Abstract
We exploited an exogenous health shock—namely, the birth of a child with a severe health condition—to investigate the effect of a life shock on homelessness in large cities in the United States as well as the interactive effects of the shock with housing market characteristics. We considered a traditional measure of homelessness, two measures of housing instability thought to be precursors to homelessness, and a combined measure that approximates the broadened conceptualization of homelessness under the 2009 Homeless Emergency Assistance and Rapid Transition to Housing Act (2010). We found that the shock substantially increases the likelihood of family homelessness, particularly in cities with high housing costs. The findings are consistent with the economic theory of homelessness, which posits that homelessness results from a conjunction of adverse circumstances in which housing markets and individual characteristics collide.
Introduction
Homelessness is a significant and often glaring social problem in the United States, particularly in urban areas (Lee et al. 2010). The magnitude is difficult to quantify because the homeless are underrepresented in household-based surveys, there is no standard methodology for counting the homeless, and homelessness is often a transient state (Link et al. 1994). Even the definition of homelessness is evolving and subject to debate (Gould and Williams 2011; Lee et al. 2010). Despite these issues, there is general agreement that U.S. homelessness rose dramatically during the 1980s (Burt 1992) and that it has gone up, particularly among families, since the 2008 foreclosure crisis and subsequent Great Recession (U.S. Conference of Mayors 2010). Aside from being direct evidence of deprivation of a basic human need, homelessness appears to adversely affect children’s health and development (Buckner 2008).
There is much interest in learning about the causes of homelessness but little population-based data with which to do so. In an aggregate city-level analysis conducted shortly after the large increase in homelessness during the 1980s, Honig and Filer (1993) found that city-level rents were strongly and positively associated with homelessness rates after controlling for a host of other city- and state-level factors. Using data from an urban birth cohort study, Fertig and Reingold (2008) found that poor physical and mental health, domestic violence, and residential mobility were significantly associated with homelessness, and that being an immigrant, receiving public housing subsidies, and having family support were negatively associated with homelessness among mothers with young children.
Recent empirical and theoretical work suggests that adverse life events can push families into homelessness. Two recent studies found that paternal incarceration is positively associated with transitions to homelessness among young children (Wildeman forthcoming) and their mothers (Geller and Walker 2012), and Curtis et al. (2010) found that infant health shocks increase the likelihood of homelessness. All three studies, in addition to Fertig and Reingold (2008), used the Fragile Families and Child Wellbeing Study data, which as far as we know is the only contemporary data set that allows for rigorous longitudinal analysis of life events on homelessness.
Other work particularly relevant to the current study is the economic theory of homelessness developed by O’Flaherty (2004:2), who argued that homelessness results from a “conjunction of bad circumstances” occurring when market conditions and individual characteristics collide. O’Flaherty further argued that the transitory component of income has been largely overlooked as a potential determinant of homelessness and that reducing income volatility to buffer the potential effects of adverse life events—such as health shocks, relationship dissolution, or unemployment—may be the best way to prevent homelessness (O’Flaherty 2008, 2009). He found that the most common shocks experienced by families involve income or health and that the main shocks preceding homelessness involve income; however, he cautioned that these descriptive findings should not be interpreted as causal (O’Flaherty 2009).
In this article, we exploit an exogenous health shock—namely, the birth of a child with a severe health condition that is considered random by the medical community—to investigate the effect of a life shock on homelessness as well as the interactive effects of that shock with housing market characteristics. By exploiting a shock that recent literature indicates has economic repercussions, we indirectly test O’Flaherty’s proposition that income volatility is a key factor affecting homelessness. By interacting the shock with measures of housing prices and subsidized housing availability, we test O’Flaherty’s theory that homelessness results from a conjunction of bad circumstances occurring when market conditions and individual characteristics collide. We consider both a traditional measure of homelessness (which includes lack of a permanent residence or residence in temporary shelters) and a broader and more contemporary measure that also includes doubling up without paying rent and residential instability. This research builds on the Curtis et al. (2010) study, which investigated the effects of infant health shocks on a broad array of housing outcomes, including homelessness. The two major contributions of the current study are that it tests O’Flaherty’s proposition that high housing costs exacerbate the effects of adverse life circumstances on homelessness and that it incorporates the broadened conceptualization of homelessness recently adopted by the U.S. Department of Housing and Urban Development (HUD).
Background
Defining and Measuring Homelessness
Annual homelessness counts conducted by HUD produce point-in-time counts of all sheltered and unsheltered homeless persons on a single night in January as well as one-year estimates of the total sheltered homeless population based on information from local Homeless Management Information Systems. These estimates provide information to legislators about the magnitude of homelessness and are used to inform service delivery.
According to HUD (2010), there were 649,917 unsheltered and sheltered homeless individuals in the United States in January 2010, including 79,446 family households with 241,951 people in those families. These estimates do not include individuals who are “doubled up” with family or friends because of economic difficulties or who confront eviction or other forms of housing instability short of homelessness as it is traditionally defined (HUD 2010). Although most homeless individuals are single male adults, children and families make up a larger percentage of homeless individuals today than in the past (Lee et al. 2010). One study of annual prevalence rates—rather than point-in-time measurements—found a higher risk of sheltered homelessness among young children than among men (Culhane and Metraux 1999).
