Alternatives to RCTs: 9 Clinical Trial Designs (2026)

By Dr. Frank Yap, MD · One Day Media Network · Originally published June 2025 · Updated October 3, 2026

Medical disclaimer: This article explains research methods for educational purposes. It is not medical advice, and no treatment mentioned here is recommended or endorsed. Talk to a qualified clinician before changing any treatment. 

Quick Answer

Randomized controlled trials (RCTs) remain the strongest design for testing whether a treatment works, but they are costly and slow, especially for off-patent drugs with no commercial sponsor. Nine alternatives or complements exist: pragmatic, platform, N-of-1, synthetic-control, real-world-evidence, case-series, Bayesian adaptive, digital, and in silico designs. Each answers a different question, and none is a free substitute for randomization.

This article began as a June 2025 conversation with ChatGPT (shared transcript). For the October 2026 update, the key figures, dates, and citations were re-checked against primary sources, corrected where needed, and extended with grading of the evidence behind each design.

Table of Contents

Why are RCTs hard for low-cost and repurposed therapies?
How to read the evidence tiers (CEBM)
Comparison table: 9 trial designs
1. Pragmatic clinical trials
2. Adaptive platform trials
3. N-of-1 trials
4. Synthetic control arms
5. Real-world evidence and target trial emulation
6. Case series and retrospective case reviews
7. Bayesian adaptive trials
8. Digital and decentralized trials
9. In silico (computer-simulated) trials
A staged strategy: from case series to randomized proof
Case report: the first insulin treatment in 1922
Do observational studies agree with RCTs?
Reported barriers to repurposed-drug trials
How to use this article with Claude, ChatGPT, Gemini and Perplexity
Frequently asked questions
References

Why are RCTs hard for low-cost and repurposed therapies?

Short answer: RCTs are expensive and slow, and an off-patent drug has no company with a financial reason to fund one. That is a funding and design problem, not proof that randomization cannot work.

The cost is not fixed, though. The UK RECOVERY trial randomized 2,104 hospitalized COVID-19 patients to dexamethasone 6 mg daily and 4,321 to usual care. Among ventilated patients, 28-day mortality was 29.3% versus 41.4% (rate ratio 0.64, 95% CI 0.51 to 0.81) [3]. A cheap, widely available drug was tested at scale inside one shared protocol.

The same randomized platforms also tested ivermectin for COVID-19 outpatients and did not find a meaningful benefit [4] [5]. That result is the reason this article treats the alternatives below as complements that generate and sharpen hypotheses, not as shortcuts that replace controlled testing.

How to read the evidence tiers (CEBM)

Each design below carries an evidence tier based on the Oxford CEBM (OCEBM 2011) levels of evidence for treatment benefits. The tier shows how much weight a typical result can bear. It is not a verdict on any individual study.

CEBM level	Typical source (treatment benefit)	How much weight it carries
Level 1	Systematic review of randomized trials, or of N-of-1 trials	Strongest basis for a treatment decision
Level 2	Randomized trial, or observational study with a dramatic effect	Strong; check size, bias, and applicability
Level 3	Non-randomized controlled cohort or follow-up study	Moderate; confounding is the main risk
Level 4	Case series, case-control, or historically controlled study	Weak; mainly generates hypotheses
Level 5	Mechanism-based reasoning, expert opinion, simulation	Weakest; a starting point for testing

The mapping of each design to a tier in this article is an editorial judgment, not an official OCEBM classification. Real studies can sit higher or lower depending on how well they were run.

Comparison table: 9 trial designs

Design	Best for	Main limitation	Typical tier
1. Pragmatic trial	Real-world effectiveness of licensed drugs	Still needs enrollment and follow-up funding	Level 2
2. Adaptive platform trial	Testing many therapies under one protocol and shared controls	Needs central infrastructure and funding	Level 2
3. N-of-1 trial	Chronic stable conditions with symptom-type outcomes	Unsuitable for progressive disease or survival endpoints	Level 1 (series of N-of-1)
4. Synthetic control arm	Rare diseases where placebo is hard to justify	Era and population differences bias comparisons	Level 3 to 4
5. Real-world evidence / target trial emulation	Long-term safety and effectiveness at scale	Confounding by indication; data gaps	Level 3
6. Case series	Rare outcomes and early hypotheses	No denominator; selection and reporting bias	Level 4
7. Bayesian adaptive trial	Efficient learning with interim looks and borrowing	Complex design; regulators want pre-specification	Level 2 if randomized
8. Digital / decentralized trial	Lowering burden and widening access	Self-reported data; uneven adherence	Depends on design
9. In silico simulation	Designing trials and exploring subgroups	Only as good as the input data and model	Level 5 (hypothesis)

1. Pragmatic clinical trials (PCTs)

CEBM Level 2Randomized

A pragmatic clinical trial tests whether a treatment works under ordinary care conditions, with broad eligibility and few extra visits. It keeps randomization but cuts cost by using routine records for outcomes.

