📊 statistics

Unbiased estimation of normalized scale-invariant indices under the gamma distribution

This paper introduces a broad class of normalized scale-invariant indices (NPRIs) and derives a simple, unbiased U-statistic estimator for them under gamma distributions, demonstrating its effectiveness through theoretical proofs, simulation studies, and a real-world application to GDP per capita data.

Original authors: Roberto Vila, Helton Saulo, Felipe Quintino

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Roberto Vila, Helton Saulo, Felipe Quintino

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to measure how "uneven" a pile of sand is. You could look at the total weight of the pile, but that doesn't tell you if the sand is spread out evenly or if it's all piled up in one corner. To measure the unevenness (or inequality) itself, you need a tool that ignores the total weight and focuses only on the shape of the pile.

This paper introduces a new, super-flexible tool for measuring that shape, specifically for data that behaves like a "Gamma distribution" (a common pattern in nature where small values are frequent, but large values can happen, like income or rainfall).

Here is the breakdown of their work using simple analogies:

1. The Problem: Too Many Rulers

For years, statisticians have had many different rulers to measure inequality:

The Gini Coefficient (the most famous one).
Entropy (measuring randomness).
Variability indices (measuring how much things jump around).

The problem is that everyone was building their own ruler from scratch. If you wanted to measure a new type of inequality, you had to invent a whole new mathematical formula and a new way to calculate it. It was like having a different screwdriver for every single screw in the house.

2. The Solution: The "Universal Adapter"

The authors (Roberto Vila, Helton Saulo, and Felipe Quintino) created a Universal Adapter called NPRIs (Normalized Pairwise Ratio Indices).

Think of this adapter as a single, magical function that can snap onto any of those old rulers. They proved that almost all these famous inequality measures are actually just special cases of this one big family.

The Magic Trick: They use a mathematical rule called "homogeneity." In plain English, this means if you double the size of every person's income in a room, the "inevenness" score stays exactly the same. It only cares about the ratios (who has more than whom), not the actual dollar amounts.

3. The Secret Ingredient: The "Gamma" Cake

The paper focuses on data that follows a Gamma distribution. You can think of this distribution as a specific type of "cake" that statisticians love to bake because it's so common in the real world (like income, insurance claims, or rainfall).

The authors discovered a special property of this "Gamma cake":

If you take a sample of data, the Total Sum (the whole cake) and the Proportions (how the cake is sliced) are completely independent.
The Analogy: Imagine you have a pizza. The total size of the pizza doesn't tell you how the slices are cut. You could have a giant pizza with equal slices, or a tiny pizza with one huge slice and one tiny slice. The authors realized that for Gamma data, you can mathematically separate the "size of the pizza" from the "shape of the slices."

4. The New Tool: The "Unbiased U-Statistic"

Because they could separate the size from the shape, they built a new calculator (an estimator) that is Unbiased.

What is "Unbiased"? Imagine you are trying to guess the average height of students in a school. If your method consistently guesses 5 feet when the real average is 5 feet 2 inches, your method is "biased." If your method hits the bullseye on average every time, it is "unbiased."
How it works: Their calculator takes a sample of data, looks at all possible combinations of groups (like looking at every possible trio of people), and averages them out. They proved mathematically that this method gives the exact right answer for any of the inequality rulers they built, provided the data follows the Gamma pattern.

5. Testing the Tool

The authors didn't just write the math; they put it to the test in two ways:

The Simulation Lab: They created thousands of fake datasets on a computer that looked like Gamma distributions. They ran their new calculator against these fakes and found that it was incredibly accurate, with very little error, even when the sample sizes were small.
The Real World Test: They applied their tool to real data: the Gross Domestic Product (GDP) per capita for 34 countries in the Americas.
- First, they checked if the income data actually looked like a Gamma "cake." It did.
- Then, they used their calculator to measure the inequality.
- The Result: They found that economic inequality in the Americas is "moderate to high." They also showed that their new method gives slightly different, but very precise, numbers compared to older methods, and it works well even when you change the specific type of inequality ruler you use (like switching from Gini to Power indices).

Summary

In short, the authors built a universal mathematical machine that can measure any kind of inequality (Gini, entropy, etc.) in one go. They proved that if your data follows a Gamma pattern (which is common in economics), this machine is perfectly accurate and doesn't make systematic errors. They tested it on fake data and real-world income data, and it worked beautifully.

What they did NOT do:

They did not claim this works for every type of data in the universe (only Gamma and a specific extension to Generalized Gamma).
They did not propose new economic policies based on these numbers; they only provided a better way to measure the numbers.
They did not apply this to medical or clinical data; the only real-world example used was economic income.

Technical Summary: Unbiased Estimation of Normalized Scale-Invariant Indices under the Gamma Distribution

Problem Statement
Measures of inequality, variability, and concentration—such as the Gini coefficient, entropy-based measures, and order-statistic functionals—are fundamental in economics, finance, and actuarial science. While these indices are often studied individually, they share common structural features: they are scale-invariant and depend on relative proportions rather than absolute magnitudes. A unified framework capable of encompassing this broad collection of measures while facilitating the derivation of their statistical properties, particularly unbiased estimators under specific distributional assumptions, has been lacking. This paper addresses the need for a general class of indices and a corresponding unified estimation methodology, specifically within the context of the gamma distribution, which is widely applicable in these fields.

