Adjustment cannot repair a broken causal question

#Question

Why does an adjusted estimate not automatically have a causal interpretation?

#Answer in brief

Because adjustment is an analytical operation, not a causal design. A regression model can estimate an association conditional on selected variables. Whether that association estimates a causal effect depends on the target question, timing, eligibility, treatment strategies, follow-up, confounding structure, selection, measurement and assumptions.

Target-trial frameworks make this ordering explicit: specify the causal experiment you would ideally conduct, then design the observational analysis to emulate its protocol as closely as possible.^[1][2]

A model cannot repair a question that was never coherently defined.

#Start with the estimand, not the software

Consider a routine-health-data study asking whether vaccination reduces hospitalization.

Before choosing covariates, the study must define which population, which vaccine, what counts as vaccinated, when follow-up starts, what comparator, which hospitalization outcome, what causal contrast and over what period.

Without those choices, "the effect of vaccination" is underspecified. An adjusted odds ratio cannot resolve that ambiguity.

#Confounders are causal roles, not columns with small p-values

A confounder is a variable related to the causal structure connecting exposure and outcome.

It is not simply any baseline variable, any variable associated with outcome, any variable that changes the coefficient, or any variable with a small p-value.

The correct adjustment set depends on the causal question. Causal diagrams are useful precisely because they make assumptions about common causes and adjustment paths explicit rather than treating confounder selection as a purely statistical exercise.^[3] A data-driven variable-selection procedure can therefore choose the wrong variables with great statistical confidence.

#Adjustment can open bias

Imagine that health-seeking behaviour affects both vaccination and hospitalization. That variable may confound the vaccination–hospitalization relationship.

But a variable measured after vaccination can be a mediator, consequence, or collider depending on the structure. Adjusting mechanically for every available variable can block part of the effect of interest or open a noncausal association. Overadjustment can therefore introduce bias rather than simply make a model more conservative.^[4]

The problem is structural, not computational.

#Mediators answer a different question

Suppose vaccination reduces hospitalization partly by preventing infection.

If infection is included as a covariate in a model estimating the total effect of vaccination on hospitalization, the model may remove part of the pathway through which vaccination works.

That does not mean mediators should never be adjusted for. It means the estimand must say whether the target is a total effect, a direct effect or another causal contrast.

Different questions require different assumptions.

#Colliders can create associations

A collider is a variable influenced by two causes.

Conditioning on it can create an association between those causes even if they were independent before conditioning. Causal-diagram methods formalize this collider problem and help distinguish variables that should from those that should not be conditioned on for a given estimand.^[3]

Health-care datasets frequently select people through care pathways. Studying only hospitalized patients may condition on hospitalization, a variable influenced by exposure and other determinants. No amount of additional adjustment necessarily removes bias created by that selection mechanism.

#Temporal ordering is part of causality

Variables called baseline in a data table are not necessarily pre-exposure in the causal process.

An EHR may store diagnoses entered after treatment but backdated, laboratory results collected before treatment but recorded later, comorbidity codes accumulated over years, or utilization variables spanning pre- and post-exposure periods.

A causal analysis needs time semantics, not only field names.

#Routine data make design mistakes easy to scale

Large datasets reduce random error. They do not automatically reduce systematic error.

If an exposure is misclassified, follow-up begins at the wrong time, or eligibility is defined using future information, millions of records can estimate the wrong target very precisely.

Hernán and colleagues have shown how specifying a target trial can prevent immortal-time bias and other design-induced errors in observational analyses.^[5]

#Target-trial thinking

A useful discipline is to write the hypothetical trial protocol:

1. eligibility criteria;
2. treatment strategies;
3. assignment procedure;
4. start and end of follow-up;
5. outcome;
6. causal contrast;
7. analysis plan.

Observational data then emulate these components as far as the data and design allow.^[1][2]

This does not magically make observational data randomized. It makes deviations from the idealized experiment visible.

#A synthetic example: treatment and survival

Suppose a database analysis compares mortality among people who received treatment within 30 days with those who did not.

A patient must survive long enough to receive treatment.

If follow-up starts at diagnosis and people are classified according to what happens in the next 30 days, treated patients necessarily contribute a period during which they must remain alive in order to later be classified as treated.

That immortal period is a design error.

Adding age, sex, comorbidities and fifty laboratory variables does not repair it. The time-zero problem must be fixed.

#Reporting quality is not causal validity

STROBE asks observational studies to describe potential confounders and explain how they were addressed.^[6]

That transparency is important. But STROBE is a reporting guideline. A complete checklist does not establish that the chosen adjustment set is causally correct.

Likewise, a model diagnostics report can show that an algorithm converged and standard errors were estimated correctly while the causal estimand remains unidentified.

#What adjustment can do

When the design and assumptions are appropriate, adjustment can help estimate a causal contrast by controlling measured confounding.

Methods include standardization, regression, inverse-probability weighting, matching, propensity-score approaches and g-methods.

The method should follow the data structure and estimand. None creates information about an unmeasured confounder simply by being sophisticated.

#Five questions before interpreting an adjusted estimate

First, what is the causal contrast? Second, is time zero aligned across eligibility, exposure assignment and follow-up? Third, why is each adjusted variable included? Fourth, what important causes are unmeasured or poorly measured? Fifth, what selection produced the database, cohort and analysis sample?

If those questions are unanswered, adjusted should not be used as shorthand for causal.

#A disciplined conclusion

Prefer a statement that the exposure was associated with lower outcome risk after adjustment for prespecified baseline covariates, while stating the assumptions required for a causal interpretation.

Avoid saying that the adjusted model proved the exposure reduced risk.

The difference is not timid writing. It is correct claim calibration.

#Measurement error can survive perfect adjustment

Suppose an analyst has correctly identified the confounders but the exposure is measured poorly.

If vaccinated people are misclassified as unvaccinated because outside-provider records are missing, adjustment for age, sex and comorbidity does not restore the missing exposure information. Measurement error and misclassification can materially bias epidemiological estimates and require their own assumptions and, where possible, validation information or sensitivity methods.^[7]

Likewise, an outcome based on a billing code can differ from the clinical event the study claims to measure.

Causal inference therefore requires both structural assumptions and measurement assumptions.

The model conditions on the variables it receives. It does not know whether those variables mean what the researcher thinks they mean.

#Complete-case analysis can change the target population

Another common trap is to fit an adjusted model only among records with complete covariate data.

If missingness is related to health status, care setting, exposure or outcome, the complete-case sample can be a selected population. Complete-case analysis is valid only under assumptions about the missingness process; alternative approaches such as multiple imputation also require explicit assumptions and careful implementation.^[8]

The resulting estimate may no longer describe the original eligible cohort.

Calling the model adjusted does not make that selection disappear.

The analysis should document how missing data arose, what assumptions support the chosen strategy, and whether the estimand changes after restriction.

#More variables can mean less identification

It is tempting to believe that a larger covariate set is always safer.

In causal work, the opposite can happen.

Adding post-exposure variables can block causal pathways. Adding colliders can create associations. Adding variables on causal pathways can change the estimand, while unnecessary adjustment can increase bias or variance depending on the structure.^[4][3]

The objective is not maximum adjustment. It is an adjustment strategy justified by a causal model and the target estimand.

#Limitations and uncertainty

Target-trial emulation is a framework for improving causal design with observational data; it does not guarantee identification or remove unmeasured confounding.^[2] DAG-based reasoning and target-trial thinking depend on substantive knowledge and assumptions that may be contested. STROBE improves reporting transparency but should not be interpreted as a quality score or proof of causal validity.^[6] The examples are synthetic.