Collider Bias in Repeated-Measures Studies

Why Conditioning Can Change an Association

In a repeated-measures study, a variable recorded after treatment may be affected both by earlier treatment and by a person’s underlying prognosis. Restricting an analysis to a particular value of that variable—or including it as a conditioning variable—can make treatment history and prognosis associated, even if they were independent before follow-up began. If prognosis also predicts the outcome, this induced association can distort a treatment comparison.

The central issue is the variable’s position in the causal graph. A variable that is a common effect of two variables is a collider on the path between them. Conditioning on a collider can open that path. This differs from conditioning on a common cause, which can block a noncausal path. The fact that a variable is measured before the outcome does not, by itself, tell us which role it plays.

Definition: On a path in a directed acyclic graph, a node is a collider if the two arrows on that path both point into it; the basic pattern is $A_0 \rightarrow L_1 \leftarrow U$. Conditioning on a collider, or on a descendant of that collider, can make its causes statistically associated. This is called collider bias when the induced association distorts an analysis or interpretation.

Here $A_0$ denotes treatment at an earlier time, $L_1$ a later covariate, $U$ a prognostic factor, and $Y$ a later outcome. The key longitudinal structure is $A_0 \rightarrow L_1 \leftarrow U \rightarrow Y$. Before conditioning on $L_1$, the path from $A_0$ to $Y$ is blocked at the collider. Conditioning on $L_1$ can open it. The resulting association between $A_0$ and $U$ can be mistaken for evidence about treatment and the outcome, even though it was created by the conditioning.

A directed acyclic graph represents the assumed causal structure; the direction of an arrow encodes causal influence, not merely association or measurement order. In the examples below, $U$ is a baseline prognostic factor. It may be unobserved, or it may be measured but inadvertently conditioned on a collider in a way that induces the relevant association.

A General Probability Derivation

The graphical pattern describes a possibility, not a guarantee: conditioning on every common effect will not necessarily produce dependence. The following calculation states exactly when the conditional distribution fails to factorize in a simple case.

Result: Suppose $A$ and $U$ are independent, and a collider $C$ has conditional probability $P(C=c\mid A=a,U=u)$. For values with $P(C=c)>0$, $$ P(A=a,U=u\mid C=c) = \frac{P(C=c\mid A=a,U=u)P(A=a)P(U=u)}{P(C=c)}. $$ Thus conditioning can associate $A$ and $U$: their conditional joint distribution need not equal the product of their conditional marginal distributions. It remains independent only when the displayed joint distribution factorizes over $a$ and $u$.

Proof. Independence gives $P(A=a,U=u)=P(A=a)P(U=u)$. By Bayes’ rule, $$ P(A=a,U=u\mid C=c) = \frac{P(C=c\mid A=a,U=u)P(A=a,U=u)}{P(C=c)}. $$ Substituting the independence factorization yields the stated expression. Its numerator includes the joint dependence of $C$ on its two causes. In general, that dependence does not separate into a function of $a$ multiplied by a function of $u$; after conditioning, the distribution of one cause can therefore depend on the value of the other. The denominator is a constant with respect to $a$ and $u$, so it cannot generally restore factorization. Special models can still yield independence, which is why the conclusion is that collider conditioning can induce association, not that it must. $\square$

Worked Example: A Binary Post-Treatment Covariate

Let baseline treatment $A_0$ and prognosis $U$ be independent Bernoulli variables, each with probability $1/2$ of being 1. Define a follow-up covariate $L_1=1$ if either $A_0=1$ or $U=1$, and $L_1=0$ otherwise. Let the later outcome be $Y=U$. The graph has $A_0 \rightarrow L_1 \leftarrow U \rightarrow Y$. What association appears if the analysis conditions on $L_1=1$?

The four joint pairs $(A_0,U)$ are equally likely, each with probability $1/4$. The condition $L_1=1$ retains $(1,0)$, $(0,1)$, and $(1,1)$, so $P(L_1=1)=3/4$. Each retained pair consequently has conditional probability $(1/4)/(3/4)=1/3$.

