Associations of Housing Mobility Interventions for Children in High-Poverty Neighborhoods With Subsequent Mental Disorders During Adolescence

IMPORTANCE Youth in high-poverty neighborhoods have high rates of emotional problems. Understanding neighborhood influences on mental health is crucial for designing neighborhood-level interventions. OBJECTIVE To perform an exploratory analysis of associations between housing mobility interventions for children in high-poverty neighborhoods and subsequent mental disorders during adolescence.

O bservational studies have consistently found that youth in high-poverty neighborhoods have high rates of emotional problems even after controlling for individual-level risk factors. 1 These findings raise the possibilities that neighborhood characteristics affect emotional functioning 2 and neighborhood-level interventions may reduce emotional problems. Available data from observational studies are unclear and subject to selection bias and the possibility of reverse causality (ie, families with emotional problems end up in poorer neighborhoods). Despite this uncertainty, presumptive neighborhood effects have been characterized, 3 causal pathways have been hypothesized, 4 and interventions have been implemented. 5 It is important to evaluate these causal claims regarding neighborhood effects experimentally. The US Department of Housing and Urban Development (HUD) enacted a housing mobility experiment known as the Moving to Opportunity for Fair Housing Demonstration by randomizing volunteer low-income public housing families with children to receive vouchers to move to lower-poverty neighborhoods. 6,7 An interim evaluation 4 to 7 years after randomization showed that the intervention caused families to move to better neighborhoods with lower poverty and crime rates and increased social ties with more affluent people. 8 Significant reductions in psychological distress and depression were also found among adolescent girls in the intervention group vs the control group but increased behavior problems were found among adolescent boys in the intervention group vs the control group. [9][10][11] Given the importance of these sex differences, clinically significant mental disorders were included in a long-term ( 12,13 reduced adult extreme obesity and diabetes, 14 and improved adult subjective well-being. 13 No detectable effects on economic self-sufficiency were found. 13 Although long-term evaluation found significantly reduced psychological distress among adolescent girls, 15 measures of mental disorders were not examined in previous reports.
The primary objectives of the Moving to Opportunity study were to move families to lower-poverty neighborhoods and increase educational achievement and economic selfsufficiency. Mental disorders were measured as post hoc outcomes. The current report presents the first exploratory analyses evaluating long-term associations of housing mobility randomization with mental disorders among participants who were in early childhood at randomization and adolescence at follow-up.

Study Design
Families (n=4604) in the Moving to Opportunity study were recruited by public housing authorities from 1994 to 1998. 16 Families had to reside in public or project-based assisted housing in high-poverty census tracts (>40% of families in poverty) in Baltimore, Boston, Chicago, Los Angeles, or New York; be eligible for Section 8 housing; and have 1 or more children age younger than 18 years. Census tracts contain 2500 to 8000 people and are defined by the US Census Bureau to be "homogeneous with respect to population characteristics, economic status, and living conditions." 17 Housing authorities sent recruitment letters, held information sessions, and asked families to complete applications within 4 weeks of the invitation. Signed consents and baseline questionnaires were obtained during intake sessions prior to randomization. Families were then randomized into 1 of 3 groups using a computerized random-number generator: a low-poverty voucher group, a traditional voucher group, or a control group. In the low-poverty voucher group, families were offered a standard rent-subsidy voucher restricted to low-poverty census tracts (<10% of families in poverty). Vouchers provided subsidies for private-market housing equal to the difference between a rent threshold and the family's rent contribution (30% of income, identical to public housing). 18 Families remained eligible for vouchers as long as they met income and other criteria. Families also received short-term housing counseling during their initial housing search. 6,7 After 1 year, families in the low-poverty voucher group could use their voucher to relocate to a different tract, including those with higher-poverty rates, or could remain in the tract where they originally moved even if the poverty rate of that tract fell out of the low-poverty range. In the traditional voucher group, families were offered a standard rent-subsidy voucher without restriction on location as well as standard mobility counseling. 6,7 In the control group, families were offered no new assistance. Enhanced mobility counseling was offered to low-poverty voucher group families because of restrictions on where they could move. The protocol was approved by the Office of Management and Budget and HUD. Twenty-three percent of invited families applied, 6 and ultimately 48% of low-poverty voucher families and 63% of traditional voucher families used their vouchers to move. 7 Interim (4-7 years after randomization) and long-term (10-15 years after randomization) evaluation surveys were carried out with household heads and residents who were children at baseline randomization and adolescents at follow-up. Most adolescents in the interim evaluation were in middle childhood or early adolescence (ages 9-16 years) at randomization, whereas most adolescents in the long-term evaluation were in early childhood (ages 0-8 years) at randomization. Long-term assessments were performed June 2008 through April 2010 by interviewers blinded to group assignment. All adolescents from households with 1 to 3 baseline children and 3 randomly selected adolescents from households with 4 or more baseline children were targeted for interview. Large households were purposely undersampled to reduce household burden.
Recruitment for long-term assessment began with telephone tracking and networking to locate hard-to-recruit participants. Potential respondents were offered $50 for completing interviews. Although most interviews were performed face-to-face, some were performed by telephone. Of

