Why Confounder Identification Is Decision-Specific
In a longitudinal study, treatment may be decided repeatedly as a person’s health and circumstances change. A covariate recorded before one decision may help explain that decision and be related to what happens afterward. At a later decision, the same named covariate may have a different value, a different history around it, and a different role. The useful question is therefore not simply “Is this variable a confounder?” but “Is this measurement relevant to this treatment decision, given what was already known?”
The first requirement is temporal: the covariate must be observed before the decision being studied. The next is conditional: its relationship with treatment and subsequent outcomes should be considered among people with comparable prior histories. These checks help identify candidate confounders. They do not, by themselves, prove that a covariate is a cause of treatment or outcome, nor do they establish that all confounding has been removed.
What to Check at Each Decision
Start by reconstructing the ordering of measurements. If \(L_t\) was measured after \(A_t\), it cannot explain the choice made at that decision, even if it is strongly associated with treatment or outcome. A measurement taken before \(A_t\) can be considered, but timing alone does not make it a confounder.
Next, compare treatment decisions across values of \(L_t\) within levels of prior history. For a binary treatment, one descriptive comparison is the conditional treatment probability \(P(A_t=1\mid L_t,H_t)\). If that probability differs by \(L_t\) for people with the same relevant prior history, the covariate may help explain who receives treatment. The comparison should represent history adequately; a crude association pooled over very different histories can be misleading.
Then consider whether \(L_t\) is related to outcomes occurring after the decision, again accounting for prior history. For example, compare outcome risks across \(L_t\) levels within history groups. Because treatment itself is related to subsequent outcomes, it can also be informative to describe the relationship within treatment groups. Such observed comparisons are diagnostic, not definitive causal tests: the way treatment is assigned can itself make a within-treatment comparison misleading. Subject-matter knowledge about how the covariate could affect treatment and outcome remains important.
These checks separate three common patterns. A covariate associated with treatment but not with subsequent outcomes, conditional on relevant history, may be a treatment predictor without being an important confounder. A prognostic covariate that does not help explain treatment may still be useful for other analyses, but does not meet the basic two-part screening idea for this decision. A covariate related to both is a stronger candidate, though its causal role and the adequacy of the measured history still require judgment.
Worked Example: A Covariate Related to Both Treatment and Outcome
Suppose a hypothetical clinic records a patient’s severity level \(L_t\) immediately before a treatment decision. Within one prior-history group \(H_t\), 100 patients have high severity and 100 have low severity. The following counts show treatment and a later adverse outcome.
| Severity | Treated, adverse outcome | Treated, no adverse outcome | Untreated, adverse outcome | Untreated, no adverse outcome |
|---|---|---|---|---|
| High | 28 | 42 | 12 | 18 |
| Low | 6 | 24 | 14 | 56 |
Treatment association. Among patients with high severity, 70 of 100 receive treatment, so \(P(A_t=1\mid L_t=\text{high},H_t)=70/100=0.70\). Among those with low severity, \(30/100=0.30\) receive treatment. Within this history group, severity is strongly associated with the decision.
Outcome association. Among treated patients, the adverse-outcome risk is \(28/(28+42)=28/70=0.40\) for high severity and \(6/(6+24)=6/30=0.20\) for low severity. Among untreated patients, the corresponding risks are \(12/(12+18)=12/30=0.40\) and \(14/(14+56)=14/70=0.20\). In both treatment groups, high severity is associated with a 20-percentage-point higher observed risk.
This pattern makes pre-decision severity a strong candidate confounder for this decision: it is related to treatment and to subsequent outcome within the stated prior history. It does not prove that changing severity would cause either difference, and it does not rule out other unmeasured causes. The example illustrates a classification screen, not a complete causal analysis.
Why Conditioning on Prior History Matters
A conditional relationship can differ from the pooled relationship because people with different covariate values may have different distributions of prior history. If prior history predicts treatment, and the mix of histories differs across \(L_t\), a pooled comparison can make \(L_t\) look treatment-related even when treatment rates are identical at each level of history.
Worked Example: A Pooled Treatment Difference That Disappears Within History
Consider a hypothetical study with a binary prior-history measure \(H_t\) and a binary covariate \(L_t\). The counts of treatment decisions are:
| Prior history | \(L_t\) | Number of people | Number treated | Treatment proportion |
|---|---|---|---|---|
| Lower-risk history | 0 | 80 | 40 | 40/80 = 0.50 |
| Lower-risk history | 1 | 20 | 10 | 10/20 = 0.50 |
| Higher-risk history | 0 | 20 | 18 | 18/20 = 0.90 |
| Higher-risk history | 1 | 80 | 72 | 72/80 = 0.90 |
Within the lower-risk history group, the treatment proportion is 0.50 for both covariate levels. Within the higher-risk group, it is 0.90 for both. Thus, after comparing people with the same \(H_t\), there is no treatment difference by \(L_t\).
Pooling the groups gives a different result. For \(L_t=1\), \(82\) of \(100\) people are treated, so the pooled proportion is \(82/100=0.82\). For \(L_t=0\), \(58\) of \(100\) are treated, so it is \(58/100=0.58\). The pooled difference is \(0.82-0.58=0.24\), or 24 percentage points. That difference arises because the \(L_t=1\) group contains more people with higher-risk history, where treatment is more common. It is not evidence that \(L_t\) is associated with treatment within prior history.
The lesson is to examine treatment assignment conditional on the relevant history, rather than treating a crude comparison as decisive. Which elements belong in \(H_t\) depends on the timing and substantive structure of the study.
