Modification of Housing Mobility Experimental Effects on Delinquency and Educational Problems: Middle Adolescence as a Sensitive Period.

Abstract

Residential mobility is one documented stressor contributing to higher delinquency and worse educational outcomes. Sensitive period life course models suggest that certain developmental stages make individuals more susceptible to the effects of an exposure, like residential mobility, on outcomes. However, most prior research is observational, and has not examined heterogeneity across age or gender that may inform sensitive periods, even though it may have important implications for the etiology of adolescent development. Moreover, there are important translational implications for identifying the groups most vulnerable to residential mobility to inform how to buffer adverse effects of moving. In this study, low-income families were randomized to residential mobility out of public housing into lower poverty neighborhoods using a rental subsidy voucher (“experimental voucher condition”), and were compared to control families remaining in public housing. The sample was comprised of 2829 youth (51% female; 62% Non-Hispanic Black, 31% Hispanic, 7% other race). At baseline, youth ranged from 5 to 16 years old. This study hypothesized that random assignment to the housing voucher condition would generate harmful effects on delinquency and educational problems, compared to the control group, among boys who were older at baseline. The results confirmed this hypothesis: random assignment to the experimental voucher condition generating residential mobility caused higher delinquency among boys who were 13 to 16 years old at baseline, compared to same-age, in-place public housing controls. However, residential mobility did not affect delinquency among girls regardless of age, or among boys who were 5 to 12 years old at baseline. The pattern of results for educational problems was similar but weaker. Families with teenage boys are particularly vulnerable to residential transitions. Incorporating additional supports into housing programs may help low-income, urban families to successfully transition to lower poverty neighborhoods.

Introduction

The deleterious effects of living in socially disadvantaged neighborhoods are well-recognized. Individuals living in areas with a high degree of poverty, crime, and other social ills display higher rates of health related problems, emotional issues such as depression, higher unemployment, and higher rates of crime, victimization, and drug use (Duncan and Kawachi 2018; Sampson et al. 2002). It is, therefore, not surprising that one posited solution to the effects of living in disadvantaged areas on individuals’ physical, economic, behavioral and emotional well-being is to facilitate movement out of those areas (Goering and Feins 2003; Briggs 1997; Keels et al. 2005). The Moving to Opportunity program is illustrative of programs designed to combat the adverse effects of living in high poverty areas (Goering and Feins 2003). Responding to the distressed conditions found in the federally funded housing projects built in the 1960’s, the Moving to Opportunity program provided housing vouchers for families to access higher opportunity neighborhoods, including neighborhoods where less than 10% of households in the census tract lived below the poverty level – this was the experimental condition hypothesized to promote the best outcomes. The program was designed as an experiment to evaluate a comparison of three groups: those who received a rental voucher to move from the distressed area to an apartment in a low-poverty neighborhood (a constrained voucher move), with those who received a rental voucher to subsidize rent in any neighborhood (an unconstrained voucher move), with those who did not receive a voucher but remained eligible to continue living in public housing (the control group).

Findings from the Moving to Opportunity experiment are mixed. Limited research has found beneficial effects of moving on household economic outcomes, but only among children who were aged 12 or younger at baseline (Chetty et al. 2016). A much larger set of studies have documented positive effects on health, such as obesity, diabetes, self-reported physical limitations, psychological distress, and major depression (Ludwig et al. 2012; Ludwig et al. 2011; Orr et al. 2003; Sanbonmatsu et al. 2011; Sanbonmatsu et al. 2012). However, one predominant conclusion from the evidence to date suggests that the program has not been beneficial for everyone. Certain groups did not experience the positive effects of relocation, while others experienced detrimental effects of relocation (Kling et al. 2007; Orr et al. 2003; Osypuk et al. 2012b; Osypuk et al. 2012a; Kessler et al. 2014; Sanbonmatsu et al. 2011).

The question arises, why does a program designed to ameliorate the adverse effects of distressed neighborhoods have a problematic effect on some of those who should benefit from moving? It is possible that moving to a better neighborhood facilitates improvement in financial situations and overall health of the family, but the move itself may be problematic for children because of the adverse effects of residential mobility. A sizeable literature has documented that residential mobility is a stressor associated with worse youth developmental outcomes (Leventhal and Newman 2010). However, this literature largely relies on observational studies (Jelleyman and Spencer 2008; Leventhal and Newman 2010), which may limit the ability to infer that residential mobility causes problematic outcomes such as lower educational attainment and delinquency. Importantly, life course models suggest that individuals may be more susceptible to the effects of a certain exposure at different points in the life course, called sensitive periods, which are common in behavioral development (Ben-Shlomo and Kuh 2002). Dramatic changes occur during adolescence, which may leave youth more susceptible to social, psychological, and biological factors that influence development, compared to other developmental stages (Kuh et al. 2003). Therefore, the effects of residential mobility may be most harmful during adolescence (Tønnessen et al. 2013; Pettit 2004; Fauth et al. 2005). If residential mobility effects on education and delinquency are heterogeneous by age and/or gender, this may have important implications for targeting prevention efforts at the most vulnerable age and gender groups. However, scarce literature exists testing whether residential mobility may be more or less harmful for certain age and gender groups. To address these gaps in the literature, this study conducted a secondary data analysis on an existing experimental housing voucher relocation program to assess whether there are developmentally sensitive periods where moving is more or less important for predicting educational outcomes and delinquency, and whether these patterns differ for boys and girls. General Strain Theory provides a framework to guide this study.

General Strain Theory

Moving may be beneficial if it substantially improves living conditions, or it may represent a substantial stressor due to disruption and change. The possible contradictory and heterogeneous effects of moving by gender and age, can be understood when viewed through the lens of General Strain Theory (Agnew 2006). Agnew recognized that strain resulted from more than just the failure to achieve positively-valued goals, as suggested by Merton and other theorists employing similar perspectives (Merton 1938; Cloward and Ohlin 1960; Cohen 1955). Strain was also the result of an inability to escape negative stimuli and/or the removal of positively valued stimuli. An example of the inability to escape negative stimuli related to living in a distressed area would be the strain girls experienced by being subjected to regular sexual harassment and victimization (Popkin et al. 2008). From a general strain theory perspective, allowing girls to escape this negative stimulus reduced their strain. The benefits of escaping sexual harassment and other contextual stressors outweighed the strain of moving, as evidenced by the beneficial effects of being randomized to one of the voucher treatment groups on lower risk of mental health and substance use. An example of the removal of positively valued stimuli that may have occurred for those in the voucher groups would be the loss of a valued friend or father figure that boys experienced because of relocation (Clampet-Lundquist et al. 2011). Losing this positive stimulus means boys were likely to experience negative emotional states leading to problematic behavior. This manifested in boys who moved as inferred from their worse outcomes, compared to controls, on mental health, substance use, and potentially education.

