Social Science & Medicine 75 (2012) 754e760

Contents lists available at SciVerse ScienceDirect

Social Science & Medicine

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

Short report

Correlation or causation? Income inequality and infant mortality in fixed effects

models in the period 1960e2008 in 34 OECD countries
Mauricio Avendano a, b, c, *
a
London School of Economics and Political Science, LSE Health and Social Care, London, UK
b
Center for Population and Development Studies, Harvard School of Public Health, Cambridge, MA, USA
c
Department of Public Health, Erasmus Medical Center, Rotterdam, The Netherlands

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

Article history: Income inequality is strongly associated with infant mortality across countries, but whether this asso-
Available online 11 May 2012 ciation is causal has not been established. In their commentary in this issue of Social Science & Medicine,
Regidor et al. (2012) argue that this association has disappeared in recent years, and question the
Keywords: premise of a causal link. This paper empirically tests the impact of income inequality on infant mortality
Infant mortality in a fixed effects model that exploits the evolution of income inequality over a 38-year period, controlling
Income inequality
for all time-invariant differences across countries. Data came from the Standardized World Income
Socioeconomic
Inequality Database, containing yearly estimates for the period 1960e2008 in 34 countries member of
Population health
Social policy
the Organization for Economic Co-operation and Development (OECD), linked to infant mortality data
from the OECD Health database. Infant mortality was modelled as a function of income inequality in
a country and year fixed effects model, incorporating controls for changing economic and labour
conditions. In a model without country fixed effects, a one-point increase in the Gini coefficient was
associated with a 7% increase in the infant mortality rate (Rate ratio[RR] ¼ 1.07, 95% Confidence Interval
[CI] 1.04, 1.09). Controlling for differences across countries in a country fixed effects model, however,
income inequality was no longer associated with infant mortality (RR ¼ 1.00, 0.98, 1.01). Similar results
were obtained when using lagged values of income inequality for up to 15 years, and in models that
controlled for changing labour and economic conditions. Findings suggest that in the short-run, changes
in income inequality are not associated with changes in infant mortality. A possible interpretation of the
discrepancy between cross-country correlations and fixed effects models is that social policies that
reduce infant mortality cluster in countries with low income inequality, but their effects do not operate
via income. Findings highlight the need to examine the impact of more specific social policies on infant
mortality.
Ó 2012 Elsevier Ltd. All rights reserved.

Introduction Wilkinson & Pickett, 2011). Although income inequality is indeed

During the last decades, a wide array of studies has examined
the association between income inequality and health in high-
income countries. The rationale behind these studies is that
income inequality, independently of individual income, is associ-
ated with population health, so that more equal societies have
better health and lower mortality (Hales, Howden-Chapman, Sal-
mond, Woodward, & Mackenbach, 1999; Kaplan, Pamuk, Lynch,
Cohen, & Balfour, 1996; Lynch et al., 2001; Navarro et al., 2003;
* LSE Health and Social Care, Cowdray House, London School of Economics and
Political Science, Houghton Street, London WC2A 2AE, United Kingdom. Tel.: þ44
20 7955 7203.
E-mail addresses: M.Avendano-Pabon@lse.ac.uk, mavendan@hsph.harvard.edu.
consistently correlated with overall mortality across countries,
whether this association is causal has been brought into question
by a series of studies showing that in many instances, the associ-
ation does not consistently hold when controlling for potential
confounders (Mellor & Milyo, 2001). There is disagreement,
however, on the right choice of confounders and methodological
approaches, as this often determines the direction and strength of
the association (Glymour, 2008; Kawachi & Blakely, 2001; Mellor &
Milyo, 2001; Zimmerman, 2008).
A noticeable exception is the association between income
inequality and infant mortality. As suggested by Regidor et al.
(2012) commentary in this issue of Social Science & Medicine,
analyses based on data for the last decades of the 1970’s and 1980’s
show a consistent association of income inequality and mortality,
presumably as a result of welfare policies that promote income

