DPH 810 Module 6 Bias, Confounding and Alternatives Example

Is It Real? Bias, Confounding and Alternative Explanations for Food Environment Effects on Diabetes

Student Name

Doctor of Public Health Program, Aspen University

DPH 810: Advanced Epidemiology in Public Health

Instructor Name

Month Day, Year

Is It Real? Bias, Confounding and Alternative Explanations for Food Environment Effects on Diabetes

Finding an association is the beginning, not the end, of epidemiological reasoning. Before concluding that food swamps raise diabetes risk, a doctoral researcher must consider whether chance, bias or confounding could explain the result, and plan methods to test or limit each. This paper examines alternative explanations for the planned cohort study and the methods chosen to address them.

Chance

With a large cohort, random error will be small, but multiple comparisons across exposure measures and subgroups could produce chance findings. The analysis plan specifies one primary exposure and outcome, reports confidence intervals rather than relying on significance thresholds and treats secondary analyses as exploratory, in line with guidance on interpreting statistical tests (Greenland et al., 2016).

Confounding

Individual income, education, age, race and ethnicity, insurance type and neighborhood deprivation may influence both where people live and their diabetes risk. A review of neighborhood research warned that failure to account for individual characteristics can make area effects appear larger than they are (Diez Roux, 2001). The analysis will adjust for these factors and use a causal diagram to avoid adjusting for mediators such as body mass index. Neighborhood deprivation will be included at the tract level to separate area poverty from food environment.

Residential Self-Selection

People may choose neighborhoods based on preferences related to diet and health, so observed associations could reflect who lives in food swamps rather than effects of the environment. Sensitivity analyses restricted to long-term residents, and analyses of people whose environments changed without moving, help address this. A Swedish cohort used such changes to strengthen inference about food outlets and diabetes (Mezuk et al., 2016).

Differential Outcome Detection

Diabetes is often undiagnosed. If residents of food swamps visit clinicians less often, fewer cases would be detected there, biasing results toward no effect; if they are screened more because of known risk, the bias would run the other way. The analysis will adjust for visit frequency and examine A1C testing rates by quartile.

Threats and Methods

The table summarizes threats, their likely direction and methods.

Exposure Misclassification

Commercial business databases contain errors: closed stores listed as open, new stores missing and misclassified outlet types. Store audits in a sample of tracts will estimate error rates. If misclassification is nondifferential, it would tend to bias results toward no effect, making any observed association conservative.

Reverse Causation

People developing diabetes symptoms might change where they shop or live, though this is unlikely to change neighborhood food environments. Excluding diagnoses in the first year of follow-up and lagging exposure by one year reduce concerns about reverse timing. Sensitivity analyses will vary the lag from one to three years.

Negative Controls

A negative control outcome, one not plausibly caused by food environments but sharing confounders, such as incident hearing loss diagnoses, can reveal residual confounding: if food swamp exposure predicts hearing loss, confounding or detection bias is likely at work.

Quantitative Bias Analysis

Instead of only naming limitations, quantitative bias analysis asks how powerful a hidden confounder would have to be to account for the whole observed association. Reporting such estimates helps readers judge robustness.

Triangulation

Evidence is stronger when different designs with different biases point the same way. The cohort, the survey and the natural experiment around new supermarkets have different weaknesses; consistent findings across them would strengthen causal inference, as would consistency with prior cohort evidence on neighborhood resources and diabetes.

Effect Modification Versus Confounding

Confounding distorts an association; effect modification is a real difference in the association across groups. Income may act as both: a confounder of the overall association and a modifier if food swamps affect lower-income residents more. The analysis will address both through adjustment and stratification.

Selection Into the Cohort

Only people receiving care in the two health systems are included. If food swamp residents without regular care are excluded, and they differ in diabetes risk, estimates could be biased. Comparing cohort demographics with census data will help assess how representative the cohort is.

Transparency in Reporting

All sensitivity analyses will be reported, whether they strengthen or weaken the main finding. Transparent reporting allows readers to judge robustness and avoids selective emphasis on favorable results.

What Would Change the Conclusion

If adjustment for individual income eliminates the association, if the negative control shows a similar association or if results among long-term residents differ sharply from the main analysis, confidence in a causal effect would drop. Stating these in advance guards against post hoc reasoning.

Planning Versus Post Hoc Analysis

Specifying these methods before seeing results protects against choosing analyses that produce favorable findings. A public registration of the analysis plan will come before any data access, which reviewers increasingly expect for observational studies.

Communicating Uncertainty

Findings will be presented with their uncertainty and with the results of sensitivity analyses, so that decision makers understand how confident they can be. If the association weakens substantially under plausible biases, that will be stated plainly.

Learning From Prior Studies

Previous cohort studies adjusted for individual income and education and still found associations between neighborhood food resources and diabetes, which suggests confounding by these factors alone does not explain the pattern, though residual confounding remains possible.

Summary of the Approach

The overall strategy is to anticipate each alternative explanation, choose methods that address it, report sensitivity analyses transparently and weigh evidence across designs rather than relying on a single estimate.

Conclusion

An association between food swamps and diabetes could reflect chance, confounding, self-selection, detection differences, misclassification, loss to follow-up or reverse causation. The study plans specific methods for each: prespecified analyses, adjustment guided by a causal diagram, sensitivity analyses, validation audits, negative controls, bias analysis and triangulation. Planning these in advance makes any conclusion more credible.

References

Diez Roux, A. V. (2001). Investigating neighborhood and area effects on health. American Journal of Public Health, 91(11), 1783-1789. https://doi.org/10.2105/AJPH.91.11.1783

Greenland, S., Senn, S. J., Rothman, K. J., Carlin, J. B., Poole, C., Goodman, S. N., & Altman, D. G. (2016). Statistical tests, P values, confidence intervals, and power: A guide to misinterpretations. European Journal of Epidemiology, 31(4), 337-350. https://doi.org/10.1007/s10654-016-0149-3

Mezuk, B., Li, X., Cederin, K., Rice, K., Sundquist, J., & Sundquist, K. (2016). Beyond access: Characteristics of the food environment and risk of diabetes. American Journal of Epidemiology, 183(12), 1129-1137. https://doi.org/10.1093/aje/kwv318