• What it is: Randomized comparison run in real clinics, with broad inclusion criteria and simple protocols.
• Advantage: More generalizable than a tightly controlled efficacy trial, and usually cheaper per patient.
• Main limitation: Lower internal control (adherence, blinding) can dilute or blur effects.
• Best use case: Licensed, off-patent drugs where the question is whether they help in everyday practice. Whether a given repurposed agent works is a separate evidence question; see the site's case-study pages for the reported experience (Level 4).

2. Adaptive platform trials

CEBM Level 2Randomized

An adaptive platform trial tests several treatments under one master protocol, shares control patients, and adds or drops arms as results arrive. RECOVERY, PRINCIPLE, and ACTIV-6 are examples that evaluated inexpensive repurposed drugs.

• What it is: A standing trial infrastructure with a shared control group and pre-specified rules for adding and stopping arms.
• Advantage: Spreads fixed costs across many questions; RECOVERY identified dexamethasone benefit and also found other candidates ineffective.
• Main limitation: Requires sustained funding, central coordination, and a recruiting network.
• Best use case: Communities of researchers who want randomized answers on many low-cost drugs without running a separate trial for each.

3. N-of-1 trials

CEBM Level 1 (as a series)Randomized crossover

An N-of-1 trial randomizes one patient through repeated, blinded treatment periods to find out which option works best for that person. It fits stable chronic conditions with quickly measurable symptoms, not progressive disease.

• What it is: A single patient receives different therapies in random order across several periods, ideally blinded, with the same outcome measured each time.
• Advantage: Fast and cheap per decision; gives individualized answers when average trial results do not apply.
• Main limitation: Needs a stable condition, a treatment that starts and stops working quickly, and a short washout. Progressive cancer and survival endpoints generally fail these conditions.
• Best use case: Symptom-type questions such as pain, sleep, blood pressure, or digestive symptoms.

Correction in this update: the original version suggested N-of-1 trials for patients unresponsive to standard cancer therapy. A randomized crossover is rarely feasible there; single-patient cancer reports are better classed as case-level evidence (Level 4). See also our N-of-1 example and The Crisis in Evidence-Based Medicine.

4. Synthetic control arms

CEBM Level 3 to 4External controls

A synthetic control arm replaces or supplements the randomized control group with data from earlier trials or health records. It can reduce how many participants receive placebo, but differences between eras and populations can bias the comparison.

• What it is: External or historical data used as the comparator for patients receiving the new treatment.
• Advantage: Saves time and sample size and eases ethical concerns about placebo in serious or rare disease.
• Main limitation: Confounding from differences in patient mix, standard of care, and how outcomes were measured.
• Best use case: Rare diseases and some oncology settings where standard-of-care outcomes are well documented. The FDA's January 2026 Bayesian draft guidance discusses augmenting concurrent controls with external or nonconcurrent data [6].

5. Real-world evidence (RWE) and target trial emulation

CEBM Level 3Observational

Real-world evidence uses electronic records, registries, and claims data to study treatments in everyday practice. Done well, it first writes down the randomized trial it is trying to imitate (a target trial) and then emulates that protocol with observational data.

• What it is: Retrospective or prospective cohorts built from routine care data, analyzed with explicit eligibility, treatment strategies, and follow-up rules.
• Advantage: Large, inexpensive datasets that reflect real practice, useful for long-term safety and rare events.
• Main limitation: Confounding by indication, missing data, and design errors such as immortal-time bias.
• Best use case: Questions an RCT cannot answer soon enough or cheaply enough. Target trial emulation provides a structured way to criticize such studies [7].

6. Case series and retrospective case reviews

CEBM Level 4Hypothesis-generating

A case series compiles outcomes from several patients to flag possible effects or safety signals. It is fast and cheap but has no control group and no denominator, so it can suggest a hypothesis and cannot prove one.

• What it is: A structured review of existing clinical cases, ideally with consistent fields for diagnosis, prior treatment, regimen, follow-up, and outcome.
• Advantage: Quick to assemble; good for rare outcomes and for deciding what deserves formal testing.
• Main limitation: Selection and reporting bias: people with good outcomes are more likely to be reported, and failures are often missing.
• Best use case: Early evaluation of repurposed or natural treatments, as long as the write-up states clearly that the evidence tier is low. See the site's case-series compilation for an example organized this way.

