Does the Temporal Pattern of Moving to a Higher-Quality Neighborhood Across a 5-Year Period Predict Psychological Distress Among Adolescents? Results From a Federal Housing Experiment

Abstract

Using data from the Moving to Opportunity (MTO) experiment (1994–2002), this study examined how a multidimensional measure of neighborhood quality over time influenced adolescent psychological distress, using instrumental variable (IV) analysis. Neighborhood quality was operationalized with the independently validated 19-indicator Child Opportunity Index (COI), linked to MTO family addresses over 4–7 years. We examined whether being randomized to receive a housing subsidy (versus remaining in public housing) predicted neighborhood quality across time. Using IV analysis, we tested whether experimentally induced differences in COI across time predicted psychological distress on the Kessler Screening Scale for Psychological Distress (n = 2,829; mean β = −0.04 points (standard deviation, 1.12)). The MTO voucher treatment improved neighborhood quality for children as compared with in-place controls. A 1-standard-deviation change in COI since baseline predicted a 0.32-point lower psychological distress score for girls (β = −0.32, 95% confidence interval: −0.61, −0.03). Results were comparable but less precisely estimated when neighborhood quality was operationalized as simply average post–random-assignment COI (β = −0.36, 95% confidence interval: −0.74, 0.02). Effect estimates based on a COI excluding poverty and on the most recent COI measure were slightly larger than other operationalizations of neighborhood quality. Improving a multidimensional measure of neighborhood quality led to reductions in low-income girls’ psychological distress, and this was estimated with high internal validity using IV methods.

Abbreviations

CI

confidence interval

COI

Child Opportunity Index

ITT

intention to treat

IV

instrumental variable

MTO

Moving to Opportunity

SD

standard deviation

SE

standard error

Neighborhood quality may be an important distal cause of health and health inequity, because it patterns an array of other social determinants of health (1). Poorer-quality neighborhoods are associated with worse depressive symptoms and psychological distress (2–4). However, this literature is based predominantly on observational and cross-sectional studies, meaning associations may be confounded or attributable to other biases. Residential selection, a form of confounding bias, may be one of the most important threats to internal validity in neighborhood health-effect studies (1, 5, 6). For example, preexisting health problems or lower individual socioeconomic status may promote moving to disadvantaged neighborhoods, thereby confounding estimated effects of neighborhood quality on health.

Moreover, neighborhood quality is typically measured as a static, single-point-in-time characteristic instead of as dynamic (7), despite the fact that mobility is common among poor households (8, 9). Dynamic trajectories may better capture neighborhoods across time (7, 10), but the small body of literature on how neighborhood quality influences health over time (7, 11–14) is observational, with little or no information on (or control for) housing-related preferences, tradeoffs, or constraints (15). Moreover, neighborhood quality is predominantly modeled as neighborhood socioeconomic status (16), which reflects composition, rather than concrete aspects of neighborhoods. Finally, few neighborhood effect studies are policy-relevant, designed to model specific exposures or mechanisms for actually modifying neighborhood environments (17).

Moving beyond narrowly defined, static measures, this study operationalized neighborhood quality as a multidimensional, validated neighborhood opportunity index over time. Moreover, we leveraged a housing policy experiment, the Moving to Opportunity (MTO) Study, with an exposure comparable to the largest current federal affordable housing voucher program (18). In MTO, families were randomly allocated to receipt of a Section 8 housing voucher (now called a Housing Choice Voucher), allowing moves to lower-poverty neighborhoods. Girls in the “treatment” group (i.e., those who moved to lower-poverty neighborhoods) experienced lower levels of psychological distress than their public housing control group counterparts (19–22). This association is thought to be attributable to improvements in neighborhood context, making this an ideal setting in which to explore what aspects of neighborhood context are most important for girls’ mental health. Meanwhile, treatment group boys experienced increased psychological distress (21, 22), which was thought to be due to the disruptive effects of moving (23, 24) rather than a result of neighborhood context (25). Therefore, evidence for boys cannot speak to the effects of neighborhood quality on lowering distress per se. Since Housing Choice Vouchers serve over 2 million low-income families in the United States (18), MTO has high policy relevance, making it important to discern which mechanisms explain these effects. In the current study, we hypothesized that improved neighborhood quality over time would lower psychological distress among adolescent girls but not boys.

