Neighborhood Contextual Influences on Depressive Symptoms in the Elderly

To understand whether neighborhood contexts contribute to the onset or maintenance of mental health problems independently of individual characteristics requires the use of multilevel study designs and analytical strat-egies. This study used a multilevel analytical framework to examine the relation between neighborhood context and risk of depressive symptoms, using data from the New Haven component of the Established Populations for Epidemiologic Studies of the Elderly, a community-based sample of noninstitutionalized men and women aged 65 years or older and living in the city of New Haven, Connecticut, in 1982. Neighborhoods were characterized by census-based characteristics and also by measures of the neighborhood service environment using data abstracted from the New Haven telephone book Yellow Pages. Living in a poor neighborhood was associated with higher levels of depressive symptoms in older adults, above and beyond individual vulnerabilities. In addition, the presence of more elderly people in the neighborhood was associated with better mental health among older adults. The authors found no evidence that access to services hypothesized to promote social engagement, to provide health services, or to affect the reputation of a neighborhood explained (i.e., mediated) neighborhood variations in depressive symptoms.

In their classic work, Mental Disorders in Urban Areas: An Ecological Study of Schizophrenia and Other Psychoses, Faris and Dunham (1) argued that residence in deprived neighborhoods with high residential turnover contributed to the onset of schizophrenia and substance abuse. Since then, researchers have sought to refine both the theory and the empirical evidence linking urban residential environments to risk of mental illness. For example, the social stress model posits that disadvantaged neighborhoods with high residential mobility make it difficult for individuals to sustain supportive social contacts with others, thereby rendering them more vulnerable to stress and mental illness (2).
The original work by Faris and Dunham (1) was based on ecologic evidence from the city of Chicago, Illinois. Skeptics have argued that the apparent link between neighborhood environments and mental health arises because certain characteristics lead individuals both to reside in deprived neighborhoods and to experience poor mental health. To test the proposition that neighborhood contexts contribute to the onset of mental health problems independently of individual characteristics requires the use of multilevel study designs and analytical strategies. Yet only a handful of multilevel investigations have been carried out to date (3)(4)(5)(6)(7)(8)(9). These studies found that neighborhood deprivation is associated with increased risk of schizophrenia (3,7) and depression (4)(5)(6)9), after taking account of individual characteristics. These studies tested primarily the associations among middle-aged adults, and whether they are broadly applicable to other age groups (e.g., the elderly) is as yet unknown.
One recent study considered two specific aspects of neighborhood environment in relation to depression among older Mexican Americans (9). Findings suggested that greater neighborhood poverty was associated with higher levels of depressive symptoms, while a greater neighborhood concentration of Mexican Americans was associated with lower levels, after taking account of a range of individual-level factors. Such findings suggest that elderly mental health may be particularly sensitive to ambient neighborhood conditions because, as a group, elderly persons tend to be less mobile and more reliant on locally provided services and amenities, as well as sources of social support and contact. Moreover, depression in people aged 65 years or older is a major public health problem (10). High levels of depressive symptoms (subclinical depression) are associated with increased risks of major depression, physical disability, medical illness, and high use of health services, with US prevalence rates estimated from 13 percent to 27 percent among community-dwelling elderly (10).
Existing multilevel studies have just begun to explore the mechanisms underlying the relation between neighborhood contexts and mental illness. Most multilevel studies have been limited to documenting the contextual effect of neighborhood deprivation, as measured by such census-derived variables as the poverty rate (11). Studies are needed to unpack the mechanisms by which neighborhood disadvantage leads to adverse mental health outcomes. That is, we need to understand what it is about neighborhood deprivation that produces differential patterns of risk and protection. This task requires going beyond the use of census-derived indicators of disadvantage and moving toward defining, operationalizing, and measuring specific neighborhood characteristics (e.g., local access to services and amenities), specific exposures (e.g., crime and vandalism), and social processes (e.g., behavioral contagion and social cohesion) (11).
The present study has two aims. First, we sought to test the relation between neighborhood context and risk of depressive symptoms in a representative survey of community-dwelling elderly in New Haven, Connecticut. The goal of this part of the study was to assess the independent contributions of neighborhood disadvantage and individual characteristics to depressive symptoms in the elderly within a multilevel analytical framework. Relevant characteristics were chosen on the basis of prior findings that factors defining or shaping exposure to adverse circumstances and the availability of resources may strongly influence the likelihood of experiencing depressive symptoms (12). These factors include socioeconomic status, gender, marital status, race/ethnicity, age, and physical disability, and they have been demonstrated to remain potent among adults aged 75 years or older (13,14). Second, we sought to develop a set of indicators to characterize the neighborhood service environment using data abstracted from the New Haven telephone book Yellow Pages. We developed neighborhood-level density measures of three types of services: 1) services that promote social interactions, 2) services that provide health care, and 3) services that adversely affect the reputation of a neighborhood (e.g., liquor outlets, pawnbrokers). We examined the relation of each of these service types with depressive symptoms, testing for both main effects and mediating effects, that is, whether differential ''exposure'' to these services mediated any relation between neighborhood socioeconomic characteristics and depressive symptoms.

