📈 economics

From Prediction to Causation: Heterogeneous Policy Effects of Rent Stabilization and Rental Assistance on Severe Housing Cost Burden in New York City

This paper employs an AI-driven causal inference framework combining machine learning with quasi-experimental designs to demonstrate that New York City's housing policies, particularly the 2019 rent stabilization expansion, effectively reduce severe housing cost burden but exhibit significant heterogeneous effects across different neighborhood contexts, thereby enabling more precise, targeted policy interventions.

Original authors: Nakib Uddin Ahmed, Azizur Rahman, Mehjabin Ferdous, Kadirur Rahman Chowdhury

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nakib Uddin Ahmed, Azizur Rahman, Mehjabin Ferdous, Kadirur Rahman Chowdhury

Original paper licensed under CC BY 4.0 (https://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

In the complex landscape of modern cities, housing is more than just a place to sleep; it is the foundation upon which health, education, and economic stability are built. When a family spends more than half of their income just to keep a roof over their heads, they face what experts call a severe housing cost burden. This condition forces impossible choices between paying rent and buying food, or between keeping the lights on and seeing a doctor. For decades, city planners and policymakers have tried to fix this problem with various tools, from rent control laws that cap how much landlords can charge, to emergency cash grants for those who fall behind. However, a persistent gap in our understanding has remained: while we can predict which neighborhoods are struggling, we have struggled to prove exactly which specific policies actually work to relieve that pressure, and for whom. The challenge lies in separating the effects of a specific law from the many other forces at play, such as the local economy or the age of the buildings, to see the true cause-and-effect relationship.

A team of researchers has now tackled this difficult question by developing a new way to look at housing data, one that blends the predictive power of artificial intelligence with rigorous methods for finding causes. Focusing on New York City, a place where housing costs have long been a crisis, they examined eight major policy instruments designed to help residents. Instead of relying on a single method, they built a framework that first uses machine learning to understand the complex patterns of the city, and then applies three different types of statistical tests to determine which policies actually reduce the burden on families. Their approach allows them to move beyond simple predictions to identify not just what is happening, but why it is happening, and how the results change depending on the specific neighborhood.

The researchers began by training a sophisticated computer model on data from 155 distinct neighborhoods across New York City, covering the years from 2015 to 2023. This model learned to forecast housing cost burdens with high accuracy, identifying that factors like the age of the housing stock, the rate of evictions, and the economic opportunities available in a neighborhood were the strongest predictors of financial stress. By using these machine-learning insights, the team created a detailed map of the city's housing landscape, which they then used to test the impact of specific laws. They did not treat all policies the same; instead, they matched each policy with the most appropriate method to measure its effect, ensuring that the results were as reliable as possible.

One of the most significant findings concerned a major expansion of rent stabilization laws passed in 2019. This law extended protections to more apartment buildings, preventing landlords from raising rents as freely as they had before. The study found that this expansion successfully reduced the rate of severe housing cost burden by 6.2 percentage points. In practical terms, this means that in the neighborhoods where the law applied, a noticeably smaller share of families were forced to spend more than half their income on rent. The researchers were able to confirm this result by showing that the trend was stable before the law changed and shifted only after it took effect, suggesting a direct causal link rather than a coincidence.

However, the study also uncovered a surprising and counterintuitive result regarding emergency rental assistance. During the pandemic, a program was launched to provide immediate cash help to households struggling to pay rent. Contrary to the intention of the program, the data showed that eligibility for this emergency aid was associated with an increase in severe housing cost burden. The researchers suggest this was not because the money failed to help, but likely because the program arrived at a time when housing instability was already skyrocketing due to job losses, and the assistance often reached families who were already in deep crisis rather than preventing the crisis from starting. In contrast, a different type of aid, known as CityFHEPS, which provides ongoing monthly vouchers rather than a one-time emergency grant, showed a strong positive effect, reducing the burden by over 9 percentage points. This distinction highlights that the design and timing of a program matter just as much as the money itself.

The researchers also discovered that the effectiveness of these policies is not uniform across the city; it depends heavily on the context of the neighborhood. The rent stabilization expansion worked best in neighborhoods that already had strong economic opportunities but were facing high rates of eviction and displacement. In these areas, the laws acted as a shield, preventing families from being pushed out of thriving communities. Conversely, other policies, such as reforms to housing court access and tenant protection ordinances, showed their greatest impact in the most vulnerable neighborhoods, where the risk of losing a home was highest and the legal system was often a barrier for residents. This suggests that a single "one-size-fits-all" solution cannot solve the housing crisis; different tools are needed for different parts of the city.

By combining these findings into a comprehensive policy matrix, the authors illustrate that the most effective strategy is a coordinated approach. They found that while individual policies like rent stabilization, tenant protections, and housing court reforms all made a difference, they work best when layered together. Stabilization prevents rents from rising too fast, protections keep people from being illegally evicted, and ongoing vouchers bridge the gap between income and rent. The study concludes that to truly solve the problem of housing insecurity, cities must move beyond isolated fixes and adopt a multi-pronged strategy that addresses the different ways families lose their homes. This research provides a clear roadmap for policymakers, showing that with the right combination of tools and a deep understanding of local conditions, it is possible to build a more stable and affordable future for urban residents.

