Randomized Controlled Trials

What is a randomized controlled trial?

A randomized controlled trial (RCT) is a prospective experimental study in which participants are randomly allocated to two or more groups, each receiving a different intervention (one of which may be a placebo or standard care), and outcomes are compared across groups under conditions designed to isolate the effect of the intervention.

Four features define an RCT:

Random allocation — assignment to arms is determined by chance, removing systematic differences between groups at baseline.
A comparison group — at minimum one control arm, against which the intervention's effect can be measured.
Manipulation of the intervention — the investigator controls what is given, when, to whom. (Distinguishing it from observational designs.)
Prospective follow-up — outcomes are measured after the intervention is delivered, not retrospectively reconstructed.

The modern clinical RCT traces to Sir Austin Bradford Hill's 1948 MRC trial of streptomycin for pulmonary tuberculosis — the first published trial to use concealed random allocation in human medicine. Hill's design replaced the informal "clinical impression" comparisons that had dominated for centuries and established the methodological template still used today.

When to use an RCT (and when not to)

Use an RCT when:

You need to establish causal evidence that an intervention works — not just association.
Observational evidence is contradictory or confounded (e.g., the WHI hormone-therapy reversal).
The intervention is novel and regulatory approval requires it (most drug and device approvals).
You're comparing two interventions both already in use ("comparative effectiveness").
The expected effect size is small-to-moderate, where confounding could plausibly explain observational signals.

Don't use an RCT when:

The intervention has a self-evident, dramatic effect (the famous "parachute trial" argument — though see Smith & Pell's 2003 BMJ satire).
Randomization would be unethical (giving placebo for life-threatening disease with a known effective treatment).
The condition is so rare that recruitment is infeasible (n-of-1 or basket designs may help).
The intervention is a complex system-level change where individual randomization is impractical — consider a stepped-wedge cluster RCT instead.
Generalizability concerns dominate — sometimes a pragmatic trial or large observational study is more useful than a tightly controlled RCT.

The 16 RCT subtypes

RCT Minds organizes RCT designs into four families. Picking the right one is the single most consequential decision in your protocol — most methodological failures trace back to a mismatch between design and research question.

Parallel-group

Two or more groups receive different interventions concurrently. The default RCT design.

Standard Parallel — Simple 1:1 head-to-head comparison.
Superiority — Test whether the new treatment is better than the comparator.
Equivalence — Test whether two treatments fall within a pre-specified margin of clinical sameness.
Non-Inferiority — Test whether a new treatment is no worse than the comparator (often used to justify a cheaper or safer alternative).

Factorial

Multiple interventions tested simultaneously in the same trial — efficient when interactions matter.

2×2 Factorial — Two binary interventions crossed. Four arms, four research questions.
2×3 Factorial — One binary + one three-level intervention. Six arms.
3×3 Factorial — Two three-level interventions. Nine arms — only feasible with large samples.
Fractional Factorial — Subset of all factorial combinations — sacrifices some interaction estimates for sample efficiency.

Crossover

Each participant receives every treatment in sequence, separated by washout. Powerful for chronic, stable conditions.

Two-period Crossover — AB / BA design. Most common crossover.
Three-period Crossover — Three treatments cycled — adds carryover-detection power.
Four-period Crossover — Used for complex dose-response or four-arm comparisons.
Latin Square — n×n design balancing treatment order across n participants and n periods.

Specialized

Designs for hard cases — cluster-level interventions, biomarker-driven oncology, single-patient questions.

Stepped-Wedge Cluster — Clusters cross over from control to intervention in randomized order over time — common in health-system rollouts.
Basket Trial — Single targeted therapy tested across multiple tumor types sharing a biomarker.
Umbrella Trial — Multiple therapies tested within a single tumor type, each matched to a molecular subtype.
N-of-1 — Single-patient multiple-crossover trial. Personalized evidence for stable chronic conditions.

Anatomy of a well-designed RCT

Five elements separate a publishable RCT from a methodologically vulnerable one:

1. A specific, pre-registered research question (PICO)

Population, Intervention, Comparator, Outcome — defined precisely before enrollment begins. Vague questions invite post-hoc fishing. Pre-registration in ClinicalTrials.gov, the WHO ICTRP, India's CTRI, or (for systematic reviews) PROSPERO is the default expectation of any major journal.

2. Sample size justified by power calculation

Pre-specify your minimum clinically important difference, the variance you expect in the outcome, your alpha (usually 0.05) and beta (usually 0.20 → 80% power). Underpowered trials waste participants and produce inconclusive results that pollute meta-analyses.