Researchers using secondary data sets to study homelessness have typically followed federal guidelines (General Definition of Homeless Individual 2007), defining homelessness as the lack of a fixed, regular, and adequate nighttime residence, or residence in a temporary accommodation (shelter, transitional housing, or welfare hotel) or in a public or private space not intended for residence (e.g., a car or an abandoned building). Most studies of family homelessness in particular have focused on families who are currently living in emergency or transitional shelters (Howard et al. 2009; Kerker et al. 2011; Miller 2011; Swick and Williams 2010;) or who report having lived on the street, in abandoned property, or in a shelter (Coker et al. 2009; Fertig and Reingold 2008).
Studies focusing on housing insecurity (housing hardships short of homelessness and that may be precursors to it) have considered measures of eviction; frequent moves; difficulty paying rent, mortgage, or utilities; spending > 50 % of household income on housing; living in overcrowded conditions; or moving in with others because of financial difficulties (Desmond 2012; Gilman et al. 2003; Kushel et al. 2005; Pavao et al. 2007; Phinney et al. 2007). The majority of homeless people, as traditionally defined, experience periods of housing insecurity (Reid et al. 2008; Sosin 2003).
Although the definition of homelessness has been relatively consistent across existing studies, the U.S. federal government recently expanded the official definition in order to recognize and serve individuals and families who previously were not considered homeless but who are close to meeting those conditions. As described in the Federal Register (2010), the 2009 Homeless Emergency Assistance and Rapid Transition to Housing (HEARTH) Act (2010) expanded the definition of homelessness to include the “imminently homeless,” outlining several broad categories under which one can qualify for homeless assistance. One category is consistent with prior definitions as an “individual or family that lacks a fixed, regular and adequate nighttime residence,” including places that are uninhabitable for humans, emergency shelters, and motels or hotels provided through social service agencies. Another is “individuals or families who will imminently lose their primary nighttime residence” within a 14-day period, wherein evidence of eviction is provided, the individual or family has no other resources, and no other residence has been identified. Another includes families with young children that have moved at least twice and have not been named on a lease within the last 60 days. In general, the HEARTH Act expanded the concept of homelessness to include experiences such as eviction, residential instability, and not having one’s own home.
The Economics of Homelessness
To understand the dramatic rise in homelessness in the 1980s, O’Flaherty (1996) formulated a microeconomic theory of homelessness in which high-priced housing markets lead landlords to disinvest in (or poorly maintain) low-priced rental units. Consumers at the lowest end of the income distribution, therefore, must choose between very low-quality housing at a certain price or homelessness. Under severe income constraints and holding preferences constant, a rational consumer would be indifferent between spending a substantial proportion of his/her income on very low-quality housing or being homeless. Homelessness, then, would be dependent on the housing markets faced by individuals at the bottom of the income distribution.
In studies based on O’Flaherty’s framework, Quigley et al. (2001) found that the demand for the lowest-quality housing explains much of the variation in rates of homelessness; and Early (2005) found that families with children, younger heads of household, or substance abusers, and families who face higher rental prices for low-quality housing are at increased risk for homelessness. Glomm and John (2002) predicted homelessness as a function of low income and borrowing constraints in order to estimate the effects of homelessness on labor productivity, providing some evidence that exogenous income parameters can lead to homelessness.
Together, the intriguing arguments and observations in the aforementioned literature point to the questions of whether and to what extent adverse health shocks lead families into homelessness. Three recent studies, all based on longitudinal population-based data, have produced varying degrees of evidence bearing on this question. Fertig and Reingold (2008) examined correlates of homelessness among mothers with young children with family incomes at <50 % of poverty and found that both poor health status and depression of mothers were positively associated with homelessness. Phinney et al. (2007) found that mental and physical health problems were positively associated with homelessness among mothers who received cash assistance. Neither study explicitly addressed the potential endogeneity of health: that is, they did not attempt to isolate the effects of health shocks. In contrast, Curtis et al. (2010) focused specifically on the effects of infant health shocks and found that among mothers with young children, poor child health increases the likelihood of both overcrowding and homelessness and may also increase the likelihood of having inadequate utilities and poor housing quality.
Contribution of This Study
We exploit an exogenous life shock—namely, the birth of a child with a severe health condition considered by the medical community to be random—to investigate the effect of a health shock on homelessness and interactive effects between the shock and housing market characteristics. Like Curtis et al. (2010), we use survey data from the Fragile Families and Child Wellbeing birth cohort study augmented with information from hospital medical records that are used to create measures of infant health shocks. Although child health may not directly affect the family’s income given that most children do not work, poor child health has been shown to affect a variety of family resources (Reichman et al. 2008). Children in poor health require substantial time resources, which can limit parents’ ability to maintain employment; indeed, numerous studies have found adverse effects of poor child health on parents’ labor supply (e.g., Corman et al. 2005; Gould 2004; Noonan et al. 2005; Powers 2003). In addition, studies have found that poor child health makes it less likely that the father will live with the child (Reichman et al. 2004) and more likely the father will become incarcerated (Corman et al. 2011). The labor supply and household structure consequences of poor child health can have negative financial ramifications for the child’s household, as can the direct costs associated with child disability, such as specialized child care. However, the effects may be offset, at least to some extent, by increased access to public support. Reichman et al. (2006) found that families with young children in poor health are more likely than those with healthy children to receive Temporary Assistance for Needy Families, Supplemental Security Income (SSI), and housing subsidies.
