Abstract
Changes in household composition are detrimental to children’s well-being and outcomes. Unaffordable or unstable housing may lead to changes in household composition. The author uses data from the 1995 through 2015 waves of the Panel Study of Income Dynamics and the linked Assisted Housing Database to estimate the effect of receipt of project- or tenant-based housing assistance on changes in household composition. The author uses matching methods to compare changes in household composition among 182 surveyed households that received housing assistance between 1997 and 2009 versus 1,549 households that did not receive assistance. In general, households receiving assisted housing have a significantly lower likelihood of experiencing changes in household composition in the six years after receipt. Providing material resources through assisted housing is one way in which policymakers could feasibly intervene to encourage housing and household stability, with longer term benefits for individual, child, and family well-being and outcomes.
Keywords
Instability in household composition has negative consequences for children’s well-being and outcomes, on average (Cavanagh and Fomby 2019; Mollborn, Fomby, and Dennis 2012; Perkins 2019, 2023, 2024). Household composition changes involving children’s parents and their partners are associated with worse social and emotional behavior and well-being among children, lower academic performance and cognition, earlier sex and union formation, and higher rates of fertility among adolescents and young adults (Cavanagh and Fomby 2019). Much research has focused on family instability involving parents and their partners, yet broader household instability, including changes in household composition involving extended family and nonrelatives, is also associated with worse outcomes among children. Very young white and Black children who experience changes in household composition involving nonparents score lower on cognitive tests than children in stable household arrangements (Mollborn et al. 2012). Changes in household composition involving extended family and nonrelatives are consequential for longer term outcomes as well: experiencing these changes during childhood or adolescence predict lower probability of high school graduation and college enrollment (Perkins 2019, 2024) and, among girls, higher probability of teen childbearing (Perkins 2023).
Unaffordable or unstable housing is one factor that may lead to changes in household composition. Rising housing costs and a supply of subsidized housing that falls far short of demand mean that more than 40 million families in the United States are housing cost burdened (Joint Center for Housing Studies 2024). In an environment in which affordable housing is scarce, families may double up with other families to ease housing cost burdens, or move in with extended family members or friends to find stability (Harvey and Dunifon 2023; Pilkauskas 2012). Once formed, however, these shared households tend to dissolve relatively quickly (Pilkauskas 2012), leading to more exposure to changes in household composition. Subsidized housing is designed to provide affordable and stable housing to families and individuals in need (HUD 2025), and does in fact result in more affordability (Cai 2024; Gold 2020) and less housing instability compared with low-income unsubsidized households (Kang 2021; Lundberg et al. 2021). Does the increased housing stability provided by subsidized housing result in fewer changes in household composition for families that receive it?
I use administrative records of housing assistance linked to the Panel Study of Income Dynamics (PSID) to estimate the effect of receipt of two types of housing assistance on changes in household composition. Receiving housing assistance may result in immediate changes in household composition as doubled up and extended family households dissolve to establish independent households, but may ultimately result in more stability in household composition if housing assistance enables households to maintain housing affordability and stability. Eligibility for and receipt of assisted housing is not evenly or randomly distributed across the population. To address selection into the receipt of housing assistance, I estimate propensity scores for the likelihood of receiving assisted housing and use these propensity scores to construct overlap weights to balance characteristics of households across treatment groups that receive assisted housing and a control group that does not receive assisted housing. I estimate the effects of housing assistance on changes in household composition separately for recipients of project- versus tenant-based housing assistance and I find that within six years after receipt of project-based assisted housing, recipient households are significantly less likely to experience changes in household composition than are similar households who do not receive assisted housing. I find fewer significant differences in likelihood of experiencing changes in household composition among recipients of tenant-based housing assistance compared with unassisted households.
On the basis of these results, I argue that receiving assisted housing can reduce changes in household composition that may be detrimental to children’s well-being. Affordability and stability are likely mechanisms connecting receipt of housing assistance and stable household composition. Admittedly, there are many other characteristics of families and households that prompt changes in household composition. The provision of affordable housing, however, is a feasible way in which policymakers can intervene to promote stability in the lives of individuals and families, with the potential for positive longer term consequences for individuals, families, and children.
Consequences and Causes of Changes in Household Composition
Changes in household composition can be disruptive and detrimental to children’s well-being and later outcomes. A long line of research finds negative effects on children of parental divorce and relationship dissolution (Amato and Cheadle 2005; Cherlin, Kiernan, and Chase-Lansdale 1995; Kim 2011), which typically involves the departure of a parent from the child’s household. Parental remarriage and repartnering that brings a new parental figure into a child’s household is also associated with worse outcomes among children, and exposure to repeated family structure transitions involving parents and their partners is particularly harmful to children’s development and well-being (Cavanagh and Fomby 2019; Cavanagh and Huston 2006; Fomby and Cherlin 2007; Fomby and Osborne 2010; Lee and McLanahan 2015; McLanahan 2011). Negative effects of changes in household composition, however, are not limited to changes involving parents (Mollborn et al. 2012; Perkins 2019), and the association between changes in household composition involving extended family and nonrelatives on girls’ teen childbearing is statistically indistinguishable from the association between changes involving parents and teen childbearing (Perkins 2023). Research estimating the association between and effects of changes in household composition on more distal outcomes have used various approaches, including counterfactual matching estimators, marginal structural models with inverse probability of treatment weights (IPTWs), fixed effects, and heterogeneous treatment effects to try to isolate the independent contribution of household changes to later outcomes (Cavanagh and Fomby 2019; Kim 2011; Perkins 2019, 2023, 2024), though all of these methods rely on strong assumptions to facilitate causal inference.
