Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
50,972 characters · 8 sections · 80 citation commands
Income Inequality and Economic Growth: A Meta-Analytic Approach
\setcounter{page}{0} \pagenumbering{Alph}
\doublespacing
{Abstract. The empirical literature on the relationship between income inequality and economic growth has produced highly heterogeneous and often conflicting results. This paper investigates the sources of this heterogeneity using a meta-analytic approach that systematically combines and analyzes evidence from relevant studies published between 1994 and 2025. We find an economically small but statistically significant negative average effect of income inequality on subsequent economic growth, together with strong evidence of substantial heterogeneity and selective publication based on statistical significance, but no evidence of systematic directional bias. To explain the observed heterogeneity, we estimate a meta-regression. The results indicate that both real-world characteristics and research design choices shape reported effect sizes. In particular, inequality measured net of taxes and transfers is associated with more negative growth effects, and the adverse impact of inequality is weaker -- or even reversed -- in high-income economies relative to developing countries. Methodological choices also matter: cross-sectional studies tend to report more negative estimates, while fixed-effects, instrumental-variable, and GMM estimators are associated with more positive estimates in panel settings.} {\noindentKey Words: Meta-Analysis; Inequality; Growth; Publication-Bias.\\ JEL Codes: D31; O40; O15; C13.}\\
{ $^{(a)}$ CEIS, University of Rome Tor Vergata, Italy. Email: [email removed]\\ $^{(b)}$ Department of Economics, Management and Quantitative Methods, University of Milan, Italy. Email: [email removed] (corresponding author) \\
\thispagestyle{empty} \setcounter{footnote}{0} \pagenumbering{arabic} \doublespacing
In recent decades, the unprecedented availability of income distribution data has fueled a surge in studies examining the role of income inequality in economic growth. This body of literature is characterized by substantial heterogeneity in results and has not converged to a clear consensus regarding either the sign or the magnitude of the effect of income inequality on subsequent economic growth. This heterogeneity may reflect differences in econometric specifications, data sources, and country samples, as well as the existence of multiple underlying mechanisms through which income distribution affects growth, whose relative importance may vary across countries. This paper aims to shed light on the factors behind this heterogeneity by means of meta-analytic tools, which allow for the systematic combination and quantitative analysis of results from the existing empirical studies.
The theoretical literature has identified several channels through which income inequality can influence economic growth. Most of these channels imply a negative impact of the former on the latter. Prominent examples include credit constraints galor1993income, banerjee1993occupational, piketty1997dynamics, galor2004physical, fiscal policy persson1991inequality, fertility rates de2003inequality, domestic demand murphy1989income, social capital knack1997does, crime rates josten2003inequality, social instability alesina1996income, knack1997does, and corruption esteban2006inequality, galor2009inequality. In contrast, other contributions suggest that income inequality may foster growth. For instance, bottom-up redistribution could stimulate capital accumulation if the savings function is convex bourguignon1981pareto, galor2004physical, or provide greater incentives for research and development foellmi2006income. More generally, income distribution could serve as an incentive for the optimal allocation of productive factors, resulting in an equity-efficiency trade-off mirrlees1971exploration, lazear1981rank, rebelo1991long, okun2010equality.
The task of identifying the sign and magnitude of the net effect of income inequality on economic growth has fallen to the empirical analysis. The literature has developed along two main lines.\footnote{For a detailed discussion see baselgia2023inequality.} The first strand focuses on single transmission channels and tipically employs two-stage structural models, in which a mediator variable is first regressed to income inequality and subsequently linked to economic growth. The second strand employs a reduced-form approach, estimating the overall effect of income inequality on growth by regressing GDP per capita growth on an inequality index and a set of control variables. This paper exclusively focuses on the reduced-form strand for two key reasons. Firstly, results from the two approaches are not directly comparable. Secondly, the number of studies relying on two-stage structural models is too limited to conduct a separate meta-analysis.
