The Impact of Public Housing on Social Networks: A Natural Experiment

Abstract

Objectives. We assessed whether 2 types of public housing—scattered among market-rate housing developments or clustered in small public housing projects—were associated with the perceived health and health behaviors of residents’ social networks.

Methods. Leveraging a natural experiment in Montgomery County, Maryland, in which residents were randomly assigned to different types of public housing, we surveyed 453 heads of household in 2011. We asked residents about their own health as well as the perceived health of their network members, including their neighbors.

Results. Residents in scattered-site public housing perceived that their neighbors were more likely to exercise than residents of clustered public housing (24.7% of network members vs 14.0%; P < .001). There were no significant differences in the proportion of network members who were perceived to have major health problems, depressed mood, poor diet, or obesity. Having more network members who smoked was associated with a significantly higher likelihood of smoking.

Conclusions. Different types of public housing have a modest impact on the health composition of one’s social network, suggesting the importance of housing policy for health.

Multiple housing policies aim to reduce concentrated poverty in neighborhoods for low-income residents who receive federal housing subsidies. The US Department of Housing and Urban Development has attempted to disperse concentrated poverty through its Housing Choice Vouchers program and initiatives such as HOPE VI and Choice Neighborhoods, which replace public housing complexes with mixed-income developments.¹ Court cases have instigated housing relocation programs intended to increase access to opportunity.^2,3 Some municipalities have adopted inclusionary zoning policies in which developers set aside a portion of homes to be sold or rented at below-market rates. Policies that deconcentrate poverty may improve residents’ health and well-being. Most prominently, the Department of Housing and Urban Development’s Moving to Opportunity randomized experiment found that recipients of vouchers to move to low-poverty neighborhoods experienced reduced obesity, diabetes, and psychological distress^4,5 and improved mental health and happiness compared with those who remained in public housing developments.⁶

One way that public housing may influence health is by shaping social networks. Social networks represent the web of relationships that exists among people; they consist of social ties that link individuals in a social network.⁷ Over the past century, social network theories and analytic methods have developed and been increasingly applied in public health.^8–10 Research suggests that multiple factors influence the formation of social ties including similarity between individuals (homophily), having relationships in common, and the frequency and duration of contact with one another.^11–13

Theoretically, public housing may affect social networks by changing the neighbors with whom residents come into contact and the frequency of these contacts. Previous research has shown that residents living in subsidized housing next to more affluent neighbors may have more socioeconomically diverse social networks than individuals living in public housing developments.^14,15 Different public housing arrangements such as clustering housing into projects or scattering units among market-rate developments, may affect the supportive quality and emotional intimacy of relationships within public housing residents’ social networks.^16–21

Social networks and ties have been increasingly shown to influence a wide range of conditions and behaviors including obesity,^22–25 physical activity,^26–31 alcohol and drug use,^32–35 and smoking.^36,37 Researchers postulate that social networks may induce changes in health and behavior through altering social norms and beliefs.^11,38 Studies suggest that social networks’ influence extends beyond a single degree of separation,^22,36 and research on vulnerable populations has highlighted the influence of social network composition on health behaviors.^39,40

Although social networks may be an important mechanism through which public housing policies affect health, to our knowledge, only 1 study has explicitly examined the connection between social networks and health behaviors among public housing residents. Shelton et al. found social network size to be associated with physical activity among Boston public housing residents.⁴¹

We sought to address 2 research questions regarding the potential relationship between public housing policy and social networks and health. First, we asked whether the type of public housing (scattered vs clustered) influenced the composition of adult public housing residents’ social networks with a focus on perceived health and health behaviors of respondents’ social network members. Second, we determined whether characteristics of these network members were associated with residents’ health behaviors.

Our study was set in Montgomery County, a Maryland suburb of Washington, DC. Unlike earlier studies such as Moving to Opportunity, in which participants initially lived in high-poverty neighborhoods, in this study the public housing residents live in low-poverty neighborhoods in an affluent county. The median household income from 2007 to 2011 in Montgomery County was $96 000 compared with the national average of $72 000, and the poverty rate was 6% compared with 9% nationally.

