{
  "id": "gini-coefficient",
  "title": "Gini Coefficient",
  "lang": "en",
  "source_lang": "en",
  "summary": "The Gini coefficient measures inequality in a distribution through relative differences between its members, most commonly in income or wealth.",
  "category": "economics",
  "subcategory": "macroeconomics",
  "type": "concept",
  "aliases": [
    "Gini index",
    "Gini ratio"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/gini-coefficient/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/gini-coefficient/",
      "json": "https://aiwiki.page/zh/gini-coefficient.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "The Gini coefficient is a statistical measure of inequality, most commonly used to describe how income, consumption, or wealth is distributed across a population. For nonnegative quantities with a positive mean, it ranges from 0, indicating complete equality, toward 1, indicating extreme concentration. It is also reported on a scale from 0 to 100. Developed by the Italian statistician Corrado Gini in 1912, it became one of the most widely used indicators in the study of income inequality.\n\nThe coefficient condenses an entire distribution into a single number. It can be understood geometrically through the Lorenz curve, or algebraically as the average absolute difference between two randomly selected observations, divided by twice the population mean. Its relative simplicity makes it useful for comparisons across places, periods, and policy settings, but these comparisons depend heavily on the underlying data and definitions.\n\nAn income Gini does not measure poverty, average living standards, or the causes of inequality. Nor does it identify which population groups receive particular shares of income. Its interpretation therefore requires attention to the measured resource, the population unit, taxes and transfers, sampling uncertainty, and other features of the distribution."
    },
    {
      "heading": "Definition and mathematical foundations",
      "text": "Consider a population whose members have nonnegative incomes, with mean income \\(\\mu>0\\). If \\(X\\) and \\(X'\\) are independent draws from the same income distribution, the Gini coefficient is\n\n\\[G=\\frac{\\mathbb{E}|X-X'|}{2\\mu}.\\]\n\nThe numerator is the mean absolute difference between all possible pairs of incomes. The denominator makes this difference relative to the average income. The expected value notation expresses an average over the underlying probability distribution, rather than necessarily over a particular observed sample. Finite mean income is sufficient for the numerator to be finite when incomes are nonnegative.\n\nFor a finite population of \\(n\\) equally weighted observations \\(x_1,\\ldots,x_n\\), the corresponding formula is\n\n\\[G=\\frac{\\sum_{i=1}^{n}\\sum_{j=1}^{n}|x_i-x_j|} {2n^2\\bar{x}},\\]\n\nwhere \\(\\bar{x}\\) is the arithmetic mean. Each pair is counted in both directions, while self-comparisons contribute zero. Consequently, this formula is not simply the average difference between distinct pairs divided by twice the mean: the inclusion of self-pairs affects its finite-population normalization.\n\nIf all members have the same positive income, every pairwise difference is zero and \\(G=0\\). If one person receives all income and the other \\(n-1\\) people receive none, the coefficient is \\((n-1)/n\\). Thus, for a finite, equally weighted population, the maximum under this convention is less than 1. Some applications multiply the coefficient by \\(n/(n-1)\\) so that this finite-population maximum becomes 1. That adjustment changes the normalization; it is not a universal correction for all statistical estimation problems.\n\nThe standard interpretation assumes nonnegative observations. If every income is zero, the coefficient is undefined because its denominator is zero. Negative incomes, which can arise from business losses, and negative net worth, which occurs when debts exceed assets, complicate the interpretation. A positive-mean distribution containing negative values can produce a coefficient above 1. A zero or negative mean makes the usual relative-inequality interpretation still more problematic. These cases matter especially in studies of wealth inequality, where debt is economically meaningful and cannot be treated as a mere recording error.\n\nThe coefficient is dimensionless. A value of 0.35 and a value of 35 on the 0–100 scale represent the same measured inequality. A change from 35 to 36 is an increase of one index point, not an increase of one percent: the proportional increase in the index is approximately 2.86 percent."