What predicts changes in household composition? Changes in household composition are more common among households with certain characteristics. Individuals living in doubled-up and multigenerational households are particularly susceptible to changes in household composition (Pilkauskas 2012), with 60 percent of multigenerational households experiencing some change in composition within one year and 90 percent experiencing at least one change over five years (Glick and Van Hook 2011). Material resources are associated with household composition and changes in these resources may prompt changes in composition. Doubling up can be a coping mechanism to account for low levels of material resources and social support (Bengtson 2001; Harvey and Dunifon 2023) such that lower income households are more likely to experience changes in household composition. Beyond socioeconomic status, aging and events in the life course, such as divorce and remarriage, and unpredictable changes in employment and health also prompt changes in living circumstances and household composition.
Very few, if any, of the life-course and family-related events that prompt household change provide opportunities for policymakers to intervene to potentially stop or reduce exposure to changes in household composition. The aim of this study is to assess whether having affordable and stable housing can reduce changes in household composition by addressing challenges associated with material resources, thereby providing a more stable environment for individuals, children, and families. Providing more resources to families who would otherwise double up with other individuals or families to make ends meet is a way in which programs and policies could potentially contribute to stability in household composition. In this article, I am specifically examining whether the additional resources provided by housing assistance contribute to household stability by reducing the frequency of changes in household composition.
Housing Assistance, Affordability, and Housing Stability
In the United States, the federal government, through the U.S. Department of Housing and Urban Development (HUD), and state and local governments have multiple programs designed to provide safe, adequate, and affordable housing to low-income families and households. These programs can be categorized into project-based assistance and tenant-based assistance. Project-based assistance is provided when income-eligible households move into a dwelling owned or whose development was subsidized by federal, state, or local governments. Project-based programs include public housing, owned by local public housing authorities (PHAs), and privately owned multifamily buildings or developments funded through project-based Section 8, Section 236, Section 221(d)(3), and the Low Income Housing Tax Credit (HUD 2025). Tenant-based assistance through the Housing Choice Voucher Program requires participants to find a suitable unit to rent in the private market and enter into a three-way contract between the participant, the landlord of the private unit, and the local PHA, which will pay a portion of the rent. In both types of housing assistance, recipients generally pay rent equal to 30 percent of their income.
If affordability and housing stability are the mechanisms connecting housing assistance to fewer changes in household composition, we must first establish that housing assistance results in affordability and stability. In 2024, the average HUD expenditure per month across programs was $1,067, suggesting that to the average recipient, federal housing assistance has a monetary value of more than $12,000 per year (HUD 2025). Assisted housing in the form of vouchers is also likely to increase participation in other public benefits (Carlson et al. 2011), which may result in more disposable income to put toward other expenses. Housing assistance is associated with substantial reductions in probability of being rent burdened and living below the poverty line, though these benefits are concentrated among households with sustained housing assistance (Cai 2024).
Assistance type matters for affordability. Moving into public housing is associated with a 21 percentage point decrease in the probability of being rent burdened the following year (Gold 2020). Descriptively, households assisted by the Housing Choice Voucher Program, however, have higher rent burdens than households living in HUD-administered public housing (Mast 2012). The mean rent burden at the end of admission year was relatively low, at 34 percent, among households that received tenant-based assistance between 2000 and 2009, but both the probability of rent burden and the proportion of households with excess burden grew over time (Mast 2014). Thus, assisted housing is associated with improved affordability, with qualifications suggesting sustained residence in public housing may more effectively address rent burden than intermittent assistance or tenant-based assistance.
In addition to affordability, assisted housing could reduce changes in household composition because it provides families with stable housing. Receiving government-assisted housing is associated with lower levels of residential mobility in a sample of low-income families in Boston, Chicago, and San Antonio (Kull, Coley, and Lynch 2016) and lower probability of being homeless in a sample of low-income New Yorkers (Cai 2024). Housing assistance is also associated with less housing insecurity, defined as involuntary moves, being behind on rent, experiencing eviction or episodes of homelessness, or moving in with others to share expenses, in a low-income sample drawn from southeastern Michigan (Kim, Burgard, and Seefeldt 2017). Among low-income New Haven residents, rental assistance is positively associated with housing stability, affordability, quality, and autonomy (Schapiro et al. 2022).
Here, too, the format of housing assistance appears to be consequential for subsequent stability. Public housing reduces eviction among low-income families with children compared with unassisted renters with children, whereas tenant-based rental assistance programs do not reduce the risk for eviction (Lundberg et al. 2021). Voucher recipients are 2.6 times more likely to experience housing instability (defined as churning residential mobility and nonprogressive moves) than public housing residents (Kang 2021). Living in public housing is associated with a lower probability of moving (Gold 2018) and longer length of residence (Beck 2019) compared with unassisted households, but there is no significant difference in probability of moving or length of residence between voucher-assisted and unassisted households in these national samples. In a Chicago sample, voucher recipients moved neither more nor less compared with applicants who did not receive a voucher (Jacob and Ludwig 2012), suggesting that vouchers did not increase housing stability.
These studies suggest that assisted housing, and public housing in particular, could reduce changes in household composition at least partially through reducing residential mobility. Residential mobility often co-occurs with changes in household composition (Desmond and Perkins 2016). Thus, studies on the association between housing assistance and residential mobility and other forms of housing insecurity suggest that housing assistance may reduce changes in household composition. If housing assistance reduces mobility and housing instability, it may as a result reduce the number of changes in household composition experienced by assisted residents.
