Why a Post-Treatment Covariate Can Create Bias
Adjusting for more variables does not always make a causal comparison fairer. A covariate measured after treatment may be affected both by treatment and by a separate cause of the outcome. Conditioning on that covariate can make treatment and that other cause associated, even if they were independent before treatment. The resulting association can distort an estimate of treatment’s effect.
The key structure is a collider: a variable with two arrows pointing into it. In the path \(A \rightarrow L \leftarrow U \rightarrow Y\), \(L\) is a common effect of treatment \(A\) and a factor \(U\) that also affects outcome \(Y\). Before conditioning on \(L\), this path is blocked at the collider. Conditioning on \(L\) opens the path, connecting \(A\) to \(Y\) through \(U\). This is a noncausal path: it does not describe an effect of intervening on \(A\) that travels along the path to \(Y\).
Here \(A\) denotes treatment, \(L\) a post-treatment covariate, \(U\) another cause of that covariate and the outcome, and \(Y\) the outcome. To isolate collider bias, initially consider a graph in which \(A\) and \(U\) are independent before treatment, and the outcome is affected by \(A\) and \(U\). The total causal effect compares outcomes under treatment and no treatment, averaged over the population’s distribution of \(U\). A comparison made only among people with the same observed \(L\) need not estimate that effect: conditioning can change the distribution of \(U\) across treatment groups.
The opening of the path is a property of the graph, not a guarantee that every numerical estimate will be biased by a particular amount. In some special data-generating processes the induced association may be absent or may cancel another bias. But when the opened path carries association to the outcome, adjustment can introduce bias where none existed through that path before conditioning.
Worked Example: Selection on a Treatment-Affected Collider
Worked Example: Selection on a Treatment-Affected Collider
Suppose a hypothetical randomized study assigns \(A=0\) or \(A=1\). A binary baseline factor \(U\) is equally common in both treatment groups and raises the outcome score. A follow-up measurement \(L\) is more likely when either treatment is given or \(U=1\). For simplicity, let the outcome score be \(Y=2A+5U\), with no direct effect of \(L\) on \(Y\). Thus, under an intervention setting \(A\), changing \(A\) from 0 to 1 raises every person’s score by 2 points, regardless of \(U\).
In a hypothetical data-generating model, treatment and \(U\) are independent, and each of the four \((A,U)\) combinations occurs 1,000 times. The following counts show how many in each combination have \(L=1\).
| \(A\) | \(U\) | People in cell | People with \(L=1\) |
|---|---|---|---|
| 0 | 0 | 1,000 | 50 |
| 1 | 0 | 1,000 | 800 |
| 0 | 1 | 1,000 | 800 |
| 1 | 1 | 1,000 | 800 |
Among people with \(L=1\), the proportion with \(U=1\) is different across treatment groups. For \(A=1\), the two \(U\) cells contribute 800 people each, so
For \(A=0\), there are 800 people with \(U=1\) and 50 with \(U=0\), giving
The treatment groups were independent of \(U\) before conditioning; within \(L=1\), they are not. Because \(Y=2A+5U\), the mean outcome score for \(A=1,L=1\) is
For \(A=0,L=1\), the mean is
The within-\(L=1\) treatment contrast is therefore \(4.500-4.706=-0.2059\) points, despite the true causal effect being \(+2\) points. The difference arises because the \(A=0\) group selected into \(L=1\) contains a much larger proportion with \(U=1\), which raises the outcome score. Without conditioning on \(L\), each treatment group has \(U=1\) in half its members: their mean scores are 4.5 and 2.5, a difference of 2 points.
Why the Opened Path Creates an Association
The selection mechanism in the example gives an intuitive account of the path \(A \rightarrow L \leftarrow U \rightarrow Y\). Among people with \(L=1\), observing that one cause of \(L\) is absent makes the other cause more informative about why \(L=1\) occurred. Here, if a person in the selected group did not receive treatment, they are more likely to have \(U=1\), which also raises their outcome. This dependence between the causes of \(L\) is the source of the induced treatment–outcome association.
The pattern is not limited to binary variables. The following linear Gaussian model gives a compact derivation. Suppose \(A\) and \(U\) are independent, each has mean zero and variance one, and \(\varepsilon\) and \(\eta\) are independent mean-zero errors, also independent of \(A\) and \(U\). Let
Under an intervention setting \(A=a\), the mean of \(Y\) is \(\tau a\), because \(U\) and \(\eta\) have mean zero. The causal effect per one-unit increase in \(A\) is therefore \(\tau\). But after conditioning on \(L\), the conditional association between \(A\) and \(Y\) also reflects the association between \(A\) and \(U\).
Worked Example: Size and Direction in a Gaussian Model
Use the model above with \(\tau=1\), \(\gamma=3\), and \(\sigma^2=1\). The causal effect of a one-unit increase in \(A\) is \(\tau=1\) outcome unit. The conditional association after fixing \(L\), however, is
Thus, in this hypothetical model, the conditional slope is negative one-half outcome unit per one-unit increase in treatment, while the causal effect is positive one outcome unit. The difference comes from the opened collider path: at a fixed value of \(L\), larger \(A\) tends to coincide with smaller \(U\), and \(U\) raises \(Y\). The calculation also shows that the size of the induced association depends on the effect of \(U\) on \(Y\) and on the noise in \(L\); it is not a universal correction.