Public housing residents in Montgomery County live in homes that are either scattered among market-rate housing developments or clustered in small public housing projects. The Housing Opportunities Commission (HOC), the county’s public housing authority, has purchased 670 scattered-site public housing homes through Montgomery County’s inclusionary zoning program. Through inclusionary zoning, developers set aside 12% to 15% of homes to be sold or rented at below-market prices in exchange for a density bonus that offsets the financial loss. In the developments where the HOC has purchased homes created through inclusionary zoning, no more than 5% of residents live in public housing. The HOC also operates 321 public housing homes that are clustered within 7 developments ranging in size from 19 to 71 homes. In these developments, all residents live in public housing, creating microneighborhoods of poorer people. Although both scattered and clustered public housing units are located in wealthy neighborhoods (Table 1), we explored potential differences among residents’ more immediate neighbors.

TABLE 1—

Sociodemographic Characteristics of Respondents by Housing Type Among Adults Living in Scattered- and Clustered-Site Public Housing: Montgomery County, MD, 2011

Characteristic	All Public Housing	Clustered Public Housing	Scattered Public Housing	P
Unweighted no. (%)	453	161 (36)	292 (64)
Weighted no. (%)	452	153 (34)	299 (66)
Mean age,a y	44	44	44	.721
Female,a %	88	87	88	.769
Race/ethnicity,a %
Hispanic	15	14	16	.609
Black	69	71	68	.534
Asian	3	3	2	.895
White	13	12	14	.707
Has a spouse or partner, %	18	21	16	.158
Citizen,a %	84	83	85	.556
Language other than English spoken at home, %	27	25	28	.45
Parent lived in public housing, %	21	23	20	.547
Time lived in neighborhood, y, %
0–2	28	30	27	.514
2.1–6	25	27	24	.5
7–11	22	19	24	.249
12–37	24	23	25	.767
Income-to-poverty ratio,a 2011	1.15	1.05	1.19	.131
Unemployed, %	27	24	29	.276
Education, %
< high school	36	39	34	.234
Completed vocational school	31	28	33	.282
Completed high school, some college, or associate’s degree	16	13	18	.158
≥ completed college	17	20	16	.283
Census-tract median income, $	95 454	92 722	96 866	.134

Note. Results are weighted to reflect characteristics of the broader Montgomery County public housing family population.

^aData derived from the county housing authority's annual recertification records.

The random assignment of households to scattered and clustered public housing creates a natural experiment. Households are offered homes as homes become available through computerized, rolling lotteries. The strong demand for public housing—with long wait lists and a large difference between public housing and market-rate rent—helps minimize bias in housing assignment. The appendix (available as a supplement to the online version of this article at http://www.ajph.org) provides detail on the natural experiment, and supplementary Table A demonstrates the comparability of families in scattered- versus clustered-site public housing.

With residents of scattered public housing units having greater exposure to higher-socioeconomic-status (SES) neighbors and with the well-established association between SES and health,^42–44 we hypothesized, for our first research question, that residents of scattered public housing would perceive that their network members were healthier.

Our second research question explored whether characteristics of public housing residents’ network members were associated with residents’ health. This question did not directly leverage the natural experiment of Montgomery County public housing assignment, but instead examined associations within the entire public housing cohort.

METHODS

We performed an in-person, computer-assisted survey of adults living in scattered- and clustered-site public housing within Montgomery County. The results of this survey were combined with household-level data from the county’s public housing authority and neighborhood characteristics from the 2006–2010 American Community Survey.