    },
    {
      "heading": "The Lorenz curve and practical calculation",
      "text": "The Lorenz curve orders people from the lowest to the highest income and plots their cumulative population share against their cumulative income share. If \\(p\\) is the population share and \\(L(p)\\) its corresponding income share, complete equality is represented by \\(L(p)=p\\). For nonnegative incomes, the Lorenz curve lies on or below this equality line. The Gini coefficient equals twice the area between the two:\n\n\\[G=1-2\\int_0^1 L(p)\\,dp.\\]\n\nThe integral is the area beneath the Lorenz curve. Since the triangle beneath the equality line has area \\(1/2\\), the coefficient is also the ratio of the area between the two curves to that triangle. This geometric description explains why greater separation from the equality line generally corresponds to greater inequality.\n\nFor individual observations sorted in ascending order, \\(x_{(1)}\\leq\\cdots\\leq x_{(n)}\\), an efficient equivalent formula is\n\n\\[G=\\frac{2\\sum_{i=1}^{n}i\\,x_{(i)}} {n\\sum_{i=1}^{n}x_{(i)}}-\\frac{n+1}{n}.\\]\n\nUnlike the direct pairwise formula, which involves \\(n^2\\) comparisons, this expression requires sorting followed by a weighted sum. It is therefore much more practical for large datasets. Both expressions yield the same finite-population coefficient when their assumptions and normalization agree.\n\nFor example, consider incomes of 10, 20, 30, and 40. Total income is 100, the mean is 25, and the cumulative income shares are 0.10, 0.30, 0.60, and 1.00 at successive population quarters. The sum \\(\\sum i x_{(i)}\\) is 300, giving\n\n\\[G=\\frac{2(300)}{4(100)}-\\frac{5}{4}=0.25.\\]\n\nA second population with incomes of 25, 25, 25, and 25 has the same total income but a Gini of zero.\n\nHousehold surveys usually require unequal weights. If observation \\(i\\) represents a population weight \\(w_i\\), with \\(W=\\sum_i w_i\\) and weighted mean \\(\\mu_w=\\sum_iw_ix_i/W\\), the weighted expression is\n\n\\[G=\\frac{\\sum_i\\sum_jw_iw_j|x_i-x_j|} {2W^2\\mu_w}.\\]\n\nThe weights must correspond to the intended population: weights representing households are not interchangeable with weights representing people.\n\nWhen only grouped income shares are available, analysts often approximate the Lorenz curve by joining reported points with straight lines. If \\(P_k\\) and \\(L_k\\) are cumulative population and income shares, the trapezoidal approximation is\n\n\\[G\\approx1-\\sum_{k=1}^{m}(L_k+L_{k-1})(P_k-P_{k-1}).\\]\n\nThis calculation ordinarily misses inequality within each group. With exact grouped shares and nonnegative incomes, linear interpolation therefore gives a lower bound on the coefficient associated with the underlying individual distribution."
