Predicting Homelessness Among U.S. Army Soldiers No Longer on Active Duty

Abstract

Introduction:

The ability to predict and prevent homelessness has been an elusive goal. The purpose of this study was to develop a prediction model that identified U.S. Army soldiers at high risk of becoming homeless after transitioning to civilian life based on information available before the time of this transition.

Methods:

The prospective cohort study consisted of 16,589 observations of soldiers after separating or being deactivated from service who had previously participated in 1 of 3 baseline surveys of the Army Study to Assess Risk and Resilience in Servicemembers in 2011–2014. A machine learning model was developed in a 70% training sample and evaluated in the remaining 30% test sample to predict self-reported homelessness in 1 of 2 Longitudinal Study surveys administered in 2016–2018 and 2018–2019. Predictors included survey, administrative, and geospatial variables available prior to separation/deactivation. Analysis was conducted in November 2020–May 2021.

Results:

Twelve-month prevalence of homelessness was 2.9% (SE=0.2%) in the total Longitudinal Study sample. Area under the receiver operating characteristic curve in the test sample was 0.78 (SE=0.02) for homelessness. The 4 highest ventiles (top 20%) of predicted risk included 61% of respondents with homelessness. Self-reported histories of depression, lifetime trauma of having a loved one murdered, and post-traumatic stress disorder were the 3 strongest predictors of homelessness.

Conclusions:

A prediction model for homelessness can accurately target soldiers for preventive intervention before transition to civilian life.

INTRODUCTION

Homelessness is an increasingly widespread and intractable crisis in the U.S. Affecting approximately 580,000 people at one point in time in January 2020, mass unemployment and evictions during the coronavirus disease 2019 (COVID-19) pandemic have put even more people at risk of homelessness.^1–3 The issues associated with homelessness are not limited to being without a home but include myriad additional problems, some of which might be causes rather than or in addition to being consequences of homelessness, including high rates of comorbid physical, mental, and substance use disorders,⁴ premature mortality,⁵ healthcare costs and utilization,⁶ suicide,⁷ crime,⁸ and victimization.⁹

Although efforts to address homelessness have most often focused on housing and health interventions for people who are already homeless, preventing homelessness has emerged as an area of increasing policy focus.¹⁰ However, initiatives have most often focused on preventing recurring homelessness in those who have already experienced it, whereas attention to primary prevention of homelessness, that is, preventing it before it ever occurs, has received less attention.¹⁰ Furthermore, efforts focused on homelessness prevention have been hindered by the challenge of efficiency, that is, identifying high-risk populations to maximize targeted provision of interventions to this group but not to those at lower risk.¹¹ Developing systems to predict homelessness accurately before it occurs has been an elusive goal.

One group at high risk for homelessness are military veterans.¹² In 2009, the Obama Administration announced a national imperative to end veteran homelessness within 5 years and dedicated extensive resources to the Department of Veterans Affairs (VA) to achieve this goal.¹³ Although the prevalence of homelessness among veterans has declined since this time, more than a decade later, veterans remain over-represented in the U.S. homeless population.¹⁴ Furthermore, the number of homeless veterans increased for the first time in years in 2020 and is expected to surge following the expiration of the federal moratorium on evictions.¹⁵

Prior literature shows that the strongest and most consistent predictors of homelessness among veterans are substance use disorders and mental illness, followed by low income and other income-related factors.^14,16,17 However, only a few of the studies that documented these associations assessed the predictors before the veterans left active duty.¹⁸ Prior studies also generally used 1 data source (e.g., survey or administrative data but not both), rather than combining multiple data sources. In addition, most of these studies used traditional regression methods to develop prediction models rather than machine learning models. The latter are more flexible in allowing inclusion of nonlinearities and interactions among predictors and in adjusting for overfitting.¹⁹ Although related research has been carried out with veterans,²⁰ no prior prospective study has combined use of multiple predictor data sources with machine learning methods to determine which soldiers are at risk of becoming homeless after returning to civilian life. This is done in the current study using data from the Study to Assess Risk and Resilience in Servicemembers-Longitudinal Study (STARRS-LS), a large epidemiological–neurobiological study of U.S. Army soldiers.