There were 991 total public housing units in Montgomery County for nonelderly households in 2011, of which 670 lived in scattered housing and 340 in clustered housing. In August 2011, 948 of the 991 homes were occupied. After an introductory mailing explaining the study, 148 of 948 households (16%) opted out of the survey. An additional household moved in during the study period, leaving 801 households eligible for recruitment. We contacted households first via telephone calls, then door knocking, and then postcards for an average of 5.3 contacts per household. We conducted interviews with the head of household (all public housing residents per federal regulation designate a head of household for records), which we performed in English and Spanish between October 1 and December 23, 2011. The response rate was 57% (453 out of 801). Responders were generally representative of the entire sample of households in Montgomery County.

Participant Indicators

We measured self-rated health with a single item with responses categorized as excellent, very good, or good versus fair or poor. We assessed depression with 8 items from the Patient Health Questionnaire with a score of 10 or higher indicating symptoms of depression.⁴⁵ We classified respondents who stated that they smoked every day or some days as current smokers.

To assess diet, we asked respondents, on a typical day, how many servings of fruit and vegetables they ate. On the basis of current recommendations,⁴⁶ we classified those reporting more than 4 servings as having an adequate intake. We also asked respondents how often they drank a sugar-sweetened beverage (e.g., soda, sports drinks) on a typical day, and we dichotomized responses on the basis of current recommendations as none versus 1 or more.

We asked respondents how often they engaged in vigorous and moderate physical activity for at least 20 minutes over the past week. We coded participants as physically active if they engaged in approximately 75 minutes per week of vigorous or 150 minutes per week of moderate physical activity.⁴⁷

We calculated body mass index (defined as weight in kilograms divided by the square of height in meters) on the basis of self-reported height and weight with overweight or obese characterized as a body mass index greater than 25.

Using the HOC’s annual recertification records for public housing households, we obtained data about household size, composition, race, ethnicity, date of first entry into the home, citizenship, and income. We constructed household income-to-poverty ratio on the basis of household income and size. From the survey, we obtained data on unemployment and educational attainment.

Social Network Assessment

Social network analysis provides an important tool for identifying and understanding the social and contextual factors relevant to engagement in particular behaviors, and has 2 distinct approaches.⁸ The sociocentric, or complete, network approach involves quantifying relationships between people within a defined group. The egocentric, or personal, network approach focuses on the network members that surround a focal individual (referred to as the “ego” and a member of the study sample; the ego’s social ties are known as “alters”).⁴⁸ We used the personal approach to create summary variables measuring composition and structure of each respondent’s network.

We first asked respondents to name up to 20 people (alters) with whom they had contact sometime over the past year. We asked respondents to start by naming the people most important to them. We then asked respondents which of the alters lived in their neighborhood. Participants who named fewer than 5 neighbors were prompted to name additional alters (up to 5 neighbors). Thus, respondents could have named up to 25 alters; the mean number of alters was 17.6.

To assess network members’ health and health behaviors, respondents identified whether each alter had major health problems, was generally sad or depressed, smoked cigarettes, ate a healthy diet, exercised regularly, or was overweight or obese.

To establish the structure of the social network, we asked respondents whether each pair of alters knew and had contact with one another. We considered a pair of alters who sometimes or often had contact with one another in the past year to be connected to one another.⁴⁸ We examined social network density, which represents the proportion of existing connections in a network relative to all possible connections. Consistent with the previous literature,⁴⁹ we excluded connections between the respondent (ego) and each of the alters from the density calculation.

Statistical Analysis

We first examined whether housing type—clustered versus scattered public housing—was associated with differences in the composition of one’s social network. Our outcomes were the percentage of alters with each health characteristic. Because we hypothesized that housing type may have the largest impact on the characteristics of neighbors within one’s social network, our primary analysis focused on alters who were either current neighbors or people met through the neighborhood. Although there was appropriate randomization on observable characteristics between scattered and clustered households, we opted to adjust for “pretreatment” characteristics (i.e., characteristics that would not change with public housing assignment) to increase the precision of our models. In our first model, we included age, gender, race/ethnicity, marital status, citizenship, English spoken at home, and number of years in the neighborhood (in quartiles to allow for nonlinearities). Our second model added potential confounders of the relationship between social network composition and housing type. Specifically, we added household socioeconomic characteristics—income-to-poverty ratio, unemployment, and educational status—and social network density. In this model, it is possible that household socioeconomic factors and network density may be both affected by housing type (e.g., housing type may lead to higher levels of household income) and related to the composition of one’s social network. Sensitivity analyses included all alters in respondents’ social networks.