We expand upon the Curtis et al. (2010) study, the only previous study to specifically investigate the effects of a health shock on housing, in two major ways. First, we test O’Flaherty’s theory that homelessness results from a conjunction of adverse circumstances occurring when individual circumstances and market conditions collide. As such, we consider not only the effects of infant health shocks but also their interactive effects with housing costs and availability of subsidized housing. The Curtis et al. (2010) study, which did not consider such interactions or otherwise refer to or test O’Flaherty’s theory, was framed around ascertaining the importance of “reverse pathways” between housing and health. Second, we use a new measure of homelessness that approximates the expanded definition of homelessness under the HEARTH Act, as well as the “component” measures of doubling up without paying rent and residential instability that distinguish this measure from the traditional measure of homelessness. Curtis et al. (2010) did not incorporate this broader conceptualization of homelessness, instead considering homelessness (traditionally defined) as part of a broad array of housing outcomes in the domains of quality, crowding, and stability. As far as we know, this is the first study in the social science literature to approximate the broadened definition of homelessness under the HEARTH Act by incorporating measures of eviction or multiple moves and doubling up without paying rent or to estimate the effects of a life shock on these precursors to homelessness.
Data and Measures
We used data from the Fragile Families and Child Wellbeing (FFCWB) study, which follows a cohort of parents and their children born in 1998–2000 in 20 large U.S. cities (in 15 states). By design, approximately three-quarters of the births were out of wedlock. Face-to-face interviews were conducted with 4,898 mothers in the hospital after giving birth (Reichman et al. 2001). Follow-up interviews were conducted approximately one, three, and five years later (plus a nine-year follow-up not used for this study). The baseline response rate was 86 % among eligible mothers; the response rates at the one-, three-, and five-year follow-ups were, respectively, 89 %, 86 %, and 85 % of mothers interviewed at baseline.
As part of an “add on” study, data from medical records (from the birth hospitalization) of the mother and child were collected using a detailed instrument based on the U.S. Standard Certificate of Live Birth. The availability of medical record data mostly depended on administrative processes of hospitals rather than decisions on the part of survey respondents to make their records available. Medical record data, which were needed to characterize infant health shocks, are available for 3,684 (75 %) of the 4,898 births in the FFCWB sample. Of those, 3,192 mothers completed the three-year interview, and 28 had missing data on key analysis variables, leaving an analysis sample of 3,164 cases.
Outcome Measures
We considered four different homelessness-related outcomes: a traditional measure of homelessness, two measures of housing instability thought to be precursors to homelessness, and an expanded measure that combines the first three and approximates the broadened conceptualization of homelessness as reflected in the HEARTH Act. We focus primarily on these outcomes measured at three years but refer to auxiliary analyses that expand the observation window to five years.
Homeless—Traditional Measure (3 % of Sample)
This measure characterizes homelessness as lack of a fixed, regular, and adequate nighttime residence, or residence in a temporary accommodation or space not intended for residence. In the three-year interview, the mother was asked where she currently lived. Response choices included living in a temporary shelter or being homeless. She was also asked whether she had stayed at a shelter, in an abandoned building, in an automobile, or any other place not meant for regular housing, even for one night, in the past 12 months. If she responded affirmatively to either question, she was coded as having been homeless.
Evicted or Multiple Moves (EMM) (6 %)
This measure includes either (1) an affirmative response to the following question in the three-year interview, “Were you evicted from your home or apartment for not paying the rent or mortgage [in the past year]?” or (2) situations in which the mother moved three or more times between the one- and three-year interviews (four or more times if she did not complete the one-year interview), based on the following questions at three years: “Have you moved since date of last interview?,” and “How many times?” The three-move cutoff seems reasonable given findings by Wood et al. (1990) and Weinreb et al. (1998) that moving more than once per year is a risk factor for homelessness.
Doubling Up and Not Paying Rent (DUNR) (6 %)
The mother was coded as DUNR if she indicated that she (1) currently lives with family or friends and pays no rent or (2) lives in a home owned by another family member, pays no rent, and, as determined by her household roster, is living with an adult other than a spouse or a partner.
Homeless—Expanded Measure (13 %)
This measure combines the first three and approximates the expanded definition of homelessness under the HEARTH Act. That is, it includes not only the conventional definition of homelessness but also potential precursors to homelessness captured by EMM and DUNR. As indicated earlier, the expanded definition under the HEARTH Act includes situations in which families are living with others and not paying rent, have evidence of eviction within the next 14 days, and/or have moved two or more times within 60 days. We created the DUNR and EMM variables to capture, as much as possible, these three types of situations, with EMM representing a composite of eviction and residential instability.
For DUNR, the difference between our measure and the HEARTH construct is that we cannot know, because of the retrospective nature of the survey, whether those coded as DUNR are in imminent danger of losing housing; however, families that we coded as DUNR represent a group clearly identified by the HEARTH Act as vulnerable. For EMM, we combined having experienced any eviction and/or having moved three or more times in the past two years, to approximate the combined HEARTH Act components of facing eviction within 14 days and having moved two or more times within 60 days. The timing inconsistencies make our EMM measure a less-than-perfect proxy for the relevant HEARTH constructs, although having moved more than three times in the past three years was relatively rare in our sample, with fewer than 1 % of families having moved four or more times.
