When Conditioning Creates an Association
Adjustment is often used to make comparisons more informative, but adding a variable to an analysis does not automatically improve it. If the variable is caused by both the exposure and the outcome, comparing people at the same value of that variable can create an exposure–outcome association that was not present in the population. This is called collider bias.
The key is the direction of the arrows in a causal diagram. A common cause points toward both exposure and outcome; a common effect receives arrows from both. These roles are different, and conditioning on them can have different consequences.
In this diagram, the path between \(E\) and \(Y\) is closed when we do not condition on \(C\). Conditioning means comparing within levels of \(C\), such as analyzing only individuals with \(C=1\), or including \(C\) as a covariate to estimate a conditional association. That operation opens the path. The resulting association is not necessarily causal: it can reflect the way the analysis has selected or grouped observations rather than an effect of \(E\) on \(Y\).
A Numerical Example: Selection on a Common Effect
Consider a hypothetical population with a binary exposure \(E\) and a binary outcome \(Y\). Suppose \(E\) and \(Y\) are independent in the full population: each of the four combinations occurs equally often. Let \(C=1\) indicate eligibility for a special program. Both having the exposure and having the outcome make eligibility more likely, so the diagram is \(E \rightarrow C \leftarrow Y\). The program itself does not cause the exposure or the outcome.
Worked Example: An Association Appears Among Eligible Individuals
Imagine 4,000 hypothetical individuals, with 1,000 in each \(E,Y\) combination. Eligibility is assigned so that the following counts result. The counts are illustrative, not observations from a real study.
| Eligibility | \(E=1,Y=1\) | \(E=1,Y=0\) | \(E=0,Y=1\) | \(E=0,Y=0\) |
|---|---|---|---|---|
| \(C=1\) | 900 | 650 | 650 | 50 |
| \(C=0\) | 100 | 350 | 350 | 950 |
| Total | 1,000 | 1,000 | 1,000 | 1,000 |
Check the unconditioned association. In the full population, \(E=1\) occurs for 2,000 people, of whom 1,000 have \(Y=1\). Thus \(P(Y=1\mid E=1)=1,000/2,000=0.5\). For \(E=0\), the corresponding calculation is also \(1,000/2,000=0.5\). The risk difference is \(0.5-0.5=0\): there is no marginal association in these data.
Now condition on eligibility. Among eligible individuals with \(E=1\), there are \(900+650=1,550\) people, and 900 have \(Y=1\). Their outcome proportion is \(900/1,550\). Among eligible individuals with \(E=0\), there are \(650+50=700\) people, and 650 have \(Y=1\), giving proportion \(650/700\). The conditional difference, calculated from the exact fractions, is
Thus, among eligible individuals, the outcome proportion is about 0.3479 lower for \(E=1\) than for \(E=0\). This is a conditional association in the hypothetical data; it does not show that the exposure reduces the outcome.
Why does this fit the collider structure? In the full population, all four exposure–outcome combinations are equally common. But eligibility is more likely when either \(E=1\) or \(Y=1\). Among eligible people without the exposure, having the outcome is one way to be eligible, so \(Y=1\) is common. Among eligible people with the exposure, the exposure itself helps explain eligibility, so the outcome is less necessary. This creates a negative association within the eligible group.
The selection probabilities also confirm that both variables increase eligibility in this example: they are \(0.90\) when \(E=1,Y=1\), \(0.65\) when exactly one is 1, and \(0.05\) when both are 0. For either fixed value of the other variable, changing \(E\) from 0 to 1 raises the probability of \(C=1\); the same is true when changing \(Y\). The example therefore has the stated common-effect structure.
Why Conditioning Opens the Path
The numerical pattern can be expressed with conditional probabilities. Suppose \(E\) and \(Y\) are independent before conditioning, and \(C\) is caused by both. For a fixed exposure level \(e\), the chance of \(Y=1\) among people with \(C=1\) is
The numerator includes the probability of eligibility when \(Y=1\); the denominator averages eligibility probabilities over both outcome values. If eligibility depends on \(E\), the relative contribution of the outcome values among eligible people can differ across exposure levels. Independence before conditioning does not guarantee independence after conditioning.
For the example, the eligibility probabilities among the two outcome values give the same results as the counts. For \(E=1\), the selected outcome proportion is \(0.90/(0.90+0.65)=900/1,550\). For \(E=0\), it is \(0.65/(0.65+0.05)=650/700\). The conditioning operation changes the composition of the groups being compared.