0277-9536/$ e see front matter Ó 2012 Elsevier Ltd. All rights reserved.
doi:10.1016/j.socscimed.2012.04.017
 M. Avendano / Social Science & Medicine 75 (2012) 754e760
redistribution and reduce inequality (Hales et al., 1999; Navarro
et al., 2003). However, Regidor et al. (2012) show that the corre-
lation between income inequality and infant mortality across
countries observed prior to the 1990’s is no longer strong and
significant for the 1990’s and onwards. Their explanation for this is
that in many European countries, secular declines in infant
mortality have occurred in the absence of changes in income
inequality, so that the previously observed correlation has to a large
extent vanished in recent years (in this issue).
The commentary by Regidor et al. (2012) raises a fundamental
issue that appears obvious but is often forgotten. Although corre-
lations may hint at possible causal associations, correlations across
countries are particularly prone to innumerable sources of potential
confounding. We are primarily interested in this correlation to the
extent that it may reveal a potential causal association. If the asso-
ciation is causal, we expect that a change in income inequality alone
will lead to a change in infant mortality. In contrast, if the association
is non-causal, this may suggest that it is not income inequality per se
what underlies this association, but perhaps other characteristics or
policies that cluster in countries with low income inequality. The
inconsistency of the correlation over the last few decades suggests
that there may indeed be joint processes driving variation in both
variables, which generates uncertainty about whether this correla-
tion reflects a causal association.
Most studies examining the income inequality hypothesis,
including the commentary by Regidor et al. (2012) rely primarily on
correlations across countries or regions (Hales et al., 1999; Kaplan
et al., 1996; Lynch et al., 2001; Mackenbach, 2002). This approach
relies on the assumption that income inequality is randomly
distributed, so that, on average, countries are equal in all measures
and unmeasured characteristics. However, income inequality is
most likely the result of joint processes that are associated with both
infant mortality and income inequality, which may or may not be
related to the social policies that are believed to generate this
association. Some recent cross-national studies on infant mortality
have used multi-level models (Navarro et al., 2003), but this
approach suffers from the same source of bias because it does not
control for differences across countries in unmeasured covariates. A
series of important studies used first-difference models across US
states or selected countries (Clarkwest, 2008; Mellor & Milyo, 2001),
relating 10-year changes in life expectancy to changes in income
inequality. However, these studies were based on few and spaced
observations of income inequality, they were hampered by limited
variability in income inequality over a few set of observations
(Kawachi & Blakely, 2001), and they did not specifically examined
associations with infant mortality.
In this paper, I re-examine the association between income
inequality and infant mortality exploiting their evolution over 48
years in a country fixed effects model using yearly data for the period
1960e2008 from 34 countries member of the Organization for
Economic Co-operation and Development (OECD). As in a first-
difference model, a fixed effect approach effectively controls for all
measured and unmeasured time-invariant differences across coun-
tries because estimation relies solely on changes within countries
over time. This addresses a key concern in cross-country compari-
sons, and provides a potentially more robust estimate of the effect of
income inequality on infant mortality (Zimmerman, 2008).
Methods
Data
Data on income inequality came from the Standardized World
Income Inequality Database (SWIID), which maximises the
comparability of income inequality statistics for the largest possible
755
sample of countries and years, making it suitable for the purpose of
cross-national research (Solt, 2009). The SWIID departs from the
World Income Inequality Database (WIID), the most comprehen-
sive international dataset on income inequality developed by the
World Institute for Development Economics Research of the United
Nations University (United Nations University (UNU)-WIDER,
2008). The SWIID uses the Gini index as measure of income
inequality, and standardizes it using as much information as
possible from proximate years within the same country. Departing
