted effort, which basically has to do with
individual control variables. We hypothesize that these two types of
inequality affect growth in opposite ways.7
On the one hand, IO would reduce growth as it favors human
capital accumulation by individuals with better social origins, rather
than by individuals with more talent or skills (Chiu, 1998; Loury,
1981, and Mejia and St-Pierre, 2008). This result can be easily linked
to the credit market imperfection literature mentioned above (Galor
and Zeira, 1993 or Banerjee and Newman, 1993). In a multiple steady
state framework with borrowing constraints, initial heterogeneity in
wealth and certain circumstances would reduce the opportunity of
accessing credit to achieve higher levels of education or to promote
profitable investment projects, which would generate a misallocation
of talent, underinvestment in human capital and a negative conse-
quence on growth. On the other hand, income inequality among
those who exert different effort reflects the reward of unequal effort
or unequal talent, hence it would give incentive to people to do
hard work, saving or to invest in education (Mirrlees, 1971), which
would stimulate growth. The main goal of this paper is to test empir-
ically this hypothesis. For this task, we combine the empirical growth
literature from macroeconomics and the inequality-of-opportunity
literature from microeconomics. A discussion on both these litera-
tures and how they can be linked is presented in Section 2.
Data requirements for comparing inequality of income across
states or countries are severe (Deininger and Squire, 1998), but com-
parisons of IO and its estimation are even more stringent (Lefranc et
al., 2008). This is because empirical analysis of IO requires not only
comparable measures of individual disposable income but also indi-
vidual circumstances or social origins measured in a comparable and
homogeneous way. Furthermore, there are only few databases with
information on individual circumstances or social origins and in
these cases the number of circumstances is usually small. In addition,
to test for long-term effects on growth, we also need the value of IO
for at least two distant periods of time, generally ten years. This last
requirement limits even more the availability of databases. To the
best of our knowledge, the Panel Study of Income Dynamics (PSID)
database is the only exception that satisfies both requirements and
is rich enough in terms of cross-sectional heterogeneity, variables,
observations and circumstances.
Given these database limitations, to fulfill the overall objective of
the paper, we measure total inequality, IO, and IE at the U.S. state level
5
Galor and Moav (2004) emphasizes that the relationship between inequality and
growth changes with the replacement of physical capital by human capital accumula-
tion as a prime engine of growth along the process of development.
6 8
Though not considered in this paper, another possible source of inequality is luck
(Lefranc et al., 2009).
7
A similar reasoning is found in World Bank (2006) and Bourguignon et al. (2007b).
(Section 3). For this task, we follow widespread used procedures in the
inequality-of-opportunity literature and use refined data from the PSID
to decompose total inequality into inequality across groups (classified
by race and parental education, the two observed circumstances),
and inequality within groups, by applying the Theil 0 decomposition
technique. The first component will be the proxy for IO, while the
second component will be the proxy for IE.
Following Durlauf and Quah (1999), Panizza (2002), Partridge
(1997 and 2005), and Barro (2000), among others, we carry out an
extensive sensitivity analysis to show how robust are the results
found across model specifications and alternative econometric tech-
niques. Thus, we present in Section 4 the system-GMM estimations
for IO and IE based on a small model, where only a few number of
controls are specified, and a base model, where a large set of controls
is considered. As a matter of robustness, we also report the standard
linear pooled-OLS, Fixed Effects (FE) and Random Effects (RE) esti-
mates and a further sensitivity analysis is developed in Section 5.
In sum, we find that the impact of overall inequality on growth is
positive, as in Partridge (1997 and 2005), although it is not robust and
significant to alternative specifications, as emphasized by Panizza
(2002). On the contrary, the impact of the IO component is negative
and significant, while the impact of the IE component is positive and
significant. These results are highly robust to alternative model specifi-
cations, econometric techniques and other robustness checks. Thus, our
results would offer a unified explanation for the two opposite streams
of empirical results by suggesting that the overall impact of total in-
equality on growth could be positive, negative or zero depending on
which of the above two sources of income inequality dominates in the
data.
A final comment is worth noting. Our results call for a proper design
of policy. In particular, general redistributive policies may increase in-
vestment across individuals, but it may also discourage unobservable
effort borne by agents. Thus, its final impact on growth would be
unclear. However, selected policies reducing IO will promote not only
equity – in the sense of opportunity – but also economic efficiency
and growth. We will further discuss this issue in the final Section 6.
2. The inequality–growth debate and inequality of opportunity
The last decade has witnessed an intensive debate about the effects
of inequality on growth. However, as commented in the Introduction,
from a theoretical perspective the prevalence of a positive or negative
relationship between overall inequality and growth depends on which
channel predominates. This fact is clearly reflected by the empirical
evidence linking income inequality to economic growth: cross-sectional
and panel data studies are generally inconclusive.
Cross-sectional analysis showing a negative relationship between
both dimensions include, among others, Alesina and Rodrik (1994),
Persson and Tabellini (1994), Clarke (1995), Perottti (1996) and
Alesina and Perotti (1996). However, other authors find a positive re-
lationship between growth and income inequality, such as Partridge
(1997 and 2005), Zou and Li (1998) and Frank (2009). Barro (2000)
shows a very slight relationship between both variables when using
panel data, while Forbes (2000) finds a positive relationship.8 Panizza
(2002), using a cross-state panel, finds some evidence in support of a neg-
ative relationship between inequality and growth, though this relation-
ship is not robust to different econometric methods and regression
specifications. And Barro (2000), using a non-linear specification, finds a
negative and significant effect of inequality on growth for poor countries,
while a positive but insignificant effect of inequality on growth for rich
Forbes' results suggest that overall inequality has a significant positive effect upon
growth in the short and medium term, while Partridge (2005) finds that inequality is pos-
itively related to long-term growth, this relationship being less clear in the short-run.
 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122 109

countries. Given these different findings in the literature, we propose to IO would imply suboptimal levels of investment in human capital, with
analyze the inequality and growth relationship using the IO concept. a negative consequence on growth.
Another general argument to support the negative effect of initial
IO on growth can be inferred from Heathcote et al. (2008), which
2.1. The inequality of opportunity concept
focus on the implications of missing insurance markets rather than
on borrowing constraints. They emphasize that it is reasonable to be-
Equality of opportunity has been traditionally understood as the
lieve that society cannot provide insurance against one's circumstances
absence of barriers to access positions, education and jobs, thus hiring
(in our case, race and parental education) and the non-pecuniary dis-
should be meritocratic and characteristics like the economic class,
crimination associated with them. As a consequence, income allocation
gender and race should have no bearing on the merit of the individual
based on those uninsurable factors would be inefficient and would
(Lucas, 1995). Rawls (1971) and Sen (1980), among others, invoked a
harm growth.
more general notion and argued that equality of opportunity would
Arguments based on one particular circumstance can also be found
require compensating persons for a variety of circumstances (i.e., socio-
in the literature. Regarding parental initial wealth or human capital,
economic background, ethnicity, place of birth) whose distribution is
Chiu (1998) shows that, under liquidity constraints and decreasing
morally arbitrary. Roemer (1993) brings that philosophical debate
marginal utility, if wealth is redistributed from rich to poor, rich parents
into economics and formalizes a precise definition of equality of oppor-
would stop sending their less talented children to college, while, among
tunity (see also Van de Gaer, 1993). He emphasizes that an individual's
poor parents, more talented children would be afforded the chance to
outcome (income, welfare, health, etc.) is a function of variables within
go to college (see also Loury, 1981 and Bénabou, 1996). As a conse-
and beyond the individual's control, called effort (occupational choice,
quence, aggregate human capital will increase and, therefore, so will
number of hours worked or investment in human capital) and circum-
growth. Another example can be found in Ferreira (2001), where pro-
stances (socioeconomic and cultural background or race), respectively.
ductivity and wages are determined by the quality of the school one at-
As a consequence, total inequality can be seen, in reality, as a combina-
tends (see also Bénabou, 2000). Under borrowing constraints, children
tion of inequality of effort (IE) and inequality of opportunity (IO).
from poorer families would not be able to attend private schools, and
The important role of circumstances has been emphasized in the
would then go to public schools. On the contrary, richer families
inequality literature. For instance, Arrow et al. (2000), Hertz et al.
would send their children to private schools and, because they are
(2008), Blume and Durlauf (2001), and Loury (1989) have found
willing to pay high fees, these schools would provide high-quality edu-
strong evidence for persistent inequality, which is not attributable
cation. Then, if public school budgets are determined by the level of
to discrimination between individuals, but rather to factors such as
taxation, and the pivotal voter is wealthy enough to send their children
social networks, poor education quality and intergenerational inertia.
to private school, the funds received by public schooling would be low.
Furthermore, Roemer (1998) and Bowles et al. (2005), among others,
As a result, children from initially rich families will stay rich because
have emphasized that circumstances can affect the realization of
they attend high-quality schools, while children from initially poor
talent and, thus, the full achievement of a purely meritocratic society.
families will stay poor because they attend low-quality schools. Using
These authors have shown that even if individuals have high inborn
a similar argument than in Chiu (1998), removing this inequality trap
talent, the likelihood of their being able to realize the benefits of that
could enhance the overall level of human capital and hence growth.
talent (for example, in terms of admission to university or access to
If we focus on a theoretical framework where the history-dependent
employment) is strongly affected by social conditions.
initial conditions yield path-dependent outcomes, an obvious example
In this respect, the inequality-of-opportunity literature has identified
in the U.S., which could be applied to other countries, is the history of
the following channels through which parents can affect the income
race. A large initial wealth inequality between black slaves, indigenous
earning capacity of their children (Dardanoni et al., 2006): wealth;
natives and Latinos, and whites, combined with subsequent racial
provision of social connections which are relevant in the labor market;
barriers for accessing credit, would justify a negative effect of inequality
formation of skills in children, through family culture and investment;
of opportunity (due to race) on growth under a model à la Galor and
genetic transmission (like native ability and race); instillation of prefer-
Zeira (1993). Alternatively, we can advocate the model proposed by
ences and aspirations. Given these factors and the restrictions imposed
Easterly and Levine (1997), which report a negative impact of ethnic
by data availability, the empirical literature has widely used the level
heterogeneity on growth; or the model by Gradstein and Justman
of parental education or occupational status as a proxy for the first
(2002), which emphasizes the negative effect of racial and ethnic
three factors, and the ethnic group or race as a direct measure of the
heterogeneity on social norms that, in turn, lower the effectiveness of
fourth factor. In this paper, we use parental education and race as the
education on growth; or the model by Galor et al. (2009), where land
causal determinants of IO.
concentration, which is highly correlated with the proportion of income
Next, we provide several existing arguments in the growth–
inequality explained by race, adversely affects the implementation of
inequality and inequality-of-opportunity literature that can give sup-
human capital promoting institutions like public schooling and child
port to our main empirical finding: larger inequalities associated with
labor regulations.
differences in circumstances (parental education and race) affect neg-
atively to economic growth.
3. Inequality of opportunity in the U.S.

