Why a Negative Control Exposure Needs More Than a Null Effect
A negative control exposure is an exposure that should not causally affect the outcome of interest. If it is nevertheless associated with that outcome after appropriate adjustment, the association can warn that the treatment analysis retains bias. But the absence of a causal effect is only one part of the argument: the control must also be connected to the sources of bias that threaten the treatment estimate. A control unrelated to those sources can show no association even when the treatment effect is confounded.
Let \(A\) denote the treatment of interest, \(Y\) the outcome, \(Z\) a candidate negative control exposure, \(C\) measured pre-exposure covariates, and \(U\) one or more unmeasured factors. The causal question concerns the effect of \(A\) on \(Y\); \(Z\) is used as a diagnostic, not as a replacement treatment or an automatic adjustment variable.
The potential-outcome statement concerns interventions, whereas a negative-control test examines observed data. Connecting the two requires consistency (observed outcomes agree with the potential outcomes under the exposures actually received), no interference between units, and sufficient overlap to compare relevant exposure levels within covariate strata. It also requires a plausible reason that the control shares a source of bias with the treatment analysis. These requirements are substantive; a test statistic cannot establish them.
When an Association Can Signal Confounding
Suppose an unmeasured factor \(U\) affects both treatment assignment and the outcome. Then, even after conditioning on \(C\), treated and untreated units can differ in \(U\), so the observed treatment contrast need not equal a causal effect. A negative control exposure can be informative if it is also associated with \(U\), while having no causal effect on \(Y\). The intended signal is an association between \(Z\) and \(Y\) that remains after accounting for \(A\) and \(C\), under conditions where that association cannot be attributed to a causal effect of \(Z\).
A simple conditional mean model makes the logic explicit. Within a fixed treatment and covariate stratum, suppose \(Y=\tau A+g(C)+\beta U+\varepsilon\), where \(\beta\ne 0\), and assume \(E[\varepsilon\mid Z,A,U,C]=0\). The term for \(Z\) is absent: changing \(Z\) does not change \(Y\) in this model. The conditional mean-zero assumption is stated given \(Z,A,U,C\), not merely given \(A,C\); that distinction is needed for the covariance calculation below.
For the last equality, iterated expectations give \(E[\varepsilon\mid Z,A,C]=0\) from the stated assumption. Consequently \(E[Z\varepsilon\mid A,C]=0\) and \(E[\varepsilon\mid A,C]=0\), so \(\operatorname{Cov}(Z,\varepsilon\mid A,C)=0\). Thus, in a stratum where \(Z\) varies, a nonzero association can arise when \(Z\) and the outcome-relevant factor \(U\) remain associated after conditioning on \(A,C\). This is a justification in a particular model, not a universal formula for all negative-control analyses.
To make that association informative about confounding of the treatment effect, \(U\) must also be relevant to treatment assignment: it must contribute to differences in treatment receipt, and it must affect \(Y\). If \(Z\) is associated with a factor that affects \(Y\) but not treatment, the control may flag outcome prediction without demonstrating that the treatment contrast is confounded. Conversely, a confounder that does not predict \(Z\) can bias the treatment effect while leaving this negative control unassociated with \(Y\).
Worked Example: A Shared Unmeasured Cause Creates a Signal
Consider hypothetical data with binary \(A,Z,U\), no varying \(C\), and outcome measured in points. Suppose the outcome model is \(Y=10+2A+3U+\varepsilon\), with \(E[\varepsilon\mid Z,A,U]=0\). Thus, changing \(Z\) has no causal effect on \(Y\). Within each treatment group, suppose \(P(U=1\mid Z=1,A)=0.80\) and \(P(U=1\mid Z=0,A)=0.30\).
At either fixed value of \(A\), the outcome mean difference comparing \(Z=1\) with \(Z=0\) is \(3(0.80-0.30)=1.50\) points. For example, when \(A=0\), the means are \(10+3(0.80)=12.4\) points and \(10+3(0.30)=10.9\) points, respectively. The 1.50-point difference is not a causal effect of \(Z\) in this model; it arises because the groups differ in \(U\).
