Why a Control Can Create the Problem It Is Meant to Reveal
Negative controls are used to probe assumptions behind causal analyses, but a candidate control is not automatically safe to adjust for or condition on. In particular, a variable can be a collider: a common effect of two variables. Conditioning on a collider, or on a descendant of one, can make its causes statistically associated even when they were independent before conditioning. If those causes are connected to treatment and outcome, the induced association can resemble confounding or a treatment effect.
This issue is distinct from the usual question of whether a negative control is associated with the outcome. A negative-control exposure may be associated with an outcome because of shared causes, as intended in a diagnostic analysis. The concern here is that conditioning on that exposure, its causes, or a selection variable can open a new path and alter other associations. A resulting nonnull diagnostic is not, by itself, proof of residual confounding.
How Conditioning Opens a Path
Suppose \(A\) and \(B\) are initially independent causes of \(C\), with \(A \to C \leftarrow B\). Before conditioning, the collider blocks the path between \(A\) and \(B\). Once \(C\) is fixed, learning about one cause can provide information about the other. For example, if \(C\) occurs whenever either cause is present, observing \(C=1\) and \(A=0\) implies \(B=1\), while observing \(C=1\) and \(A=1\) does not.
This induced association can travel onward to treatment and outcome. If \(A\) causes treatment \(T\), and \(B\) causes outcome \(Y\), conditioning on \(C\) opens the path \(T \leftarrow A \to C \leftarrow B \to Y\). The path was blocked at \(C\) before conditioning. After conditioning, it can transmit an association between \(T\) and \(Y\), even if treatment has no causal effect on the outcome and \(A\) and \(B\) were independent at baseline.
If \(A\) and \(B\) are independent before selection, then \(P(B \mid A)=P(B)\). But after conditioning on \(C\), the factor \(P(C \mid A,B)\) can depend jointly on \(A\) and \(B\). Thus the conditional distribution of \(B\) can vary with \(A\), creating dependence. This is the core mechanism; no causal effect of \(A\) on \(B\) is required.
A variable being a cause of a proposed negative control does not, by itself, make conditioning on that cause unsafe. The relevant question is the cause’s place in the full graph: is it a collider, or a descendant of a collider on a path of interest? Conditioning on a noncollider can block a path; conditioning on a collider or its descendant can open one. The arrows and the conditioning set—not the label “control”—determine the result.
Worked Example: Selection Creates an Apparent Treatment Effect
Consider a hypothetical population in which randomized treatment \(T\) and a binary risk factor \(U\) are independent, each with probability \(0.5\) of being 1. Let selection \(S=1\) occur if either \(T=1\) or \(U=1\). The outcome is \(Y=U\), so treatment has no causal effect: \(Y\) is determined entirely by \(U\). Analyze the association after restricting to \(S=1\).
The selection rule is \(S=1\) when \(T=1\) or \(U=1\), so \(T \to S \leftarrow U\) is a collider structure. In the selected group, a treated person can have either value of \(U\), because treatment alone guarantees selection. Since \(T\) and \(U\) were independent, \(P(U=1 \mid T=1,S=1)=0.5\). An untreated person is selected only when \(U=1\), so \(P(U=1 \mid T=0,S=1)=1\).
Within the selected sample, treated people have an outcome mean 0.5 units lower than untreated people. In the full randomized population, both treatment groups have \(E(Y)=0.5\), so the treatment–outcome difference is 0. The selected-sample difference is entirely induced by conditioning on \(S\), not by a treatment effect. The example assumes the stated selection rule and outcome model exactly; it also shows why restricting an analysis to people who meet a treatment- or risk-dependent inclusion criterion can undermine the protection randomization provides.
Collider Controls and Their Causes
A proposed negative control can itself be a collider. Consider independent binary variables \(U\) and \(V\), each with probability \(0.5\) of being 1. Let \(U\) determine treatment \(T\), let \(V\) determine outcome \(Y\), and let a candidate negative-control exposure \(C\) equal 1 whenever either \(U=1\) or \(V=1\). There is no arrow from \(C\) to \(Y\), so \(C\) has no causal effect on the outcome in this model. However, \(C\) is a common effect: \(U \to C \leftarrow V\). Conditioning on it can associate \(T\) with \(Y\).
The next example calculates that association. It also clarifies what is and is not being claimed: a candidate control’s association with an outcome may have several explanations, but conditioning on the candidate as a covariate or restricting the sample by its value can create a separate collider problem.
Worked Example: Adjusting for a Collider Negative-Control Exposure
Use the hypothetical model above: \(U\) and \(V\) are independent Bernoulli variables with probability \(0.5\); \(T=U\); \(Y=V\); and \(C=1\) if \(U=1\) or \(V=1\). Compare the outcome means for the two treatment groups among people with \(C=1\).
Among those with \(C=1\) and \(T=1\), \(U=1\). The value of \(V\) remains equally likely to be 0 or 1, so \(E(Y \mid T=1,C=1)=0.5\). Among those with \(C=1\) and \(T=0\), \(U=0\). The selection rule then requires \(V=1\), so \(E(Y \mid T=0,C=1)=1\).
The induced difference is \(-0.5\), although \(T\) and \(Y\) are independent in the full population and \(C\) has no causal effect on \(Y\). The open path is \(T \leftarrow U \to C \leftarrow V \to Y\). Conditioning on \(C=1\) creates an association between \(U\) and \(V\), which is then reflected in the treatment–outcome comparison. This calculation does not show that every negative-control exposure is a collider; it shows why its causal parents and the proposed adjustment strategy must be examined before conditioning on it.