METHODS

Data

We linked individual-level data from residential histories spanning 1994–2002 with neighborhood-level indicators of opportunity.

The MTO Study.

Individual-level data were obtained from the MTO Study (1994–2002), a randomized controlled trial of voluntary neighborhood relocation sponsored by the US Department of Housing and Urban Development in 5 US cities: Baltimore, Maryland; Boston, Massachusetts; Chicago, Illinois; New York, New York; and Los Angeles, California (26). Eligible low-income families were those that had children under age 18 years, qualified for rental assistance, and lived in public housing in high-poverty neighborhoods (27). Applicants were drawn from waiting lists. They signed enrollment agreements and gave informed consent, completed baseline surveys, and were evaluated for eligibility (28).

Treatment assignment.

A total of 4,610 eligible volunteer families were randomly assigned to: 1) a “regular Section 8” group that received a Section 8 housing voucher to move to a subsidized rental apartment in any neighborhood; 2) a “low-poverty-neighborhood” group that received a Section 8 voucher to move only to a low-poverty neighborhood (<10% of census tract poverty) and received housing counseling to aid relocation; and 3) a control group that could remain in public housing (28). Treatment compliance—whether families successfully leased an apartment with the offered voucher—was 47.2% for the low-poverty group and 61.4% for the Section 8 group.

Assessment.

Baseline (1994–1998) and interim (2001–2002) follow-up surveys were conducted via in-person interviews with household heads and children (19, 28). Our analytical sample included 2,829 adolescents (1,426 girls, 1,403 boys) who responded to the MTO interim survey (90% response rate) (19). Our fundamental hypothesis was that improved neighborhood quality would reduce psychological distress. Although aspects of neighborhoods improved for boys (25), treatment increased their psychological distress; thus, the instrumental variable (IV) analysis for boys was inconsistent with our hypothesis that improved neighborhood quality is a mechanism for reducing distress. Analyses for boys are presented in Web Tables 1–3 (available online at https://doi.org/10.1093/aje/kwaa256), but it is likely that mechanisms other than neighborhood quality were operating to explain the harmful effects on boys’ psychological distress. Therefore, we focused our main analysis on girls.

Reconstructing residential history.

We reconstructed the address histories of MTO families (n = 3,526) coded to census tracts at baseline (1994–1997), the MTO first voucher-related move, and location at the 1997, 2000, and 2002 canvasses (19). Treatment group vouchers expired within 90 days of random assignment, so the first voucher move corresponded approximately to 90 days after random assignment.

Measures

Outcome (2001–2002).

Past-month psychological distress was measured using the Kessler Screening Scale for Psychological Distress (29), a 6-item scale (being so depressed that nothing could cheer you up; being nervous; being restless or fidgety; feeling hopeless; feeling that everything is an effort; feeling worthless) with Likert responses (none of the time, a little, some of the time, most of the time, or all of the time during the past 30 days). We used item response theory latent variable methods to construct the Kessler Screening Scale, which reduces measurement error in comparison with a straight mean. Cronbach’s α coefficient was 0.80. The mean Kessler Scale scores were −0.04 (standard deviation (SD), 1.12) overall, 0.11 (SD, 0.93) for girls, and −0.11 (SD, 0.89) for boys, approximating a standardized measure.

MTO treatment instrument.

MTO treatment was modeled as a categorical variable (intention-to-treat (ITT) analysis: low-poverty-neighborhood group, Section 8 group, or control group (referent)).

Baseline covariates.

