Strengths and weaknesses of different epidemiological study designs - randomised controlled trial

Why randomisation matters

Randomisation is the only design feature that controls for UNMEASURED confounders - this is the entire reason RCTs sit at the top of the hierarchy

Expensive, slow, and often underpowered for rare or long-latency outcomes
External validity: highly selected populations, expert centres, run-in periods -> efficacy, not effectiveness
Unethical or impractical for harms (smoking, asbestos) and for rare exposures
Vulnerable to attrition, crossover, and post-randomisation bias
Composite endpoints and surrogate outcomes can conceal a null effect on what matters

Design	Use
Parallel group	Standard
Crossover	Chronic stable disease; each patient is their own control. Needs a washout; carryover effect is the risk
Cluster	Randomise practices/wards - for interventions that cannot be individually allocated. Requires design-effect inflation of sample size
Factorial	Two questions in one trial. Assumes no interaction
Non-inferiority	New agent is cheaper/safer. Requires a pre-specified margin; ITT is anti-conservative here, so per-protocol matters
Adaptive / platform	Interim-driven changes; efficient in pandemics (RECOVERY)

Only ~10-20% of clinical practice recommendations are supported by high-quality RCT evidence - the rest rests on observational data and consensus

Random allocation makes treatment assignment independent of every baseline characteristic, measured or not
- -> any baseline imbalance is due to chance alone
- -> hence formal significance testing of a baseline characteristics table is meaningless and should not be done
Allocation concealment protects randomisation at the point of enrolment (see separate note)
Blinding protects it afterwards

Intention to treat: analyse every patient in the group to which they were randomised, regardless of what they received
- Preserves randomisation, prevents post-randomisation selection bias
- Conservative for superiority trials; anti-conservative for non-inferiority trials (non-adherence pushes arms together, favouring "non-inferior")
Per-protocol analysis breaks randomisation and is inherently observational - report both

Effect estimate + 95% confidence interval, not the p value alone
Absolute risk reduction and NNT, not only the relative risk reduction
Pre-specified primary endpoint vs the endpoint reported in the abstract - check they are the same
Fragility index - how many events would need to change to lose significance

The vertical distance between the curves at a given time point is the absolute difference in survival probability at that time
The point where the curves begin to diverge indicates when the treatment effect starts to manifest
- Late divergence: a delayed mechanism (immunotherapy, disease modification). Early divergence with later convergence: a transient effect
Number at risk beneath the x-axis - the tail of a KM curve is built from very few patients and is unstable
Interpret with a formal statistical comparison (log-rank test) and an effect estimate such as a hazard ratio with confidence interval - never by visual inspection alone
- A hazard ratio assumes proportional hazards. Crossing curves violate this and make the single HR misleading

Randomised and concealed? Blinded to whom?
Groups similar at baseline?
Complete follow-up? (>20% loss is a serious threat)
Analysed by intention to treat?
Was the trial stopped early for benefit? (early stopping systematically overestimates effect size)
Is my patient like the trial population?

CONSORT - reporting standard for RCTs, with the participant flow diagram
Cochrane Risk of Bias 2 (RoB 2) - five domains: randomisation process, deviations from intended intervention, missing outcome data, outcome measurement, selective reporting
GRADE - rates certainty of evidence; RCTs start high and are downgraded for risk of bias, inconsistency, indirectness, imprecision and publication bias
- Observational studies start low but can be upgraded for large effect, dose-response, or when all plausible confounding would reduce the effect
Prospective registration (ANZCTR, ClinicalTrials.gov) before enrolment - the main defence against outcome switching

A statistically significant result in a huge trial may be clinically trivial - always look at the absolute difference
A non-significant result is not evidence of no effect - look at the confidence interval width
Subgroup analyses are hypothesis-generating; with 20 subgroups, one will be "significant" by chance
Composite endpoints are usually driven by the softest and commonest component (often hospitalisation, not death)
Surrogate endpoints have repeatedly misled - CAST (antiarrhythmics suppressed ectopy and increased mortality), and the fate of many lipid surrogates
Industry-funded trials report favourable results more often - not usually through fraud, but through comparator choice, dose, and outcome selection
Trials stopped early for benefit overestimate the treatment effect, often substantially

13 more sections, plus exam facts

Premium unlocks every note across every specialty, and the full exam fact library behind it.