2.2. Theoretical frameworks for the role of IO on growth
Following the pioneering works of Galor and Zeira (1993) and
Banerjee and Newman (1993), a consensus is emerging that, in the
presence of borrowing constraints and fixed costs of education, a
higher initial inequality might have a negative effect on long-term
growth. These authors emphasize the existence of multiple steady-
state equilibrium paths, which cause a convergence trajectory that
will depend, among other things, on initial inequality. In this context,
people with unfavorable initial circumstances (because of their race
and/or parent's education) will face considerable barriers for accessing
credit, regardless of their talent and degree of effort exerted. As a result,
In this section we estimate the IO in the U.S. by using refined data
of the PSID database for 26 states in the 1970s, 1980s and 1990s. First,
we describe the database; next, we present the method; finally, after
describing the sample design, we show the IO estimates.
3.1. The PSID database
We need microdata of comparable measures of individual disposable
income and observed circumstances that span at least two decades and
cover a large enough cross-section of countries or states. While many
databases satisfy some of the above restrictions, the PSID database (for
U.S. states) is, to the best of our knowledge, the only exception that
110 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122
satisfies all of the above requirements and hence can be used to charac-
terize the IO–growth relationship.9
The PSID is a household panel which began in 1968 and is still
running. The survey was conducted annually from 1968 to 1997,
and then every other year. The initial sample for the PSID consists of
two independent probability samples. The first one is an equal prob-
ability sample of households from the 48 contiguous states (based on
a stratified multistage selection of the civilian noninstitutional popu-
lation of the U.S.) drawn by the Survey Research Center (SRC); the
second one is a national sample of low-income households drawn by
the Survey of Economic Opportunity (SEO). The combination of both
is also a probability sample, with unequal selection probabilities and,
as a result, compensatory population weighting would be needed in
the estimation of inequality indices. The PSID supplies these weights,
which indicate how many persons in the U.S. population are represented
by a given observation in the sample. These weights are designed to com-
pensate for unequal selection probabilities and differential attrition.10
The quality of the PSID database has been successfully assessed
by comparing different distributions from this database with their
equivalent in other sources.11 However, we are aware that two main
problems remain. First, sample sizes in the PSID are not very large
and, second, the data, which is representative on a national level, may
not be at the state level. To overcome these problems, we have made
a reasonable selection of data, states and decades, as commented
below (subsection 3.3), and an extensive robustness analysis will be
carried out in Sections 4 and 5.
3.2. The estimation approach
Consider a finite population of discrete individuals indexed by
i ∈ {1, …, N}. As is standard in the inequality-of-opportunity litera-
ture, the individual income, yi, is assumed to be a function of the amount
of effort, ei, and the set of circumstances, Ci, that the individual faces,
such that yi = f(Ci,ei). Effort is treated as a continuous variable, while,
for each individual i, Ci is a vector of J elements, each element corre-
sponding to a particular circumstance. Finally, circumstances are exog-
enous because they cannot be affected by individual decisions, while
effort is influenced, among other factors, by circumstances.
In order to estimate IO, the population is partitioned into a mutually
exclusive and exhaustive set of types Γ = {H1, …, HM}, where all indi-
viduals in each type m share the same set of circumstances. Further-
more, let us assume that em(π) is the level of effort exerted by the
individual at the πth quantile of the distribution of effort Fm where
π ∈ [0,1]. Given the type m, we can define the level of income obtained
by the individual at the πth quantile as vm(π) = ym[em(π)]. Now, let


v = (v1, …,vΜ) be a partition of income into M groups, and v ¼ μ i m
9
The Survey on Income, Social Inclusion and Living Conditions in Europe (EU-SILC)
database gives information on individual disposable income and circumstances for
most European countries (see Marrero and Rodríguez, 2012). However, this survey is
valid only for 2005, since it is the only year for which information is available on paren-
tal level of education. For the U.S., the Integrated Public Use Microdata Series (IPUMS-
CPS and IPUMS-USA) databases contain more information than the PSID but, unfortu-
nately, they do not provide information on parent's education. Furthermore, Roemer et
al. (2003), Lefranc et al. (2008), Bourguignon et al. (2007a), Rodríguez (2008) and
Cogneau and Mesplé-Somps (2009) have considered heterogeneous data with circum-
stances, but for a few countries and specific years.
10
A representative sample of 2043 Latino (Mexican, Cuban, and Puerto Rican) house-
holds was added to the PSID data in 1990. However, this sample missed out Asians, and
because of this crucial shortcoming, and a lack of sufficient funding, the Latino sample
was dropped after 1995. To avoid longitudinal inconsistencies, we have not considered
the Latino sample in our study. For more information about the PSID database visit:
http://psidonline.isr.umich.edu/Guide/.
11
For instance, Gouskova and Schoeni (2010) have compared estimates of family in-
come between the PSID and the March Current Population Survey (CPS) for the entire
history of the PSID (1968–2007). They find that the distributions are in close agree-
ment throughout the 39-year history of the PSID, above all in the range between the
5th and 95th percentiles.
be an N-dimensional smoothed distribution where each individual in
type m receives that type's mean income μm.
In order to fulfill the aim of this paper, we need to decompose
overall inequality into IO and IE components. Following Ruiz-Castillo
(2003) and Ferreira and Gignoux (2011), among others, we define
IO ¼ I ðv Þ, where I is an inequality index. In this manner, whenever
total inequality can be additively decomposed by population groups
according to a set of circumstances, the between-group inequality com-
ponent can be seen as the IO term, while the within-group inequality
component can be interpreted as the IE term. Among all the possible
inequality indices, only those of the Generalized Entropy class are addi-
tively decomposable into a between-group and a within-group compo-
nent (Bourguignon, 1979 and Shorrocks, 1980). Consequently, we
adopt the mean logarithmic deviation or Theil 0 (T), because it belongs
to the Generalized Entropy class, has a path-independent decomposi-
tion (Foster and Shneyerov, 2000), and uses weights based on the
groups' population shares. The Theil 0 index can be exactly decomposed
as follows:
X
M  m
T ðvÞ ¼ T ðv Þ þ pm T v ; ð1Þ
m¼1
where T ðv Þ is the between-group component (the IO term), which is
calculated by applying the Theil 0 index to the vector v, and the second
term is the within-group component (the IE term), which captures the
income inequality within each type m, weighted by pm, the frequency of
type m in the population. In this manner, the two sources of income
differences, circumstances and effort, can be included separately in
the inequality–growth regressions conducted in Section 4.
The IO and IE components can be estimated non-parametrically
(Checchi and Peragine, 2010, and Lefranc et al., 2008) and/or paramet-
rically (Bourguignon et al., 2007a, and Ferreira and Gignoux, 2011). The
convenience of using one or another approach generally depends on the
available database and the issue under analysis. In this paper, we use the
first approach because it does not assume any particular functional form
and its application is straightforward for our set of circumstances.
3.3. The sample design
In order to estimate IO, we need to refine the PSID samples. First,
we consider individuals who are household heads (male in married-
couple families, but female or male, otherwise).12 Correspondingly,
gross income is computed as the household head's labor income plus
the household capital income divided by the number of adults in the
household (Rodríguez, 2008; Roemer et al., 2003).
Second, we remove the so-called composition effect: individuals
with different ages are in different phases of the wage-earning time se-
ries. To do this, the common practice in the inequality-of-opportunity
literature is to consider only household heads in a given age group, for
example, 25 to 50 years old. However, in our case, this strategy could
significantly reduce the number of available observations and, as a
result, the accuracy of IO estimates. Alternatively, we just restrict the
samples to household heads between 18 and 65 and, following
Checchi and Peragine (2010), we regress actual gross incomes on expe-
rience, experience squared and survey years.13 Then, we take residuals
from this regression and, because they are centered around zero, we
add a constant to match the minimum of the actual series.
Third, we consider two circumstances: the father's education and
race. For the father's education, we assume four groups: no education,
12
Notice that information on father's education for wives is only available from 1974
onwards.
13
Due to the lack of information about actual experience, we have calculated potential
experience as: age – age when finished education. The results of these partial regressions
are available upon request.
 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122
primary, secondary and tertiary education14; for race, we consider
two groups: white and non-white. Then, combining both circum-
stances, the sample is partitioned into 8 groups or types (i.e., M = 8),
and the estimated inequality-of-opportunity index is called “IO-8group”.
Despite the limitations of the PSID database to using more circum-
stances, parental education (as a proxy of provision of social connec-
tions, transmission of culture, parents' wealth and investment on
formation of skills in children) and race are two of the main factors
used in the inequality-of-opportunity literature.
Four, to neutralize possible outliers, inequality indices (overall
inequality, IO and IE) in 1970, 1980 and 1990 are the average up to
2 years, that is, 1969 and 1970, 1979 and 1980, and 1989 and 1990,
respectively. Finally, we disregard those states with fewer than 50
observations at any decade so as to have enough information in
each group to estimate IO. In this respect, we will show that the
main results do not vary significantly when the selection criterion
changes to 20, 30, 70 or 100 observations (Section 5). Following these
criteria, our final sample for 1970, 1980 and 1990 reduces to a set of 26
states distributed throughout the whole U.S. territory: West (California,
Missouri, Oregon and Washington), Midwest (Illinois, Indiana, Iowa,
Michigan, Minnesota and Ohio), South (Arkansas, Florida, Georgia,
Kentucky, Louisiana, Maryland, Mississippi, North and South Carolina,
Tennessee, Texas and Virginia) and Northeast (Massachusetts, New
Jersey, New York and Pennsylvania).
3.4. Inequality of opportunity in the U.S. by states
For our refined database, Table 1 shows the estimations of total
inequality (Theil 0), the absolute IO and the relative IO (IO as a per-
centage of Theil 0) for 1970, 1980 and 1990.15 Fig. 1a, b and c
shows the decade averages of these inequality measures, respectively,
sorted from the highest to the lowest index.16
By comparing the income inequality and IO average results
(Fig. 1a, b and c), we observe substantial differences between their rank-
ings. For example, there exists a group of states with high total inequal-
ity and rather low IO (absolute and relative), such as Massachusetts and
California, while the opposite happens for states like Louisiana and
Virginia. Additionally, there exist some states whose relative position re-
mains basically unchanged, such as Oregon and Minnesota, which are at
the lowest levels of both dimensions, and New Jersey, South Carolina
and Florida, which are at the top of the three rankings. This finding is
consistent with Fig. 2a and b, which, for the entire pool of observations,
show the relationship between the Theil 0 index and the estimated IO,
and between IE and IO indices, respectively. Their coefficients of deter-
mination (R2) are 0.45 in Fig. 2a, positive but far from unity, and only
0.21 in Fig. 2b. This result highlights how those factors affecting these
two dimensions of inequality should be different. Note that Fig. 2a and
b shows several important outliers (Louisiana and South Carolina in
1970, and California, Maryland, Massachusetts and New Jersey in
1990), which will be analyzed in more detail in the analysis of robust-
ness in Section 5.
14
Information on mother's education is not available for the whole period. “No education”
means 5 grades or less; “primary” education goes from 6 grades to 11; “secondary” educa-
tion refers to 12 grades and 12 grades plus non-academic training; and, “tertiary” education
refers to college with or without a degree. The results do not change significantly when
three groups for the father's education (primary or no education; secondary education;
and, tertiary education) are considered.
15
Due to space constraints, we have omitted from Table 1 the standard errors com-
puted by bootstrapping (Davison and Hinkley, 2005). The estimated standard errors
for the income inequality and IO indices are rather precise. They are available from
the authors upon request.
16
Note that a direct comparison of these inequality indices with those published by
the U.S. Census Bureau would be misleading. Data from the PSID were refined to esti-
mate IO and not to estimate total inequality. Thus, Census data refer to families, while
our data refer to individuals (household heads), our samples consider only individuals
between 18 and 65 years old, and our income data are corrected to consider that indi-
viduals are at different phases of the wage-earning time series.
111
Going back to Fig. 1c, we observe two relevant facts. First, IO esti-
mates represent a modest percentage of the total inequality, which is
consistent with results in the literature.17 The existence of additional
circumstances capturing differences in opportunity, other than race
and parental education, could explain this result.18 Second, the
range of inequality on income opportunities across states is large,
from 5% to 31% of total inequality, with the highest shares being
systematically found in the Southern states. In this respect, it is inter-
esting to note that the majority of the Southern states are above the
regression line in Fig. 2a and b. These findings could be explained
by the history of race in the U.S. In keeping with the works mentioned
earlier by Easterly and Levine (1997), Gradstein and Justman (2002)
and Galor et al. (2009), among others, large initial wealth inequality,
mainly in the form of land concentration, between whites and black
slaves (who were extensively used in the Southern states) could be
effectively causing our share estimates.
Finally, we comment the time evolution of the indices. As a general
trend, we observe that total inequality between the 70's and the 80's
is stable, while it increases between the 80's and the 90's. Note that
this progression is consistent with the information gathered from the
U.S. Census Bureau. With respect to IO values, its temporal progression
is less clear, and depends on each state. For example, between the 80's
and 90's, IO drops significantly in Louisiana and Mississippi, while it
increases considerably in Maryland, Georgia, Illinois, New Jersey,
Massachusetts, Tennessee, Texas and Virginia. In all other states in the
sample, IO is relatively constant between the 80's and the 90's. Accord-
ingly, IO and IE evolve in one direction or another depending on the
case. This finding is consistent with Fig. 3. Looking at the entire data
pool, the relationship between IO and IE annual changes is positive
though small, showing a coefficient of determination of 0.35, and of
just 0.15 if the two extreme values shown in the figure are omitted.
Thus, a clear-cut endogenous individual behavior response of IE to IO
changes (or vice-versa) cannot be inferred from this simple data analy-
sis. Many other factors such as institutions or policy actions might be
affecting the evolution of both components in one way or another. An
extensive analysis of this issue deserves much more attention, which
goes beyond the scope of this paper.
4. Inequality, inequality of opportunity and growth: an
empirical investigation
In this section we carry out the main task of this paper, which is to
empirically characterize the effect of IO on growth. The benchmark
analysis is limited to three decades, the 1970–2000 period, and to a
selected set of 26 U.S. states.19 Our empirical strategy relies on estimating
of a reduced-form growth equation with a set of inequality indices (total,
IO and IE) added to an otherwise standard set of growth determinants,
GY iðt−s;t Þ ¼ α i þ φ′T t þ βY it−s þ ϕ′INEQ it−s þ λ′Χ it−s þ εit ; ð2Þ
where GYi(t − s,t) = ln Yi,t − ln Yi,t − s is the logarithmic growth rate
of real personal income (adjusted by CPI) divided by total mid-year
17
Ferreira and Gignoux (2011) find that between one fifth and one third of all in-
come inequality is explained by opportunities in six countries in Latin America.
Marrero and Rodríguez (2012) find that between 2% and 22% of overall inequality is
explained by opportunities in a set of 23 European countries. See also Checchi and
Peragine (2010) and Marrero and Rodríguez (2011).
18
The true IO requires the observation of all circumstances, but this is unfeasible in
practice. Consequently, our IO estimates must be interpreted as a lower bound
(Ferreira and Gignoux, 2011). Another possibility is that income-based IO tends to un-
derestimate IO because the higher measurement error and variance for transitory com-
ponents in the distribution of income (in comparison with the distribution of
consumption) may be effectively counted as inequality of effort.
19
An advantage of dealing with states instead of with countries is that heterogeneity
within states does not stem from the political process because, for the most part, it is sim-
ilar across the different states. More importantly, institutional, cultural, religious and other
differences are less intensive for U.S. states than for different countries (Partridge, 1997).
112 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122