In mediation, we are concerned about confounders of the exposure-outcome association and the mediator-outcome association. Randomization ameliorates concerns about the former, although we did adjust for any imbalanced baseline covariates. Since the mediator was not randomized, we still considered confounders of the latter, and thus adjusted for any baseline variable associated with psychological distress that could conceivably confound the mediator-outcome association. We adjusted for study site, child characteristics (child’s age, race/ethnicity, enrollment in advanced classes, expulsion, and behavioral/learning problems in school), household size, the presence of a disabled household member, and head-of-household characteristics (never being married, being a teenage parent, being employed, length of residence in the baseline neighborhood, and chatting with neighbors).

Child Opportunity Index data and study design.

Following a geography of opportunity framework (30), we operationalized neighborhood quality using the Child Opportunity Index (COI) (31). The COI includes 19 indicators within 3 domains (educational, health and environmental, social and economic; see Web Table 4) that assesses the degree to which opportunities are available to children and their families in each neighborhood. The COI was constructed at the census tract level for the 100 largest metropolitan areas in the United States, including the 5 MTO metropolitan areas, and provides a standardized index quantifying the population-based level of opportunity by neighborhood. Each indicator was standardized to the metropolitan area; variables belonging to each subdomain were averaged within the subdomain (education, social/economic, health/environment). The average of the scores for the subdomains was then calculated for each tract, to summarize overall opportunity. A higher COI score indicates higher neighborhood opportunity (i.e., quality), with zero indicating the mean opportunity level for the metropolitan area population. We merged these neighborhood-level population-based COI data with the residential histories that we constructed for each MTO household from baseline to interim. The COI is a valid and reliable measure of neighborhood opportunity (31, 32). The COI was operationalized in approximately 2010 (range, 2007–2011), and we merged the census tract COI levels with the MTO census tract of residence at 5 time points: baseline, 90 days postrandomization, 1997, 2000, and 2002. We assumed that the relative value of the COI, had it been operationalized before 2010, would have been at least as relevant as the 2010 value, since neighborhoods are relatively stable over time (33, 34) and change happens slowly (10).

Modeling neighborhood quality across time.

Since there are many strengths of a multidimensional measure of neighborhood quality (31), we hypothesized that the COI would improve upon conventional measures of neighborhood poverty and that change-since-baseline measures would operate differently than simple averages. Therefore, we created 2 different variables: 1) average postrandomization COI, the average COI from 90 days postrandomization through 2002 (omitting baseline COI, since all families lived in comparable very-high-poverty neighborhoods at baseline), and 2) change from baseline COI, the difference between average postrandomization COI and baseline COI, to create a continuous measure of change since baseline. Both COI measures were restandardized within the MTO sample, so that a 1-unit change represents a 1-SD change in COI.

The low-poverty-neighborhood experimental condition required participants to move to neighborhoods with less than 10% poverty, so MTO treatment closely aligned with neighborhood poverty. To rule out the possibility that our findings might have been driven by neighborhood poverty (1 indicator in the COI) and to assess what the COI contributes without neighborhood poverty, we removed poverty and recreated an 18-item “modified COI” following the procedures outlined above. We then recreated the 2 variables average postrandomization COI and change from baseline COI, standardized to the MTO sample. Results from both COI measures are presented side-by-side.

Neighborhood poverty.

We compared neighborhood assessments of census-tract–level poverty with measures based on the COI, since neighborhood poverty is one of the most common measures of neighborhood quality and was directly targeted by the MTO intervention. As with COI, neighborhood poverty was merged with MTO census tracts at baseline, 90 days postrandomization, 1997, 2000, and 2002. We created the variables average postrandomization neighborhood poverty (90 days postrandomization through 2002) and change from baseline neighborhood poverty (difference between average postrandomization and baseline neighborhood poverty), and standardized them within the MTO sample. When modeling average neighborhood poverty, the original poverty measure is multiplied by −1 for interpretability, meaning that a higher value indicates lower poverty; for the decrease in poverty (change measure), a higher value indicates a larger decrease in poverty. So, in all analyses for girls, we expected a negative relationship.