When one considers all three sources of strain posited by Agnew, e.g., failure to achieve a positive goal, inability to escape negative stimuli, or removal of positive stimuli, one might conclude that the overall family’s health and well-being might improve by relocating to a better neighborhood. Yet the situation for boys, older boys in particular, might not be as sanguine, and the relationships for both delinquency and educational outcomes may be more nuanced than if one expected the treatment to affect everyone equally. Developmental and life-course theories (Thornberry and Krohn 2001; Farrington 2005) emphasize the importance of the youth’s developmental stage for moderating exposures, suggesting adolescence is a sensitive period for exposures to be more influential for outcomes than at younger ages (Ben-Shlomo and Kuh 2002; Kuh et al. 2003). Friendships and the influence of friends becomes particularly important in predicting outcomes during middle adolescence (Thornberry 1987). During this time, youth begin to assert their independence from their parents and, at least in terms of day to day issues, rely on friends for advice and companionship. Thus, separating youth from their friends would be particularly problematic at this developmental stage, and may result in anger and frustration as they attempt to establish new social networks in a different neighborhood. Moreover, youth may perceive delinquent behavior as a way to garner favor in the new network. Evidence supports this idea, suggesting that treatment group boys were more likely to fall in with riskier peer groups and have friends who used drugs, because that may be who accepted them in the new neighborhoods (Schmidt et al. 2017b; Clampet-Lundquist et al. 2011). Educational performance might also be impacted by the anger and frustration youth may feel about being relocated. Once youth enter high school, educational performance and school-related activities become increasingly important for youth. Removing youth from the established routine of the educational system to which they have become accustomed, and reestablishing them in a new system, may adversely affect their performance in, and commitment to, education. Again, one might expect the effect of relocation to be greater on older adolescents who recognize that high school is coming to an end and preparation for what comes next is increasingly important.

In summary, General Strain Theory and the developmental and life course literature jointly provide a framework for understanding how residential mobility may impact positive and negative stimuli, which may be differentially important for different gender and age groups. Residential mobility may remove adolescent boys from positively valued stimuli (friends and educational routine), thereby increasing strain and the resulting negative emotional states and problematic outcomes. In contrast, mobility may remove adolescent girls from negative stimuli (sexual harassment and victimization), thereby reducing strain and improving outcomes. Children who are younger at baseline may have developed fewer important ties to the neighborhood and to peers, given that younger children have less autonomy and the family is more central at this developmental stage. Thus younger children may better adapt to mobility than older children. The prior empirical findings from Moving to Opportunity are consistent with General Strain Theory, particularly with respect to gender, for other outcomes including substance use, risky behavior, and mental health outcomes (reviewed below); however few prior studies have jointly considered both gender and age in understanding differential treatment effects for any outcomes. This is one contribution of the current research study.

Residential Mobility, Delinquency, and Education

Turning to empirical findings from prior research, a broad literature has found, overwhelmingly, both short- and long-term negative associations between residential mobility and a host of health and developmental outcomes, including behavioral and emotional problems, teen pregnancy, substance use, and mental health (Jelleyman and Spencer 2008; Leventhal and Newman 2010; Brown et al. 2012; Mok et al. 2016; DeWit 1998; Hoffmann and Johnson 1998). A large body of work has focused specifically on the links between mobility and crime and delinquency at both the macro-level and the individual-level (Gasper et al. 2010). At the individual level, studies have established that residential mobility is associated with increases in delinquency (South et al. 2005), behavior problems (Wood et al. 1993; Simpson and Fowler 1994; Adam and Chase-Lansdale 2002), violence (Haynie and South 2005), and peer deviance (Haynie et al. 2006a). Moreover, mobility may increase recidivism among youth who were adjudicated for a juvenile offense, however this relationship differs for boys and girls depending on the type of neighborhood (Wolff et al. 2017). The largest effect of mobility on boys’ recidivism was seen for moves to better neighborhoods, while the largest effect on girls’ recidivism was seen for moves to worse neighborhoods (Wolff et al. 2017). Yet, a few recent studies find no effect of mobility on delinquency (Gasper et al. 2010; Porter and Vogel 2014).

Relocation may also adversely affect education (Leventhal and Newman 2010). Residential mobility is associated with lower educational achievement (Adam and Chase-Lansdale 2002), repeating a grade (Simpson and Fowler 1994; Wood et al. 1993), suspension or expulsion (Simpson and Fowler 1994), lower test scores (Pribesh and Downey 1999), and school drop out (Astone and McLanahan 1994; Haveman et al. 1991). Hango (2006), however, found a positive long-term relationship between mobility (from birth to age 15) and high school completion. The discrepancy between this finding and prior work may be because families in this study made high-quality moves to better housing and/or neighborhoods that offset the stress of moving (Hango 2006), or it may be a methodological artifact.

Methodological limitations of prior work (Haelermans and De Witte 2015) mean the causal relationship between mobility and delinquency and education is unclear, and may explain some of the mixed findings of mobility on delinquency and education. Observational studies that employ traditional regression-based methods of covariate adjustment represent the bulk of the mobility literature (Leventhal and Newman 2010; Jelleyman and Spencer 2008), yet they often fail to control for important threats to internal validity that may bias results. Residential selection is particularly salient for studies involving residential mobility, where families choose whether and where to move, yet this selection is seldom modeled or fully controlled (Oakes 2004; Sampson 2008). In particular, poor and minority households move often, and to similarly poor neighborhoods (Quillian 2003; South and Crowder 1997), thus the effects of residential mobility may be confounded by family socioeconomic status or race/ethnicity, among other factors related to residential preferences or mobility constraints.

Studies applying stronger designs to control for confounding by factors predicting residential selection generally find mixed effects of mobility on delinquency and education. For example, using fixed and random effects models, Gasper and colleagues concluded that any association between mobility and delinquency is a result of unobserved individual and/or family characteristics that are different between mobile and non-mobile youth, and not of mobility itself (Gasper et al. 2010). Likewise, a study applying propensity score matching to control for selection by matching movers and non-movers found no effect of mobility on delinquency after matching (Porter and Vogel 2014). However, another study applying propensity score matching found that mobility was associated with recidivism of delinquent youth (Wolff et al. 2017). Yet another study, applying nearest-neighbor matching to create a comparison group for examining the association between mobility and school drop out, found that mobility increased the likelihood of dropping out in the first few years after the move, and then again 6 years after the move, but lowered the risk in the middle time period (Haelermans and De Witte 2015).

Although recent studies have made important advances to more strongly model residential selection when examining effects of residential mobility, they are not panaceas. In contrast to intervention studies that generate change in the types of moves as part of the design, within-person observational studies rely on the naturally-occurring variability in residential mobility over time. However, it is commonly observed that when households move, these moves are often lateral (Oakes 2014), weakening the ability to make strong inference about the effects of moves to different places, particularly in differentiating moves from low to high opportunity areas from other mobility trajectories. Aside from interventions, analytic features like propensity score matching can also enhance causal inference by better controlling for confounding, and ensuring positivity and avoidance of off-support inference. Yet propensity score matching can only control for measured, not unmeasured, confounders. Notably, some of the limitations of prior work do not stem from design or analytic issues, but from failure to test developmental heterogeneity. Indeed, although some of the observational evidence suggests variation in the consequences of residential mobility by gender (Wolff et al. 2017; Haynie et al. 2006b; DeWit 1998), there is a gap in the literature with respect to testing for heterogeneity by age. Experimental studies, which we discuss next, overcome several of these serious threats to internal validity, and may therefore provide the strongest evidence on how residential mobility causes delinquency and educational outcomes, and how these effects vary by gender and age.

Experimental and Quasi-Experimental Effects of Residential Mobility

A few studies have tested the effects of allocation to residential mobility programs on youth outcomes, demonstrating beneficial effects of residential mobility. Gautreaux was a court-ordered residential mobility program aimed at desegregating Chicago public housing projects. This program helped volunteer families move from public housing to primarily white, suburban neighborhoods, or to neighborhoods within the city of Chicago, using a Section 8 housing voucher along with housing counseling (Keels et al. 2005). Evidence from Gautreaux showed that youth in families who were offered the opportunity to move to a rental unit in suburban areas exhibited better educational outcomes (Rosenbaum 1995) and lower homicide mortality rates (Votruba and Kling 2009), compared to families who were offered the opportunity to move to rental units within the city of Chicago.