Infant Health Shocks
We considered three different measures of poor infant health that have been used successfully to study effects of life shocks on crime (Corman et al. 2011), housing conditions (Curtis et al. 2010), and social capital (Schultz et al. 2009). With our goal of isolating causal effects of health shocks on homelessness and related outcomes, the ideal measure of poor infant health would (1) characterize a health shock that was present at birth and unlikely a function of parental behaviors, and (2) capture conditions strongly associated with long-term morbidity. A pediatric consultant classified each condition in the infants’ medical record or reported by the mother at one year according to degree of severity (in terms of expected significant long-term morbidity) and likelihood, according to the medical community, of having been caused by parental behavior (see Table 5 in the appendix). Our goal was to capture severe conditions that are largely random (e.g., Down syndrome, congenital heart malformations), given that the pregnancy resulted in a live birth.
The first measure—severe infant health condition (SIHC)—includes conditions that are severe; chronic; unlikely caused by prenatal behavior; and, in the case of one-year maternal reports, likely present at birth. This measure, which is relatively rare (2 % of sample), best meets our “gold standard” and is thus our preferred measure.
The second measure is severe infant health condition or very low birth weight (SIHC/VLBW). VLBW, defined as < 1,500g, is associated with a number of serious and long-term health conditions (Reichman 2005). The advantage of this measure is that we gain a few additional analysis cases with poor infant health (3 % of the sample are categorized as having a SIHC/VLBW). The disadvantage is that the VLBW component may not be exogenous because birth weight is inversely associated with poverty (Reichman 2005). All but three of the VLBW infants also had moderately severe infant health conditions, defined as conditions not considered to be related to parental behavior that may or may not have long-term health consequences.
The third measure—moderate or severe infant health condition (MSIHC)—includes any condition that meets the criteria for SIHC or is less severe but still considered random (not a function of parental behavior). The disadvantage of this measure, which characterizes 20 % of the sample, is that it is very broad: that is, it includes shocks that may or may not have poor long-term prognoses (e.g., hydrocephalus).
Because there is neither a standard for measurement nor a consistent reporting of child disability (Reichman et al. 2008), it is difficult to provide a national comparison for our measures of poor infant health. Our rates are generally consistent with the range of 6 % to 18 % of children in the United States that have special health care needs as reported by Stein (2005). It is not unexpected that our strictest measure (SIHC) is lower than this range, since it includes only very serious conditions and excludes conditions known to be related to parental behaviors. Likewise, it is not unexpected that our broadest measure (MSIHC) is slightly higher than the upper-bound estimate, since it is defined to include conditions that are not necessarily disabling in the long run.
Covariates
Sample means (all measured at baseline unless specified otherwise) are presented in Table 1. We included many sociodemographic characteristics: maternal age, race/ethnicity, nativity, education, employment, parity, parents’ relationship status, Medicaid or other public insurance (proxy for poverty), and census-tract poverty. By design, only one-quarter (24 %) of the mothers were married. One-half (48 %) were non-Hispanic black, 28 % were Hispanic, more than one-third had less than a high school education, and 65 % had publicly financed births. The mean age was 25.
We included indicators for whether the mother’s medical record included documentation of any preexisting physical health condition (20 %) and diagnosed mental disorder (11 %). We controlled for multiple birth, child’s gender, and child’s age at the time of the three-year interview. We included indicators for noncompletion of the one-year interview and for missing data on census-tract poverty, both of which may be related to housing instability. Our main specifications included indicators for the mother’s city of residence in order to control for housing markets or other city- or state-level characteristics that may be associated with both infant health and housing instability. Eight cities with fewer than 100 observations were aggregated into a single indicator. In some specifications, we included city-level unemployment rates in 2000 instead of city indicators.
In some models, we controlled for the mother’s housing conditions and/or living arrangements prior to the birth, to some extent allowing us to capture changes in, rather than levels of, housing instability. The first measure captures homelessness or poor housing quality based on information abstracted from the mother’s prenatal medical record (2 % of sample). The abstractors were instructed to record any mention (in progress notes or elsewhere in the chart) of specific situational challenges, including “homelessness or threatened eviction” and “inadequate heat, electricity, or running water or other poor housing/living condition,” which were combined to create the baseline measure of homelessness or poor housing quality. The second measure is whether the mother lived with any adult other than the baby’s father during the pregnancy (one-third of sample), based on information provided by the mother in her household roster (which included information about the age of all household members and their relationship to her) at baseline.
Housing Market Conditions
For housing costs, we use two different measures: Fair Market Rents (FMRs), and a more targeted index of rental housing prices developed by Carrillo et al. (2010). FMRs are gross rental estimates compiled by HUD to set the payment standard for its Housing Choice Voucher program and are therefore appropriate for capturing costs at the low end of the market. We measured monthly FMRs in 2000 for a two-bedroom unit at the Primary Metropolitan Statistical Area (PMSA) level in cities for which HUD made data available at that level (9 of the 20 cities) and at the (broader) Metropolitan Statistical Area (MSA) level in the other cities. FMRs in our sample were significantly higher than the national average as a result of the sampling frame (cities with more than 200,000 people). The national average FMR for a two-bedroom unit in 2000 was $443 (HUD 2011), compared with $728 in the FFCWS cities. FMRs varied considerably across the sampled cities, ranging from less than $550 to more than $1,200 (see Fig. 2 in the appendix).