Sensitivity analyses.

Since the COI and poverty were correlated (r = 0.43), we fitted a model where we instrumented for COI while controlling for poverty. We also tested 2 alternative specifications of COI, modified COI, and neighborhood poverty (standardized within MTO), to see if either the short-term or long-term change (compared with baseline) was more informative than the change in the postrandomization average. We operationalized COI and poverty as the difference between 90 days postrandomization and baseline, then as the difference between 2002 (at interim) and baseline, and then refitted all models. We also modeled adherence (or use of the voucher, sometimes called “lease-up”) and COI simultaneously as a test of validity for the IV analysis, to confirm that the effect of lease-up was operating through COI and not other pathways (i.e., addressing the IV exclusion restriction assumption) (35). Finally, since we used a model-based outcome that might have had some uncertainty, we conducted sensitivity analyses in which we jointly fitted the measurement model and the exposure-outcome model (see the Web Appendix for a description of these analyses and Web Table 5 for results).

Analytical methods

We fitted ITT models to test whether randomized MTO treatment predicted psychological distress and separately neighborhood quality. The causal diagram depicted in Web Figure 1 illustrates the novel hypothesis that neighborhood quality over time influenced psychological distress. To test this with high internal validity, we apply IV using 2-stage least squares regression (36–38) with the 3-category treatment variable as the instrument and neighborhood quality as the endogenous variable (Â):

(1)

In the second stage, the predicted value of COI (Â) predicts psychological distress (Y):

(2)

Under IV assumptions, is a consistent estimator of the effect of COI on distress. Note that identical sets of covariates are included in both stages to improve efficiency without introducing endogeneity. We tested whether treatment (the instrument) met IV assumptions, including a first-stage association F test value of 10 or higher (39). The amount of missing data on endogenous variables was small (1%–9%), so we conducted complete-case analyses. We applied survey weights in the analyses and adjusted for household clustering (19). We report robust standard errors with P values from 2-sided tests. All analyses were carried out using ivreg2 in STATA 14.0 (StataCorp LLC, College Station, Texas).

RESULTS

No baseline variables differed by treatment group, confirming proper randomization and exogeneity of treatment (Table 1).

Table 1.

Baseline Characteristics (%) of Girls in the Moving to Opportunity Study, by Treatment Group, United States, 1994–2002^a