from the original WIID, information is added from two series of the
Luxemburg Income study (LIS), providing information on net and
gross income, which serve as the baseline to which the WIID data is
standardized to construct the SWIID.
Full details on the dataset are available elsewhere (Solt, 2009). In
summary, to construct the SWIID, data from the WIID were first
classified according to the combination of reference unit (household
per capita, household adult equivalent, household without adjust-
ment, employee, and person) and income definition (net income,
gross income, expenditures and unidentified). The ratio of the Gini
index between each pair of categories was then used in order to
impute missing observations based on proximate years using loess
regression, which fits a smooth curve point-by-point through the
available data. Predictions were then generated using multi-level
modelling, with ratios modelled as a function of country-decade,
country and region-decade, country, region-decade, region and
advanced/developing country. The final dataset covers 153 countries,
with 3351 country-year observations on net income inequality. Data
have been assessed to represent a time-series of comparatively high-
quality. In addition, correlations with social and health variables,
including infant mortality and life expectancy, are similar between
the SWIID, WIDD and Deiningner and Squire datasets (Solt, 2009).
The SWIID represents a particular choice in the balance between
comparability and coverage. As an alternative option, I conducted
supplementary analysis using data from the newly available
University of Texas Inequality Project eUnited Nations Develop-
ment Organization (UTIP-UNIDO) dataset of gross household
income inequality developed by Galbraith and Kum (Galbraith &
Kum, 2005a, 2005b). The UTIP-UNIDO is a global dataset on
inequality based on the Industrial Statistics database published
annually by the United Nations Industrial Development Organiza-
tion (UNIDO). The EHII-GINI is estimated on the basis of the
systematic relationship between the UTIP-UNIDO measures of the
dispersion of pay across industrial categories in the manufacturing
sector, and the Deiningner and Squire database (Deininger & Squire,
1996a, 1996b). The result is a dataset of estimated household
income inequality that is more comprehensive and addresses some
of the problems of comparability in the Deiningner and Squire
database. It contains approximately 3200 observations over 36
years in the period 1963e1999 (Galbraith & Kum, 2005a, 2005b).
Data on infant mortality rates came from the OECD Health data-
base 2011 (Organization for Economic Co-operation and Develop-
ment, 2011), which contains yearly data for the period 1960e2008.
The infant mortality rate is defined as the number of deaths of chil-
dren under one year of age in a given year, expressed per 1000 live
births. The Eurostat database is the main data source for European
countries, except for Poland before 1978 (Central Statistics Office for
Poland), and the United Kingdom (UK Health Statistics, Office for
National Statistics). Data in the OECD Health database for other
countries is compiled from national health statistics offices
(Organization for Economic Co-operation and Development, 2011).
Analytical approach
Analyses on the SWIID focus on the GINI coefficient for net
household income, but results were virtually the same when using
756 M. Avendano / Social Science & Medicine 75 (2012) 754e760
the GINI for gross household income. To assess whether previously
reported associations were reproduced in the data, I first estimated
the Pearson correlation between the Gini coefficient and infant
mortality. I started by examining correlations separately for each
year, exploiting only variability across countries. Subsequently, I
examined the correlation within each country, exploiting variation
over the 48-year period.
Correlations provide a first glance of the association but they do
not allow controlling for potential confounders. Therefore, I
implemented a country fixed effects model to assess whether
annual changes in income inequality are associated with annual
changes in infant mortality rates in the period 1960e2008. The
major advantage of fixed effects methods (Wooldridge, 2002, pp.
461e468) is that by differencing out variability within countries, it
is possible to control for all time-invariant differences across
countries (characteristics that vary across countries but not over
time). Based on Poisson regression, the model estimated had the
following generic form:

Iceland, some European countries and Japan, and highest in Poland,
Log IMRjt ¼ a þ bGjt þ gXjt þ Cj þ Yt
where IMR is the infant mortality rate in country j and year t; G
stands for the Gini index of income inequality; X is a vector of
country time-variant control variables; C is a country fixed effect
that controls for time-invariant country characteristics; and Y is
a vector of year fixed effects that controls for time-variant
confounders that vary systematically across all countries. Time-
invariant country characteristics drop out of the model by incor-
porating C so that only variation within countries is utilized for
estimation. In supplementary analyses, I incorporated country-
specific linear time trends to account for differential trends across
countries. I used Poisson regression because mortality typically
follows a Poisson distribution, and our outcome is a count variable,
i.e., counts of death per 1000 live births. However, results using
regular linear models were qualitatively identical to those based on
Poisson regression.
Model estimation was done in five steps. First, I estimated
a model with only household income inequality. Subsequently, I
incorporated year fixed effects, country fixed effects, and country-
specific linear time trends. Since our identification strategy relies
on within-country variation over time in income inequality,
a possible limitation is that by including country-specific linear
time trends, some of the effect of income inequality on mortality is
absorbed in the country-specific trend coefficient. I therefore esti-
mated separate models without country-specific time trends,
incorporating yearly estimates for the following variables obtained
from the OECD StatExtracts database (Organization for Economic
Co-operation and Development, 2010) as controls: national
unemployment rates; share of employment in agriculture, industry
and services as percentage of total civilian employment; and
percentage of females out of total civilian employment. These
variables were chosen to roughly represent changes in the
economy and labour market structure of each country. Supple-
mentary models including Gross Domestic Product (GDP) and GDP
growth rate yielded very similar results as those with only the
variables above (results not shown). Standard errors were clustered
by country using Generalized Estimating Equations (GEE)
modelling.
A common criticism of fixed effects models refers to the use of
the appropriate etiologic period, the time interval required for
inequality to influence infant mortality (Glymour, 2008). This
relates to a more fundamental concern regarding fixed effects
models: this approach is particularly suited to identify short-term,
non-permanent effects of income inequality ‘shocks’ on mortality,
but it is less appropriate to identify long-run, permanent effects of
income inequality on health over a long period. Nevertheless, to
provide an estimation of the robustness of results to different
assumptions regarding plausible etiological periods, I implemented
models that related infant mortality to income inequality lagged 1,
5, 10 and 15 years. In models that incorporated controls for factors
that characterize the labour market structure, the corresponding
lagged values contemporary to the measure of income inequality
were used.
All analyses were conducted using SAS version 9.2 (Cary, NC).
Results
Table 1 shows ranges and means of variables for each country.
For 29 countries, data were available since the early 1960’s until
around 2005e2008. In some Central European countries and
Switzerland, data were available since the 1980’s, while data for
Iceland were only available since 2003. There were large variations
across countries in all variables. Infant mortality was lowest in
Mexico and Turkey. There was also large variation in average
unemployment rates, the proportion of the working population in
different sectors, and the proportion of females as a share of total
employment.
Income inequality varied greatly both across countries as well as
within countries over time (Table 1). The mean GINI coefficient was
lowest in Finland (23) and highest in Chile (51), but over the years
the GINI ranged from 20 to 36 in Finland and from 32 to 52 in Chile
(Table 1). Fig. 1 shows the GINI coefficient for a selection of coun-
tries in Europe (Panel A) and other OECD countries (Panel B), and
illustrates the major changes in income inequality experienced by
countries. For example, in some European countries, the U.S. and
Japan, inequality declined from 1960 to 1980, and increased since
the 1980’s. Mexico and Turkey have experienced up- and down-
turns in income inequality over the years, with a general trend of
decline over the entire period towards convergence with the other
OECD countries. These results show that there appears to be
sufficient variability over time as to identify an effect on infant
mortality.
Fig. 2 confirms the well-documented finding that there is
a positive correlation between income inequality and infant
mortality across countries. Results show that the correlation has
been persistent since the early 1960’s, with values around 0.8 until
1980, and values ranging between 0.6 and 0.7 in the period
thereafter. The correlation was significant at the 1% level in all
years. Associations were of a similar magnitude when restricting
analysis to European countries only (results not shown). Fig. 3
shows the within-country correlation between income inequality
and infant mortality exploiting year-to-year variation in the period
1963e2008. Earlier years are excluded because only a few countries
had valid data. Strikingly, results are now very different: in more
than half of the OECD countries, periods of high-income inequality
were correlated with lower infant mortality. In 13 countries higher
income inequality was associated with higher infant mortality,
with varying levels of strength and significance.
Table 2 shows rate ratios for the estimated change in infant
mortality for a one-point increase in the GINI coefficient. In a model
without controlling for differences across countries and years
(model 1), a one-point increase in income inequality was associated
with a 7% significant increase in infant mortality, which remained
after introducing year fixed effects (model 2 RR ¼ 1.07, 95%CI
[Confidence Interval] 1.04, 1.09). However, the association dis-
appeared after incorporating country fixed effects, which
controlled for time-invariant differences across countries (model
4). Results were similar in models introducing country-specific
time trends (model 5 RR ¼ 1.00, 95%CI 0.99, 1.02).
 M. Avendano / Social Science & Medicine 75 (2012) 754e760 757

