Mind the Income Gap: Bias Correction of Inequality Estimators in Small-Sized Samples

Technical Summary: Mind the Income Gap: Bias Correction of Inequality Estimators in Small-Sized Samples

Problem Statement
The estimation of income inequality in small domains (e.g., specific regions or socio-demographic groups) is frequently hindered by data scarcity. While Small Area Estimation (SAE) techniques are employed to improve reliability by "borrowing strength" across areas, they often rely on area-level models (such as the Fay-Herriot model) that assume input survey estimators are unbiased. However, inequality estimators (e.g., Gini, Generalized Entropy, Atkinson) are inherently non-linear functions of income data. In small samples, particularly under complex survey designs involving stratification and multi-stage sampling, these estimators exhibit significant bias, typically leading to an underestimation of inequality. Neglecting this bias results in model misspecification and misleading inferences, especially when comparing inequality across time or locations with disparate sample sizes. Furthermore, existing bias correction literature is largely limited to independent and identically distributed (iid) samples or specific measures like the Gini index, leaving a gap for complex survey designs and broader classes of inequality measures.

Methodology
The authors propose a design-based bias correction framework that generalizes previous iid approaches (specifically Breunig and Hutchinson, 2008) to finite populations with complex sampling designs. The core of the methodology relies on second-order Taylor series expansions of inequality estimators around their population values.

Key methodological components include:

1. General Bias Formulation: The authors derive a general expression for the approximate bias of an inequality estimator $\hat{\theta} = f(\hat{\mu}, \hat{\gamma})$, where $\hat{\mu}$ is the sample mean and $\hat{\gamma}$ is a Horvitz-Thompson estimator of a transformed income variable. The bias is expressed as a function of the design variances and covariances of these components ($V[\hat{\mu}]$, $V[\hat{\gamma}]$, $Cov[\hat{\mu}, \hat{\gamma}]$).
2. Scope of Measures: The framework is applied to a wide class of measures, including the Gini Index, the Generalized Entropy (GE) family (covering Mean Log Deviation, Theil index, and others), and the Atkinson family.
3. Complex Survey Adaptation: To handle complex designs, the authors utilize the "Ultimate Clusters" technique for variance estimation, accommodating both Self-Representing (SR) and Non-Self-Representing (NSR) strata. For the Gini index, where the rank estimator is non-linear, a linearization approach (Graf, 2011; Graf and Tillé, 2014) is employed to facilitate variance estimation.
4. Non-Parametric Nature: The approach does not require parametric assumptions regarding the underlying income distribution, enhancing its flexibility.

Key Contributions

• First Application to Atkinson and GE in Complex Surveys: To the authors' knowledge, this is the first proposal for bias correction of Atkinson and Generalized Entropy indexes specifically within the context of complex survey designs.
• Extension of Gini Correction: The paper extends existing Gini bias corrections (e.g., Davidson, 2009; Fabrizi and Trivisano, 2016) by utilizing a second-order Taylor expansion rather than a first-order one, resulting in a more refined adjustment term.
• Design-Based Framework: The methodology explicitly accounts for the correlation structures introduced by stratification, multi-stage sampling, and post-survey treatments (calibration/imputation), moving beyond the limitations of iid assumptions.

Results
The efficacy of the proposed framework was evaluated through a design-based simulation study using 2017 Italian EU-SILC data across 21 NUTS-2 regions. Two scenarios were tested: one with raw data and one with extreme values treated (using semi-parametric Pareto-tail modeling).

• Bias Reduction: The correction strategy demonstrated a noticeable reduction in Relative Bias (RB) for all considered measures. In the treated population scenario with a 1.5% sampling rate, bias reductions ranged from approximately 5% (for the Coefficient of Variation) to 14% (for the Gini index).
• Sample Size Dependence: The bias was confirmed to be a small-sample phenomenon. For sample sizes greater than 20 individuals, the bias-corrected estimators approached approximate unbiasedness.
• Robustness to Outliers: Measures sensitive to extreme values (e.g., GE with $\alpha=2$, Atkinson with $\epsilon=2$) exhibited higher bias in untreated data. While the correction reduced bias, the authors note that for populations with extreme values, focusing on robust measures (Gini, GE with $\alpha=0$, Atkinson with $\epsilon=1$) remains advisable.
• Variance Trade-off: The correction induced a slight, generally negligible increase in Absolute Relative Error (ARE) for most measures, except for the Gini index, where the variance inflation was more pronounced due to the specific form of the unbiased estimator adjustment.

Small Area Estimation Application
The authors conducted an illustrative SAE exercise using the Fay-Herriot model to demonstrate the practical consequences of ignoring bias.

• Model Misspecification: Using uncorrected (biased) survey estimates as inputs led to model misspecification.
• Underestimation: The resulting model-based estimates (EBLUP) consistently underestimated inequality levels compared to those derived from bias-corrected inputs. This underestimation was particularly severe for the Gini index.
• Conclusion on SAE: The exercise confirms that preliminary bias correction is a necessary step before applying area-level SAE models to inequality data to avoid misleading policy inferences.

Significance and Claims
The paper claims that ignoring the small-sample bias of inequality estimators in complex surveys leads to systematic underestimation of inequality, a problem exacerbated in small domains and when extreme values are present. The proposed framework offers a flexible, non-parametric solution that significantly reduces this bias without requiring strong distributional assumptions. The authors position this work as a critical refinement for researchers and policymakers relying on local inequality estimates, arguing that accurate measurement requires addressing the bias ex-ante before applying small area models. They modestly note that while the simulation results are promising, they are based on specific data (Italian EU-SILC) and should be viewed as a reference rather than a universal conclusion, suggesting future research directions such as multivariate SAE frameworks and extensions to quintile-based measures.