Construct and Variable	Treatment Group
	Low-Poverty-Neighborhood Voucher (n = 594)	Section 8 Voucher (n = 389)	Controls (n = 443)
Poverty rate in 1990 Census tract	49.3	48.6	50.1
Family Characteristics
Study site
Baltimore, Maryland	15.2	14.5	14.0
Boston, Massachusetts	17.2	23.2	21.3
Chicago, Illinois	23.5	21.9	19.9
Los Angeles, California	16.7	15.0	19.7
New York, New York	27.4	25.5	25.1
Household size, no. of people
2	7.6	7.5	9.6
3	23.9	21.9	24.4
4	27.9	28.3	23.8
≥5	40.6	42.4	42.2
Household member victimized by crime during past 6 months	44.0	38.2	41.5
Health
Household member had a disability or a health or developmental problem	40.3	36.6	38.4
Household member had a disability	19.6	18.0	15.0
Child Characteristics
Age, yearsb	10.0	10.1	10.0
Race/ethnicity
African-American	62.8	63.0	65.6
Hispanic (any race)	30.6	28.3	26.9
White	1.7	2.4	2.5
Other	4.5	5.0	3.8
Missing data	0.4	1.2	1.2
Special class for gifted students or advanced schoolwork	12.4	16.1	13.3
Developmental problems
Special school, class, or help for learning problem in past 2 years	12.1	12.4	11.3
Special school, class, or help for behavioral or emotional problems in past 2 years	4.8	6.7	3.4
Problems that made it difficult to get to school and/or to play active games	4.6	5.7	5.2
Problems that required special medicine and/or equipment	5.7	7.9	5.3
School asked to talk about problems child having with schoolwork or behavior in past 2 years	16.2	20.7	18.5
Child suspended or expelled from school in past 2 years	6.3	8.6	5.4
Head-of-Household Characteristics
Family structure
Never married	56.4	54.5	57.6
Teenage parent	27.0	28.6	24.4
Socioeconomic status
Employed	25.0	23.6	24.2
On AFDC (welfare)	76.1	75.8	76.0
Completed education
Less than high school	48.4	44.5	45.3
High school diploma	14.5	16.8	21.6
GED diploma	37.1	38.7	33.1
Currently in school	13.6	17.4	13.7
Neighborhood/mobility variables
Lived in neighborhood ≥5 years	62.4	70.4	66.2
No family living in neighborhood	65.6	63.2	66.8
No friends living in neighborhood	39.7	35.3	37.8
Had applied for a Section 8 voucher before	47.8	42.6	47.5
Neighbor relationships
Chatted with neighbors at least once a week	50.1	49.1	51.3
Very likely to tell neighbor if saw neighbor’s child getting into trouble	57.5	59.5	52.3

Abbreviations: AFDC, Aid to Families With Dependent Children; GED, General Educational Development.

^a Test of treatment-group differences calculated from Wald χ² tests outputted from logistic regression for dichotomous baseline characteristics and multinomial logistic regression for categorical characteristics. F tests were used with linear regression for continuous variables. The null hypothesis was that the treatment and control group proportions or means did not differ. All variables ranged between 0 and 1 except baseline age (5–16 years) and mean poverty rate, so mean values represent proportions. The analysis was weighted for varying treatment random assignment ratios across time and for attrition. All tests adjusted for clustering at the family level.

^b Values are expressed as mean.

MTO treatment predicted lower psychological distress at the time of the interim survey (2001–2002), as compared with controls, for the low-poverty-neighborhood group (ITT β = −0.14 (standard error (SE), 0.07), 95% confidence interval (CI): −0.27, −0.01) and less precisely for the Section 8 group (ITT β = −0.10 (SE, 0.07), 95% CI: −0.24, 0.05). First-stage ITT results (Table 2) indicated that MTO treatment most strongly predicted change in average postrandomization COI (model A), although it also predicted the simple average postrandomization (model B). For example (Table 2, model A), girls exhibited a 0.48-point (SE, 0.09; 95% CI: 0.31, 0.65) greater improvement in average COI score over time in the low-poverty group versus the control group and a 0.21-point (SE, 0.08; 95% CI: 0.05, 0.36) greater improvement in average COI score over time in the Section 8 group versus the control group. MTO treatment also predicted the modified COI (excluding neighborhood poverty), as well as neighborhood poverty alone, operationalized as both the change in average (compared with baseline) and the absolute average (after baseline).

Table 2.

Effects of Moving to Opportunity Treatment on Post–Random-Assignment Neighborhood Child Opportunity Index Score and on Post–Random-Assignment Neighborhood Poverty (Intention-to-Treat First-Stage Estimates), Standardized Within the Moving to Opportunity Study Sample, 1994–2002^a

