Using Linked Administrative Data for Predicted Probabilities of Homelessness: An Opportunity to Enhance Targeting In Arizona

Background and Purpose

Social workers have identified ending homelessness as one of the profession’s Grand Challenges. Homelessness is at an all-time high in the United States and is increasing. In Arizona specifically, the homelessness rate increased 71% from 2017-2025, with entries into the system consistently outpacing exits, reaching a record high of 15,259 in 2025.

Losing a home or living without one is traumatic and increases the risks for morbidity and mortality. Moreover, homelessness disproportionately impacts People of Color and other minoritized populations due to overlapping structural factors. In this context of rising harm and inequity, homeless prevention is crucial. However, studies show that traditional prevention efforts often do not target those at highest risk of homelessness. One logical space where targeted interventions might be embedded are statewide call and referral programs, such as 2-1-1. This study employed multiple logistic regression to learn which populations are at highest risk of homelessness among housed callers to 2-1-1 Arizona.

Methods

We integrated administrative data from Arizona’s 2-1-1 referral system and the state’s Homeless Management Information System, creating a dataset of nearly 240,000 calls from housed adults between 2020-2024. Logistic regression analyses were conducted and interpreted as predicted probabilities to account for all covariates with this linked data to examine theoretically important variables, demographics, and service use associated with homelessness in the year after a call. R (v. 4.5.2) was used for analyses.

Results

The rate of homelessness among all calls in the dataset was 16.06% in the year after the call. Several predictors had significantly higher or lower probabilities of homelessness than this base rate. Among calls from those who previously experienced homelessness, the adjusted predicted probability was 28% (CI [.27-.28], p < .001). Among calls from those who stayed in hospitals or psychiatric facilities the night before calling, probabilities were 29% and 28%, respectively (CI [.27-.31], p < .001; CI [.24-.30], p < .001), and from those who stayed in foster care the probability was 28% (CI [.22-.34], p < .001). Sociodemographic disparities were also evident: probabilities were 21% among calls from Black/African Americans (CI [.20-.21], p < .001), 20% among calls from gender expansive callers (CI [.17-.22], p = .004), and 21% among calls from Hispanic callers (CI [.21-.22], p < .001). Regarding service use, among calls from those who previously called 2-1-1, the probability was 20% (CI [.19-.20], p = .004). On the other hand, among calls from those who accessed 2-1-1’s virtual case management service, the probability was 11% (CI [.11-.12], p < .001).

Conclusions and Implications

Outcomes, shared with Arizona 2-1-1 for targeting and program enhancement, can inform other prevention efforts using linked administrative data. By embedding data science into program design, social services systems can shift to proactive, targeted prevention. This could involve offering specific groups of 2-1-1 callers enhanced services, such as virtual case management. This study illustrates how linked administrative data can address a Grand Challenge for Social Work and has the potential to reduce harm and inequity.