Gini Coefficient

Gini coefficient (also known as the Gini index when expressed as a percentage from 0 to 100) is the standard quantitative metric of statistical dispersion and economic inequality used across macroeconomics, welfare economics, demography, and ecology to measure how far an empirical distribution of income, consumption, or wealth departs from an egalitarian distribution. Formulated by the Italian statistician and demographer Corrado Gini in his 1912 monograph Variabilità e mutabilità (building on the graphical Lorenz curve introduced by American economist Max O. Lorenz in 1905), the Gini coefficient maps any non-negative distribution onto the unit interval G ∈ [0, 1]. A value of G = 0 signifies complete equality (every household or individual receives the same income or wealth), whereas a value of G = 1 (or (N − 1) / N in a finite population of size N) signifies maximal concentration (a single individual receives 100% of total national income and all others receive zero).

Within business and economics theory, the Gini coefficient complements aggregate macroeconomic output metrics such as gross domestic product (GDP) per capita and purchasing power parity by summarizing the distributional geometry of living standards, evaluating the redistributive impact of progressive taxation and fiscal policy transfers, and anchoring Amartya Sen's (1976) inequality-adjusted social welfare function W_Sen = μ(1 − G), where μ is mean income. Beyond economics, the Gini coefficient is applied in machine learning (CART decision-tree splitting purity), biodiversity ecology, credit-scoring Bayesian inference model validation (where G = 2·AUC − 1 relative to the Receiver Operating Characteristic curve), and genomic expression dispersion.

Max Lorenz's Cumulative Share Curve and the Area Ratio A / (A + B)

Let a continuous population have non-negative income y ≥ 0 governed by cumulative distribution function F(y) = p ∈ [0, 1], probability density f(y) = F′(y), quantile (inverse CDF) function y(p) = F⁻¹(p), and finite positive population mean μ = ∫₀^∞ y f(y) dy = ∫₀¹ y(p) dp > 0. Ordering all individuals from poorest (p = 0) to richest (p = 1), the Lorenz curve L(p) plots the cumulative share of total income earned by the bottom p fraction of the population:

By construction, L(0) = 0, L(1) = 1, the first derivative is L′(p) = y(p) / μ ≥ 0 (so the Lorenz curve has slope 1 at the population mean income y(p) = μ), and the second derivative is L′′(p) = y′(p) / μ ≥ 0, ensuring that every Lorenz curve is convex and lies on or below the 45° Line of Equality L_eq(p) = p.

Geometrically, let A denote the lens-shaped area between the 45° Line of Equality (p) and the Lorenz curve L(p), and let B denote the area underneath the Lorenz curve L(p) over p ∈ [0, 1]:

The Gini coefficient G is defined as the ratio of area A to the total right-triangle area A + B = 0.5 under the Line of Equality:

Relative Mean Absolute Difference, Covariance Formula, and Parametric Distributions

Corrado Gini (1912) proved that the geometric area ratio 2A equals half of the relative mean absolute difference—the average absolute income gap |y_i − y_j| between all N² ordered pairs of individuals drawn independently from the population, normalized by twice the population mean 2μ:

When a finite sample of N observations is sorted in non-decreasing order y₍₁₎ ≤ y₍₂₎ ≤ ... ≤ y_(N), the double summation simplifies to an O(N log N) rank-weighted linear sum (and can also be expressed via the statistical covariance between income y and its cumulative rank F(y)):

For standard parametric probability distributions used to model income and wealth in econometrics, the Gini coefficient admits closed-form analytical expressions:

• Uniform Distribution on [a, b] (0 ≤ a < b): G = ( b − a ) / [ 3 ( a + b ) ], which equals 1 / 3 ≈ 0.333 when a = 0.
• Exponential Distribution f(y) = λ exp(−λy): G = 1 / 2 = 0.500 for all scale parameters λ > 0.
• Pareto Power-Law Distribution F(y) = 1 − (y_min / y)^α (for tail exponent α > 1):

As the Pareto tail becomes heavier (α → 1⁺), G_Pareto → 1; for Vilfredo Pareto's classical 80–20 rule (α = ln 5 / ln 4 ≈ 1.161), G ≈ 0.756.