The Carrillo et al. (2010) rental housing price index (RHPI) is based on FMRs as well as neighborhood characteristics of units occupied by families with tenant-based housing vouchers. The RHPI is thus particularly appropriate when studying disadvantaged populations. As for FMRs, we used rents in 2000 at the PMSA level when available and at the MSA level otherwise. The RHPI is scaled so that the mean across all areas is 1, with higher-priced markets greater than 1 and lower-priced markets less than 1. As such, the magnitude of the scale has no intuitive interpretation. The variation in RHPI across cities was similar to that for FMR (Fig. 2 in the appendix).
We used a measure of housing subsidies, constructed by Curtis (2007) and applied by Curtis and Waldfogel (2009) and Curtis (2011), that characterizes the availability of subsidized housing (ASH) in 1998 at the MSA level as the proportion of subsidized units available per household with incomes at or less than 50 % of area median income. It includes project-based assistance (public housing), tenant-based assistance (certificates and Section 8 vouchers), and low-income housing tax-credit units. ASH varied widely across cities, from a low of about 2 % to a high of more than 18 % (Fig. 2).
Expected Effects
Based on the theory and empirical literature reviewed earlier, we hypothesize that poor infant health will increase homelessness (traditionally defined) and that the magnitude will increase with severity of the shock. O’Flaherty’s framework would predict that the adverse effect of a family health shock on homelessness is stronger for individuals living in areas with high housing costs and low housing subsidies. We would expect weaker interactive effects with housing subsidies than with housing costs because while everyone confronts housing markets, subsidies are not an entitlement and thus directly affect only those who know about them, are eligible, apply, and rise to the top of waitlists. In addition, vouchers (one component of overall housing subsidies) could potentially increase homelessness by driving up rents facing nonrecipients (Susin 2002).
Given the lack of previous research incorporating the broader conceptualization of homelessness, we do not have prior expectations vis-à-vis the relative magnitudes of estimated effects across the four outcomes, or even the expected direction in certain cases. EMM and DUNR are considered precursors to, or risk factors for, homelessness; as such, we would generally expect health shocks to increase the likelihood of experiencing these situations, with the same general pattern across measures of infant health as for the traditional measure of homelessness. However, it is important to consider the expected effects for the various outcomes in light of the relevant competing risks. For the traditional measure of homelessness, the competing risk is not having experienced homelessness (but potentially having experienced EMM or DUNR). For the expanded measure of homelessness, the competing risk is having lived independently (that is, not having experienced any of the unfavorable housing situations). For EMM and DUNR, the competing risks are not experiencing those situations but possibly having experienced one or more of the others. For example, for EMM, the competing risk includes having lived independently, having doubled up and not paid rent, or having been homeless, but not having experienced EMM.
Overall, we have clear expectations for the direction of the effects of infant health shocks on both the traditional and expanded measures of homelessness (i.e., positive). However, for EMM and DUNR, our expectations regarding relative magnitudes, and even direction, of the effects are ambiguous by virtue of the heterogeneous nature of the competing risks. Nevertheless, it is useful to separately consider the individual components of the expanded measure of homelessness to understand how and why the estimated effects for the traditional and expanded measures of homelessness—our main outcomes—may differ. The relative magnitudes across the four outcomes will also depend, at least in part, on which transitions are affected most. For example, the shock could push individuals with unstable housing (EMM or DUNR) into homelessness and/or those with stable housing into an unstable situation or homelessness. As such, the relative magnitudes across outcomes would depend on which group (the already vulnerable or the less vulnerable) is more strongly affected.
The expected effects for DUNR are further complicated by the fact that this arrangement requires not only the need or desire for coresidence on the part of the visiting family but also the availability of friends or relatives who are willing to host them. Taking in a family with a young child can be a large sacrifice for the host family (including potential loss of public assistance) and may be even more of a sacrifice when the child has health problems. Having a child in poor health, therefore, would be expected to increase the neediness of the child’s family but decrease their chance of getting support in the form of a free place to live. The net effect for this outcome will depend on the strength of these countervailing forces.
A final consideration is that although DUNR is considered an undesirable situation under the HEARTH Act from the standpoint of service provision (because it is often a precursor to homelessness), doubling up even without paying rent may not always represent an unfavorable situation for families. Extended families are often formed as a result of hardship (Angel and Tienda 1982; Hogan et al. 1990), and for most people, independent residence is the preferred arrangement (Cohen and Casper 2002; Ruggles 1996). However, some studies have found favorable effects of living in multigenerational households for children (e.g., DeLeire and Kalil 2002). The advantages and disadvantages of living in extended-family households are complex and no doubt vary across family members and situations.
Analyses
Preliminary Analyses
We compared characteristics of the mothers included in our sample to those (from the 4,898) who were not included and found that the two groups were very similar in terms of marital status, education, and Medicaid coverage for the birth. That is, those in the sample and those excluded because of missing data items (particularly from medical records) were very similar in terms of observed socioeconomic disadvantage (Online Resource 1, Table S1).