Meta-analysis provides a natural framework to reconcile the diverging findings in the inequality–growth literature. By systematically collecting results from studies that meet explicitly stated inclusion criteria and applying standardized statistical procedures, it reduces the scope for selective interpretation and allows a robust assessment of the differences in reported results stanley2012meta. Given the pronounced variability that characterizes empirical estimates of the inequality–growth relationship, a meta-analytic approach is particularly well suited to identify the driving factors of this heterogeneity.
Previous meta-analyses provide conflicting evidence regarding the sources of heterogeneity in the inequality-growth relationship. While de2008meta identify the econometric estimator -- and specifically fixed effects -- as the main driver of variation, neves2016meta attribute heterogeneity primarily to the dataset structure (panel vs. cross-section) rather than the estimator employed. Furthermore, the two studies disagree on the role of data quality. However, they agree that income inequality is more detrimental to growth in developing countries and both document the presence of publication bias, albeit in different forms: selective publication based on statistical significance in neves2016meta and a bias toward negative estimates in de2008meta. Given these inconsistencies, we contribute to this debate by providing an updated and methodologically refined meta-analytic assessment. In particular, our analysis improves upon previous works in several ways. First, we update the pool of observations to include studies published up to 2025, thereby incorporating more than a decade of recent research. Compared to neves2016meta, our dataset adds 11 years of literature, while relative to de2008meta it adds 19 years.\footnote{The larger number of studies in de2008meta reflects the inclusion of working papers, whereas we focus exclusively on peer-reviewed journal articles. See Section (ref) for details.} Second, unlike neves2016meta, we collect all reported estimates from each study rather than relying on author-selected preferred specifications. This results in a substantially larger dataset of 531 estimates (compared to 49), increasing statistical power and reducing the influence of individual studies. Third, in contrast to neves2016meta, we restrict attention to estimates based exclusively on the Gini coefficient, thereby ensuring both conceptual and quantitative comparability across studies. Different inequality measures capture distinct aspects of the income distribution and are not simple rescalings of one another. Combining estimates based on heterogeneous inequality indices implicitly assumes that these measures are interchangeable, an assumption that is difficult to justify and may confound true economic heterogeneity with measurement differences. Fourth, we employ state-of-the-art methods to address publication bias, relying on the RoBMA–PSMA framework, and conduct a meta-regression analysis that includes a broader set of moderators.
Our results indicate that the average effect of income inequality on economic growth is negative and statistically significant, although economically small. We also find substantial between-study heterogeneity, consistent with the presence of multiple true effect sizes. The publication-bias analysis reveals strong evidence of selection based on statistical significance, pointing to an over-representation of significant results in the literature, but no evidence of systematic directional bias. To investigate the sources of heterogeneity, we estimate multilevel meta-regressions that account for dependence among multiple estimates reported within the same study. These results show that both real-world characteristics and research design choices contribute to variation in reported effects. In particular, inequality measured net of taxes and transfers is associated with more negative growth effects, and the adverse impact of inequality is stronger in developing countries. Methodological choices also matter: cross-sectional studies tend to report more negative estimates, an effect that is mitigated by the inclusion of regional controls, while Instrumental Variable (IV), fixed-effects, and Generalized Method of Moments (GMM) estimators are associated with more positive estimates in panel settings. By contrast, growth horizon length, data quality, and journal characteristics do not appear to systematically influence reported results.
The remainder of the paper is organized as follows. Section (ref) reviews the empirical literature estimating the effect of income inequality on economic growth. Section (ref) describes the study selection process and presents descriptive statistics summarizing the characteristics of the meta-analytic dataset. Section (ref) reports the meta-analysis. We first estimate the average effect size (Section (ref)), then test for the presence of publication bias (Section (ref)), and finally conduct a meta-regression to investigate the sources of heterogeneity across studies (Section (ref)). Section (ref) concludes.
