Why an Adjustment Can Create an Association
Adjustment is often used to make treatment groups more comparable. But conditioning on an additional variable does not always improve a comparison: if that variable is a collider, conditioning on it can create an association that was absent before adjustment. This matters because an adjusted treatment–outcome association may then reflect the way the data were selected or stratified, rather than a causal effect.
Consider a treatment \(A\), an outcome \(Y\), and a variable \(U\) that causes the outcome. Suppose treatment and \(U\) both cause a third variable \(C\). The corresponding directed acyclic graph (DAG) is \(A \rightarrow C \leftarrow U \rightarrow Y\). The arrows into \(C\) make it a common effect of \(A\) and \(U\), and therefore a collider on the path from \(A\) to \(Y\).
The intuition is that once we know a person has a particular value of \(C\), knowing whether \(A\) contributed to that value can reveal information about whether \(U\) did. If treatment makes \(C\) more likely, for example, then among people with \(C=1\), those without treatment may tend to have more of the other cause of \(C\). If that other cause also affects \(Y\), treatment and outcome can become associated within the \(C\) group.
This is different from controlling for a common cause of treatment and outcome, which can block a noncausal path. It is also not the same question as adjusting for a mediator: a collider is identified by the directions of the arrows into it, not merely by whether it occurs after treatment. A variable can have more than one role in a larger DAG, so the relevant paths must be examined.
A Numerical Demonstration
The following hypothetical model isolates the collider mechanism. Let \(A\) and \(U\) be independent binary variables, each with probability \(0.5\) of being 1. Let the outcome be \(Y=U\), so treatment has no causal effect: under either treatment value, \(Y(a)=U\). Let \(C=1\) exactly when \(A\) and \(U\) differ. Thus \(A\) and \(U\) both cause \(C\), while \(U\) causes \(Y\).
Worked Example: Conditioning on a Common Effect
Before conditioning, the four possible pairs \((A,U)\) are equally likely. Since \(Y=U\), the joint distribution of treatment and outcome is:
| \(A\) | \(Y\) | \(P(A,Y)\) | \(C\) |
|---|---|---|---|
| 0 | 0 | 0.25 | 0 |
| 0 | 1 | 0.25 | 1 |
| 1 | 0 | 0.25 | 1 |
| 1 | 1 | 0.25 | 0 |
Marginally, \(P(Y=1\mid A=1)=0.5\) and \(P(Y=1\mid A=0)=0.5\). The risk difference is \(0.5-0.5=0\), consistent with the model’s zero causal effect.
Now condition on \(C=1\). The only possible pairs are \((A,U)=(0,1)\) and \((1,0)\), each with conditional probability \(0.5\). Because \(Y=U\), this gives:
The conditional risk difference is \(0-1=-1\). In the \(C=1\) group, treatment is perfectly associated with a one-unit lower outcome, even though it has no causal effect. Conditioning on \(C\) has connected \(A\) to \(U\), which predicts \(Y\).
Conditioning on \(C=0\) creates an association in the other direction: the possible pairs are \((0,0)\) and \((1,1)\), so the conditional risk difference is \(1-0=1\). These stratum-specific associations do not contradict the marginal difference of zero; they arise from the imposed conditioning.
Why Conditioning Opens the Path
The example can be expressed more generally. If \(A\) and \(U\) are initially independent, then their joint distribution factors as \(P(A,U)=P(A)P(U)\). After conditioning on \(C=c\), their joint distribution is proportional to the probability of that value of \(C\) under each pair of causes:
Even though the factors \(P(A)\) and \(P(U)\) separate, the factor \(P(C=c\mid A,U)\) generally depends on both variables jointly. The resulting conditional distribution need not factor into \(P(A\mid C=c)P(U\mid C=c)\). That is the induced dependence between the collider’s causes.
The first example uses a deterministic rule for \(C\). The same mechanism does not require a deterministic collider or perfect associations.
Worked Example: A Probabilistic Collider
Suppose again that \(A\) and \(U\) are independent Bernoulli variables with probability \(0.5\) of being 1, and \(Y=U\). In a hypothetical setting, let \(C\) be a selection indicator with \(P(C=1\mid A,U)=0.2\) when \(A=U\) and \(0.8\) when \(A\ne U\).
Each \((A,U)\) pair has prior probability \(0.25\). The joint probabilities for \(C=1\) are therefore \(0.25(0.2)=0.05\) for \((0,0)\) and \((1,1)\), and \(0.25(0.8)=0.20\) for \((0,1)\) and \((1,0)\). Summing gives \(P(C=1)=0.50\). Dividing each cell by \(0.50\), the conditional probabilities given \(C=1\) are \(0.10, 0.40, 0.40, 0.10\), respectively.
Among those with \(C=1\) and \(A=1\), the total probability is \(0.40+0.10=0.50\), and \(Y=1\) occurs in the \((1,1)\) cell with probability \(0.10\). Thus \(P(Y=1\mid A=1,C=1)=0.10/0.50=0.20\). For \(A=0\), the corresponding risk is \(0.40/0.50=0.80\). The conditional risk difference is \(0.20-0.80=-0.60\), while the marginal risk is \(0.50\) in each treatment group and the marginal risk difference is zero.
Here the induced association is not perfect, but the conclusion is the same: conditioning on the common effect creates an association between treatment and a cause of the outcome. The numbers also show that the strength and direction of the induced association depend on how \(C\) is generated.