Measures
Baseline head-of-household questionnaires focused largely on sociodemographics and neighborhood experiences (eg, social networks and crime victimization). Mental disorders were not assessed. Item-level missing data on the variables assessed was less than 5% for all but 5 variables (low birth weight; hospitalization before first birthday; baseline health problems that restricted normal activities; parent educational level; whether someone read to the child more than once daily during his/her early childhood; 5.5% to 11.2% missing). There were no missing values on the intervention variables. Item-level missing data were imputed using multiple imputation 20 using SAS software. 21 The long-term assessment included the Composite International Diagnostic Interview (CIDI), 22 a widely used psychiatric diagnostic evaluation tool known to have good concordance with clinical diagnoses of mental disorders based on the Diagnostic and Statistical Manual of Mental Disorders (Fourth Edition) (DSM-IV). 23 The CIDI questions were read word-forword and responses recorded in prespecified (mostly yes/no) format. Diagnoses were generated by CIDI algorithms operationalizing DSM-IV inclusion criteria. Diagnoses were made for DSM-IV disorders present within the past 12 months. Itemlevel missing data were less than 1% for each symptom question and were recoded conservatively to assume the symptom was absent. We focused on 6 DSM-IV/CIDI disorders: mood (major depression), anxiety (panic, posttraumatic stress), and disruptive behavior (oppositional-defiant, intermittent explosive, conduct). Bipolar disorder was also assessed, but was not analyzed due to low prevalence and insufficient statistical power to detect meaningful associations (eTable 1 in the Supplement).

Statistical Analysis
Sample size was determined by the Moving to Opportunity study budget ($70 million Congressional authorization, additional vouchers from local housing authorities, and counseling donated by nonprofit agencies). Randomization was designed to yield equal numbers of families within cities using vouchers in each intervention group. The number of families in the control group invited was set to equal the mean number invited in the 2 intervention groups. As voucher use percentages were determined only after randomization, proportions randomized across groups were modified during the study to adjust for observed rates of voucher use. The HUD determined that this design would yield 80% power to detect an effect of $2000 in increased earnings in each intervention group with a one-sided α of .05. 6(pE-4,exhibit,E4) Post hoc power calcu-lations showed that the long-term follow-up sample of adolescents had at least 80% power to detect an odds ratio (OR) for each of the 6 mental disorders considered herein of 1.4-1.8 (eTable 1 in the Supplement).
Intention-to-treat 24 logistic regression analysis 25 was used to estimate associations of the interventions with the outcomes. Across-time variation in the intervention vs control group selection ratios from 1994 to 1998 was corrected for by weighting. Case-level multiple imputation based on 20 pseudosamples was used to adjust for the fact that not all baseline participants completed follow-up interviews. The Taylor series method 26 implemented in SUDAAN 27 was used to adjust for weighting and clustering (cities, housing projects, families). The significance of sex differences was assessed by estimating a logistic regression equation to predict each disorder that included dummy variables for each intervention, a dummy variable for sex, and 2 dummy variables for the interactions of interventions with sex. A 2-degree-of-freedom χ 2 test was used to evaluate the significance of the interactions. In cases for which the test was significant, associations of the interventions with the disorder were considered separately for each sex. The evaluation of sex differences was carried out because significant sex differences had been found in previous interim evaluations. [9][10][11] The 6 mental disorders were considered separately because risk factors vary across these disorders. 28,29 The Benjamini-Hochberg method 30 was used to adjust significance tests across outcomes for the false discovery rate.
Logistic regression coefficients and standard errors were exponentiated to create ORs and 95% confidence intervals. Mental disorder prevalence estimates in the intervention and control groups were used to calculate absolute risk (AR) and absolute risk reduction (ARR). The jack-knife repeated replications method in SAS 26 was used to generate confidence intervals for the estimates of AR and ARR. Statistical significance was evaluated using a 2-sided α of .05.

