The Effect of Adverse Housing and Neighborhood Conditions on the Development of Diabetes Mellitus among Middle-aged African Americans

Baseline sample

The sampling design of the African-American Health Study has been described elsewhere (34). Briefly, the African-American Health Study is a population-based cohort study of 998 African Americans. All subjects lived in either a poor, inner-city area (St. Louis, Missouri) or less impoverished and more heterogeneous suburbs just northwest of the City of St. Louis. Sampling proportions were set to recruit approximately equal numbers of subjects from both areas (sampling strata), which resulted in higher probabilities of selection in the inner city because it had fewer eligible subjects. Sampling involved random selection of first-area street segments within block groups and then housing units within each selected street segment. Professional interviewers (two thirds of whom were African American) with extensive project-specific training contacted households in person. Within each noninstitutional housing unit, interviewers screened for eligibility criteria, which were self-reported Black or African-American race and birth date from January 1936 through December 1950. If the household contained two or more eligible persons, one of them was selected by using Kish tables (35).

As a result, we used weighted data in the analysis. The overall weight for each African-American Health Study subject was constructed by using three components representing 1) the probability of selection based on the proportion of area segments, housing units, and (when appropriate) number of eligible persons in the household; 2) sample nonresponse; and 3) a poststratification weight for population nonresponse or noncoverage based on the 2000 US Census. When these weights are applied, the African-American Health Study cohort represents the noninstitutionalized African-American population in the two areas as of the 2000 US Census.

Inclusion criteria also involved Mini-Mental State Examination scores of 16 or greater and willingness to sign informed consent. We used the inclusion score of 16 because previous studies showed that subjects were able to make vital decisions about their medical care and there was the potential for false-positive results when using the standard scores of 23/24 (36). All subjects received in-home, baseline evaluations that averaged 2.5 hours, which occurred between September 2000 and July 2001. The response rate was 76 percent (998/1,320).

Follow-up sample

In-home interviews were conducted 36 months after baseline assessments. Of the 998 persons who participated at baseline, 853 were successfully interviewed at follow-up. Since 51 persons had died between baseline and follow-up, the response rate for surviving subjects was 90.1 percent (853/947). In-home follow-up assessments took an average of 1.5 hours to complete. Attrition analysis through four waves and involving all major variables indicated that drop-out status was associated with diagnoses of cancer (adjusted odds ratio = 2.57) and heart disease (adjusted odds ratio = 0.48) only and better vision (adjusted odds ratio = 0.90 per point on a 13-point scale. Thus, no attrition bias during waves 1–4 is evident for any of the major variables involved in the current analysis.

Incident diabetes

Both the baseline and follow-up interviews asked respondents about the presence of diabetes: “Did a doctor ever tell you that you have diabetes, high blood sugar, ‘sugar,’ sugar in your urine, or ‘sweet blood’?” The possible responses were “yes,” “suspect or possible,” “no,” “don’t know,” or “refuse.” Test-retest reliability of this question in a subsample of African-American Health Study participants was very high (κ = 0.94) (37). We limited subjects in this study to those reporting “no” to this question at baseline. At follow-up, we defined incident diabetes as answering “yes” to the same question.

Adverse neighborhood and housing conditions

Neighborhood and housing conditions were based on interviewer observations at the block face of the location of study participants and on participants’ self-report of neighborhood desirability. Housing conditions also were observed by the interviewers at the home of the study participant.

The survey team “objectively” assessed the external appearance of the block face (neighborhood) where the respondent lived during the earlier process of household enumeration by using a previously published assessment tool (38). On a four-point scale (1 = excellent, 4 = poor), observers rated each of five characteristics: condition of houses, amount of noise (from traffic, industry, etc.), air quality, condition of the streets, and condition of the yards and sidewalks in front of homes where participants lived. Of all block faces, 84.8 percent were rated by two independent observers. The average of the scores of two raters was used in the analysis, when available. Kappa was very high, ranging from a low of 0.66 (condition of the streets) to a high of 0.84 (conditions of the yards and sidewalks) (39).

We also obtained a subjective measure of neighborhood conditions from respondents at baseline by using a four-item scale of the neighborhood as a place to live, general feelings about the neighborhood, attachment to the neighborhood, and neighborhood safety from crime (40). Participant responses were dichotomized for each condition. Questions were modified from the Behavioral Risk Factor Surveillance System and are similar to those from other studies (41).

Assessment of housing conditions was an observed five-item scale based on the interviewer’s ratings at the baseline interview regarding the cleanliness inside the building, physical condition of the interior, condition of furnishings, condition of the exterior of the building, and a global rating (all rated as excellent, good, fair, or poor). In the analysis, each condition was dichotomized as either fair or poor versus good or excellent. The housing condition of 80 subjects was reassessed 5–45 days after the first assessment by the same assessors (42). Kappa showed moderate agreement for all five housing conditions (the condition of furnishings (κ = 0.79), cleanliness inside the building (κ = 0.74), condition of the exterior of the building (κ = 0.74), the global rating of the building (κ = 0.71), and physical condition of the interior (κ = 0.68)).

