![Sample our Health and Social Care journals, sign in here to start your FREE access for 14 days]({"type":"image" , "src" : "/sda/5938/TOC-Subject-Sample-2021-Health-Social-Care.jpg"})
Abstract
Filtering of housing units—the process through which housing units over time serve occupants with lower or higher incomes—is a primary source of low-cost housing supply in the United States. However, the extent of filtering can vary widely in response to local housing market conditions, and this variation carries implications for the affordable housing strategies used in different areas. This paper uses the American Housing Survey (AHS) panels for 1985–2013 and 2015–2021 to construct repeat income measures of filtering. The analyses then describe the extent of variation in filtering outcomes across time periods, price points, and metropolitan areas. The results show significant variation across all three domains. Temporal analyses document significant changes in the estimated filtering outcomes across multiple time periods. In particular, the estimates for 2015–2021 suggest that downward filtering of housing units stalled or reversed as housing markets tightened in recent years. The direction and extent of filtering is also shown to vary significantly across metropolitan areas with higher and lower home price appreciation. These findings highlight the importance of heterogeneity in filtering outcomes to the conclusions drawn for policy. They also provide insight into the potential limitations of filtering as a source of affordable housing supply.
Disclosure Statement
No potential conflict of interest was reported by the author(s).
Notes
1 Refer to Hankinson (Citation2018) for survey evidence measuring voter perceptions of the tradeoffs between housing development and affordability in San Francisco following the 2015 election, which included multiple propositions related to housing policy.
2 Refer to Baer and Williamson (Citation1988) for a more extensive review of the early filtering literature.
3 Refer to Mast (Citation2023) for a recent example of an analysis describing the chain of moves resulting from new additions to housing supply.
4 This method is similar to the approach taken in the Rental Market Dynamics report series produced by the US Department of Housing and Urban Development. Refer to, e.g., Eggers and Moumen (Citation2020).
5 Refer also to Somerville and Holmes (Citation2001), who empirically examine filtering at the housing unit level, identifying determinants of whether a unit is likely to filter downward or upward over time. While several housing unit attributes are predictive of filtering, the authors conclude that neighborhood conditions have a stronger influence on filtering outcomes.
6 An important limitation of using longitudinal panels of this length is that analyses are not able to account for all changes over time in survey data collection, editing, imputation, and other processing steps. For more information about historical changes in the AHS panels, refer to the AHS Historical Changes documents.
7 The estimated filtering rates are not highly sensitive to this choice. For more information about AHS weights, refer to the AHS’s source and accuracy document (US Census Bureau & US Department of Housing and Urban Development, Citation2022).
8 An important distinction is that the analyses in this paper stop short of constructing an index based on the age of the unit. Instead, the repeat income approach is used to construct measures of filtering outcomes, and the analyses then examine heterogeneity in survey-based estimates of these measures.
9 This issue is similar to the potential for selection bias in repeat-sales home price indices (Gatzlaff & Haurin, Citation1997).
10 This definition relies initially on the samehh (1985–2013) and samehhld (2015–2021) variables, which identify units in which some or all household members changed since the prior occupied interview. Additionally, it uses the move-in year variables to require that no household member reports moving in prior to the year of the last occupied interview. This definition excludes units in which some household members turn over in multiple years such that all household members are eventually different. It may also slightly underestimate the number of turnovers to the extent that responses or imputations produce measurement error in the move-in date variable and households are excluded based on the earliest move-in date.
11 This approach allows the turnover pair to reflect a slightly longer duration for units that turn over frequently. It also mitigates the potential for selection bias due to the inclusion of multiple turnover pairs from such units. While the first–last approach is preferred for these reasons, comparison of the first–last approach to the use of all turnover pairs shows relatively small differences in filtering estimates. For example, the mean change in log income for all 87,000 turnover pairs in the 1985–2013 panel is −.032, compared to −.037 using the 35,000 turnover pairs produced by the first–last approach.
12 For example, shows that 26 percent of turnover pairs include household incomes that differ by a multiple of 3 or more. These large changes are most common when one of the households has a very low income, which allows modest differences in income amounts to produce large values when measured as a percent change. The paper’s primary findings are robust to the exclusion of these cases. For example, excluding turnover pairs where either arriving occupant household has income below $20,000 reduces the share of turnover pairs with incomes that differ by a multiple of 3 or more from 26 percent to 13 percent—and it reduces the share of turnover pairs where the absolute value of the mean log change in household income is larger than 5 from 2.8 percent to 0.4 percent. Using this reduced sample, the estimates in for the mean change in log income are robust in sign and significance, continuing to show significant temporal heterogeneity. Similarly, the estimates for the high-appreciation West category in are robust in sign and significance, continuing to show significant geographic heterogeneity in filtering outcomes.
13 Changes in the cost-burden ratio are topcoded at 100 percent to reduce the influence of outlier values.
14 Sample sizes throughout the paper are rounded in accordance with the Census Bureau’s disclosure review board (DRB) requirements.
15 Because this variable is logged, the 5-year percent change in income must be calculated by applying the exponential function to this value and subtracting 1. In this case, the estimate of −.013 corresponds with a change of e(−.013) – 1 = −.013, which is a decrease of 1.3 percent. Similarly, the estimate of −.330 for own-to-rent transitions corresponds with a change of e(−.330) – 1 = −.281, which is a decrease of 28.1 percent.
16 One possible explanation for this finding might be that changes in the number of tenure transitions across decades contribute to the differences in the filtering rates. However, this hypothesis is not borne out in decomposition regressions. For example, the simple difference between the mean change in log income estimates for 2005–2013 and 2015–2021 is .154, compared to .142 after controlling for tenure transitions and other housing unit characteristics. This result suggests that the changes in filtering outcomes across decades are instead driven primarily by differences in filtering rates among units making the same types of tenure transitions.
17 The exception is that the mean change in log income estimate for own-to-own transitions for 2005–2013 is not significantly different than the prior decade, although the direction of the change in filtering speeds is consistent with the change observed for rent-to-rent transitions.
18 Within each of the overlapping five-wave panels, the process of identifying turnover pairs and constructing the filtering measures is repeated using the process described in the Filtering Measures section.
19 The housing unit counts for the decennial censuses are drawn from Social Explorer’s county-level data, which uses reallocation fractions from the longitudinal tract database to account for any changes in county boundaries across decades. The county-level counts are then aggregated into CBSAs using the 2010 boundaries consistently for all decades.
20 This restriction removes approximately 7 percent of the sample.
21 The latitude and longitude associated with the central business district of the primary city in each CBSA come from Fee and Hartley (Citation2013). Refer to Holian (Citation2019) for discussion of the tradeoffs between alternative methods for determining the location of city centers.
22 The reduction in sample occurs because the housing cost measure does not use values from the same survey wave as either turnover in a pair. The rationale is that measurement error in the income and housing cost measures within the same survey wave may be correlated due to the imputation process. Because the AHS does not use longitudinal imputation, using values from other survey waves eliminates this concern.
23 Housing cost is defined as the total monthly housing cost including utilities. For large CBSAs, units are rank ordered by housing cost within each CBSA and then separated into quartiles with equal sample sizes. For small CBSAs, this process is repeated using the CBSA categories in .
Additional information
Notes on contributors
Jonathan Spader
Jonathan Spader is manager of the National Mortgage Database at the Federal Housing Finance Agency. He primarily completed this paper while working in the Social, Economic, and Housing Statistics Division of the US Census Bureau.