    },
    {
      "heading": "Historical development and adoption",
      "text": "The Gini coefficient emerged from early twentieth-century work on statistical variation, income distribution, and social measurement. Max O. Lorenz introduced the curve that bears his name in 1905, in an article on measuring the concentration of wealth. Its graphical representation made it possible to compare distributions without reducing their differences immediately to a single average or dispersion statistic.\n\nGini presented his measure in the 1912 work Variabilità e mutabilità. His research treated inequality as part of a broader statistical problem: how to quantify differences among observations. The mean absolute difference was central to this approach. Unlike the variance, which squares deviations from the mean, the Gini construction uses absolute differences between observations and normalizes them by the mean. Its close connection to the Lorenz curve helped make an abstract measure accessible to applied researchers.\n\nThe development of inequality measurement was also linked to debates in welfare economics. Arthur Cecil Pigou and Hugh Dalton helped establish the importance of transfers between richer and poorer people in judging distributions. Dalton’s 1920 article on measuring economic welfare brought this transfer-based reasoning into systematic discussion. The resulting Pigou–Dalton transfer principle became a standard criterion for assessing inequality measures, including the Gini.\n\nLater work clarified that choosing an inequality index is not entirely a matter of statistical convenience. Anthony Atkinson’s 1970 article “On the Measurement of Inequality” examined how comparisons depend on judgments about social welfare and inequality aversion. This work helped distinguish measures whose ethical assumptions are explicit from those, such as the conventional Gini, whose weighting of distributional differences is built into their mathematical form.\n\nThe expansion of household surveys and administrative statistics during the twentieth century made routine estimation increasingly feasible. Statistical agencies and international organizations, including the World Bank and the Organisation for Economic Co-operation and Development, incorporated Gini measures into comparative datasets. Their use spread across studies of development, redistribution, living standards, and social conditions.\n\nThis adoption did not produce a single worldwide measurement convention. Some statistical systems emphasize consumption expenditure, while others emphasize disposable household income. Historical series can also change when surveys are redesigned or income definitions are broadened. The history of the coefficient is therefore both a history of a mathematical measure and a history of the institutions and data systems that make its calculation possible."
    },
    {
      "heading": "Inequality properties and their interpretation",
      "text": "Several properties explain the coefficient’s usefulness. Anonymity means that exchanging the incomes of two people leaves the coefficient unchanged. The measure depends on the distribution of amounts, not on the identities of recipients. It therefore cannot distinguish distributions that are numerically identical but allocate resources differently across age groups, regions, or other socially significant categories.\n\nScale invariance means that multiplying every income by the same positive constant leaves the coefficient unchanged:\n\n\\[G(cx_1,\\ldots,cx_n)=G(x_1,\\ldots,x_n),\\qquad c>0.\\]\n\nChanging the currency unit, or doubling every income, does not change relative inequality. Uniform adjustment for inflation likewise leaves it unchanged, although differences in the prices faced by different households can affect the distribution of purchasing power. Scale invariance also explains why the Gini alone cannot reveal whether living standards are high or low.\n\nThe coefficient is not invariant to adding the same absolute amount to every income. A common positive addition leaves all pairwise differences unchanged but raises the mean, reducing the Gini. For the earlier distribution of 10, 20, 30, and 40, adding 10 to every income raises the mean from 25 to 35 and reduces the coefficient from 0.25 to approximately 0.179. By contrast, multiplying all four incomes by two leaves it at 0.25. Equal absolute additions and equal proportional increases are therefore different distributional changes.\n\nPopulation replication invariance means that copying every member of a distribution the same number of times leaves the conventional coefficient unchanged. A population containing one person at each of four income levels has the same Gini as a population containing a thousand people at each of those levels. This property permits comparisons across differently sized populations. It should be distinguished from the finite-population maximum adjustment, which depends explicitly on population size.\n\nThe coefficient also satisfies the Pigou–Dalton transfer principle: a transfer from a richer person to a poorer person, holding total income constant and without reversing their ranks, reduces measured inequality. For equally weighted observations, a transfer \\(\\delta\\) from rank \\(j\\) to rank \\(i<j\\), with all ranks otherwise unchanged, reduces the coefficient by\n\n\\[\\frac{2\\delta(j-i)}{n^2\\bar{x}}.\\]\n\nThus its response depends on rank distance as well as the amount transferred. Descriptions of the Gini as relatively sensitive to the middle of the distribution are useful comparative shorthand, but not a complete account of its weighting. Its exact response depends on the population shares affected, their ranks, and the accompanying changes in the mean."