Table 1
Inequality of income and inequality of opportunity (8 groups) in 1970, 1980 and 1990.

Observations Total inequality IO (8 groups)a IO to total inequality ratiob

State 1970 1980 1990 1970 1980 1990 1970 1980 1990 1970 1980 1990

Arkansas 88 85 100 0.0916 0.0894 0.0819 0.0172 0.0157 0.0100 18.8 17.7 12.3
California 353 438 434 0.0945 0.1046 0.1854 0.0068 0.0046 0.0061 7.2 4.4 3.3
Florida 86 135 176 0.1286 0.1464 0.2046 0.0384 0.0341 0.0280 29.9 23.3 13.7
Georgia 91 129 144 0.0798 0.1064 0.1278 0.0176 0.0201 0.0311 22.0 18.9 24.3
Illinois 120 158 147 0.0813 0.0966 0.1480 0.0079 0.0045 0.0215 9.8 4.7 14.6
Indiana 84 133 130 0.0553 0.0767 0.1207 0.0031 0.0055 0.0143 5.7 7.1 11.9
Iowa 59 89 89 0.0901 0.1025 0.1681 0.0056 0.0129 0.0160 6.2 12.6 9.5
Kentucky 77 85 97 0.0692 0.0653 0.1062 0.0036 0.0097 0.0163 5.2 14.9 15.3
Louisiana 79 119 94 0.0587 0.1251 0.0981 0.0156 0.0578 0.0191 26.6 46.2 19.5
Maryland 111 168 211 0.0549 0.1062 0.2355 0.0206 0.0184 0.0590 37.6 17.3 25.0
Massachusetts 74 98 124 0.0780 0.1289 0.2668 0.0077 0.0020 0.0244 9.9 1.6 9.1
Michigan 177 220 245 0.1031 0.1368 0.2094 0.0175 0.0299 0.0350 17.1 21.9 16.7
Minnesota 58 74 88 0.0712 0.0946 0.1223 0.0069 0.0079 0.0066 9.8 8.4 5.5
Mississippi 101 195 204 0.0987 0.0782 0.0885 0.0309 0.0205 0.0097 31.3 26.3 11.0
Missouri 134 140 143 0.0736 0.0880 0.1180 0.0019 0.0077 0.0056 2.7 8.8 4.8
New Jersey 92 116 142 0.0911 0.1488 0.4078 0.0103 0.0285 0.0841 11.3 19.2 20.6
New York 193 217 228 0.0849 0.1196 0.1416 0.0107 0.0080 0.0084 12.6 6.7 5.9
N. Carolina 151 209 272 0.0881 0.1074 0.1455 0.0284 0.0295 0.0338 32.3 27.5 23.3
Ohio 170 234 200 0.0911 0.0869 0.1325 0.0272 0.0069 0.0118 29.9 8.0 8.9
Oregon 60 74 76 0.0675 0.0984 0.0910 0.0031 0.0063 0.0035 4.6 6.4 3.9
Pennsylvania 192 236 263 0.0765 0.1042 0.1195 0.0026 0.0078 0.0076 3.5 7.5 6.4
S. Carolina 155 215 288 0.1675 0.0926 0.1170 0.0686 0.0142 0.0238 40.9 15.4 20.3
Tennessee 51 75 88 0.0506 0.0568 0.1521 0.0054 0.0041 0.0355 10.8 7.3 23.3
Texas 210 286 308 0.0902 0.0940 0.1224 0.0135 0.0063 0.0196 15.0 6.8 16.1
Virginia 120 155 164 0.0587 0.0698 0.1161 0.0066 0.0061 0.0243 11.3 8.7 20.9
Washington 53 74 72 0.1152 0.0961 0.1414 0.0160 0.0111 0.0062 13.9 11.6 4.4
USA 3464 4706 5124 0.0916 0.1058 0.1749 0.0055 0.0046 0.0099 6.1 4.4 5.7

The values for 1970, 1980 and 1990 are actually the averages of 1969 and 1970, 1979 and 1980, and 1989 and 1990, respectively.
a
IE estimates are the difference between total inequality and IO values. This decomposition is only exact for the Theil 0 index.
b
In percentage.

population in the entire decade; αi and Tt are country- and time-specific
effects, respectively; Yi,t-s is real per capita lagged income (in logs),
which controls for conditional convergence across states; INEQi,t-s
represents the inequality indices (overall inequality, IE and/or IO,
depending on the case); Xi,t-s groups a set of control variables other
than lagged income; and εit is an i.i.d error term.
Depending on the controls included in X, we have a small or a base
model. The small or parsimonious version includes human capital var-
iables, the percentage of people who live in metropolitan areas and
the percentage of the population over 65 years of age. The base model
accounts for human capital, industry mix, farm employment, welfare
public expenditures, lag employment growth and the fertility rate
(fertility) at the beginning of the period as an additional explanatory
variable.20 While the small specification suffers from omitted-variable
20
As is standard in the literature, we consider three categories to measure human
capital: the percentage of the population over 24 years of age who have graduated
from high school but do not have a four-year college degree (high school); the percent-
age who have graduated from a four-year college (college); and the omitted category,
which is the percentage of individuals who have not graduated from high school. To
control for the initial economic sectoral composition of each state, the shares of
nonagricultural employment are included for the following sectors: mining, construc-
tion, manufacturing, transportation and public utilities, finance, insurance and real estate,
and government. Traded goods and services are the omitted sector, and thus the employ-
ment share coefficients should be interpreted as being relative to this sector. The per-
centage of the population who worked on a farm (farm) is included to account for the
different importance of agriculture across states. In order to account for the possibility
that growth in the previous decade could, in turn, influence growth in the following
decade and be correlated with past inequality, we include the percentage change in
nonagricultural employment (emp_growth) in the preceding decade (e.g., employment
growth in the 70s is used to explain per capita income growth in the 80s). Welfare ex-
penditure as a percentage of personal income (welfare) is included as a proxy for the
degree of distributive policy. More inequality would imply greater distribution, more
welfare expenditure and that, in turn, would imply lower average growth rate (as
some political economy models suggest). Hence, a negative growth−inequality rela-
tionship could appear if we omit this variable. Fertility has been proved to be an impor-
tant channel through which initial inequality may reduce growth. For the case of U.S.
states, Panizza (2002) and Partridge (1997, 2005) omit fertility in their reduced-form
regressions, thus, although minor, this inclusion is a contribution of the paper.
bias problems, the base model could introduce significant collinearity
problems in the regression. For these reasons, especially in a reduced-
form exercise, Perottti (1996) and Panizza (2002), among others, em-
phasize the importance of considering this distinction and comparing
the results for both specifications.
Fig. 4a, b and cillustrate the main results of the paper. They show
the scatter plots between decennial growth and each measure of
lagged inequality (total, IO and IE), conditioning both variables on
all other explanatory variables included in the base model. Consistent
with Partridge (1997, 2005) and Forbes (2000), Fig. 4a indicates a
positive relationship between total inequality and growth, although
somewhat weak, with a slope of 31.4. The two other figures consider
the cases of IE and IO. Fig. 4b suggests that the relationship between
growth and IE becomes more positive (its coefficient is now 93.7)
and more significant (i.e., data are less dispersed around the regres-
sion line) than for total inequality. However, Fig. 4c shows that the
relationship between IO and growth is significant but negative, with
a coefficient of − 107.9.21 Moreover, note that in Fig. 4a–c the differ-
ent regions (South, North-East, West and Mid-West) are more mixed
in the scatter plots than in Fig. 2a and b, which suggests that the base
model is capturing quite well the regional clustering pattern originally
observed in the data.
4.1. Econometric issues
The easiest way to estimate a panel data model like Eq. (2) is to
ignore any unobserved country specific heterogeneity – i.e., set αi = α
for all i – and then apply OLS to pooled data. However, this strategy
may result in seriously biased estimates when region heterogeneity
21
Correcting for the main outliers in these scatter plots does not change significantly
the results. In particular, estimated slopes become 22.5 for total inequality
(R2 = 0.028), 79.0 for IE (R2 = 0.126) and −88.9 for IO (R2 = 0.06).
 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122 113

a) Income inequality in U.S. (Theil 0)

(average 1970-1980-1990)
0.25

0.20

Total Inequality
0.15

0.10

0.05

0.00

CA

USA

WA
NY
NC
NJ

LA
GA
OH

OR
MO

TN
MA
MI
MD

MN

MS
FL

TX
SC

PA

AR

VA
KY
IA

IN
IL
States

b) Inequality of opportunity (8 groups) in U.S.

