How do negative-control outcomes and exposures expose residual confounding in an observational study?

## The logic You cannot check a causal estimate against the truth, because the truth is what you are trying to learn. But you can sometimes construct a nearby analysis where the truth is known in advance - specifically, known to be zero. Run your entire pipeline on that, and whatever comes out is bias, because there is no signal for it to be measuring. This turns an unfalsifiable claim into a testable one. ## Negative-control outcomes A negative-control outcome satisfies two conditions. First, the exposure cannot cause it - no biological, behavioural, or mechanical pathway exists. Second, it is subject to the same confounding structure as the real outcome, so the unmeasured variables that would bias the main analysis would bias this one too. The canonical demonstration comes from studies of influenza vaccination in older adults. Observational data repeatedly showed vaccinated seniors having dramatically lower mortality, including in the off-season, and lower rates of outcomes such as hospitalisation for broken bones. A vaccine cannot prevent a fracture. The apparent protection was healthy-vaccinee bias: people who are mobile, engaged with the health system and not already frail are both more likely to get vaccinated and less likely to fall ill or die of anything. The negative control did not merely raise a doubt; it quantified the bias and showed the headline mortality benefit was largely an artifact of who gets vaccinated. The second condition is the one people skip. An outcome that the exposure cannot cause but that is *not* driven by the same confounders is uninformative - it will be null regardless of how badly confounded the main analysis is. Choosing a good control requires thinking explicitly about which unmeasured variable you fear and asking whether it would also move the control. ## Negative-control exposures The mirror construction holds the outcome fixed and swaps the exposure. You look for a treatment that cannot affect the outcome of interest but that people select into for the same reasons. If a study claims a drug reduces mortality, a negative-control exposure might be an unrelated medication prescribed to a similar population for a similar level of engagement with care. If that unrelated drug also appears to reduce mortality, the association is about who takes medication reliably, not about pharmacology. A related trick is the negative-control *period*: estimate the effect in a window where the exposure could not yet have acted - before it was available, or before enough time has passed for any mechanism to operate. A non-zero estimate there means the estimator, not the treatment, is producing the number. This is the placebo-test form of the same idea: apply the estimator where the answer must be zero and check that it returns zero. ## Interpreting the result The asymmetry of evidence is the crux, and it is what separates a good answer from a rote one. **A non-null negative control is strong evidence of a problem.** You have caught the pipeline producing an effect that cannot exist. The main estimate is then suspect, and the magnitude of the control effect gives a rough scale for the contamination. Some methods go further and use the control estimate to calibrate or partially correct the main one, though that requires assuming the bias acts similarly on both. **A null negative control is weak evidence of correctness.** It says only that this particular bias pathway left no detectable trace in this particular check. The control may be underpowered - fractures are rarer than deaths, so the interval may be too wide to exclude anything meaningful. Or the control may not actually share the confounder you fear. Report power, not just the point estimate: a null with an interval spanning 0.6 to 1.7 has ruled out nothing. **Direction matters.** If the negative control shows apparent protection and the main estimate also shows protection, the bias points the same way and the true effect is smaller than reported. If the control's bias runs against the main effect, the true effect may be larger. That directional reasoning is often more actionable than the magnitude. ## Designing them in advance The practice that distinguishes a serious analysis is pre-specifying negative controls before seeing the main result. Chosen afterwards, they are vulnerable to selection: an analyst who runs six candidate controls and reports the two that came back null has demonstrated nothing. Write down the candidate controls, the confounder each is meant to probe, and the threshold that would count as a failure, then run them. A final practical note: a failed negative control is not a reason to quietly drop the check. It is the most useful thing the analysis produced. The correct response is to report it, diagnose the mechanism, and either fix the design - better matching, a restricted comparator, an active-comparator design that removes the selection - or state that the question cannot be answered from this data.