Among those with $A_0=1$ and $L_1=1$, the possible pairs are $(1,0)$ and $(1,1)$. They have equal conditional probability, so $$ P(Y=1\mid A_0=1,L_1=1) = P(U=1\mid A_0=1,L_1=1) = \frac{1}{2}. $$ Among those with $A_0=0$ and $L_1=1$, only $(0,1)$ is possible. Therefore $$ P(Y=1\mid A_0=0,L_1=1)=1. $$

Within the selected group, the observed outcome risk difference comparing $A_0=1$ with $A_0=0$ is $1/2-1=-1/2$. Yet $Y=U$ and $A_0$ was initially independent of $U$, so intervening on $A_0$ has no effect on $Y$ in this model. Selection on $L_1$ has made treatment and prognosis negatively associated: when $A_0=0$, $U=1$ is required to enter the selected group; when $A_0=1$, either value of $U$ is possible.

How the Pattern Appears Over Time

A collider need not be the outcome or a selection indicator. It can be an ordinary repeated-measures covariate. For instance, a patient’s early treatment may affect a follow-up biomarker, while baseline severity also affects that biomarker. If severity predicts a later outcome, conditioning on the biomarker can induce an association between early treatment and severity. The same logic applies if the common effect is a later treatment decision: earlier treatment and prognosis may both influence that decision.

Worked Example: A Later Treatment Decision as a Common Effect

Suppose $A_0$ is an initial treatment, $U$ is baseline prognosis, and $A_1$ is a later treatment decision. For illustration, let $A_0$ and $U$ be independent Bernoulli variables with probabilities $1/2$, and set $A_1=1$ when either $A_0=1$ or $U=1$. Let $Y=U$. This gives the longitudinal graph $A_0 \rightarrow A_1 \leftarrow U \rightarrow Y$. Consider comparing initial-treatment groups only among people with $A_1=1$.

The selected histories are $(A_0,U)=(1,0),(0,1),(1,1)$, each with conditional probability $1/3$, because the selection event has probability $3/4$. Therefore, among those with $A_1=1$, $$ P(Y=1\mid A_0=1,A_1=1)=\frac{1}{2}, \qquad P(Y=1\mid A_0=0,A_1=1)=1. $$ The conditional risk difference is again $-1/2$, although $A_0$ does not cause $Y$ in the specified model. Conditioning on the later treatment decision has opened the path through $U$.

This example isolates the collider mechanism; it does not imply that later treatment should always be excluded from every analysis. The role of a variable depends on the causal question and the graph. The relevant lesson here is narrower: conditioning on a later decision that is affected by both earlier treatment and prognosis can alter their association.

Exact Dependence in a Gaussian Model

A continuous example shows that the induced association is not an artifact of binary variables. Suppose $A_0$, $U$, and $\varepsilon$ are mutually independent, each normally distributed with mean zero; let $\operatorname{Var}(A_0)=\operatorname{Var}(U)=1$, and let $\operatorname{Var}(\varepsilon)=\sigma^2>0$. Define $L_1=A_0+U+\varepsilon$ and $Y=U$. The common-effect structure is again $A_0 \rightarrow L_1 \leftarrow U \rightarrow Y$.

Result: In this Gaussian model, $$ \operatorname{Cov}(A_0,U\mid L_1) = -\frac{1}{2+\sigma^2}. $$ Thus $A_0$ and $U$, though marginally independent, are negatively associated conditional on the measured value of $L_1$.

Worked Example: Conditional Covariance With Measurement Noise

To derive the result, first calculate $\operatorname{Var}(L_1)=1+1+\sigma^2=2+\sigma^2$. Also, $\operatorname{Cov}(A_0,L_1)=1$, $\operatorname{Cov}(U,L_1)=1$, and $\operatorname{Cov}(A_0,U)=0$. For jointly normal variables, the conditional covariance is $$ \operatorname{Cov}(A_0,U\mid L_1) = \operatorname{Cov}(A_0,U) - \frac{\operatorname{Cov}(A_0,L_1)\operatorname{Cov}(L_1,U)} {\operatorname{Var}(L_1)}. $$

Substituting the calculated quantities gives $$ \operatorname{Cov}(A_0,U\mid L_1) = 0-\frac{1\cdot1}{2+\sigma^2} = -\frac{1}{2+\sigma^2}. $$ Since $\sigma^2>0$, the denominator is positive and the conditional covariance is strictly negative. For example, if $\sigma^2=2$, the covariance is $-1/4$. More noise weakens the induced association: as $\sigma^2$ increases, its magnitude $1/(2+\sigma^2)$ decreases. The calculation relies on joint normality and the stated independent errors; it is not a universal formula for arbitrary distributions.