We considered whether attrition from the study differed by infant health and how those patterns might affect our estimated effects of poor infant health on homelessness. We compared the health at birth of children in our sample with that of children whose mothers did not complete the three-year follow-up and found that those in our sample were significantly less likely to have a SIHC, SIHC/VLBW, and MSIHC (Online Resource 1, Table S2). Given that we may be losing some children with housing problems owing to poor health, this finding suggests that adverse effects of poor infant health on homelessness-related outcomes—should we find any—would be underestimates.
To explore the exogeneity of poor infant health, we compared sociodemographic characteristics of mothers of healthy children to those of unhealthy children (Table 6 in the appendix). If the measures of poor infant health are indeed random, they should be unrelated to most maternal characteristics. We found that the measures were not significantly associated with maternal race/ethnicity, relationship status, education, Medicaid use, parity, nativity, or employment. The only significant differences were for maternal age (mothers of unhealthy children were slightly older) and census-tract poverty (mothers of children with MSIHCs lived in slightly wealthier tracts). Overall, these results provide convincing evidence that our measures of poor infant health have large exogenous components.
Multivariate Analyses
Table 2 presents the estimated effects of each of the three measures of poor infant health on each of the four housing outcomes. The figures in each cell are from a separate probit model that includes all variables classified as maternal or child characteristics in Table 1, plus city indicators and corresponding baseline controls for the particular outcome (see table note). In each cell, the probit coefficient of the effect of poor infant health is in the top row; the standard error of the probit coefficient, corrected for city clustering of observations using the Huber-White method, is in parentheses; and the marginal effect is in brackets. (Full multivariate results based on the SIHC measure are presented later in the article, and those based on SIHC/VLBW and MSIHC are in Online Resource 1, Tables S3 and S4.) The similarity of the marginal effects and significance levels in Table 2, which are based on models that include a large set of covariates, to those from the corresponding unadjusted models (see Table 7 in the appendix, along with the relevant cross-tabulations) underscores that we have been successful at capturing shocks with large random components and provides reassurance that the multivariate models have not overtaxed the data (because some of the outcomes and infant health measures are relatively rare).
For the most severe measure of poor infant health (SIHC) and the most severe outcome (homeless, traditional), the shock has a strong and significant positive impact: the marginal effect is 6 percentage points, which represents a very large increase relative to the sample mean of 0.03 for this outcome. The other two measures of health shocks have positive and significant effects as well. As expected, the magnitude decreases as the severity of the shock decreases, with marginal effects of 0.03 and 0.01 for SIHC/VLBW and MSIHC, respectively.
The health shocks do not have significant effects on either measure of “pre-homelessness” (EMM, DUNR), although the estimated effects for EMM are positive; additionally, the two most severe health measures have t values greater than 1. The estimated effects for DUNR are negative but not significant, perhaps reflecting opposing effects of increased need and decreased availability. In both cases, the estimates decrease in magnitude as the shock becomes less severe, as expected. The findings are insensitive to the use of two or four moves as the cutoff in the measure of EMM (not shown).
The estimated effects of poor infant health on the expanded measure of homelessness are of smaller magnitude (relative to the relevant sample mean) than the effects on homelessness as traditionally defined. They also exhibit the expected decline with severity of the shock, and MSIHC is no longer statistically significant. Notably, the “bottom line” inferences vis-à-vis the effects of infant health shocks on homelessness are the same whether the traditional or expanded measure is used.
Table 3 presents the full multivariate results (except for city indicators, which are included in all models) that correspond to the first row in Table 2. That is, the estimates pertain to SIHC and each of the four outcomes. The child’s age, which is equivalent to the time interval between the mother’s baseline and three-year interviews, is positively associated with both homelessness and EMM and negatively associated with DUNR. Maternal age is negatively associated with EMM and the expanded measure of homelessness, consistent with prior work suggesting that older mothers are better able to maintain independent households (Curtis 2007). Compared with non-Hispanic white mothers, non-Hispanic black mothers are more likely to become homeless but less likely to experience EMM, and Hispanics are less likely to experience EMM and are more likely to DUNR. Immigrant mothers are less likely than native-born mothers to experience homelessness. Education is negatively related to EMM, and Medicaid use (proxy for poverty) is positively associated with both homelessness and EMM. Mothers who were employed prior to the birth are more likely than those who were not employed to experience EMM and less likely to DUNR. Mothers who were married or cohabiting at the time of the birth were much less likely than their nonmarried, noncohabiting counterparts to DUNR. As expected, mothers who did not complete the one-year survey were more likely to experience all the homelessness-related situations. Prenatal mental illness is positively associated with the expanded measure of homelessness, suggesting that long-term struggles with mental health issues may generally erode the ability to remain stably housed over time. The measures of poor housing quality and doubling up during pregnancy are positively and significantly associated with the relevant outcomes. The city indicators were jointly and highly significant in all models (not shown), suggesting that residential location is an important predictor of homelessness. The covariate estimates were very stable across the measures of poor infant health, which can be seen by comparing the estimates from Table 3 with those in Tables S3 and S4 of Online Resource 1.
