How Public Health Studies Tackle the Homelessness-Drug Crisis

A typical exam passage on a public health study of homelessness and the drug crisis will not ask whether homelessness and substance use are serious problems. It will ask what the study design permits the researchers to conclude. If the passage describes a one-time survey, the answer cannot be a clean causal claim. If it describes death records linked to homelessness counts, the answer may involve mortality rates but not individual life histories. If it describes a randomized intervention, the causal claim may be stronger, but narrower.

That is the main reading problem: public health researchers do not study homelessness and drug use with one master method. They use representative surveys, administrative data linkage, observational cohorts, scoping reviews, and randomized controlled trials. Each design sees a different part of the road. The skill for MCAT and GRE readers is not to rank them by prestige, but to match the design to the question and then name the validity threat that remains.

Start with the question the design can actually answer

The California Statewide Study of People Experiencing Homelessness, usually shortened to CASPEH, is the cleanest working example because it shows how much method is hidden inside a phrase like “people experiencing homelessness.” The study surveyed 3,198 adults across eight California counties and combined survey data with in-depth interviews. UCSF’s Benioff Homelessness and Housing Initiative describes it as the largest representative study of homelessness in the United States since the mid-1990s.[1]

For an exam reader, that sentence contains several moving parts. “Representative” is not decorative. It means the investigators were trying to make claims about a broader population, not merely about the people easiest to find. “Across eight California counties” gives the study geographic reach inside one state, but it does not turn California evidence into national evidence. “Survey data plus in-depth interviews” means the study can estimate patterns and also describe experiences that a checkbox might flatten.

CASPEH is especially useful because homelessness is not a stable, easily sampled status. A housed person can be reached through an address-based survey. An unsheltered person may be sleeping outdoors, in a vehicle, in a shelter, or in a temporary arrangement that changes before a study team can classify it. The sampling problem is not a side note; it shapes who becomes visible in the data.

The same is true for measurement. Substance use, mental health symptoms, housing history, income, and service use are not all measured with the same reliability. Some are self-reported. Some may be remembered imperfectly. Some may be underreported because the respondent fears stigma, punishment, or loss of services. A strong passage answer would not simply say “self-report is biased” and stop there. It would ask which variable depends on self-report and whether that weakness affects the claim being made.

What CASPEH can show—and what it cannot

CASPEH can describe the characteristics, experiences, and reported needs of adults experiencing homelessness in California during the study period. It can compare subgroups. It can identify commonly reported barriers to housing. It can put behavioral health into the same frame as income, rent, services, and institutional contact. Those are substantial strengths.

It cannot, by itself, prove that substance use caused homelessness, or that homelessness caused substance use. A cross-sectional design observes variables at a point or period close enough in time that temporality becomes difficult. If a respondent reports both homelessness and substance use, the study may show co-occurrence, association, or reported sequence depending on the exact question. It does not automatically establish a causal pathway.

This distinction matters because public discussion often reaches too quickly for a single cause. CASPEH complicates that reflex. Among participants who wanted housing, 89% reported housing costs as a barrier, and 90% reported that there was no housing available that they could afford.[1] Those figures do not erase behavioral health needs. They do make it harder to defend the simpler claim that substance use or mental illness is the primary explanation for homelessness as a whole.

Prevalence figures can help set context, but they should not be asked to carry more than they can. KFF reported that 22% of adults experiencing homelessness had serious mental illness and 18% had substance use disorder, compared with about 5% to 6% and 3%, respectively, in the general population.[2] Those are important disparities. They are not, on their own, a causal model.

Administrative linkage sees deaths that surveys may miss

A mortality linkage report starts from a different problem. Instead of asking living respondents to report experiences, it uses administrative records to identify deaths and connect them to a population denominator. Los Angeles County’s 2025 Homeless Mortality Report used Medical Examiner data and state death records, then linked those deaths to point-in-time count data. The report identified 2,508 deaths among people experiencing homelessness in 2023 and a mortality rate of 3,326 per 100,000.[3]

That design has a blunt strength: death is an outcome surveys can miss, especially when the people most at risk are also the hardest to follow. Administrative linkage can reveal outcomes that would be difficult to capture through interviews alone. It can also reduce some forms of recall bias because it does not depend on a respondent remembering and reporting an event.

But a death record is not a biography. It can record cause and circumstance more systematically than rumor or anecdote, but it cannot fully explain the sequence of eviction, shelter access, drug use, illness, policing, hospital contact, and service gaps that preceded death. The denominator matters too. If the point-in-time count undercounts unsheltered people, the mortality rate can be affected. An exam answer should therefore treat the report as strong evidence about recorded mortality, not as a complete causal explanation for why each death occurred.

Reviews summarize a field, not a single clean experiment

A scoping review answers another kind of question: what has the literature, as a whole, studied and found? Lin and colleagues reviewed 50 studies on housing instability and injection drug use. The review reported that unstable housing was associated with 44% longer time to cessation of injection drug use and about 50% faster relapse to injection drug use.[4]

Those findings are useful because they do not depend on one city, one service site, or one sample. A review can show whether a pattern appears repeatedly across different research settings. It can also expose gaps: inconsistent definitions of homelessness, uneven measurement of drug use, and differences in how studies follow participants over time.