Selection as a Collider
A frequent practical setting is selection into a study or into the subset with complete follow-up. Suppose treatment \(A\) and the outcome \(Y\) are independent in the population, but both affect whether a person is included, represented by \(S\). The DAG \(A \rightarrow S \leftarrow Y\) has \(S\) as a collider. An analysis restricted to \(S=1\) can therefore show a treatment–outcome association that is absent in the population.
Worked Example: Association Created by Study Inclusion
For a hypothetical illustration, let \(A\) and \(Y\) be independent binary variables, each with probability \(0.5\) of being 1. Let \(S=1\) be more likely when \(A\) and \(Y\) differ: \(P(S=1\mid A,Y)=0.8\) for unequal values and \(0.2\) for equal values.
Before selection, each pair has probability \(0.25\), so the marginal risk of \(Y=1\) is \(0.5\) for both \(A=1\) and \(A=0\). For \(S=1\), the unnormalized probabilities of \((A,Y)=(0,0),(0,1),(1,0),(1,1)\) are \(0.05,0.20,0.20,0.05\). Their sum is \(0.50\), so the conditional probabilities are \(0.10,0.40,0.40,0.10\).
Within the selected sample, \(P(Y=1\mid A=1,S=1)=0.10/(0.10+0.40)=0.20\), while \(P(Y=1\mid A=0,S=1)=0.40/(0.40+0.10)=0.80\). The selected-sample risk difference is \(-0.60\), despite a population risk difference of zero. The association is a consequence of conditioning on study inclusion, not evidence that treatment caused the outcome.
How to Decide Whether to Adjust
There is no reliable rule that says to adjust for every measured variable, every predictor of the outcome, or every variable that improves model fit. Instead, use the causal structure to ask what paths conditioning on a variable would block or open. The earlier material on choosing pretreatment covariates emphasizes that adjustment decisions concern causal roles, not merely predictive strength.
Identify treatment, outcome, and the population comparison of interest before selecting adjustment variables.
Use subject-matter knowledge to consider whether a candidate variable is caused by both treatment and another cause of the outcome, or by treatment and the outcome itself.
Check whether conditioning would open a path through a collider. A descendant of a collider can also open that path when conditioned on.
Do not include a variable solely because it predicts the outcome. Prefer an adjustment strategy justified by the causal structure and the assumptions needed for the target comparison.
In observational data, a variable’s observed associations do not by themselves establish whether it is a collider, a common cause, or a mediator. Those roles depend on the assumed causal ordering. In a randomized trial, randomization makes treatment independent of pretreatment causes on average, but conditioning on a post-randomization collider can still induce an association among the selected or stratified participants.
Diagnostics, Alternatives, and Limitations
Collider bias is not generally diagnosed by a standard balance table, a small p-value, or a model-fit statistic. Those summaries describe observed data; they do not reveal whether conditioning has opened a causal path. Useful checks are causal: identify why the variable is measured, when it occurs, what may cause it, and whether study inclusion or complete-case analysis depends on both treatment and outcome-related factors.
A comparison without conditioning on the collider avoids this particular induced path, but it is not automatically a valid causal analysis: other noncausal paths may remain. Conversely, adjustment for appropriate pretreatment common causes can address confounding when its identifying assumptions hold. Randomization provides a different design-based route to treatment comparability, although it does not protect an analysis that conditions on a post-treatment collider or selects participants in a way that opens a path.
The main limitation is that causal diagrams encode assumptions rather than reveal them mechanically. If the arrows are wrong or important causes are omitted, the adjustment decision may also be wrong. Further, conditioning on a collider does not always produce a large or detectable association; the effect depends on the causes’ relationships and the collider mechanism. A small observed association is not proof that collider bias is absent.
Common Mistake
Wrong reasoning: “This variable predicts the outcome, so adjusting for it must reduce bias.” Prediction alone does not establish that adjustment is beneficial. If treatment and an outcome cause both affect that variable, conditioning on it can connect treatment to the outcome cause and introduce bias.
Correction: Ask what causes the candidate variable and trace the paths in a causal diagram. Adjust for a variable because the causal structure supports that adjustment for the intended question, not simply because it is measured or predictive.
Check Your Understanding
Answer each question, then check the brief solutions below.
- In \(A \rightarrow C \leftarrow U \rightarrow Y\), which variable is the collider, and what path can conditioning on it open?
- If \(A\) and \(U\) are independent, \(Y=U\), and \(C=1\) exactly when \(A\ne U\), what is the marginal risk difference comparing \(A=1\) with \(A=0\)? What is the risk difference among those with \(C=1\)?
- Why can restricting an analysis to people with complete follow-up create an association between treatment and outcome?
- True or false: any variable that predicts the outcome should be included in an adjustment model. Explain.
- Does finding no strong treatment–outcome association within collider strata prove that collider bias is absent? Why or why not?
Solutions. (1) \(C\) is the collider. Conditioning on it can open the path \(A-C-U-Y\), associating \(A\) with \(U\) and thus with \(Y\). (2) The marginal risks are both \(0.5\), so the marginal risk difference is \(0\). Given \(C=1\), \(A=1\) implies \(U=0\) and \(Y=0\), while \(A=0\) implies \(U=1\) and \(Y=1\); the risk difference is \(0-1=-1\). (3) If treatment and outcome-related factors both affect follow-up, conditioning on follow-up is conditioning on their common effect and can make them dependent. (4) False. Prediction does not determine causal role; a predictive variable may be a collider, and conditioning on it may introduce bias. (5) No. Collider bias depends on the causal structure and mechanism and may be weak or obscured in observed data; a null-looking association does not establish that conditioning was harmless.