7. Bayesian adaptive trials

CEBM Level 2 if randomizedAdaptive

Bayesian adaptive trials update the probability that a treatment works as data accumulate, allowing early stopping or dropping weak arms. In January 2026 the FDA published draft guidance on using Bayesian methods in drug and biologic trials.

• What it is: A pre-specified statistical design in which interim results change enrollment or stopping decisions.
• Advantage: More efficient use of data, and a natural way to borrow information from earlier studies.
• Main limitation: Complex to design and simulate; regulators expect pre-specified operating characteristics, and the guidance is still a draft (comment period closed March 13, 2026).
• Best use case: Small oncology programs and rare diseases testing several arms with limited patients. Bayesian designs can be randomized, so this is an efficiency tool, not a way to skip controls.

8. Digital and decentralized trials

CEBM Level depends on designRemote

Digital and decentralized trials use apps, wearables, telemedicine, and home delivery so participants do not need frequent clinic visits. They can be randomized and blinded: ACTIV-6 ran as a decentralized, double-blind, placebo-controlled platform trial.

• What it is: Remote enrollment, medication delivery, and patient-reported or device-collected outcomes.
• Advantage: Lower infrastructure cost, wider geographic reach, and potentially better retention.
• Main limitation: Self-reported outcomes, uneven digital access, and data-quality checks are harder.
• Best use case: Lifestyle, dietary, and symptom-based interventions; long-term safety follow-up.

9. In silico (computer-simulated) trials

Hypothesis-generating (about Level 5)Model-based

In silico trials use real-world data and models to simulate trial arms, test protocol choices, and explore subgroups. They add depth to RCT findings but do not replace them, because a model is only as reliable as the data and assumptions behind it.

• What it is: Computer simulation of trial protocols on large medical-record datasets or mechanistic models.
• Advantage: Nearly free to repeat across scenarios; can show how results change in more diverse populations.
• Main limitation: Garbage in, garbage out: biases in the source data carry into the simulation.
• Best use case: Designing better RCTs and stress-testing results. In one published example, a donepezil simulation matching the original trial population gave a serious adverse event rate of 8.9% (trial: 8.3%), while a more diverse population gave 15.5% [9]. See also AI Simulations Reveal Treatment Synergies.

A staged strategy: from case series to randomized proof

Short answer: for controversial or off-label interventions, build evidence in stages and decide in advance what result would stop or advance the work. Each stage answers a narrow question and is cheaper than the next.

Stage	Design	Question answered	Tier	Advance only if
1	Structured case series	Is there a consistent signal worth testing?	Level 4	Cases are complete, denominators are reported, and failures are included
2	Matched cohort or target trial emulation	Does the signal persist against a comparison group?	Level 3	Effect survives adjustment and sensitivity analyses
3	Pragmatic or platform RCT (Bayesian where suitable)	Does it work compared with usual care or placebo?	Level 2	Pre-registered outcome is met with acceptable safety
4	Replication and systematic review	Is the result consistent across settings?	Level 1	Independent trials agree

A pre-registered protocol and stop rules matter at every stage. Without them, it is easy to keep going on a weak signal because early results looked promising.

Case report: the first insulin treatment in 1922

Short answer: insulin's first human use was a single-patient report, not an RCT, and it still changed medicine. That works only because the effect was dramatic against a very predictable prognosis.

On January 11, 1922, 14-year-old Leonard Thompson, close to death from type 1 diabetes at Toronto General Hospital, received the first pancreatic extract prepared by Banting and Best. His blood sugar fell somewhat, but the contaminated extract caused a sterile abscess and the injections were stopped. After James Collip substantially purified the extract, Thompson received it on January 23, 1922; his blood glucose fell to normal within a day and the ketones disappeared. He lived 13 more years on insulin [10].

The lesson is not “case reports versus RCTs.” It is that different evidence answers different questions. Glasziou and colleagues proposed a signal-to-noise rule: when the treatment effect is very large relative to a predictable prognosis (a rate ratio often above 10), bias is unlikely to explain it, and randomization is less necessary [8]. The same authors list hormone replacement therapy and beta-carotene as warnings where non-randomized evidence misled because the effects were modest. Most repurposed-drug signals look more like the second group than the first.

In evidence terms, an early case report is hypothesis-generating (Level 4), not proof of efficacy, and that does not make it unimportant.

Do observational studies agree with RCTs?

Short answer: on average, closely, but not on every question. A 2024 Cochrane meta-epidemiological review pooled 34 of 47 reviews and found a ratio of ratios of 1.08 (95% CI 1.01 to 1.15), with low certainty and substantial heterogeneity (I² 69%) [1].

