DPH 810 Module 4 Choosing a Study Design Example

Matching the Design to the Question: Choosing a Study Design for Food Environments and Diabetes

Student Name

Doctor of Public Health Program, Aspen University

DPH 810: Advanced Epidemiology in Public Health

Instructor Name

Month Day, Year

Matching the Design to the Question: Choosing a Study Design for Food Environments and Diabetes

No single study design is best for every question. The right design balances causal strength against feasibility, cost and ethics. This paper compares candidate designs for testing whether food swamp exposure raises the incidence of type 2 diabetes and justifies the choice of a retrospective cohort built from linked health records.

What the Question Requires

The hypothesis concerns incidence, so the design must establish that exposure preceded disease and follow people free of diabetes over time. Exposure is environmental and cannot ethically be assigned by investigators, so randomized trials are not feasible. Confounding by individual characteristics must be addressed.

Candidate Designs

The table compares five options.

Why a Retrospective Cohort

Linked health records from two large systems allow identification of adults without diabetes at baseline, geocoding of their home addresses to measure food environments and follow-up for new diagnoses over five years. The design establishes temporal order, measures incidence directly and can adjust for individual confounders recorded in the records. Cohort studies of neighborhood resources and diabetes have used similar approaches (Auchincloss et al., 2009).

Emulating a Target Trial

Framing the observational study as an emulation of a hypothetical randomized trial clarifies eligibility criteria, the start of follow-up, exposure definition and the comparison. Reanalyses of observational data using this approach have reduced biases arising from mismatched time zero and from comparing prevalent with new exposures (Hernán et al., 2008). Here, time zero is the first visit in the study window at which a patient is diabetes-free with a geocoded address.

Handling Residential Moves

People move, changing their exposure. The analysis will treat exposure as time-varying, updating the food swamp ratio when addresses change, and will conduct a sensitivity analysis restricted to people who did not move. Moves themselves may relate to health, which will be noted. About one resident in ten moves each year, so this matters.

Sample Size Reasoning

Sample size depends on the expected incidence, the smallest effect worth detecting, the significance level and the desired power. Standard approaches to sample size calculation set these inputs explicitly (Whitley & Ball, 2002). With about 80,000 eligible adults and expected incidence near 9 per 1,000 person-years, the study would have high power to detect a hazard ratio of 1.2 between the extreme quartiles. The calculation will be repeated once exact counts are available from the data systems.

A Secondary Natural Experiment

During the study period, two supermarkets opened in neighborhoods classified as food swamps. Comparing diabetes incidence before and after the openings in nearby residents with trends in similar neighborhoods without openings offers a natural experiment that strengthens causal inference for one specific change.

Why Not Other Designs

An ecological design would repeat existing county analyses and cannot address individual confounding. A cross-sectional survey cannot establish timing. A case-control study would struggle to reconstruct past food environments reliably. These designs remain useful for hypothesis generation but are weaker for the question posed.

Ethical and Practical Considerations

Using existing records avoids burdening participants but requires data use agreements, de-identification and review board approval. Geocoded data must be protected carefully. The design is feasible within a doctoral timeline because data already exist.

Limitations of the Chosen Design

The cohort includes only people receiving care in the two systems, who may differ from uninsured residents. Diabetes may go undiagnosed in people with few visits, causing outcome misclassification that may differ by neighborhood. Unmeasured confounding by factors such as diet preference or wealth may remain. Later modules address these.

Data Linkage

Linking records requires matching patient addresses to coordinates, then to outlet data and tract characteristics. Geocoding match rates, typically above 90% for urban addresses but lower for rural routes, will be reported, and unmatched records compared with matched ones to check for bias.

Exposure Windows

Food environments change. The study will measure exposure annually and consider cumulative exposure over the preceding years, since diabetes develops gradually. Sensitivity analyses will compare baseline-only and time-varying definitions.

Comparison With the Survey

The community survey planned in the next module complements the cohort by measuring behaviors and perceptions that records lack. Together, the two designs allow both stronger causal inference and richer interpretation.

Design Trade-Offs

A prospective cohort with measured diet would provide richer data but would take years and considerable funding. The retrospective design trades some measurement detail for speed and size, a trade-off appropriate for a doctoral project and for informing timely policy decisions.

Analytic Approach Implied by the Design

The cohort design implies survival analysis, with Cox models estimating hazard ratios and time-varying exposure updated at address changes. Clustering by census tract will be addressed with robust standard errors or multilevel models, since residents of the same tract share exposure.

Review and Approvals

The design requires approvals from both health systems' research offices and the university review board, as well as data use agreements specifying de-identification and geocoding procedures. Starting these early protects the timeline.

Choosing Comparison Groups

Comparing the highest with the lowest quartile maximizes contrast but ignores the middle. The analysis will also treat the food swamp ratio as a continuous measure, estimating the change in hazard per standard deviation, which uses all the data and tests for a dose-response pattern.

Pilot Work

Before the full analysis, a pilot using one year of data from one health system will test data linkage, geocoding and exposure calculations, revealing problems early when they are easier to fix.

Consistency With the Question

Each design choice traces back to the question: incidence requires a cohort, environmental exposure rules out randomization and confounding by individual characteristics requires rich covariate data.

Conclusion

A retrospective cohort built from linked health records, framed as a target trial emulation with time-varying exposure and supplemented by a natural experiment, offers the best balance of causal strength and feasibility for the food swamp question. Its limitations are known and can be addressed through careful analysis and sensitivity checks.

References

Auchincloss, A. H., Diez Roux, A. V., Mujahid, M. S., Shen, M., Bertoni, A. G., & Carnethon, M. R. (2009). Neighborhood resources for physical activity and healthy foods and incidence of type 2 diabetes mellitus: The Multi-Ethnic Study of Atherosclerosis. Archives of Internal Medicine, 169(18), 1698-1704. https://doi.org/10.1001/archinternmed.2009.302

Hernán, M. A., Alonso, A., Logan, R., Grodstein, F., Michels, K. B., Willett, W. C., Manson, J. E., & Robins, J. M. (2008). Observational studies analyzed like randomized experiments: An application to postmenopausal hormone therapy and coronary heart disease. Epidemiology, 19(6), 766-779. https://doi.org/10.1097/EDE.0b013e3181875e61

Whitley, E., & Ball, J. (2002). Statistics review 4: Sample size calculations. Critical Care, 6(4), 335-341. https://doi.org/10.1186/cc1521