Outcome	Model A: Change in Average (Compared With Baseline)			Model B: Average (After Baseline)
	βb (SE)	P Value	95% CI	βb (SE)	P Value	95% CI
COI (neighborhood poverty included)
LPN voucher	0.48 (0.09)	<0.001	0.31, 0.65	0.36 (0.07)	<0.001	0.22, 0.50
Section 8 voucher	0.21 (0.08)	0.01	0.05, 0.36	0.11 (0.08)	0.17	−0.05, 0.28
Joint test P value		<0.001			<0.001
Modified COI (neighborhood poverty removed)
LPN voucher	0.36 (0.08)	<0.001	0.21, 0.52	0.25 (0.07)	<0.001	0.11, 0.38
Section 8 voucher	0.14 (0.08)	0.08	−0.02, 0.30	0.04 (0.08)	0.59	−0.12, 0.21
Joint test P value		<0.001			<0.001
Neighborhood poverty
LPN voucher	0.79 (0.07)	<0.001	0.66, 0.92	0.85 (0.07)	<0.001	0.72, 0.99
Section 8 voucher	0.64 (0.06)	<0.001	0.52, 0.76	0.74 (0.07)	<0.001	0.61, 0.87
Joint test P value		<0.001			<0.001

Abbreviations: CI, confidence interval; COI, Child Opportunity Index; LPN, low-poverty neighborhood; MTO, Moving to Opportunity.

^a Analytical sample sizes: n = 1,380 for change in average COI; n = 1,422 for average COI; n = 1,411 for change in average neighborhood poverty; n = 1,426 for average neighborhood poverty. We used complete-case analysis to handle missing data. Results were adjusted for baseline characteristics associated with psychological distress: child’s age, child’s race/ethnicity, child’s history of expulsion/suspension, placement in a special class for gifted students/doing advanced schoolwork, or the school calling about a behavioral/learning problem; study site; household size; the presence of a disabled household member; and the household head never being married, being a teen parent, being employed, having lived in the neighborhood for ≥5 years, and chatting with neighbors.

^b Beta coefficients represent standard deviation units.

Both COI measures and neighborhood poverty upheld the IV F test first-stage diagnostics. IV models (Table 3, model A) found that improving average neighborhood COI since baseline by 1 SD lowered psychological distress (β = −0.32 (SE, 0.15), 95% CI: −0.61, −0.03). For the simple average COI value over time after baseline (Table 3, model B), the IV effect estimate was similar but less precisely estimated (β = −0.36 (SE, 0.19), 95% CI: −0.74, 0.02).

Table 3.

Effects of Post–Random-Assignment Neighborhood Child Opportunity and of Post–Random-Assignment Neighborhood Poverty on Psychological Distress Among Girls (Instrumental Variable Estimates), Standardized Within the Moving to Opportunity Study Sample, 1994–2002^a

	Model A: Change in Average (Compared With Baseline)					Model B: Average (After Baseline)
Endogenous Variable	First Stage		Second Stage			First Stage		Second Stage
	F Test	P Value	βb(SE)	P Value	95% CI	F Test	P Value	βb(SE)	P Value	95% CI
COI (neighborhood poverty included)	30.56	<0.001	−0.32 (0.15)	0.03	−0.61, −0.03	27.18	<0.001	−0.36 (0.19)	0.06	−0.74, 0.02
Modified COI (neighborhood poverty removed)	20.33	<0.001	−0.42 (0.20)	0.04	−0.83, −0.02	13.67	<0.001	−0.52 (0.30)	0.08	−1.11, 0.07
Neighborhood poverty	118.83	<0.001	−0.19 (0.09)	0.03	−0.36, −0.02	135.63	<0.001	−0.17 (0.08)	0.04	−0.33, −0.01

Abbreviations: CI, confidence interval; COI, Child Opportunity Index.

^a The instrument of analysis was Moving to Opportunity treatment (0 = control group, 1 = Section 8 group, 2 = low-poverty-neighborhood group). Analytical sample sizes: n = 1,380 for change in average COI; n = 1,422 for average COI; n = 1,411 for change in average neighborhood poverty; n = 1,426 for average neighborhood poverty. We used complete-case analysis to handle missing data. Results were adjusted for baseline characteristics associated with psychological distress: child’s age, child’s race/ethnicity, child’s history of expulsion/suspension, placement in a special class for gifted students/doing advanced schoolwork, or the school calling about a behavioral/learning problem; study site; household size; the presence of a disabled household member; and the household head never being married, being a teen parent, being employed, having lived in the neighborhood for ≥5 years, and chatting with neighbors.