The Role of Housing Markets
FMR is our preferred area-specific housing market characteristic: as discussed earlier, the RHPI has no intuitive interpretation, and we would expect weaker interactive effects of the infant health shock with ASH than with measures of rental prices. In Table 4, we present estimates from adjusted and unadjusted linear probability models that interact SIHC (the most exogenous measure) with FMRs to predict each measure of homelessness (traditional and expanded). The adjusted models control for all covariates in Table 3 (with the exception of city indicators) plus city-level unemployment rate in 2000. The unadjusted estimates are presented to alleviate concerns about power. Let β1 be the coefficient on SIHC, β2 be the coefficient on FMR, and β3 be the coefficient on the interaction term. For each $100 increase in FMR, the probability of homelessness will increase by β2 if the child does not have a severe health condition and by β2 + β3 if he/she does. Figure 1 displays these interactive effects in graphical form, with predicted probabilities of homelessness at specific percentiles of FMR assuming mean values of all covariates. Corresponding graphs for adjusted and unadjusted MSIHC × FMR interactions are presented in Fig. S1 in Online Resource 1, and graphs for adjusted and unadjusted SIHC × RHPI, MSIHC × RHPI, SIHC × ASH, and MSIHC × ASH interactions are in Online Resource 1, Figs. S2–S5.
Figure 1 clearly shows that living in a city with higher FMR is positively associated with homelessness and appears to exacerbate the adverse effects of an infant health shock. For families experiencing a SIHC, the probability of homelessness (traditional) increases almost sixfold across the rental distribution, from about 3 % to almost 18 %; it increases much less so (from about 2 % to 4 %) for families with healthy children (panels a and c). For the expanded measure of homelessness, the probability of homelessness more than doubles across the rental distribution for families experiencing a SIHC, but does not increase with FMRs for families with healthy children (panels b and d). The interactions between SIHC and homelessness (either measure) were statistically significant at conventional levels, regardless of whether we adjust for covariates. When we use MSIHC as the infant health measure, we find that although the probability of homelessness increases as rents increase, the interaction effect of rents with MSIHC is not significant (Online Resource 1, Fig. S1). When we use RHPI as the measure of housing costs, we find the same pattern of interactive effects as when using FMR (Online Resource 1, Figs. S2 and S3). We find no evidence that subsidized housing availability, as we have characterized it, buffers the adverse effects of poor infant health on homelessness (Online Resource 1, Figs. S4 and S5); the relevant interactions never reached statistical significance.
Overall, these results—which are designed to test a theory rather than to isolate causal effects of area-specific factors—support O’Flaherty’s proposition that homelessness results when adverse individual circumstances and housing markets collide. The pattern of estimates strongly suggests that housing prices are an important contributor to homelessness, particularly when coupled with an unexpected life challenge. The results do not lend strong support to the proposition that housing subsidies play an important buffering role.
Supplemental Analyses
Our main analyses rely on the assumption that infant health shocks cause homelessness rather than the other way around. As “falsification tests,” we estimated effects of poor infant health on the mother’s pre-delivery housing situation. The logic is that a shock that takes place at the time of the birth cannot affect the mother’s housing situation or living arrangements prior to the shock; finding significant associations would indicate spurious correlation. We estimated models with each baseline housing measure as an outcome, as a function of poor infant health and all covariates in Table 3 except city indicators, baseline housing controls, and age of child (Online Resource 1, Table S5). We found no evidence that any measure of poor infant health is associated with prenatal housing or living arrangements, validating that we were successful at characterizing poor infant health as an exogenous shock. These results also suggest that housing arrangements at the time of the birth are not a function of problems with the pregnancy that may result in poor infant health.
We used the three-year observation window for two reasons. The first involves a qualitative judgment about what the appropriate time horizon would be for studying the effects of the shock on the outcome. In that regard, it seems to us that one year might be too short, that five years might be too long, and that nine years would certainly be too long. The second issue is a more practical one. As we discussed earlier, attrition from the study appears to be systematically related to the life shock of having a child born in poor health. Attrition is also likely related to housing problems or homelessness because people with less-stable housing tend to be more difficult to locate and interview. We conducted supplementary runs estimating the traditional measure of homelessness at one or three years and at one, three, or five years (Online Resource 1, Table S6) and obtained estimates similar to the relevant cells in Table 2, providing some evidence that three years is the most appropriate time horizon and that there would be little benefit to going out nine years.
We assessed the sensitivity of our main estimates to a number of additional model specifications (Online Resource 1, Tables S7–S10): (1) restricting the sample to mothers who participated in all interviews; (2) estimating models that excluded cities with <100 cases, excluded city indicators, or excluded the baseline housing measures; (3) estimating seemingly unrelated regression (SUR) models, which simultaneously estimated the four outcomes, and linear probability models; and (4) estimating models with measures of local housing costs, availability of subsidized housing, and city-level unemployment in place of the city indicators. For (1), the marginal effects of SIHC were somewhat smaller than those in Table 2 for both measures of homelessness, although they were less precisely estimated because of much smaller samples. In all other cases, the estimated effects of poor infant health on the various housing outcomes were at least as large as, or highly consistent with, the corresponding probit estimates in Table 2.