Housing Assistance and Household Composition
Previous research on housing assistance and household composition generally finds that housing assistance and household stability are significantly associated. In a sample of Supplemental Nutrition Assistance Program and Temporary Assistance for Needy Families applicants in Wisconsin, receipt of a housing voucher predicts losing adult household members (Carlson et al. 2012). We do not know from this study who these adult household members are—whether they are partners of the head of household, extended family members, or other nonrelatives—and thus whether the receipt of a voucher is related to a relationship dissolution or to leaving a doubled-up or multigenerational arrangement and establishing an independent household. The Welfare to Work Voucher experiment provides another opportunity to assess the association between vouchers and changes in household composition; results from this experiment show significant reductions in multigenerational households and increases in single parent households as single parents leave their parents’ homes to establish independent households, but no effects on the likelihood of living with a spouse or partner (Mills et al. 2006). Vouchers are the biggest source of assisted housing and thus important to analyze as a category, yet programs providing project-based assisted housing are excluded from these studies. Research assessing affordability and stability suggest that project- and tenant-based assistance may be differently associated with changes in household composition. Research on crowding provides additional evidence supporting differential association: public housing significantly reduces the likelihood of living in an overcrowded housing unit and the number of persons per room among children because it reduces household size, whereas receipt of a housing choice voucher reduces persons per room by helping families move into bigger housing units (Zhu et al. 2025).
Housing assistance defined more broadly is positively correlated with being a single parent and negatively associated with marrying on the basis of cross-sectional and longitudinal analysis of residents of New York (Freeman 2005). This study incorporates more forms of housing assistance than just vouchers, but is limited to a correlational analysis and does not identify the relationship of other adults living in and entering and exiting the household.
Hypotheses
Prior research informs a set of hypotheses about assisted housing and changes in household composition involving extended family and nonrelatives, partners, and children. If housing assistance results in affordability and housing stability, recipients may no longer need to seek support from the private safety net offered by doubling up with extended family and nonrelatives. Individuals receiving housing assistance may be less likely to join their adult children’s households for the same reason, as older adults generally prefer to live in nonshared arrangements when they have the resources to do so (Engelhardt, Gruber, and Perry 2005). The association between housing assistance and household changes involving partners is more ambiguous: we might see fewer entrances or exits of partners because the additional resources provided by assisted housing eliminate one common motivation to cohabit (Sassler and Miller 2011), or, alternatively, increase the cost of leaving a cohabiting relationship, but we may see increased likelihood of changes involving a partner if the stability offered by assisted housing encourages the partner to move in or enables someone to leave a bad relationship. I therefore have competing hypotheses about household changes involving adults: if adult changes predominantly involve extended family and nonrelatives, assisted housing may decrease the likelihood of adults entering or exiting the household (hypothesis 1a). If, however, partners are the primary adults exiting or entering, assisted housing may instead increase the likelihood of changes involving partners (hypothesis 1b). Minor children may be less likely to leave the home of an adult receiving housing assistance if housing safety or instability are factors in whether grandparents or child welfare authorities encourage family separation (Pittman 2023). To the extent that minor children of extended family and nonrelatives would accompany their parent into or out of a doubled up housing arrangement, the pattern of changes involving individuals younger than 18 years would align with changes involving extended family or nonrelatives, discussed previously. Thus, I expect that minor children are less likely to enter and leave assisted compared with unassisted households (hypothesis 2).
The Present Study
Informed by previous research, in this article I examine how changes in resources available for housing, through receipt of housing assistance, are related to changes in household composition. Changes in resources are but one characteristic that could predict changes in household composition, but providing material resources for housing is one of the ways in which policymakers could feasibly intervene to encourage household stability. To assess the role of additional resources for housing expenditures, I take advantage of a database of assisted housing records linked to the PSID family survey. These linked data allow me to identify when households first receive housing assistance and track changes in their household composition longitudinally following receipt of assistance. This approach is similar to Kang (2021) and Gold (2018), who rely on the linked PSID data, but I am explicitly focused on changes in household composition, whereas Kang focused on measures of housing instability and Gold focused on residential moves. These linked data also allow me to account for housing assistance in the form of vouchers in addition to assistance provided by other programs. I take advantage of the longitudinal nature of the PSID to use a matching method that improves upon estimates of the effect of housing assistance on changes in household composition from observational data alone.
Subgroup Analysis
After estimating the effect of assisted housing on changes in household composition for all households I assess whether any effect I find is driven by households with children or households without children. Much of the research on the consequences of changes in household composition focuses on children’s outcomes, and households with children have been a focus of research on the effects of assisted housing as well (Fenelon et al. 2018, 2021; Fenelon, Slopen, and Newman 2023; Gold 2018; Kucheva 2018).
Data
I use data from the 1995 through the 2015 waves of the PSID to estimate the effects of housing assistance on changes in household composition. Data collection for the PSID began in 1968 with a panel of approximately 4,800 families containing approximately 18,000 individuals (PSID 2018). The survey was designed to be nationally representative and has followed these individuals and their descendants for more than 50 years. The PSID interviewed families annually from 1968 until 1997 and biennially starting in 1997.
