Allocating Interventions Based on Predicted Outcomes: A Case Study on Homelessness Services

Amanda R. Kube, Sanmay Das, Patrick J. FowlerAAAI 2019 · DOI

Introduction

Homelessness service providers operate in an environment of extreme scarcity, where the demand for housing assistance consistently outstrips available supply. Traditionally, the allocation of resources—ranging from emergency shelter beds to long-term rental assistance—has relied on human intuition or simplified risk assessment tools. While these methods aim to identify "high-risk" individuals, they often fail to determine which specific intervention is most effective for a particular household. This paper addresses this gap by framing resource allocation as a causal inference problem: to optimize the system, we must predict not just how a family performed with the service they received, but how they would have performed under alternative interventions.

Methodology: Counterfactual Modeling and Optimization

The researchers propose a two-stage framework combining predictive modeling with constrained optimization.

First, they employ Bayesian Additive Regression Trees (BART) to estimate individual treatment effects. BART is an ensemble method particularly suited for capturing the non-linearities and complex interactions inherent in social services data. By using BART, the authors generate counterfactual estimates—the predicted probability of "re-entry" (returning to the homeless system within two years) for every household across four distinct intervention types.

Second, the problem is framed as an Integer Programming (IP) optimization. The objective is to minimize the total expected number of re-entries across the entire population, subject to the real-world capacity constraints of the service providers.

The mathematical formulation of the optimization problem, aiming to minimize total expected re-entries subject to capacity constraints.

To address ethical concerns, the authors also introduced a fairness constraint. A purely efficient system might sacrifice the outcomes of specific individuals to lower the global re-entry rate. The constrained IP ensures that no household is assigned a service that increases their predicted probability of re-entry by more than 5 percentage points compared to their original, human-assigned service.

Data and Features

The study utilizes administrative data from the Homeless Management Information System (HMIS) of a major U.S. metropolitan area between 2007 and 2014. The dataset includes 7,474 households and 68 variables, including demographics, housing history, disability status, and income. The primary target variable is system re-entry within two years of exiting a service.

Key Results and the Efficiency-Fairness Trade-off

The results demonstrate that data-driven allocation could significantly improve system performance:

  • Baseline Performance: The actual re-entry rate under human-led allocation was 43.04%.
  • Unconstrained Optimization: The BART-optimized model predicted a re-entry rate of 31.88%, representing a relative reduction of 27.08%.
  • Constrained Optimization: With the 5% fairness constraint, the predicted re-entry rate was 37.38%, still a 14.66% improvement over the status quo.

One of the most significant findings is the "Pareto Problem." To achieve the global minimum re-entry rate, the unconstrained model often reassigns households to services where they are predicted to do worse than their original assignment. This occurs because another household is predicted to benefit far more from that specific slot. In the unconstrained model, approximately 32.62% of households were assigned to "worse" outcomes for the sake of global efficiency.

A histogram showing the distribution of changes in predicted reentry probability when moving from the original allocation to the optimal allocation.

Discussion and Future Directions

The analysis suggests that current systems may be under-utilizing high-impact services for the most vulnerable populations. The optimization algorithm frequently prioritized households with lower incomes and longer wait times for intensive interventions like "Homelessness Prevention."

However, the authors acknowledge several limitations. Observational bias remains a risk, as unrecorded factors may have influenced original assignments. Furthermore, the model is currently static; future iterations could incorporate dynamic matching to account for the continuous arrival of households over time. Ultimately, the authors argue that these models should serve as "human-in-the-loop" tools, providing social workers with empirical insights to supplement their qualitative expertise rather than replacing human judgment entirely.

Summary updated

BibTeX
@inproceedings{kube2019allocating,
  author = {Amanda R. Kube and Sanmay Das and Patrick J. Fowler},
  title = {{Allocating Interventions Based on Predicted Outcomes: A Case Study on Homelessness Services}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.3301622},
}
Figure