A clinical trial reports an average treatment effect. A guideline committee turns that average into a rule. Audit data check whether outcomes improved. Each step is methodologically defensible. None of them, on its own, answers the question a clinician faces with a specific patient: what works for this one. The chain treats “treatment X reduces outcome Y” as a population fact and applies it to the next patient — without naming the structural assumption that lets it.
Each case below shows the gap between an average effect and the individual-patient decision it was supposed to inform. Model files ship with every case.
▾Hide / show all
▸Statins & Hospitalisation — The correct average applied to the wrong individual.
A policy that works on average is applied uniformly — including to patients for whom it is harmful, neutral, or unnecessary. Population statistics used to make individual decisions.
▸Iatrogenic Medications — The statin isn’t fighting the disease. In many cases, it’s fighting another drug.
An audit attributed a large share of adverse events to underlying conditions. The causal model found they were drug-drug interactions in the prescribing cascade. The reserve was overstated; the clinical liability was understated.
▸Adverse Event Attribution — The NSAID raises AKI risk in the population. Did it cause this patient’s AKI?
A pharmacovigilance signal is a population-level association. The clinical and legal question is whether this exposure caused this harm. Two different questions, two different formal structures.
▸Sepsis Dynamic Treatment Regimes — The vasopressor was given to the patients who needed it — so observational data shows it kills.
Treatment confounded by indication is the classic Rung 1 trap. The covariates that selected for treatment also selected for outcome; standard adjustment cannot reach the counterfactual. A dynamic regime model can.
▸Immunotherapy vs Chemotherapy — The biomarkers that select the patient also select the response.
Selection effects make immunotherapy look better than it is in head-to-head observational comparisons. The model separates selection from effect — the patients who would have responded to chemo, do.
▸Drug Repurposing — Transportability — The trial population is not the target population.
A drug’s effect estimated in one population does not transport unaltered to another. The transportability formula closes the gap when the causal structure is known and the differences are characterized — and is silent on what to do when they aren’t.
▸Treatment-Resistant Depression — After two failed antidepressants, “what works on average” is no longer the right question.
The patient has demonstrated what doesn’t work for them. The next decision is a Rung 3 counterfactual: which option, on this individual, given their non-response pattern. Population statistics applied to a non-representative individual.
Next
For the methods behind these cases, see Causal Modeling, Pearl’s Ladder (population vs. individual reasoning), and Medicine (the cluster taxonomy with sub-branches). For the wider portfolio across all five risk types, see About Risk.