Why Covariate Choice Is a Causal Decision
In an observational study, adjusting for a variable is not automatically helpful just because it predicts the outcome or strongly distinguishes treated from untreated people. The central question is whether the chosen pretreatment variables make treatment assignment conditionally comparable with respect to the potential outcomes. Covariate selection therefore requires a causal argument, not just a search for statistically significant associations.
Let \(A\) be a binary treatment, with \(A=1\) denoting treatment and \(A=0\) control; let \(Y\) be the observed outcome; and let \(Y(1)\) and \(Y(0)\) be the potential outcomes under treatment and control. A candidate covariate is pretreatment if it is measured before treatment assignment or initiation and is not itself changed by treatment. Write \(C\) for a set of pretreatment covariates considered for adjustment.
These roles are defined relative to a particular treatment, outcome, population, and causal understanding. A variable can play different roles in different studies, and a variable that predicts both treatment and outcome is not necessarily a confounder unless its causal role supports that conclusion. Covariate selection also depends on whether the measured variable is a useful representation of the relevant cause: a poorly measured confounder may not adequately control confounding.
What Should Adjustment Aim to Achieve?
For an observational causal comparison, adjustment should prioritize a set of pretreatment covariates \(C\) that makes treatment conditionally exchangeable: within levels of \(C\), treatment assignment is independent of the potential outcomes. It must also be possible to observe both treatment conditions at the covariate values relevant to the target population. These are substantive assumptions, not facts established merely by fitting a regression model.
The first condition is conditional exchangeability; the second is positivity. Consistency links the observed outcome to the potential outcome under the treatment actually received: \(Y=Y(A)\). Under these assumptions, the mean potential outcome under treatment \(a\) can be identified by standardizing the observed outcome means across covariate values:
For continuous covariates, the sum is replaced by an integral. The weights \(P(C=c)\) describe the chosen target population’s covariate distribution. The expression makes the purpose of confounder adjustment clear: compare treatment groups within covariate values, then average those comparisons over a common distribution rather than allowing the groups’ different covariate compositions to determine the contrast.
Classifying Candidate Covariates
Begin by asking what happens before treatment and how that feature relates causally to both treatment and outcome. Subject-matter knowledge, study procedures, and timing matter. A variable’s observed association with treatment or outcome can inform the investigation, but it does not by itself establish whether the variable is a common cause. Likewise, a non-significant association in a small sample does not show that a variable has no causal role.
Worked Example: Classifying Variables in a Training Study
Suppose a hypothetical observational study compares adults who enroll in a job-training program with adults who do not. The outcome is monthly earnings one year later. Before enrollment, investigators record prior education, baseline health, and distance to the training site. Based on substantive knowledge for this hypothetical setting, prior education affects both enrollment and later earnings; baseline health predicts earnings but does not affect enrollment after accounting for the study’s other baseline information; and distance affects enrollment but has no effect on earnings after that information is accounted for.
Prior education is a candidate confounder. It is a common cause of enrollment and earnings, so differing education distributions could make the unadjusted comparison misleading.
Baseline health is an outcome predictor in this setting. It helps predict earnings but does not explain why adults enroll. Its prognostic value alone does not make it necessary for confounding control.
Distance is a treatment predictor in this setting. It helps explain enrollment but does not predict earnings once the stated information is accounted for. Its association with treatment alone does not make it a confounder.
For illustration, suppose 30 of 50 adults living near the site enroll, while 10 of 50 adults living farther away enroll. The enrollment proportions are \(30/50=0.60\) near the site and \(10/50=0.20\) farther away. This shows that distance predicts enrollment in the hypothetical data; it does not, by itself, show that distance confounds the earnings comparison. The causal claim that distance has no outcome role comes from the stated substantive assumptions, not from this arithmetic.
Suppose attendance is also recorded after enrollment. It is not a pretreatment covariate, so it should not be treated as a baseline confounder. The timing and causal role of variables affected by treatment require particular care.
A Justification for Standardizing Over Confounders
The standardization formula follows from the law of total expectation and the identification assumptions. For a binary treatment level \(a\), first average \(Y(a)\) over the covariate distribution. Conditional exchangeability permits replacing the potential-outcome mean at each \(c\) with the corresponding observed outcome mean among people with \(A=a\) and \(C=c\). Consistency then equates that observed mean with the mean of \(Y(a)\) in that group:
The second equality uses conditional exchangeability; the final equality uses consistency. Positivity is needed so that the conditional observed mean in every weighted covariate stratum is estimable. If an important common cause is omitted, conditional exchangeability may fail, and this identification argument does not justify the adjusted contrast.
Worked Example: A Common Cause Reverses the Crude Comparison
Consider a hypothetical sample of 80 patients. Severity is recorded before treatment and is assumed to affect both treatment choice and the outcome. There are 40 high-severity and 40 low-severity patients. The table gives the number in each treatment-severity group and the mean outcome score in that group.
| Severity | Treated, count | Treated mean | Control, count | Control mean |
|---|---|---|---|---|
| High | 30 | 8.2 | 10 | 11.5 |
| Low | 10 | 4.3 | 30 | 5.2 |
The treated group contains \(30+10=40\) patients, and the control group also contains \(10+30=40\). Severity composition differs: 30 of 40 treated patients are high severity, whereas 10 of 40 control patients are high severity.