The 2014 version of the review (Anglemyer, Horvath and Bero; not “Engelmayer,” as an earlier version of this article had it) found a pooled ratio of odds ratios of 1.08 (95% CI 0.96 to 1.22) in 14 reviews and reached a similar conclusion [2]. In the 2024 update, the pharmaceutical-only subgroup showed a slightly larger difference (1.12, 95% CI 1.04 to 1.21).

A separate analysis posted by c19early.org and relayed on Substack reports a risk ratio of 1.00 (0.92 to 1.08) between RCTs and observational studies across 102 COVID treatments (their post). We have not verified it against a peer-reviewed source, so we note it as a claim, not a finding to rely on.

Pooled averages cannot tell you whether observational data and RCTs agree for one specific drug. For ivermectin in COVID-19 outpatients, large placebo-controlled randomized trials found no meaningful benefit, so each treatment needs its own evidence check.

Reported barriers to repurposed-drug trials

Short answer: funding structure, approval rules, and timing can all keep inexpensive combinations from being tested. The account below comes from one physician-scientist and is expert opinion (Level 5).

In a recorded conversation, Dr. Robert W. Malone described trying to study combinations of approved drugs, including famotidine, celecoxib, and ivermectin, as early COVID-19 treatment. His account of the barriers:

• Approval requirements shaped what reached trials. He reported that proposals were rejected until laboratory antiviral data were supplied for ivermectin, and that ivermectin was removed from the protocol so the study could proceed.
• Delays shaped which treatments gained attention. While approvals stalled, public-health policy moved ahead, widening the gap between early ideas and formal evidence.
• Combination strategies fit poorly in single-drug frameworks. Frameworks that assess each drug separately can miss effects that depend on synergy, like testing a key and lock in separate rooms.
• Funding concentrated on some pathways. Once large trials and funding locked onto particular approaches, alternatives received fewer resources and slower evidence accumulation.

These are one participant's recollections and interpretations, not independently verified findings. The structural point, that off-patent drugs lack a sponsor, is widely recognized. Related commentary: Why You Saw Some COVID Treatments, and Not Others (Level 5) and Paul Marik on RCTs and repurposed drugs (Level 5).

How to use this article with Claude, ChatGPT, Gemini and Perplexity

Short answer: paste the question you are trying to answer, plus the comparison table above, and ask an AI assistant to match the question to a design, name the main bias, and state the evidence tier. Always check any cited study yourself.

Assistant	Useful for	Example prompt
Claude	Critiquing a study plan and spotting bias	“Here is my question and data source. Which of the nine designs fits, and what are the three biggest threats to validity?”
ChatGPT	Drafting a protocol outline or a plain-language summary	“Draft a one-page outline of a retrospective case series with fields for diagnosis, prior treatment, regimen, follow-up, and outcome.”
Gemini	Comparing designs across sources you supply	“Using the table I pasted, explain when a synthetic control arm is a weaker choice than a platform trial.”
Perplexity	Finding and listing primary sources	“Find peer-reviewed trials or reviews on [treatment] and label each by design and CEBM level, with links.”

Do not enter identifiable health information. AI tools can misstate numbers or invent citations, and none of them replaces a clinician or a biostatistician. Treat their output as a draft to verify, not as medical advice.

Frequently asked questions

Are randomized controlled trials always the best evidence?

RCTs are the strongest design for estimating whether a treatment causes an outcome, because randomization balances known and unknown confounders. They are not always feasible, ethical, or affordable. Other designs can answer different questions, such as long-term safety, rare outcomes, or real-world effectiveness, but they carry more risk of bias and should be graded accordingly.

What is the cheapest alternative to a full RCT?

Retrospective case series and real-world-data cohort studies are usually cheapest, because they reuse existing records. They are also the most vulnerable to selection bias and confounding. A pragmatic or platform RCT that reuses routine clinical data and shared control arms costs more but keeps randomization, so it can support much stronger conclusions.

Can a case report prove that a treatment works?

No. A case report or case series is hypothesis-generating evidence, typically CEBM Level 4. The exception is a very dramatic effect against a predictable prognosis, as with insulin in 1922, where a rate ratio above about 10 is unlikely to be explained by bias. Most repurposed-drug signals are far smaller and need controlled testing.

Do observational studies give the same answers as RCTs?

On average they come close. A 2024 Cochrane meta-epidemiological review of 47 reviews found a pooled ratio of ratios of 1.08 (95% CI 1.01 to 1.15), with low certainty and substantial heterogeneity. Averages can hide disagreement on individual questions, so each treatment still needs its own evidence check.

What is a synthetic control arm?