If \(U\) also predicts treatment receipt, this pattern is consistent with \(Z\) tracking a source of treatment-outcome confounding. It is a warning to investigate residual bias, not a numerical correction for the treatment effect. The example assumes overlap in both \(Z\) levels within each \(A\) group and the stated conditional error mean; without those conditions, the comparison would not have the same interpretation.
Pooled Associations and Model Assumptions
Conditional associations can differ across levels of \(A\) and \(C\). Therefore, a single adjusted regression coefficient should not automatically be described as the conditional covariance ratio in every stratum. For a model that fits separate slopes within strata, the within-stratum linear-projection slope of \(Y\) on \(Z\), given \(A=a,C=c\), is \(\operatorname{Cov}(Z,Y\mid A=a,C=c)/\operatorname{Var}(Z\mid A=a,C=c)\), provided the conditional variance is positive. Under the model above, this equals \(\beta\operatorname{Cov}(Z,U\mid A=a,C=c)/\operatorname{Var}(Z\mid A=a,C=c)\).
A pooled regression has a different interpretation. Let \(W\) contain the intercept and all included adjustment terms for \(A,C\), let \(\widetilde Z\) be the residual from the linear projection of \(Z\) on \(W\), and let \(\widetilde U,\widetilde Y,\widetilde\varepsilon\) be defined analogously. By the Frisch–Waugh–Lovell result, the pooled coefficient of \(Z\) is \(\operatorname{Cov}(\widetilde Z,\widetilde Y)/\operatorname{Var}(\widetilde Z)\), when the denominator is positive. If the conditional mean model is correctly represented by the included \(W\) terms, and \(\operatorname{Cov}(\widetilde Z,\widetilde\varepsilon)=0\), then
The pooled ratio aggregates residualized variation; it need not equal any one stratum-specific ratio. The error orthogonality condition is essential. For example, \(E[\varepsilon\mid A,C]=0\) alone does not imply \(\operatorname{Cov}(Z,\varepsilon\mid A,C)=0\): within a stratum, if \(Z\) is Bernoulli with probability \(1/2\) and \(\varepsilon=Z-1/2\), then \(E[\varepsilon\mid A,C]=0\) but \(\operatorname{Cov}(Z,\varepsilon\mid A,C)=1/4\). Model assumptions must support the calculation actually used.
Worked Example: A Pooled Slope Is Not Each Stratum’s Slope
Suppose two equally frequent hypothetical strata have equal within-stratum variances of residualized \(Z\). In the first stratum, the conditional linear-projection slope of \(U\) on \(Z\) is 1; in the second, it is 2. Let \(Y\) follow \(Y=\tau A+g(C)+3U+\varepsilon\), with the included adjustment terms representing \(\tau A+g(C)\) and with the residualized error uncorrelated with residualized \(Z\).
The implied stratum-specific slopes of \(Y\) on \(Z\) are \(3(1)=3\) and \(3(2)=6\) outcome units per unit of \(Z\). With equal stratum weights and equal residualized \(Z\) variances, the pooled adjusted slope is their weighted average, \((3+6)/2=4.5\), not 3 in both strata or 6 in both strata. This illustrates why the pooled regression coefficient must be interpreted through its residualized pooled variation rather than assigned a stratum-specific covariance formula.
Diagnostics, Failure Modes, and Alternatives
A nonnull association is interpretable as a confounding signal only if the control’s no-effect condition, the analysis specification, and the shared-bias rationale are credible. Potential alternative explanations include a genuine causal effect of \(Z\), an indirect effect through \(A\), selection into the analyzed sample, conditioning on a common consequence of \(Z\) and \(U\), measurement error, or model misspecification. In particular, conditioning on treatment can itself induce an association between \(Z\) and \(U\); an observed conditional association is not automatically evidence that the same association existed before treatment assignment.