We then sought to examine whether the composition of one’s social network was associated with the respondent’s health and health behavior. This research question examined associations within the entire cohort, with adjustment for housing type. We ran separate logistic regression models for each dependent health variable. Our main independent variables were the corresponding compositional feature of one’s social network and included all alters (i.e., not only those identified as neighbors). For example, in the model that examined the odds of depression, our main independent variable was the percentage of alters whom respondents perceive as sad or depressed. We additionally included age, gender, race/ethnicity, marital status, citizenship, English spoken at home, whether a parent lived in public housing, number of years in the neighborhood, and housing type, as well as household SES and network density.

All models included sampling weights, which made the households who enrolled representative of the county’s public housing population. To make our results more interpretable, we used recycled predictions to estimate the average outcome associated with different types of public housing and varying composition of one’s social networks.⁵⁰ For each analysis, we also present odds ratios in appendix tables (available as a supplement to the online version of this article at http://www.ajph.org). Because of the number of hypotheses tested, we have adjusted all reported tests of statistical significance by using Holm–Bonferroni corrections for the possibility of finding statistically significant differences by chance.

RESULTS

In total, we interviewed 453 out of 801 adult heads of household for a response rate of 57%. Responders were not statistically different from eligible Montgomery County public housing residents on observable characteristics with the exception of percentage of Asian households and percentage of US citizens (Table A, available as supplement to the online version of this article at http://www.ajph.org). Two thirds of the adult respondents (n = 292) lived in scattered-site public housing and one third in clustered public housing (n = 161). The mean age was 44 years, 88% were female, and more than two thirds were Black (Table 1). There were not significant sociodemographic differences between respondents who lived in clustered- and scattered-site public housing, and the median Census tract income was high regardless of housing type.

Table 2 presents the health and social network characteristics of our sample. Nearly one third (32%) reported fair or poor self-rated health; more than a quarter (28%) were current smokers; and 15% had depressive symptoms. With regard to diet, 26% reported adequate fruit and vegetable intake and more than half (53%) drank sugar-sweetened beverages daily. Physical inactivity (57%) and obesity (49%) were common. These characteristics were not significantly different between housing types.

TABLE 2—

Respondent Health and Social Network Characteristics by Housing Type Among Adults Living in Scattered- and Clustered-Site Public Housing: Montgomery County, MD, 2011

Characteristics	All Public Housing Homes	Clustered Public Housing	Scattered Public Housing	P
Respondent, %
Self-rated health is fair or poor	32	34	32	.588
Smokes	28	31	26	.24
Has depressive symptoms	15	16	15	.66
Adequate fruit and vegetable intake	26	27	26	.72
Consumes ≥ 1 sugar-sweetened beverage per day	53	58	50	.145
Is physically active	43	43	43	.961
Is obese	49	54	47	.153
Social network
Number of alters	17.6	17.5	17.7	.778
Alters who are female, %	62	64	62	.338
Alters who live in neighborhood or who were met because respondent lives in neighborhood, %	21	21	21	.927
Alters who are friends, %	52	53	52	.823
Alters who are family, %	34	34	34	.92
Network density	0.30	0.29	0.30	.749

On average, 62% of alters were female, 34% were family members, 52% were friends, and 21% were neighbors or were met because the respondent lived in the neighborhood. The average network density was 0.30. These social network characteristics did not significantly vary by whether the respondent lived in scattered-site or clustered public housing.