Table 1
Descriptive characteristics of data for 34 OECD countries in the period 1960e2008.

Country Range Mean

Years Net income Net income Infant Unemployment % females share % employed % employed % employed
Gini Gini mortality rate of employment in agriculture in industry in services
Australia 1960e2008 22.5e33.5 29.2 9.6 5.9 39.3 6.2 27.8 65.6
Austria 1964e2008 22.7e35.1 26.4 7.7 3.6 41.7 7.7 36.9 55.5
Belgium 1963e2008 22.3e27.9 23.3 8.4 8.5 40.4 2.7 28.3 68.9
Canada 1961e2007 25.3e32.3 28.6 8.1 7.5 42.3 5.0 25.7 69.3
Chile 1963e2006 32.3e52.2 51.3 18.7 8.5 32.8 13.2 23.9 62.8
Czech 1987e2008 19.5e25.9 25.2 5.6 5.3 44.0 5.7 41.5 52.9
Denmark 1961e2008 21.8e27.2 25.1 7.9 5.5 45.6 6.7 28.2 65.9
Estonia 1981e2008 22.5e37.5 33.3 12.4 8.8 49.2 8.5 33.8 58.5
Finland 1962e2008 20.5e35.8 22.6 6.1 5.4 47.7 10.7 31.6 57.7
France 1962e2008 25.9e391 28.7 7.3 8.4 43.7 5.1 27.5 67.4
Germany 1960e2008 24.4e34.2 27.1 8.7 6.3 39.5 4.5 40.9 54.7
Greece 1967e2008 30.6e34.9 33.0 11.0 7.7 35.1 26.6 26.3 46.2
Hungary 1962e2008 21.2e32.3 25.7 19.0 7.5 45.8 6.7 33.9 58.9
Iceland 2003e2009 24.6e28.6 26.3 2.3 3.0 46.9 6.3 21.6 72.3
Ireland 1963e2008 28.5e33.6 32.4 8.9 7.1 32.7 15.7 28.6 55.8
Israel 1961e2005 23.3e37.0 30.7 8.5 8.9 45.4 2.2 23.7 74.2
Italy 1967e2008 29.0e36.4 33.7 9.6 8.7 34.3 10.2 33.8 57.3
Japan 1961e2005 25.1e35.9 28.5 6.2 2.2 39.9 9.3 34.3 56.0
Korea 1963e2008 26.8e39.4 32.8 9.9 3.8 39.5 19.5 28.1 47.6
Luxembourg 1963e2008 23.1e28.4 25.6 7.9 1.6 34.9 3.0 30.2 67.0
Mexico 1963e2008 44.5e53.9 47.9 39.6 3.2 33.6 20.1 24.9 55.5
Netherlands 1962e2008 23.1e30.3 25.9 7.6 5.4 38.9 4.8 26.6 68.6
New Zealand 1963e2007 23.2e36.4 26.6 11.4 3.8 36.1 10.7 32.1 56.8
Norway 1963e2008 21.1e25.6 23.3 7.8 2.6 44.7 7.0 27.1 65.9
Poland 1960e2008 22.2e32.0 26.8 21.1 13.8 45.1 19.1 30.8 50.5
Portugal 1968e2008 26.7e37.2 32.2 8.9 6.7 44.6 13.6 34.0 51.8
Slovak Republic 1987e2008 15.1e26.6 23.6 8.8 13.6 44.8 6.2 38.9 55.9
Slovenia 1987e2008 17.3e25.4 24.5 5.1 6.5 45.8 9.6 37.4 53.1
Spain 1963e2008 23.0e35.3 31.1 8.9 11.5 31.6 15.3 32.8 53.0
Sweden 1960e2008 19.7e27.5 22.9 6.8 2.7 47.2 5.1 29.9 65.1
Switzerland 1980e2005 26.8e38.7 30.8 5.9 2.4 42.0 4.6 28.4 67.3
Turkey 1963e2005 42.2e55.2 45.3 75.4 8.3 28.7 48.3 21.0 30.0
UK 1960e2008 25.7e35.8 27.8 9.6 5.2 41.6 2.6 34.6 62.9
USA 1960e2008 30.0e37.3 33.5 10.8 5.6 43.7 3.3 28.0 68.5

A potential concern is that country-specific linear time trends
partly pick up the effect of income inequality. Therefore, in model 6,
I excluded country-specific time trends and incorporated selected
measures of changes in the economy and labour market conditions
that may act as time-varying confounders (Table 2). The rate ratio
for income inequality was around the null and suggested no asso-
ciation between income inequality and infant mortality. Findings
were robust to additional controls for GDP per capita and GDP
growth (results not shown).
Supplementary models that experimented with lagged values
for income inequality up to 15 years showed qualitatively the same
results (Table 3). There was no significant effect of lagged income
inequality on infant mortality. There was a small but significant
effect for the 15-year lagged income inequality, but this was not
robust in models that excluded country-specific time trends and
included control variables for the corresponding lagged year.
Results were similar in models that simultaneously incorporated
the Gini coefficient for all years in a single model (results not
shown).
A potential concern is that the results are specific to the dataset
used. Therefore, analyses were replicated using the UTIP-UNIDO
database, and are presented in online Appendix Table 1. Overall,
results were very similar as for the SWIID database. In models that
incorporated both country and fixed effects (model 4), there was
a small but significant positive association between income
inequality and infant mortality (RR ¼ 1.02, 95%CI 1.00, 1.04).
However, this association disappeared when controlling for
country-specific time trends (RR ¼ 0.99, 95%CI 0.98, 1.00), or in

Fig. 1. The Gini coefficient in a selection of European (panel A) and other OECD (panel B) countries in the period 1960e2008.
758 M. Avendano / Social Science & Medicine 75 (2012) 754e760

