Data sources
- HUD Point-in-Time (PIT) Counts, 2007–2024
- The numerator for all rate calculations. HUD's annual count of sheltered and unsheltered persons experiencing homelessness on a single night in late January, reported by each Continuum of Care. Source: HUD User AHAR Reports.
- UCSF BHHI CoC Data Project, 2011–2019
- Provides the population denominator for each CoC, derived from American Community Survey 5-year estimates and mapped to CoC geography via census tracts. Also includes covariates used in our drivers analysis: poverty rate, rent burden, vacancy rate, unemployment, eviction rates, etc. Source: github.com/ucsf-bhhi/coc-data.
- HUD CoC Program Awards, 2018–2024
- Federal funding allocated to each Continuum of Care under the CoC Program. Used to compute funding per capita and funding per homeless person. Source: HUD Exchange.
- HUD Housing Inventory Counts (HIC), 2013–2024
- Bed inventory data: emergency shelter, transitional housing, and permanent supportive housing beds available in each CoC. Used as a capacity covariate.
The data quality framework
Every CoC in our analyses carries a data-quality tier. This isn't optional. Communities that count well can appear to have higher homelessness rates than communities that count poorly. A "low" rate in a poorly-funded rural CoC may reflect undercounting rather than actual housing stability.
The undercount risk score awards points for each red flag:
| Signal | Points |
|---|---|
| Avg sheltered share > 95% (last 4 years), without right-to-shelter law | 2 |
| Avg sheltered share 90–95%, without right-to-shelter law | 1 |
| HUD-classified rural CoC | 1 |
| Year-over-year volatility > 40% standard deviation in % change | 1 |
| Federal funding under $1,000 per homeless person | 1 |
0 points: Higher confidence · 1–2 points: Possible undercount · 3+ points: Likely undercounting
Right-to-shelter jurisdictions (NYC, all Massachusetts CoCs, DC) are exempted from the sheltered-share flag because their high sheltered share reflects legal entitlement to shelter, not undercounting of unsheltered persons.
Rate calculations
Rate per 10,000 = (PIT count ÷ CoC population) × 10,000. We display rates in four equivalent formats: per 10,000, percentage, per 1,000, and "1 in N residents." No single format is right for every audience. "Per 10,000" is the public-health convention for rare events. Percentages are intuitive but compress the variation. "1 in N" is most plain-English.
The CoC population denominator is held constant per CoC at the most recent ACS estimate. Annual variation in our reported rates therefore reflects changes in the homeless count, not in population. A rigorous time-varying denominator would require pulling fresh ACS each year (Census API key territory).
Known caveats
- PIT counts undercount. Our rates are floors, not true prevalence. The 2020 GAO report to Congress documented this.
- The 2021 dip is not a real decline. HUD waived unsheltered counting requirements during COVID. Many CoCs partially counted or skipped entirely. Treat 2021 as noise.
- Territories lack ACS coverage. Puerto Rico, Guam, US Virgin Islands, and Northern Mariana Islands appear in PIT data but cannot have rates computed because we don't have population denominators.
- CoC boundaries change occasionally. Mergers and splits introduce small distortions in trend lines for affected CoCs.
- Correlations are not causation. The drivers analysis identifies statistical association; it does not establish that variable X causes Y. Funding-per-capita correlates positively with homelessness rate because HUD allocates by need, not because money causes homelessness.
Reproducibility
Source datasets are saved locally in the project's /datasets/ folder. Processing scripts,
rate calculations, and dashboard generation are tracked in git at
github.com/GaitherStephens/gaither-research.
If you reproduce a result and find it differs from ours, open an issue on GitHub or email
[email protected].