Adjustment Methods

A family of analysis tools used to control confounding when you cannot randomise. The main ones are multivariable regression (put the confounders in the model alongside the exposure), propensity scores (model each patient’s probability of receiving the treatment from their measured covariates, then match, stratify or weight on that single score), and inverse-probability-of-treatment weighting (IPTW) (up-weight patients who got the “unexpected” treatment to build a pseudo-population in which treatment is independent of the covariates). Every one of them adjusts for measured confounders only.

The measured covariates are condensed into a single propensity score — one number summarising how likely each patient was to get the treatment. That score is then spent in one of three ways (matching, stratification, weighting), all converging on the same goal: groups that look alike on the things you measured. The dashed callout is the whole exam point — anything off the diagram is never balanced.

Read it as a funnel, not magic. A propensity score does not add information; it is just a tidy repackaging of the same measured covariates a regression would use — it collapses many variables into one, which helps when events are few. Check the paper reports covariate balance after adjustment (standardised mean differences) and adequate overlap between groups; if treated and untreated patients barely overlap, the method is extrapolating into a region with no real comparators.

Propensity methods look more sophisticated than regression and are often presented as if they approximate a randomised trial. They do not. They balance only the measured confounders, exactly like regression, so a “propensity-matched” observational result is still observational — vulnerable to the unmeasured factor that drove both the treatment decision and the outcome. Treat any causal claim from these methods with the same caution you give any non-randomised comparison.

• Regression = confounders in the outcome model
• Propensity score = P(treatment | measured covariates), then match / stratify / weight
• IPTW = re-weight to a pseudo-population free of measured confounding
• All adjust for measured confounders only — never unmeasured

Propensity-matched registry studies — a recurring pattern in EM and critical care: a large database (e.g. a trauma or resuscitation registry) compares an intervention (say, prehospital advanced airway management, or early vasopressors) against usual care, using propensity matching to balance age, injury severity, comorbidity and physiology. The matched analysis reports an apparent mortality difference. When the same question is later tested in a randomised trial, the effect often shrinks or vanishes — because the sickest patients were selected for (or against) the intervention in ways no recorded variable fully captured. Indication by severity is the classic unmeasured confounder: clinicians choose aggressive treatment because a patient looks sick, and no propensity model holds the bedside gestalt that drove the decision.

• Claiming propensity matching “mimics randomisation” — it balances only measured covariates, so unmeasured confounding persists.
• Over-confidence in fancy methods — IPTW and propensity scores are no more causal than the regression they replace.
• Ignoring overlap / positivity — if treated and untreated groups barely share covariate space, the estimate is built on extrapolation.