The Assisted Housing Database (AHD) is a restricted-use dataset provided by the PSID that can be linked to the public-use PSID family data file to identify PSID households living in housing units subsidized by HUD, the Farmer’s Home Administration, through tax credits administered by the U.S. Department of Treasury (known as the Low Income Housing Tax Credit), or state programs. The AHD is a result of standardizing addresses both in the PSID and in assisted housing records and then matching these addresses by year to determine which PSID households lived in state- and federally assisted housing. Linking the AHD to the PSID family file allows me to determine the years in which PSID families lived in assisted housing and identifies the type of assisted housing benefiting these families (e.g., public housing, other project-based housing, or vouchers). 1 The AHD matched assisted housing addresses to PSID addresses from 1968 through 2009, but for the purposes of this analysis I am using AHD data from 1995 to 2009. Distinguishing between project- and tenant-based housing assistance is a critical element of this analysis and is possible using the most recent years of linked AHD data. Before 1995, the AHD did not distinguish between project- and tenant-based assistance and severely undercounted addresses where tenant-based assistance was used (see Newman and Schnare (1997) for a detailed description of the AHD). It is possible that households appear to receive assistance for the first time in 1995 because the type of assistance they receive was not captured in the AHD before 1995. Therefore, to be more confident that I am observing a new spell of housing assistance, I measure receipt of housing assistance beginning in 1997 and treat 1995 as the baseline year for households that receive assistance beginning in 1997.
Sample
My analytic sample includes 1,731 heads of household (182 who receive assisted housing and 1,549 unassisted low-income heads of household) that I follow for eight years. Assisted housing data are linked to the survey data through the 2009 wave, which means that the most recent year in which I can observe that a household receives the “treatment” of assisted housing is 2009. I restrict my treatment sample to households whose address matches the AHD for at least two waves (meaning treated households had to receive assistance starting no later than 2007) so that I can estimate the effect of sustained receipt of housing assistance. 2 Households receiving housing assistance for just one wave (i.e., their first year of housing assistance is equal to their last year of housing assistance) may experience instability and changes in household composition following a different pattern from households receiving sustained assistance. I return to this decision in the discussion.
I limit my sample to individuals who were heads of household in the wave that their address was identified as assisted housing. To account for characteristics of households prior to treatment, I observe household characteristics at the wave prior to receipt of assisted housing. I do not restrict my sample to individuals who were also heads of household in the wave prior to receiving assisted housing (though most were). This means that the baseline head of household and household characteristics reflect the living conditions of individuals who received assisted housing prior to receipt and do not in all cases represent those individuals’ own or independent household characteristics. 3
I follow everyone who receives assisted housing between 1997 and 2007 for at least three waves after receiving assistance. This provides six years of follow-up after receipt of housing assistance to observe household composition and change (approximately one quarter of assisted households in my sample maintain assistance for more than six years). For example, if a household receives housing assistance for the first time in 1999, baseline for that household is 1997 and I track changes in that household’s composition in 2001, 2003, and 2005.
Measures
In addition to determining which PSID households received housing assistance, I must also document their household composition and how it does or does not change after receiving assistance. To do this I construct household rosters at each wave of the survey and track which household members enter or leave across waves. I use the relationship to head variable to identify how each member of a household is related to the head so that I can determine who enters and leaves households over time. Biennial data collection means that I may miss some changes in household composition if a household member enters and exits between waves. Surveys with more frequent contact with respondents, such as the Survey of Income and Program Participation, report more changes in household composition than annual or biennial surveys (Perkins 2017).
My models predicting receipt of housing assistance and changes in household composition include a number of covariates representing characteristics of the head of household and household. All of these covariates are measured as baseline characteristics in the wave prior to receipt of assisted housing (1995 at the earliest). The head of household characteristics I include at baseline are sex (female = 1), age, race (Black and other race, with white as the reference), educational attainment (less than high school and high school diploma, with more than high school as the reference), employment (unemployed and retired or disabled, with employed as the reference), marital status (single and widowed, divorced, or separated, with married as the reference), and whether the head was born outside the United States. The household characteristics I include are housing tenure (owned, with rented or neither rented nor owned as the reference), household income (in 2014 dollars), an indicator for whether the household income is below the poverty line, household size, and number of children in the household.
Methods
The goal of this project is to compare households that received assisted housing with similar households that did not receive assisted housing and track changes in household composition over time. The PSID is a nationally representative survey; by comparison, households receiving assistance are a select group with a distribution of characteristics that differs from a nationally representative sample. To address selection into receipt of housing assistance, I first estimate the propensity to receive project- or tenant-based assisted housing using a multinomial logistic model. The covariates I include in the propensity model are the head and household characteristics listed previously. The control group is limited to heads of household who do not receive housing assistance anytime between 1994 and 2015 (some could have had spells of housing assistance before 1994) and who, as a proxy for income eligibility, have a household income equal to or less than 200 percent of the federal poverty line (following Cai 2024 and Lundberg et al. 2021).
Matching heads of household who received project- or tenant-based housing assistance and have similar characteristics with heads who did not receive housing assistance using a propensity score is one way to address the different distributions of characteristics across treatment and control groups. I use the predicted probabilities of receiving project- or tenant-based assistance from the multinomial logistic model to create an IPTW in an attempt to balance covariates in the treatment and control groups. The IPTWs are calculated as the inverse of the predicted probability for the treatment groups who received project- or tenant-based assisted housing and the inverse of one minus the predicted probability for the control group who did not receive any assisted housing. The IPTWs, however, are not an ideal solution for this sample as considerable imbalance in covariates can remain even after applying the weight (see Kim et al. 2017).
