Why a Candidate Control Outcome Must Be Untouched by Treatment
A negative control outcome is useful only if the treatment does not cause it. This condition is separate from whether the outcome shares relevant causes with the primary outcome: a candidate may be associated with the same background factors and still be unusable if treatment can change it. The danger is not limited to an obvious, immediate effect. Treatment may change the control outcome through an intermediate process, at a later time, or through effects on other people or units.
Let \(A_i\) be a binary treatment indicator for unit \(i\), and let \(W_i\) be a proposed negative control outcome. Write \(W_i(a)\) for the outcome that unit \(i\) would have under treatment status \(a\). In a setting with no interference, the central requirement is that changing treatment does not change the control outcome for the relevant units and time period.
The condition is stated at the individual level. A zero average treatment effect on \(W\), \(E[W_i(1)-W_i(0)]=0\), is weaker: positive effects for some units could cancel negative effects for others. Such cancellation does not make the outcome causally untouched. For falsification, the relevant question is whether the treatment can affect the candidate outcome through a plausible pathway, not merely whether one estimated average contrast is near zero.
The potential-outcome notation presumes that treatment conditions are sufficiently well defined and that the observed outcome under the received condition agrees with its corresponding potential outcome (consistency). The no-interference form also assumes one unit’s outcome does not depend on other units’ treatment. Where that assumption is implausible, the potential outcome must include relevant exposure to others’ treatment as well. These definitions do not by themselves identify a causal effect from observational data; they specify the causal condition the proposed control must satisfy.
Four Ways Treatment Can Reach a Control Outcome
A causal graph is a useful first audit: if a directed path runs from treatment to the proposed outcome, the outcome is not a negative control for that treatment in the relevant setting. The path can take several forms. A direct path runs from \(A\) to \(W\). An indirect path runs through a mediator \(M\), as in \(A\rightarrow M\rightarrow W\). A delayed path represents an effect of treatment at one time on a control outcome measured later. A spillover path runs from treatment assigned to someone else or another unit to the candidate outcome.
“Direct” and “indirect” describe causal pathways, not automatically identifiable numerical components. To estimate a direct or mediated effect separately requires additional assumptions. For deciding whether a candidate is suitable, it is enough to ask whether a credible pathway from treatment to \(W\) exists at all.
Worked Example: A Direct Biological Effect
Suppose a hypothetical randomized trial evaluates a drug intended to reduce joint pain. A researcher proposes blood concentration of a liver enzyme, measured in units per liter, as a negative control outcome because it is not part of the pain outcome. Assume the same measurement procedure and follow-up time are used in both randomized groups. The hypothetical mean enzyme concentrations are 12 U/L under the drug and 15 U/L under placebo.
The estimated treatment contrast is
The negative contrast means the treated group’s mean is 3 U/L lower in this hypothetical comparison. Because assignment was randomized, consistency holds, and the outcome is measured comparably, the group contrast can estimate the average causal effect under the stated trial conditions. More importantly for the control’s suitability, a plausible direct pharmacological effect on the enzyme means the drug can cause \(W\) to change. The enzyme outcome is therefore unsuitable as a negative control, even though it is not the trial’s primary endpoint.
Worked Example: An Effect Transmitted Through a Mediator
Consider a hypothetical program intended to reduce blood pressure. A researcher proposes liver-disease diagnosis during follow-up as a negative control outcome. The program does not directly alter the liver, but it increases follow-up visits, which can increase detection of liver disease. Let \(M=1\) indicate receiving the additional follow-up. Suppose the program changes the probability of follow-up from 0.30 to 0.80. For this illustration, assume follow-up itself causally changes the probability of diagnosis from 0.05 to 0.20, and assume there is no other effect of the program on diagnosis.
Under these simplifying causal assumptions, the diagnosis risk if assigned the program is
The risk under no program is
Thus the hypothetical total causal contrast is \(0.170-0.095=0.075\), or 7.5 percentage points. The pathway is \(A\rightarrow M\rightarrow W\): treatment changes follow-up, and follow-up changes diagnosis. The outcome is causally affected even though the effect is indirect and the treatment has no direct biological effect on the liver. The numerical calculation relies on the stated mediator model; it illustrates the pathway, not a generally valid recipe for estimating mediation from observed data.
Worked Example: A Delayed Treatment Effect
In a hypothetical study, a short course of treatment ends at week 2. A researcher proposes a symptom measured at week 8 as a negative control outcome, reasoning that the treatment has ended before measurement. Suppose the mean symptom score at week 8 is 4 points among those assigned treatment and 6 points among those assigned no treatment. Assume random assignment, consistent outcome measurement, and no loss to follow-up that would compromise the comparison.
The treatment contrast at week 8 is \(4-6=-2\) points. If the treatment can produce a persistent biological change that lowers the symptom score by week 8, then the treatment affects the later outcome. The fact that measurement occurs after treatment has ended does not remove the causal path: \(A_{\text{week 0--2}}\rightarrow W_{\text{week 8}}\). The candidate is invalid for this treatment and time horizon. A later measurement is not automatically a negative control; the relevant test is whether earlier treatment can affect it.