Correlation: Inequality and infant mortality, 1963-2008

1

0.9

0.8

0.7
Pearson coefficient

0.6

0.5

0.4

0.3

0.2

0.1

Year
Fig. 2. Pearson correlation between household income inequality (Gini) and infant mortality rate for each year in the period 1963e2008 in 34 OECD countries.

models without country-specific time trends but with controls Discussion

(RR ¼ 1.00, 95%CI 0.99, 1.01). Results from models that included up
to 15-year lags were qualitatively similar to those based on the This paper extends previous empirical tests of the income
SWIID (online Appendix Table 2). inequality hypothesis by implementing a fixed effects model,

Year-to-year correlation: Income inequality and infant mortality, 1960-2009

1

0.8

0.6

0.4
Pearson coefficient

0.2

-0.2

-0.4

-0.6

-0.8

-1

Fig. 3. Within-country year-to-year Pearson correlation between household income inequality (Gini) and infant mortality rates in 34 OECD countries for the period 1960e2008.
 M. Avendano / Social Science & Medicine 75 (2012) 754e760 759

Table 2
Rate ratios of infant mortality as a function of net household income inequality for the period 1960e2008 in 34 OECD countries.

Model 1 (n ¼ 1280) Model 2 (n ¼ 1280) Model 3 (n ¼ 1280) Model 4 (n ¼ 1280) Model 5 (n ¼ 1280) Model 6 (n ¼ 1068)
Household income inequality (GINI) 1.07 (1.04, 1.10) 1.07 (1.04, 1.09) 1.00 (0.92, 1.09) 1.00 (0.98, 1.01) 1.00 (0.99, 1.02) 1.00 (0.99, 1.01)
Year (fixed effects) e Yes e Yes Yes Yes
Country (Fixed effects) e e Yes Yes Yes Yes
Country*year (linear trend) e e e e Yes e
Unemployment rate e e e e e 0.99 (0.98, 1.00)
% Females as share of total employment e e e e e 1.01 (1.00, 1.03)
% labour force in agriculture e e e e e 1.07 (1.00, 1.14)
% labour force in industry e e e e e 1.05 (0.98, 1.12)
% labour force in services e e e e e 1.04 (0.98, 1.11)

‘Yes’ indicates that variable is included in the model, but coefficients are omitted from Table.

effectively controlling for country characteristics that are constant
over time. Using a new dataset with a large coverage, I found that
changes in income inequality in the period 1960e2008 were not
associated with changes in infant mortality in OECD countries.
Results seem to be robust to different specifications and controls for
time trends or time-varying country characteristics. These findings
question the premise that changes in the distribution of income per
se are sufficient to generate changes in infant mortality.
Despite the potential advantages of fixed effects models, some
important limitations should be considered. Fixed effects models
control for time-invariant country characteristics, but results may
still be biased by time-varying confounders that are correlated with
changes in income inequality. For example, the introduction of
technology or changes in policies other than income redistribution
policies may occur contemporaneously or immediately precede
changes in income inequality. Although this remains a possibility,
results were robust across models including year fixed effects that
controlled for factors that varied systematically in all countries,
country-specific time trends, and time-varying measures of
changes in the economy and the labour market structure. All-in-all,
the impact of time-varying confounding is likely to be considerably
smaller than confounding in studies that exploit variability across
countries.
A more important limitation of the analysis refers to the diffi-
culty of identifying the right time interval or ‘etiologic period’
required for inequality to influence infant mortality in fixed effects
models (Glymour, 2008). The main analysis in this paper relates
income inequality and infant mortality at a given year, therefore
testing the hypothesis of short-term, non-permanent effects of
income inequality on infant mortality. In extended models,
however, I tested the hypothesis of long-term effects with a lag of
up to 15 years. Similarly, these models suggest that income
inequality many years earlier is not consistently associated with
infant mortality. On the other hand, fixed effects models do not
enable us to appropriately examine the hypothesis of long-run,
permanent cumulative effects, because identification relies on
fluctuations in income inequality within the same population, so
that infant mortality in later years is potentially ‘contaminated’ by
both cumulative effects and short-term shocks. Thus, the models
presented here do not reject the hypothesis of permanent effects.
Another limitation refers to the data on income inequality.
Ideally, data would come from a standardized assessment including
observations for each year and country. The SWIID database,
despite its advantages, does not contain information for each year,
but uses data from proximate years to impute missing observations.
This could potentially have led to a downward bias due to
smoothing of potentially more disruptive variation in income
inequality across years, on which identification in the fixed effects
model relies. I identified no ideal data alternatives to test the
impact of this potential bias. However, using the UTIP-UNIDO
database, which is based on measured dispersion of pay across
industrial categories in each year roughly capturing variations in
household income inequality, we found virtually the same results.
It is useful to discuss the findings from this study in the context
of previous interpretations of the income inequality hypothesis. It
has been argued that there is a curvilinear association between
income and health, due to marginal health returns to income. This
would predict that redistributing income from the rich to the poor
would lead to larger mortality reductions at the bottom than at the
top of the income distribution, resulting in a decline in total
mortality (Gravelle, 1998), and an association between income
inequality and total mortality. In addition, it has been argued that
income inequality per se may generate a special form of human
capital that leads to improvements in mortality for everyone in the
population, beyond the effects of income alone (Kawachi & Blakely,
2001; Wilkinson & Pickett, 2011). If at least one of these mecha-
nisms would operate in the short-term for infant mortality, we
would expect to observe an association with inequality in a fixed
effects model. Contrary to previous studies (Hales et al., 1999;
Kaplan et al., 1996; Navarro et al., 2003), however, results from this
study suggest that income inequality alone may not generate short-
term declines in overall infant mortality. A possible explanation is
that the correlation observed in cross-national studies is attribut-
able to social policies that vary systematically across countries. For
example, strong social protection policies such as universal access
to care and favourable maternity leave benefits may reduce infant
mortality directly and may at the same time cluster in countries
with strong redistribution policies, without necessarily having an
impact through reducing income inequality. The impact of these
and other policies may not necessarily be captured in a measure

