Two Tunable Gini-Type Measures with U-Statistic Estimation: Theory, Simulation, and an Empirical Application to GDP per Capita in the Americas

Technical Summary: Two Tunable Gini-Type Measures with U-Statistic Estimation

Problem Statement
The classical Gini coefficient, while a standard benchmark for measuring income and wealth inequality due to its interpretability and ease of estimation, treats all pairwise differences between observations equally. This uniform weighting can obscure nuanced patterns of dispersion, particularly regarding how extreme disparities versus moderate gaps contribute to overall inequality. The authors address the need for inequality measures that allow analysts to tune sensitivity to tail-inequality without abandoning the theoretical framework of the Gini coefficient.

Methodology
The paper introduces two new families of inequality measures, denoted $G_p$ and $H_q$, which generalize the classical Gini coefficient ($G$) through tunable parameters $p > 1$ and $q > 0$, respectively.

1. Definitions:

   • $G_p$: Defined via a logarithmic kernel. It replaces the raw absolute difference $|X_1 - X_2|$ with a scaled kernel $T(p)$ involving logarithmic terms. As $p \to \infty$, $G_p$ converges to the classical Gini.
   • $H_q$: Defined via a power-mean difference kernel $K(q)$, utilizing the difference between the $q$-th power mean and the $-q$-th power mean of two observations. As $q \to \infty$, $H_q$ also converges to the classical Gini.

2. Theoretical Properties:

   • Both measures are shown to be increasing functions of their respective parameters ($p$ and $q$), implying that higher values increase the measure's sensitivity to extreme disparities (inequality aversion).
   • Both measures converge to 0 as parameters approach their lower bounds ($p \to 1^+$ and $q \to 0^+$) and to the classical Gini as parameters approach infinity.
   • Scale Invariance: A key distinction is noted: $H_q$ is ratio-scale invariant (invariant under positive scalar multiplication of income), similar to the classical Gini. In contrast, $G_p$ is not ratio-scale invariant; scaling the income variable is mathematically equivalent to changing the tuning parameter $p$.
   • Both measures are bounded between 0 and 1.

3. Estimation and Asymptotics:

   • The authors derive closed-form U-statistic plug-in estimators ($\hat{G}_p$ and $\hat{H}_q$) based on sample data.
   • Using standard results for U-statistics (Hoeffding, 1948), they establish strong consistency and asymptotic normality for both estimators under mild moment conditions (specifically, finite first and second moments of the underlying distribution).
   • The asymptotic variance is derived using the multivariate delta method, providing a basis for constructing confidence intervals and hypothesis tests.

Empirical and Simulation Results

• Monte Carlo Simulation: A simulation study was conducted using data generated from Gamma distributions with varying sample sizes ($n \in \{30, 50, 100, 200, 500\}$) and parameter values ($p, q \in \{1.1, 2, 5, 10, 50\}$).
  • Results indicate that the Mean Absolute Relative Error (MARE) and Root Mean Squared Error (RMSE) decrease as sample size increases.
  • The asymptotic normal approximation is found to be adequate for sample sizes $n \geq 100$.
  • The study confirms that larger parameter values yield estimates closer to the classical Gini, inheriting its favorable estimation properties in those specific configurations.
• Empirical Application: The measures were applied to GDP per capita data for 34 countries in the Americas (2023).
  • The data were validated against a Gamma distribution using Kolmogorov–Smirnov and Cramér–von Mises tests.
  • The analysis of the sensitivity profiles (plotting estimates against $p$ and $q$) revealed that $G_p$ rises steeply at low $p$ values before leveling off, while $H_q$ rises more gradually. This suggests that extreme pairwise disparities are a significant driver of inequality in the dataset, and that moderate values of $p$ capture most of this effect.

Key Contributions

1. Novel Measures: The introduction of two tunable inequality indices ($G_p$ and $H_q$) that bridge the gap between standard dispersion measures and the classical Gini coefficient.
2. Statistical Theory: The derivation of closed-form U-statistic estimators and the rigorous proof of their strong consistency and asymptotic normality.
3. Practical Utility: The provision of a framework for sensitivity analysis. By examining the profile of estimates across a range of parameters, analysts can determine whether inequality is uniformly distributed or driven by extreme outliers, thereby informing the choice of an appropriate summary metric.

Significance
The paper claims that these measures are designed not to replace the Gini coefficient but to complement it. By allowing the tuning of sensitivity to tail-inequality, these indices offer a more granular view of distributional structures. The authors emphasize that the choice of parameter ($p$ or $q$) should be guided by the specific normative framework or the nature of the data (e.g., whether the measurement unit should affect the assessment). The ability to construct confidence intervals via the derived asymptotic distributions further enhances their applicability in policy analysis and empirical research.