Sample and procedure
This study is based on the New Haven component of the Established Populations for Epidemiologic Studies of the Elderly (EPESE), a community-based sample of noninstitutionalized men and women aged 65 years or older and living in the city of New Haven, Connecticut, in 1982. The New Haven EPESE cohort is based on a stratified probability sample of 2,812 noninstitutionalized men and women. The sampling design has been described elsewhere (14). Briefly, the sampling design was based on three housing strata: community housing, age-restricted private housing, and age-and income-restricted public housing for the elderly. The overall response rate was 82 percent. The original sample comprised 79 percent non-Hispanic Whites and 19 percent non-Hispanic Blacks. Two percent gave some other response and have been excluded from these analyses.
Participants were interviewed in person at baseline with subsequent in-person or telephone interviews at almost yearly intervals. The analyses presented here are based on data from 1985, the time closest to the earliest availability of relevant Yellow Pages data for characterizing the neighborhoods where participants lived. Information on age, gender, education, and race/ethnicity was obtained at the baseline interview in 1982. All other individual-level variables included in the analyses were collected during the 1985 interview. After exclusion of those who had died since 1982 (17 percent), refused to participate, or were lost to follow-up (5 percent), 2,109 individuals were available for analysis.

Measures
Assessment of depressive symptoms. Depressive symptoms were assessed using the 20-item Center for Epidemiologic Studies Depression (CES-D) Scale (15). Of the 2,109 respondents in 1985, sufficient responses to the CES-D questionnaire were available for 1,926. Those who failed to answer at least 17 items were excluded (9 percent). Of these respondents, most were unable to complete most parts of the interview, or the questionnaire was responded to by a proxy. A small number of subjects (2 percent) had data missing on covariates other than income (see description of covariates below). Thus, 1,884 individuals remained available for analyses. We did not detect systematic differences in relation to gender or race/ethnicity between those who were and were not included in the analyses. Individuals included in the analyses were younger, were more likely to be married, had higher levels of income and education, and were less likely to be missing data on income than those excluded from the analyses.
Each item is scored on a standard four-point scale (0-3 points), with scores for positively worded items reversed; high scores represent more depressive symptoms. Total CES-D scores were derived by taking a sum of the items. For those with three or fewer missing items, we imputed the overall score by assigning the average based on the remaining items. The potential range of the scale is 0-60, while the actual range was 0-47, with a mean of 7.9 (standard error: 0.25). Scores were somewhat skewed toward the bottom of the scale. The internal reliability consistency coefficient in this sample was 0.88.
Assessment of covariates. Information on age, gender, marital status, years of education, level of household income, and race was obtained from structured interviews. Age was measured in years. Gender was a dichotomous variable coded 1 for females and 0 for males. Education was categorized on the basis of whether an individual had less than a high school education, completed high school, had some college education, or completed a college education or beyond (referent group). Household income was coded as less than $5,000 per year, $5,000-$9,999 per year, $10,000-$14,999 per year, or greater than or equal to $15,000 per year (referent group). We also included a category for missing data on income, given the sizable percentage of respondents missing income data (n ¼ 278). Race was a dichotomous variable coded 1 for non-Hispanic Blacks and 0 for non-Hispanic Whites. Marital status was categorized into three groups: 1) married (referent group), 2) widowed, and 3) unmarried, divorced, or separated. Functional disability was a continuous measure based on 15 items designed to assess basic activities of living, gross mobility function, and physical performance (14).
Assessment of neighborhood characteristics. Following previous studies, we used census tracts as proxies for neighborhoods. This follows existing convention in epidemiologic research, although it may not be the ideal approach. On the other hand, census tracts were created by the Bureau of the Census to represent reasonably homogeneous sociodemographic groupings of residents. Tract data were obtained from the 1980 US Census summary tape files for the 28 census tracts of New Haven to summarize neighborhood differences in structural characteristics. On the basis of previous research and theory, five neighborhood characteristics were identified for study, including neighborhood socioeconomic disadvantage (percentage of people living in poverty), racial/ethnic heterogeneity (percentage of Black residents), residential stability (percentage of individuals who have been in their homes longer than 5 years), age structure (percentage of individuals aged over 64 years), and socioeconomic advantage (affluence; percentage of individuals with income greater than $75,000 per year) (4,5,16).
An alternative method for characterizing neighborhoods was also developed. Using information abstracted from the Yellow Pages, we developed neighborhood-level density measures to assess access to a wide variety of services that theoretically affect health outcomes (e.g., functional independence and mobility). Two of the authors developed a coding scheme for abstracting a list of neighborhood services and amenities from the 1985 New Haven Yellow Pages. Each investigator independently developed a listing of such services, based on a page-by-page analysis of the Yellow Pages. Examples of services included health services (e.g., hospitals, audiologists), financial services (e.g., banks), social organizations (churches), recreational facilities, grocer-ies and food outlets, and places of social interaction (e.g., beauty parlors, cafes). An index of potentially ''undesirable'' amenities was also developed (e.g., liquor outlets, pawnbrokers). After these lists were independently developed, a third independent investigator compared the two lists to obtain a measure of interrater agreement (>85 percent). Categories that resulted in disagreement were later reconciled by the investigators.
Every establishment that made it to the finalized list of services and amenities was then abstracted from the Yellow Pages by a research assistant, and its 1985 street address was geocoded. These data were entered, checked, and validated. We then developed census tract-level measures of service density (i.e., number of services per total population), with the denominators obtained from the census. Services were grouped a priori into one of three categories: 1) services promoting social engagement, that is, places where elderly residents could potentially engage in social interactions (e.g., beauty salons, the public library); 2) services providing care, that is, those that provide health services; and 3) ''undesirable'' amenities, that is, those that may promote perceptions of lack of safety or decline in a neighborhood's reputation (e.g., guns and gunsmiths). A complete listing of the services and their relevant categories may be found in appendix table 1. Service density in each category was calculated by dividing the count of the services in a given neighborhood by its population.

