Conditioning on a common effect can create a misleading association. Contrast colliders with confounders and trace a fictional agreement-only strategy archive.
Collider bias
Collider bias is distortion introduced by conditioning on a variable that is a common effect of other variables. Such a common effect is called a collider. Filtering a strategy archive by a result influenced by both backtest and forward performance can create an apparent relationship between them, even when the unfiltered outcomes are independent. Adding that result as a regression control can create the same kind of causal problem.
A third variable associated with both X and Y is not automatically a useful control. Compare two simple causal models:
Common cause:X←Z→Y. The shared cause Z creates a noncausal route between X and Y. Conditioning on Z blocks this route. This is the basic confounding adjustment argument.
Common effect:X→C←Y. This route is blocked while the collider and its descendants are unconditioned. Conditioning on C opens the route, allowing an association that does not represent an effect of X on Y.
Two stipulated causal graphs contrast a common cause Z, where conditioning blocks X left-arrow Z right-arrow Y, with a common effect C, where conditioning opens X right-arrow C left-arrow Y. Blue-filled nodes are conditioned on.
Arrow direction changes the adjustment decision. These are three-node causal models; an observed correlation cannot establish their arrows.
This is a path rule, not a universal instruction to remove every variable called a collider. A variable can have different roles along different paths. A causal adjustment decision concerns the whole graph, the effect of interest and the other controls. Pearl's d-separation rules explain when a path is blocked; Cinelli, Forney and Pearl's examples show why “more controls” is not a sufficient design principle.
An archive with suspiciously perfect agreement
Consider a fictional strategy-research archive. Define X=1 when a candidate passes its backtest screen and Y=1 when it passes a later paper-trading screen. A zero means failure. Stipulate an intentionally artificial causal model: the two verdicts are independent fair binary outcomes, with no persistent skill factor, shared market driver, or causal effect of one verdict on the other. They each equal one with probability 1/2.
After both tests, an analyst keeps only candidates whose verdicts agree, perhaps to assemble examples of consistent passes and consistent failures. Let C=1 mean a retained record:
C=1{X=Y}.
Both verdicts cause inclusion; inclusion causes neither verdict. Each full-population pair has probability 1/4. The agreement screen retains (0,0) and (1,1), so P(C=1)=1/2. Each retained pair therefore has conditional probability(1/4)/(1/2)=1/2.
An artificial agreement-only archive retains states 00 and 11 and excludes 01 and 10. Each full-population state has probability one quarter; each retained state has conditional probability one half. Correlation is zero before filtering and plus one after filtering.
Original stipulated model, not measured strategy performance. Perfect agreement is built into the archive's inclusion rule.
In the archive,
E[X∣C=1]=E[Y∣C=1]=21,E[XY∣C=1]=21.
Thus the conditional covariance is 1/2−(1/2)(1/2)=1/4. Each conditional variance is 1/4, so
Corr(X,Y∣C=1)=(1/4)(1/4)1/4=1.
The population correlation was zero. Within the archive, a backtest pass predicts a paper-test pass perfectly because every disagreement was removed. This says nothing about backtests predicting paper-trading success among all candidates. Under the stipulated structural model, forcing X to either value leaves P(Y=1) at 1/2; observing X inside the agreement archive is a different operation.
The positive sign is deliberate. The selection-bias article works a different screen that retains independent fair binary inputs when at least one equals one. Its selected correlation is −1/2. The common-effect graph warns of a possible distortion; the selection rule and probability law determine its sign and size.
Filtering is already conditioning
No regression is needed for the archive example: reading only records with C=1 has already conditioned the analysis. Deleting the inclusion column afterwards leaves exactly the same selected rows.
Conversely, retaining the full dataset does not make every adjustment safe. Including a collider as a control in a regression, stratifying by its value, or restricting the sample can open a noncausal path. These procedures need not give identical numerical estimates; the shared issue is conditioning on the wrong causal variable. Cinelli, Forney and Pearl's Models 16 and 17 develop this control-variable problem.
A descendant also matters. Suppose C affects a later review flag R. Selecting on R can open the route through C even if C itself is never used as a control. In the special case R=C, the retained archive is identical. A noisy descendant need not create the same magnitude of association. Pearl's collider rule explicitly includes descendants.
For a real research dataset, trace how each filter and proposed control was produced before interpreting a coefficient causally. In the fictional archive, recovering the discarded disagreements would restore the original joint law. When those records are unavailable, simply omitting C from the model cannot recover them. Describing the agreement archive is legitimate; extending its association to the full candidate population requires additional evidence or assumptions.
For the probability-table mechanics, Conditioning and Independence for Discrete Random Variables develops conditional laws and the distinction between conditional and unconditional independence. A public preview is available; the full lesson requires membership.
