TLDR: This research introduces a framework using the Mallows model to study how the spatial distribution of inequality affects the effectiveness of AI-powered resource allocation, specifically for door-to-door eviction prevention outreach. It defines the RENT metric to compare targeting vs. non-targeting approaches. The study, calibrated with St. Louis eviction data, finds that targeted policies significantly outperform non-targeted ones in reaching high-risk households, even in highly segregated areas, reconciling previous conflicting findings by emphasizing deployment costs and the true concentration of risk.
In the realm of social service provision, the question of how to effectively allocate scarce resources is paramount. With the rise of Artificial Intelligence (AI), prediction-based policies are increasingly being used to target aid, whether to specific individuals or broader settings like neighborhoods. However, previous research has presented conflicting views on the usefulness of individual-level targeting, especially when societal inequality is high.
A new research paper, titled “Who pays the RENT? Implications of Spatial Inequality for Prediction-Based Allocation Policies,” by Tasfia Mashiat, Patrick J. Fowler, and Sanmay Das, delves into this discrepancy. The authors introduce a novel framework, based on the classic Mallows model, to understand how the geographical distribution of inequality impacts the effectiveness of door-to-door outreach policies. Their focus is on preventing tenant eviction, a critical issue with significant negative implications for low-income households.
The core of their analysis revolves around a new metric called RENT (Relative Efficiency of Non-Targeting). This metric helps assess how effective targeting approaches are compared to neighborhood-based, non-targeting methods in reaching high-risk households. The research calibrates its model using real-world eviction court records from a medium-sized city in the USA, specifically St. Louis, Missouri, a city known for its high levels of segregation.
Understanding the Approach: Targeting vs. Non-Targeting
Prediction-based resource allocation systems typically generate a ‘score’ for each individual, indicating their need for intervention. In situations with limited resources, these scores help prioritize who receives aid until the budget is exhausted. Non-targeting methods, on the other hand, proportion resources based on the aggregated welfare of a larger unit, like a neighborhood. For instance, in door-to-door canvassing for eviction prevention, non-targeting might involve approaching all residents in a high-risk area, regardless of individual precarity. While easier to coordinate, this can be inefficient for low base-rate events like eviction filings.
The paper examines two types of prediction-based targeting policies: High-Risk Property Targeting (HPT), which focuses only on high-risk properties within prioritized neighborhoods, and Top-k Property Targeting (TPT), which targets the ‘k’ highest-risk properties overall. A crucial aspect of their model is accounting for deployment costs, particularly the higher cost of traveling between neighborhoods compared to within one.
Key Findings: Targeting’s Surprising Value
The research reveals considerable gains in the number of high-risk households reached through individually targeted policies. Surprisingly, this holds true even in a highly segregated metro area like St. Louis, where risks of eviction might seem concentrated. The study found that the dispersion of risk in St. Louis, even with its high segregation, corresponds to a scenario where targeting provides significant value.
This finding helps reconcile the apparent discrepancies in prior literature by highlighting two critical factors: first, the source of deployment costs (e.g., travel costs between locations), and second, the difference between intuitively perceived concentrations of risk and their mathematical representation in formal models. When the cost of outreach is considered, and high-risk properties are spatially dispersed (even if segregation is high by other measures), targeting policies substantially outperform non-targeting ones.
Also Read:
- AI and Homelessness: Policymakers’ Views on Integrating Technology for Housing Services
- Advanced Optimization Techniques for Fairer Electoral Maps
Real-World Implications
Applying their framework to St. Louis City and County, the authors demonstrate that targeted policies consistently lead to a greater reduction in the expected number of evictions compared to non-targeting policies. For example, with a 30% reduction in eviction probability and a budget equivalent to visiting properties in three neighborhoods, a targeting policy in St. Louis County could lead to an expected reduction of 67 evictions, versus only 12 for a non-targeting policy.
This research provides valuable insights for policymakers deploying AI-based solutions in social service provision. It underscores the importance of accounting for specific applications and geographies, suggesting that even in areas with high spatial inequality, precision targeting can lead to more efficient and impactful resource allocation. For more details, you can refer to the full research paper here.