(average 1970-1980-1990)
0.05
Inequality of Opportunity

0.04

0.03

0.02

0.01

0.00
NJ

NC

NY

CA
WA

USA
LA

GA

OH

OR
FL
MD

MS

TN

MA

MN

MO
MI

TX
SC

AR

VA

KY

PA
IA

IL

IN
States

c) Inequality of opportunity Ratio (8 groups) in U.S.

(average 1970-1980-1990)
0.35
0.30
0.25
Ratio

0.20
0.15
0.10
0.05
0.00
WA

USA
GA

OH

OR
FL

TN

TX
MD

MS

MN
MA

MO
MI
SC

AR

VA

KY

PA
IL
IA
NC

NY
IN

CA
NJ
LA

States

Fig. 1. a. Income inequality in the U.S. (Theil 0). b. Inequality of opportunity (8 groups) in the U.S. c. Inequality of opportunity ratio (8 groups) in the U.S.

exists. The standard alternatives are the Fixed Effects (FE) and Random
Effects (RE) estimators, which wipe out the αi term by transformation.
However, in a dynamic model like Eq. (2), there is a problem of
endogeneity because (at least) the lagged transformed income variable
is correlated with the transformed error term. As a result, the FE estimate
of β is typically downward biased, while the RE estimate is upward biased
in the same direction as pooled OLS (Anderson and Hsiao, 1982). Another
potential issue in our case is the simultaneous determination of inequality
and growth. To reduce this problem INEQ and X in Eq. (2) are measured at
the beginning of each decade.
In the absence of suitable external instruments, to address all these
potential problems we employ the system-GMM estimator (Arellano
and Bover, 1995; Blundell and Bond, 1998). This approach is based
on the use of internal instruments (lagged levels of regressors) and
estimates a system of equations in both first-differences and levels. To
implement the system-GMM estimator, we first rewrite Eq. (2) as a
Dynamic Panel Data (DPD) model:
0 0 0
ln Y it ¼ α i þ β̃ ln Y it−s þ ϕ INEQ it−s þ φ T t þ λ X it−s þ εit ; ð3Þ
β̃ ¼ ð1 þ βÞ;
where the main difference with Eq. (2) is the interpretation of the coef-
ficient for the lagged income term; the requirement for conditional con-
vergence in Eq. (2) is β b 0, while it is β̃ b 1 in Eq. (3).
The first-difference transformation of Eq. (3) eliminates the fixed
term αi. For the resultant set of equations in first-differences, the basic
idea is to employ the levels of the series lagged two or more periods
(i.e., Yit-s, INEQit-s and Xit-s, for s ≥ 2) as instruments. However, using
the model only in first-differences may lead to important finite sample
bias problems when variables are highly persistent, which is the case for
114 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122

a) The scatter plot of total inequality and IO in U.S. (Theil 0).

(pool of observations: 1970,1980 and 1990)
0.09

0.08 NJ-90
IO = -0.0055 + 0.1963 ·Ineq

Inequality of Opportunity (IO)

R² = 0.45
0.07
SC-70
0.06 MD-90
LO-80
0.05

0.04

0.03 South
North
North-east
MA-90 West
0.02
Mid-west

0.01
CA-90
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45
Total Inequality, Theil 0
Outliers: New Jersey, 1990 (NJ-90); Maryland, 1990 (MD-90); South Carolina, 1970 (SC-70);
Louisiana, 1980 (LO-80); Massachusetts, 1990 (MA-90); and, California, 1990 (CA-90).

b) The scatter plot of IE and IO in U.S. (Theil 0).

(pool of observations:1970, 1980 and 1990)
0.09

0.08 NJ-90
IO = 0.0013+ 0.1604 ·IE
Inequality of Opportunity (IO)

R² = 0.21
0.07
SC-70

0.06 MD-90
LO-80
0.05

0.04

0.03 South
North east
North-east
0.02 MA-90 West
Mid-west

0.01
CA-90
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Inequality of Effort (IE)

Outliers: New Jersey, 1990 (NJ-90); Maryland, 1990 (MD-90); South Carolina,1970 (SC-70);
Louisiana, 1980 (LO-80); Massachusetts, 1990 (MA-90); and, California, 1990 (CA-90).

Fig. 2. a. The scatter plot of total inequality and IO in the U.S. (Theil 0). b. The scatter plot of IE and IO in the U.S. (Theil 0).

variables such as income and inequality (Blundell and Bond, 1998). first-difference model, where the instruments of the levels are suitable
Under these conditions, lagged levels of the variables are only weak in- lags of their own first differences (i.e., ΔYit-s, ΔINEQit-s and ΔXit-s).23
struments for subsequent first-differences.22 To overcome this problem, Consistency of the system-GMM estimator depends on the validity
the system-GMM procedure adds a set of equations in levels to the of the instruments set constructed in this way, which is basically

22
Forbes (2000), a particularly influential paper in the growth–inequality literature,
used the first-difference GMM (Arellano and Bond, 1991) for identification. We focus
on the system-GMM because it also addresses the weak instruments problem. Bond
et al. (2001) and Marrero (2010), among many others, have emphasized the relevance
of using system-GMM when working with growth models.
23
For the set of level equations, most of the associated moment conditions are math-
ematically redundant with the instruments of the first-differenced equations. As a re-
sult, for the set of level equations, only one lag is ordinarily used for each period and
instrumental variable. In our case, we have two first-difference equations, one for
1990–2000 (the instruments could be the levels in 1980 and 1970) and another for
1980–1990 (the instruments could be the levels in 1970); and two equations in levels,
one for 2000 (the instruments are the first-differences in 1980–1990), and another for
1990 (the instruments are the first-differences in 1970–1980).
 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122 115

The scatter plot of IE and IO variations in U.S. (Theil 0).

(pool of observations: 1980-1970 and 1990-1980)
0.25

Inequality of Opportunity (IO,change)

0.2
change(IO)=0.0269+1.2708·change(IE) NJ 90-80
R² = 0.35
0.15

MA 90-80
0.1 MD 90-80
CA 90-80

0.05

LO 80-70 South
0 North-east
SC 80-70 West
Mid-west
-0.05

-0.1
-0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08
Inequality of Effort (IE,change)
Outliers: New Jersey, 1990 (NJ-90); Maryland, 1990 (MD-90); South Carolina, 1970 (SC-70);
Louisiana, 1980 (LO-80); Massachusetts, 1990 (MA-90); and, California, 1990 (CA-90).

Fig. 3. The scatter plot of IE and IO variations in the U.S. (Theil 0).

determined by the autocorrelation structure of the error term in
Eq. (3). The Hansen's J-test of over-identifying restrictions is the test
most commonly used to assess the joint validity of the proposed
instruments set.24 This test basically examines the correlation be-
tween the instruments and the regression residuals, where the null
hypothesis is the absence of such correlation. In addition, the absence
of autocorrelation of the disturbances εit in Eq. (3) is required for the
validity of internal instruments lagged two or more periods.25 To test
for this, Arellano and Bond (1991) developed the m1 and m2 tests,
which examine the first and the second-order correlation of the residuals
in first differences. Absence of autocorrelation requires that the m1 test
rejects the null while the m2 does not. Unfortunately, we cannot com-
pute the m2 test because we need observations for at least five time
periods; hence we only report the m1 test for illustrative purposes.
System-GMM estimates are shown for the one-step estimator case. In
contrast with the two-step version, the one-step system-GMM estimator
has standard errors that are asymptotically robust to heteroskedasticity
and, more importantly in our case, are more reliable for finite sample
inference (Blundell and Bond, 1998; Bond et al., 2001). In addition, errors
in panel models like Eq. (2) or (3) are generally heteroskedastic and,
more likely, serially correlated, as some unobservable variables cor-
related with the growth of states might persist over time. Following
Green (2008), we correct for these problems by considering panel-
robust standard errors, that is, clustering by states.26
24
If errors are homoskedastic, the Hansen's J-test is equivalent to the older Sargan's
test. Since homoskedasticity is an unrealistic assumption in these sorts of dynamic
models, we show only the results for the Hansen's J-test as it is standard in the
literature.
25
Including time dummies in Eq. (2) or (3) makes the assumption of no correlation
across states in εit more likely to hold. For this reason, we include time dummies in
all specifications estimated in the paper.
26
We have estimated the following variance–covariance matrix (Green, 2008, page 188):
" #−1 " ! n !′ #" #−1
XG
G X G Xng
X g X G
X ′g X g xig ε^ ig xig ε^ ig X ′g X g ;
g¼1
G−1 g¼1 i¼1 i¼1 g¼1
where G is the number of clusters in the sample (in our case, the number of states, N = 26),
and each cluster consists of ng (g = 1, … G) observations (in our case 3); the matrix X con-
tains the regressors used; and, ε^ refers to the residuals estimated initially.
4.2. Estimation results
For the reasons discussed above, we now show the estimated
results for our preferred system-GMM (one-step) model (Table 2).
In all specifications, we first note that the p-value of the Hansen
J-test is above 0.10, which is a symptom of the validity of the set
instruments used. Moreover, the AR(1) tests for first-difference resid-
uals are consistent with the absence of autocorrelation. Accordingly,
we consider that the set of instruments used is valid in correcting
the endogeneity problem posed in Eq. (3).
We can now start commenting on the estimated results of model (3)
for the Theil 0 index (panel 1 in Table 2). For the small and base models,
we find a positive, although weak, relationship between total inequality
and per capita income. With respect to the other controls, the results
are, in general, in line with the literature. First, the positive – and
lower than one – coefficient for lagged per capita income reflects condi-
tional convergence. Second, future economic growth is expected to be
positively correlated with human capital. As in Partridge (1997 and
2005) and Panizza (2002), the relevant variable for education is
college, which is highly positive and significant with respect to the omit-
ted category (non-graduated). The coefficient of the high school variable
is negative but never significant.
Third, the estimates for the percentage of the population who
worked on a farm and the initial industrial mix variables are negative
and significant except for the variables finance, insurance and real estate
with a negative but non-significant coefficient and transportation and
public utilities, with a positive and non-significant coefficient. Thus, in
general, states with greater initial shares in services and traded goods
(the omitted category) have experienced higher economic growth.
Fourth, the labor growth in the preceding year variable is negative
and significant, a result consistent with previous studies. Fifth, the
percentage of public welfare expenditure to personal income has a
negative, though non-significant coefficient. According to mainstream
political economy models, the expected result would be a negative
and significant coefficient. However, we must note that welfare expen-
ditures (and the implied redistribution) could promote growth by
relaxing the credit constraints on the opportunity disadvantaged indi-
viduals, which would partially offset the traditional negative effect.
Finally, the relationship between lagged fertility and subsequent growth
116 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122

a) Total inequality and growth in US states

(adjusted pooled data, base model)
15
growth(adjusted) = -0.0001+31.4·Theil 0 (adjusted)
R² = 0.0653
LO-70
10