Modeling overview
A major issue in neighborhood studies is being able to distinguish true contextual effects from compositional effects. Briefly, compositional effects refer to ''the difference that people make to neighborhoods,'' whereas contextual effects refer to ''the difference that neighborhoods make to people.'' For instance, if it is observed that neighborhood X has a higher mortality rate than neighborhood Y, the difference could be entirely due to the fact that neighborhood X has more residents (compared with neighborhood Y) who are individually at higher risk of mortality (e.g., poor, smokers, obese, and so on). This exemplifies a compositional effect; that is, the worse health in neighborhood X is entirely explained by the poor health risk of the residents who ''make up'' that neighborhood. On the other hand, a contextual effect is implied if neighborhood X has a higher mortality rate even after taking account of all known relevant differences (between neighborhoods X and Y) in the characteristics of individual residents. Examples of contextual neighborhood effects on health include exposures to pollution, crime, or violence. Disentangling compositional effects from contextual effects requires a multilevel analytical approach, in which information is available at both the individual and neighborhood levels (17). As a result, our analyses include key individual characteristics that might be linked with depressive symptoms: gender, age, race/ethnicity, marital status, disability, household income, and educational attainment. Because of the somewhat skewed distribution of areabased service measures, scores were dichotomized into high or low for each of the three categories on the basis of a median split.

Analytical approach
To address the complex sample design, we assigned sampling weights to respondents to adjust for differential sampling, response, and coverage rates, and analytical models and tests of significance used estimates of variance that take account of the complex sampling design (14). This strategy makes it possible to draw inferences to the larger defined population of elderly in New Haven. Reported sample characteristics are based on the underlying sample design.
We used a multilevel linear regression model with the structure of 1,884 individuals (level 1) nested within 28 neighborhoods (level 2). These models allow the estimation of 1) the conditional relation between depressive symptoms and individual predictors (''fixed parameters''), 2) variation between census tracts that cannot be accounted for by individual predictors (''random parameters''), and 3) the main effect of neighborhood predictors on depressive symptoms (''fixed parameters''), conditional upon the individual-level relation between depressive symptoms and individual sociodemographic indicators. Models were calibrated using maximum likelihood estimation, as implemented within MlwiN software, version 1.10.006 (Institute of Education, University of London, London, United Kingdom), that utilizes the iterative generalized least-squares algorithm (18). Since the survey data oversampled the elderly from particular housing units (stratified by age and gender), model estimates are weighted to a comprehensive sampling weight. Four types of models were developed. The first examined the unadjusted fixed effect of each neighborhood-level marker on individual depressive symptoms. The second examined the fixed relation between individual demographic markers and depressive symptoms, conditional on a random effect for census tracts, but unadjusted for neighborhood characteristics. In the third type of model, the fixed effect of each neighborhood characteristic was reestimated after adjustment for the individual-level relation between depressive symptoms and sociodemographic markers. The final type of model considered the extent to which the service density characteristics derived from the Yellow Pages additionally contribute to the prediction of individual depressive symptoms, after adjustment of the model for individual and structural (census-based) neighborhood measures. This model also allows for consideration of possible mediating effects (19). All models consider effects for a one-unit change in depressive symptoms. All reported tests of statistical significance are two sided.