METHODS

Study Sample

Army STARRS administered baseline surveys to 3 separate samples in 2011–2014. Field procedures have been reported extensively elsewhere.^21–24 All participants gave written informed consent approved by the human subjects committees of collaborating organizations to have their de-identified survey data linked to their Army administrative data. Two STARRS-LSs were then administered in 2016–2018 (LS1; n=14,508, weighted response rate of 35.6%) and 2018–2019 (LS2 with LS1 respondents; n=12,156, conditional response rate of 83.7%). A detailed methodological appendix on all STARRS surveys, sample sizes, and field procedures is available upon request. The focus of the current report is on the LS respondents who were no longer on active duty at the time of survey, defined as either separated completely from Army service or deactivated/no longer activated in a Reserve or National Guard Component. LS2 respondents who reported being homeless in the 12 months before LS1 were excluded to avoid double counting cases of homelessness in the pooled LS1–LS2 analysis. The full analysis sample included N=16,589 observations (n=8,797 LS1, n=7,792 LS2).

Measures

The LS surveys included a modified question from the Veterans Health Administration’s (VHA’s) Homelessness Screening Clinical Reminder²⁵ on how much time over the past 12 months respondents were living in stable housing that you own, rent, or stayed in as part of a household. If not all of the time, a follow-up question asked how many of the past 12 months respondents were homeless. A response of ≥1 was coded as homeless in past 12 months and ≥3 as persistent homelessness. Respondents missing the duration question (n=5) were coded as not homeless.

Indicators of 9 categories of predictors of homelessness based on a literature review were selected from the STARRS baseline surveys (136 variables) and Army/Department of Defense (DoD) administrative data systems (566 variables): sociodemographics, Army career variables, personality characteristics, adverse childhood experiences, other lifetime traumatic events, chronic stressors, self-injurious thoughts and behaviors, physical health problems, and mental disorders.^14,26,27 Information was also included about characteristics of the counties where respondents moved after leaving active duty (1,368 variables), which previous research suggests might predict homelessness.^28–31

Statistical Analysis

Analysis (carried out in November 2020–May 2021) used a generalized stacking ensemble machine learning method with 10-fold cross-validation implemented in the Super Learner (SL), version 2.0–28 package^29,32 to predict homelessness. The model was developed in a 70% training sample and then tested in the remaining 30% test sample. As the number of respondents with persistent homelessness was too small to estimate a separate model, prediction accuracy of the overall model was evaluated separately for any homelessness and persistent homelessness. A diverse set of algorithms was included in the SL ensemble (Appendix Table 1).^33,34 The number of predictors was restricted using lasso regression for linear models and random forests for tree-based models. Fifty was the maximum number of predictors allowed in each algorithm, as this is about 1.5 times the maximum number recommended relative to number of respondents with the outcome to avoid overfitting.³⁵

Variable importance was examined using the model-agnostic kernel SHapley Additive exPlanations (SHAP) method to estimate marginal contributions of predictors to overall model accuracy.³⁶ A case-control sampling scheme was used in the training sample to adjust for class imbalance caused by the rarity of homelessness. In addition to estimating area under the receiver operating characteristic curve (AUC) to evaluate model accuracy, calibration with isotonic regression³⁷ was used to determine how well model predicted probabilities approximated observed probabilities of homelessness.^38–40 Model fairness was also evaluated across important population segments.⁴¹ Sensitivity (the proportion of homelessness) and positive predictive value (PPV; prevalence of homelessness) were evaluated within and across ventiles of predicted risk.

RESULTS

Prevalence of homelessness in the 12 months before the survey was 2.9% (SE=0.2%) in the total sample, higher in LS1 (3.7%, SE=0.3%) than LS2 (2.0%, SE=0.2%) and higher among the separated (3.4%, SE=0.3%) than deactivated (1.7%, SE=0.2%) (Table 1). Persistent homelessness was much less common (1.3%, SE=0.1%).

Table 1.

12-Month Prevalence of Homelessness in the Pooled Sample (n=16,589)^a

	Prevalence of homelessnessb		Unweighted homelessness frequency (n)
LS wave and Army status	Any % (SE)	Persistent % (SE)	Any	Persistent	Total
LS1
Separated	4.5 (0.5)	2.0 (0.2)	(243)	(118)	(5,825)
Deactivatedc	1.7 (0.4)	0.6 (0.2)	(49)	(18)	(2,972)
All	3.7 (0.3)	1.6 (0.1)	(292)	(136)	(8,797)
LS2d
Separated	2.1 (0.3)	0.9 (0.2)	(117)	(58)	(5,484)
Deactivated	1.7 (0.3)	0.8 (0.2)	(34)	(16)	(2,308)
All	2.0 (0.2)	0.9 (0.2)	(151)	(74)	(7,792)
LS1/LS2 combined
Separated	3.4 (0.3)	1.5 (0.1)	(360)	(176)	(11,309)
Deactivated	1.7 (0.2)	0.7 (0.1)	(83)	(34)	(5,280)
All	2.9 (0.2)	1.3 (0.1)	(443)	(210)	(16,589)

^aSee the text for a description of weighting.