Decennia lgrowth ,adjusted

NY-80

(basemodel) NJ-90

0 OH-70
MA-90 SC-70
MD-90
-5 South
North
North-east
CA-90 West
Mid-west
-10 MA-70

-15
-0.1 -0.05 0 0.05 0.1 0.15 0.2
Initial period Total Inequality, adjusted (basemodel)

b) Inequality of Effort and growth in US states

(adjusted pooled data, base model)
15
growth(adjusted) = -0.0002 + 93.7·IE (adjusted)
R² = 0.1543
NY-80 LO-70
10
Decennial growth, adjusted

5
(basemodel)

NJ-90
0 OH-70

MA-90
SC-70
-5 MD-90 South
North
North-east
West
CA-90 Mid-west
-10 MA-70

-15
-0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05 0.06
Initial period IE, adjusted (basemodel)

c) Inequality of Opportunity and growth in US states

(pooled and adjusted data, base model)
12
growth (adjusted) = 0.0004-107.9 IO (adjusted)
R² = 0.0577
10
LO-70 NY-80
8
NJ-80
Decennial growth, adjusted

6
South
CA-90
4 North-east
(basemodel)

West
2 Mid-west
OH-70
0

-2
MA-90
-4 TX-80
MD-90
-6 SC-70

MA-70
-8
OR-90
MO-70
-10
-0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04
Initial period IO, adjusted (basemodel)

Fig. 4. a. Total inequality and growth in US states. b. Inequality of effort and growth in US states. c. Inequality of opportunity and growth in US states.
 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122 117

Table 2
Growth, inequality and inequality of opportunity (system-GMM).

Theil 0 Theil 0 + IO Theil 0 + IO/Theil 0 IE + IO

Small Base Small Base Small Base Small Base

Total inequality (Theil 0) 18.528 24.216*** 53.941** 59.393*** 25.190* 58.012** – –

(1.22) (2.43) (1.97) (3.21) (1.39) (2.06)
Inequality return to effort (IE) – – – – – – 53.607** 60.736***
(2.09) (3.17)
Ionequality of opportunity (IO) – – −154.380** −133.940** −16.694** −22.909** −98.852** −80.448**
(−1.85) (−2.28) (−1.98) (−1.84) (−1.82) (−1.89)
Income lag 0.566*** 0.283** 0.620*** 0.327** 0.606*** 0.245* 0.617*** 0.351***
(6.70) (1.92) (5.65) (2.24) (5.58) (1.65) (5.09) (3.05)
High-school (% total graduated) 0.023 0.016 −0.180 −0.059 −0.171 0.044 −0.174 −0.077
(0.16) (0.07) (−1.10) (−0.30) (−1.04) (0.19) (−1.06) (−0.45)
College (% total graduated) 1.117*** 1.794*** 0.856*** 1.586*** 0.928*** 1.712*** 0.869*** 1.521***
(4.13) (4.09) (3.29) (3.64) (3.43) (3.39) (3.33) (3.99)
Farm population (% total) – −67.920** – −70.560** – −87.322*** – −68.348***
(−2.33) (−2.32) (−2.76) (−2.45)
Mining empl. (% nonagric.) – −117.830* – −117.150** – −114.050 – −114.750**
(−1.51) (−1.83) (−1.28) (−1.79)
Construction empl. (% nonagric.) – −215.470*** – −145.520** – −103.970 – −140.940*
(−2.74) (−1.74) (−0.93) (−1.65)
Manufact. empl. (% nonagric.) – −63.800*** – −52.929** – −48.037 – −51.011**
(−2.51) (−1.96) (−1.38) (−1.87)
Transp. & pub. util. empl. (% nonagric.) – 40.610 – 65.829 – 78.052 – 69.487
(0.45) (0.67) (0.72) (0.72)
Fin. inst. & real estate empl. (% nonagric.) – −136.750 – −150.820 – −156.620 – −153.280
(−1.11) (−1.26) (−1.03) (−1.33)
Government empl. (% nonagric.) – −61.280* – −42.174 – −36.647 – −38.778
(−1.37) (−0.93) (−0.64) (−0.87)
Change nonagric. empl. previous decade – −0.087* – −0.087* – −0.088 – −0.084
(−1.45) (−1.35) (−1.18) (−1.32)
Welfare exp. (% personal income) – −0.870 – −0.958 – −0.458 – −0.958
(−0.62) (−0.74) (−0.32) (−0.76)
Fertility rate – −0.364*** – −0.369*** – −0.367*** – −0.369***
(−3.20) (−4.26) (−3.24) (−4.36)
Population age 65 or above (% total) 0.265 – 0.075 – 0.122 – 0.079 –
(0.64) (0.32) (0.53) (0.33)
Population in Metropolitan (% total) 0.129** – 0.108** – 0.115** – 0.109** –
(2.02) (1.83) (1.90) (1.78)
Hansen J-test (p-value) 0.262 0.442 0.390 0.690 0.420 0.671 0.365 0.598
m1-test (p-value) 0.003 0.084 0.008 0.016 0.008 0.070 0.008 0.010
Num. observ. 78 78 78 78 78 78 78 78

t-statistics in parenthesis: * significant at 10%; ** significant at 5%; *** significant at 1% (one-sided test).
A constant term and time dummies are included in all models.