^b Beta coefficients represent standard deviation units.

Modified COI (excluding poverty) reduced psychological distress, with slightly larger effect sizes in comparison with the COI that included neighborhood poverty. Improving the modified average COI since baseline by 1 SD lowered distress by 0.42 SD (Table 3, model A). IV results derived using neighborhood poverty were very similar to those obtained using neighborhood COI, but with smaller effect sizes (Table 3). Improving average neighborhood poverty since baseline by 1 SD lowered distress by 0.19 SD (Table 3, model A).

Sensitivity analyses

In models controlling for poverty (Web Table 6), the effect of change in average modified COI was larger but less precise (β = −0.69 (SE, 0.39), 95% CI: −1.46, 0.08). The effect of average modified COI, adjusting for poverty, was uninterpretable, as high multicollinearity led to a weak F test and imprecise estimates. Alternative specifications of COI and neighborhood poverty (Web Table 7) showed similar patterns as those in the main analysis—for example, a 1-SD increase in COI 90 days postrandomization since baseline significantly predicted lower distress (β = −0.24 (SE, 0.11), 95% CI: −0.47, −0.02)—but the effect was smaller than that estimated for average improvement in COI over 4–7 years (main analysis, −0.32). Improvement in COI in 2002 demonstrated similar patterns with a larger effect size, although treatment was not as strong an instrument (first-stage F = 9.1).

Treatment adherence alone was negatively associated with psychological distress (first-stage F = 190.24; β = −0.32 (SE, 0.16), 95% CI: −0.63, −0.01). When including both COI and adherence as endogenous variables, the estimated effects of adherence were close to the null, but the estimated effects of COI changed little from the estimate using COI as the only endogenous variable. This suggests that adherence had no effect independent of the influence on COI (Web Table 8).

DISCUSSION

Consistent with our hypothesis, moving to higher-quality neighborhoods reduced psychological distress in low-income adolescent girls. Improving neighborhood quality by 1 SD of the COI reduced girls’ psychological distress by 0.29 SD (Cohen’s d (40) = β/sample SD = 0.32/1.12). This effect is small to moderate (41) and is slightly smaller than effect sizes seen from psychotherapy treatments administered to clinical populations (approximately 0.5) (42). This aligns with observational studies (2, 4) which find that better neighborhood quality is associated with better mental health. Our internally valid causal effect was estimated with IV analysis using experimental data, setting it apart from prior neighborhood effect studies using predominantly observational and cross-sectional designs.

To improve upon unidimensional, narrow compositional measures, we defined neighborhood quality using a multidimensional opportunity index, based on conceptual and empirical evidence that multiple neighborhood factors combine to influence families (31). These factors tap concrete aspects of neighborhoods that are sources of both stressors and resources. Guided by the geography of opportunity framework (1, 31), this presents a policy-relevant way to summarize overlapping opportunities, to target neighborhoods for investment and service delivery (31, 43–45). Notably, the multidimensional neighborhood quality measure had a larger effect on distress than neighborhood poverty (Cohen’s d = 0.19/1.12 = 0.17), and COI estimates were even larger when neighborhood poverty was excluded from the COI. Although confidence intervals were wide, effect estimates were larger for 2002 COI (4–7 years after randomization) than average COI across follow-up, or an earlier exposure. Estimates were most precise when evaluating change in average COI over follow-up compared with baseline COI. This suggests that a broader characterization of quality measured over time may be important for evaluating how neighborhoods can improve mental health. Although neighborhoods improved for boys, neighborhood quality increased psychological distress, unlike the case for girls. The adverse effects of treatment on distress for boys are probably attributable to the traumatic effects of moving (23, 24), which may overshadow any benefits of moving to higher-opportunity neighborhoods.