^bWeighted to correct for nonresponse bias.

^cNo longer activated but still in the National Guard or Army Reserve Component.

^dLS2 respondents who were homeless at any time in the 12 months before the LS1 survey were excluded from the LS2 sample for purposes of this analysis.

STARRS-LS, Study to Assess Risk & Resilience in Servicemembers-Longitudinal Study; LS1, STARRS-LS Wave 1; LS2, STARRS-LS Wave 2.

The AUC of the SL model in the test sample after optimally constraining number of predictors (Appendix Table 2) was 0.78 (SE=0.02) compared to 0.75 (SE=0.02) for a benchmark lasso penalized logistic regression model fit and tuned using 10-fold cross-validation in the training sample and then applied in the test sample. The SL model was superior or equal to the lasso model across all prediction thresholds (Appendix Figure 1).

Soldiers in the 4 highest-risk ventiles (i.e., the 20% of respondents with highest predicted risk of homelessness determined in the training sample, which characterized 22.9% of respondents in the test sample) accounted for 61.1% of homelessness and 63.2% of persistent homeless (Table 2, Appendix Figure 2). PPV among high-risk soldiers was 5.7%–12.2% compared with 0.0%–4.7% among others.

Table 2.

Observed 12-Month Homelessness and Persistent Homelessness in the Test Sample (n=4,977)^a

		Any homelessness				Persistent homelessness
		Sensitivity (SN)		Positive predictive value (PPV)		Sensitivity (SN)		Positive predictive value (PPV)
	Distribtionc	Within-ventile	Cumulative	Within-ventile	Cumulative	Within-ventile	Cumulative	Within-ventile	Cumulative
Risk ventileb	% (SE)	SN (SE)	SN (SE)	PPV (SE)	PPV (SE)	SN (SE)	SN (SE)	PPV (SE)	PPV (SE)
1	5.1 (0.5)	21.2 (3.8)	21.2 (3.8)	12.2 (2.5)	12.2 (2.5)	25.2 (6.4)	25.2 (6.4)	7.8 (2.2)	7.8 (2.2)
2	5.8 (0.5)	14.9 (3.8)	36.1 (5.0)	7.4 (2.4)	9.6 (1.7)	12.1 (5.3)	37.3 (7.5)	3.9 (1.8)	5.9 (1.4)
3	7.1 (0.8)	15.4 (5.3)	51.5 (6.2)	6.3 (3.1)	8.3 (1.7)	15.9 (8.2)	53.3 (8.1)	3.6 (1.9)	5.0 (1.1)
4	4.9 (0.5)	9.6 (3.3)	61.1 (6.2)	5.7 (2.1)	7.8 (1.3)	9.9 (6.1)	63.2 (7.3)	3.5 (2.3)	4.7 (1.0)
5	4.0 (0.5)	4.0 (4.2)	65.2 (5.7)	3.0 (2.5)	7.0 (1.1)	3.4 (2.6)	66.5 (7.2)	1.6 (1.2)	4.2 (0.9)
6	3.3 (0.3)	5.3 (2.2)	70.5 (5.7)	4.7 (1.3)	6.8 (0.9)	4.5 (2.6)	71.0 (7.0)	1.9 (1.1)	3.9 (0.8)
7	6.1 (0.6)	4.3 (1.3)	74.8 (5.5)	2.0 (0.7)	6.0 (0.8)	5.1 (2.7)	76.1 (6.4)	1.8 (0.9)	3.6 (0.7)
8	4.8 (0.4)	8.1 (4.1)	82.8 (4.4)	4.9 (2.5)	5.9 (0.8)	5.0 (3.2)	81.2 (5.9)	1.4 (0.9)	3.3 (0.6)
9	4.7 (0.4)	6.3 (3.2)	89.2 (3.3)	3.9 (1.9)	5.7 (0.7)	5.1 (3.6)	86.3 (4.9)	2.0 (1.4)	3.2 (0.6)
10	4.6 (0.6)	1.2 (1.4)	90.4 (3.0)	0.8 (0.8)	5.2 (0.7)	2.4 (2.2)	88.7 (4.6)	1.1 (1.0)	3.0 (0.5)
11	4.5 (0.5)	2.9 (1.9)	93.4 (2.4)	1.9 (1.1)	5.0 (0.6)	0.0 (0.0)	88.7 (4.6)	0.0 (0.0)	2.7 (0.5)
12	4.4 (0.5)	1.7 (1.2)	95.1 (2.0)	1.1 (0.7)	4.7 (0.6)	4.7 (3.4)	93.4 (3.2)	2.0 (1.4)	2.7 (0.5)
13	5.3 (0.7)	0.0 (0.0)	95.1 (2.0)	0.0 (0.0)	4.3 (0.5)	0.0 (−)	93.4 (3.2)	0.0 (−)	2.4 (0.4)
14	4.7 (0.6)	0.7 (0.7)	95.8 (1.9)	0.4 (0.4)	4.0 (0.5)	2.2 (1.7)	95.6 (2.8)	0.8 (0.6)	2.3 (0.4)
15	4.3 (0.4)	1.4 (1.6)	97.2 (1.1)	0.9 (0.9)	3.8 (0.5)	0.0 (−)	95.6 (2.8)	0.0 (−)	2.2 (0.4)
16	4.5 (0.4)	1.1 (0.5)	98.2 (1.0)	0.7 (0.3)	3.7 (0.4)	2.4 (2.3)	98.0 (1.6)	0.9 (0.9)	2.1 (0.4)
17	5.0 (0.5)	0.4 (0.0)	98.8 (1.0)	0.2 (0.0)	3.4 (0.4)	0.4 (0.4)	98.4 (1.5)	0.1 (0.1)	2.0 (0.3)
18	4.7 (0.4)	0.0 (0.3)	98.8 (0.9)	0.0 (0.2)	3.3 (0.4)	0.0 (−)	98.4 (1.5)	0.0 (−)	1.9 (0.3)
19	6.9 (0.8)	1.2 (0.9)	100.0 (0.0)	0.5 (0.5)	3.1 (0.4)	1.6 (1.5)	100.0 (0.0)	0.4 (0.4)	1.8 (0.3)
20	5.4 (0.7)	0.0 (0.0)	100.0 (0.0)	0.0 (0.0)	2.9 (0.2)	0.0 (−)	100.0 (0.0)	0.0 (−)	1.7 (0.3)