Worked Example: Spillover to an Untreated Unit
Suppose a hypothetical outreach program is offered at some clinics. Investigators use waiting time at a nearby untreated clinic as a negative control outcome for patients who do not receive the program. The program redirects patients from the nearby clinic, changing its workload. Let \(G_i=1\) mean that unit \(i\) is exposed to a neighboring clinic’s program and \(G_i=0\) mean no such exposure. Suppose the nearby clinic’s mean wait is 18 minutes when \(G_i=1\) and 12 minutes when \(G_i=0\), under otherwise comparable hypothetical conditions.
The spillover contrast is
Here the first argument indicates the unit’s own treatment status, held at 0, and the second indicates neighboring exposure. The 6-minute difference illustrates that another unit’s treatment can affect the candidate outcome. A control unit’s own untreated status is not sufficient to guarantee that its outcome is unaffected. If this spillover pathway is plausible, the proposed outcome cannot serve as an unaffected control without a design and analysis that address the relevant exposure structure.
A Practical Causal-Path Audit
A disciplined audit begins by specifying the treatment, outcome, population, and measurement window. “Treatment does not affect the outcome” is incomplete unless these are clear: a treatment may affect one version of an outcome, a particular subgroup, or a later measurement even if it has no effect on another. Then trace possible mechanisms from treatment to the candidate outcome, including pathways through changes in behavior, care, surveillance, or other intermediate events. Finally, ask whether treatment of other units could alter the outcome being measured.
Define what counts as treatment, which outcome is proposed as the control, who is observed, and when it is measured. A vague treatment definition can conceal effects that differ across treatment versions.
Ask whether treatment can change the outcome itself or change an intermediate variable that then changes the outcome. An indirect route still violates the causal-null requirement.
Consider effects that persist until the measurement time and effects of treatment assigned to connected people, clinics, or areas. State the relevant treatment history or neighboring exposure when needed.
If a credible treatment-to-outcome pathway remains, do not interpret the candidate as an outcome that treatment cannot affect. Seek a better candidate or use a different falsification strategy.
Observed data can reveal warning signs, but a nonsignificant association between treatment and \(W\) does not establish causal immunity. The estimate may be imprecise, effects may cancel in the average, or a causal effect may be offset by other differences. Conversely, an observed association alone does not prove a treatment effect, because noncausal differences between groups can also produce it. The causal-path audit relies on design knowledge, temporal ordering, and substantive mechanisms; an association test is evidence to interpret, not a substitute for that reasoning.
What to Do When a Candidate Is Affected
A negative control outcome that treatment can change no longer plays the intended role of an outcome that is causally untouched. Its treatment association can reflect the treatment pathway, so it cannot be read simply as a falsification signal about other bias. The first response is to reconsider the candidate’s causal status, not to explain away a nonzero result as evidence of confounding.
The main alternative is to choose a different candidate outcome for which credible direct, mediated, delayed, and spillover pathways can be ruled out for the stated intervention and follow-up period. A negative control exposure offers a different diagnostic route, but it requires its own causal restrictions; it is not a repair for an outcome known to be treatment-affected. A pretreatment outcome may avoid effects of a later treatment, but only if treatment truly occurs after its measurement and there is no relevant anticipation or prior exposure. These approaches answer related but distinct diagnostic questions.
There are limits to what a causal-path audit can establish. A biologically plausible pathway may be uncertain in size, and absence of a known pathway is not proof that no effect exists. The relevant treatment versions, outcome windows, and spillover boundaries may also be difficult to define. State the scope of the claim precisely: a candidate may be unaffected by one treatment version over one follow-up window, without being unaffected under every version or at every time.
Common Mistake: “It Is Unrelated to the Main Outcome”
Wrong reasoning: “This outcome is not part of the treatment target, so treatment cannot affect it and it is a valid negative control.”
Correction: A treatment can alter an outcome outside its intended target, directly or through a mediator, delayed response, or spillover. The required condition is causal invariance under the relevant treatment conditions, not merely that the outcome is different from the primary endpoint or was not intended to change. Likewise, failing to detect an association does not prove that invariance.
Check Your Understanding
For each item, decide whether the proposed outcome is defensible as a negative control under the stated facts, and compare your reasoning with the solutions.
- A treatment lowers pain, and a proposed control outcome is a biomarker that the treatment is known to alter directly. Is the outcome suitable?
- A program changes the frequency of screening, and the proposed control outcome is a condition detected more often under screening. Identify the relevant pathway and explain the consequence.
- A study measures a proposed control outcome six months after treatment ends. What must be checked before treating it as unaffected?
- A unit receives no treatment, but nearby units receive a program that may change the unit’s workload. What causal condition is missing from a no-interference analysis?
Solutions. (1) No. A direct treatment-to-biomarker path means the treatment can change the candidate outcome. (2) The pathway is treatment \(\rightarrow\) screening \(\rightarrow\) detection. The outcome is indirectly affected and is not a valid unaffected control. (3) Check whether treatment can have a persistent or delayed effect at the six-month measurement time; elapsed time alone does not establish causal invariance. (4) The unit’s potential outcome may depend on neighboring treatment as well as its own treatment. The spillover exposure must be considered, and a plausible spillover effect makes the outcome unsuitable as an unaffected control.