Other quasi-experimental studies have evaluated scattering public housing residents (using inclusionary zoning policies) and/or scattering public housing project sites (vs. concentrating them). For example, the Yonkers project, a court-ordered desegregation program in Yonkers, NY, offered families rental units in scattered-site, low-rise newly built housing in lower-poverty areas of the city. Youth in families who moved reported fewer behavior problems and delinquency, compared to those who remained in high-poverty public housing developments (Fauth et al. 2005). Like Yonkers, the Monitoring Mt. Laurel Study evaluated the effects of moving into a newly built public housing project in a low-poverty, predominantly white, suburban town (Mt. Laurel, NJ), compared to applying to but not being accepted to move into this housing. Using propensity score methods to improve the comparability of the treatment groups, the adolescent children who moved into suburban affordable housing exhibited improved school quality (e.g., higher graduation rates, SAT scores), reduced exposure to violence and social disorder at school, and marginally increased the time children spent reading (Casciano and Massey 2012b). However, there were no total effects of the affordable housing program on grade point average (GPA), and results of the program on negative life events (e.g., serious illness, deaths, imprisonment) were conflicting across two different analyses (Casciano and Massey 2012b, 2012a). In wealthy suburban Montgomery County, MD, researchers tested differences among residents who were offered via lottery a placement in a home that was scattered among market-rate housing, or a placement in a home clustered in small public housing projects (Pollack et al. 2014). The school district used neighborhood-based attendance zones, so this housing treatment also translated to randomly assigning school context. Children assigned to attend low-poverty elementary schools performed better than peers in moderate-poverty schools in math and reading, after 5–7 years (Schwartz 2010).

Aside from the Montgomery County study, these studies lacked a randomly-allocated control group, or a strong non-experimental comparison group, and thus study findings may still be confounded by selection bias (Acevedo-Garcia et al. 2004). The Monitoring Mt. Laurel study additionally had limited power with small sample sizes, and experienced differential attrition across treatment groups, given the difficulty in identifying nonresident families who applied but did not succeed in receiving an offer to move there. The intervention design of these quasi-experimental studies strengthens their causal inference, but other threats to internal validity persist. The Moving to Opportunity experiment remains the strongest study available for testing the effects of residential mobility (Acevedo-Garcia et al. 2004).

The Moving to Opportunity study was designed and implemented after promising effects emerged from the quasi-experimental Gautreaux program (Goering and Feins 2003). Its main goal was to combat the effects of very distressed, high poverty housing projects, by randomly-assigning housing vouchers to subsidize rent in lower poverty neighborhoods, compared to in-place public housing controls (Briggs et al. 2010). Moving to Opportunity’s randomized design controls for all confounders, whether measured or unmeasured, except those imbalanced by chance. The Moving to Opportunity treatment, in essence, is a residential mobility effect that captures all aspects of moving: changing housing, neighborhoods, schools, and social networks, contrasted with remaining in highly distressed public housing. Therefore, this is a strong design for causal inference. Findings from the Moving to Opportunity study have been only partially consistent with prior work on the effects of residential mobility.

Gender.

Gender heterogeneity in the effects of Moving to Opportunity is well established. Mobility negatively affected multiple outcomes for boys, including psychological distress, behavior problems, risky behaviors, substance use, and asthma, which aligns with the hypothesis that residential mobility is a stressor for boys. Yet mobility positively impacted girls’ outcomes including psychological distress, depressive symptoms, and substance use (Orr et al. 2003; Kling et al. 2007; Osypuk et al. 2012b; Osypuk et al. 2012a; Schmidt et al. 2014; Schmidt et al. 2017a), suggesting that the destination of the move may have offset the stress of the move for girls. Although early follow up results (2 years after random assignment) in New York and Boston showed promise, with initial reductions in treatment group boys’ behavior problems (Leventhal and Brooks-Gunn 2003; Katz et al. 2001), girls’ violent and property crimes (Kling et al. 2005), and violent crime arrests for both genders (Ludwig et al. 2001), these effects did not persist at the interim follow up (4–7 years later) when all five sites were evaluated. The effects of Moving to Opportunity on adolescent delinquency were largely nonsignificant at the interim and final evaluations for both boys and girls (4–7 years and 10–15 years after random assignment, respectively) (Orr et al. 2003; Kling et al. 2005; Sciandra et al. 2013; Sanbonmatsu et al. 2011; Graif 2015). One study used Moving to Opportunity to test whether children who moved to clusters of disadvantaged neighborhoods would exhibit more delinquency. This study found that boys who moved to neighborhoods of locally concentrated disadvantage (i.e., disadvantaged neighborhoods surrounded by non-disadvantaged neighborhoods) exhibited higher delinquency than control group boys; however, the study documented no effects of the intervention on delinquency for either gender in intention-to-treat analyses and they omitted youth who were 12 to 14 years old from their analysis (Graif 2015).

Findings with respect to education were similar, insofar as early Moving to Opportunity evidence documented beneficial effects initially that did not persist in later evaluations. Early evaluation results in New York noted improvements in treatment group boys’ achievement scores (Leventhal and Brooks-Gunn 2004), but effects on achievement test scores, and most other education outcomes, at the interim survey were nonsignificant when examined overall and by gender (Sanbonmatsu 2006; Orr et al. 2003). The one exception indicating harmful effects on treatment group boys’ scholastic performance emerged at the interim survey, where they were less likely to report grades of mostly Bs or higher than controls (Orr et al. 2003).

Age.

Fewer studies have examined Moving to Opportunity program effects by age. The early evaluation documented that the improvements among boys’ achievement test scores were concentrated among children who were 11–18 years old (versus 6–10 year olds) at follow up (Leventhal and Brooks-Gunn 2004), however this study did not look at variation in treatment effects by age at random assignment, and focused on only the New York site. The interim evaluation study investigators did examine program effects on education outcomes among subgroups defined by age at random assignment (5 or younger, 6–11, and 12 or older at baseline), not considering gender. Treatment effects were homogeneous by age at randomization for all outcomes, except that children 12 or older at baseline were more likely to have repeated a grade compared to similar aged control group youth (Orr et al. 2003). Another recent study found that residential mobility in Moving to Opportunity improved college attendance and earnings when youth were younger than 13 years old when their families moved, but mobility had non-significant or negative impacts in the long-term for youth who were 13 or older when their families moved (Chetty et al. 2016). No studies have examined youth delinquency by subgroups of age at randomization.

Gender and Age.

Despite the strong evidence for heterogeneity of Moving to Opportunity treatment effects by gender, and the suggestive evidence that effects may vary by the youth’s age when their family moved, only one study to date has explicitly examined treatment effects by gender and age together. This study tested whether residential mobility had differential impacts on adolescents’ risky behaviors (substance use and risky sexual behavior) depending on the age when children moved. Results supported the conclusion that adolescence may be a sensitive period for residential mobility, where treatment group boys who moved at age 10 years and older exhibited more risky substance use and risky sexual behavior than same-age controls, while treatment group girls 9 years and older exhibited lower risky substance use than same-age controls (Schmidt et al. 2017a). Treatment effects were nonsignificant for boys younger than 10 and girls 10 and older.