^aThe n=4,977 respondents in the test sample represent 30% of the n=16,589 in the total sample, including n=113 of the n=443 total sample respondents who reported being homeless in the 12 months before their STARRS-LS survey. The remaining 70% of the total sample were in the training sample.

^bDefined in terms of thresholds in the calibrated training sample to separate the sample into 20 subsamples of equal size rank ordered in terms predicted risk. Based on the distribution of predicted probabilities in the training case control sample (n=1,860).

^cAs the thresholds defining ventiles of predicted risk were based on the training sample, the proportions of test sample respondents in each ventile do not equal 5%.

Model performance varied meaningfully across important subsamples, with the AUC higher for men (AUC=0.79, SE=0.02) than women (AUC=0.74, SE=0.04) and among the separated (AUC=0.78, SE=0.02) than deactivated (AUC=0.67, SE=0.07) (Figure 1). Methodological variables related to the study design were also important, with AUCs higher among those who separated/deactivated 0–2 years before their LS survey (AUC=0.81, SE=0.03) than ≥3 years (AUC=0.73, SE=0.02) and among respondents whose baseline survey was <4 years before their LS survey (AUC=0.85, SE=0.05) than ≥4 years (AUC=0.77, SE=0.02) (Figure 1).

Figure 1.

Receiver operating characteristic curves in subsamples of the test sample based on (a) substantive and (b) methodological variables.

AUC, area under the receiver operating characteristic curve.

Model calibration in the test sample was excellent, with standard measures of discrepancy between model and observed data close to 0. Fairness of calibration was also excellent, as indicated by the fact that relative risk of homelessness based on predicted probabilities from the model was comparable across test sample subgroups defined by sex and race–ethnicity (Appendix Table 3).