is negative and significant at a 1% level, in line with the related literature
(de la Croix and Doepke, 2003; Galor and Zang, 1997). This relationship
is strongly robust to all the specifications considered.
Next, we can comment on the impact of IO and IE on per capita
income growth (panels 2–4 in Table 2). We consider different ways
to include IO in the INEQ vector in (2). The easiest way is to consider
IO together with total inequality (panel 2). When including the IO
term, we control for the observed circumstances (father's education
and race) and, consequently, total inequality would now more clearly
reflect the IE component. As a result, the coefficient for total inequal-
ity should be higher and more significant than before, which is what
effectively happens. Moreover, IO coefficients are always negative
and significant at the 5% level. However, IO and total inequality can
be highly correlated, because IO is a part of total inequality, which
can affect the regression results.
Alternatively, as proposed by Ferreira and Gignoux (2011), we can
consider the IO to total inequality ratio in the regression (panel 3).
We see that the sign for relative IO remains negative and significant.27
Nevertheless, the use of this relative index is problematic because the
IO to total inequality ratio index depends, by construction, not only on
opportunities but also on the IE component. For this reason, the third
alternative (fourth panel in Table 2) is the most appealing one: using
the breakdown of total inequality into a between-groups component
27
Note that coefficients for IO in panel 2 and panel 3 are not comparable to each other
because in panel 2 the IO term is a level, while it is a ratio in panel 3.
(the IO term) and a within-groups component (the IE term). In this
manner, we attempt to isolate the effect of each component on growth.
The impact of the IO component is significantly negative and the impact
of the IE component is significantly positive.28 Moreover, the coefficient
associated with IE is higher and more significant than the one associated
with overall inequality in almost all the previous columns, which sup-
ports our hypothesis and is consistent with Fig. 4a–c. According to this
finding, the overall impact of total inequality on growth can be posi-
tive, negative or zero depending on the source of income inequality
(opportunity or effort) that dominates in the data.
After presenting these results, it would be beneficial to have a more
quantitative assessment of the main coefficients we estimate (those
for total inequality, IO and IE). The ϕ coefficient in Eqs. (2) and (3) re-
flects the immediate impact of the associated index on decade growth.
Since the resultant β estimated in Eq. (3) is always negative, we cannot
reject conditional convergence among the U.S. regions, and we can then
28
The IO component might be picking up the proportion of poor individuals in each
state, because the latter variable could have a negative correlation with income growth
(Ferreira et al., 2010). To check for this possibility, we have compared the IO coeffi-
cients with and without the proportion of poor individuals in each state (the poverty
rate as provided by the U.S. annual census data base) in the ‘base’ model. The negative
coefficient (and its significance) prevails for IO. Moreover, the poverty coefficients are
negative in most cases and significant in some cases. Thus, although poverty and IO
must be related, one does not exclude the other in explaining economic growth perfor-
mance. A complete analysis of this issue goes beyond the scope of this paper and it is a
promising extension.
118 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122
calculate the implied effect of inequality on the steady-state level of per
capita income. This is obtained by dividing the corresponding coeffi-
cient of inequality ϕ by β. In addition, we can compute the effect of a
one standard deviation change in inequality (total, IO, IO ratio and IE)
on growth and steady-state income in absolute and relative terms
(with respect to one standard deviation of growth or steady-state in-
come). Table 3 shows this assessment for system-GMM under the
base model.
Focusing first on the results in relative terms (last two columns in
Table 3), the effect of total inequality on growth is about 21%, while it
is 8.3% for the steady-state level of per capita income. This effect is
much higher (more than double) if we look at the impact of IE: 43.3%
for growth and almost 19.2% for the steady-state of income. For IO,
the effect is negative and significant, although lower in magnitude
than for IE: it is about −20% for growth and −9% for the steady-state
of income. Looking at the absolute effects, the results indicate that
increasing IE by one standard deviation could raise decade growth by
about 2.6 percentage points (the average decade growth in the 1970–
2000 period was 20.2%), and about 586 real US$ per person (the average
income in the 1970–2000 period was 14,363 US$ per person). At the
same time, decreasing IO by one standard deviation might raise growth
by about 1.24, and steady-state income by about 274 real US$ per
person.
These estimates tell us that gains coming from a reduction in IO
and an improvement in IE might be significant in terms of growth
and personal income. Given the importance of this finding, in the
next section we carry out a sensitivity analysis.
5. Robustness and sensitivity analysis
Durlauf and Quah (1999), Panizza (2002), Partridge (2005) and Barro
(2000), among others, have emphasized the importance of including
an extensive sensitivity analysis to show how robust the findings of a
reduced form regression are across alternative econometric techniques
or model specifications.
Using the IO and IE decomposition, Table 4 shows the estimated
results of our growth model (2) for pooled-OLS, (long-run) cross-
section OLS, FE and RE.29 In spite of the shortcomings of these methods,
we find that our previous results are robust to the econometric tech-
nique used. In particular, IO coefficients are always negative and IE
coefficients are always positive. In both cases, the coefficients are highly
significant. With respect to the estimated coefficients for all other
regressors, we find that they are also highly robust to the alternative
econometric method used.30
29
A battery of standard specification tests has been carried out (see the bottom of
Table 4): pooled-OLS is rejected against FE when the F-test is considered; the RE ap-
proach is rejected against pooled-OLS by the Breush–Pagan test; and, RE is rejected
against FE by the Haussman test. Therefore, among the three alternative methods,
the FE specification seems to be the most convenient. In this respect note that Banerjee
and Duflo (2003) point out that the FE technique basically uses within-state variability,
hence their estimates may produce inaccurate results for variables that mostly vary
cross-sectionally. To check for this possible problem, Panizza (2002) suggests
regressing the inequality variable with respect to time and regional dummies. A high
R2 would reflect that there is a limited within-state and within-decade variability,
which would lead to unstable FE estimates of the inequality coefficient. For our Theil
0, IE and IO estimates, the associated R2 are 0.38, 0.50 and 0.18, respectively, which
are clearly smaller than the 0.76 found for the Gini index in Panizza (2002).
30
Following Partridge (2005) and Panizza (2002) and in relation to these traditional
methods, we check the robustness of our results to the way the income lagged term is
included in model 2. In an inequality–growth regression, this is an important sensitiv-
ity analysis because the estimated coefficient of inequality may change dramatically
depending on whether the income lag is included or not and, if so, whether it is in
levels or logs (Panizza, 2002). In our case, the relevant finding is that while this change
might happen when including overall inequality alone, the results are highly robust
when overall inequality is broken down into IO and IE (i.e., estimates are always neg-
ative and significant for IO, while they are always positive and significant for IE). These
results are available on the web (see Appendix B).
Apart from this sensitivity analysis, we study the robustness of the
main results to the number of states considered in the regression, the
impact of anomalous states and the role of key regressors. To simplify
displaying the results, from now on we just show the IE and IO coef-
ficients for the system-GMM model.31
First, we run the regression in Eq. (3) changing the criterion of the
sample state selection. As discussed in Section 3, the 26 states for
which the analysis is conducted are those with no fewer than 50
observations for any decade. Table 5 shows the results when the
selection criterion changes to 20, 30, 70 or 100 observations. This ex-
ercise is interesting because population and inequality of opportunity
might be correlated by, among other factors, early settlement patters
and slavery.32 The coefficients of IE and IO always have the expected
sign, positive and negative, respectively. Moreover, these estimates
are in the majority of cases significant at least at the 10% level. We
observe, therefore, that the main results in Table 2 holds under samples
of different sizes.33
Second, we evaluate the change in our results when those states
with outliers highlighted in Fig. 2a and b (New Jersey, California,
Louisiana, Massachusetts, South Carolina and Maryland) are dropped
one state at a time (Table 6). We observe that the IO and IE coeffi-
cients are similar and their significance remains unchanged. Hence,
possible measurement errors in the IO and IE estimates for these
states seem not to be significantly affecting our regression results.
Finally, we run regression in Eq. (3) for the ‘base’ model by dropping
one main control variable at a time (Table 7). Significant changes in the
IO and IE coefficients might indicate that the excluded variable is an
important – indirect – channel through which these inequalities are
affecting growth.34 In this exercise, we find that the most important
variables are fertility and college, while the results remain basically
unchanged when welfare expenditure and industry mix are dropped. IO
and IE estimates remain basically unchanged when all other variables
in the base model are dropped (not shown in the Table).
By excluding the fertility (first two columns in Table 7), we obtain
a similar coefficient for IO and a clearly less positive coefficient for IE.
However, by excluding the college variable, we observe that the IO
coefficient becomes clearly higher, while the IE coefficients are simi-
lar. These results could imply that fertility and college are important
indirect channels affecting growth either through IE or IO. Excluding
welfare expenditure basically has no influence on IO and IE estimates.
A possible explanation for this result is that the impact of welfare
expenditures on growth and its correlation with IO and IE depend
on the financing scheme and the composition of total public expendi-
ture. Thus, a more precise analysis of this indirect channel would re-
quire considering the tax mix and the public expenditure composition
as well, which goes beyond the scope of this paper. Finally, excluding
the industry mix basically keeps the estimated IO and IE coefficients
unchanged.
31
In Appendix B (available on the web), the entire sensitivity analysis is also
performed for pooled OLS, FE and RE. Again, the results are highly robust to the econo-
metric method used.
32
Considering all the states with 20 or more observations, the sample increases from
26 to 33 states, where the new states in the sample are: Alabama, Arizona, Colorado,
Connecticut, Nebraska, Utah and Wisconsin. The last four states mentioned are deleted
from the sample when only those states with 30 or more observations are considered.
When considering 70 or more observations, the sample decreases from 26 to 21, and
the states excluded are: Iowa, Minnesota, Oregon Tennessee, and Washington. In addi-
tion, we also exclude from the sample Arkansas, Florida, Indiana, Kentucky, Louisiana,
Massachusetts, Georgia and New Jersey when we only consider states with 100 or
more observations in each decade.
33
Nevertheless, we must bear in mind that working with only 20 or 30 observations
makes the analysis based on the corresponding IO and IE indices less reliable. We con-
sider these cases as a matter of robustness.
34
An alternative interpretation would be that the IO and IE coefficients and the omit-
ted variable are spuriously correlated because they are driven by a third variable. We
acknowledge this point to an anonymous referee.
 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122 119

Table 3
Effects of one standard deviation change in inequality indices on growth and steady-state per capita income (system-GMM).

One std. dev. change of the One std. dev. change, relative to the
associated inequality index std. dev. of growth or pc real income

Mean Std β ϕ ϕ/−β On growth On s.s. income On growth On s.s. income

(level)

Theil 0 0.1128 0.0527 −0.7173 24.216 0.338 1.28 255.5 20.9 8.3
IO ratio 0.1457 0.0948 −0.7551 −22.909 −0.303 −2.17 −413.1 −35.5 −13.5
IE (IO-IE model) 0.0957 0.0436 −0.6486 60.736 0.936 2.65 586.4 43.3 19.2
IO (IO-IE model) 0.0172 0.0154 −0.6486 −80.448 −1.240 −1.24 −274.3 −20.3 −9.0
Growth (×100) 20.190 6.110
Pc real income 14363 3062
Pc real income, logs 9.5501 0.2134

For the steady-state (s.s.) income, the resultant ϕ/−β is multiplied by the average pc real income.

6. Concluding remarks
Models exploring the effect of income inequality upon economic
growth do not reach a clear-cut conclusion. We postulate in this paper
that one possible reason for this inconclusiveness is that income
inequality indices are indeed measuring at least two different sorts of
inequality: inequality of opportunity and inequality of effort. Though
this distinction has already been emphasized in the inequality-of-
opportunity literature, it has not yet been considered in the growth
literature. In this manner, the present paper contributes to the literature
on the relation between inequality and economic growth by incorporat-
ing the notion of inequality of opportunity in macro studies.
Using refined data of the PSID database for 26 U.S. states in 1970,
1980 and 1990, and applying system-GMM and, as a matter of robustness,

Table 4
Growth, IO and IE: alternative econometric methods.

Pooled OLS Long-run OLS FE panel regression RE panel regression

Small Base Small Base Small Base Small Base

Inequality of returns to effort (IE) 72.152*** 93.775*** 64.660 212.660** 72.026*** 52.455*** 77.725*** 77.939***
(2.63) (4.47) (0.62) (2.46) (3.81) (3.01) (3.68) (4.07)
Inequality of opportunity (IO) −105.870* −107.930*** −148.370** −277.87** −97.370** −87.462*** −117.624** −85.130**
(−1.50) (−2.48) (−1.89) (−2.37) (−2.01) (−2.76) (−2.10) (−1.97)
Income lag −0.006*** −0.005*** −0.006* −0.013*** −0.015*** −0.011*** −0.005*** −0.005***
(−6.66) (−6.91) (−1.43) (−6.94) (−10.11) (−11.16) (−7.04) (−6.90)
High school (% total graduated) −0.075 −0.244 −0.399 0.935 0.484** −0.383 −0.208** −0.329***
(−0.38) (−0.87) (0.25) (1.08) (2.08) (−0.98) (−2.24) (−2.58)
College (% total graduated) 2.075*** 1.950*** 5.025*** 10.800*** 6.099*** 5.034*** 1.566*** 1.815***
(5.50) (6.06) (2.74) (7.02) (7.73) (5.72) (7.12) (5.42)
Farm population (% total) – −24.420* – 39.950 – 45.626 – −18.797
(−1.58) (1.08) (0.38) (−0.89)
Mining empl. (% nonagric.) – −77.460 – −266.630 – −78.579 – −94.818
(−1.04) (−1.23) (−0.75) (−1.25)
Construction empl. (% nonagric.) – −44.217 – −407.180 – −163.587* – −67.085
(−0.69) (−1.26) (−1.53) (−1.09)
Manufact. empl. (% nonagric.) – −11.616 – 106.390* – −78.958 – −28.703
(−0.40) (1.49) (−1.02) (−1.18)
Transp. & pub. util. empl. (% nonagric.) – 141.860** – 581.320** – 344.277* – 94.809
(1.92) (1.97) (1.65) (1.14)
Fin. inst. & real estate empl. (% nonagric.) – −90.532 – −228.690** – −580.218*** – −131.678*
(−0.96) (−1.89) (−2.44) (−1.41)
Government empl. (% nonagric.) – −17.474 – −195.960*** – 40.357 – −25.463
(−0.48) (−2.81) (0.48) (−0.65)
Change nonagric. empl. previous decade – −0.101* – 0.149 – 0.014 – −0.097**
(−1.56) (0.44) (0.25) (−1.75)
Welfare exp. (% personal income) – 0.647 – 3.022 – 1.398 – −0.012
(0.60) (1.08) (0.83) (−0.01)
Fertility rate – −0.376*** – −0.218 – −0.259** – −0.341***
(−4.36) (−1.04) (−1.95) (−5.38)
Population age 65 or above (% total) 0.002 – −1.266* – 2.162** – 0.169 –
(0.01) (−1.47) (2.04) (0.65)
Population in Metropolitan area (%) 0.133*** – −0.243 – 0.204 – 0.118*** –
(3.21) (−1.23) (1.15) (3.07)
Num. observ. 78 78 26 26 78 78 78 78
Tests1
R2 0.4784 0.6409 0.6945 0.8954 – – – –
F-test – – – – 9.32 (0.00) 136.80 (0.00) – –
Breusch–Pagan test – – – – – – 1.53 (0.22) 0.74 (0.39)
Hausman test – – – – 105.52 (0.00) 51.67 (0.00) – –

Cross-sections included: 26; total pool (balanced): 78; robust cluster standard errors; t-statistics in parenthesis.
(*) significant at the 10% level; (**) significant at the 5% level; (***) significant at the 1% level (one-sided test).
A constant term and time dummies are included in all models. Regional dummies are also included in pooled-OLS and long-run OLS.
(1): p-values in parenthesis.
120 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122

Table 5
Sensitivity analysis of IO estimates by selecting the sample of states (system-GMM).