Current Study

Both General Strain Theory and research suggest that residential mobility may impact delinquency and educational outcomes, and also that this relationship may vary across gender and age, with adolescence being a particularly sensitive period for the effects of mobility to manifest, compared to younger ages. Yet no research has explicitly tested whether residential mobility effects on delinquency and educational problems differ jointly by both gender and age. This study is a contribution to the literature for several reasons. First, understanding why certain stressors differ by gender or developmental stage informs life course/sensitive period theory and the etiology of adolescent development. Despite that developmental scientists emphasize the contingency of effects, most of the neighborhood effects literature, for example, does not probe gender-specific phenomena. From a sociological perspective, this makes little sense given that gender is considered a master status, and as such is exceptionally important for patterning one’s social identity and interactions. Moreover, because this study is an experimental design, it can inform etiology with a high degree of internal validity, unlike prior residential mobility studies. This study also has important translational implications for buffering adverse effects of residential mobility. Low-income populations such as Moving to Opportunity families are highly mobile, and they are also the target of affordable housing policies in the US (National Low Income Housing Coalition 2011). Knowing whether, when, and how to tailor mobility programs, like housing choice vouchers, to fit families’ needs, and to provide extra support to families with certain vulnerabilities, may help to inform the future of housing voucher policies, including to prevent or cushion any adverse effects of mobility.

Recognizing the various sources of strain suggested by Agnew (2006) and the potential for sensitive periods (Ben-Shlomo and Kuh 2002), this study hypothesized that exposure to residential mobility may have differential effects on individuals depending on both their gender and their developmental stage at the time of the move. Specifically, it is expected that moving between the ages of 13 and 16 (middle adolescence) may occur during a sensitive period for boys where this strain increases delinquency and educational problems, while moving at younger ages will not. Further, this study hypothesized that residential mobility would have little effect on girls’ delinquency and educational problems, regardless of age, because 1) girls already engage in low levels of delinquency (Farrington 1986; Steffensmeier and Schwartz 2009), and 2) the avoidance of negative stimuli afforded by the move outweighs the costs that girls experienced by moving.

Methods

Data and Sample

The Moving to Opportunity for Fair Housing Demonstration Project (MTO) was a randomized controlled trial (RCT) sponsored by the US Department of Housing and Urban Development (HUD) in five US cities (Baltimore, Boston, Chicago, Los Angeles, New York) (Goering et al. 1999). Families volunteered for the experiment, and were eligible if they qualified for rental assistance, had children under 18 years old, and lived in public housing (Feins and McInnis 2001). Public housing authorities in each city drew applicants from the waiting list, evaluated families for eligibility, administered enrollment forms, obtained informed consent, and fielded the baseline survey (Goering et al. 1999). Over 4600 volunteer families were eligible and enrolled in the study. This study analyzes children from these families who were 12 to 19 years old at the interim survey in 2002 (aged 5 to 16 at baseline). Of the 3537 youth randomized, 2829 completed the interim evaluation (N=1950 in treatment group, N=879 in control group), for an 89% effective response rate (Orr et al. 2003).

Random Assignment

Families were randomly assigned to one of three treatment groups using special software on a rolling basis from 1994 to 1998 (Goering et al. 1999); this study analyzes families that were randomized through December 1997. The first randomly-assigned group, called the low poverty treatment group, received a Section 8 voucher that could only be used to subsidize rent of a private market apartment in a neighborhood with less than 10% of families living in poverty; they also received housing counseling to assist families in relocating. The second randomly-assigned group, the Section 8 treatment group, received a traditional Section 8 voucher that could be redeemed to subsidize rent of an apartment in any neighborhood and this group did not receive housing counseling. Families in both voucher groups had 90 days to use the Section 8 voucher, or it was revoked. Finally, the third randomly-assigned group, the control group, could remain in public housing, but was offered no other assistance.

Survey Collection

Public housing authorities administered the baseline survey at the time of random assignment from (1994–1997), the interim survey 4 to 7 years after random assignment (2001–2002), and the final survey 10 to 15 years after random assignment (2008–2010). Surveys were conducted using in-person interviews with household heads and up to two randomly-selected children per family (Orr et al. 2003; Sanbonmatsu et al. 2011). This study uses data from the baseline and interim surveys. The final survey was not used because outcome data was not collected among children at the final evaluation who were between the ages of 5 to 16 at random assignment.

Measures

Delinquency.

Self-reported delinquency, collected in 2002 at the interim survey (Orr et al. 2003), is measured as the fraction of nine items (yes/no) youth reported ever engaging in, with zero indicating the youth engaged in no delinquent acts, and one indicating the youth engaged in all nine. Items include: destroying property, stealing an item worth less than $50, stealing an item worth $50 or more, committing other property crimes (fencing, receiving, possessing stolen property), attacking/assaulting someone, carrying a gun, being in a gang, selling drugs, and getting arrested. Ten youth (<1%) who were missing on five or more items were omitted, leaving an analytic sample of 2819 youth (N=1941 in treatment group, N=878 in control group). Variety scores like this are highly reliable and valid for measuring delinquency (Sweeten 2012).

Educational Problems.

Educational problems was measured as the fraction of four items (yes/no) youth exhibited: disconnection from school, school discipline, truancy, and poor grades. Disconnection from school is a measure of whether the youth self-reported no current school enrollment at the interim survey (versus being enrolled in school or having completed high school). School discipline is a parent-reported measure of whether the child was expelled or suspended in the past 2 years. School absence is a measure of whether the youth self-reported missing school more than 3 weeks. Poor grades is a measure of whether the youth self-reported having grades that were mostly below a 75 on average during the last completed school year. Youth missing on 2 or more items (1.6%) were omitted, leaving an analytic sample of 2783 youth (N=1922 in treatment group, N=861 in control group).

Treatment Assignment and Take Up.

Families were randomly assigned to either a “low-poverty” treatment group (housing voucher restricted to low-poverty neighborhoods, with housing counseling), a “Section 8” treatment group (unrestricted housing voucher), or a public housing control group. Only 51% of the treatment voucher group complied, and actually moved using one of the offered Section 8 vouchers, called “take up”. Therefore, models were also estimated using treatment take up to predict delinquency and educational problems, to differentiate the effects of actually using the voucher to move from the effects of being randomly offered a voucher.

Age at Random Assignment and Gender.

Effect modification of MTO treatment by age at random assignment (range 5 to 16 years old) was tested using a threshold effect for age at 13 or older (1), compared to younger than 13 (0). Gender modification was tested by comparing girls (0) to boys (1).

Baseline Covariates.

In a balanced RCT, the treatment-outcome relationship is unconfounded by design, making covariate adjustment unnecessary to achieve internal validity. Imbalances may exist by chance, and imbalanced covariates may be associated with the outcome of interest (Tsiatis et al. 2008). Baseline covariate imbalance was estimated using bivariate logistic regression models predicting treatment from the baseline covariates, and regression models adjusted for all imbalanced covariates. Balanced baseline covariates were not included because: 1) including covariates selected on a post hoc basis may lead to biased and imprecise effect estimates (Pocock et al. 2002; Kahan et al. 2014; Freedman 2008; Tsiatis et al. 2008); and 2) the effects on subgroups of the data are defined by three strata (treatment versus control, girls versus boys, and 5–12 year olds versus 13–16 year olds), reducing the subgroup sample size each time the data is subset. A fundamental assumption of causal inference, positivity, requires sufficient variability in each level of the treatment exposure across all strata of adjusted confounders (Petersen et al. 2012). As the number of covariates increases, the risk for violating the positivity assumption because of data sparsity in the treatment group exposure compounds (Cole and Hernan 2008), increasing bias and variance (Cole and Hernan 2008), while decreasing power to detect treatment effect heterogeneity in an RCT (Pocock et al. 2002). Given that confounding in the treatment-outcome relationship is not reduced by additional covariate adjustment because of the RCT design, and bias and imprecision may be introduced, only the imbalanced covariates were modeled.