3. Concealed randomization with allocation concealment

Randomization assigns by chance; allocation concealment prevents the enrolling clinician from knowing the next assignment. Both are essential — Schulz et al.'s 1995 meta-analysis found that trials without adequate allocation concealment overestimated treatment effects by 41% on average. Read the full deep dive on randomization →

4. Blinding wherever feasible

Single-blind (participants), double-blind (participants + investigators), triple-blind (+ outcome assessors), or open-label. Blinding reduces performance bias, detection bias, and post-randomization differential attrition. Some interventions (surgery, behavioral therapy) cannot be blinded to participants — outcome-assessor blinding still helps.

5. A pre-specified statistical analysis plan

Primary analysis: intention-to-treat (ITT) — every participant analyzed in their assigned arm regardless of adherence. Per-protocol analysis is supplementary. Pre-specify your handling of missing data, interim analyses, multiplicity adjustment for secondary outcomes, and subgroup analyses. Deviations from the SAP after data lock require disclosure.

CONSORT 2010 — the reporting standard

The Consolidated Standards of Reporting Trials (CONSORT) is the 25-item checklist that journals use to evaluate whether an RCT is reported transparently enough for readers to judge risk of bias. It is not optional — most major journals (BMJ, Lancet, JAMA, NEJM, PLoS Medicine) require CONSORT compliance for publication.

Key items most often under-reported (per meta-research):

Item 7a: sample size calculation — how it was derived.
Item 8a/8b/9: randomization sequence generation, type, allocation concealment mechanism.
Item 11a/11b: blinding — who was blinded and how.
Item 12a: primary outcome analysis method (and changes from protocol).
Item 13a/13b: participant flow diagram with reasons for loss to follow-up.

The matched protocol standard is SPIRIT 2013 for the protocol phase, and TIDieR for intervention description so others can replicate it.

The pitfalls that invalidate trials

Underpowered design.Calculating sample size to detect an implausibly large effect, then publishing the "negative" result as evidence of no effect. A null finding with wide confidence intervals tells you nothing.

Pseudo-randomization. Alternation by day of week, hospital number, or odd/even ID is not random and is detectable in meta-analyses as a marker for inflated effect estimates.

Selective outcome reporting. The primary outcome in the protocol becomes a secondary outcome in the paper because the result was unfavorable, and a secondary becomes the headline. Cross-check protocol pre-registration vs. publication.

Loss to follow-up >20%.Cochrane risk-of-bias tool downgrades any trial losing >20% of participants without convincing imputation. Differential loss between arms is worse.

Spin in conclusions.A non-significant primary outcome described as showing a "trend toward benefit" in the abstract. Reviewers and readers should compare the abstract conclusion to the actual point estimate and confidence interval.

Methodology deep dives

8 min read

Randomization in clinical trials

Methods (simple, block, stratified, cluster, minimization), allocation concealment, common pitfalls, CONSORT reporting.

More deep dives coming — sample size calculation, blinding strategies, PICO framework, statistical analysis plans, intention-to-treat vs per-protocol.

References

Medical Research Council. Streptomycin treatment of pulmonary tuberculosis. BMJ 1948;2(4582):769–82. (The first modern RCT.)
Schulz KF, Chalmers I, Hayes RJ, Altman DG. Empirical evidence of bias. JAMA 1995;273(5):408–12.
Moher D, Hopewell S, Schulz KF, et al. CONSORT 2010 Explanation and Elaboration: updated guidelines for reporting parallel group randomised trials. BMJ 2010;340:c869.
Chan A-W, Tetzlaff JM, Altman DG, et al. SPIRIT 2013 Statement: defining standard protocol items for clinical trials. Ann Intern Med 2013;158(3):200–7.
Hoffmann TC, Glasziou PP, Boutron I, et al. Better reporting of interventions: TIDieR checklist. BMJ 2014;348:g1687.
ICH Harmonised Tripartite Guideline. Statistical Principles for Clinical Trials (E9). International Conference on Harmonisation, 1998.
Smith GCS, Pell JP. Parachute use to prevent death and major trauma related to gravitational challenge: systematic review of randomised controlled trials. BMJ 2003;327(7429):1459–61.
Writing Group for the Women's Health Initiative Investigators. Risks and benefits of estrogen plus progestin in healthy postmenopausal women. JAMA 2002;288(3):321–33.