Worked Examples of Other Covariate Roles
Worked Example: A Treatment Predictor With Little Outcome Association
In a hypothetical history group, 100 people have \(L_t=1\) and 100 have \(L_t=0\). Among \(L_t=1\), 70 receive treatment and 30 do not; among \(L_t=0\), 30 receive treatment and 70 do not. The treatment proportions are \(70/100=0.70\) and \(30/100=0.30\), a 40-percentage-point difference.
Suppose the subsequent adverse-outcome counts are as follows: for \(L_t=1\), 14 of 70 treated people and 9 of 30 untreated people have the outcome; for \(L_t=0\), the corresponding counts are 6 of 30 and 21 of 70.
Within the treated group, risks are \(14/70=0.20\) and \(6/30=0.20\). Within the untreated group, they are \(9/30=0.30\) and \(21/70=0.30\). Thus, in these data, \(L_t\) strongly predicts treatment but shows no outcome-risk difference within treatment groups. It is not supported as a confounder by the two-part observed-association screen, though measurement limitations or other causal structure could change the assessment.
Worked Example: A Prognostic Covariate That Does Not Predict Treatment
Now suppose a hypothetical prior-history group contains 50 people with \(L_t=1\) and 50 with \(L_t=0\). In each group, 20 people receive treatment and 30 do not. Therefore, the treatment proportions are equal: \(20/50=0.40\) for both covariate levels.
Suppose 8 of the 20 treated people with \(L_t=1\) have a subsequent adverse outcome, compared with 4 of the 20 treated people with \(L_t=0\). The treated-group risks are \(8/20=0.40\) and \(4/20=0.20\). Among untreated people, suppose 12 of 30 with \(L_t=1\) and 6 of 30 with \(L_t=0\) have the outcome. The untreated-group risks are \(12/30=0.40\) and \(6/30=0.20\).
The covariate is prognostic in the sense that its levels are associated with different subsequent outcome risks in both treatment groups. But it does not predict treatment in this history group. On these comparisons, it is not a candidate confounder for this particular treatment decision under the two-part screen. Prognostic importance and confounding are related but distinct ideas.
A Practical Classification Sequence
The following sequence keeps the classification tied to a specific decision and makes the reasoning explicit.
Name the treatment \(A_t\), the outcome \(Y\), and the follow-up period in which \(Y\) occurs. Do not classify a covariate without specifying which treatment decision it is being assessed against.
Check that \(L_t\) was measured before \(A_t\) and that \(Y\) is subsequent to the decision. A post-decision measurement cannot be a cause of that already-made decision.
Ask whether treatment proportions or probabilities differ across \(L_t\) values among people with comparable \(H_t\). Avoid relying only on pooled associations.
Compare outcomes across \(L_t\) values conditional on prior history, and describe treatment-specific comparisons where useful. Interpret these as observed associations, not automatic proof of a causal effect.
State whether the covariate is a plausible candidate for this decision, explain which comparisons support that judgment, and note important uncertainties such as measurement error, unmeasured history, or limited data.
A Short Justification for Conditional Comparisons
The reason for comparing within prior-history strata can be seen directly from the law of total probability. For a binary covariate \(L_t\), the pooled treatment probability is a weighted average of the history-specific probabilities:
Even when \(P(A_t=1\mid L_t=l,H_t=h)\) is the same for \(l=0\) and \(l=1\) at every value of \(h\), the pooled probabilities can differ if \(P(H_t=h\mid L_t=l)\) differs across covariate groups. In the preceding example, the treatment probabilities were 0.50 and 0.90 in the two history strata for either covariate value, but the covariate groups had different proportions in those strata. The pooled comparison therefore reflected history composition rather than a within-history treatment association.
This calculation justifies conditioning on measured prior history when describing the covariate-treatment relationship. It does not show that the chosen history is complete or that conditioning alone identifies a causal effect. Those are broader design and identification questions.
Common Mistakes and Limits of the Screen
- Calling every treatment predictor a confounder. A variable can predict who receives treatment without being related to subsequent outcomes after relevant history is considered. Treatment prediction alone is not enough for the basic confounder screen.
- Calling every prognostic variable a confounder. A variable related to outcomes but unrelated to the treatment decision within prior history is prognostic, but may not confound that decision.
- Using only a pooled comparison. A crude association can be created, hidden, or reversed by differing prior histories. Examine conditional comparisons and explain which history is being held comparable.
- Inferring causal status from association alone. Observed conditional associations are evidence for screening, not proof that the covariate causes treatment or outcomes. Causal knowledge, measurement quality, and unmeasured factors matter.
- Ignoring measurement timing. A covariate measured after \(A_t\) may be associated with treatment, but it cannot be a pre-decision confounder of that decision.
- Assuming a permanent label. A variable can be relevant to one treatment decision and not another. Reassess its timing and conditional relationships at each decision.
Check Your Understanding
Use the decision-specific, conditional approach to answer the following.
- A symptom measure is recorded after a treatment decision and predicts later outcome. Can it be a confounder of that decision? Explain using measurement order.
- Within each prior-history group, treatment proportions are the same for two levels of a covariate, but pooled proportions differ. What should you conclude about the pooled association?
- A covariate predicts treatment within prior history but has no observed association with subsequent outcomes within treatment groups. What role is supported by these comparisons, and what limitation remains?
- Why is a covariate associated with outcome but not treatment not automatically a confounder for the current decision?
- What should be reconsidered when assessing the same named covariate at a later treatment decision?