Analytic Approach

Given the randomized controlled design, models estimated intention-to-treat (ITT) linear regression analyses to obtain the average effect of being randomly-assigned a Section 8 voucher, compared to public housing controls. The main hypothesis test of interest is the three-way interaction between treatment, gender, and age (13+ vs. younger than 13), where a significant 3-way interaction indicates that the treatment effects were heterogeneous across these subgroups (Wang and Ware 2013). Age and gender specific estimates of the treatment-control group differences in delinquency and educational problems were obtained using post-estimation commands. The patterns were displayed by graphing the treatment-control differences by gender and age at randomization with 95% confidence intervals (CI). To interpret the magnitude of the effect, an effect size is presented using Cohen’s d by dividing the age-gender specific coefficients by the standard deviation (SD) of the total sample (Cohen 1969).

Models were then estimated using instrumental variable analysis to estimate the effect of actually moving with the voucher. ITT analyses preserve the randomized design and produce unbiased effects of being offered a voucher, but in cases where there is imperfect compliance, as in MTO, ITT effects will be watered down by the noncompliance. For example, if families with older children are more or less likely to use the voucher to move, then the ITT estimates will capture this pattern. Very early impact assessments in Boston and Baltimore suggested that baseline factors may be associated with treatment take up, including behavior problems (Katz et al. 2001) and pre-program arrests (Ludwig et al. 2001). To account for the possibility of selective take up, adherence-adjusted treatment-on-treated (TOT) models were estimated using instrumental variables (IV). When TOT patterns are similar to ITT, this rules out selective use of the voucher by baseline characteristics. To estimate IV models, one obtains the probability of using the voucher for the whole sample by predicting treatment take-up in the first-stage from MTO treatment and imbalanced covariates. By design, no controls can take-up the voucher. In the second stage, the probability of treatment take-up predicts the effect of moving with the voucher on delinquency and educational problems. IV methods are unbiased estimates of the effect of voucher use on the outcomes (Angrist et al. 1996), and readily accommodate treatment effect modification using interactions (Schmidt et al. 2017a; Osypuk et al. 2012a; Osypuk et al. 2012b). All analyses were weighted for changes to random assignment ratios over time and for attrition, and estimated using robust standard errors to account for household-level clustering of children (Kling et al. 2007) in Stata 14.

Sensitivity Analyses.

A series of sensitivity analyses were undertaken to assess the robustness of findings. First, it is possible that results may be an artifact of studying an outcome (delinquency, poor school performance) that is more prevalent among older youth. To test whether the results for age at baseline may be conflated with age at the interim survey (which are correlated at .9), two additional sets of models were estimated: 1) all models were adjusted for age at the interim survey, in addition to the age-at-baseline key modifier of interest, and 2) the sample was restricted to youth who were 15 to 19 years old at the time outcomes were collected (who were 8 to 16 year olds at baseline). Second, it is possible the results are sensitive to the skewed nature of the delinquency outcome, so all models were estimated using: 1) a logged version of the outcome, 2) a dichotomous measure of delinquency prevalence (any delinquent act vs. none), and 3) a count measure of delinquency estimated using negative binomial regression. Instrumental variable models for the binary and count outcomes were estimated using 2-Stage Residual Inclusion (2SRI) models, which are appropriate for non-linear outcomes (Terza et al. 2008). Results from these sensitivity analyses estimating the 3-way interaction models are presented in figures, with the subgroup patterns displayed graphically, and the 3-way interaction and subgroup effects reported in figures notes.

Finally, it is possible that other baseline factors may modify any treatment effects on delinquency and educational problems, such as race/ethnicity or baseline behavior problems (as a proxy for baseline delinquency). Restricting the sample to boys, since the hypothesized treatment effects would manifest for only boys, models were estimated testing for: 1) a 2-way interaction between MTO treatment and race/ethnicity, and 2) a 3-way interaction between MTO treatment, age, and race/ethnicity. This was repeated, testing for a 2-way or a 3-way interaction between MTO treatment, age, and behavior problems.

Results

Compared to the control group, families randomized to the low-poverty and Section 8 treatment groups lived in substantially better neighborhoods in 2002 (4–7 years after random assignment) on characteristics including economic conditions, crime, collective efficacy, the physical environment, and health (Nguyen et al. 2017). The neighborhoods that the low poverty group lived in were significantly better than the Section 8 group, and the effect estimates among those who actually moved using the voucher were double the size of those who were randomly assigned to receive the voucher (Nguyen et al. 2017). However, despite that the low poverty group moved to better neighborhoods than the Section 8 group, a formal statistical test of the effect of the two voucher treatment groups documented that the low poverty and Section 8 groups have statistically indistinguishable effects on delinquency and educational problems in all of the models (p>.05). In other words, the treatment effects on the outcomes are the same, regardless of what type of voucher families were offered and where they moved, which is consistent with prior work documenting homogeneous effects on other youth outcomes (Osypuk et al. 2012b; Osypuk et al. 2012a; Schmidt et al. 2017a). Given this homogeneity, the two voucher treatment groups were combined (compared to controls).

Descriptive Statistics

Table 1 displays the means/proportions for baseline covariates, overall and by treatment group, as well as the odds ratio (OR) from logistic regression models documenting the magnitude of non-equivalence (OR > 1 signifies covariate odds is higher among treatment than controls, OR <1 signifies covariate odds is lower among treatment than controls). As expected with an RCT, most baseline covariates were balanced for the interim analytic sample (Table 1); results were similar for the delinquency and educational problems analytic samples (not shown). Significant treatment-control differences emerged for two variables, which were entered as covariates in all regression analyses: youth having a problem requiring special medicine; and youth enrolling in a special class/school for health with a learning or behavioral problem.

Table 1.

Moving to Opportunity Youth, Baseline Variables, Overall and by Treatment Group

			Treatment Group
Construct	Variable	Overall	Treatment	Controls	OR
Analytic Sample Size	Total N in Interim Survey (2002)	2829	1950	879
Baseline Modifiers
Age	Age in years	9.94	9.96	9.88	1.01
Gender	Male	49.9%	49.5%	51.0%	0.94
	Female	50.1%	50.5%	49.0%	(ref)
Household Characteristics
Baseline mean poverty rate	Percent poverty rate in baseline census tract	49.6%	49.3%	50.2%	1.00
Site	Baltimore	15.5%	16.0%	14.2%	(ref)
	Boston	18.9%	18.1%	20.7%	0.78
	Chicago	22.4%	23.3%	20.4%	1.02
	Los Angeles	18.6%	17.5%	21.2%	0.73
	New York	24.6%	25.1%	23.5%	0.95
Household size	2 people	7.3%	6.9%	8.3%	(ref)
	3 people	22.3%	22.1%	22.9%	1.17
	4 people	25.4%	26.2%	23.4%	1.36
	5 or more people	45.0%	44.8%	45.4%	1.20
Youth Characteristics
Race/ethnicity	African American	62.8%	63.2%	62.1%	(ref)
	Hispanic ethnicity, any race	30.0%	30.3%	29.5%	1.01
	White	1.1%	1.0%	1.2%	0.81
	Other race	2.2%	2.4%	1.9%	1.24
	Missing race	3.8%	3.2%	5.3%	0.59
Gifted	Special class for gifted students or did advanced work	15.4%	14.7%	16.8%	0.85
Developmental Problems	Special school, class, or help for learning problem in past 2 years	16.6%	16.7%	16.3%	1.03
	Special school, class, or help for behavioral or emotional problems in past 2 years	7.7%	8.7%	5.3%	1.69*
	Problems that made it difficult to get to school and/or to play active games	6.5%	7.1%	5.0%	1.46
	Problems that required special medicine and/or equipment	9.1%	10.0%	7.0%	1.50*
	School asked to talk about problems child having with schoolwork or behavior in past 2 years	26.3%	26.7%	25.4%	1.07
Household Head Characteristics
Family Structure	Never married	55.9%	55.2%	57.5%	0.91
	Teen parent	25.9%	26.4%	25.0%	1.08
Socioeconomic Status	Employed	25.8%	26.1%	25.3%	1.04
	On AFDC (welfare)	76.0%	75.5%	76.9%	0.92
Education	Less than high school	47.1%	47.2%	46.7%	(ref)
	High school diploma	36.2%	36.6%	35.3%	0.89
	GED	16.7%	16.1%	17.9%	1.03
	In School	13.9%	14.4%	12.6%	1.18
Neighborhood/Mobility Variables	Lived in neighborhood 5 or more years	65.7%	65.8%	65.5%	1.02
	No family living in neigh	64.1%	63.1%	66.3%	0.87
	No friends living in neigh	37.3%	36.8%	38.5%	0.93
	Had applied for section 8 voucher before	44.3%	43.6%	45.8%	0.91
Neighbor Relationships	Chats with neighbors at least once a week	51.9%	51.3%	53.2%	0.92
	Respondent very likely to tell neighbor if saw neighbor's child getting into trouble	56.7%	56.8%	56.4%	1.02