Potential mediating pathways

Pathways that may account for any observed association of adverse neighborhood and housing conditions with development of diabetes examined in this study consisted of health behavior factors, psychosocial factors, health status, access to medical care, and sociodemographic characteristics. First, health behavior factors consisted of a seasonally adjusted activities dimensions summary index on the Yale Physical Activity Scale (continuous variable) (43), smoking status, and risk of alcohol abuse (score of at least 2 on the CAGE alcohol screening instrument) (44). All were obtained at baseline, except for risk of alcohol abuse, which was assessed 1 year after baseline.

Second, psychosocial factors consisted of social support and depressive symptoms. Social support was measured by using five items (someone to confide in, get together with, help with daily chores, turn to for suggestions, and love and make you feel wanted; range: 5–25) from the Medical Outcomes Study social support instrument (45). The resulting scale score was recoded to contrast being in the lowest quintile versus all others. Depressive symptoms were measured by using the 11-item version of the Center for Epidemiologic Studies Depression scale (α = 0.836) (46). A score of 9 or greater on the 11-item version is equivalent to 16 or greater on the 20-item scale (46) and was considered to indicate a high level of depressive symptoms.

Third, health status at baseline consisted of self-reported health, body mass index, hypertension, number of severe chronic conditions, presence of one or more lower-body functional limitations, number of medications used, and whether the subject had been hospitalized overnight in the year prior to baseline. Self-reported health was measured by the SF-36 Health Survey’s self-rated health status question (Medical Outcomes Trust, Inc., Waltham, Massachusetts). Whether the respondent ever experienced any of nine severe chronic conditions (47) was based on self-report of physician diagnosis. The chronic conditions included asthma, chronic airway obstruction, heart failure, heart attack, angina, stroke, chronic kidney disease, arthritis, and cancer other than a minor skin cancer. The presence of one or more lower-body functional limitations at baseline was noted by using the Nagi physical performance scale (48, 49). Subjects were asked whether they had taken any medications, prescribed by a physician or not, in the past 2 weeks (34).

Fourth, access to medical care consisted of having health care insurance at the time of or during the 12 months prior to interview and not being able to see a physician because of cost during 12 months prior to baseline interview. Fifth, demographic baseline factors included age, gender, income categories, perceived income adequacy, educational attainment, marital status, employment status, length of time at the present address, home ownership, and sampling stratum.

Statistical analysis

We assessed the correlation between block face and housing conditions with the phi coefficient. We used logistic regression to determine the unadjusted risk of block face, housing, perceived neighborhood conditions, and each of the covariates on self-reported diabetes at follow-up. Next, we assessed the potential mediating effect of each of the five hypothesized pathways of the association of adverse block face, housing, and perceived neighborhood conditions with diabetes incidence by adding all variables constituting a potential pathway to the logistic model containing only neighborhood and housing conditions. This procedure was conducted for each of the pathways separately. Reduction in odds ratios for the adverse neighborhood and housing conditions relative to unadjusted analysis was evidence for mediation. The fit of the models was analyzed by calculation of Hosmer-Lemeshow goodness-of-fit statistics and Nagelkerke R².

We used the propensity score method to obtain adjusted estimates of the effect of neighborhood and housing conditions on development of diabetes (50, 51). Propensity score methods produce estimates that are more accurate than multivariable logistic regression estimates when there were seven or fewer events per confounder, as was the case in the present study (52). We included only those variables associated with diabetes development when calculating propensity scores (53). The propensity score was defined as the conditional probability of a person living under a certain neighborhood/housing condition given the covariates included. We then grouped the subjects into five strata representing quintiles of the propensity score. This method is usually adequate to remove more than 90 percent of the bias due to each of the covariates in a fully specified model (50).

We conducted a series of analyses to challenge the robustness of the findings. First, since mediators are required to be associated with both neighborhood/housing conditions and incident diabetes (54), we examined the mediational effects of only those variables that were associated with the risk of diabetes. Doing so also reduces the potential for collinear effects between variables as part of a pathway. Second, we expected that persons who resided longer in their neighborhood/home would have more exposure or opportunity to be affected by environmental conditions than persons who resided there for a shorter period of time. We determined whether the associations of neighborhood and housing conditions with incidence of diabetes were similar when limiting the analysis to persons who reported living at the same address for at least 5 years before the baseline interview and when limiting the analysis to persons who resided at the same address during the 3-year study period.

Third, previous studies have shown that persons who owned their home were more likely to report better health status than those who rented (55). Therefore, we determined whether the associations of housing conditions with incidence of diabetes were similar when limiting the analysis to persons who owned their home.

Fourth, one of the issues in neighborhood-based observational research has been how to deal with selection bias (56). We used propensity scores in an attempt to address the fact that persons are not randomly assigned to housing or neighborhood conditions (57). The propensity score is defined as the conditional probability of a person living under a certain neighborhood/housing condition given all observed socio-demographic covariates. We then modeled the association of neighborhood/housing conditions for the entire sample, adjusting for propensity score group in an attempt to control for selection bias. Next, we determined whether the adjusted odds ratio for the neighborhood/housing condition changed when adding all variables as part of the five hypothesized pathways to a model containing only the propensity score group and the adverse neighborhood/housing condition. This procedure was again performed separately for each of the five potential mediating pathways.