20 (1) 30 (2) 70 (3) 100 (4)

Small Base Small Base Small Base Small Base

IE 60.30* 104.48*** 50.09** 50.80*** 51.04** 70.96*** 16.51 105.33**

(1.37) (2.75) (2.03) (2.80) (2.32) (3.45) (0.53) (2.39)
IO −188.04*** −179.09** −99.55** −83.09** −93.93** −85.79** −26.26 −199.92**
(−2.37) (−1.81) (−1.75) (−1.97) (−1.68) (−1.81) (−0.39) (−1.69)

t-statistics in parenthesis. (*) Significant at the 10% level; (**) Significant at the 5% level; (***) Significant at the 1% level (one-sided test). Robust cluster standard errors.
(1) Cross-sections included: 33; apart from the states in the benchmark, the sample include: Alabama, Arizona, Colorado, Connect icut, Nebraska, Utah and Wisconsin.
(2) Cross-section included: 29; apart from the states in the benchmark, the sample include: Alabama, Arizona and Colorado.
(3) Cross-section included: 21; from the states in the benchmark, the sample excludes: Iowa, Minnesota, Oregon, Tennessee and Washington.
(4) Cross-section included: 13; in addition to those states excluded in (3), Arkansas, Florida, Indiana, Kentucky, Lousiana, Massachusetts, Georgia and New Jersey are also excluded.

Table 6
Sensitivity analysis of IO and IE estimates by dropping some states
(system-GMM).

New Jersey California Louisiana Massachusetts South Caroline Maryland

Small Base Small Base Small Base Small Base Small Base Small Base

IE 60.26** 86.20*** 67.83*** 53.62*** 36.81** 61.26*** 42.12* 57.74*** 50.40** 69.80*** 55.73** 82.93***
(1.88) (3.27) (2.64) (2.87) (2.14) (3.31) (1.55) (2.72) (1.91) (2.96) (2.01) (3.11)
IO −109.05** −82.10** −137.86*** −67.85* −54.88* −81.06** −88.37* −79.39** −93.69* −108.34* −97.06** −133.00**
(−2.10) (−1.82) (−2.68) (−1.46) (−1.35) (−1.67) (−1.60) (−1.81) (−1.51) (−1.52) (−1.67) (−1.95)

t-statistics in parenthesis. (*) Significant at the 10% level; (**) significant at the 5% level; (***) significant at the 1% level (one-sided test).