^aP-value for testing the null hypothesis that the treatment and control group proportions or means did not differ were calculated from Wald chi-square tests (logistic regression models for dichotomous characteristics and multinomial logistic regression for categorical characteristics), or F-tests (linear regression for continuous variables).

NOTES: Variable means represent proportions, except baseline age (5-16) and mean poverty rate. Analyses weighted for varying treatment random assignment ratios across time and attrition, and standard errors were adjusted for family-level clustering.

^*p<.05

The overall mean for delinquency at 4–7 years after random assignment was .09 (Table 2), indicating on average youth reported engaging in less than one delinquent act; 60% of the sample reported no delinquency, 40% reported 1 or more delinquent acts, and the sample median was 0, all of which indicate a skewed outcome. The mean for educational problems was .21, meaning on average youth reported fewer than one educational problem; 48% of the sample reported no educational problems, 52% reported 1 or more educational problems, and the sample median was .25. Boys reported more delinquency and educational problems than girls, in line with prior research. For both outcomes, the means did not differ across treatment groups for the total sample or for girls only. Treatment group boys had a slightly higher mean delinquency and educational problems (.13 and .25 respectively) compared to control group boys (.11 and .23 respectively), but these differences were not statistically significant.

Table 2.

Moving to Opportunity Outcomes at Interim (2002) Descriptive Statistics, by Gender and Treatment Group

	Total Sample						Treatment Group						Control Group
Delinquency	N	M	SD	Q1	Med	Q3	N	M	SD	Q1	Med	Q3	N	M	SD	Q1	Med	Q3
Total Sample	2819	0.09	0.19	0	0	0.11	1941	0.09	0.19	0	0	0.11	878	0.09	0.18	0	0	0.11
Girls	1424	0.06	0.15	0	0	0.11	981	0.06	0.15	0	0	0.11	443	0.06	0.14	0	0	0.11
Boys	1395	0.12	0.21	0	0.11	0.22	960	0.13	0.22	0	0.11	0.22	435	0.11	0.21	0	0	0.22
Educational Problems	N	M	SD	Q1	Med	Q3	N	M	SD	Q1	Med	Q3	N	M	SD	Q1	Med	Q3
Total Sample	2783	0.21	0.30	0	0.25	0.25	1922	0.21	0.30	0	0.25	0.33	861	0.20	0.30	0	0.25	0.25
Girls	1405	0.17	0.27	0	0	0.25	971	0.16	0.27	0	0	0.25	434	0.17	0.28	0	0	0.25
Boys	1378	0.24	0.31	0	0.25	0.50	951	0.25	0.31	0	0.25	0.50	427	0.23	0.32	0	0.25	0.50

Note: Delinquency and educational problems are continuous, ranging from 0 to 1. M=Mean (weighted); SD=Standard Deviation; Q1=First Quartile; Med=Median; Q3=Third Quartile.

Intention-to-Treat Results

The MTO treatment main effect (collapsed on gender) on both outcomes was nonsignificant (delinquency B(SE) = .004(.01), p=.59, 95% CI: −.01, .02; educational problems B(SE) = −.003(.01), p=.81, 95% CI: −.03, .02). Although prior studies have documented treatment effect modification by gender on mental health and risky behaviors (Orr et al. 2003; Kling et al. 2007; Osypuk et al. 2012a; Osypuk et al. 2012b), this study did not detect significant modification of treatment effects on delinquency or educational problems by gender alone (Supplementary Table 1).

The primary hypothesis of a 3-way interaction between treatment, gender, and age at random assignment was statistically significant for delinquency, B(SE) = .08(.04), p=.03, 95% CI: .01, .15. The subgroup effects underlying the interaction (Figure 1) showed no significant effects of MTO treatment on delinquency for girls of any age, or for boys who were 12 or younger at baseline. There were significant and harmful effects of MTO treatment on delinquency for boys in the voucher treatment group who were in middle adolescence (13 to 16 years old) at baseline, B(SE) = .05(.03), p=.05, 95% CI: .001, .10. Calculating a standardized effect size (Cohen’s d), the harmful treatment effect equates to a .28 SD increase in delinquency when boys are offered a housing voucher to move during middle adolescence, compared to control group boys; this is a small effect. The pattern of effects for educational problems mirrored that of delinquency, but there was no significant 3-way interaction between treatment, gender, and age at random assignment, B(SE) = .11(.08), p=.14, 95% CI: −.04, .26 (Figure 2).

Fig 1.

Moving to Opportunity intention-to-treat and instrumental variable treatment effects on delinquency among youth aged 12–19 at interim (4–7 years after randomization), modification by gender and age at random assignment

Notes ITT results 3-way interaction effect B(SE)= .08(.04), p=.03, 95% CI: .01, .15; IV 3-way interaction effect B(SE)=.17(.08), p=.03, 95% CI: .02, .33. ITT subgroup effect: Girls 5–12 years old B(SE)= .01(.01), p=.55, 95% CI: −.01, .02; Girls 13–16 years old B(SE)= −.03(.02), p=.14, 95% CI: −.07, .01; Boys 5–12 years old B(SE)= .01(.01), p=.65, 95% CI: −.02, .03; Boys 13–16 years old B(SE)= .05(.03), p=.05, 95% CI: .001, .10. IV subgroup effects: Girls 5–12 years old B(SE)= .01(.02), p=.55, 95% CI: −.02, .04; Girls 13–16 years old B(SE)= −.06(.04), p=.15, 95% CI: −.15, .02; Boys 5–12 years old B(SE)= .01(.03), p=.65, 95% CI: −.04, .07; Boys 13–16 years old B(SE)= .11(.06), p=.05, 95% CI: −.001, .23. P-values reported for each bar test each subgroup effect against a null hypothesis of zero. Models adjusted for unbalanced baseline covariates.

Fig 2.