A total of 373 variables in the predictor set had significant (p<0.05, 2-sided test) zero-order associations with homelessness in the training sample (50.7% of survey variables, 38.1% of administrative variables, 11.1% of geospatial variables). However, only 26 predictors were selected for the optimal final SL model (25 for linear algorithms, 14 for tree-based algorithms, and an overlap of 13). Twelve of these were survey variables. The others were administrative or geospatial variables. However, survey variables were by far the most important (Figure 2). Self-reported lifetime depression and post-traumatic stress disorder were 2 of the 3 most important predictors, along with lifetime trauma of having a loved one murdered. Other mental health indicators, including lifetime generalized anxiety disorder and self-reported indicators of suicidality (lifetime ideation and ≥2 attempts) were also among the significant predictors, all associated with increased homelessness risk. Other key predictors included non-military lifetime traumas (exposure to natural disaster, ≥4 interpersonal losses), indicators of adverse childhood experiences (childhood homelessness, being on welfare as a child, and physical neglect) and 4 geospatial variables (county level Medicaid eligibility rate, high SE disadvantage, very low MVC mortality rate, high % food insecurity), all associated with increased risk of homelessness (results available on request).

Figure 2.

Predictor importance based on kernel SHAP values in the test sample (n=4,977)^a

^aSurvey predictors were measured retrospectively in the time period prior to leaving or being released from active duty, administrative predictors were defined as the earlier of the 2 times of leaving or being released from active duty or December 31, 2016, given that our access to administrative data was only up to the end of 2016. Geospatial predictors were based on the geographic area of residence at the time of the LS1/LS2 survey.

^bThe SHAP value for an individual is the extent to which the predicted probability of homelessness changes when a single variable is deleted from the prediction model averaged across all logically possible combinations of the 26 predictors. The model-agnostic kernel SHAP method was used to estimate SHAP values. As these values can be either positive or negative at the individual level, the mean of the absolute SHAP value across all respondents in the test sample is reported here.

(a), administrative predictor; AC, Army career; ACE, adverse childhood experiences; CL, county level; Ext, external; hosp, hospitalization; (g), geospatial predictor; GAD, generalized anxiety disorder; in/out, inpatient admission/outpatient visit; ins, insecurity; LT, lifetime; LTT, lifetime trauma; MDE, major depressive episode; med, median; mort, mortality; MVC, motor vehicle; NCOs, non-commissioned officers; non-viol, non-violent; perp, perpetrator; PTSD, post-traumatic stress disorder; (s), survey predictor; SE, social-economic; serv, service; SHAP, SHapley Additive exPlanations; stab, stabilized; stnd, standardized; txt, treatment; UL, unit level; yr, year; 12m, 12-month.

DISCUSSION

Results show clearly that homelessness in the 12 months before the LS surveys can be predicted significantly using a flexible ensemble machine learning model that captures about 60% of the soldiers who become homeless in the top 20% of the predicted risk distribution. The fact that the model was based only on a small set of survey predictors (n=12) has meaningful practical implications, as a brief self-report questionnaire could be included feasibly in the Department of the Army Career Engagement Survey, which all soldiers are required to complete before leaving the Army.^42,43 Given that accuracy was inversely proportional to time since leaving active duty, model results based on a survey completed shortly before separation/deactivation would presumably be even stronger than in the current study.

Identifying soldiers at highest risk for homelessness prior to separation/deactivation could facilitate provision of targeted preventive interventions for homelessness. It might even be possible eventually to include this information in DoD electronic health records, which will be linked with VHA electronic health records once the 2 agencies complete implementing a shared system currently in development, creating an electronic risk flag for VHA clinical encounters.

It is striking that of the approximately 2,000 potential predictor variables considered, indicators of mental health emerged as among the most important. Mental illness has been a consistent risk factor for homelessness in prior studies of veterans, although most such studies did not find depression, post-traumatic stress disorder, and generalized anxiety disorder to be particularly important relative to other mental illnesses as was found here.¹⁴

In addition, several significant predictors rarely included in previous studies were documented, including certain non-military lifetime traumas, childhood homelessness, lifetime suicidality, and characteristics of counties where soldiers relocate after separation/deactivation. It is notable that substance use disorder, found in prior studies to be a strong, consistent correlate of homelessness among veterans, was not identified as important in this study. This might be due to substance use problems emerging or worsening after separation/deactivation, given the Army’s careful monitoring for drug abuse while on active duty. However, caution is needed in interpreting these results regarding predictor importance, as this ranking depends on associations of predictors with each other as well as with the outcome.