Housing Type as Related to Network Characteristics

Figure 1 presents the results of the predicted percentage of alters for respondents in clustered- and scattered-site public housing who were perceived to have each health characteristic (odds ratios are presented in Table B, available as a supplement to the online version of this article at http://www.ajph.org). The analyses focused on those alters who were either current neighbors or met through the neighborhood. We observed significant differences in the proportion of alters who smoked and who exercised regularly by housing type. Among respondents who lived in clustered public housing, 23.2% of alters smoked compared with 17.1% of alters among scattered-site housing residents (based on recycled predictions from our first model). We observed lower rates of regular exercise among alters of clustered housing compared with alters of scattered-site residents (14.0% vs 24.7%). After we adjusted for multiple testing, only differences in the rates of regular exercise remained significant. Results for exercise remained significant in our second model, which also adjusted for socioeconomic factors and network density. We did not observe a significant association by housing type for other measures of alter health characteristics.

FIGURE 1—

Predicted percentage of alters with each characteristic by respondent housing type among adults living in scattered- and clustered-site public housing: Montgomery County, MD, 2011.

Note. Alters are limited to current neighbors or alters who were met through the neighborhood. Predicted probabilities adjust for age, gender, race/ethnicity, marital status, citizenship, English spoken at home, whether a parent lived in public housing, and number of years in the neighborhood. We computed predicted probabilities by using our outcome model to predict outcomes for each individual in our data assuming that they had lived in scattered housing and then again for each individual in our data assuming that they had lived in clustered housing. For each individual, all other covariates remained fixed at their observed values when we computed the predicted values.

*P < .05 after accounting for multiple testing.

When we examined the characteristics of all alters in respondents’ social networks, we found that respondents in clustered-site public housing reported a lower percentage of alters who exercised regularly compared with scattered-site respondents (19.7% vs 24.7%; P = .018; based on recycled predictions from our first model), though results did not remain significant after we adjusted for multiple testing. Other comparisons in alter characteristics were not statistically significant.

Network Characteristics as Related to Health

In our next series of analyses, we examined whether characteristics of one’s network members were associated with the respondent’s own health and health behaviors. Figure 2 shows the predicted probability of the respondent’s own health based on (1) the average proportion of his or her alters with each characteristic and (2) a 10-percentage-point increase in alters with each characteristic (for odds ratios see Table C, available as a supplement to the online version of this article at http://www.ajph.org). The corresponding alter characteristics were not significantly associated with respondent’s health problems, fruit and vegetable intake, sugar-sweetened beverage consumption, or obesity.

FIGURE 2—

The association between alter characteristics and respondent health and health behaviors among adults living in scattered- and clustered-site public housing: Montgomery County, MD, 2011.

Note. The circle represents the predicted probability for a respondent whose alters have the average percentage of the corresponding health characteristic, and the triangle represents the predicted probability for a respondent whose alters have a 10-percentage-point increase in the corresponding health characteristic. With respondent health status as an example, the corresponding alter health characteristic is the proportion of alters whom the respondent perceives to have a major health problem. Predicted probabilities adjust for age, gender, race/ethnicity, marital status, citizenship, English spoken at home, whether a parent lived in public housing, number of years in the neighborhood, housing type, income-to-poverty ratio, unemployment, educational status, and network density. We computed predicted probabilities by using our outcome models to predict outcomes for each individual in our data assuming that they had the population average of the network composition variable and then again for each individual in our data shifting this percentage by 10 percentage points to understand how a 10-percentage-point change in these measures would affect outcomes. All control covariates remained fixed at their observed values.

*P < .05.

An increase in the proportion of alters who smoked was associated with a higher likelihood of a respondent smoking. Respondents with the mean proportion of alters who smoked had a predicted rate of smoking of 26%. A hypothetical increase in the proportion of alters who smoked by 10 percentage points—or by approximately 2 more smokers in one’s social network—increased a respondent’s predicted rate of smoking to 34% (P < .001). A 10-percentage-point increase in the proportion of alters with depressive symptoms was associated with a higher rate of respondent depression (a 3-percentage-point increase; P = .015), though this result did not remain significant when we adjusted for multiple testing. Examination of models in which respondents’ network characteristics were (1) dichotomized as having at least 1 alter with the corresponding health characteristic versus none and (2) placed into quartiles on the basis of the percentage of alters with the corresponding health characteristic yielded qualitatively similar results (Table D, available as a supplement to the online version of this article at http://www.ajph.org).