Methodology
The authors introduce a broad class of Normalized Pairwise Ratio Indices (NPRIs) generated by homogeneous functions. Let $X$ be a non-negative random variable with finite mean $\mu$ . An NPRI is defined as:
$I(g) = \frac{E[g(X_1, \dots, X_m)]}{mC\mu}$
where $g: (0, \infty)^m \to [0, \infty)$ is a measurable function satisfying:

Homogeneity of degree one: $g(\lambda x_1, \dots, \lambda x_m) = \lambda g(x_1, \dots, x_m)$ for all $\lambda > 0$ .
Boundedness: $g(x_1, \dots, x_m) \leq C \sum_{i=1}^m x_i$ for some finite constant $C$ .

This definition ensures that $I(g)$ is scale-invariant and normalized to the interval $[0, 1]$ . The framework encompasses the classical Gini coefficient, generalized Gini indices, entropy measures, product-type indices, and various order-statistic-based measures.

The core methodological contribution is the derivation of an unbiased estimator for any NPRI when the underlying population follows a Gamma distribution ( $X \sim \text{Gamma}(\alpha, \lambda)$ ). The derivation exploits a key property of gamma random samples: the independence between the total sum $S = \sum X_i$ and the associated Dirichlet vector of proportions $(X_1/S, \dots, X_n/S)$ .

Based on this independence and the homogeneity of $g$ , the authors propose an estimator $\hat{I}_{g,m}$ based on a U-statistic of order $m$ :
$\hat{I}_{g,m} = \frac{1}{\binom{n}{m} C m \bar{X}} \sum_{1 \leq i_1 < \dots < i_m \leq n} g(X_{i_1}, \dots, X_{i_m})$
where $\bar{X}$ is the sample mean. Theoretical properties, including strong consistency and asymptotic normality, are established using the classical theory of U-statistics, the law of large numbers, and the delta method.

Key Contributions

Unified Framework: The paper formalizes a general class of NPRIs that subsumes numerous well-known inequality and variability indices under a single mathematical formulation.
Explicit Expressions: For the gamma distribution, the authors derive explicit analytical expressions for the true values of various NPRIs (e.g., Gini, Power, Asymmetric, Shannon Entropy indices) in terms of the shape parameter $\alpha$ .
Unbiased Estimation: The primary theoretical contribution is the proof that the proposed U-statistic-based estimator is unbiased for any NPRI when the population is gamma-distributed. This provides a single estimation procedure applicable to the entire class of indices.
Generalizability: While the theoretical unbiasedness is proven for the gamma distribution, the paper investigates the estimator's performance under the Generalized Gamma (GG) distribution via simulation, suggesting robustness beyond the strict gamma assumption.

Results

Simulation Study: Monte Carlo simulations ( $M=5,000$ $M = 5, 000$ replications) were conducted for various indices (Gini, Power, Asymmetric, SCV, etc.) under Gamma populations with different shape parameters ( $\alpha \in \{1, 3, 5\}$ $α \in {1, 3, 5}$ ) and sample sizes ( $n \in \{10, 20, 50\}$ $n \in {10, 20, 50}$ ).
- The results demonstrate that the estimator is effectively unbiased, with bias and Root Mean Square Error (RMSE) decreasing as sample size increases.
- Sensitivity analyses regarding the order $m$ of the index show stable performance across different configurations.
- In the Generalized Gamma setting, the estimator converges and maintains low bias, indicating potential applicability beyond the strict gamma model.
Real-Data Application: The methodology was applied to a dataset of Gross Domestic Product (GDP) per capita for 34 countries and territories in the Americas (2023).
- Goodness-of-fit tests (Kolmogorov–Smirnov, Cramér–von Mises, Anderson-Darling) confirmed that the Gamma distribution is a suitable model for the data.
- The authors calculated estimates for various NPRIs (Gini, Power, Maximum, Minimum, etc.) using the proposed estimator.
- Comparisons between non-parametric estimates and parametric estimates (derived from the explicit formulas in Table 2) showed consistency, with parametric confidence intervals generally being narrower.

Significance and Claims
The paper claims to provide a unified estimation framework for a broad family of normalized scale-invariant indices. By leveraging the specific decomposition properties of the gamma distribution, the authors achieve a simple, analytically tractable, and unbiased estimator that applies simultaneously to diverse measures of inequality and variability.

The significance lies in:

Theoretical Unification: Moving away from ad-hoc derivations for specific indices to a general theory based on homogeneous functions.
Practical Utility: Offering a reliable method for estimating these indices in economic and reliability contexts where gamma distributions are prevalent.
Empirical Validation: Demonstrating through simulation and real-world GDP data that the estimator performs well, providing a tool for analyzing economic inequality in the Americas with distinct interpretations offered by different index choices within the same framework.

The authors conclude that while the current work focuses on the gamma distribution, future research may extend these asymptotic properties and variance estimation techniques to other distributional families.