The first empirical studies investigating the relationship between income inequality and economic growth emerged in the early 1990s, following the release of the first large-scale datasets on income distribution. This early strand of the literature alesina1994distributive, persson1991inequality, perotti1996growth, clarke1995more relies on ad hoc datasets constructed by aggregating inequality measures from multiple sources. Due to limited time coverage, these studies primarily employ cross-sectional models. Typically, the growth rate of GDP per capita is regressed on an inequality index and a set of control variables, as illustrated by the following specification:
where $\alpha$ is a constant term, $y_i$ denotes the growth rate of GDP per capita in country $i$, typically measured over a long horizon of 25–40 years, and $I_{ji}$ represents inequality index $j$ for country $i$, usually observed at the beginning of the sample period to mitigate concerns of reverse causality. The associated coefficient $\theta_j$ captures the effect of inequality on growth. Although the Gini coefficient is the most commonly used measure, several studies also rely on income shares, decile ratios, and other summary indicators of inequality. The vector $X_{ki}$ includes standard growth controls such as initial income, education, and investment. Overall, this early literature predominantly finds that higher income inequality is associated with lower subsequent economic growth.
This negative relationship was challenged in the late 1990s with the emergence of studies based on panel data. The release of the deininger1996new dataset provided repeated observations of income distribution over time, enabling researchers to exploit within-country variation. This development gave rise to panel regressions in which the sample period is divided into shorter subperiods - typically five to ten years - rather than a single long growth spell. Studies in this second wave of the literature li1998income, partridge1997inequality, forbes2000reassessment, often employing fixed-effects estimators, tend to find a positive effect of income inequality on subsequent economic growth.
Subsequent research, however, has questioned the robustness and interpretation of these findings. Some contributions argue that the positive relationship uncovered in early panel studies is spurious. banerjee2003inequality depart from the linear framework in equation ((ref)) and document a non-linear relationship, whereby any change in inequality - regardless of its direction - reduces subsequent growth. They attribute this pattern to measurement error in inequality, which tends to be more severe during periods of economic distress characterized by sharp output declines. Similarly, scholl2019re challenge the foundations of the early panel evidence, showing that the positive association between inequality and growth is largely driven by transition economies in Eastern Europe during the 1990s. In their interpretation, the collapse of the Soviet Union led simultaneously to rising inequality and subsequent economic recovery, generating a correlation that does not reflect a causal relationship.
Other studies suggest that the relationship between inequality and growth is inherently context-dependent. Several authors find that once countries are grouped by their level of development, inequality promotes growth only in high-income economies, while it hampers growth in less developed ones barro2000inequality, khalifa2010income, grundler2018growth. In contrast, brueckner2018inequality report the opposite pattern, with inequality fostering growth in low-income countries but reducing it in high-income economies.
The impact of inequality may also depend on which segment of the income distribution is considered. voitchovsky2005does argues that relying on a single inequality index masks important distributional dynamics and shows that inequality at the top of the distribution can stimulate growth, whereas inequality at the bottom is detrimental. Related evidence is provided by bartak2020inequality, who document heterogeneous effects across different parts of the income distribution, although their overall results point to a negative impact of inequality on growth. Along similar lines, marrero2022growth find that inequality has no direct effect on growth but operates indirectly through its association with poverty, which significantly reduces economic development.
Finally, the time horizon over which growth is measured appears to matter. herzer2012inequality criticize the common practice of regressing growth rates on inequality levels and instead apply cointegration techniques using variables in levels. Their results indicate a negative long-run relationship between inequality and economic growth. Comparable conclusions are reached by berg2018redistribution, who identify a negative long-term effect and a positive short-term effect of inequality. These findings are broadly consistent with those of halter2014inequality and davis2011institutional. Evidence in el2019impact further suggests that inequality may stimulate growth over intermediate horizons.
This section describes the construction of our meta-analytic dataset and summarizes its main characteristics. We searched for English-language peer-reviewed articles indexed in the Scopus database using the keyword “growth” in combination with either “inequality” or “distribution” in titles or abstracts. We complemented this search with a title-based query on Google Scholar. Overall, this procedure yielded 3,461 records (3,245 from Scopus and 216 from Google Scholar). We additionally screened references cited in the selected articles and included all studies analyzed by de2008meta and neves2016meta that were not already part of our sample and satisfied our inclusion criteria -- described below in detail -- adding 25 further studies.