To address this shortcoming of the IPTW approach I turn to the generalized overlap weights using the R package PSweight (Zhou et al. 2022). The generalized overlap weights accommodate multiple treatments (in this case, project- and tenant-based assistance) by focusing on the part of the distribution with substantial probabilities to be assigned to any of the treatment groups (Li and Li 2019). They have the useful properties of being bounded between 0 and 1, result in improved balance between the means of covariates in the treatment and control groups, and enable me to compare the effects of project-based assistance versus tenant-based assistance on household changes compared with unassisted households. These weights are calculated by multiplying the inverse probability weights by the harmonic mean of the generalized propensity scores (Li and Li 2019). Overlap weights have been employed to assess the effect of housing assistance on housing insecurity (Kim et al. 2017), in an analysis with similar aims as the one I present here. I additionally use the PSID-provided longitudinal weights to account for sample design and attrition. My descriptive statistics and the multinomial logistic model predicting the propensity to receive project- or tenant-based assisted housing are weighted with the longitudinal survey weights. My models estimating the effect of project- or tenant-based assisted housing on changes in household composition are weighted with the product of the longitudinal survey weight and the generalized overlap weight.
I run separate models predicting six types of changes in household composition: changes involving adults, changes involving children, entrances of adults, entrances of children, exits of adults, and exits of children. Children here mean individuals younger than 18 years, they may or may not be the child of the head of household. These models are generally represented by the following equation:
where P represents the treatment of receipt of project-based housing assistance, T represents the treatment of receipt of tenant-based housing assistance, and
Results
Table 1 presents the baseline characteristics of the heads and households in my sample, stratified by receipt of assisted housing, weighted with the longitudinal survey weight to account for sampling and attrition. The first set of columns show means and standard deviations for the comparison group of households not receiving assisted housing. The second set is for the group receiving project-based assisted housing, and the third is for the group receiving tenant-based assisted housing. A larger proportion of the assisted groups—project- or tenant-based—have a female head of household and a Black head of household at baseline. 5 Educational attainment among assisted heads is somewhat lower, a smaller share is employed at baseline, and a smaller share married. A smaller share of individuals who later received assisted housing live in an owned home at baseline, yet their household income is somewhat higher than those in the unassisted group (given the income restriction in the control group). 6 The household size, number of adults and children in the household, and proportion of households containing children are fairly consistent across groups.
Descriptive Statistics by Type of Housing Assistance.
Note: Weighted with longitudinal survey weight to account for sampling and attrition.
There are also notable differences between the project- and tenant-based groups. A larger share of the tenant-based assistance group is married compared with the project-based group. A larger share of tenant-based recipients has income below the poverty line and is foreign born compared with project-based recipients.
It is important to keep in mind that these baseline characteristics represent the head and household where individuals in the assisted housing groups were living before they received assisted housing. For most of them, this is their own independent household and they are the head of household represented in these statistics. A small number, however, were living with someone else at baseline—a parent or other relative, perhaps—and these statistics represent those heads of household. I decided not to restrict my sample to only individuals who were heads of household prior to receiving housing assistance as this would introduce another element of selection.
Results from the multinomial logistic model predicting treatment of project- or tenant-based assisted housing with baseline covariates suggest that living in an owned home is significantly negatively associated with later receipt of project-based assisted housing compared with living in a rented or neither rented nor owned home. Some of the head of household’s characteristics also predict later receipt of project-based assisted housing. Being Black compared with white and being unmarried compared with married (or living with a Black or unmarried head of household) are positively associated with project-based assisted housing. Being female compared with male and Black compared with white both positively predict receipt of tenant-based assistance. Being foreign born compared with native born is negatively associated with project- and tenant-based assisted housing (coefficients are presented in Table A1 in the Appendix).
Weighting with the product of the overlap weight and the longitudinal survey weight results in improved balance on the means of covariates across respective treatment and control groups compared with unweighted and inverse probability of treatment–weighted means. In general, the treatment and control groups are weighted such that the covariate means across groups more closely match the original distribution in the treated groups. Once balance is achieved it is possible to assess the effect of project- or tenant-based assisted housing on changes in household composition (see Figure A1 for a Love plot demonstrating improved balance achieved with the generalized overlap weights compared with unweighted and inverse probability of treatment–weighted means).
Table 2 presents the proportion of the sample that experienced various types of changes in household composition, again stratified by receipt of assisted housing. The first two rows capture any changes—exits or entrances—involving adults or individuals younger than 18 years (excluding entrances of individuals ages 0, 1, and 2 years, to exclude births, and excluding exits of individuals ages 16 or 17 years, to exclude exits of individuals transitioning to own residence, not restricted to children of householder). These statistics represent the percentage of households that experienced each type of change at least once anytime in the 6 years after receiving assisted housing (or equivalent among the unassisted households). Forty percent of households receiving project-based assistance, and 49 percent of tenant-based assisted households experienced an adult join or leave their household at least once compared with 61 percent of unassisted households. The shares of households experiencing at least one change involving children are 23 percent, 14 percent, and 23 percent, respectively.
Percentage Experiencing Household Change within Six Years.
Note: Weighted with longitudinal survey weight to account for sampling and attrition. The household change involving individuals younger than 18 years categories exclude entrances of individuals ages 0, 1, and 2 years (presumed births) and exits of individuals ages 16 and 17 years (presumed transitions to own residence).
The remainder of the table categorizes household exits separately from entrances. The first two categories under exit or entrance represent a change involving anyone younger than 18 years (same restrictions as described earlier) or a change involving an adult (anyone ≥18 years of age). The remaining categories of partner, child, sibling, parent, other relative, and nonrelative are mutually exclusive and reflect the relationship of the person exiting or entering to the head of household. Approximately 50 percent of households in the unassisted group and one third of households in both assisted groups experienced an adult leave the household at least once over the observation period and a significant share of household heads across groups experienced the exit or entrance of a child (both minor and adult children of the householder are included in this category). Between 8 percent and 18 percent of households experienced the exit or entrance of the head’s partner over 6 years. Smaller shares of household heads experienced changes involving their siblings, parents, other relatives, or nonrelatives. Taking the difference between household changes involving adults and household changes involving partners suggests that the majority of changes involving adults are when a nonpartner exits or enters (Walter et al.’s [2024] findings are similar in a sample of assisted renters). Notably, households receiving tenant-based assistance have higher rates of household changes involving other relatives compared with unassisted and project-based assisted households.
