How robust is the association between neighborhood socioeconomic status and coronary heart disease among women?
Keywords: Omitted variables; cox proportional hazard models; propensity scores; neighborhood socioeconomic status; coronary heart disease; sensitivity analyses
Objective: To assess how robust the relationship between neighborhood socioeconomic status (NSES) and coronary heart disease (CHD) among women is to observed and unobserved covariates. Study Design: Using 2-level hierarchical Cox proportional hazard regression models, we first analyzed the Women’s Health Initiative Clinical Trials data, merged with tract-level Census data on neighborhoods, using traditional multivariable regression. Participants age 50–79 were recruited (1993–1998) at 40 clinical centers and 36 satellite or remote sites, and followed through March 2005. Outcomes were age at first CHD event, age at CHD death or first myocardial infarction (MI), and age at CHD death. The NSES index, our primary exposure variable of interest, was formed from six census tract-level socioeconomic measures. After implementing traditional multivariable regression models, we performed two types of sensitivity analyses. First, propensity score methods developed for continuous treatments were utilized to balanced different levels of NSES on observed individual-level covariates. Then, an omitted variable sensitivity analysis (i.e., simulation study) was performed to determine the maximum magnitude of correlations the omitted variable would have to have with NSES and CHD outcomes to eliminate the effect of NSES on our outcomes. Principal Findings: In our multivariable models, women in lower NSES neighborhoods experienced significantly higher risk of adverse CHD outcomes after controlling for individual-level characteristics. Propensity analyses confirmed that the effect sizes seen in the multivariable regression models were not sensitive to adjustment for observed characteristics. However, the results were sensitive to inclusion of a hypothetically omitted binary variable. For example, the omitted binary variable would have to be correlated with CHD event at -0.12 or greater and with NSES at 0.11 or greater to eliminate the statistically significant association between NSES and CHD event. While absolute value correlations of at least 0.11 are found between a number of our observed covariates and NSES, for an omitted variable to explain our results, it would require a stronger correlation with CHD event than all the other observed variables in our data, except for diabetes. Conclusions: Living in a lower NSES neighborhood was independently associated with greater CHD risk, suggesting policies that improve NSES may also yield health returns.