After screening abstracts, we excluded 3,438 articles that did not examine the relationship between income inequality and economic growth. The full texts of the remaining 48 studies were then assessed in detail. We excluded studies and estimates that did not conform to the reduced-form framework described in equation ((ref)) or its panel-data counterpart.\footnote{While acknowledging their contribution to the literature, studies estimating non-linear relationships were excluded because their results are not directly comparable with the bulk of the empirical literature and their number is insufficient to support a separate meta-analysis.} We also excluded studies conducted exclusively at the national level to avoid confounding effects driven by country-specific institutional or regional factors. Furthermore, studies measuring inequality using concepts other than income - such as wealth, land, or human capital - were not considered.\footnote{The limited number of such studies does not allow us to control for those specific concepts of inequality in the meta-analysis.}
To ensure comparability across studies, all included estimates were required to meet a set of minimum specification criteria. First, we retained only estimates based on the Gini coefficient.\footnote{Alternative inequality measures, such as the Theil index or decile ratios, capture different features of the income distribution. Pooling estimates based on heterogeneous inequality measures would therefore confound substantive heterogeneity with measurement differences. For comparison purposes with neves2016meta, we replicate the meta-regression analysis including non-Gini-based estimates in Appendix.} Second, we excluded estimates that did not control for initial GDP per capita, given its well-known role in the growth dynamics since the seminal contribution of solow1956contribution. Finally, we removed outliers to prevent extreme values from disproportionately influencing the results. An estimate was classified as an outlier if either its effect size or its standard error deviated from the sample median by more than ten interquartile ranges mccracken2016fred. Applying these minimum criteria led to the exclusion of an additional 15 studies. Figure (ref) in Appendix (ref) summarizes the search and screening process using a PRISMA flow diagram.
The final dataset comprises 33 published articles and a total of $N = 531$ estimates. The effect size extracted from each study is the coefficient $\theta_j$ associated with the Gini index $I_{ji}$ in equation ((ref)). This coefficient measures the change in the annual growth rate of GDP per capita associated with a one-point increase in the Gini index, measured on a 0–100 scale.\footnote{For example, a coefficient of $\theta = -0.0345$ implies that a one-point increase in the Gini index reduces annual GDP per capita growth by 0.0345 percentage points. When studies employ alternative scaling conventions, estimates are rescaled to ensure consistency across observations.}
Table (ref) reports key characteristics of the studies included in the meta-analysis, including the number of estimates per study (Observations, Column 3), the structure of the dataset (cross-sectional or panel, Column 4), the average effect size within each study $\theta$ (Column 5), whether inequality is measured before or after taxes and transfers (Income Concept, Column 6), the type of countries included (Column 7), and the estimation methods employed (Column 8). Additional information on the collected variables and their summary statistics is provided in Appendix (ref).\footnote{All authors jointly conducted the literature search, independently reviewed the included studies, and collaborated in coding the extracted estimates. Coding decisions were cross-validated to ensure accuracy and consistency.}
\newgeometry{left=1.5cm,right=1.5cm,top=3cm,bottom=3cm}
\newgeometry{left=3cm,right=3cm,top=3cm,bottom=3cm}
The number of studies included in our dataset differs from de2008meta and neves2016meta, even when the period of analysis overlaps. In particular, some studies included in de2008meta do not appear in our dataset since they are unpublished working papers instead of peer-reviewed journal articles. We restrict attention to published studies in order to ensure minimum standards of methodological scrutiny concerning data quality, econometric set-up and robustness transparency. Additionally we exclude, some studies analyzed by neves2016meta as they rely on inequality indices other than the Gini index.
Figure (ref) displays the distribution of the collected effect sizes, sorted by magnitude. The estimates span both positive and negative values and cluster around zero, offering no clear indication of the sign or magnitude of the average effect. This dispersion further motivates the meta-analytic and meta-regression analyses presented in the next section.