Technical Summary: From Prediction to Causation in Housing Policy

Problem Statement
Severe housing cost burden, defined as households spending more than 50% of their income on rent, affects approximately one in four New York City (NYC) households, with detrimental consequences for health, education, and economic mobility. While extensive policy responses exist, there is a critical gap in urban policy research regarding rigorous causal evidence on which specific interventions reduce housing insecurity. Existing literature suffers from three primary limitations: (1) a reliance on predictive modeling without causal identification; (2) a neglect of heterogeneous effects across different neighborhood contexts; and (3) a failure to integrate machine learning (ML) prediction with causal inference to simultaneously forecast outcomes and estimate policy impacts.

Methodology
The authors propose an AI-driven causal inference framework that integrates ensemble machine learning for prediction with a "triple" quasi-experimental design for causal identification. The study utilizes a neighborhood-level panel dataset covering 155 Neighborhood Tabulation Areas (NTAs) in NYC from 2015 to 2023, aggregating data from the NYC Housing and Vacancy Survey, HPD administrative records, NYS HCR, the U.S. Census ACS, and the NYC Open Data Portal.

The empirical strategy proceeds in three stages:

Prediction Stage: An ensemble ML model (stacking LightGBM, XGBoost, and Random Forest with elastic net meta-learning) is trained to predict severe housing cost burden. This stage identifies 30 key predictive features (including neighborhood opportunity index, eviction rates, and median income) which serve as the confounder set for subsequent causal analyses. The ensemble achieves an out-of-sample $R^2$ of 0.87.
Causal Identification Stage: The framework applies distinct quasi-experimental designs tailored to the institutional features of eight specific policy instruments:
- Difference-in-Differences (DiD): Used for the 2019 Housing Stability and Tenant Protection Act (HSTPA) rent stabilization expansion, exploiting the policy's application to buildings with three or more units.
- Regression Discontinuity (RD): Applied to the Emergency Rental Assistance Program (ERAP) and CityFHEPS, utilizing sharp income eligibility cutoffs (80% Area Median Income and specific poverty thresholds, respectively).
- Double/Debiased Machine Learning (DML) with Causal Forests: Employed for five additional policy instruments (tenant protection ordinances, housing court access, preservation initiatives, construction subsidies, and inclusionary zoning) to estimate heterogeneous treatment effects while controlling for the 30 ML-derived features.
Heterogeneity Analysis Stage: The framework characterizes how causal effects vary across neighborhood contexts (e.g., opportunity indices, eviction rates) using the rich set of controls derived from the ML model.

Key Results
The study quantifies the causal effects of eight housing policy instruments on severe housing cost burden:

Rent Stabilization (HSTPA 2019): The expansion of rent stabilization to 3+ unit buildings reduced severe housing cost burden by 6.2 percentage points (95% CI: [-6.8, -5.7]), a 17.6% reduction from the baseline mean. Event study analysis confirmed no pre-trends, supporting the causal interpretation.
Rental Assistance (ERAP vs. CityFHEPS): The results revealed a stark contrast. Eligibility for the Emergency Rental Assistance Program (ERAP) was associated with a 6.2 percentage point increase in severe burden during the pandemic period. Conversely, the CityFHEPS voucher program eligibility was associated with a 9.1 percentage point reduction in severe burden.
Other Interventions:
- Tenant Protection Ordinances: Reduced burden by 5.6 percentage points.
- Housing Court Access Reform: Reduced burden by 5.1 percentage points.
- Housing Preservation Initiatives: Reduced burden by 4.4 percentage points.
- Construction Subsidies (e.g., LIHTC): Reduced burden by 2.4 percentage points.
- Inclusionary Zoning/Upzoning: Showed negligible average effects (0.0 percentage points).

Heterogeneity of Effects
The analysis uncovered significant effect heterogeneity based on neighborhood context:

Stabilization Policies (HSTPA): Were most effective in high-opportunity, high-eviction neighborhoods. This suggests stabilization works best where economic foundations are strong but market pressures threaten affordability.
Protection and Access Reforms: Showed stronger impacts in vulnerable areas with higher baseline housing insecurity and lower opportunity scores.

Significance and Contributions
The paper claims three primary contributions to the literature:

Methodological Integration: It demonstrates how ensemble machine learning enhances traditional econometric approaches by improving prediction accuracy ( $R^2 = 0.87$ ) and providing data-driven feature selection for causal model specification, achieving more flexible confounder adjustment than traditional parametric methods.
Heterogeneity Characterization: By employing causal forests and ML-derived controls, the study reveals complex interaction patterns, showing that policy effectiveness is not uniform but depends heavily on neighborhood opportunity structures and eviction rates.
Comprehensive Policy Matrix: The authors construct a policy matrix comparing the causal effects of eight instruments using the most appropriate identification strategy for each. This provides evidence-based guidance for policy combinations, suggesting that effective housing policy requires a coordinated approach combining stabilization (to prevent cost spikes), protection (to prevent unlawful displacement), and access (to bridge income-cost gaps).

The authors conclude that their AI-enhanced causal inference framework offers a template for evaluating urban housing policies, moving beyond pure prediction to establish credible causal estimates and guiding precision policy design in cities facing affordability crises. They note limitations regarding potential unobserved confounding and the aggregation of data at the neighborhood level, suggesting future work should validate findings with individual-level data and explore long-term supply effects.

Technical Summary: From Prediction to Causation in Housing Policy

More like this