The same principle applies to a cause of a candidate control. Suppose a proposed control \(K\) is caused by \(Z\), and \(Z\) is itself a collider \(A \to Z \leftarrow B\). If \(A\) predicts treatment and \(B\) predicts outcome, conditioning on \(Z\)—a cause of \(K\)—can open the path between them. Conditioning on a descendant \(K\) of \(Z\) can also open that path. What matters is that \(Z\) or its descendant is a collider on the relevant path, not simply that \(Z\) causes a control.
Worked Example: Conditioning on a Cause of a Control
For a numerical illustration, let independent binary variables \(A\) and \(B\) each have probability \(0.5\) of being 1. Let \(Z=1\) if \(A=1\) or \(B=1\), let a proposed control \(K\) be a downstream measure of \(Z\), and let \(T=A\) and \(Y=B\). Compare treatment groups after restricting to \(Z=1\), the cause of \(K\).
When \(Z=1\) and \(T=1\), \(A=1\), and \(B\) is still equally likely to be 0 or 1. Therefore \(E(Y \mid T=1,Z=1)=0.5\). When \(Z=1\) and \(T=0\), \(A=0\), so the rule defining \(Z\) forces \(B=1\). Therefore \(E(Y \mid T=0,Z=1)=1\).
The path \(T \leftarrow A \to Z \leftarrow B \to Y\) is opened by conditioning on \(Z\). Since \(Z\) causes \(K\), examining only a stratum defined by \(Z\) can produce a misleading association even if \(K\) itself is not directly included in the outcome model. The example uses a deterministic downstream relationship only to make the structure clear; in real settings, a cause and its measured consequence need not be identical.
How to Diagnose and Avoid Collider Bias
Begin with the causal role of each variable, not with whether it is called a confounder, negative control, or selection criterion. Draw plausible arrows into the candidate control and its causes, then trace paths linking treatment to outcome. Ask whether the planned analysis conditions on a collider, a descendant of a collider, or a selection variable that can be a collider. Check whether the parents of that collider connect to treatment and outcome. If so, conditioning may create an association not present in the population.
Specify whether the control is being tested for association, included as an adjustment variable, or used to define an analysis subset. These are different operations.
Identify plausible causes of treatment, outcome, the control, and sample inclusion. Mark common effects and their descendants.
Check whether the proposed control or selection variable opens a path. Compare the intended analysis with an analysis that does not condition on that variable, while recognizing that the latter may have other biases.
A changed estimate or control association may reflect collider bias, confounding, selection, or model differences. It does not isolate one explanation without additional assumptions.
A useful diagnostic is to compare the observed relationship before and after conditioning, when both comparisons are meaningful, and to examine whether the proposed conditioning creates sparse or empty strata. Such a comparison can reveal sensitivity to the adjustment choice; it cannot establish the causal graph. In particular, a dramatic association in a selected subset is a warning, not proof that the selection mechanism is exactly the one hypothesized.
| Analysis choice | Potential benefit | Collider-related risk |
|---|---|---|
| Use the negative control as a diagnostic without adjusting for it | Preserves the intended association check | Its association can still have multiple causal explanations |
| Adjust for the candidate control | May address a needed source of confounding if the variable is appropriate | Can open a path if the control is a collider or descendant of one |
| Restrict the sample by control status or selection | May define a relevant target population | Can induce dependence among causes of selection and change the estimand |
| Adjust for measured pre-treatment covariates chosen from a causal model | Can block backdoor paths under suitable assumptions | Still requires avoiding colliders and their descendants |
The main alternative is not “adjust for everything else.” It is to choose an adjustment set using a defensible causal model and the target estimand, while keeping a negative-control test distinct from covariate adjustment. If selection is unavoidable, describe the selected population as the target and examine how selection could affect the relevant paths. Neither a nonnull control result nor a null result resolves these structural questions on its own.
Common Pitfall and Limits
Wrong reasoning: “The control has no causal effect on the outcome, so controlling for it cannot bias the treatment estimate.” Correction: absence of a causal arrow from control to outcome does not establish that the control is safe to condition on. It may be a common effect of causes related to treatment and outcome; conditioning on it can open a noncausal path. Likewise, a cause of a control is not automatically safe or unsafe: its role in the complete graph determines the consequence.
Collider explanations depend on assumptions about the graph and the selection process, which are rarely known with certainty. A graph can omit causes, arrows can be misspecified, and observed associations can have more than one explanation. Conditioning can also change the population being described, even apart from collider bias. Therefore, report which variables were conditioned on and why, and treat sensitivity to plausible alternative structures as a limitation. A negative-control analysis is a diagnostic, not a substitute for the assumptions required for causal identification.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- If \(A\) and \(B\) are independent causes of \(C\), why can conditioning on \(C\) associate them?
- A randomized treatment \(T\) and risk factor \(U\) are independent. Selection occurs only when \(T=1\) or \(U=1\), and \(Y=U\). What is the outcome-mean difference between treated and untreated people among those selected?
- Does being a cause of a proposed negative control automatically mean that a variable is unsafe to condition on? Explain.
- A control has no causal effect on the outcome, but is a common effect of one cause of treatment and another cause of outcome. What should be checked before adjusting for it?
Solutions. (1) Conditioning on the common effect makes information about one cause informative about the other, opening the path \(A \to C \leftarrow B\). (2) Among selected treated people, \(P(U=1)=0.5\); among selected untreated people, \(U=1\) necessarily. Thus the difference is \(0.5-1=-0.5\) outcome units. (3) No. The cause’s own graph position matters: conditioning is especially concerning if it is a collider or descendant of one on a path of interest. (4) Check whether conditioning opens the path from the treatment cause through the collider to the outcome cause, and whether the proposed adjustment is needed for the estimand. No causal effect from the control to the outcome does not make adjustment automatically safe.