Collider checks: a control variable can create an association
Adding another control variable can create an association that was absent before adjustment. The key question is what generates that variable. If two factors both affect selection into a dataset, restricting analysis to selected cases can make those factors associated even when they are independent in the full population.
The deliverable is a documented causal diagram, a full-population and selected-sample table, and an explanation of which association the selection creates. A diagram is a structural assumption informed by substantive knowledge. It is not a conclusion that an algorithm can read automatically from correlations.
Identify the common result being conditioned on
In a hypothetical company registry, let X indicate high demand and Y indicate high management capability. Let S indicate inclusion in a research showcase. Assume for the example that X and Y are independent inputs and that the showcase includes a company if either factor is high: S = X OR Y.
The relevant diagram is X → S ← Y. S is a collider because arrows from both inputs meet there. The example assumes no causal arrow from X to Y or Y to X and no common cause of those two inputs. Those exclusions are essential; they are not inferred just because the table below looks balanced.
Hernán and Robins, Causal Inference: What If, August 19, 2026 version, chapter 8 section 8.1 and Figures 8.1–8.2, explains selection on a common effect and its descendants. The calculation here is an original company example of that structure, rather than an estimate of an actual management-demand relationship.
Reconstruct the unselected population first
Assume a finite illustrative population of 400 companies, with 100 in each combination:
| Demand X | Capability Y | Population count | Showcase S |
|---|---|---|---|
| 0 | 0 | 100 | 0 |
| 0 | 1 | 100 | 1 |
| 1 | 0 | 100 | 1 |
| 1 | 1 | 100 | 1 |
In the full population, capability is high for 100/200=0.50 among high-demand companies and 100/200=0.50 among low-demand companies. X and Y are independent in this constructed table. Their covariance and correlation are zero.
Selection removes the 100 companies for which both factors are low. The remaining 300 form the showcase sample. Among selected high-demand companies, capability is high in 100/200=0.50. Among selected low-demand companies, capability is high in 100/100=1.00.
Thus the selected-sample difference P(Y=1|X=1,S=1)−P(Y=1|X=0,S=1) is −0.50. A negative association has appeared without a causal effect of demand on capability. At low demand, selection requires high capability; at high demand, selection can occur with either capability level. The alternative routes into the dataset explain the pattern.
As an independent arithmetic check, within the selected sample both marginal high proportions are 2/3 and the joint high proportion is 1/3. Their covariance is 1/3−(2/3)²=−1/9; each binary variance is 2/9, so the correlation is −1/2. These are descriptive properties of the selected population, not a causal effect estimate.
Why a selected regression does not establish a tradeoff
A regression of Y on binary X within S=1 would reproduce the conditional mean difference −0.50 in this simple table. It would not show that increasing demand lowers management capability. Under the stipulated structural model, intervening on X leaves the independently generated Y distribution at 0.50 in the full population.
The apparent “capability compensates for demand” pattern is a consequence of the entry rule. Changing demand can also change which companies qualify, so comparing selected sets under different demand values changes their composition. That is not the same intervention contrast on a fixed target population.
The example separates three layers. The structural layer stipulates the arrows and entry rule. The identification layer asks whether the desired intervention contrast can be recovered from available selected data under additional assumptions. The estimation layer computes probabilities or a coefficient from the observed table. Accurate estimation of a selected association cannot replace the missing identification argument.
Selection is not automatically collider bias
For a different boundary, suppose inclusion is an independent sampling process that happens to retain 75 companies from each original cell. The selected total is again 300, the same 75% fraction as before. Yet the two conditional capability proportions remain 0.50, and no demand-capability association is created in that constructed sample.
The fraction selected alone therefore does not diagnose the direction or magnitude of bias. It matters why selection occurs and how it relates to the variables. Conversely, if selection were based only on X and retained only X=1, the selected sample would lack a demand comparison; the problem would not be an estimable negative contrast like the main example.
A common cause also differs from a common effect. If a pre-existing factor L affects both X and Y, adjusting for L may be needed to address confounding. Labeling every available control as dangerous would be as mistaken as assuming every additional control helps. Determine its temporal and causal role for the specific target.
Audit filters, controls and their timing
Conditioning can occur before a regression is fitted. Restricting a dataset to approved cases, complete records, survivors or showcase participants is already a conditioning operation. Removing the S column after filtering does not restore the excluded companies.
A variable derived from S can also carry selection information. For example, a response available only to showcase invitees may define a selected subset even when the word “showcase” appears nowhere in the final analysis. Trace how the dataset was assembled, including upstream dashboards and manual review rules.
| Review step | Evidence to obtain | If the role is unresolved |
|---|---|---|
| Define the target | population and intervention or association | avoid an unspecified causal claim |
| Map timing | when X, Y and selection were determined | do not call a later result a baseline cause |
| Document selection | eligibility, approvals and missingness rules | retain competing plausible diagrams |
| Rebuild cells | counts before and after filters | label absent population information |
| Classify each control | common cause, mediator, collider or other role | justify adjustment rather than maximizing controls |
| Interpret output | selected association versus identified contrast | keep the two labels distinct |
If selection rules are unavailable, the 400-company table is a diagnostic possibility, not proof that the real dataset suffered exactly this bias. Seek enrollment records, rejected cases or information about both selection inputs. Do not infer the OR rule from a negative fitted coefficient and then cite the same coefficient as confirmation.
Decide what new evidence could resolve the problem
Access to excluded cases can reveal how the observed association differs from the original population. A well-designed sampling study may support a selection model, but correcting selection requires its own assumptions and support. In the deterministic showcase example, the both-low cell never appears in S=1; reweighting the observed sample alone cannot manufacture its missing outcome information.
The population table was supplied by construction here. In an actual research project it may be unknown. Preserve that distinction: a complete hypothetical demonstration explains a failure mode without establishing that the full population has the same counts.
Revise the analysis if documentation changes the arrows, if new data recover excluded cells, or if the target is explicitly narrowed to a selected population. A narrower descriptive question can be legitimate, but it must not inherit a causal interpretation from the original broader question. The final adjustment record should state why each control is present and what comparison its inclusion permits, rather than presenting a long control list as evidence of reliability.
Filtering on S differs from intervening on S
In the stipulated showcase mechanism, conditioning on S=1 retains only 300 original companies and yields the negative association. Consider instead a hypothetical admission-policy intervention that sets S=1 for all 400 while leaving the generation of demand and capability unchanged. Both inputs retain their original distribution; the capability proportion remains 0.50 at each demand level.
That difference separates an observational restriction from a causal intervention on the selection mechanism. It depends on the explicit assumption that changing admission itself does not change X or Y. If an actual showcase program improves capability or changes demand, the proposed intervention has additional pathways and requires a different diagram. Do not translate “hold S constant in a regression” into “make every company eligible” without identifying which operation the study actually represents.
If the intended question is purely descriptive of showcase participants, their negative association is a valid feature of that selected population. Bias is assessed relative to a stated target, such as the original population association or an intervention contrast. Report that target before diagnosing the difference as an error.
FAQ
Does a negative association among selected companies prove a causal tradeoff?
No. The original independent inputs acquire a negative association under the stipulated showcase rule.
Should all controls and filters be removed?
No. Their roles differ. Common causes, mediators and colliders require different decisions tied to the causal question and timing.