Recognizing the Structure in a Clinical Timeline
Consider a hypothetical study in which an initial treatment \(A\) affects whether a patient receives a follow-up test \(L\). A baseline symptom burden \(U\) also affects whether the test is given, and independently predicts a later outcome \(Y\). The relevant path is \(A \rightarrow L \leftarrow U \rightarrow Y\). If an analysis is restricted to tested patients, or includes \(L\) as an adjustment variable, it conditions on the common effect and can connect treatment to baseline symptom burden within the analyzed groups.
This example is about the causal structure, not a claim that every follow-up test is a collider. The arrows must be justified by the timing and substantive knowledge: treatment must affect the test, and another cause of the test must also affect the outcome. If those conditions do not hold, this particular opened path is not present. Conversely, the same logic applies to selection into follow-up, complete-case analysis, or any other restriction based on a post-treatment common effect.
Worked Example: A Different Bias Magnitude Under the Same Mechanism
Consider another hypothetical randomized setting with \(A\) and baseline \(U\) independent and equally frequent. Let \(L=1\) occur with probabilities 0.10 when \(A=0,U=0\), and 0.70 in each of the other three \((A,U)\) groups. Let the outcome score be \(Y=4A+6U\). Larger scores are simply defined to be better in this example; the purpose is to calculate how conditioning changes the comparison.
In a hypothetical sample of 1,000 people per \((A,U)\) group, the counts with \(L=1\) are 100, 700, 700, and 700, respectively. For \(A=1,L=1\), there are 700 people with each value of \(U\), so the mean outcome is
For \(A=0,L=1\), there are 100 people with \(U=0\) and 700 with \(U=1\). Their mean outcome is
The selected-group contrast is \(7.000-5.250=1.750\) outcome units. Yet before conditioning, \(U\) is equally common at each treatment level, so the mean outcome is 7 for \(A=1\) and 3 for \(A=0\); the causal effect is 4 units. Collider bias has attenuated the contrast here, whereas the previous examples showed more severe distortion. The direction and size depend on the outcome model and the selection probabilities.
Collider Bias Is Not the Only Consequence of Adjustment
A post-treatment covariate can have more than one causal role. For example, if \(L\) also affects \(Y\), then \(A \rightarrow L \rightarrow Y\) is a causal pathway. Conditioning on \(L\) can block that pathway, changing a total-effect comparison, while simultaneously opening \(A \rightarrow L \leftarrow U \rightarrow Y\) if \(U\) is also a cause of \(L\) and \(Y\). These are distinct mechanisms: blocking a causal path removes part of an effect, whereas opening the collider path creates a noncausal association. Both may occur in the same analysis.
The presence of the collider path alone does not determine the net bias of an adjusted estimate. Other paths may be open or blocked, and their contributions can reinforce or offset one another. The useful diagnostic question is narrower: does the proposed adjustment condition on a common effect of treatment and another cause of the outcome? If so, the adjusted comparison may no longer preserve the original balance between treatment groups.
Common Mistakes
- Assuming more adjustment must reduce bias. This is wrong because conditioning on a collider can create an association that was absent before adjustment. Draw the arrows into the covariate and check whether the resulting open path reaches the outcome through another cause.
- Calling every post-treatment covariate a collider. Being measured after treatment is not enough. A collider must be a common effect on the path, with arrows into it from both treatment and another variable.
- Confusing collider bias with blocking mediation. Blocking a path such as \(A \rightarrow L \rightarrow Y\) changes the causal pathways included in the comparison. Opening \(A \rightarrow L \leftarrow U \rightarrow Y\) creates a noncausal path. A covariate may do both, but the mechanisms are different.
- Assuming the bias must point in one direction. Its direction depends on how treatment and \(U\) affect \(L\), and how \(U\) affects \(Y\). The worked examples produce different magnitudes, including a reversal.
- Interpreting a conditional association as an intervention effect. A contrast among people with the same observed post-treatment \(L\) is not automatically the effect of assigning treatment while holding \(L\) fixed. The selected groups can differ in \(U\) precisely because they were selected on \(L\).
Check Your Understanding
Use the path structure and calculations in this tutorial to answer the following questions.
- In \(A \rightarrow L \leftarrow U \rightarrow Y\), which variable is the collider, and what happens to the path when we condition on it?
- Suppose \(A\) and \(U\) are independent before treatment, \(L\) is caused by both, and \(U\) raises \(Y\). Explain how conditioning on \(L\) can associate \(A\) with \(Y\).
- In the Gaussian model, what is the conditional slope when \(\tau=2\), \(\gamma=2\), and \(\sigma^2=1\)? How does it compare with the causal effect?
- Give one way a post-treatment covariate could both block a causal pathway and open a collider path.
- Why does the graph identify a potential collider problem without, by itself, specifying the numerical direction or size of the resulting bias?