567,000+ Persons Experiencing Homelessness, and Counting

The Department of Housing and Urban Development (HUD) issued a press release on December 20 stating that homelessness in the United States increased by 2.9 percent in 2019. The press release stated:

The study found that 567,715 persons experienced homelessness on a single night in 2019, an increase of 14,885 people since 2018. Meanwhile, homelessness among veterans and families with children continued to fall, declining 2.1 percent and 4.8 percent, respectively, in 2019.

Where do these statistics come from, how accurate are they, and what do they mean?

Every January, state and local agencies responsible for coordinating homeless services (called Continuums of Care, CoCs) conduct a point-in-time count. Each CoC participating in the count (CoCs applying for grant funding are required to participate) chooses one night from the last ten nights in January* to conduct the count. They attempt to gather a snapshot of how many persons are experiencing homelessness at that particular point in time. In most localities, the count is conducted by teams of volunteers.

Who is included in the count? HUD guidelines distinguish between sheltered and unsheltered populations:

Sheltered population: Persons in emergency shelters, transitional housing, or “safe havens” (providing temporary shelter and services for persons with special needs such as mental illness or substance abuse).
Unsheltered population: Persons staying in a place not intended for human habitation, such as on the streets, under bridges, or in cars, tents, bus stations, or abandoned buildings.

HUD has not yet issued the full report for 2019 with more detailed breakdowns, but the 2019 spreadsheet, available on the HUD website, lists 356,422 persons in shelters and 211,293 persons not in shelters. The corresponding statistics from 2018 were 358,363 persons in shelters and 194,467 persons not in shelters. All of the increase reported for 2019 was in the unsheltered population, the population for which it is most difficult to obtain an accurate count.

Some CoCs attempt to count all persons experiencing homelessness in their area, while others may use counts obtained in a sample of subareas to estimate the total number.** Either procedure, though, results in an estimate, not an exact count. In fact, CoCs that take a statistically representative sample may well end up with a more accurate estimate of the number of persons experiencing homelessness than CoCs that attempt to cover the entire area, because it may be possible to obtain a more comprehensive listing if the volunteers can concentrate on smaller areas.

Regardless of method, no one believes that the estimate from the point-in-time count captures every person experiencing homelessness on the day of the count (and of course it does not capture persons who experience homelessness at some other point in the year, or those in areas that do not participate in the January count). It thus always underestimates the extent of homelessness in a community. If the same methodology is used every year, however, the estimates can be used to evaluate changes in homelessness over time, even though they underestimate the number of persons experiencing homelessness every individual year.***

Any comparison, however, needs to be done carefully, and with awareness of some of the other factors that affect the statistics. These include:

Resources available. A CoC with few volunteers may have more of an undercount than one with many volunteers (although the CoC with few resources could still improve its count by using sampling methods to strategically deploy those resources). Point-in-time counts are used to allocate funding, which may incentivize some communities to expend more effort on the count than other communities.
Procedures used to count people. Although HUD specifies general guidelines for conducting the count, there is a lot of variation in how individual CoCs carry it out. City A might have a higher estimated rate of homelessness than City B because City A does a better job of finding people during the count, rather than because it actually has more homelessness.
Some CoCs attempt to canvass the entire area; others canvass only, or primarily, locations known to be frequented by persons experiencing homelessness. And volunteers are usually instructed to avoid locations that might pose a safety risk, such as abandoned buildings.
Judgments about which persons encountered are experiencing homelessness. Some volunteers may decide, without approaching a person, that he or she “does not look homeless.”
Technology. Until 2018, volunteers conducting the count in Maricopa County in Arizona used paper forms to record the counts of temporary shelters and persons, as well as the information about demographic, health, veteran status and other experiences with homelessness provided by persons who were willing to answer survey questions. In 2019, however, Maricopa County switched from paper forms to a cell phone app, which may have affected the count (by making it easier to record data) or the quality of the data (the app ensures that volunteers do not forget to ask some of the questions).
Changes in a community that might shift the distribution of persons experiencing homelessness from sheltered to unsheltered, or vice versa. The point-in-time count of the persons in shelters is easier to obtain, and more accurate, than the count of persons experiencing homelessness who are not in shelters that night. If a shelter closed in a city during 2018 but the total number of people experiencing homelessness was unchanged, the 2019 point-in-time count would likely be smaller than the 2018 count because in 2019 a higher proportion of the persons experiencing homelessness were in the harder-to-count unsheltered population.
Other events affecting the count in one particular January. The federal government shutdown from December 22, 2018 through January 25, 2019 had a disproportionate effect on persons experiencing homelessness and those who were one financial setback away from being homeless. Social services provided by federal agencies (such as low-income rent assistance and block grants to organizations that help the poor) were reduced, delayed, or suspended. One needs to look at data from other months to see the effect of events such as the shutdown.