The conditional-covariance identity follows from the multivariate normal conditional distribution: for a jointly normal vector, conditioning on a scalar $L_1$ subtracts $\operatorname{Cov}(A_0,L_1)\operatorname{Cov}(L_1,U)/\operatorname{Var}(L_1)$ from the original covariance. Here all conditional variances are well-defined because $\operatorname{Var}(L_1)=2+\sigma^2>0$. $\square$

Diagnostics, Alternatives, and Limits

A useful diagnostic is to inspect the arrows into every variable on which the analysis conditions. Ask whether a conditioned variable is caused by both treatment history and a prognostic factor, and whether that factor predicts the outcome. If so, mark the collider path explicitly. A change in an estimated association after adjustment may be a warning, but it does not establish collider bias: the change could have other explanations, and a numerical diagnostic alone cannot identify the graph.

Compare three distinct structures before deciding what to condition on:

Structure	Role of conditioned variable	Graphical implication
$A_0 \leftarrow U \rightarrow Y$	$U$ is a common cause	Conditioning on $U$ can block this noncausal path.
$A_0 \rightarrow L_1 \leftarrow U \rightarrow Y$	$L_1$ is a common effect	Conditioning on $L_1$ can open the path through $U$.
$A_0 \rightarrow L_1 \rightarrow Y$	$L_1$ lies on a directed path	This is a different graphical role; collider reasoning alone does not settle whether to condition on it.

The third row is included only to distinguish graph shapes: deciding how to handle a covariate on a directed path requires a separate causal question. A common mistake is to adjust for every measured variable that predicts the outcome. Correction: outcome prediction alone is not sufficient justification for conditioning. First determine whether the variable is a cause, a common effect, or part of a directed path under the assumed graph. Conversely, the presence of a collider does not mean that every analysis conditioned on it is necessarily biased; the association may be absent under special distributions, or the target question may require a carefully specified conditional contrast.

This reasoning has limits. A DAG encodes assumptions and cannot, by itself, show that a proposed arrow is true. Collider bias can be difficult to diagnose from observed data alone, especially when the relevant prognostic factor is unmeasured. Also, the examples isolate one path; real longitudinal studies may contain several paths, feedback across treatment and covariates, and other sources of bias. The purpose of the collider check is to identify a potential mechanism, not to replace a full analysis of the causal structure.

Locate the conditioned node.
List variables fixed by restriction, stratification, or inclusion in a conditional comparison.

Inspect its incoming arrows.
Check whether treatment history and prognosis are both causes of that node.

Trace the opened path.
Ask whether the prognostic factor also predicts the outcome, creating a noncausal treatment-outcome association after conditioning.

Interpret the comparison cautiously.
Do not treat a conditional association as a causal effect without considering the full graph and the target contrast.

Check Your Understanding

Answer each question, then compare with the solutions below.

In $A_0 \rightarrow L_1 \leftarrow U \rightarrow Y$, which node is the collider, and what happens to the path when conditioning on it?
If $A_0$ and $U$ are initially independent, does conditioning on their common effect always make them dependent? Explain.
In the binary example, what is $P(Y=1\mid A_0=0,L_1=1)$, and why?
In the Gaussian model with $\sigma^2=6$, calculate $\operatorname{Cov}(A_0,U\mid L_1)$.
Why is “this variable predicts the outcome” not, by itself, a sufficient reason to condition on it?

Solutions. (1) $L_1$ is the collider because both arrows point into it. Without conditioning, it blocks the path; conditioning can open the path $A_0-L_1-U-Y$. (2) No. The conditional joint distribution need not factorize, but special distributions or causal mechanisms can preserve independence. (3) The probability is 1: if $A_0=0$ and $L_1=1$ under $L_1=A_0\lor U$, then $U=1$, and $Y=U$. (4) The covariance is $-1/(2+6)=-1/8$. (5) The variable may be a common effect or lie on another causal path; conditioning can open a noncausal path rather than remove confounding.

Structure	Role of conditioned variable	Graphical implication
\(A_0 \leftarrow U \rightarrow Y\)	\(U\) is a common cause	Conditioning on \(U\) can block this noncausal path.
\(A_0 \rightarrow L_1 \leftarrow U \rightarrow Y\)	\(L_1\) is a common effect	Conditioning on \(L_1\) can open the path through \(U\).
\(A_0 \rightarrow L_1 \rightarrow Y\)	\(L_1\) lies on a directed path	This is a different graphical role; collider reasoning alone does not settle whether to condition on it.