An important issue in considering preventive interventions is that model PPV is relatively low even among high-risk soldiers. This is inherent in probabilistic models of rare outcomes, leading some commentators to raise concerns about the feasibility of preventive interventions for homelessness.¹¹ It is important to recognize, though, that other widely used preventive interventions, such as interventions to prevent heart attacks and strokes by prescribing statins, set intervention thresholds at lower values of PPV than those in the model developed here.⁴⁴ However, these other decisions are based on 2 factors lacking in research on homelessness: solid evidence for the effectiveness of a rigorously evaluated preventive intervenion with a known cost and known effect size, and broad agreement among policymakers on the societal value of preventing one case of the negative outcome. Once this information become available, conventional decision science methods could be used to set principled cost-effective intervention thresholds for a model such as the one developed here.⁴⁵

The VA is actively involved in a series of new initiatives to help transitioning veterans adjust to life back in the civiian world.⁴⁶ This work is in the early stages, focuses largely on universal interventions, and has not been evaluated to determine effects on reducing homelessness. It might be that more intensive interventions targeted at the new veterans who are estimated to be at high risk would be a useful addition. The only way to know if this is the case, though, would be to implement such an intensive intervention. The latter might build on the resources the VA is now making available to transitioning veterans by helping those targeted as high-risk to overcome barriers to taking advantages of existing resources and determining which, if any, of these resources might be most helpful to them prior to requiring just-in-time interventions.^10,47

The current results only address an initial component of such an agenda by making it clear that a practical risk targeting system could be created.

Limitations

Several limitations need to be noted. Two involve the assessment of homelessness. First, the sample was limited to soldiers who participated in Army STARRS surveys in 2011–2014 and could be traced and resurveyed in 2016–2019, raising the possibility of sample bias. Adjustment for nonrandom response was made to the extent possible, but this is no guarantee against bias. Second, homelessness was assessed exclusively in a self-report question that left it up to respondents to determine what they meant by the term “homeless.” It would have been better to expand the question series to ask specifically about living situations defined officially as homeless that respondents might not consider being homeless (e.g., sleeping on the couch of a friend). This problem has been rectified in subsequent LS waves but presumably led to underestimation here. In a related way, it would have been desirable to link baseline STARRS data to VHA administrative data on homelessness. This would have addressed both the likelihood that LS surveys were less successful in tracing and assessing baseline respondents who became homeless than those who did not become homeless and in addressing the possibility of bias related to variables in the model (e.g., mental illness) in self-reports of homelessness. Importantly, however, survey data among veterans find different correlates of homelessness than those found in VA administrative data of veterans seeking homelessness services, indicating that some segments of the homeless population are under-represented as well in administrative data.⁴⁸ It was not possible to link STARRS survey data with VA administrative data, though, because informed consent was not obtained from baseline STARRS survey respondents to make such linkages.

A third limitation involves the predictors. Specifically, information about the county of residence of the soldier after separation/deactivation was included among the predictors based on evidence of substantial geographic variation in risk of homelessness.⁴⁹ Although this information was collected in the LS surveys, it is noteworthy that DoD records information on planned residential addresses of all separating and deactivating servicemembers prior to these individuals leaving active duty. However, to the extent that residential relocation across county lines occurred after becoming homeless, results might overestimate the importance of county characteristics as predictors of homelessness.

Finally, the study is limited by methodological features of the study design, which was not developed to optimize the prediction of outcomes after separation/deactivation. As a result, substantial variation existed in the time between the baseline Army STARRS surveys and the LS surveys as well as time between separation/deactivation and the LS surveys. This variation meant that instances of homelessness shortly after separation/deactivation were missed that did not persist into the 12 months before LS1. It also meant, as shown in Figure 1, that model prediction accuracy was lower than it may have been if the baseline surveys were administered shortly before separation/deactivation and the LS surveys were administered at uniform times in the first years after separation/deactivation. These methodological limitations suggest that future efforts to refine this model could yield even stronger results than those reported here.

CONCLUSIONS

If the U.S. is to end veteran homelessness, as pledged but not fulfilled more than a decade ago, better strategies are needed to identify individuals at highest risk prior to separation or deactivation for implementation of targeted preventive interventions. The novel findings presented here mark a path forward to accurately predict homelessness in soldiers. Whether a comparable model would work for servicemembers in other branches is unclear, especially in light of evidence that veteran homelessness varies by branch of service.¹⁸ Nor is it clear how much the prediction accuracy of the current model could be improved in a survey that was administered to servicemembers shortly before separation/deactivation for the specific purpose of targeting preventive interventions. The results are clear, though, that even the lower-bound accuracy documented here is adequate to have practical significance for isolating a substantial proportion of high-risk soldiers for targeted interventions that could help reduce prevalence of homelessness after leaving active duty.