A central objective of a meta-analysis is to estimate the average effect size implied by the empirical literature. Let $\theta_q$ denote the $q$th reported effect size and let $w_q$ be a precision weight. The precision-weighted mean is given by
where $\hat{\mu}$ is the overall size effect, and $w_q$ is the associated weight given by the inverse of its variance, $1/\sigma^2_q$.
We estimate $\mu$ using the Unrestricted Weighted Least Squares (UWLS) approach of stanley2015neither. UWLS treats each reported estimate as an observation of an underlying average effect and allows for heteroscedasticity proportional to the reported sampling variances. Operationally, UWLS is implemented as a weighted regression of $\theta_q$ on a constant, using weights proportional to $1/\sigma_q^2$, and inference is based on study-level cluster-robust standard errors to account for multiple, non-independent estimates reported within the same study.
The UWLS results yield $\hat{\mu} = -0.019$ (p-value $<0.01$), implying that a one-point increase in the Gini index (on a 0--100 scale) is associated, on average, with a 0.019 percentage-point reduction in the annual growth rate of GDP per capita. Although statistically significant, the estimated average effect is economically small.
Publication bias arises when the probability that a result appears in the published literature depends on its statistical significance or on how well it conforms to prevailing expectations. Such selectivity can distort the evidence base and bias estimates of the average effect. A common diagnostic is the small-study effect: estimates from less precise studies (with larger standard errors) are expected to be more dispersed around the underlying effect. In the absence of selective publication, the distribution of effect sizes should be approximately symmetric around the mean across the range of standard errors. If, however, statistically insignificant or undesired results are less likely to be published, the observed distribution may become asymmetric.
Figure (ref) presents a funnel plot of the collected estimates, plotting each effect size against its standard error. Visual inspection suggests some asymmetry, with a concentration of estimates in the statistically significant negative region, although the pattern is not sufficiently sharp to draw firm conclusions from the plot alone.
Since funnel-plot interpretation is inherently subjective and small-study patterns may also arise from heterogeneity, we implement a formal publication-bias assessment using the Robust Bayesian Meta-Analysis Publication-bias and Small-study effects Model Averaging (RoBMA--PSMA) framework, following bartovs2023robust.\footnote{We restrict attention to two-sided selection models because there is no strong a priori basis for assuming one-sided selection in a specific direction in this context.} RoBMA--PSMA combines multiple candidate meta-analytic models that differ in their assumptions about the presence of the effect, heterogeneity and selective publication and averages across them using posterior model probabilities. In particular, the model set includes standard meta-analytic specifications, selection models based on two-sided $p$-value intervals (weight functions), and regression-based adjustments for small-study effects (Precision-Effect Test, PET and Precision-Effect Estimate with Standard Errors, PEESE). Further implementation details are reported in Appendix (ref).
Table (ref) summarizes the RoBMA--PSMA results. They provide overwhelming posterior support for a non-zero overall effect, substantial heterogeneity, and selective publication. The Bayes Factors (BF) in Table (ref) (upper panel) imply that models allowing for heterogeneity and publication bias fit the observed distribution of effect sizes substantially better than models that omit these components.\footnote{They are largely above the conventional threshold of $BF>10$ bartovs2023robust.}
The model-averaged estimate of the overall effect $\hat{\mu}$ is negative and close in magnitude to the UWLS estimate, indicating that the average effect, albeit significant, remains economically modest even after accounting for publication bias and heterogeneity. At the same time, the strong evidence for heterogeneity ($\hat{\tau}> 0$) suggests that a substantial share of the variation in $\theta_q$ reflects differences in underlying true effects across studies rather than sampling error alone.
The estimated publication weight function $\omega$ (also reported in Figure (ref)) summarizes relative publication probabilities across two-sided $p$-value intervals. Estimates with $p$-values below 0.05 receive the highest relative publication weight, whereas less statistically significant results are assigned substantially lower weights. This pattern is consistent with selective publication, whereby authors or journals favor statistically significant findings irrespective of their sign, leading to an over-representation of statistically significant results in the published literature. This finding is in line with neves2016meta.
