Observational studies and confounding
In one sentence
An observational study records care or exposures without assigning them, so other differences between groups can distort the apparent effect of a treatment.
The intuition
Suppose hikers who carry expensive boots finish more difficult trails. The boots might help. But experienced hikers may both buy those boots and choose harder trails. Comparing boot owners with everyone else mixes the equipment's effect with the people wearing it.
Cancer treatment comparisons face a similar problem: people receive different care for reasons that also affect their outcomes. Recognizing those reasons helps you read a result without dismissing the whole study.
How it works
An observational study follows or reconstructs what happened without assigning the treatment or exposure being compared. A cohort follows a defined group over time. Other designs compare people with and without an outcome, or take a snapshot at one time.
A confounder is a factor that helps explain both who receives an intervention and how the outcome develops, outside the effect being estimated. Disease burden, fitness, access to care and prior treatment can make groups different before an intervention begins.
The diagram shows two possible paths to the outcome. Observing a treatment–outcome association does not tell us how much traveled along each path. Confounding by indication means the reason for choosing a treatment also affects the outcome—for example, giving a more intensive treatment to people with more severe disease.
Researchers can restrict eligibility, match participants or statistically adjust for measured differences. Each method depends on assumptions and adequate data. Recording a variable poorly, missing an important difference, or having groups with little overlap can leave the comparison misleading. Adjusting for every available variable is not a universal remedy; a consequence of treatment is not automatically a baseline confounder.
A worked fictional example
Two groups each contain 100 people. A fictional beneficial outcome occurs equally often within each baseline-health category:
| Baseline health | Treatment A | Treatment B |
|---|---|---|
| Better health | 60 of 80: 75% | 15 of 20: 75% |
| Poorer health | 5 of 20: 25% | 20 of 80: 25% |
| Everyone combined | 65 of 100: 65% | 35 of 100: 35% |
The combined comparison favors A, although this constructed example has no difference within either category. More people with better baseline health received A. Real studies rarely have just one perfectly measured source of difference, so the table illustrates a problem rather than providing a recipe for correcting every study.
Why it matters in cancer
Observational evidence can describe uncommon harms, care patterns and outcomes in people who were underrepresented in trials. A large sample does not remove systematic bias. Ask why the groups received different care, when follow-up began, and which differences the analysis could address.
Common confusions
- Association is a relationship in the data; causation is a claim about changing an outcome by changing an intervention.
- An adjusted estimate can still be confounded by unmeasured or poorly measured factors.
- Reverse causation is a related problem: developing disease can change an exposure, making its direction of influence easy to misread.
- Randomization helps form comparable groups; it does not guarantee flawless follow-up or universal applicability.
Related concepts
Sources and scope
Source check: October 9, 2026. The table is invented to explain selection and confounding, not a treatment result. Expert and learner review remain pending.
- National Cancer Institute: observational studies — purposes and designs.
- STROBE statement — reporting eligibility, potential confounders, methods and limitations.
- FDA: research on bias and confounding in real-world evidence — why bias assessment matters in treatment comparisons.