The limitation is built into the method. A scoping review does not magically upgrade weak primary studies into strong causal evidence. If the included studies define unstable housing differently, measure drug use differently, or lose participants to follow-up at different rates, the review must work across that uneven terrain. On exams, the tempting wrong answer is often the one that treats “review” as if it means “settled causal proof.”

Adjusted odds ratios still need a careful verb

Observational studies often sit between descriptive surveys and experiments. They can adjust for measured covariates and estimate associations, but they usually cannot eliminate all confounding. Milaney and colleagues studied 432 people and used logistic regression to examine hospital care. Unstably housed individuals were more likely to require hospital care, with an odds ratio of 2.04 and a 95% confidence interval from 1.29 to 3.21.[5]

The exam-useful reading is precise: the study found an adjusted association between unstable housing and hospital care. The confidence interval does not include 1, which is consistent with a statistically significant association under conventional interpretation. But “twice as likely” should not be quietly converted into “unstable housing caused the hospital care” unless the design and analysis justify that causal step.

The remaining threat is confounding. People who are unstably housed may differ from stably housed people in severity of illness, access to outpatient care, exposure to violence, income, social support, or other factors that are hard to measure fully. Regression can adjust for variables included in the model. It cannot adjust for what the researchers did not measure or measured poorly.

Why randomization is powerful, and why it is not the whole field

A randomized controlled trial is the cleanest design when the question is whether a specific intervention changes a specific outcome for an eligible group. In a 2004 Housing First randomized controlled trial, 79% of participants in the Housing First group retained housing at six months, compared with 27% in the treatment-first group.[6]

Randomization matters because, in principle, it distributes both measured and unmeasured characteristics across groups at baseline. If the groups are comparable and the trial is well conducted, a later difference in housing retention can more plausibly be attributed to the assigned intervention. That is a stronger causal structure than a cross-sectional comparison of people who happened to receive different services.

The tradeoff is scope. An RCT of Housing First is not designed to estimate the prevalence of substance use disorder among all people experiencing homelessness. It is not designed to count deaths across a county. It may not include the people most disconnected from services, depending on eligibility and recruitment. Its causal strength comes from narrowing the question.

The hardest person to count is often the person most important to the conclusion

In passages on homelessness, sampling bias deserves more attention than it usually gets. People staying in shelters may be easier to locate than people living outdoors. People already connected to clinics or service agencies may be easier to recruit than those avoiding institutions. People cycling between a friend’s couch, a motel, a car, and the street may not fit neatly into a single housing category.

That is why operational definitions matter. “Homelessness” can mean unsheltered homelessness, shelter use, literal homelessness, unstable housing, or broader housing insecurity, depending on the study. “Drug use” can mean any use, injection use, opioid use, substance use disorder, overdose, or treatment history. If two studies use different definitions, their results may not be directly comparable even when their abstracts sound similar.

Older or narrower statistics should be handled with the same caution. A dated city-perception statistic about homelessness and substance use is not the same as direct measurement in a representative sample. A single community sample showing that drug use preceded homelessness for many participants can be useful for a specific subpopulation, but it should not be stretched into a universal causal sequence. The direction can vary: substance use may increase housing instability for some people, and homelessness may worsen or initiate substance use for others.

How to read these studies in MCAT and GRE terms

When an exam passage presents a public health study on homelessness and drug use, first identify the unit of evidence. Is the passage describing individuals surveyed once, individuals followed over time, records linked after the fact, studies summarized in a review, or participants assigned to an intervention? That one move prevents many overclaims.

• If the design is cross-sectional, look for prevalence, association, sampling, operational definitions, and self-report bias. Be cautious about temporality.
• If the design uses administrative linkage, ask what records were linked, how the population denominator was defined, and whether the data explain mechanisms or mainly document outcomes.
• If the design is observational with regression, read adjusted odds ratios as adjusted associations unless the passage gives stronger causal support.
• If the design is a scoping review, ask whether the included studies used comparable definitions, samples, and outcomes.
• If the design is an RCT, identify the intervention, comparison group, outcome, follow-up period, and eligible population before generalizing.

This is the same evidence-evaluation habit students need when reading other major scientific reports. A companion example is our guide to the National Academies climate report for students, where the task is to separate an institution’s conclusion from the types of evidence supporting it.

The practical rule is simple enough to remember and demanding enough to matter: do not rank study designs abstractly. Ask what question the researchers asked, whether the design can answer that question, and which validity threat would make the conclusion too strong. The homelessness-drug crisis is urgent, but urgency does not change what cross-sectional surveys, mortality linkages, observational models, reviews, and randomized trials can bear.

References

1. California Statewide Study of People Experiencing Homelessness, UCSF Benioff Homelessness and Housing Initiative, 2023.
2. Five Key Facts About People Experiencing Homelessness, KFF, 2025.
3. Los Angeles County Releases 2025 Homeless Mortality Report, Los Angeles County, March 6, 2025.
4. Housing instability and injection drug use cessation and relapse: A scoping review, PMC, 2024.
5. Drug use, homelessness and health: responding to the opioid overdose crisis with housing and harm reduction services, PMC, 2021.
6. Housing First, Johns Hopkins Bloomberg School of Public Health.