HLTH 8035 Week 7 Healthy-User Confounding Analysis Example

Confounding Analysis: Healthy-User Effects Behind a Multivitamin Cohort's Lower Mortality, Three Ways the Authors Could Have Controlled Them, and a Negative Control Outcome as a Test of What Remains

Student Name

Walden University

HLTH 8035: Epidemiology: Decoding the Science of Public Health

Instructor Name

Month Day, Year

Confounding Analysis: Healthy-User Effects Behind a Multivitamin Cohort's Lower Mortality, Three Ways the Authors Could Have Controlled Them, and a Negative Control Outcome as a Test of What Remains

The Cohort

The composite cohort enrolled 24,000 adults aged 50 to 74 who reported at enrollment whether they took a multivitamin daily, then followed them for death over ten years through a national death index. Daily users had lower all-cause mortality than non-users, with an age- and sex-adjusted hazard ratio of 0.82. The authors suggested that multivitamins may reduce mortality. Randomized evidence points the other way: large trials of multivitamins for prevention have generally found little or no effect on mortality, which makes the observational result a candidate for confounding rather than a discovery.

The Causal Diagram, in Words

A causal diagram for this question places multivitamin use as the exposure and death as the outcome, with an arrow between them that the study hopes to estimate. Several variables have arrows pointing into both, which is what makes each a confounder: a common cause of exposure and outcome (Porta, 2014). Smoking lowers the chance of taking supplements, since people who smoke are less often engaged in health-promoting habits, and raises the risk of death. Physical activity raises the chance of supplement use and lowers mortality. Income raises both the ability to buy supplements and access to the conditions of long life. Routine preventive care, such as screening and vaccination, travels with supplement use and independently reduces death. These common causes open backdoor paths from multivitamin use to survival that do not pass through the supplement itself. Together they make up what is often called the healthy-user effect: people who adopt one preventive behavior tend to adopt others, and observational comparisons credit the one under study with the benefits of all of them (Shrank et al., 2011).

One variable should stay out of the adjustment set. Diagnosed chronic disease after enrollment could be caused by supplement use or its absence and could lead to death; adjusting for it would block part of any real effect. The diagram therefore guides not only which confounders to control but which variables to leave alone.

Strategy 1: Restriction

Restricting the analysis to never-smokers removes confounding by smoking completely, because smoking no longer varies. It is simple and transparent, and it requires no assumptions about how smoking relates to death. The cost is generalizability: the estimate then applies only to never-smokers, who are healthier than the population, and the sample shrinks by perhaps a third. Restriction also handles only the variable restricted on; physical activity, income and preventive care still differ between users and non-users within the never-smokers.

Strategy 2: Matching

Matching each user to a non-user of the same age, sex and income band balances those variables by design. It improves efficiency when the matching variables are strong confounders, and it makes the comparison easy to explain. But matching can only balance what was measured, and the healthy-user effect lives largely in what was not: motivation, health literacy and habits that no questionnaire captured. Matching on income also leaves physical activity and preventive care unbalanced unless they too are matched on, which quickly makes suitable matches scarce.

Strategy 3: Stratification

Stratifying by a composite health-behavior index, combining smoking, activity, alcohol, diet and preventive care into low, middle and high bands, lets the analysis ask whether the association weakens within strata. If the hazard ratio of 0.82 moves toward 1 within each band, the crude association was carried largely by health behavior. In the composite data, the within-stratum hazard ratios ranged from 0.90 to 0.95, which suggests that much, though perhaps not all, of the apparent benefit reflected the behaviors that travel with supplement use. Stratification shows the pattern openly, but strata become thin with many variables, and a composite index can hide which component matters.

A Negative Control Outcome

None of these strategies reaches unmeasured confounding, so the analysis adds a test. A negative control outcome is an outcome that the exposure could not plausibly affect but that shares the confounding structure of the outcome of interest; if the exposure appears to "protect" against it, confounding is the likely explanation (Lipsitch et al., 2010). Death from accidental injury serves here. Multivitamins have no credible mechanism for preventing car crashes, falls or drowning, but the same cautious, health-conscious habits that lead people to take supplements also lead them to wear seat belts and avoid risks. In the composite cohort, daily users had a hazard ratio of 0.85 for accidental death after the same adjustments. An apparent protection against injury deaths nearly as large as the protection against all deaths is strong evidence that residual healthy-user confounding remains.

What the Analysis Concludes

The cohort's association between multivitamins and lower mortality is largely explained by the habits and circumstances that accompany supplement use. Restriction and matching address measured confounders at a cost; stratification shows the association shrinking within behavior bands; and the negative control outcome indicates that unmeasured confounding persists. The study cannot support the claim that multivitamins reduce mortality, and its result is consistent with the trial evidence that they do not.

References

Lipsitch, M., Tchetgen Tchetgen, E., & Cohen, T. (2010). Negative controls: A tool for detecting confounding and bias in observational studies. Epidemiology, 21(3), 383-388. https://doi.org/10.1097/EDE.0b013e3181d61eeb

Porta, M. (Ed.). (2014). A dictionary of epidemiology (6th ed.). Oxford University Press.

Shrank, W. H., Patrick, A. R., & Brookhart, M. A. (2011). Healthy user and related biases in observational studies of preventive interventions: A primer for physicians. Journal of General Internal Medicine, 26(5), 546-550. https://doi.org/10.1007/s11606-010-1609-1