Table 3 presents coefficients from overlap-weighted logistic models estimating the effect of project- and tenant-based assisted housing on six types of changes in household composition: (1) any change (exit or entrance) involving an adult, (2) any change involving an individual younger than 18 years, (3) an adult exited the household, (4) an individual aged 0 to 16 years exited the household, (5) an adult entered the household, and (6) an individual aged 3 to 17 years entered the household. I focus here on changes involving any adults or individuals younger than 18 years and report results for specific relationship categories in the Appendix (Table A2). This categorization is more parsimonious compared with the relationships listed in Table 2 and corresponds to different potential mechanisms connecting assisted housing with changes in household composition (outlined earlier). Each row in Table 3 reports the odds ratios associated with each treatment condition from a different logistic model and the odds ratios represent the effect of receipt of project- or tenant-based assisted housing on experiencing change involving at least one person in each category within six years of receipt. Each model also controls for the baseline covariates that were used in the model to predict the propensity score and calculate the overlap weights (following Kim et al. 2017), though these coefficients are not shown in the table.
Effect of Assisted Housing on Household Change.
Note: The table shows the output from six logistic regression models predicting one of six types of household change. Each row shows the treatment odds ratios and 95 percent CIs predicting household change within six years after receipt, controlling for the covariates in Table 1, weighted with the product of the survey and overlap weights. The results are consistent when modeled with rare events logistic models. CI = confidence interval.
p < .05. ***p < .001.
All of the estimated odds ratios in Table 3 are less than 1, suggesting that after accounting for selection with the overlap weights, households receiving project- and tenant-based assisted housing have a lower likelihood of experiencing changes in household composition in the subsequent 6 years compared with unassisted households. In aggregate, receipt of project-based assistance reduces the likelihood of having an adult or child exit or enter the household over the subsequent 6 years. All of the project-based odds ratios are rather large in magnitude and all are significant at the .05 level. These results suggest that project-based housing assistance reduces household instability. Receipt of tenant-based assistance significantly reduces the probability of a household change involving an adult, and an adult entrance in the aggregate sample compared with unassisted households. The odds ratios for adult exits, and child entrances and exits are less than one, but not significant. The multinomial logistic model I used to construct the generalized overlap weights means I can compare the odds ratios for project-based assisted households with those for tenant-based assisted households using postestimation tests. Despite only two significant odds ratios among tenant-based households compared with six significant odds ratios among project-based households, none of the tenant-based odds ratios presented in Table 3 are significantly different from the corresponding project-based odds ratios. Together, the results in Tables 2 and 3 support hypothesis 1a, that assisted housing reduces the likelihood of household changes involving adults. Table 2 reveals that the majority of changes involving adults are the household head’s adult child, sibling, parent, other, or nonrelative exiting or entering. Partners account for a minority of adult changes. Although Table 3 does not disaggregate by relationship status, the odds ratios for changes involving any adult suggest assisted housing, especially project based, reduces adult changes. Table 3 also supports hypothesis 2 for project-based assistance: project-based assisted housing reduces changes, exits, and entrances involving children younger than 18 years.
Figure 1 shows predicted probability of household change of a given type within six years of receiving housing assistance for individuals who received project-based (blue) or tenant-based (orange) assistance compared with unassisted households (gray). Unassisted household heads had a .68 and .32 probability of experiencing a household change involving an adult or child in the six years after receipt compared with a .41 and .15 probability among heads who received project-based assistance and a .51 and .24 probability among heads who received tenant-based assistance. Predicted probabilities of adult (.52 vs. .35) and child exit (.25 vs. .11) and adult (.51 vs. .30) or child entrance (.21 vs. .12) also show particularly dramatic differences between unassisted and project-based assisted groups. Following the results in Table 3, the differences in predicted probabilities between unassisted and tenant-assisted groups are less dramatic and less consistently significant (nonsignificant differences represented by the partially transparent orange bars).
Predicted probability of household change within six years of receiving project-based housing.
Supplemental Analyses
Table A2 presents results from models predicting changes categorized by relationship to the household head rather than age (adult vs. child). These models are necessarily based on smaller numbers of households experiencing these particular changes so results should be interpreted with caution. Project-based assistance reduces the likelihood of experiencing household changes involving the head of household’s partner, children, and other and nonrelatives in the six years following receipt compared with unassisted households, and five of the six odds ratios are significant. The results for partner exit versus changes involving children and other relatives or nonrelatives underscore the ambiguity expressed by competing hypotheses 1a and 1b: when we look at the relationship of household members rather than simply age, we see that project-based housing reduces changes among the head of household’s (adult or minor) children and other relatives or nonrelatives, reduces entrances of partners, but does not significantly reduce partner exits. The odds ratios for tenant-based assistance are all below one, yet only entrances involving the head’s partner or child are significant. The odds ratios for other relative or nonrelative exit and other relative or nonrelative entrance are significantly lower for project- compared with tenant-based assistance. Again, disaggregating by relationship type suggests that tenant-based housing may reduce some types of changes, but not others, providing support for hypotheses 1a and 1b.
