Questions: Causal Inference in Epidemiology

Question 1

A researcher studying exercise's effect on cardiovascular disease adjusts for resting heart rate in their model. Resting heart rate is actually on the causal pathway from exercise to CVD — lower resting heart rate is one mechanism through which exercise reduces CVD risk. What is the consequence of this adjustment?

• A. The analysis becomes more accurate by removing a confounder that distorted the estimate
• B. The analysis introduces selection bias by restricting to a specific subgroup
• C. The estimated effect of exercise is attenuated, underestimating the total causal effect of exercise on CVD
• D. No consequence — adjusting for any variable associated with both exposure and outcome improves causal estimates

Answer: (C)

Resting heart rate is a mediator — a variable on the causal pathway from exercise to cardiovascular outcomes. Adjusting for a mediator blocks part of the causal effect you are trying to estimate. The remaining estimate captures only the portion of exercise's effect that does not operate through resting heart rate, underestimating the total causal effect. DAGs make this pitfall explicit: conditioning on a node on the directed path from exposure to outcome removes that path from the analysis. Adjusting for all variables associated with both exposure and outcome is not a safe rule — the DAG-based approach requires distinguishing confounders (should be adjusted) from mediators (should not, for total effect estimation) and colliders (must not be adjusted).

Question 2

In a DAG, a researcher conditions on a collider — a variable with arrows pointing INTO it from both the exposure and the outcome. What happens to the exposure-outcome association?

• A. A spurious association between the exposure and outcome is introduced where none previously existed
• B. The analysis is improved because conditioning on a collider blocks a non-causal backdoor path
• C. The association is unchanged — colliders have no effect on estimates when conditioned on
• D. The true causal association is revealed more clearly by removing the collider's distorting effect

Answer: (A)

Conditioning on a collider opens a path between its causes — in this case, the exposure and outcome. If the exposure and outcome both independently cause the collider, then knowing the collider's value makes the exposure and outcome statistically dependent even if they have no causal relationship. This 'collider stratification bias' is the opposite of what most researchers intuit: unlike confounders (which should be adjusted) and mediators (which usually should not), colliders must NOT be conditioned on to obtain unbiased causal estimates. DAGs are the primary tool for identifying colliders and recognizing when conditioning introduces rather than removes bias.

Question 3

Adjusting for most measured variables that are associated with both the exposure and outcome will eliminate confounding and yield an unbiased causal estimate.

• T. True
• F. False

Answer: False

This is a common and dangerous misconception. Adjusting for all associated variables will inadvertently include mediators (attenuating causal effects) and colliders (introducing spurious associations where none existed). Furthermore, even perfect adjustment for measured confounders cannot address unmeasured confounding, and residual confounding from measurement error can leave substantial bias. The DAG-based approach requires reasoning about the causal structure first, then selecting a valid adjustment set based on the backdoor criterion — not mechanically adjusting for everything associated with exposure and outcome.

Question 4

Temporality — the requirement that a cause precede its effect — is the only one of Hill's criteria that is logically necessary for a causal interpretation.

• T. True
• F. False

Answer: True

The other Hill criteria (strength, consistency, specificity, biological gradient, plausibility, coherence, experimental evidence, analogy) increase or decrease confidence in a causal interpretation but none is logically necessary — a causal relationship could be weak, inconsistent across studies, or biologically unexplained and still be causal. Temporality is the exception: by definition, a cause cannot follow its effect. If the outcome is shown to precede the putative exposure, the hypothesized causal direction is ruled out. This is why longitudinal study design — which establishes exposure measurement before outcome — is methodologically fundamental in epidemiology.

Question 5

In the counterfactual framework, why can an observational study never directly answer a causal question in the same way a randomized controlled trial can?

Model answer: The counterfactual definition of causation asks: what would have happened to this person had they not been exposed? For any individual, we can only observe one state — exposed or unexposed — never both simultaneously. In an RCT, random assignment makes exposed and unexposed groups identical in expectation on all other variables, so the unexposed group serves as a valid counterfactual for the exposed group at the population level. In an observational study, people who are exposed differ from those who are not in ways often related to the outcome (confounding, selection), so the unexposed group is not a valid counterfactual. All observational causal inference methods — regression adjustment, instrumental variables, matching — are attempts to construct a valid counterfactual comparison, but they depend on untestable assumptions that RCT randomization satisfies by design.

This is why epidemiologic causal inference is described as making the observational comparison 'as if randomized' through analytical means — it is an approximation to the counterfactual, not the counterfactual itself. The strength of a causal claim from observational data depends on the plausibility of the assumptions underlying the approximation, which is why DAGs, sensitivity analyses, and explicit assumption-naming are essential rather than optional.