What a Negative-Control Outcome Can Check
A treatment model can produce weights that look plausible while still leaving important problems undetected. One supplementary check is to examine an outcome that the treatment cannot plausibly cause. If treatment groups remain associated with that outcome after weighting, the result can be a warning: the weights may not have removed some source of treatment–outcome association. But this check is informative only when the negative-control outcome is relevant to the problem being diagnosed.
The idea is distinct from checking balance on measured covariates. A negative-control outcome is an outcome variable chosen because, under the scientific setting and time ordering, it should not respond causally to the treatment regimes being compared. Its observed association with treatment after weighting can reveal residual structure that measured-covariate balance alone may not show. It cannot, by itself, establish that the treatment model is correctly specified.
A pre-treatment outcome can sometimes serve as a negative control for a later treatment, provided it is measured before all treatment being assessed. An outcome measured after treatment requires stronger substantive justification: it must not be affected by any treatment component in the contrast, directly or through downstream variables. For longitudinal treatment, an outcome unaffected by the final treatment may still be affected by earlier treatment. The negative-control claim must therefore match the treatment history whose weights are being checked.
For a useful check, the negative-control outcome should also be capable of exposing a relevant bias. In particular, its causes should overlap in a meaningful way with causes of treatment assignment or with the sources of residual association one is concerned about. This relevance condition is substantive; it does not follow from the definition of a negative-control outcome.
A Weighted Null Result Under Explicit Conditions
The simplest derivation uses one binary treatment and measured baseline covariates \(X\). Write \(e(X)=P(A=1\mid X)\). Suppose the treatment weights are the unstabilized inverse-probability weights \(1/e(X)\) for treated individuals and \(1/\{1-e(X)\}\) for untreated individuals. The weighted mean of \(Z\) in each group is a ratio of weighted expectations.
Proof. By consistency and treatment invariance, the observed \(Z\) is the same under either treatment. Conditional exchangeability gives \(E[Z\mid A=1,X]=E[Z\mid X]\) and \(E[Z\mid A=0,X]=E[Z\mid X]\). Therefore,
and \(E\{A/e(X)\mid X\}=1\). Taking expectations yields \(E\{AZ/e(X)\}=E[Z]\) and \(E\{A/e(X)\}=1\). Replacing \(A,e(X)\) with \(1-A,1-e(X)\) gives the same two identities for the untreated group. The ratios are therefore both \(E[Z]\). Integrability ensures these expectations exist. \(\square\)
This result explains why a non-null weighted contrast can be a warning: at the population level, it contradicts at least one condition needed for the derivation. Possible explanations include a misspecified treatment model, unmeasured confounding, a failure of positivity, an invalid negative-control claim, or a failure of conditional exchangeability for \(Z\). In a finite sample, random error and estimated weights can also produce a nonzero contrast even when the population contrast is zero.
For longitudinal treatment, the analogous diagnostic compares weighted summaries of \(Z\) across treatment histories or regimes using the relevant treatment-history weights. The same caution applies: a null contrast is expected only under conditions that make the weighted comparison meaningful, including correct treatment probabilities, positivity, and sufficient exchangeability for the negative-control outcome. The single-treatment proposition above gives the complete calculation without implying that every longitudinal weighted contrast has the same simple form.
Worked Example: Measured Confounding Removed by Weighting
Let \(X\) be binary with \(P(X=1)=P(X=0)=0.5\). Treatment probabilities are \(P(A=1\mid X=1)=0.8\) and \(P(A=1\mid X=0)=0.2\). Let the negative-control outcome be \(Z=X\), so treatment cannot cause \(Z\).
The joint probabilities for treated individuals are \(P(X=1,A=1)=0.5(0.8)=0.4\) and \(P(X=0,A=1)=0.5(0.2)=0.1\). Thus \(P(A=1)=0.5\), and \(E[Z\mid A=1]=0.4/0.5=0.8\). For untreated individuals, the joint probabilities are \(P(X=1,A=0)=0.5(0.2)=0.1\) and \(P(X=0,A=0)=0.5(0.8)=0.4\), so \(E[Z\mid A=0]=0.1/0.5=0.2\). The unweighted difference is \(0.8-0.2=0.6\), entirely due to measured confounding.
For treated individuals, the weights are \(1/0.8=1.25\) when \(X=1\) and \(1/0.2=5\) when \(X=0\). The weighted treated mass in each covariate group is \(0.4(1.25)=0.5\) and \(0.1(5)=0.5\). The total weighted mass is 1, and the weighted mean of \(Z=X\) is \(0.5/1=0.5\).
For untreated individuals, the weights are \(1/0.2=5\) when \(X=1\) and \(1/0.8=1.25\) when \(X=0\). The weighted masses are \(0.1(5)=0.5\) and \(0.4(1.25)=0.5\). The weighted mean is again \(0.5/1=0.5\). The weighted contrast is zero, as predicted by the proposition.