This study has greater policy relevance than prior observational studies for a few reasons. First, the randomized exposure was comparable to the largest current federal affordable housing program, Housing Choice Vouchers (18). Second, this study sheds light on the most relevant modifiable exposures related to affordable housing location (46). Although the objective of the Housing Choice Voucher Program is to increase housing affordability, tailoring the vouchers to promote opportunity moves may substantially impact the health of low-income families. Third, this study describes effects in a low-income, racial/ethnic minority, urban population. Although this may limit generalizability, these are the families targeted by housing assistance/fair housing policies addressing historical housing inequities and discrimination.

Limitations and strengths

The time window for operationalizing the COI (2007–2011) was later than the baseline and interim individual-level MTO data (1994–2002). We assume the relative value of the COI, had it been measured earlier, would have been similar, given prior work documenting the temporal stability in neighborhood quality (33, 34). We acknowledge that there may have been measurement error and that effects may have been underestimated. As an older study, MTO may lack generalizability to other settings and times. However, given that randomized controlled trials are rare, that neighborhood change is slow (10), and that housing vouchers remain the primary avenue for obtaining affordable housing in the United States, MTO still provides the best available evidence to date.

IV analysis relies on several assumptions. The exclusion restriction, which assumes that no path other than the one explicitly modeled (e.g., the COI) connects the instrument with the outcome, is perhaps the greatest threat (35). As a multidimensional measure, we believe the COI blocks other neighborhood-related paths linking the treatment to health better than a single measure like neighborhood poverty. Voucher adherence was unrelated to psychological distress when accounting for COI, suggesting that we blocked all paths emanating from adherence. Theoretically, the further downstream one travels from the instrument, the higher the risk of violation of the exclusion criteria. Operationalizing neighborhood quality over time, as we did, strengthens our confidence in blocking other paths.

It remains possible that the estimated IV association of COI with psychological distress actually reflects that families with successful socioeconomic trajectories were able to move to higher-opportunity neighborhoods. This would be consistent with the larger point estimates based on the 2002 neighborhood. However, confidence intervals were wide because neighborhood COIs at all post–random-assignment time points were strongly correlated, making it difficult to disentangle which time point was most important for the influence of neighborhood opportunity, and whether COI itself is important or is a proxy for family socioeconomic status.

MTO did not measure baseline psychological distress. This may have reduced precision, but it did not bias results because the experimental design ensures balance across treatment groups in baseline psychological distress. Moreover, children were younger than the age at which mental health problems typically emerge at randomization (10 years old, on average). We also used a dimensional measure of mental health (psychological distress) rather than a diagnostic measure favored by clinicians. MTO had comparable effects for dimensional and diagnostic measures, although the prevalence of diagnostic measures (major depressive disorder) was low, limiting statistical power (21, 22). Dimensional measures are appropriate for population-based analyses, aligning with our objective here (47, 48).

Lastly, although IV analysis improves the causal inference for the second-stage association, it is still subject to bias if the endogenous variable (neighborhood quality) is not the actual causal mechanism linking randomization and the outcome but is merely strongly correlated with the true etiological exposure. To mitigate this risk, we used a multidimensional measure of neighborhood quality. Despite the strength of our randomized trial design and application of IV analysis, we still cannot claim with certainty that neighborhood quality is the active ingredient that improved mental health for girls, since it is so difficult to disentangle effects of highly correlated aspects of neighborhoods.

Conclusion

Leveraging policy-induced neighborhood change, our study documented how a multidimensional index of neighborhood opportunity improved mental health for low-income adolescent girls. Our results suggest the importance of looking beyond simplistic measures of poverty to consider other dimensions of social and educational opportunity as important determinants of girls’ mental health outcomes. Policies enacted outside of the health sector, like MTO in the housing sector, may be important levers with which to change social determinants of health in order to improve population health and promote health equity.