    },
    {
      "heading": "Income concepts, population units, and comparability",
      "text": "A Gini coefficient is meaningful only after specifying what is being distributed. Earnings, market income, disposable income, consumption expenditure, and wealth describe different resources. Earnings concern payment for work; market income also includes other income sources; disposable income generally accounts for direct taxes and cash transfers. Wealth is a stock of assets less liabilities at a particular date, rather than a flow received over a period. Coefficients for these concepts are not interchangeable.\n\nThe population unit is equally consequential. A distribution across individual workers excludes people without earnings unless they are explicitly included with zero values. A distribution across households gives each household one place in the population, regardless of its size. A person-based distribution of household income typically assigns each person the adjusted income of the household in which that person lives. These approaches answer different questions and can produce different coefficients from the same records.\n\nHousehold resources are commonly adjusted for needs through an equivalence scale. Dividing income by household size gives per capita income, while other scales recognize that household members can share housing, heating, and durable goods. The modified OECD scale assigns a weight of 1 to the first adult, 0.5 to each additional person aged 14 or over, and 0.3 to each younger child. Equivalized income is household income divided by the resulting total. Person-weighted inequality then counts each household member separately at that equivalized income level.\n\nIncome and consumption can differ substantially because households save, borrow, and draw down assets. Consumption may be less volatile than annual income, but measuring it requires decisions about home-produced goods, durable purchases, and housing services. Income measurement faces its own decisions about self-employment profits, irregular payments, capital gains, and imputed rent for owner-occupied housing. Publicly provided education and health services are generally outside cash disposable income, although broader resource measures can assign them monetary values.\n\nTime and place create additional comparability issues. Monthly income can be more variable than annual income, while lifetime resources differ from either. A common currency conversion leaves a country’s Gini unchanged, but region-specific price adjustments can change the ranking of real incomes. International comparisons using purchasing power parity are essential when combining people into a worldwide distribution, rather than simply comparing national coefficients.\n\nConsequently, two published Ginis should not be treated as directly comparable merely because they refer to the same year. Their income concept, treatment of taxes, population coverage, equivalence scale, and weighting must also align. A global coefficient cannot be obtained by simply averaging national Ginis, even when those national figures are individually comparable."
    },
    {
      "heading": "Data quality, estimation, and statistical uncertainty",
      "text": "Most income Ginis are estimated from sample surveys rather than complete population records. The resulting number is an estimate subject to sampling error. Because the coefficient depends on the full distribution, its uncertainty reflects more than uncertainty in the mean. Clustering, stratification, unequal selection probabilities, and nonresponse adjustments all affect the precision of an estimate.\n\nStatistical agencies can estimate uncertainty through linearization, replicate weights, or a bootstrap. A suitable bootstrap reproduces relevant features of the survey design rather than merely drawing individual records at random. For example, a clustered survey may require resampling primary sampling units within strata. Ignoring design features can produce misleadingly narrow confidence intervals. A small difference between two estimated coefficients therefore does not automatically establish that inequality changed.\n\nUndercoverage and reporting errors may be more important than sampling variation. Wealthy households can be difficult to reach, may decline participation, or may underreport certain income sources. Surveys may also top-code high incomes to protect confidentiality. These problems can reduce measured inequality, although the magnitude and direction of total error depend on which people and resources are missing. Underreporting by low-income households or omitted benefit income can have different effects.\n\nAdministrative records can improve coverage of wages, taxable income, and transfers, but they introduce other limitations. A tax unit may not correspond to a household, and nontaxable income may be absent. People outside filing requirements can be missing, while changes in tax rules alter recorded income even without equivalent changes in underlying resources. Combining survey and administrative data requires consistent definitions, secure record linkage, and attention to data privacy.\n\nResearchers sometimes adjust the upper tail using tax records, rich lists, or statistical models. Pareto-type models, associated with Vilfredo Pareto, are often used to represent highly concentrated top incomes or wealth. The resulting coefficient depends on the threshold above which the model is applied and on assumptions about people missing from the original dataset. These adjustments are estimates of an incompletely observed distribution, not directly observed replacements for every missing value.\n\nMissing income components, survey breaks, and discrepancies with national accounts create further uncertainty. Household survey totals need not match national-account aggregates because coverage and definitions differ as well as because of reporting error. Scaling all observations by a common factor cannot repair distribution-specific omissions: it leaves the Gini unchanged. Methods that allocate missing income unevenly can change the coefficient, but their allocation assumptions become part of the estimate."
