What are the different ways to account for SES in an analytic model when investigating racial/ethnic health disparities? (Hint: you should have three options). Discuss the interpretations/implications of each approach as it relates to the interest in understand health disparities by race/ethnicity. Draw a DAG for each option and reference it in your response (you do not have to post this!).
Option 1:
Treat SES as a confounding variable. In this situation, SES would affect both race/ethnicity and the outcome of interest. An analytic method such as this would better characterize an observed association between the outcome and race/ethnicity that was in fact due to previously unmeasured factors. This in turn would provide a more detailed picture of the myriad of factors that impact the outcome; helping to identify true associations and those that are in fact false or incompletely characterized. In a DAG, directed edges would go from SES to race/ethnicity and the outcome of interest.
Option 2:
Treat SES as a mediating factor, which would at least partially explain the association between race/ethnicity and the outcome. In this analytic model, we would be able to estimate the impact of SES on the association between race/ethnicity and the outcome. In the setting of the Lorch paper, the goal was to assess how if specific factors mediated the association between race/ethnicity and fetal death rates. This allowed for the identification of factors which mediate the association and were noted to differ by race/ethnicity. In a DAG, there would be an edge from race/ethnicity to SES and another from SES to the outcome.
Option 3:
Use clustering to group factors associated to SES which may be related by geographic distribution to test the impact of a contextual environment on the outcome. In the Merlo paper, this was done to take a more ecological approach to individual level data, clustered by neighborhood.
Think about multilevel influences on a health outcome of interest to you. Discuss how you would study this, including measurement and analytic approaches you would use to account for exposures across multiple levels.
My health outcome of interest would be utilization of healthcare resources for treatment of bladder cancer. To study this, I would include basic demographic and clinical characteristics such as age, gender, race/ethnicity, clinical staging, histologic diagnosis and additional factors such as household income, insurance coverage (y/n and type of coverage), distance from hospital, level of education and an estimate of health resource weath (as defined in the AHRQ database, which identifies those areas which are ‘resource-poor’). The outcome would be defined as definitive treatment for bladder cancer. I would use a similar approach as that from the Merlo paper, looking at individual level rates of hematuria (often a starting point in the workup for bladder cancer) as it differs by county. We could then extrapolate state and national level differences in treatment of bladder cancer. As a general ecologic model, we could compare state-level data to see differences by state. To look at county and individual level data, we would have to factor in the variation of utilization by country, adjusted for the rates of hematuria diagnosis in that region.