Moving to Opportunity intention-to-treat and instrumental variable treatment effects on educational problems among youth aged 12–19 at interim (4–7 years after randomization), modification by gender and age at random assignment

Notes ITT results 3-way interaction effect B(SE)= .11(.08), p=.14, 95% CI: −.04, .26; IV 3-way interaction effect B(SE)=.25(.17), p=.13, 95% CI: −.08, .58. ITT subgroup effect: Girls 5–12 years old B(SE)= −.01(.02), p=.55, 95% CI: −.04, .02; Girls 13–16 years old B(SE)= −.05(.05), p=.34, 95% CI: −.14, .05; Boys 5–12 years old B(SE)= −.002(.02), p=.91, 95% CI: −.04, .04; Boys 13–16 years old B(SE)= .07(.05), p=.15, 95% CI: −.03, .18. IV subgroup effects: Girls 5–12 years old B(SE)= −.02(.03), p=.55, 95% CI: −.07, .04; Girls 13–16 years old B(SE)= −.10(.11), p=.34, 95% CI: −.31, .11; Boys 5–12 years old B(SE)= −.01(.04), p=.91, 95% CI: −.08, .07; Boys 13–16 years old B(SE)= .16(.11), p=.15, 95% CI: −.06, .39. P-values reported for each bar test each subgroup effect against a null hypothesis of zero. Models adjusted for unbalanced baseline covariates.

Instrumental Variable Results

Adherence-adjusted IV models reinforced the ITT patterns, with IV effect estimates that were about twice as large. There was nonsignificant modification of the treatment by gender alone on both outcomes in IV, as with ITT models (Supplementary Table 1), but there continued to be a significant 3-way interaction between treatment, gender, and age at random assignment for delinquency, B(SE) = .17(.08), p=.03, 95% CI: .02, .33. Interpreting the subgroup effects, the adherence-adjusted effects of treatment on delinquency were nonsignificant for girls and younger boys, while the effect of moving with a Section 8 voucher to lower poverty neighborhoods was marginally significant and harmful for boys who were between 13 and 16 years old at baseline, B(SE) = .11(.06), p=.05, 95% CI: −.001, .23 (Figure 1). Calculating a standardized effect size (Cohen’s d), the harmful effect of treatment equates to a .61 SD increase in delinquency when boys are moved during middle adolescence, compared to control group boys. This is a moderate effect size. The 3-way interaction for educational problems was nonsignificant, B(SE) = .25(.17), p=.13, 95% CI: −.08, .58 (Figure 2). The consistency between the ITT and IV analyses confirmed that factors that influence whether or not families with older boys choose to use the voucher do not explain the ITT results.

Sensitivity Analyses

Analyses that adjusted for the youth’s age at the time the outcome was measured, and that restricted the sample to youth who were 15 to 19 years old at the time the outcome was measured (i.e., when the outcome is most prevalent), confirmed statistically significant 3-way interactions (p<.05). Therefore, the results were not due to a conflation between random assignment and interim survey age (Supplementary Figures 1–2). Results for educational problems confirmed nonsignificant 3-way interaction effects (not shown). Models using the three alternate delinquency outcomes (logged delinquency, delinquency prevalence, and delinquency count), all confirmed a statistically significant 3-way interaction (p<.05), and identical subgroup patterns as in the primary analysis (Supplementary Figures 3-5). Finally, all models testing for a 2-way or 3-way interaction between treatment and race/ethnicity or baseline behavior problems were nonsignificant, in both ITT and IV models. This suggests that the effects of residential mobility on delinquency and education are not explained by the youth’s race/ethnicity or existence of behavior problems upon study entry. Collectively, these findings confirmed the ITT and IV findings are robust to many different model specifications.

Discussion

The neighborhood is an important ecological environment influencing healthy child development (Browning and Soller 2014). Housing mobility policy is one policy in the housing sector that promotes moving low-income households to high-opportunity neighborhoods. Since neighborhood context is strongly patterned by race and income, this affordable housing policy is a promising tool to expand the choice set of neighborhoods of residence for low-income, minority families, within an equity framework (Osypuk 2013; Osypuk and Acevedo-Garcia 2010). However, changing exposure to neighborhood context via household-based policies like rental vouchers requires the household to move. General Strain Theory posits that, although the destination neighborhoods of high opportunity may be salutary, the stress of moving may offset any benefits, particularly at certain points in development (Agnew 2006). A large literature has documented the adverse effects of residential mobility for children’s outcomes (Leventhal and Newman 2010), yet the vast majority of these studies are observational (Jelleyman and Spencer 2008; Leventhal and Newman 2010), thus are likely confounded. This limits understanding whether residential mobility actually causes youth outcomes, as well as limiting the ability to inform promising strategies to mitigate any harm of residential mobility. Further, few studies have probed whether there are developmentally sensitive periods (Ben-Shlomo and Kuh 2002; Kuh et al. 2003) where moving is more or less important for affecting developmental outcomes, and whether these patterns differ for boys and girls. Therefore, little is known about who might be most vulnerable to residential moves, and when in development a move might be most (or least) disruptive.

Leveraging the natural variation in the age and gender of children in enrolled families, and the randomized design that is unbiased by residential selection or its predictors, this study documented that gender and age jointly modified the Moving to Opportunity experimental effects of residential mobility for adolescent delinquency, but not for educational problems. Residential mobility moderately increased delinquency (.61 SD effect size) only for boys who moved during middle adolescence (13–16 years old at baseline), compared to a control group who remained in baseline public housing, suggesting that mobility is an important stressor for boys during middle adolescence. This finding is consistent with a body of work showing harmful effects of residential mobility on youth outcomes using observational data (Leventhal and Newman 2010; Jelleyman and Spencer 2008), although findings were not comparable with respect to education. A few studies attempting to control for residential selection did not find effects of mobility on delinquency (Gasper et al. 2010; Porter and Vogel 2014), however, these studies did not examine mobility from a life course perspective, to model amplified effects at later ages, which may explain the discrepant findings. Guided by General Strain Theory and sensitive period life course models, this study posited that the adverse effect of Moving to Opportunity emerged because boys at this age are vulnerable to the stressors of moving and act out, including in the form of delinquency. Indeed, prior work suggests that residential mobility may have been harmful for Moving to Opportunity boys because they were cut off from their social networks (friends and father figures), and fell in with riskier peer groups (Popkin et al. 2008; Schmidt et al. 2017b; Clampet-Lundquist et al. 2011). Researchers are beginning to formally test how these, and other, factors mediate the effect of mobility on other adolescent outcomes (Schmidt et al. 2014; Schmidt et al. 2017b; Rudolph et al. In press). Understanding the pathways by which residential mobility influences healthy development remains an important area of investigation for future research.

For girls, residential mobility did not predict delinquency or educational problems, nor did the effects of residential mobility vary by age. This finding is inconsistent with prior work that moving is a stressor associated with negative outcomes for girls. It is consistent with the hypothesis, based on General Strain Theory, that the benefits of moving to a low-poverty neighborhood may have outweighed the stress of moving (Agnew 2006; Hango 2006). This sample was comprised of low-income families on rental assistance, and the stressors that changed with moving from high- to low-poverty neighborhoods, e.g., reduction in sexual harassment and victimization (Popkin et al. 2008), may have differed, or the magnitude may be far greater, than what the general population of female adolescents experiences. This may explain this finding relative to other residential mobility literature for girls.