    },
    {
      "heading": "Redistribution, income sources, and subgroup decomposition",
      "text": "Gini coefficients are widely used to describe the distributional role of taxation and transfers. A common comparison contrasts market-income inequality with disposable-income inequality. If the former coefficient is \\(G_M\\) and the latter \\(G_D\\), their difference, \\(G_M-G_D\\), records a reduction in measured inequality under the selected definitions. A relative measure divides this difference by \\(G_M\\).\n\nThis comparison is descriptive rather than a complete causal estimate of government intervention. Taxes and transfers may influence employment, saving, household formation, and reported income, so the observed market-income distribution is not necessarily the distribution that would exist without them. Treatment of pensions is particularly consequential: classifying public pensions as transfers rather than deferred earnings can substantially change the measured market-income inequality of a population containing many retired people.\n\nIncome-source decomposition asks how wages, self-employment income, pensions, investment income, and other components contribute to overall inequality. If total income is the sum of sources, a standard decomposition is\n\n\\[G=\\sum_k s_k C_k,\\]\n\nwhere \\(s_k\\) is source \\(k\\)'s share of total income and \\(C_k\\) is its concentration coefficient when recipients are ranked by total income. Under regularity conditions, this is also written\n\n\\[G=\\sum_k s_kG_kR_k,\\]\n\nwhere \\(G_k\\) is the source’s own Gini and \\(R_k\\) is its Gini correlation with total income. The source’s concentration depends not only on how unequally it is distributed, but on whether its recipients are rich or poor in the overall distribution.\n\nThis distinction is important for transfers. A narrowly targeted payment may be very unequal when viewed in isolation because most people receive nothing. Nevertheless, if its recipients are concentrated at the bottom of the total-income distribution, it can reduce overall inequality. Conversely, an income source with only moderate inequality among its recipients can increase total inequality if it accrues predominantly to already affluent households.\n\nDecomposition by population subgroup is less straightforward. Overall inequality is not generally the population-weighted average of regional, occupational, or demographic Ginis. Differences between group means and differences among individuals across groups also matter. When the income ranges of groups overlap, cross-group pairwise differences depend on that overlap, not just on group averages.\n\nSome Gini decompositions separate within-group, between-group, and overlap-related components. Their interpretation depends on the decomposition method and the chosen counterfactual distributions. For analyses requiring a clean additive partition into within-group and between-group inequality, measures such as the Theil index are often more convenient."
    },
    {
      "heading": "Limits of a single-number measure and alternative indicators",
      "text": "Different distributions can have the same Gini coefficient. Incomes of \\(0,2,2\\) and \\(1,1,4\\), for example, both produce a coefficient of \\(1/3\\), although their means and distributional shapes differ. Even equal means would not make the coefficient uniquely identify a distribution. Equality of Ginis therefore does not imply equality of living conditions, poverty, top income shares, or the position of a particular population group.\n\nThe coefficient also cannot reveal poverty without additional information. Two populations can have identical relative income distributions while one has incomes ten times as high as the other. Their Ginis are equal, but the proportions below a fixed real poverty line may differ sharply. Likewise, a fall in inequality can accompany declining incomes if higher incomes fall faster. Interpreting such a change requires evidence about levels as well as shares.\n\nLorenz curves provide a stronger comparison in some cases. If one distribution’s curve is everywhere at least as high as another’s, it exhibits Lorenz dominance and has no greater inequality under a broad class of Lorenz-consistent measures. Its Gini cannot be higher. When curves cross, however, one distribution may give a larger share to the poorest people but also a larger share to the richest. Different indices can then rank the distributions differently because they weight these changes differently.\n\nThe Atkinson index makes inequality aversion explicit through a parameter and relates inequality to equally distributed equivalent income. Generalized entropy measures include the Theil index and the mean log deviation; their parameter choices emphasize different parts of the distribution and support additive subgroup decompositions. Percentile ratios and income-share measures focus directly on selected positions or groups, but omit much of the remaining distribution.\n\nTo combine distributional shape with income levels, analysts can use the generalized Lorenz curve, defined as mean income multiplied by the ordinary Lorenz curve. Generalized Lorenz dominance supports welfare comparisons under specified assumptions about increasing and concave social evaluations of income. It addresses a different question from the scale-invariant Gini: not simply how unequal resources are, but how both their level and distribution compare.\n\nThe Gini also does not measure mobility. An unchanged coefficient can conceal substantial movement of individuals between income ranks, while a changing coefficient can occur with little rank movement. Longitudinal records are required to distinguish persistent disadvantage from temporary income fluctuations. Nor does the coefficient identify discrimination, unequal opportunity, or causal mechanisms merely by recording an unequal outcome."