• Lognormal Distribution ln(y) ~ Normal(m, σ²):

where Φ(·) is the standard normal cumulative distribution function and erf(·) is the Gauss error function, depending only on the log-standard deviation σ and remaining scale-invariant to the log-mean m.

The Four Scale-and-Transfer Axioms and Sen's Social Welfare Function

The Gini coefficient satisfies four foundational axioms of inequality measurement (Foster, 1985; Sen, 1973):

1. Scale Invariance (Zero Homogeneity): Multiplying every person's income by a positive constant c > 0 (for example, converting currencies from dollars to euros or experiencing uniform inflation) leaves G unchanged: G(c·y) = G(y). (Note, however, that adding an equal absolute dollar transfer +d > 0 to everyone reduces G because d increases the denominator 2μ while leaving pairwise absolute gaps |y_i − y_j| unchanged.)
2. Population Principle (Replication Invariance): Cloning a population k times by replicating every income value k times leaves the Lorenz curve and G identical.
3. Anonymity (Symmetry): G depends only on the multiset of incomes and is invariant to permuting individual labels.
4. The Pigou-Dalton Transfer Principle: Any rank-preserving progressive transfer δ > 0 from a richer individual j to a poorer individual i (y_i < y_i + δ ≤ y_j − δ < y_j) shifts the Lorenz curve upward and reduces the Gini coefficient:

Notice that the reduction ΔG is proportional to the rank distance (j − i) between the two individuals rather than their absolute income difference (y_j − y_i). Because population density f(y) is highest near the median/mode of a unimodal income distribution, a given dollar transfer spans more rank positions (j − i) in the middle class than in the sparse tails, making the Gini coefficient most sensitive to middle-income transfers.

In 1976, Nobel laureate Amartya Sen axiomatized a rank-order weighted social welfare function that resolves the trade-off between economic growth (higher mean income μ, as in the Solow-Swan model) and distributional equity (lower G):

Under Sen's welfare criterion, a policy or technological transformation that raises per-capita gross domestic product μ by 10% while causing (1 − G) to fall by more than 10% reduces aggregate social welfare, whereas a policy that raises μ·(1 − G) is welfare-improving whenever Lorenz curves do not cross (Atkinson's theorem, 1970).

Market vs. Disposable Income Gini, Wealth Gini, and Intersecting Lorenz Curves

In empirical comparative macroeconomics (such as the Luxembourg Income Study and OECD databases), economists distinguish three fundamental Gini benchmarks within every country:

• Pre-Tax / Pre-Transfer Market Income Gini (G_mkt): Measures inequality generated by labor wages, self-employment, and capital dividends before government intervention. Across advanced OECD economies, G_mkt typically ranges from 0.44 to 0.53 (driven in part by retired households having near-zero market wage income).
• Post-Tax / Post-Transfer Disposable Income Gini (G_disp): Measures household disposable cash income after deducting direct income/payroll taxes and adding public transfers (pensions, unemployment insurance, child credits). The fiscal Reynolds-Smolensky redistribution index ΔG_fiscal = G_mkt − G_disp averages 0.15 to 0.22 Gini points in Northern and Western European welfare states (reducing G_disp to ~0.25–0.30 in Nordic nations and Germany/France), compared to ΔG_fiscal ≈ 0.09–0.11 in the United States (G_disp ≈ 0.39–0.41) and smaller fiscal reductions in Latin American and Sub-Saharan African economies (G_disp ≈ 0.45–0.63).
• Net Wealth Gini (G_wealth): Because household net worth (financial assets plus real estate minus mortgage and consumer debt) represents a cumulative stock driven by compounding returns and lifecycle savings, wealth Gini coefficients are higher than income Gini coefficients—typically ranging from 0.65 to 0.85 across developed and developing nations (and occasionally exceeding 1.0 if a large fraction of households hold negative net worth, which pulls the Lorenz curve below zero for bottom deciles).