RESULTS
The sample is described in table 1, including individual, structural, and service density characteristics. The number of individuals in each census tract ranged from 5 to 403. As demonstrated in table 1, there was a range of variability in census-derived structural characteristics across the 28 census tracts. Less variability was evident in the service density measures. Of note is that most of the variation in undesirable services was due to the presence of liquor outlets. Because of the small number of tracts and somewhat skewed distribution, Spearman's correlations were used to determine the associations between service density and structural neighborhood characteristics (table 2). None of the correlations was significant. Table 3 reports the fixed individual-level relation between various demographic and sociodemographic indicators and depressive symptoms. Consistent with results from prior work in this sample (14), women, those with less education, z Service density is calculated as the number of services divided by the total population in a census tract. and those reporting more disability were more likely to report depressive symptoms, when all individual-level factors were included in the model simultaneously. We also found that, after adjustment for the individual-level characteristics, the between-neighborhood variation (the random effect for neighborhoods) was marginally statistically significant (p ¼ 0.07).

Associations between structural characteristics and depressive symptoms
Multilevel analyses also suggest contextual effects on individual levels of depressive symptoms. In separate models that did not adjust for individual characteristics, several of the neighborhood-level structural measures were significantly associated with depressive symptoms (table 4). Individuals residing in neighborhoods with a higher rate of poverty had more depressive symptoms than did those living in census tracts with lower levels of poverty (b ¼ 8.34; p < 0.01). While this effect was somewhat attenuated after taking account of the individual-level characteristics, there remained an independent effect of neighborhood poverty (b ¼ 6.51; p < 0.05). That is, the contextual effect of neighborhood poverty is not entirely explained by the fact that individuals with a higher risk of depression live in the same neighborhoods. Several other structural characteristics were significantly associated with depressive symptoms after taking account of individual-level characteristics (table 4). For instance, individuals living in an area with more people over 64 years of age had lower levels of depression relative to areas with fewer people over 64 years of age (b ¼ ÿ13.55; p < 0.05). Individuals living in more affluent areas had marginally lower levels of depression relative to less affluent areas (b ¼ ÿ34.23; p < 0.10). We did not find associations of either residential stability or racial/ethnic heterogeneity with depression levels.
Additional analyses examined depression as a dichotomous outcome, using a score of 16 or higher as a cutpoint for high likelihood of clinical depression. Findings with the census-based data and individual-level factors were similar to those reported above, although somewhat attenuated (data not shown). Because of high intercorrelations between some of the variables, as well as power limitations, we could not examine whether the three census-based characteristics significantly associated with depressive symptoms would have independent effects when included in a model simultaneously.

Associations between service density measures and depressive symptoms
Measures of levels of neighborhood service density did not add any information to models of depression, either alone (data not shown) or in conjunction with other neighborhood characteristics, such as the percentage of poverty or the percentage of people aged over 64 years (table 5). On the other hand, the associations of structural measures with depressive symptoms were largely unchanged. In other words, our characterization of census tract service environments based on the Yellow Pages failed to provide additional predictive power in understanding how neighborhood context may influence individual-level depressive symptoms.  * Nonstandardized parameter estimates are obtained from the fixed part of a multilevel linear regression model conditional upon census-tract random effects.
y 95% confidence intervals are obtained per one-unit change in the parameter estimate.