This is the practical meaning of the diagram’s open path. A causal diagram is not a calculation of the size or direction of bias; it signals that conditioning can create a noncausal association. The direction and magnitude depend on the relationships and probabilities in the particular setting. In many settings where both causes increase the chance of the collider, a negative association is intuitive, but the direction is not guaranteed in every possible model.
Recognizing a Collider in Practice
Ask what causes the variable being considered for adjustment. If both the exposure and the outcome can cause it, the variable may be a collider on the path between them. The same concern arises if the analysis is limited to a group defined by that variable, even when the variable is not entered directly into a regression model.
Worked Example: Conditioning on Specialist-Clinic Attendance
Suppose, in a hypothetical setting, a treatment exposure and a disease outcome can each increase the chance that a person attends a specialist clinic. The relevant structure is \(E \rightarrow C \leftarrow Y\), where \(C\) denotes clinic attendance.
Identify the role. Attendance is caused by both exposure and outcome, so it is a common effect, not a common cause. It is a collider on the path between \(E\) and \(Y\).
Consider the proposed analysis. If researchers analyze only clinic attendees, they condition on \(C=1\). Exposure and outcome can then become associated among attendees even if they were unrelated in the source population. An association measured in this restricted group may therefore reflect collider bias.
Draw the appropriate conclusion. The diagram identifies a risk, not the numerical size or direction of the resulting association. Without information about the relevant probabilities, we cannot calculate that association or conclude that it must be positive or negative. The safe interpretation is that clinic attendance is not automatically an appropriate adjustment variable for estimating an exposure–outcome relationship.
Worked Example: Separating a Common Cause From a Common Effect
Suppose a hypothetical study considers whether a training course \(E\) is associated with a later test result \(Y\). Prior preparation \(U\) may affect both course participation and the test result, while receiving an award \(C\) may depend on both course participation and the test result.
Classify prior preparation. Its arrows point outward: \(E \leftarrow U \rightarrow Y\). It is a common cause of exposure and outcome. This is a backdoor path that may create confounding.
Classify the award. Its arrows point inward: \(E \rightarrow C \leftarrow Y\). It is a common effect. Conditioning on the award can open a path between course participation and the test result.
Compare the implications. These structures are not interchangeable. Adjusting for a genuine common cause can block a backdoor path under appropriate assumptions, whereas conditioning on a common effect can create an association along a path that was closed. The variable’s causal role, not merely its predictive value, matters when choosing adjustments.
Collider Bias in Regression Adjustment
A regression model that includes \(C\) estimates an association conditional on the included covariates, according to the model’s specification. If \(C\) is a collider, the conditional comparison can inherit the association opened by conditioning on \(C\). A large sample or a precisely estimated regression coefficient does not remove this structural problem: precision describes sampling uncertainty, not whether the comparison is biased for the intended question.
This does not mean that every variable affected by exposure must be treated as a collider. Its role depends on the causal structure. The relevant concern here is specifically that the variable is a common effect on a path connecting exposure and outcome, and that the analysis conditions on it. A variable caused by exposure alone does not, on that fact alone, have the \(E \rightarrow C \leftarrow Y\) structure.
Collider bias is also distinct from confounding by a common cause. In the common-cause structure \(E \leftarrow U \rightarrow Y\), the path is open before adjustment and can produce an unadjusted association. In the collider structure \(E \rightarrow C \leftarrow Y\), the path is closed before conditioning and can become open after conditioning. Thus the question is not simply “Does this variable relate to exposure and outcome?” It is “What arrows lead into or out of it, and what happens to the path when we condition on it?”
Common Mistake and Key Takeaway
Wrong reasoning: “This variable is associated with both exposure and outcome, so adjusting for it must remove bias.” That conclusion treats every associated variable as though it were a confounder.
Correction: Association alone does not determine a variable’s causal role. If both exposure and outcome cause the variable, conditioning on it can open a path and induce an association. Use subject-matter knowledge to consider the directions of causation, draw the diagram, and ask whether the proposed analysis conditions on a common effect.
Check Your Understanding
Use a diagram and the definitions in this tutorial to answer the following questions.
- In \(E \rightarrow C \leftarrow Y\), which variable is the collider, and what happens to the path between \(E\) and \(Y\) when conditioning on \(C\)?
- In the numerical example, calculate the outcome proportions for \(E=1\) and \(E=0\) in the full population. What is their difference?
- Using the exact conditional proportions in the example, what is the difference among individuals with \(C=1\), rounded to four decimal places?
- Why can two variables be independent in the full population but associated among people selected because of a common effect?
- Draw the structures for a common cause and a common effect. Explain how conditioning can have different consequences for the two structures.