Table 3
Rate ratios of infant mortality as a function of contemporaneous and lagged household income inequality for the period 1960e2008 in 34 OECD countries.

Contemporaneous Lag 1-year Lag 2-year Lag 3-year Lag 4-year Lag 5-years Lag 10-years Lag 15-years
Model 1 0.99 (0.98, 1.01) 1.00 (0.99, 1.02) 1.00 (0.99, 1.02) 1.01 (0.99, 1.02) 1.01 (0.99, 1.02) 1.01 (0.99, 1.02) 1.00 (0.99, 1.02) 1.01 (1.00, 1.02)
Model 2 1.00 (0.99, 1.02) 1.01 (0.99, 1.03) 1.01 (0.99, 1.02) 1.01 (0.99, 1.02) 1.01 (0.99, 1.02) 1.01 (0.99, 1.02) 1.00 (0.99, 1.00) 1.01 (1.00, 1.01)
Model 3 1.00 (0.99, 1.01) 1.00 (0.99, 1.01) 1.00 (0.99, 1.01) 1.00 (0.99, 1.01) 1.00 (0.99, 1.01) 1.00 (0.98, 1.01) 1.00 (0.99, 1.00) 1.00 (0.99, 1.01)

Model 1 includes the Gini coefficient, year fixed effects and country fixed effects.
Model 2 includes the Gini coefficient, year fixed effects, country fixed effects, and country-specific linear time trends.
Model 3 includes the Gini coefficient, year fixed effects, country fixed effects, and the corresponding contemporaneous or lagged national unemployment rate, share of
employment in agriculture, industry and services as percentage of total civilian employment, and percentage of females out of total civilian employment.
760 M. Avendano / Social Science & Medicine 75 (2012) 754e760
such as the Gini coefficient, because their effect does not neces-
sarily operate through income redistribution. By generating
equality of some form other than in income, such as in maternity
care, these social policies may influence infant mortality directly.
The fact that these policies are correlated with income redistribu-
tion policies, however, does not imply that a change in income
inequality will mechanically lead to a change in infant mortality.
In conclusion, in the short-run, changes in income inequality
alone are not associated with changes in infant mortality. This does
not imply, however, that several forms of redistribution of resources
may not have an impact on infant mortality, and it is likely that many
social policies that fuel resources to the poor contribute to decrease
infant mortality. A possible interpretation of the cross-country
correlation is that effective social policies that reduce infant
mortality directly tend to cluster in more equal societies. Each of
these policies is a distinct entity, however, and their effect on infant
mortality may not operate by reducing income inequality, but by
redistributing other resources towards the poor. Income inequality
may be too broad a measure and at least partly the result of processes
not fully amenable to policy, and we need evidence on the impact of
accessible social policies that reduce infant mortality, particularly
among the poor. Findings from this study highlight the need to
examine these different policies, which may have a direct impact on
infant mortality, and may explain cross-country correlations.
Randomized studies, or natural experiments that exploit reforms
during the last decades, can help us understand the impact of social
policies on infant mortality. They may also help us understanding
why although there is a correlation across countries, income
inequality alone may be insufficient to reduce infant mortality.
Acknowledgements
Mauricio Avendano is supported by a Starting Researcher grant
from the European Research Council (ERC) (grant no. 263684) and
a Eur-fellowship from the Erasmus University.
Appendix A. Supplementary material
Supplementary data associated with this article can be found, in
the online version, at doi:10.1016/j.socscimed.2012.04.017.
References
Clarkwest, A. (2008). Neo-materialist theory and the temporal relationship between
income inequality and longevity change. Social Science & Medicine, 66, 1871e1881.