DISCUSSION

On the basis of analyses from a natural experiment in public housing, we found that residents in scattered-site public housing—where the large majority of their immediate neighbors have higher incomes—perceived that their neighbors were more likely to exercise regularly compared with residents in clustered public housing. However, other health-related characteristics of respondents’ social networks were not significantly different between the 2 groups. Analyses also revealed the health composition of one’s social networks was modestly associated with one’s own health behaviors, particularly with respect to smoking. Overall, the results underscore both the opportunity and limitations of public housing’s ability to shape social networks and suggest a potential link between these networks and one’s health behaviors.

We did not observe as many significant health-behavior differences in the social networks of clustered and scattered site public housing residents as we had hypothesized. One potential explanation is that homophily—the tendency for people with similar characteristics to form relationships with one another³⁸—may limit the ability of residents of scattered housing to form relationships with people who have different health behaviors. An alternate explanation is that we may have been underpowered to detect meaningful differences in the social networks between residents of different housing types. Moreover, the location of clustered- and scattered-site public housing in neighborhoods with almost equally high median household incomes may have blunted potential differences in social networks. Given the similarity of the broader neighborhoods along observed demographic characteristics, the current study only tests how differences in immediate neighbors affect low-income adults’ social networks.

Notably, the 10-percentage-point difference in perceived rates of exercise in scattered compared with clustered public housing respondents’ social networks may have public health significance. Given that Moving to Opportunity found that public housing assignment was associated with rates of obesity and diabetes,⁴ it is plausible that perceived exercise among one’s network members may be 1 factor that contributes to the increased risk.

Rates of smoking were high among public housing residents—across housing types—compared with nationally (28% vs 19%, respectively⁵¹). Smoking among one’s social networks was significantly associated with whether an individual smoked, which may suggest the potential for interventions to target the social context and social network aspects of smoking among public housing residents. This finding builds upon earlier research, which has shown that not only do smokers cluster in social networks but also that having network members who stop smoking increases an individual’s likelihood of quitting.³⁶

Limitations

This study has multiple limitations. First, as noted previously, the egocentric approach examines perceptions of one’s network members rather than alters’ actual health and health behaviors. It may be easier for participants to observe some behaviors (e.g., exercise and obesity) than others (e.g., dietary habits), especially among network members to whom the participant is less close. Likewise, our measures of respondent health and health behaviors are self-reported and thus subject to bias. Though respondents with a particular health characteristic may be more likely to perceive the characteristics among their network members, there is no evidence to suggest that perceptions of alters or self-reported health would differ by housing type.

Second, we elicited up to 25 network alters. Although previous research has found that even lower numbers of alters results in stable measures of network structure,⁵² any cap on the number of alters may affect the composition and structure of one’s network. Third, our study is cross-sectional. We are unable to assess directionality in the associations between network characteristics and respondent outcomes; the observed associations may be attributable to influence (network members causing one’s behavior), homophily (people with like characteristics forming social ties), or context.⁵³ Fourth, nonresponse bias may have had an impact on our results, although not differentially by housing type. Finally, as noted previously, the current study took place within the context of Montgomery County and may not be generalizable to other regions.

Conclusions

Our study reinforces existing evidence that public housing residents are vulnerable from a health perspective.^54–56 We found that the health status of one’s social network members varies modestly between residents of clustered public housing projects and residents of units scattered into market-rate developments where neighbors are not poor. In particular, living in scattered public housing units—placing residents in closer proximity to higher-SES neighbors—is associated with the perception that more network members exercise regularly. We also found that increases in the proportion of smokers in a participant’s social network are significantly associated with the likelihood a participant smokes. Although this evidence is insufficient to adjudicate policymaking to favor scattered-site public housing, these results explore mechanisms that may contribute to health among public housing residents.