alternative pooled-OLS, long-run cross-sectional, random effects and
fixed effects regressions, we find robust support for a negative relation-
ship between inequality of opportunity and growth, and a positive
relationship between inequality of effort and growth. Despite these
promising results, a word of caution is in order because only two cir-
cumstances were used in the analysis. On the one hand, our results
indicate that those opportunity-related inequalities associated with
parental education and race lead to lower growth. However, nothing
can be said from our results about other opportunity-related inequal-
ities (explained by parental occupation, nationality or geographical
location, for example). On the other hand, it is likely that our IE esti-
mates still contain a certain amount of inequality of opportunity. Con-
sequently, the positive impact of effort-related inequalities on growth
found in this paper, though promising, has to be confirmed by future
studies.
With these cautions in mind, our results seem to indicate that
inequality of opportunity and inequality of effort may affect growth
through opposite channels. One possible explanation could be that
distinct returns to effort may encourage people to invest in education
and to exert an effort, while inequality of opportunity may not favor
human and physical capital accumulation in the more talented individ-
uals. In fact, Van de Gaer et al. (2001) have pointed out that inequality of
opportunity reduces the role that talent plays in competing for a posi-
tion by worsening intergenerational mobility. Moreover, our findings
would provide support for a general theoretical prediction of models
with multiple steady-states and borrowing constraints. In this context,
people with initial adverse circumstances would be likely exposed to
Table 7
Sensitivity analysis of IO and IE estimates by dropping main control variables
(system-GMM).
Fertility College Welfare
IE 42.24** 61.40** 59.64***
(2.34) (2.03) (3.15)
IO −77.47** −98.60** −78.72**
(−1.86) (−1.84) (−1.89)
Cross-sections included: 26; total pool (balanced): 78; t-statistics in parenthesis. Base
model (*) significant at the 10% level; (**) at the 5% level; (***) at the 1% level
(one-sided test). Robust cluster standard errors.
barriers for accessing credit or education, independently of their talent
or effort, which would undermine subsequent economic growth. Never-
theless, other models that are not based on borrowing constraints might
also rationalize our results (see, for example, Mejia and St-Pierre, 2008).
Making a distinction between inequality of income and inequality
of opportunity can throw some light upon several intriguing empirical
facts in the growth literature. Two examples are pointed out. Barro
(2000) shows a positive relationship between growth and inequality
within most developed countries, while this relationship is negative
when looking at the poorest countries. He proposes, as a tentative
explanation, the different role of capital markets. In particular, he con-
siders that problems of information (moral-hazard and repayment
enforcement problems) are larger in poor countries because they have
less-developed credit markets. However, he does not find empirical
evidence for this different role of capital markets. An alternative expla-
nation that would arise from the present paper is that differences in op-
portunity are more important within less-developed countries. At this
respect, some evidence is found in the inequality-of-opportunity litera-
ture as in Ferreira and Gignoux (2011), Rodríguez (2008), Cogneau and
Mesplé-Somps (2009), and Checchi and Peragine (2010).
Secondly, some empirical studies have found that the effect of
income inequality on growth is sensitive to the inclusion of some var-
iables like regional dummy variables (Birdsall et al., 1995). However,
the relationship between initial land inequality and growth is nega-
tive and robust to the introduction of regional dummies and other ex-
plicative variables (Deininger and Squire, 1998). Our proposal offers an
easy explanation for this empirical fact. Income inequality comes not
only from unequal opportunities but also from different levels of
effort. As a result, the effect of income inequality upon growth can
have a different sign depending on the kind of controls that are intro-
duced in the regressions. However, initial land inequality comes from
unequal opportunities and has a clear-cut negative effect upon growth.
Industry
It is clear from the discussion above that general redistributive
61.40***
policies may discourage unobservable effort by agents and, in this
(3.08)
−72.64** manner, decrease growth. However, policies that equalize opportunity
(−1.79) for the acquisition of the attributes necessary to compete for jobs and
careers may promote not only equity, but also economic growth. One
natural proposal for equalizing opportunity is ‘affirmative action’.
Following Roemer (1998), affirmative action can be defined as
 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122
preferential treatment in social policy of some (disadvantaged) groups
of individuals over others. Despite its appeal, some policies adopted as
affirmative action, such as racial or gender quotas for collegiate admis-
sion or targeted recruitment, have been criticized as a form of reverse
discrimination. In this respect, some authors have responded that the
equality-of-opportunity principle should be applied only when the
advantage in question is the acquisition of an attribute required to com-
pete for a position, not for the competition itself.35
Other arguments against affirmative action have been that it
undermines the achievements of minorities, encourages groups to
identify themselves as disadvantaged and may increase racial tension.
On the other hand, a highly unequal playing field might cause, among
other things, an important misallocation of talent by impeding or
discouraging disadvantage people from acquiring those attributes
required to compete in society. Our results seem to favor this second
possibility, supporting the hypothesis that reducing IO has a positive
net effect on the efficiency of the economy. However, further research
concerning these issues is clearly needed, and a complete under-
standing of the relationship between inequality and growth requires
more effort in constructing the appropriate databases that properly
represent social origins.
Appendix A. Data sources
Population (annual midyear population) and personal income data
come from the Regional Economic Accounts of the Bureau of Economic
Analysis (U.S. Department of Commerce, http://www.bea.gov/regional/
spi/drill.cfm). Personal income is the income that is received by per-
sons from all sources. It is calculated as the sum of wage and salary dis-
bursements, supplements to wages and salaries, proprietors' income
with inventory valuation and capital consumption adjustments, rental
income of persons with capital consumption adjustment, personal
dividend income, personal interest income, and personal current
transfer receipts, less contributions for government social insurance.
CPI data come from the U.S. Department of Labor (All Urban Consumers
CPI series: http://www.bls.gov/data/#prices). Employment data (total
and by type of industry) come from the Current Employment Statistics
of the Bureau of Labor Statistics (U.S. Department of Labor: http://
www.bls.gov/data/#employment). Education data comes from the
Historical Census Statistics on Education Attained in the U.S., 1940 to
2000 (U.S. Census Bureau): http://www.census.gov/population/www/
socdemo/education/introphct41.html.
Fertility is measures as the number of live births per 1000 women
between the ages of 15 and 44 years (the general fertility rate),
obtained from the Vital Statistics of the United States (http://www.
cdc.gov/nchs/products/vsus.htm). Regarding the poverty rate, we fol-
low the Office of Management and Budget's (OMB) Statistical Policy
Directive 14, and use the statistics from the Census Bureau. Poverty
rate is measured as the percent of families – and every individual in
it – whose total income is less than a family's threshold. Data comes
from the Census of Population, General Social and Economic Charac-
teristics, 1960, 1970, 1980 and 1990, http://www.census.gov/hhes/
www/poverty/census/cphl162.html. The percent of Population above
65 years old and the percent of Metropolitan population by states come
from the U.S. Census Bureau decennial Census of population, 1960–
2000. The 1970, 1980 and 1990 state and local welfare expenditure data
are from the U.S. Department of Commerce publication (Government
Finances).
35
For instance, Roemer (1998) proposes spending more educational resources, per
capita, on children from disadvantaged groups. In the same vein, Bourguignon et al.
(2007b) propose cash transfers conditional on specific behaviors, such as school atten-
dance; interventions to increase learning rates at public schools; health interventions
to increase basic knowledge of nutrition and hygiene; and, promotion of sports and
arts to reduce the appeal of violence.
121
Appendix B. Supplementary material
Supplementary data to this article can be found online at http://
dx.doi.org/10.1016/j.jdeveco.2013.05.004.
References
Aghion, P., Caroli, E., García-Peñalosa, C., 1999. Inequality and economic growth: the
perspective of the new growth theories. Journal of Economic Literature 37,
1615–1660.
Alesina, A., Perotti, R., 1994. The political economy of growth: a critical survey of the
recent literature. The World Bank Economic Review 8, 350–371.
Alesina, A., Perotti, R., 1996. Income distribution, political instability and investment.
European Economic Review 40, 1203–1228.
Alesina, A., Rodrik, D., 1994. Distributive politics and economic growth. Quarterly Journal
of Economics 109, 465–490.
Anderson, T.W., Hsiao, C., 1982. Formulation and estimation of dynamic models using
panel data. Journal of Econometrics 18, 47–82.
Arellano, M., Bond, S., 1991. Some tests of specification for panel data: Monte Carlo
evidence and an application to employment equations. Review of Economic Studies
58, 277–297.
Arellano, M., Bover, O., 1995. Another look at the instrumental-variable estimation of
error-components models. Journal of Econometrics 68, 29–52.
Arrow, K., Bowles, S., Durlauf, S., 2000. Meritocracy and Economic Inequality. Princeton
University Press.
Banerjee, A., Duflo, E., 2003. Inequality and growth: what can the data say? Journal of
Economic Growth 8, 267–299.
Banerjee, A., Newman, A.F., 1993. Occupational choice and the process of development.
Journal of Political Economy 101, 274–298.
Barro, R.J., 2000. Inequality and growth in a panel of countries. Journal of Economic
Growth 5, 5–32.
Bénabou, R., 1996. Inequality and growth. NBER Macroeconomics Annual 11, 11–74.
Bénabou, R., 2000. Unequal societies: income distribution and the social contract.
American Economic Review 90, 96–129.
Bertola, G., Foellmi, R., Zweimüller, J., 2005. Income Distribution in Macroeconomic
Models. Princeton University Press, Princeton.
Birdsall, N., Ross, D., Sabot, R., 1995. Inequality and growth reconsidered: lessons from
East Asia. World Bank Economic Review 9, 477–508.
Bleaney, M., Nishizama, A., 2004. Income inequality and growth—Does the relationship
vary with the income level? Economics Letters 84, 349–355.
Blume, L.E., Durlauf, S.N., 2001. Social Dynamics. Brookings Institution Press, Washington
D.C.
Blundell, R., Bond, S., 1998. Initial conditions and moment restrictions in dynamic panel
data models. Journal of Econometrics 87, 115–143.
Bond, S., Hoeffler, A., Temple, J., 2001. GMM estimation of empirical growth models.
Economics Papers 2001-W2 1.Economics Group, Nuffield College, University of
Oxford.
Bourguignon, F., 1979. Decomposable income inequality measures. Econometrica 47,
901–920.
Bourguignon, F., 1981. Pareto-superiority of unegalitarian equilibria in Stiglitz' model
of wealth distribution with convex savings function. Econometrica 49, 1469–1475.
Bourguignon, F., Ferreira, F., Menéndez, M., 2007a. Inequality of opportunity in Brazil.
Review of Income and Wealth 53, 585–618.
Bourguignon, F., Ferreira, F., Walton, M., 2007b. Equity, efficiency and inequality traps:
a research agenda. Journal of Economic Inequality 5, 235–256.
Bowles, S., Gintis, H., Osborne-Groves, M., 2005. Unequal Chance: Family Background and
Economic Success. Russell Sage Foundation and Princeton University Press, New York.
Checchi, D., Peragine, V., 2010. Inequality of opportunity in Italy. Journal of Economic
Inequality 8, 429–450.
Chiu, W.H., 1998. Income inequality, human capital accumulation and economic
performance. The Economic Journal 108, 44–59.
Clarke, G.R.C., 1995. More evidence on income distribution and growth. Journal of
Development Economics 47, 403–427.
Cogneau, D., Mesplé-Somps, S., 2009. Inequality of opportunity for income in five countries
of Africa. Research on Economic Inequality 16, 99–128.
Dardanoni, V., Fields, G.S., Roemer, J.E., Sanchez Puerta, M.L., 2006. How demanding
should equality of opportunity be, and how much have we achieved? In: Morgan,
S., Grusky, D., Fields, G.S. (Eds.), Mobility and Inequality: Frontiers of Research from
Sociology and Economics. Stanford University Press, Palo Alto, CA.
Davison, A.C., Hinkley, D.V., 2005. Bootstrap Methods and their Application. Cambridge
University Press, New York.
De la Croix, D., Doepke, M., 2003. Inequality and growth: why differential fertility matters.
American Economic Review 93, 1091–1113.
Deininger, K., Squire, L., 1998. New ways of looking at old issues: inequality and
growth. Journal of Development Economics 57, 259–287.
Durlauf, S.N., Quah, D.T., 1999. The new empirics of economic growth. In: Taylor, J.B.,
Woodford, M. (Eds.), Handbook of Macroeconomics, vol. 1. Elsevier Science, New
York.
Easterly, W., Levine, R., 1997. Africa's growth tragedy: policies and ethnic divisions.
Quarterly Journal of Economics 112, 1203–1250.
Ehrhart, C., 2009. The effects of inequality on growth: a survey of the theoretical and
empirical literature. ECINEQ WP 2009-107.
Ferreira, F., 2001. Education for the masses? The interaction between wealth, educa-
tional and political inequalities. The Economics of Transition 9, 533–552.
122 G.A. Marrero, J.G. Rodríguez / Journal of Development Economics 104 (2013) 107–122
Ferreira, F., Gignoux, J., 2011. The measurement of inequality of opportunity: theory
and an application to Latin America. Review of Income and Wealth 57, 622–657.
Ferreira, F.H.G., Leite, P.G., Ravallion, M., 2010. Poverty reduction without economic
growth? Explaining Brazil's poverty dynamics, 1985–2004. Journal of Develop-
ment Economics 93, 20–36.
Forbes, K., 2000. A reassessment of the relationship between inequality and growth.
American Economic Review 90, 869–887.
Foster, J., Shneyerov, A., 2000. Path independent inequality measures. Journal of
Economic Theory 91, 199–222.
Frank, M., 2009. Inequality and growth in the United States: evidence from a new state-
panel of income inequality measures. Economic Inquiry 47, 55–68.
Galor, O., Moav, O., 2004. From physical to human capital accumulation: inequality and
the process of development. Review of Economic Studies 71 (4), 1001–1026.
Galor, O., Zang, H., 1997. Fertility, income distribution and economic growth: theory
and cross-country race obviousness. Japan and the World Economy 9, 197–229.
Galor, O., Zeira, J., 1993. Income distribution and macroeconomics. Review of Economic
Studies 60, 35–52.
Galor, O., Moav, O., Vollrath, D., 2009. Inequality in land ownership, the emergence
of human capital promoting institutions and the Great Divergence. Review of
Economic Studies 76, 143–179.
Gouskova, E., Schoeni, R.F., 2010. Comparing Estimates of Family Income in the Panel
Study of Income Dynamics and the March Current Population Survey, 1968–2007.
Panel Study of Income Dynamics Technical Series Paper, 10-01.
Gradstein, M., Justman, M., 2002. Education, social cohesion and growth. American
Economic Review 92, 1192–1204.
Green, W.H., 2008. Econometric Analysis. Pearson Prentice-Hall.
Gupta, D.K., 1990. The Economics of Political Violence. The Effect of Political Instability
on Economic Growth. Praeger Publishers, New York.
Heathcote, J., Storesletten, K., Violante, G.L., 2008. Insurance and opportunities: the welfare
implications of rising wage dispersion. Journal of Monetary Economics 55, 501–525.
Hertz, T., Jayasundera, T., Piraino, P., Selcuk, S., Smith, N., Verashchagina, A., 2008. The
inheritance of education inequality: international comparisons and fifty-year
trends. The Berkeley Electronic Journal of Economic Analysis and Policy 7.
Knowles, S., 2005. Inequality and economic growth: the empirical relationship reconsidered
in the light of comparable data. Journal of Development Studies 41 (1), 135–159.
Kremer, M., Chen, D., 2002. Income distribution dynamics with endogenous fertility.
Journal of Economic Growth 7, 227–258.
Lefranc, A., Pistolesi, N., Trannoy, A., 2008. Inequality of opportunities vs. inequality of out-
comes: are Western societies all alike? Review of Income and Wealth 54, 513–546.
Lefranc, A., Pistolesi, N., Trannoy, A., 2009. Equality of opportunity and luck: definitions
and testable conditions, with an application to income in France. Journal of Public
Economics 93, 1189–1207.
Loury, G.C., 1981. Intergenerational transfers and the distribution of earnings.
Econometrica 49, 843–867.
Loury, G.C., 1989. Why should we care about group inequality? In: Shulman, S., Darity
Jr., W. (Eds.), The Question of Discrimination: Racial Inequality in the U.S. Labor
Market. Wesleyan.
Lucas, J.R., 1995. Responsibility. Clarendon Press, Oxford.
Marrero, G.A., 2010. Greenhouse gases emissions, growth and the energy mix in
Europe. Energy Economics 32 (6), 1356–1363.
Marrero, G.A., Rodríguez, J.G., 2011. Inequality of opportunity in the United States:
trends and decomposition. Research on Economic Inequality 217–246.
Marrero, G.A., Rodríguez, J.G., 2012. Inequality of opportunity in Europe. Review of
Income and Wealth 58 (4), 597–621.
Marshall, A., 1988. Income distribution, the domestic market and growth in Argentina.
Labour and Society 13 (1), 79–103.
Mason, A., 1988. Savings, economic growth and demographic change. Population and
Development Review 14, 113–144.
Mejia, D., St-Pierre, M., 2008. Unequal opportunities and human capital formation.
Journal of Development Economics 86, 395–413.
Mirrlees, J., 1971. An exploration in the theory of optimum income taxation. Review of
Economic Studies 38, 175–208.
Murphy, K.M., Schleifer, A., Vishny, R., 1989. Income distribution, market size, and
industrialization. Quarterly Journal of Economics 104, 537–564.
Panizza, U., 2002. Income inequality and economic growth: evidence from American
data. Journal of Economic Growth 7, 25–41.
Partridge, M.D., 1997. Is inequality harmful for growth? Comment. American Economic
Review 87 (5), 1019–1032.
Partridge, M.D., 2005. Does income distribution affect U.S. state economic growth?
Journal of Regional Science 45, 363–394.
Perottti, R., 1996. Growth, income distribution and democracy: what the data say. Journal
of Economic Growth 1, 149–187.
Persson, T., Tabellini, G., 1994. Is inequality harmful for growth? Theory and evidence.
American Economic Review 84, 600–621.
Rawls, J., 1971. A Theory of Justice. Harvard University Press, Cambridge, Ma.
Rebelo, S., 1991. Long-run policy analysis and long-run growth. Quarterly Journal of
Economics 107, 255–284.
Rodríguez, J.G., 2008. Partial equality-of-opportunity orderings. Social Choice and
Welfare 31, 435–456.
Roemer, J.E., 1993. A pragmatic approach to responsibility for the egalitarian planner.
Philosophy & Public Affairs 20, 146–166.
Roemer, J.E., 1998. Equality of Opportunity. Harvard University Press, Cambridge, M.A.
Roemer, J.E., Aaberge, R., Colombino, U., Fritzell, J., Jenkins, S., Marx, I., Page, M.,
Pommer, E., Ruiz-Castillo, J., San Segundo, M.J., Tranaes, T., Wagner, G., Zubiri, I.,
2003. To what extent do fiscal regimes equalize opportunities for income acquisi-
tion among citizens? Journal of Public Economics 87, 539–565.
Ruiz-Castillo, J., 2003. The measurement of the inequality of opportunities. Research on
Economic Inequality 9, 1–34.
Sen, A., 1980. Equality of what? Tanner Lectures on Human Values, vol. I. Cambridge
University Press.
Shorrocks, A.F., 1980. The class of additively decomposable inequality measures.
Econometrica 48, 613–625.
Stiglitz, J.E., 1969. The distribution of income and wealth among individuals. Econometrica
37, 382–397.
Székeli, M., 2003. The 1990s in Latin America: another decade of persistent
inequality, but with somewhat lower poverty. Journal of Applied Economics
VI, 317–339.
Van de Gaer, D., 1993. Equality of Opportunity and Investment in Human Capital.
no. 92. Catholic University of Leuven, Faculty of Economics.
Van de Gaer, D., Schokkaert, E., Martinez, M., 2001. Three meanings of intergenerational
mobility. Economica 68, 519–538.
Voitchovsky, S., 2005. Does the profile of income inequality matter for economic
growth? Distinguishing between the effects of inequality in different parts of the
income distribution. Journal of Economic Growth 10, 273–296.
World Bank, 2006. World Development Report 2006: Equity and Development. The
World Bank and Oxford University Press, Washington, DC.
Zou, H., Li, H., 1998. Income inequality is not harmful for growth: theory and evidence.
Journal of Development Economics 2, 318–334.