Deininger, K., & Squire, L. (1996a). A new data set measuring income inequality.
World Bank Economic Review, 10, 565e591.
Deininger, K. W., & Squire, L. (1996b). Measure income inequality database. http://go.
worldbank.org/UVPO9KSJJ0 Accessed on 03.02.12.
Galbraith, J., & Kum, H. (2005a). University of Texas inequality project estimated
household income inequality data set. http://utip.gov.utexas.edu Accessed on
03.02.12.
Galbraith, J. K., & Kum, H. (2005b). Estimating the inequality of household incomes:
a statistical approach to the creation of a dense and consistent global data set.
Review of Income and Wealth, 115e143.
Glymour, M. M. (2008). Sensitive periods and first difference models: integrating
etiologic thinking into econometric techniques: a commentary on Clarkwest’s
“Neo-materialist theory and the temporal relationship between income
inequality and longevity change”. Social Science & Medicine, 66, 1895e1902,
discussion 1903e1898.
Gravelle, H. (1998). How much of the relation between population mortality and
unequal distribution of income is a statistical artefact? British Medical Journal,
316, 382e385.
Hales, S., Howden-Chapman, P., Salmond, C., Woodward, A., & Mackenbach, J.
(1999). National infant mortality rates in relation to gross national product and
distribution of income. Lancet, 354, 2047.
Kaplan, G. A., Pamuk, E. R., Lynch, J. W., Cohen, R. D., & Balfour, J. L. (1996). Inequality
in income and mortality in the United States: analysis of mortality and potential
pathways. British Medical Journal, 312, 999e1003.
Kawachi, I., & Blakely, T. A. (2001). When economists and epidemiologists disagree.
Journal of Health Politics, Policy and Law, 26, 533e541.
Lynch, J., Smith, G. D., Hillemeier, M., Shaw, M., Raghunathan, T., & Kaplan, G. (2001).
Income inequality, the psychosocial environment, and health: comparisons of
wealthy nations. Lancet, 358, 194e200.
Mackenbach, J. P. (2002). Income inequality and population health. British Medical
Journal, 324, 1e2.
Mellor, J. M., & Milyo, J. (2001). Reexamining the evidence of an ecological associ-
ation between income inequality and health. Journal of Health Politics, Policy and
Law, 26, 487e522.
Navarro, V., Borrell, C., Benach, J., Muntaner, C., Quiroga, A., Rodriguez-Sanz, M.,
et al. (2003). The importance of the political and the social in explaining
mortality differentials among the countries of the OECD, 1950e1998. Interna-
tional Journal of Health Services, 33, 419e494.
Organization for Economic Co-operation and Development. (2010). OECD.Stat. Paris:
Organization for Economic Co-operation and Development. doi:10.1787/data-
00285-en. Accessed on 30.01.12.
Organization for Economic Co-operation and Development. (2011). OECD health
data 2010 [computer software and file]. Paris: Organization for Economic Co-
operation and Development. http://new.sourceoecd.org.
Regidor, E., Martinez, D., Santos, J. M., Calle, M. E., Ortega, P., & Astasio, P. (2012).
New findings do not support the neomaterialist theory of the relation between
income inequality and infant mortality. Social Science & Medicine, 75, 752e753.
Solt, F. (2009). Standardizing World income inequality database. Social Science
Quarterly, 90, 231e242, SWIID Version 233.231, December 2011.
United Nations University (UNU)-WIDER. (2008). World income inequality database,
version 2.0c, May 2008. Accessed on 03.02.12.
Wilkinson, R., & Pickett, K. (2011). The spirit level: Why equality is better for everyone.
London: Penguin Group.
Wooldridge, J. (2002). Advanced panel data methods. Econometric analysis of cross
section and panel data. Cambridge: MIT Press.
Zimmerman, F. J. (2008). A commentary on “Neo-materialist theory and the
temporal relationship between income inequality and longevity change”. Social
Science & Medicine, 66, 1882e1894, discussion 1903e1888.