A synthetic control arm uses existing data, such as prior trial arms or electronic health records, in place of or alongside patients randomized to a control group. It can reduce the number of people given placebo. Differences in patient mix, care standards, and outcome measurement between eras can bias the comparison, so methods and data quality must be reported transparently.

Are N-of-1 trials suitable for cancer treatment?

Generally not for survival outcomes. N-of-1 trials need a stable, chronic condition, a treatment whose effect starts and stops quickly, and an outcome measurable within each period. Progressive cancer usually fails those conditions. N-of-1 designs fit symptom-type questions, such as pain, sleep, or blood pressure, where each patient can serve as their own control.

Has the FDA accepted Bayesian trial designs?

In January 2026 the FDA published draft guidance on Bayesian methodology for drug and biologic trials, covering its use to support primary inference. It is a draft, not final policy, and sponsors are expected to pre-specify methods and discuss protocols with the agency. Bayesian designs can still be randomized, so they reduce data needs without removing control groups.

October 2026 update note: corrected the Cochrane citation (Anglemyer et al. 2014, updated by Toews et al. 2024), the Nature-family attribution of the in silico editorial (npj Digital Medicine, 2021), and the N-of-1 use case; added platform trials, the ivermectin RCT results, the January 1922 first-dose detail, FDA Bayesian draft guidance (January 2026), evidence tiers, and an FAQ.

References

1. Toews I, Anglemyer A, Nyirenda JLZ, et al. Healthcare outcomes assessed with observational study designs compared with those assessed in randomized trials: a meta-epidemiological study. Cochrane Database Syst Rev. 2024;1:MR000034. Link
2. Anglemyer A, Horvath HT, Bero L. Healthcare outcomes assessed with observational study designs compared with those assessed in randomized trials. Cochrane Database Syst Rev. 2014;(4):MR000034. Link
3. RECOVERY Collaborative Group. Dexamethasone in hospitalized patients with Covid-19. N Engl J Med. 2021;384:693-704. Link
4. ACTIV-6 Study Group. Effect of ivermectin 600 µg/kg for 6 days vs placebo on time to sustained recovery in outpatients with mild to moderate COVID-19 (preprint of the JAMA 2023 report). Link
5. PRINCIPLE Trial Collaborative Group. Ivermectin for COVID-19 in community settings: a randomised, controlled, open-label, platform adaptive trial (open-access full text). Link
6. U.S. FDA. Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products: Draft Guidance for Industry, January 2026 (Docket FDA-2025-D-3217). Link
7. Hernán MA, Robins JM. Using big data to emulate a target trial when a randomized trial is not available. Am J Epidemiol. 2016;183(8):758-764. Link
8. Glasziou P, Chalmers I, Rawlins M, McCulloch P. When are randomised trials unnecessary? Picking signal from noise. BMJ. 2007;334:349-351. Link
9. Wedlund L, Kvedar J. Simulated trials: in silico approach adds depth and nuance to the RCT gold-standard (editorial). npj Digital Medicine. 2021;4:121. Link
10. University of Toronto Fisher Library. Leonard Thompson: first patient to receive pancreatic extract (Insulin 100 exhibit). Link
11. Oxford Centre for Evidence-Based Medicine. OCEBM Levels of Evidence (2011). Link
12. OneDayMD. AI Simulations Reveal Treatment Synergies Clinical Trials Will Never Test (2026). Link
13. OneDayMD. The Crisis in Evidence-Based Medicine: Limitations of RCTs and the Rise of Personalized N-of-1 Trials (2026). Link
14. OneDayMD. N-of-1 Trial: Harvard Med Student Eats 720 Eggs in 30 Days (2024). Link
15. SmartCancer.org. How Precision Medicine & Genomics Are Transforming Cancer Care and Prevention (2026). Link
16. Commentary (Level 5): Paul Marik. Why Randomized Controlled Trials Are Poorly Suited to Testing Repurposed Drugs and Nutraceuticals in Cancer (Substack, 2026). Link
17. Commentary (Level 5): Sid Sijbrandij's care journey: Going Founder Mode On Cancer (Century of Bio, Substack, 2026). Link
18. Commentary (Level 5): Why You Saw Some COVID Treatments, and Not Others (Mercola, 2026). Link
19. Commentary (Level 5): Massimaux. 5 years later: the most comprehensive analysis of RCTs vs observational studies (Substack), relaying a c19early.org analysis. Link
20. Origin of this article: ChatGPT conversation, June 2025 (shared transcript). Link

About the author: Dr. Frank Yap, MD, runs the One Day Media Network, publishing medical intelligence, integrative oncology, and wellness content since 2012. See About Us.