Finally, the regression-based small-study adjustments (PET and PEESE in Table (ref)) receive negligible posterior support, suggesting that precision-related small-study effects play a limited role in explaining the observed data. Instead, selective publication based on statistical significance provides a more plausible description of the publication process in this literature. This finding aligns with neves2016meta and contrasts with de2008meta, who report evidence consistent with a left-skewed small-study effect using PET-type tests.
The strong evidence for heterogeneity implies that the average effect, although statistically significant, is likely to mask substantial variation in true effects across studies. Explaining this heterogeneity is therefore the focus of the meta-regression analysis in the next section.
Meta-regression is a useful tool that aims to explain heterogeneity in reported effect sizes by relating them to observable study characteristics, commonly referred to as moderators. Formally, meta-regression is analogous to a standard regression model in which the dependent variable is the reported effect size and the covariates are study-level or estimate-level characteristics.
Meta-regression can be estimated under either a Fixed-Effects (FE) or a Random-Effects (RE) framework. The FE approach assumes that all heterogeneity in reported effects can be fully explained by the included moderators greenland1987quantitative. In contrast, the RE approach allows for residual heterogeneity that remains unexplained by the observed moderators berkey1995random. Given the complexity of the inequality--growth relationship and the unlikelihood that the available moderators exhaust all relevant sources of heterogeneity, we adopt a random-effects specification.
Let $\theta_q$ denote the $q$th reported effect size. The random-effects meta-regression can be written as
where $\beta_0$ is the conditional mean, $\mathbf{M}_q$ is a vector of moderators, $\boldsymbol{\beta}$ is the associated vector of coefficients, and $\xi_q$ is an error term with variance $\sigma_q^2 + \tau^2$, combining sampling uncertainty and unexplained heterogeneity. Estimation is performed using inverse-variance weights.
This formulation, however, treats all effect sizes as independent observations. In our dataset, individual studies report multiple estimates, implying that effect sizes are clustered within studies. Ignoring this dependence leads to over-representation of studies reporting many estimates and may bias inference. A common solution is to retain a single, representative estimate per study, as in neves2016meta. While this approach avoids dependence, it comes at the cost of discarding a substantial amount of variation contained in the remaining estimates.
To address this issue, we adopt a multilevel (hierarchical) meta-regression model that explicitly accounts for the nesting of effect sizes within studies goldstein2011multilevel. Let $\theta_{qr}$ denote the $q$th estimate reported in study $r$, with known sampling variance $\sigma_{qr}^2$. The model is specified as
where $u_r$ captures unobserved between-study heterogeneity, assumed to be normally distributed with variance $\tau^2_{\text{between}}$, and $w_{qr}$ captures within-study heterogeneity across multiple estimates from the same study, with variance $\tau^2_{\text{within}}$. The sampling error $\varepsilon_{qr}$ has mean zero and variance $\sigma_{qr}^2$. The model is estimated by Restricted Maximum Likelihood (REML). This three-level specification decomposes total variability in reported effects into sampling uncertainty, within-study heterogeneity, and between-study heterogeneity.
Given the large number of potential moderators and the risk of multicollinearity, we use Bayesian Model Averaging (BMA) as a data-driven filtering approach to select a parsimonious set of moderators steel2020model. Specifically, we estimate a BMA regression with the reported effect size as the dependent variable and a broad set of candidate moderators as potential predictors. BMA evaluates many alternative model specifications and summarizes results by averaging across models, weighted by their posterior support in the data. We retain moderators with Posterior Inclusion Probabilities (PIP) above 0.1 and exclude the remaining candidates from the subsequent multilevel meta-regression. The full list of candidate moderators and their PIPs is reported in Appendix (ref). Table (ref) reports the included moderators along with their estimated meta-regression coefficients.
