Technical Summary: Tradeoffs are Domain Dependent: Improving Accuracy and Fairness in Property Tax Assessments

Problem Statement

Algorithmic fairness research has long operated under the assumption of a universal tradeoff between fairness and accuracy: improving a model's fairness often necessitates a reduction in predictive accuracy. This paper challenges that assumption within the specific domain of U.S. property tax assessments.

Property tax assessments in the U.S. are currently plagued by systematic regressivity, where owners of lower-valued properties pay a disproportionately higher tax burden relative to their home value compared to owners of higher-valued properties. This violates the principle of vertical equity. While regressivity is often attributed to corruption or legal caps on assessment increases, the authors argue that a primary driver is the reliance on sparse data and simplistic models by assessors. Unlike buyers and sellers who utilize rich data (inspections, appraisals) to determine fair market value, assessors often rely on limited exterior data, causing assessments to regress toward the mean. This results in low-value homes being over-assessed and high-value homes being under-assessed.

The central research question is whether efforts to increase assessment accuracy (by utilizing more data and better models) inherently lead to increased fairness (reduced regressivity), or if a tradeoff exists in this domain.

Methodology

The authors utilize a comprehensive dataset of 26 million property sales spanning 2018–2023, covering 95% of U.S. counties (2,844 out of 3,007). The data, sourced from Cotality (formerly CoreLogic), includes transaction details, property characteristics, and tax information, merged with American Community Survey (ACS) socioeconomic data.

The study employs three complementary analytical approaches:

1. Empirical Correlation Analysis: The authors analyze the relationship between assessment accuracy and fairness across jurisdictions under "status quo" practices. They measure accuracy using Mean Absolute Percentage Error (MAPE) between assessed values and actual sale prices. Fairness is measured using three distinct regressivity metrics: the Log Coefficient (LC), the Suits Index (SI), and the Price-Related Differential (PRD).
2. Counterfactual Simulation (Ablation Experiment): Using LASSO and Random Forest models trained at the county level, the authors simulate assessment pipelines. They compare a "sparse" model (using only three highly available property features) against a "rich" model (using all available property features). This tests whether adding property characteristics improves both accuracy and fairness simultaneously.
3. Census Data Integration Experiment: The authors simulate the incorporation of publicly available Census block group data (e.g., neighborhood income, education, demographics) into Random Forest assessment models. This experiment evaluates whether adding neighborhood-level socioeconomic data, which is currently underutilized by most assessors, yields Pareto improvements (gains in both accuracy and fairness) relative to status quo predictions.

The evaluation framework utilizes a Pareto frontier approach. The authors examine whether counterfactual models offer simultaneous gains in accuracy and fairness relative to baselines. If a model improves both metrics without tradeoff, it represents a Pareto improvement.

Key Results

1. Strong Correlation in Status Quo Practices

Across 98% of counties analyzed, status quo assessments were regressive. However, a strong positive correlation exists between accuracy and fairness: counties with lower assessment errors (higher accuracy) generally exhibited lower levels of regressivity (higher fairness). This suggests that methodological improvements in current practices could yield joint gains.

2. Simulated Models and Feature Addition

In the ablation experiment comparing sparse vs. rich LASSO models:

• Adding property features improved accuracy in the majority of cases.
• Crucially, in 99.9% of counties where accuracy improved, fairness also improved. Only one county (DeSoto County, Missouri) exhibited a tradeoff.
• The magnitude of fairness gains tended to correlate with the magnitude of accuracy gains.

3. Impact of Census Data

In the experiment incorporating Census block group data into Random Forest models (excluding counties with statutory assessment caps):

• Significant simultaneous improvements in accuracy and fairness were observed in 14.7% of counties.
• Among those counties showing significant change, 96.2% experienced simultaneous gains in both metrics, with only 3.8% showing tradeoffs.
• The addition of Census data generally increased assessed values for high-priced homes and decreased them for low-priced homes, thereby reducing regressivity.

4. Intersectional Fairness (Race and Income)

The inclusion of Census features also reduced regressivity with respect to neighborhood-level race and income:

• Race: Predominantly Black neighborhoods (defined as >30% Black residents) saw average decreases in assessed values, while predominantly non-Black neighborhoods saw increases. This suggests a reduction in the tax burden for Black homeowners.
• Income: Assessments decreased in lower-income neighborhoods and increased in higher-income neighborhoods, making the tax system more progressive with respect to income.

Key Contributions

• Domain-Specific Evidence: The paper provides the first large-scale empirical evidence that the fairness-accuracy tradeoff is not universal, specifically demonstrating that in property tax assessments, accuracy and fairness are strongly correlated.
• Methodological Framework: It introduces a holistic evaluation of assessment pipelines using domain-relevant metrics (vertical equity/regressivity) rather than abstract fairness constraints, showing that tradeoffs are the exception rather than the rule in this setting.
• Policy-Ready Interventions: The study identifies that incorporating publicly available Census data is a feasible, low-cost reform that can significantly improve both the accuracy and equity of property tax assessments in hundreds of U.S. counties.
• Challenge to Prevailing Assumptions: The findings challenge the implicit assumption in fair ML literature that building fairer models requires forfeiting accuracy, suggesting that in regression-based public sector systems, "fairness can come for free" through better data and modeling.

Significance and Claims

The authors claim that their findings fundamentally challenge the presumed universality of the fairness-accuracy tradeoff. They argue that the tradeoff is domain-dependent. In the context of property tax, where fairness is defined as vertical equity (proportionality to value) and accuracy is defined as proximity to market value, improving the model's ability to predict market value inherently corrects the systematic biases that cause regressivity.

The paper posits that the perceived tradeoff often arises from:

1. Data Sparsity: Current models lack the data necessary to distinguish between property values, leading to mean-reversion errors that hurt low-value owners.
2. Metric Misalignment: Using abstract fairness constraints (like statistical parity) rather than domain-specific equity principles (like vertical equity).

The authors conclude that well-designed modeling improvements—specifically the adoption of tree-based Automated Valuation Models (AVMs) and the integration of Census data—can advance both fairness and accuracy simultaneously in large-scale public sector systems. They emphasize that while tradeoffs are theoretically possible (e.g., if accuracy gains are clustered only among high-value appeals), they are not observed in the practical application of adding data to assessment models.

Limitations and Scope

The authors remain modest regarding the scope of their claims:

• Baseline Dependency: The "no tradeoff" finding is relative to current status quo models and sparse baselines. Tradeoffs might emerge if assessors were already operating near an efficiency frontier or if different fairness definitions were used.
• Data Representativeness: The study relies on brokered sales data, which may not perfectly represent the entire population of assessed properties (e.g., excluding long-term owner-occupants). However, weighting experiments suggest this does not significantly alter the findings.
• Sale Price as Ground Truth: The study assumes sale price equals fair market value. The authors acknowledge that sale prices themselves may be biased by racial discrimination (e.g., undervaluation of homes in Black neighborhoods). While their models improve accuracy relative to sale price, they note that this does not fully solve the underlying issue of racial devaluation in the market, though it does reduce the tax burden on those neighborhoods.
• Legal and Structural Barriers: The paper notes that statutory caps on assessment increases in 17 states and the lack of individual-level race data for assessors limit the full realization of these gains.