Some cities go far beyond the January point-in-time count in their data collection. The Christian Science Monitor recently described the Houston area’s Homeless Management Information System, a database of individuals receiving services throughout the year. Houston has used the point-in-time count and database to assess the effects of its “housing first” approach, where homeless families and individuals are first placed in stable housing and then provided assistance with other needs.**** New York City publishes a daily report of how many families and individuals used shelter services in the city, and outreach teams work year-round to identify, engage, and assist unsheltered New Yorkers—and, in the process, collect data that can be used to evaluate needs and progress.

How accurate are the statistics? The point-in-time count underestimates the number of persons experiencing homelessness in individual communities and for the United States as a whole, and the underestimation is likely greater for the unsheltered population than the sheltered population.***** But it’s difficult to tell how large the error is, because HUD does not publish standard errors or other measures of accuracy for the statistics. Because different communities use different procedures to conduct the count, one must be wary of using the data to compare or rank different areas. But the data say one thing clearly: in January 2019 there were at least half a million persons in the United States without a home.

Update on January 9, 2020: HUD has now released the full 2019 report.

Footnotes

*Why the last ten days in January? In many areas of the country, January is a harsh time to be homeless. The count is conducted on one of the last ten days of the month in order to capture people who may be able to pay for temporary shelter at the beginning of the month but run out of money by the end. Conducting the survey on one day reduces the chance that a person is double-counted.

**Some CoCs cover such a large geographic area that they must necessarily restrict the point-in-time count to a sample of areas. Delaware, Maine, Montana, North Dakota, Rhode Island, South Dakota, and Wyoming each have one CoC that covers the entire state. Most other states have at least one geographically large CoC that covers the parts of the state that are not in one of the urban CoCs. Arizona, for example, has three CoCs: Maricopa County (home of Phoenix), Pima County (home of Tucson), and “Balance of State” (the other 13 counties). “Balance of state” CoCs must of necessity take a sample of areas. Not all of them take a statistically representative sample, however, and not all of them publish their methods for taking the count.

The HUD guidelines state: “It is preferable for CoCs to conduct a census count when practicable, as it is by definition the most complete and accurate information available.” But a census is only the most accurate information when it actually counts everyone. A well-designed and carefully executed sample will produce better quality estimates than an incomplete census.

New York City, for example, takes a carefully designed probability sample to carry out its point-in-time count. Schneider et al. (2016) described the procedure. Each of the approximately 7,000 census block groups in the city is classified as “high-density” (deemed likely to contain unsheltered persons, based on previous counts and experiences of outreach workers) or “low-density.” Volunteers are sent to each area in the “high-density” stratum and to a random sample of the areas in the “low-density” stratum. This design makes more efficient use of the volunteers by deploying them to areas thought to contain persons to include in the survey, yet still produces a statistically valid estimate because results can be extrapolated from the sampled areas in the “low-density” stratum to the unsampled areas in that stratum.

***If the amount of underestimation is the same every year, then it cancels out when one subtracts the estimated number of persons experiencing homelessness in 2018 from the estimated number of persons experiencing homelessness in 2019.

****Houston’s 2019 point-in-time count found 3,938 persons, compared with the peak of 8,471 persons in 2011. Of course this comparison does not prove that the policies caused the decrease in homelessness—one needs to conduct a randomized experiment to establish causation—but Houston did report a sharper decrease than many other cities over the same time period.

*****Some cities use statistical methods to estimate how many unsheltered persons are missed in the January count. In New York City, researchers have estimated how many persons may have been missed by the volunteers by placing “decoys”—persons paid to be outside on the night of the count—at various locations in the city. They then estimated the percentage of unsheltered persons missed in the count by the percentage of decoys who were missed in the count. Such studies also give information for improving the count procedures in future years.

Other data sources may provide insight about the extent of underestimation. For example, the 2018 point-in-time survey for Los Angeles County reported approximately 5,000 children under 18 who were homeless, but in November 2018 the Los Angeles Unified School District reported more than 16,000 students “who have been identified as homeless, living in shelters, motels, abandoned buildings, cars, doubled up with other families, or unsheltered.” The LAUSD uses a broader definition of homelessness than HUD (also including students living in motels and doubled up with other families) and covers a longer period of time, and so it would be expected to have a larger count. But the large difference between the two numbers suggests that the point-in-time count misses a substantial number of people.

References

Schneider, M., Brisson, D., and Burnes, D. (2016). Do We Really Know How Many Are Homeless?: An Analysis of the Point-In-Time Homelessness Count. Families in Society: The Journal of Contemporary Social Services, 97, 321-329.

sample surveysSharon LohrDecember 29, 2019