Multivariable Analysis

A regression with multiple predictors, used to adjust for confounders and estimate each variable’s independent association with a single outcome. (Strictly, multivariable = one outcome, many predictors; multivariate = multiple outcomes — the terms are widely conflated.) The output is a set of adjusted estimates — each holding the others constant.

Several predictors entering a single regression together. Instead of one crude association per variable, the model returns an adjusted effect for each — the association that remains once the other measured variables are accounted for.

An adjusted OR/HR/β is the effect of one predictor with the others held constant — the model’s attempt to remove confounding. Compare it with the crude (unadjusted) estimate: a big shift after adjustment means a measured confounder was distorting the crude figure. What can’t shift are variables nobody measured.

Observational EM data are riddled with confounding; multivariable analysis is the main statistical defence. But it adjusts only for what was measured and modelled correctly, so “adjusted” is not a synonym for “causal” — reading which variables went into the model is essential to judging the claim.

• Multivariable = 1 outcome, many predictors
• Multivariate = many outcomes (often conflated)
• Rule of thumb: ≈10 events per variable

Cohort study adjusted estimates — a typical ED prognosis paper reports a crude odds ratio for, say, a treatment, then a multivariable-adjusted OR after entering age, sex, comorbidity and severity. The headline is usually the adjusted figure — “independent” of those confounders. When the adjusted OR moves sharply toward 1 from the crude value, confounding was inflating the crude association. “Adjusted” only covers measured confounders — an unmeasured one (e.g. frailty never recorded) can still drive the whole effect, so an adjusted association never proves causation.

• Confusing multivariable (one outcome, many predictors) with multivariate (many outcomes).
• Overfitting — too many predictors for the number of outcome events (breach of the ~10 EPV rule).
• Adjusting for a mediator on the causal pathway — this can mask a real effect rather than remove bias.