The next set of odds ratios are from models stratified by the presence of individuals younger than 18 years in the household in the wave housing assistance was received, an important group to isolate as both housing assistance and household instability have consequences for children’s development and outcomes (Table A3 reports percent of households experiencing household change, stratified by presence of children). In households without children younger than 18 years, project-based assistance significantly reduces the likelihood of experiencing changes involving partners, children, and other relatives or nonrelatives of the head of household, whereas tenant-based assistance marginally reduces the odds of other relative or nonrelative exits and child entrances. Odds ratios for child exit and other relative or nonrelative entrance are significantly smaller for project- compared with tenant-assisted households.
Among households with at least one child younger than 18 years at receipt of assistance, four of the six types of household change are significantly reduced for those with project-based assistance and two are significantly reduced among those with tenant-based assistance. Changes involving other relatives or nonrelatives, exits involving children, and entrances involving partners are significantly less likely among project-based assisted households with children compared with unassisted households with children. Tenant-based assistance reduces the odds of entrances involving partners and children compared with unassisted households with children. Odds ratios for exits and entrances involving other/nonrelatives are significantly smaller for project- compared with tenant-assisted households.
Discussion and Conclusion
Low-income families often face high housing cost burdens and challenges securing stable and affordable housing (Joint Center for Housing Studies 2024). Some low-income families solve affordability challenges by sharing housing with others (Harvey and Dunifon 2023), yet this can lead to consequential instability for children, in particular, as shared households tend to dissolve quickly (Pilkauskas 2012). Housing assistance is intended to provide safe and affordable housing to low-income families. Might it also lead to stability in household composition? Taken together, my results suggest that assisted housing can result in more stability in household composition compared with similar households that do not receive assisted housing, but the type of housing assistance matters. Project-based assistance reduces the likelihood of experiencing household changes involving adults and children in the six years following receipt compared with unassisted households. Project-based assistance appears to lead to more stability among both households with and without minor children. Tenant-based assistance, however, has more limited significant effects on changes in household composition, suggesting that households receiving tenant-based housing assistance are less likely to experience changes involving adults, but in aggregate, no more or less likely than unassisted households to experience changes in household composition involving children.
That the effects of project-based assisted housing on the likelihood of experiencing changes in household composition involving adults and children are negative is consistent with the idea that assisted housing provides families with stable and affordable housing that means they do not need to double up with other individuals and families to afford a place to live. Receipt of project-based assistance reduces the likelihood of household changes among households with and without children. We would expect to see a reduction in changes if assisted housing were enabling families to afford housing on their own and limiting their need to double up with other households to make ends meet.
Why might the format of housing assistance matter for subsequent changes in household composition? Perhaps the mechanism connecting housing assistance to fewer changes in household composition is stability rather than affordability. Households in project- and tenant-based assisted housing generally have affordable housing, especially in the few years immediately following receipt of assistance (Gold 2020; Mast 2014). There is a long line of research finding that housing stability, however, varies by the type of housing assistance received. Voucher recipients are more likely than public housing residents to experience housing instability (Kang 2021), and public housing is associated with lower probability of moving and longer length of residence compared with unassisted households (Beck 2019; Gold 2018). Receiving project-based housing assistance means that a household has been assigned a physical unit of housing and is likely to stay in the same unit for as long as they are receiving assistance (or at least for the six years my observation window covers). PHAs or property managers assign units on the basis of household size and attempt to match households to units with a sufficient number of, but not too many, bedrooms to fit their household. Tenant-based housing assistance also incorporates household size and number of bedrooms needed, but households must find their own units in the private market, and they may need to move if a landlord decides to stop accepting their voucher or if the unit fails a housing quality inspection. Voucher recipients may accept low-quality housing if they are short on time to lease up, which could lead to unit dissatisfaction and subsequent moves (Petrakos et al. 2025), and in turn, changes in household composition. Housing stability is significantly better for recipients of project- versus tenant-based assistance (Gold 2018; Kang 2021), and project-based assistance predicts better mental health among children compared with a similar group of unassisted children (Fenelon et al. 2018), whereas children in households receiving tenant-based assistance have better health, fewer days of school missed, and live in a more advantaged neighborhood compared with unassisted children (Fenelon et al. 2021, 2023).
Different results by type of assisted housing raises one threat to the validity of my findings. Perhaps PSID respondents who receive assisted housing are simply not reporting changes in household composition because they are concerned about losing their benefits if they reveal others have moved in with them? I acknowledge, but am relatively unconcerned about this source of bias. The PSID-AHD link is from administrative data and carried out separately from the survey. Some PSID respondents accurately report that they receive housing assistance but there is indeed some mismatch between self-report and administrative indicators of housing assistance such that some of the respondents I observe as receiving housing assistance do not self-report receipt. If PSID respondents are underreporting household changes because they are concerned about losing housing assistance, I would expect the underreporting to appear primarily in household entrances, not household exits, yet the results are fairly consistent across exits and entrances. That this study uses administrative data on assisted housing is a strength given high rates of inaccurate self-report of assistance status (Shroder 2002). The biennial structure of the PSID, however, may mean that I miss changes in household composition that occur between waves and misclassify households that experience changes as stable.
I am measuring changes in household composition that occur in the two to six years after households receive assistance. What about changes in household composition that occur simultaneous to receipt of housing assistance? As I speculate in the introduction, receiving housing assistance may result in immediate changes in household composition as doubled up and extended family households dissolve to establish independent households. Indeed, this is the case. In supplemental analyses I find that both project- and tenant-assisted households have significantly higher odds of experiencing changes involving adults and changes involving individuals younger than 18 years in the wave they receive assistance compared with unassisted households. Household changes could be part of the selection process into assisted housing if assistance is designated for a family unit, excluding extended family and nonrelatives with whom they may be living at the time of application. These changes may be positive, if they reduce crowding (Zhu et al. 2025) or permit mothers of young children to establish their own households (Harvey 2022).