Response Rates
The 3689 adolescents assessed were aged 0 through 8 years (median age, 4 years) at baseline and aged 13 through 19 years (median age, 16 years) at the time of long-term follow-up interviews. A total of 2872 adolescents were interviewed (1407 boys and 1465 girls from 2134 families), including 1165 in the low-poverty voucher group (843 families), 799 in the traditional voucher group (615 families), and 908 in the control group (676 families), out of the 3689 eligible in the baseline sample (77.8% participation rate). An additional 643 adolescents were randomly excluded (188 from families with >4 eligible respondents and 455 due to difficult recruitment) and 174 were lost to follow-up (including 18 deceased) ( Figure). The weighted response rates were 92.9% (lowpoverty voucher group), 86.9% (traditional voucher group), and 89.4% (control group) using the American Association of Public Opinion Research definition. 31(p51) Respondents were more likely to be girls and non-Hispanic black but did not

Sample Characteristics
Baseline sociodemographic characteristics of adolescents were largely comparable across the 3 groups for both boys ( between 10 and 20 years were targeted for follow-up, whereas a random 3 were targeted from baseline households with 4 or more youth. A weight of n/3, for which n equals the number of eligible youths in the baseline household, was used to adjust for the undersampling of youths from baseline households containing more than 3 eligible youth. Phase 1 data collection refers to the efforts made to contact and inter-view all target respondents until the end of the field period, at which point a random 35% of eligible target respondents (those who had not yet been interviewed, were not deceased or incapacitated, had not declined to participate) were selected for a more intensive Phase 2 data collection effort that included expanded tracing efforts andincreasedfinancialincentives.Aweightof1/.35wasusedtoadjustfortheundersampling of the Phase 2 hard-to-recruit youths who were interviewed. b For the 693 families excluded before randomization, the number excluded for each reason listed above is not known. and girls ( Table 2). Most respondents were non-Hispanic black (61.8%-66.2% of groups) or Hispanic (27.7%-33.2% of groups). The majority of respondents were ages 0 through 5 years at baseline (82.2%-87.9% of groups), with mean age of 3.6 years in each group and range of 0 through 7 years in the lowpoverty voucher group and 0 through 8 years in traditional voucher and control groups. The majority of baseline families received Aid to Families with Dependent Children (79.1%-85.1% of groups). Mean baseline neighborhood poverty rates were 53.6% to 54.9% (Table 3).

Mental Disorder Prevalence Within the Past 12 Months
The most prevalent mental disorders within the past 12 months were found to be intermittent explosive disorder (14.2% of boys and 16.0% of girls) and oppositional-defiant disorder (6.8% of boys and 8.4% of girls), followed by major depressive disorder (5.5% of boys and 7.9% of girls), posttraumatic stress disorder (PTSD) (4.4% of boys and 6.6% of girls), conduct disorder (4.3% of boys and 1.6% of girls), and panic disorder (4.1% of boys and 3.7% of girls) (eTable 3 in the Supplement).  Table 4). None of the other 11 comparisons of low-poverty or traditional voucher groups with the control group was significant. Odds ratios comparing the low-poverty voucher group with the control group were in the range 0.7-1.6 (P = .13-.84).
Odds Ratios comparing the traditional voucher group with the control group were in the range of 0.9-1.1 (P = .70).

Associations of Interventions With Mental Disorders Among Girls
Adjusting for the false discovery rate, girls in the traditional voucher intervention group had significantly reduced rates of  Table 5). The number needed to treat (NNT) (inverse of ARR) among girls was 23 for major depression and 38 for conduct disorder. None of the other 4 comparisons between intervention and control groups was significantly different, with ORs in the range 0.5-1.2 (P = .06-0.40).