That residential mobility in the context of housing mobility does not generate adverse developmental effects for girls has important implications for the prioritization of housing assistance funds, particularly for families with young children. Housing authorities often prioritize populations for housing choice vouchers or housing mobility policies, which helps them allocate limited resources more strategically (Scott et al. 2013). Prioritization can target households that are most likely to successfully use their voucher to make opportunity moves, including households that have made at least one move with a voucher before or households on the cusp of self-sufficiency in terms of meeting public housing authority work requirements. Housing assistance also can target households that are most likely to benefit from the relocation (e.g., households with children, or with chronic health conditions), and/or housing assistance can be targeted to rectify historic injustices in housing markets (e.g., programs originating from court consent decrees, civil rights settlements) (Scott et al. 2013).

These findings also have implications for developmental science that investigates the effects of social conditions: examining heterogeneity by key drivers of development may be necessary to progress the understanding of how root causes operate. Previous studies found nonsignificant effects of the Moving to Opportunity study on delinquency by gender at the interim follow up (Orr et al. 2003; Kling et al. 2005; Graif 2015). However, they did not examine heterogeneity of effects by youth age at study baseline, despite that both gender and age may intersect in complex ways to pattern development, particularly with respect to interacting and navigating neighborhood environments. In developmental science, heterogeneity is common and expected. Future studies evaluating programs that address fundamental social conditions, like the social determinants of health, need to prioritize testing heterogeneity of effects. Indeed, interventions that do test for variability of effects find that more vulnerable subgroups are less likely to benefit from an intervention, and may even exhibit harm, as was seen here for older boys (Osypuk et al. 2012b; Osypuk et al. 2012a; Nguyen et al. 2016; Ertel et al. 2007; McCormick et al. 2006). Even then, most trials are powered for detecting only a main effect of an intervention rather than probing variation, where type I error concerns often arise. Although planned (a priori) exploration is preferred based on the typical hypothesis-testing model of science, there is often insufficient understanding of the underlying phenomena to anticipate some of the developmental complexities arising from social exposures. Developmental scientists may benefit from embracing data-driven exploration and search methods, coupled with transparency of reporting (Glymour et al. 2013), that may help to illuminate some of the opacity and complexity of child development pathways.

A significant implication here is that families with adolescent boys may be particularly vulnerable to residential mobility. Research is beginning to document how policies in one sector may affect outcomes in another, both positively and negatively (Osypuk et al. 2014). Policy makers should consider the unintended consequences that social programs may have for health and development, and consider tailoring them to address the needs of vulnerable families. Moving to Opportunity was a relatively simple intervention within the housing sector to encourage using housing subsidies within higher opportunity neighborhoods. Although it aimed to promote employment, income, and education, and reduce welfare reliance of single mothers and their children, the program had few effects on these outcomes. Instead, the greatest effects of Moving to Opportunity emerged for mental health (for both mothers and children) and risky behavior outcomes among the children (Orr et al. 2003; Sanbonmatsu et al. 2011). If the housing intervention were buttressed with additional services, across sectors, targeting children and adolescents, a holistic or tailored approach might buffer any harmful effects of residential mobility. Indeed, housing voucher programs are beginning to incorporate such supportive services into their programs (Scott et al. 2013), which may be central for ensuring that all families can benefit from residential mobility. Evaluating the success of introducing these additional supports into existing mobility programs will be important moving forward.

Although Moving to Opportunity has many strengths, including its high internal validity, the study does have some limitations. Generalizability may be limited given this study targeted low-income, minority families living in distressed public housing in large US cities. Since families volunteered to enroll in the program, thereby expressing their willingness to relocate, this raises concerns about self-selection (Wolff et al. 2017). These results should be interpreted as being generalizable to the population of low-income US families residing in public housing projects who apply for housing voucher assistance, and who live in large cities in the US. Notably, that population happens to be very policy-relevant, since housing voucher subsidy programs that target these families are the largest affordable housing mechanism in the US (National Low Income Housing Coalition 2011).

Given that Moving to Opportunity was implemented over 20 years ago, one might inquire how applicable the intervention or its results are for today. Housing affordability and concentration of poverty continue to be important issues hampering the quality of life for low-income households. For example, despite the federal investment in neighborhood revitalization since the early 1990s, such as Hope VI that demolished some of the most distressed public housing developments (Popkin et al. 2004), housing affordability in America remains dire for low income populations (Joint Center for Housing Studies 2017). Although the share of impoverished Americans living in high-poverty neighborhoods declined from 1990 to 2000, it has risen again to match that of the early 90’s. Concentrated poverty remains correlated with race/ethnicity (Jargowsky 2015), so minority children continue to face double jeopardy of risk arising from both family level, and neighborhood level, poverty (Acevedo-Garcia et al. 2008). Moreover, some important findings of this study would not have been documented without a long term follow up, thereby introducing a tension between timeliness of data with latency periods necessary to allow complex social phenomena to unfold. For example, moving to higher quality neighborhoods may only improve low-income economic prospects in young adulthood when children were younger than 13 years old at baseline (Chetty et al. 2016), and may reduce middle-aged adults’ risk of metabolic disease, like diabetes and morbid obesity (Ludwig et al. 2011), that may delay chronic disease. These findings only emerged after 15 years of follow up. Despite its age, Moving to Opportunity represents one of the only opportunities for evaluating effects of housing mobility policy promoting neighborhood opportunity within a randomized design. Its results remain policy relevant, given that housing choice voucher policy (the study intervention) remains the largest affordable housing policy in the US, that housing allowances are a common housing policy globally (OECD 2017), and that government funds to support affordable housing using vouchers continue to grow across time (Kingsley 2017).

Moving to Opportunity is a bundled treatment (Ludwig et al. 2008), so one cannot separate the effects of neighborhood context, housing changes, and mobility itself. Yet many of these factors change with any move, so this is a meaningful treatment effect to operationalize residential mobility. The two voucher treatment groups were meant to separate the effect of mobility to any neighborhood, from that to a much lower poverty neighborhood. For youth, both voucher groups generated comparable effects on delinquency and educational problems, so these two treatments cannot be teased apart empirically. Moreover, because of the nature of the experimental design, Moving to Opportunity can only examine the effects of upward mobility to better neighborhoods, and not lateral or downward mobility (Wolff et al. 2017).

There are a few limitations with respect to the measurement of delinquency in Moving to Opportunity; however results from sensitivity models suggest these are not major concerns. First, the study did not measure delinquency at baseline; however, this is unlikely to bias effects. One would expect the experimental design to have balanced most measured covariates at baseline, including any baseline delinquency. It does preclude modeling whether baseline delinquency further modifies treatment effects, and it is possible that the age-gender interaction is capturing this to some extent. Sensitivity models tested for heterogeneity of treatment effects by baseline behavior problems along with age, as a proxy for delinquency, but did not document any. In addition, delinquency is a skewed outcome, but linear regression produced comparable results as logging the outcome, modeling the count using negative binomial regression, or dichotomizing it.

Conclusion

Using a randomized housing mobility program, this study examined variation in residential mobility effects on delinquency and educational problems by gender and age when children were moved. Although the importance of environmental contexts have been long theorized for child development (Bronfenbrenner 1993), strongly evaluating manipulable exposures, such as this policy that promoted residential mobility, is rare. This study documented that adolescence was a sensitive period for the effects of moving on delinquency. Delinquency increased among older boys who moved with a housing voucher, while delinquency and educational problems among girls and younger boys were not affected by residential mobility. The translational implications are important, suggesting that families with teenage boys may be particularly vulnerable to residential transitions and may need additional supports to buffer adverse effects of moving. This study contributes to the flourishing literature that policies outside of criminal justice and education sectors, including the housing sector, can profoundly impact child developmental outcomes (Osypuk et al. 2014). Evaluating such policies with strong methods, over long periods of time, can inform the etiology of complex developmental processes, and how to best promote healthy development for all children.