    },
    {
      "heading": "Applications beyond household income",
      "text": "The Gini coefficient can be applied to any suitable distribution of nonnegative quantities, not only income. In agriculture, it can describe the concentration of landholdings. In energy and environmental research, it can measure disparities in consumption or emissions. In science studies, it can describe the concentration of citations across publications. Each application requires defining both the measured quantity and the units among which it is distributed.\n\nLand inequality illustrates the importance of the denominator population. A coefficient calculated among landowners excludes landless households and measures the unequal sizes of existing holdings. A coefficient calculated across all households, assigning zero land to those without it, addresses access to land as well as concentration among owners. The two statistics may differ markedly without either being mathematically incorrect. Questions of land quality, land tenure, and shared rights require information beyond area alone.\n\nIn environmental applications, per capita emissions and total emissions across countries describe different distributions. Treating countries as equally weighted observations gives a small country the same statistical weight as a populous one. Weighting national per capita emissions by population instead describes inequality among people under the assumption that everyone within each country has the national average. Such a calculation captures between-country differences but omits within-country inequality.\n\nWealth applications require especially careful treatment of debt, valuation, and ownership. A household may have negative net worth while possessing positive assets, and pension rights may be included in one dataset but excluded in another. Because financial wealth is often heavily concentrated and difficult to observe, upper-tail assumptions can have large effects. A wealth Gini should therefore be distinguished from a coefficient of gross asset holdings or of financial assets alone.\n\nThe Gini is sometimes mentioned alongside measures of industrial concentration, but it is not interchangeable with the Herfindahl–Hirschman index. If every firm has an equal market share, the Gini of firm sizes is zero whether the market contains two firms or two hundred. The Herfindahl–Hirschman index distinguishes those cases because it sums squared market shares. Thus equal firm sizes do not by themselves imply a competitive market.\n\nA further distinction arises in machine learning. Decision-tree “Gini impurity,” commonly written \\(1-\\sum_k p_k^2\\), measures the heterogeneity of class labels at a node and is not the income-inequality coefficient. In binary classification, another quantity called a Gini coefficient is often defined as \\(2\\,\\mathrm{AUC}-1\\), where AUC is the area under a receiver operating characteristic curve. It evaluates ranking performance under a different construction; its numerical scale and interpretation should not be transferred to household-income inequality."
    }
  ],
  "links": [
    "income-inequality",
    "lorenz-curve",
    "expected-value",
    "probability-distribution",
    "net-worth",
    "wealth-inequality",
    "integral",
    "variance",
    "welfare-economics",
    "world-bank",
    "inflation",
    "disposable-income",
    "household",
    "education",
    "purchasing-power-parity",
    "bootstrap-statistics",
    "data-privacy",
    "vilfredo-pareto",
    "national-accounts",
    "taxation",
    "poverty",
    "poverty-line",
    "causation",
    "land-tenure",
    "herfindahl-hirschman-index",
    "machine-learning"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}