The selected moderators include variables that characterize four key features of the included studies: the characteristics of the dataset, the estimator employed, the covariates incorporated in the regression, and the statistics of the journal in which the study was published. Additionally, in order to control for potential direction-bias in terms of small-study effect, we included the standard error among the set of moderators.
\paragraph{Dataset characteristics}
Several features of the underlying dataset can shape the estimated inequality--growth relationship. First, the structure of the dataset matters. Cross-sectional designs are more exposed to omitted-variable bias and cannot control for time-invariant omitted variables as panel estimators. We therefore include a dummy variable $Cross-sectional$ equal to one for cross-sectional estimates and zero for panel-based estimates. The estimated coefficient is negative and highly significant, implying that cross-sectional specifications tend to report more negative effects of inequality on growth. This result is consistent with neves2016meta but differs from de2008meta, who find no systematic role for dataset structure.
Second, the time horizon over which growth is measured may affect results, as some channels may operate over different horizons herzer2012inequality, berg2018redistribution. We therefore include the variable Growth Span, defined as the number of years over which GDP per capita growth is computed. The estimated coefficient is negative, but -- in contrast with de2008meta -- small and statistically insignificant, indicating limited evidence that longer growth spells systematically shift reported effects in this sample.
Third, the effect of inequality may vary with countries’ level of development, since the relative importance of transmission channels can differ across income levels chambers2010relationship, deininger1998new, grundler2018growth. We include a dummy variable High-income equal to one when an estimate is obtained from a sample restricted to high-income economies or is associated to an interaction term with an high-income economy indicator, and zero otherwise. The coefficient in Table (ref) is positive and statistically significant, suggesting that estimated effects are less negative (or more positive) in advanced-economy settings. This finding aligns with both de2008meta and neves2016meta.
Fourth, inequality measured net of taxes and transfers may convey different information than inequality measured on market income. Net inequality reflects both market outcomes and the redistributive system and may be more closely related to channels operating through disposable income (e.g., credit constraints and human capital accumulation). By contrast, in the fiscal-policy channel persson1991inequality, market inequality may be the relevant object as it affects redistributive pressures and the size of distortionary taxation. To capture this distinction, we include a dummy variable ($Net$ $Inequality$) equal to one when inequality is measured after taxes and transfers and zero otherwise. As shown in Table (ref), the coefficient on $Net$ is negative and statistically significant, indicating that net-inequality measures are associated with more adverse estimated growth effects.
Fifth, we control for inequality data quality. Prior to the harmonized dataset introduced by deininger1996new and subsequent efforts, studies often relied on ad hoc compiled datasets with limited comparability across countries and time. We include the dummy High Quality Data, equal to one when inequality is constructed from survey-based, nationally representative sources with broad income coverage, and zero otherwise.\footnote{High-quality sources include: deininger1996new, the Luxembourg Income Study (LIS), the World Income Inequality Database (WIID), the Standardized WIID (SWIID) solt2016standardized, the Estimated Household Income Inequality (EHII) by the University of Texas Inequality Project and iradian2005inequality.} The coefficient is not statistically significant, consistent with neves2016meta but in contrast to de2008meta.
Finally, we also test whether the time dimension of the sample influences the results. Studies based on shorter sample periods might capture short-run dynamics or extraordinary events, while longer time horizons help minimize such risks. Therefore, we include Number of years, defined as the span between the first and last years of the underlying data. The estimated coefficient is close to zero and statistically insignificant, suggesting that this dimension does not systematically explain variation in reported effect sizes.
\paragraph{Estimator}
Because inequality and growth may be jointly determined, estimators addressing endogeneity -- such as IV or GMM -- may yield different estimates than methods that treat inequality as exogenous. Moreover, since cross-sectional and panel designs are associated with different signs of the effect size, we allow the impact of IV methods to differ by dataset structure. Specifically, we include the interactions $IV \times Cross$ and $IV \times (Panel)$ to distinguish IV estimates in cross-sectional and panel settings respectively. The results suggest that IV methods amplify the negative estimates in cross-sectional designs and the positive estimates in panel designs, although the latter effect is only marginally significant.