Why a Non-Null Association Is Not a Diagnosis by Itself
A non-null weighted association is a signal to investigate, not a label for the cause. Even if treatment cannot cause \(Z\), the treatment groups may differ in \(Z\) because of an unmeasured common cause. If the treatment model includes only measured variables that do not capture that common cause, weighting will not remove the association. A non-null finding can therefore indicate residual confounding without showing whether the treatment model itself is misspecified.
Worked Example: An Unmeasured Common Cause Leaves an Association
Suppose \(U\) is binary with \(P(U=1)=P(U=0)=0.5\). Let \(P(A=1\mid U=1)=0.8\) and \(P(A=1\mid U=0)=0.2\), but suppose the fitted treatment model omits \(U\) and uses \(e=0.5\). Set \(Z=U\), which is unaffected by treatment.
The treated joint probabilities are \(P(U=1,A=1)=0.4\) and \(P(U=0,A=1)=0.1\); hence \(E[Z\mid A=1]=0.4/(0.4+0.1)=0.8\). The untreated probabilities are \(0.1\) and \(0.4\), so \(E[Z\mid A=0]=0.1/(0.1+0.4)=0.2\). The fitted weights are both \(1/0.5=2\), so within each treatment group the constant weight cancels in the weighted mean. The weighted difference remains \(0.8-0.2=0.6\). This example has a genuine negative-control outcome but fails conditional exchangeability given the variables in the fitted model. The association warns that the weighted groups are not comparable on this outcome; it does not identify the omitted cause.
The converse problem is just as important. A null weighted association may occur because the control outcome does not share the relevant causes of treatment assignment or the outcome of interest. It may also result from imprecision. A control that is independent of treatment-related covariates can remain null even when the weights leave those covariates badly imbalanced.
Worked Example: A Null Control Despite Residual Imbalance
Let \(X\) be binary with probability \(0.5\), and let \(P(A=1\mid X=1)=0.8\), \(P(A=1\mid X=0)=0.2\). Suppose the analyst incorrectly uses constant weights, which do not change the observed treatment groups. Let \(Z\) be independent of both \(X\) and \(A\), with \(P(Z=1)=0.5\).
As in the first example, \(P(X=1\mid A=1)=0.8\) and \(P(X=1\mid A=0)=0.2\). Thus the treatment groups remain substantially imbalanced on \(X\). But independence gives \(E[Z\mid A=1]=E[Z\mid A=0]=0.5\), and constant weights preserve these means. The negative-control contrast is zero even though the weighting has not balanced \(X\). This control is uninformative about the particular imbalance because it does not reflect \(X\).
Interpreting and Comparing the Check
A useful report gives the weighted contrast, its uncertainty, the definition of the negative-control outcome, and the reason it is expected to be causally unaffected by treatment. It should also state the hypothesized connection between the control and the confounding structure of concern. Without that scientific rationale, a numerical null or non-null result is hard to interpret.
The diagnostic complements, rather than replaces, other checks. Post-weighting balance directly examines measured covariates used in or relevant to the treatment model. Simulation checks with known treatment mechanisms can assess whether a fitting procedure recovers known probabilities under specified data-generating conditions. A negative-control outcome instead examines an observed outcome that should have no treatment effect. Agreement among these checks can be reassuring, but none establishes all causal assumptions.
Common failure modes include selecting a control that is actually affected by earlier treatment, using a control with little connection to treatment-related confounding, interpreting a failure to reject a null contrast as proof of balance, and treating a statistically significant association as proof that one particular model term is missing. Diagnostics can also be sensitive to extreme weights and small effective sample sizes: a noisy contrast may reflect limited information rather than a clear structural failure. These concerns are reasons to investigate the result and compare it with other evidence, not reasons to discard the check.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- What condition makes \(Z\) a negative-control outcome for a binary treatment, and how does the condition change for longitudinal treatment?
- Under the proposition's assumptions, why do both weighted group means equal \(E[Z]\)?
- A weighted negative-control contrast is nonzero. Name two distinct explanations besides treatment causing the negative-control outcome.
- Why can a null negative-control contrast coexist with serious measured-covariate imbalance?
Solutions. (1) The condition is \(Z^1=Z^0\) almost surely. For longitudinal treatment, \(Z^{\bar a}=Z^{\bar a'}\) almost surely for every pair of regimes under comparison. (2) Conditional on \(X\), the true treatment probability cancels the treatment indicator in the weighted expectation; exchangeability makes the conditional mean of \(Z\) the same in each treatment group, so each weighted numerator and denominator has expectation \(E[Z]\) and 1, respectively. (3) The association may reflect unmeasured confounding, a misspecified treatment model, lack of positivity, or finite-sample variation; any two suffice. (4) The control may be unrelated to the imbalanced covariate, so its treatment-group means can agree even while measured covariate distributions differ.