Poster #21 - Automation Configurations and Public Values in Homelessness Service Prioritization

• In Event: Saturday Poster Luncheon

Abstract

Governments increasingly adopt automated and algorithmic decision-making (ADM) systems in policy domains such as criminal justice, child welfare, and homelessness. Yet research on deployed public-sector ADM systems shows that they can introduce errors, intensify administrative burdens, and reinforce existing inequalities. For public administration, the key issue is not only whether algorithms perform well, but how automation shapes the “three Es” of effectiveness, efficiency, and equity — core public values in practice. Despite the growth of ADM in public services, existing research rarely offers a systematic way to distinguish where automation occurs within decision processes, how much authority is delegated at each stage, and how those design choices relate to program-level outcomes and public values. 

To address this gap, we examine how different automation configurations across decision stages shape effectiveness, efficiency, and equity in homelessness service prioritization. We extend and adapt Parasuraman’s (2000) stage-and-level framework for automation, originally developed in the human-computer interaction literature, to the context of public-sector decision tools. The framework distinguishes four stages of decision-making: information acquisition, information analysis, decision selection, and action implementation.

We apply this framework to homelessness services, where prioritization systems are shifting from survey-based assessments to more data-driven models that use automation and prediction to streamline decision-making. In this setting, supportive housing is a scarce resource, making prioritization necessary for service allocation. Drawing on content analysis of publicly available documents and prior studies, we compare three prioritization tools that span a range of automation: the Vulnerability Index-Service Prioritization Decision Assistance Tool (VI-SPDAT), the Maricopa Assessment and Prioritization (MAP) tool, and the Allegheny Housing Assessment (AHA). 

We find that these tools are best compared not by whether they use algorithms, but by where automation is embedded and what kinds of human judgment, data quality, and accountability each configuration requires. The clearest differences lie in information acquisition and analysis rather than in final referral decisions, where human discretion remains important. VI-SPDAT relies most heavily on direct interviews, which may better capture clients’ changing circumstances but also imposes substantial staff time and administrative burden. AHA is more automated, using administrative data and predictive models to improve efficiency and standardization, yet its performance depends heavily on upstream data quality and may miss vulnerabilities not visible in formal systems. MAP occupies a middle position by combining administrative data with limited interview input. Although the tools differ in design, all three largely preserve a prioritarian approach to serving those judged most vulnerable, and equity concerns persist across them through different mechanisms. 

This study contributes to research on algorithm-assisted decision-making in public administration by offering a framework that treats automation as a continuum of governance choices rather than a binary distinction between human and algorithmic decision-making. In doing so, it shows how automation design shapes the distribution of discretion, responsibility, and burden within public service systems. More broadly, it argues that ADM should be assessed not only in terms of model performance, but also in terms of programmatic effectiveness, efficiency, and equity.

Authors

• Lingwei Cheng, University of Central Florida
  Presenting Author

• Kyungmin Lee
  Non-Presenting Co-Author

• Andrew Alfred Sullivan, University of Central Florida
  Non-Presenting Co-Author