Conclusion
This study exploited an exogenous health shock—the birth of a child with a severe health condition—to investigate the effect of that shock on the probability of homelessness during the child’s first three years and the extent to which the shock interacts with housing market conditions. We found robust evidence that the shock substantially increases the likelihood that the family experiences homelessness, particularly in cities with high housing costs. We considered a traditional measure of homelessness, two measures of housing instability thought to be precursors to homelessness (eviction or frequent residential moves, doubling up and not paying rent), and a combined measure that approximates the broadened conceptualization of homelessness under the 2009 HEARTH Act. We found the shocks to have positive but weak effects on eviction or frequent residential moves and no aggregate effects on doubling up and not paying rent, suggesting that the shocks are more likely to drive individuals with unstable housing into homelessness than to drive individuals from stable to unstable housing situations. Future research, which would require detailed histories that include both shocks and transitions between various housing situations, is needed to elucidate the pathways.
The weak, negative associations of a life shock with doubling up and not paying rent deserves particular scrutiny because that arrangement—even more so than the other outcomes considered—reflects a complicated interplay of incentives. Future research clarifying the role of public support in setting the conditions for private support that is offered in the form of a place to live would likely be fruitful. According to He et al. (2010), many federal assistance programs (including SSI, food stamps, and Section 8 vouchers) impose substantial implicit taxes on shared housing; as such, it is possible that families or friends who would otherwise be willing to provide a place to live to those confronting shocks may be dissuaded from doing so, or that those confronting the shocks are deterred from doubling up. Because doubling up, even without paying rent, is not always an unfavorable situation, some families experiencing life shocks may benefit from this arrangement without conferring offsetting decrements in well-being to their host families.
Our findings add to a growing body of evidence that housing markets are an important contributor to homelessness. We found consistent evidence that housing prices exacerbate the effects of a life shock on homelessness, but little evidence that generosity in terms of public housing subsidies buffers the adverse effects of the shock, perhaps because (1) there could be mismatch between those getting subsidies and those at risk of homelessness, (2) cities or other geographic units may be more appropriate than MSAs for characterizing housing subsidy availability, or (3) the voucher component of overall housing subsidies may drive up rents for the potentially homeless and dominate offsetting benefits of other types of subsidies. Because subsidies are an important supply-side tool in combating housing insecurity and homelessness, particular attention should be given to which policies work, for whom, and under what circumstances.
As far as we know, this is the first study to explicitly test O’Flaherty’s theory that homelessness results from a conjunction of adverse individual circumstances and housing markets, perhaps because longitudinal population-based data sets that include measures of both life shocks and homelessness, that can be appended with area-specific information, and that have sufficient sample sizes are rare. As such, it makes an important contribution to the small literature on the economics of homelessness. By focusing on a life shock that has been shown to have adverse effects on employment, we found indirect evidence that income shocks are a key factor affecting homelessness. By considering interactions with relevant area-specific characteristics, we found evidence consistent with O’Flaherty’s theory that adverse individual circumstances and high housing costs have synergistic effects. That said, this study was based on one type of life shock, did not explore underlying mechanisms, and cannot be used to obtain precise causal estimates involving specific housing market characteristics. Thus, the findings need to be complemented with those from studies of health shocks to adults; other types of life shocks; potential underlying pathways, such as work and relationship dissolution; and causal main and interactive effects of area-specific characteristics. Our attempt to capture the broadened conceptualization of homelessness under the HEARTH Act represents an important addition to research on homelessness and makes our study current and policy relevant. Although our approximations of the HEARTH Act components are not perfect, they represent a first and reasonable attempt at characterizing the expanded definition of homelessness and can serve as an important baseline attempt to inform future research and data collection.
Acknowledgments
This research was supported by Grants #R01-HD-45630 and #R01-HD-35301 from the National Institute of Child Health and Human Development. We are grateful for helpful input from Susan Averett, Dan O’Flaherty, Amy Crews Cutts, participants at Lafayette College Economics Department Seminar Series, participants at 2011 NBER Spring Health Economics Workshop, participants at Center for Homelessness Prevention Studies Grand Rounds at Columbia University, and participants at the Economics seminar series at the University of Iceland, as well as for valuable assistance from Oliver Joszt and Taťána Čepková.
Appendix 1
The coding of abnormal conditions in the FFCWB data was designed to identify cases that were at least moderately severe, unlikely caused by prenatal behavior, had a poor long-term prognosis, and were present at birth. A pediatric consultant was directed to glean information from the medical records (augmented with one-year maternal reports) and to assign all infant conditions a number between 1 and 16 according to the grid shown in Table 5. After giving the consultant the grid and clear instructions, the investigators had no further input into how particular conditions were coded. If a child had multiple conditions, each condition was assigned a separate number.
Severe Infant Health Condition (SIHC) was coded as a 1 (yes) if the child had a health condition in cell 1. Examples of conditions in cell 1 are microcephalus, renal agenesis, total blindness, and Down syndrome.
Severe Infant Health Condition or Very Low Birth Weight (SIHC/VLBW) was coded as a 1 (yes) if the child had a condition in cell 1 or the child was very low birth weight (<1500g).
Moderate or Severe Infant Health Condition (MSIHC) was coded as a 1 (yes) if the child had a condition in either cell 1 or cell 2. Examples of conditions in cell 2, which are considered random at birth but may or may not have long-term health consequences, are malformed genitalia, hydrocephalus, shoulder dystocia, pneumomediastinum, and webbed fingers or toes.
Example of high-severity conditions considered possibly related to parents’ behavior are cerebral palsy (cell 5) and likely related to prenatal behavior are fetal alcohol syndrome (cell 9). These conditions are not coded as a 1 in the preceding measures.