This analysis addresses selection into receipt of assisted housing using a weighting method that achieves covariate balance between the treatment groups of households receiving project- and tenant-based assisted housing and a control group that did not receive assisted housing. This method addresses concerns about covariate imbalance in observational data, yet interpreting my results as causal estimates requires assuming I have included all variables predicting receipt of project- or tenant-based assistance in the models I used to calculate the overlap weights. If they were available, it would be valuable to include other variables capturing characteristics of heads, household members, and households such as housing histories, criminal justice contact, and substance use, which could affect receipt of assisted housing.
A study of low-income New Yorkers finds that housing assistance sustained over two years consistently reduces rent burden, crowding, homelessness, and poverty status whereas intermittent housing assistance does not have these benefits (Cai 2024). I purposefully limited the treatment group of assisted housing to household heads who had at least two consecutive waves of housing assistance identified in the PSID because I wanted to measure the effects of sustained housing assistance. This restriction cut the sample size substantially, but I argue that it allows a better estimate of the effect of housing assistance. Household heads who had just one wave of housing assistance (i.e., their first year of housing assistance equaled their final year of housing assistance) may not have had time to achieve stability in household composition, and if gaining and then losing housing assistance involved one or more moves, changes in household composition might have been more, not less, likely than for unassisted households. Results from two supplemental approaches not shown support this conclusion: including households with only one wave of housing assistance identified or limiting the treatment group to households who lose assistance during the observation period both return coefficients that are much less consistently significantly negative. That said, more research remains to be done on how the timing and duration of housing assistance is associated with household stability.
It should also be possible to observe changes in household composition for individuals leaving housing assistance. Households leaving project- and voucher-assisted housing have more housing instability than households that remain in public housing (Kang 2021), and it would be worth exploring whether leaving assisted housing predicts doubling up or other changes in household composition. Measuring any exit from housing assistance may be less informative than capturing whether the exit was positive, negative, or neutral, as type of exit predicts subsequent experiences with homelessness (Petrakos et al. 2025), a good reminder of heterogeneous pathways into and out of housing assistance. Only about half of the assisted households in my sample exited assistance after four years, thus it was not possible for me to estimate the effects of leaving housing assistance given my small sample size. This should be a priority for future research.
These results suggest that policymakers could play a role in increasing household stability among low-income families through the provision of assisted housing. There are, however, debates about the best way to improve family stability and economic resources. Building and operating project-based subsidized housing is expensive and inefficient, according to some analyses (Early and Olsen 2025). Yet the benefits of household and housing stability for individuals, families, and children are clear (Bryan and Harvey 2025; Cavanagh and Fomby 2019; Perkins 2019, 2023) and the connection between housing assistance and housing stability is well established (Gold 2018; Kang 2021; Lundberg et al. 2021). Directly providing housing assistance is critical for addressing instability in household composition in the current housing context of the United States, characterized by an affordability crisis and limited supply of low-cost housing (Joint Center for Housing Studies 2024). Cash assistance or other material resources cannot help if low-cost housing units are unavailable. Directly providing housing assistance through project- and tenant-based programs remains a priority for supporting the country’s low-income families and households.
Footnotes
Appendix: Prediction Models,Supplemental Analyses,and Diagnostics
Percentage Experiencing Household Change over Six Years: Households with and without Children Younger Than 18 Years at Wave 2.
| Without Children (%) | With Children (%) | |
|---|---|---|
| Household change involving adults | 55 | 65 |
| Household change involving individuals <18 years of age | 12 | 35 |
| Household change: one or more exits | ||
| Adults | 43 | 56 |
| Individuals <18 years of age | 6 | 30 |
| Partner | 14 | 15 |
| Child | 17 | 50 |
| Sibling | 2 | 4 |
| Parent | 2 | 2 |
| Other relative | 5 | 14 |
| Nonrelative | 1 | 3 |
| Household change: one or more entrances | ||
| Adults | 39 | 43 |
| Individuals <18 years of age | 10 | 18 |
| Partner | 19 | 16 |
| Child | 21 | 35 |
| Sibling | 2 | 3 |
| Parent | 1 | 2 |
| Other relative | 6 | 16 |
| Nonrelative | 3 | 5 |
| Observations | 762 | 969 |
Note: Weighted with longitudinal survey weight to account for sampling and attrition. The household change involving individuals younger than 18 years categories exclude entrances of individuals ages 0, 1, and 2 years (presumed births) and exits of individuals ages 16 and 17 years (presumed transitions to own residence).
Acknowledgements
For thoughtful comments I thank Michael C. Lens, Ann Owens, Evan Roberts, and Jonathan Spader. Vernika Mrig provided excellent coding assistance. Any errors are mine alone.
Funding
The author disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The collection of data used in this study was partly supported by the National Institutes of Health (grants R01 HD069609 and R01 AG040213) and the National Science Foundation (awards SES 1157698 and 1623684).
1
The AHD identifies state-assisted housing in some survey years, but not in the years of data I use in this analysis.
2
The AHD links end in 2009, but I can observe household composition through 2015.
3
For example, if Ann was head of household and received assistance in 1997 but lived in her mother’s home in 1995, the baseline characteristics reflect her mother’s household characteristics, as Ann’s household did not yet exist.
4
5
A very small proportion of the assisted housing groups are Hispanic, which is consistent with other research using the PSID-AHD (Gold 2018; 
)
6
Results do not change when the control group is not restricted to heads of households with income less than 200 percent of the poverty line.