A null result is also limited. It may reflect weak association between \(Z\) and the relevant confounder, lack of overlap, imprecise measurement, low precision, or cancellation among several sources of association. Even a precisely estimated null association rules out only the detectable association targeted by that analysis, not all unmeasured confounding. The introductory material on falsification tests motivates using negative controls as probes; the formal conditions explain why their results are evidence to weigh rather than definitive tests of exchangeability.
Compared with a negative control outcome, a negative control exposure probes the outcome relationship using an exposure that should have no causal effect. Both approaches rely on a credible null causal pathway and a relevant connection to bias, but they can fail for different reasons. Ordinary covariate adjustment instead attempts to control measured confounders directly; a negative control does not make an unmeasured confounder measured and does not, by itself, identify or remove bias. The best interpretation is diagnostic: a signal challenges the assumed adequacy of the design, while no signal provides limited reassurance.
A Practical Conditions Checklist
State which intervention on \(Z\) is assumed not to change \(Y\), including any possible pathway through \(A\).
Identify whether the target is a conditional mean contrast, conditional independence, or a regression coefficient, and define its adjustment set.
Explain why \(Z\) could track a factor that affects both treatment assignment and the outcome after measured adjustment.
Assess consistency, no interference, overlap, selection, measurement, and the assumptions required by the chosen model.
Report the observed association as a diagnostic signal or lack of detectable signal, not as a proof or a correction of the treatment effect.
Worked Example: A Null Result Does Not Clear the Treatment Analysis
Suppose a binary unmeasured factor \(U\) affects both \(A\) and \(Y\), but a candidate \(Z\) is independent of \(U\) within each measured-covariate stratum. Suppose also that \(Z\) has no causal effect on \(Y\), and the study has adequate overlap. Then the shared-confounding mechanism is absent for this particular control: in a model \(Y=\tau A+\beta U+\varepsilon\), \(\operatorname{Cov}(Z,U\mid A,C)=0\) may yield no conditional \(Z\)-\(Y\) association under the model assumptions.
The treatment effect can still be confounded because \(U\) affects both \(A\) and \(Y\). Thus, even a correctly estimated null negative-control association does not establish that \(A\) is unconfounded. The diagnostic was uninformative about this confounder because the candidate exposure did not track it. The conclusion depends on the stated model and association condition; a null estimate alone does not verify them.
Check Your Understanding
Answer each question, then compare with the solutions.
- In potential-outcome notation, what does it mean for \(Z\) to have no causal effect on \(Y\) at fixed \(A\)?
- In the conditional mean model, let \(\beta=2\) outcome units per unit of \(U\) and \(\operatorname{Cov}(Z,U\mid A,C)=0.25\). Under the stated error condition, what is \(\operatorname{Cov}(Z,Y\mid A,C)\)?
- Why is a shared association between \(Z\) and an outcome predictor not by itself enough to show treatment-effect confounding?
- Two strata have slopes of \(Y\) on \(Z\) equal to 2 and 8, with equal weights and equal residualized \(Z\) variances. What is the pooled slope under these conditions, and why should it not be reported as each stratum’s slope?
- Give one reason a null negative-control result can occur even when the treatment effect is confounded.
Solutions. (1) \(Y(a,z)=Y(a,z')\) for all relevant \(a,z,z'\); if the analysis conditions on \(A\), also consider whether \(Z\) can affect \(A\) and thereby affect \(Y\). (2) The covariance is \(2(0.25)=0.50\) outcome-units times \(Z\)-units, because the conditional error covariance is zero under the stated assumption. (3) The associated factor must also affect treatment assignment, as well as the outcome, for it to be a source of confounding of the treatment effect. (4) The pooled slope is \((2+8)/2=5\); it aggregates both strata and equals neither stratum-specific slope. (5) The control may fail to track the confounder, so it can be null even though that confounder affects both treatment and outcome.