DISCUSSION
Findings from this study suggest that a number of neighborhood structural characteristics are associated with depressive symptoms among the elderly, even after adjustment for individual-level risk factors. Consistent with previous multilevel studies among younger or more select samples (4,5,9), the present study found that neighborhood poverty was strongly linked with poorer mental health. Living in a poor neighborhood seems to be associated with greater depressive symptoms in older adults, beyond whatever individual vulnerabilities there may be. A new finding to emerge is that the presence of more elderly people in the neighborhood is associated with better mental health among older adults; those living in neighborhoods with a greater concentration of other older adults seem to experience fewer depressive symptoms, regardless of an individual's risk for problems. We did not find the strong contextual effect of residential stability that other work has suggested (4), although the direction of effects was consistent with prior work (greater residential stability was associated with fewer depressive symptoms). A measure of service density could be considered a mediator if it accounts for variations in levels of both structural neighborhood characteristics and depressive symptoms and if, when it is included in the model, a service density measure also attenuates the strength of the direct effect of neighborhood structural characteristics on depressive symptoms (19). However, none of the service density measures was independently associated with depression, and we found no evidence that access to services that promote social engagement, provide health services, or affect the reputation of a neighborhood could explain (i.e., mediate) how neighborhood structural characteristics might affect mental health.
Other studies with EPESE data have considered individuallevel factors as they relate to depressive symptomatology in more detail (14). Consistent with this work, we found that less education, being female, and disability were strongly associated with higher levels of depressive symptoms. The present study suggests, however, that the social environment may contribute to symptoms of depression for reasons beyond simply that poorly educated or disabled individuals tend to live in the same neighborhoods. It may be that neighborhoods with a higher percentage of poor individuals have fewer material and social resources to provide support for residents in a variety of domains. In addition, such conditions may also make it difficult for individuals to sustain supportive social relationships with other individuals (1,4). Fewer resources may also make the community less effective in preventing the occurrence of negative experiences (e.g., crime, loss of income), resulting in more chronic stress for neighborhood residents, which in turn is related to mental distress (4,20).
The present study represents an attempt to understand area characteristics beyond those captured by census data to consider how neighborhoods might influence mental health. The presence of specific kinds of services may provide a proxy for a social environment that promotes or deters social integration. However, in this study, we found more services (both good and bad) present in neighborhoods with more poverty, more residential mobility, and fewer individuals aged over 64 years. Moreover, the presence of services in disadvantaged neighborhoods did not appear to provide any buffer for depressive symptoms, although we did not examine the effects of particular services separately. To our knowledge, only one other study has considered the neighborhood service environment, by use of data from the Alameda County Study (21). In that study, services were measured by identifying the number of commercial stores in a neighborhood via the telephone book Yellow Pages. Unlike the present study, only four types of commercial stores were included, and they were not categorized according to whether they might promote or impair social interactions. In the Alameda County Study, greater density of services was associated with poor social environments, and lower quality social environments were associated with increased risk of death during 11 years of follow-up (21). An independent effect of commercial stores on mortality was found, such that individuals living in neighborhoods with many commercial stores were at increased risk of death compared with people living in neighborhoods with few stores. A number of limitations should be considered in relation to the present study. Defining neighborhoods via census tracts may not always reflect meaningful neighborhood boundaries, particularly for area-based measures that characterize neighborhood service availability. Such measures may be particularly sensitive to whether people live near neighborhood boundaries, and it may be more appropriate to use other definitions of neighborhood for this work (22). However, if census tracts do not define meaningful neighborhood boundaries for residents, we would expect nondifferential misclassification, which would result in a bias toward the null. We also used a cross-sectional study design and could not evaluate the degree of individual exposure to various neighborhood conditions in conjunction with information on the onset or time course of the experience of depressive symptoms. Although we adjusted for an array of individual characteristics, associations could reflect selection, whereby there are unmeasured variables that explain how individuals select into neighborhoods. For example, although we did adjust for physical disability, because of data limitations and the focus on social factors, we did not adjust for such factors as the presence of other psychiatric disorders, a family history of depression, or substance use. Service accessibility was ascertained using listings from telephone book Yellow Pages. However, accessibility may be more influenced by the availability of transportation services than by proximity to one's residence. Unfortunately, we did not have data on transportation services.
Overall, our findings suggest that neighborhood structural characteristics are associated with individual levels of depressive symptoms. Living in disadvantaged neighborhoods was associated with more mental distress. Effects on mental health were related to not only the level of neighborhood disadvantage but also the neighborhood age composition. Additional work is needed to determine the mechanisms by which the concentration of older individuals in a neighborhood affects mental health. Given that access to various types of services could not explain the effects of structural characteristics, future research might more profitably focus on other area-based characteristics.