Discussion
Our post hoc exploratory analysis found that interventions to encourage moving out of high-poverty neighborhoods were associated with increased depression, PTSD, and conduct disorder among adolescent boys and reduced depression and conduct disorder among adolescent girls randomized at ages 0 through 8 years. These sex differences were broadly consistent with interim follow-up results. 8,9,11 Qualitative evidence a Based on multiply imputation data (described in the text) to adjust for the fact that 22.2% of eligible baseline respondents did not participate in the long-term evaluation survey. b Mean number of respondents in the group with the outcome averaged across the 20 multiply imputed pseudosamples. c Compared with controls. suggested these differences were due to girls profiting more than boys from moving to better neighborhoods because of sex differences in both neighborhood experiences and in the social skills needed to capitalize on the new opportunities presented by their improved neighborhoods. [32][33][34] The magnitudes of the protective associations of the interventions with mental disorders among girls were modest in the intentionto-treat analyses, although these estimates would be larger if the analysis was restricted to movers. However, the ORs are comparable in size with those published in studies of risk factors considered to be of high policy significance. For example, the elevated ORs of PTSD found among boys were comparable with the ORs found between combat exposure and PTSD in epidemiological studies of the military, 35 whereas the reduced ORs of major depression found among girls was comparable with the inverse of the ORs found in previous research between sexual assault and major depression in epidemiological studies of young women. 36 Furthermore, it is important to recognize that these associations were evaluated 10 to 15 years after randomization. It is not clear if the magnitudes of the associations were stable over this entire time, but if so, they would be substantial despite the relatively high NNT. For example, ARR for major depression among girls would be 58.3 person-years per 100 respondents over 15 years if ORs were temporally stable over the entire time. External validity was reduced by the fact that only 23% of eligible families volunteered for the Moving to Opportunity study. However, the public housing population is large and therefore even this small fraction represents more than 300 000 low-income US children, 37 making the volunteer families significant from a policy perspective even though they were only a minority of all public housing families. A question might be raised about whether the added costs of developing a special housing intervention for such a small proportion of public housing recipients could be justified by the small proportion accepting the offer, but this concern is mitigated by the fact that many housing economists believe the true costs of housing vouchers are actually lower than those of conventional public housing because of the increased efficiency of the open housing market. 18 It is nonetheless difficult to draw policy implications from these results, because the findings suggest that the interventions might have had harmful effects on boys but protective effects on girls. Future governmental decisions regarding widespread implementation of changes in public housing policy will have to grapple with this complexity based on the realization that no policy decision will have benign effects on both boys a Based on a series of logistic regression equations comparing respondents in the low-poverty and traditional voucher intervention groups with respondents in the control group, ignoring whether intervention families used their vouchers. The equations are based on multiply-imputed data (described in the text) to adjust for the fact that 22.2% of eligible baseline respondents did not participate in the long-term evaluation survey. b The P values evaluate the significance of odds ratios using the Benjamini-Hochberg method 30 to adjust for the false discovery rate. These P values are higher than those implied by the 95% CIs, as the latter are based on models for separate outcomes. c Mean number of respondents in the group with the outcome averaged across the 20 multiply imputed pseudosamples. and girls. One way to do so might be to develop more nuanced assignment rules than currently exist or additional intervention elements to mitigate the adverse effects of the intervention on boys while maintaining the protective effects on girls. Development of such refinements will require a better understanding of the interactions of influences among individual, family, and neighborhood characteristics leading to child and adolescent mental disorders. Although the Moving to Opportunity study was not designed to produce this kind of understanding, these results should create an impetus to do so by documenting that neighborhoods do matter. The challenge for future research is to increase understanding enough to guide allocation of the substantial amount of money spent on public housing in the United States each year (more than $36 billion in fiscal year 2012) 38 to maximize the health and well-being of all family members rather than to maximize value for some family members at the expense of other family members.
The Moving to Opportunity study had several strengths, including an experimental design, large sample size, and long follow-up. However, it also had several noteworthy limitations: only 23% of eligible families volunteered, and families offered vouchers had rather severe time limits on enrollment and practical constraints on finding new housing that might have artificially reduced uptake. 39 These factors may reduce the generalizability of the results. 40 Other limitations include the fact that nonrespondents might have differed systematically from respondents; that the experiment was implemented when the unemployment rate was much lower than it is today; 41 that the CIDI and other mental health measures were not administered at baseline; and that, as with all policy experiments, the study design made it impossible to trace intervening processes that might account for aggregate intervention effects. In addition, the Moving to Opportunity study was underpowered to detect effects of the 2 separate intervention groups on uncommon adolescent mental disorders.
Despite these limitations, we found significant associations of the study interventions to reduce neighborhoodlevel poverty with several important adolescent mental disorders, providing evidence that experimental manipulation of incentives to move is associated with adolescent emotional functioning. However, because the interventions were also associated with changes in many other aspects of neighborhoods and participant experiences, pathways accounting for the associations of the interventions with adolescent mental disorders remain unclear, creating a challenge for future research to develop nuanced decision rules for matching public housing families with neighborhoods to maximize the health and well-being of all family members.