Despite its universality, the Gini coefficient has three well-known analytical limitations:

1. Intersecting Lorenz Curves: Two distinct societies can exhibit the same Gini coefficient (for example, G = 0.40) if Country X has a poor bottom 50% sharing 10% of income while the top 50% share the remaining 90% equally, whereas Country Y has the bottom 90% sharing 50% of income equally while the top 10% capture the remaining 50%. Their Lorenz curves cross, yielding identical area A despite opposite structural distributions.
2. Non-Additive Subgroup Decomposition: Unlike the generalized entropy Theil index (T = T_within + T_between), the Gini coefficient of a combined population cannot be written as the sum of within-group Gini and between-group Gini whenever the income ranges of the subgroups overlap; Dagum (1997) showed that a third positive transvariation (overlap) term R is required (G = G_W + G_B + R).
3. Lifecycle Age Confounding: Because young students and retirees naturally earn lower annual incomes than prime-age workers (ages 45–55) even if every cohort enjoys identical lifetime net present value earnings, cross-sectional annual Gini coefficients overstate lifetime permanent-income inequality unless adjusted via Paglin-Gini age-cohort corrections.

Can the Gini coefficient ever exceed 1.0?

When all observations are non-negative (y_i ≥ 0), the Gini coefficient is bounded on [0, 1 − 1/N] ⊂ [0, 1]. However, when measuring net wealth or business net income where some households have negative net worth (y_i < 0 due to student loans, business losses, or underwater mortgages) while mean wealth μ remains positive, the Lorenz curve L(p) dips below the horizontal axis (L(p) < 0) for bottom percentiles, causing the area A between the 45° line and L(p) to exceed 0.5 and pushing the raw Gini coefficient above 1.0 (unless renormalized via Raffinetti, Siletti, and Vernizzi's 2015 [0, 1] normalization).

What is the Kuznets curve hypothesis regarding the Gini coefficient and economic development?

In his 1955 American Economic Association presidential address, Simon Kuznets hypothesized an inverted-U relationship between log per-capita gross domestic product and the Gini coefficient: during early industrialization, labor migration from low-inequality rural agriculture to high-productivity urban manufacturing initially widens income inequality (G rises), whereas in mature industrial stages, universal education, skill diffusion, and welfare-state democratization compress wage dispersion (G falls). Subsequent long-run tax-record studies by Thomas Piketty and Emmanuel Saez (2003, 2014) showed that top-income shares and Gini coefficients in several post-1980 advanced economies rose again due to skill-biased technological change, globalization, and return-on-capital dynamics (r > g).

How is the Gini coefficient related to the ROC curve in machine learning and credit risk?

In binary classification and bank credit-default modeling (such as adverse selection screening of borrowers), plotting the cumulative share of actual defaults captured against the cumulative share of all borrowers ordered by predicted risk score generates the Cumulative Accuracy Profile (CAP) curve—the mirror image of a Lorenz curve. The classifier's Gini coefficient (also called Somers' D) satisfies the linear identity Gini = 2 · AUC − 1, where AUC ∈ [0.5, 1.0] is the Area Under the Receiver Operating Characteristic (ROC) curve.

1. Lorenz, M. O. (1905). "Methods of measuring the concentration of wealth." Publications of the American Statistical Association, 9(70), 209–219. https://doi.org/10.2307/2276207
2. Gini, C. (1912). Variabilità e mutabilità: Contributo allo studio delle distribuzioni e delle relazioni statistiche. Tipografia di Paolo Cuppini, Bologna. https://archive.org/details/variabilitemutab00giniuoft
3. Atkinson, A. B. (1970). "On the measurement of inequality." Journal of Economic Theory, 2(3), 244–263. https://doi.org/10.1016/0022-0531(70)90039-690039-6)
4. Sen, A. (1976). "Real national income." Review of Economic Studies, 43(1), 19–39. https://doi.org/10.2307/2296597
5. Dagum, C. (1997). "A new approach to the decomposition of the Gini income inequality ratio." Empirical Economics, 22(4), 515–531. https://doi.org/10.1007/BF01205777