The crude mean among treated patients is \( (30\times8.2+10\times4.3)/40=(246+43)/40=7.225 \). The crude control mean is \( (10\times11.5+30\times5.2)/40=(115+156)/40=6.775 \). Thus the crude treated-minus-control contrast is \(7.225-6.775=0.450\) outcome points.
Now standardize both treatment groups to the sample’s severity distribution: \(40/80=0.50\) high severity and \(40/80=0.50\) low severity. The standardized treated mean is \(0.50(8.2)+0.50(4.3)=4.10+2.15=6.25\). The standardized control mean is \(0.50(11.5)+0.50(5.2)=5.75+2.60=8.35\). The sample-standardized contrast is therefore \(6.25-8.35=-2.10\) outcome points.
The crude and standardized contrasts differ in sign because the treatment groups have different severity compositions and the outcome means differ substantially by severity. The standardized estimate is a causal effect only if consistency, conditional exchangeability given severity, and positivity are credible. The arithmetic alone cannot verify those assumptions.
Why Predictive Strength Is Not the Selection Rule
A strong predictor of treatment can be useful for describing assignment and for diagnosing overlap, but it need not reduce confounding if it is unrelated to the outcome. Adjusting for such a variable may also make comparisons less precise or leave few comparable treated and control observations in some covariate strata. Conversely, an outcome predictor can improve prediction without being needed to make treatment groups exchangeable. In a randomized study, baseline outcome predictors may have a role in precision adjustment, but that is distinct from controlling confounding in an observational comparison.
Worked Example: A Treatment Predictor Alone Does Not Change the Contrast
Suppose a hypothetical sample has 100 people and a baseline variable \(Z\) indicating whether someone lives near a clinic. There are 50 people in each \(Z\) group. Among those living near the clinic, 40 are treated and 10 are controls; among those farther away, 10 are treated and 40 are controls. Assume, for illustration, that \(Z\) predicts treatment but does not affect the outcome, and that the treatment and control mean outcomes are 10 and 9 points, respectively, in each \(Z\) group.
The treatment proportions are \(40/50=0.80\) near the clinic and \(10/50=0.20\) farther away, so treatment assignment differs substantially by \(Z\). The treated mean is 10 in both strata, so its standardized mean is \(0.50(10)+0.50(10)=10\). The control mean is 9 in both strata, so its standardized mean is \(0.50(9)+0.50(9)=9\). The standardized contrast is \(10-9=1\) outcome point, the same as the within-stratum contrast.
In this hypothetical setup, adjustment for \(Z\) does not remove confounding because \(Z\) was not a common cause. Yet its strong relationship with treatment is relevant to overlap: only 10 controls are near the clinic and only 10 treated people are farther away. If some strata had no people in one treatment group, positivity would fail there, and a standardized contrast for the full target population could not be obtained from these data without additional assumptions or a changed target.
Diagnostics, Alternatives, and Limitations
A defensible adjustment set begins with a causal rationale and is then checked against the data and study design. Useful checks include whether each proposed covariate was measured before treatment, whether measurements are sufficiently reliable, whether treated and control participants overlap across covariate values, and whether the adjusted comparison retains adequate representation of the target population. Large differences in covariate distributions after adjustment can signal that the adjustment procedure has not achieved its intended balance; they do not prove that all confounding has been removed.
Regression adjustment, stratification, matching, and weighting are different ways to construct adjusted comparisons. Their calculations differ, but none can compensate for a covariate set that fails to address relevant confounding, a lack of overlap, or an implausible causal model. Randomization addresses treatment assignment by design rather than relying on measured covariates for exchangeability. In observational research, unmeasured common causes remain a limitation even after carefully chosen measured covariates are adjusted for.
Use extra caution with variables recorded after treatment begins. They may be affected by treatment, so conditioning on them can change the comparison being made. The earlier material on post-treatment adjustment explains this concern. Also avoid selecting covariates solely because they happen to be statistically significant in the observed sample, or including every available variable without considering its timing and causal role.
Common Mistake
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A pretreatment variable causes both treatment choice and the outcome. What role does it play, and why is it a priority for adjustment?
- A variable predicts treatment but, under the study’s substantive assumptions, does not affect the outcome. Does its treatment association alone make it a confounder?
- In a sample of 200 people, 100 have a baseline risk factor, and the other 100 do not. Suppose the risk factor is a common cause. What weights would standardization use for these two groups?
- Why does a standardized contrast not automatically establish a causal effect?
- A candidate variable is measured after treatment begins. What is the key concern with treating it as a pretreatment adjustment covariate?
Solutions. (1) It is a candidate confounder because it is a common cause of treatment and outcome; adjusting for it can help make the groups conditionally comparable. (2) No. It is a treatment predictor in the stated setting, but prediction of treatment alone does not make it a common cause. (3) The sample proportions are \(100/200=0.50\) in each group, so standardization uses weights 0.50 and 0.50. (4) A causal interpretation also requires credible consistency, conditional exchangeability for the adjusted set, and positivity; calculation does not establish these assumptions. (5) Treatment may affect the variable, so conditioning on it can change the causal comparison rather than control a baseline common cause.