We additionally include indicators for estimators used primarily in panel settings. Difference GMM captures estimates obtained with difference-GMM methods, and Fixed effects identifies fixed-effects estimators. Both coefficients are positive and highly significant, indicating that these estimators are associated with less negative (or more positive) reported effects in panel studies. This result aligns with de2008meta, but contrasts with neves2016meta.
\paragraph{Covariates}
Several contributions persson1991inequality, perotti1996growth report that the inclusion of regional dummies weakens the effect of inequality on growth, a result also found in both de2008meta and neves2016meta. To test this effect, we include the variable Regional Dummy, which takes the value of 1 if the corresponding estimate includes a regional dummy and 0 otherwise. As with other moderators, we multiply Regional Dummy by $Cross-sectional$ since the sign of the effect can vary depending on the structure of the dataset, and most studies that include regional dummies are cross-sectional. Our results confirm that controlling for regional fixed effects reduces the magnitude of the effect size in cross-sectional studies, although it is only marginally statistically significant.
\paragraph{Journal characteristics}
High quality journal might be more selective in publishing only those papers that conduct rigorous econometric analysis. To control whether the quality of a journal influences the collected effect sizes we add the moderator measuring the journal Simple Impact Factor in 2022. The coefficient in Table (ref) is statistically non-significant, suggesting no systematic relationship between journal impact and reported inequality--growth effects in our sample, in line with neves2016meta.
This paper provides a meta-analysis of the empirical literature examining the relationship between income inequality and economic growth. Building on the seminal contributions of de2008meta and neves2016meta, our analysis extends the existing evidence base along several important dimensions.
First, we update the literature by incorporating studies published over the last decade, thereby capturing recent advances in data availability and econometric practice. Second, unlike previous contributions, we collect all reported estimates from each study rather than relying on author-selected preferred specifications, resulting in a considerably larger dataset of 531 estimates from 33 studies. This approach increases statistical power and reduces the influence of individual studies. Third, in contrast to the previous meta-analysis, we restrict attention to estimates based on the Gini coefficient, ensuring conceptual and quantitative comparability across studies and avoiding the conflation of heterogeneous inequality measures that capture fundamentally different distributional concepts. Fourth, we employ state-of-the-art methods to address publication bias, relying on the RoBMA--PSMA framework, and combine this with a multilevel meta-regression that explicitly accounts for dependence among multiple estimates reported within the same study.
Our results indicate that the average effect of income inequality on economic growth is negative and statistically significant, although economically small. At the same time, we find strong evidence of substantial heterogeneity across studies, consistent with the presence of multiple true effect sizes rather than a single, context-independent inequality--growth relationship. The publication-bias analysis reveals strong evidence of selection based on statistical significance, leading to an over-representation of statistically significant results in the published literature, but no evidence of systematic directional bias.
To explain the observed heterogeneity, we estimate multilevel meta-regressions incorporating a broad set of moderators capturing dataset characteristics, econometric design choices, and journal characteristics. These results show that both real-world features and methodological decisions play a key role in shaping reported estimates. Some of them confirm the conclusion of previous meta-analyses. In particular, the adverse impact of inequality is weaker -- or even reversed -- in high-income economies compared to developing countries. Consistent with neves2016meta, we found that methodological choices also matter: cross-sectional studies tend to report more negative estimates, an effect that is attenuated by the inclusion of regional controls. Beyond these similarities, our analysis shows that instrumental-variable, fixed-effects, and difference-GMM estimators are associated with more positive estimates in panel settings. By contrast, growth-horizon length, income-data quality, and journal characteristics do not appear to influence reported results.
A central contribution of this study relies in the finding that the type of inequality measure matters: net income inequality has a more negative effect on growth than market-income inequality. This result suggests that channels operating through disposable income -- such as credit constraints and human-capital accumulation -- may be more relevant than channels operating through market-income inequality alone, such as redistributive fiscal pressures.
\addcontentsline{toc}{chapter}{References}