Abundant Housing LA researched the relationship between homelessness rates and measures of the housing market across seven metropolitan regions. We focused on the average rent burdens (percentage of household income spent on rent) incurred by households at various percentiles of household income, as well as a measure of housing unit scarcity. We found that while rent burdens can be predictive of growth in the number of homeless residents in a region, housing unit scarcity is not a statistically significant predictor. Furthermore, we find that the predictive power of rent burdens only exists within the three California metropolitan areas studied (Greater Los Angeles, San Diego, and San Francisco).
While these results are based on a fairly limited amount of data, we observe remarkably strong fit in several California cases, bolstering our confidence.
This technical addendum supplements the AHLA blog post available here.
Homelessness counts were obtained from the Department of Housing and Urban Development (HUD) which publishes the point-in-time counts annually submitted by regional Continuums of Care. Data on populations, household incomes, and available market units were sourced from the American Community Survey, a Census Bureau program that collects granular data from a sample of households every year. Nominal dollar amounts were converted to real amounts using local CPI and CPI-housing deflators produced by the Bureau of Labor Statistics. The combined dataset used for the analysis is available in the AHLA Github.
In our analysis, we pair the independent variables with the homelessness count for the following year. We do this because the count is conducted in late January each year, meaning that it will reflect the results of changes that occur during the previous year rather than during the same year, which will have only just begun. For Los Angeles, San Francisco, and San Diego, we exclude homeless counts taken before 2011 because of issues with the apparent accuracy of the counts.
An extensive academic and anecdotal literature explains that the point-in-time counts conducted each January by HUD Continuums of Care (CoC) persistently undercount the true number of homeless by a significant amount. For the sake of the analysis, we assume that the undercounting happens in a systematic fashion, such that the trends are accurate, even if the point estimates are not themselves accurate.
We estimated linear regressions for two related dependent variables: (1) growth in the number of homeless, and (2) change in the homeless rate.
In our first set of regressions, we regress the dependent variables against the changes in the rent burdens faced by the 10th, 20th or 50th percentile. We also test whether the inclusion of the housing unit scarcity variable affects the parameter estimate on the rent burden variable. All of the specifications include fixed effects for year and continuum of care.
In the set of regressions that include all 7 CoCs, changes in the 50th percentile rent burden appear predictive of growth in homelessness. The estimated coefficient on the change in the median rent burden (2.753) is significant at the 95% confidence level and seems fairly insensitive to the addition of the scarcity variable. (Compare specifications 3 and 6 in the table above—the numbers in parentheses are t-values, and an asterisk notes statistical significance at the 95% level.) However, the significance of this estimate seems to be driven by San Francisco and Los Angeles. When we re-run the regressions on a subset of the data that excludes those two cities, the coefficient on the change in median rent burden is diminished from 2.7 to 0.9, and there is no statistical significance. On the other hand, when we subset the data to only include LA and San Francisco, we see a much larger coefficient (3.9) that is statistically significant. We also recover a significant coefficient of 2.1 on the change in rent burden at the 20th percentile.
Examining only the trend in Los Angeles, we observe the highly significant coefficient on the 10th percentile rent burden of 3.219, suggesting that each percentage point increase in the rent burden predicts 3.2% additional growth in the homeless count the following January. This is the headline result in the blog post.
In San Francisco, the 10th percentile rent burden is not a statistically significant predictor, but the 20th percentile rent burden has a significant coefficient of 1.9.
We were unable to identify a relationship between homelessness and housing unit scarcity or a change in scarcity.
In addition to revealing a predictive relationship between rent burdens and homelessness, our analysis raises the question of why many cities outside of California seem to have uncoupled changes in rent burdens from changes in homelessness. This may be the result of effective programs or dynamics that prevent homelessness inflows in the first place despite rising rents, or it could be the result of more effective rehousing programs that have enough slack to absorb people who become unhoused. Alternatively, many homeless people in non-California cities may not be properly counted because they double-up with other households.
Assuming that the variation is not an issue of data quality, we attribute it largely to differences in housing and homeless policy across US cities. For example, while cities in California tend to focus efforts on developing affordable housing, New York focuses muchall of their efforts on providing homeless shelters. Of the 56,000 homeless people in Los Angeles in 2019, 42,000 were unsheltered, while only 3,600 out of 78,000 homeless people in New York are unsheltered.
Miami-Dade County experienced a decrease in homelessness between 2016 and 2019, largely due to a focus on constructing permanent supportive housing, although the county still needs 130,000 housing units to cover its gap. A study by researchers David Lee and Michael McGuire found that County governments that align homelessness programs with federal and state government policies have better results in homelessness reduction and affordable housing production than those that do not. It is, however, unlikely that the same policy plan will work for every city, because homelessness is caused by a diverse array of factors, including access to affordable housing, overall poverty rate, and proximity to public transit, depending on the city or geographic region.
There are a number of ways to expand this analysis for greater robustness. These include:
- Devising a more accurate measure of the homeless population. Glynn et al. address the uncertainty in the point-in-time count by constructing a Bayesian nonparametric model to estimate homelessness rates in different CoCs.
- Measuring homelessness inflows rather than a point-in-time count. By estimating against the point-in-time count, we unintentionally capture the effect of government programs that rehouse the homeless (thus removing them from the PIT count) but do not bear on our predictor variables. Therefore, the effect of rent burden on homelessness that we recover is muddled by how good or bad a job the government happens to do in a given year at rehousing the homeless. While we struggled to find consistent annual data on this topic, adding the number of people rehoused to the point-in-time count would more accurately document changes in homelessness inflows.
- Collecting data on additional continuums of cares nationwide, beyond those analyzed in this model.