What Randomized Controlled Trials Prove and What They Do Not

A randomized controlled trial, usually shortened to RCT, is the most rigorous design for testing whether a treatment or intervention causes an effect. Participants are randomly assigned to receive the intervention or a comparison condition, and outcomes are measured afterward. Because chance decides who gets what, the groups should differ only in the intervention on average. That single feature gives RCTs a strength that observational studies rarely match.

This guide explains how RCTs are designed, why random assignment and blinding matter, what the results can and cannot tell us, and why even a well-run trial is one piece of evidence rather than a final answer.

Why randomization is the key step

Suppose a researcher wants to test whether a study technique improves exam scores. If students choose the technique themselves, the group that chooses it may already be more motivated, more organized or better prepared. Any difference in scores could reflect those traits rather than the technique. This is called selection bias.

Random assignment solves this problem by letting chance divide the sample. With enough participants, the groups should be similar on both measured and unmeasured characteristics, such as age, prior knowledge and personality. Any systematic difference that remains is more likely to come from the intervention.

Randomization works on average. In a small study, chance can still produce imbalanced groups, which is one reason sample size matters.

Control conditions

An RCT needs a meaningful comparison. The control condition might receive no treatment, a placebo, a standard treatment or an alternative intervention. The right choice depends on the question. A placebo control tests whether the active ingredient or technique does more than expectation and attention. A standard-care control tests whether a new option is better than what people already receive.

Poorly chosen controls can mislead. If the control group receives nothing at all, a trial may show that a treatment beats doing nothing without showing that it beats an existing option.

Blinding and expectation

Blinding means keeping participants, clinicians or outcome assessors unaware of who received which condition. Participant blinding reduces the influence of expectation. In psychology, expectation is a powerful force. People who believe they are receiving an effective treatment may report improvement even when the active component is absent.

Assessor blinding reduces bias in how outcomes are measured. A rater who knows the condition may score ambiguous answers differently. In many psychological trials, full blinding is hard, because a therapy is difficult to disguise. Researchers then use other safeguards, such as independent raters or standardized measures.

Reading results: effect size and uncertainty

A positive trial result is not a yes or no answer. Researchers report effect sizes, which describe how large the difference between groups is, and confidence intervals, which show the range of plausible values for that difference. A statistically significant result can still have a small effect, and a non-significant result can still be compatible with a meaningful effect if the trial was too small to detect it.

The table below shows how the same statistical significance can mean different things depending on the size of the effect and the precision of the estimate.

Trial feature	Strong reading	Weak reading
Effect size	Large and clinically relevant	Tiny, barely above noise
Confidence interval	Narrow and far from zero	Wide, crossing zero
Sample size	Large, adequately powered	Small, prone to chance results
Preregistration	Outcomes declared in advance	Outcomes chosen after data

The replication problem

Psychology has faced a replication crisis. Large projects that attempted to repeat published studies found that a substantial share of original results did not replicate with similar strength. Causes include small samples, flexible analysis choices, publication bias that favors positive results and incentives that reward novelty.

Preregistration, larger samples and open data have become standard responses. Registering a trial's hypotheses and analysis plan before data collection makes it harder to change the question after seeing the results. The same reforms apply to randomized trials, which are not immune to bias just because they use randomization.

What one trial cannot prove

A single RCT rarely settles a question. Participants in a trial are often selected for the study, which may limit how well results apply to people in ordinary clinical or everyday settings. Trials also usually run for a fixed period, so they may not show whether effects last or whether harms appear later.

Replication across independent teams, populations and settings matters. Systematic reviews and meta-analyses combine many trials to estimate an overall effect and to examine whether results depend on context. Guidelines from health bodies usually rest on this kind of body of evidence, not one study.

Limits of the design

Some questions cannot be randomized for practical or ethical reasons. Researchers cannot randomly assign people to smoking, to childhood trauma or to a severe disease. In those cases, observational designs, natural experiments and quasi-experimental methods are used, and they require more caution when making causal claims. The strength of the design should match the strength of the conclusion.

Randomization also tests a specific intervention in a specific setting. A program that works in one school may fail in another if local conditions differ. Good trial reports describe participants, procedures and outcomes in enough detail for readers to judge whether results transfer.

How to read a trial headline

When you see a claim that a study shows a treatment works, ask a few questions. Was assignment randomized and was the control meaningful? Were participants or assessors blinded where possible? How large is the effect and how precise is the estimate? Were the outcomes and analysis plan declared in advance? Has the finding been replicated?

These questions do not require a statistics degree. They separate a claim that has been tested carefully from one that has only been reported.

Intention to treat and dropouts

Real trials lose participants. Some people move away, decline follow-up or stop the treatment early. The standard analytic approach, called intention to treat, analyzes everyone according to the group they were originally randomized to, whether or not they completed the protocol. This keeps the groups comparable, because dropping people who did poorly would make the treatment look better than it is.

Dropout patterns themselves are informative. If many people leave one arm of a trial, the reason matters. High dropout due to side effects or burden is a result in its own right, and it tells readers something about how the treatment works in practice.

Trial registries and reporting

Many countries and journals now require trials to be registered in public databases before enrollment, with primary outcomes listed in advance. Registration allows readers to compare what was planned with what was published. Studies that switch outcomes or leave out prespecified measures are considered less reliable, and reviewers often check for this discrepancy.

Reporting guidelines, such as the CONSORT statement for randomized trials, ask authors to describe their randomization method, blinding, participant flow and analysis in a standard format. Reading a trial report alongside its registration and guideline checklist is a practical way to judge its quality.

Key takeaways

Randomized controlled trials are the strongest tool for testing causal effects because random assignment balances groups on average. Good control conditions and blinding protect against expectation and bias.

Results should be read through effect size and confidence intervals, not significance alone. Preregistration and large samples make findings more trustworthy.

A single trial is one piece of evidence. Replication and systematic reviews give a more reliable picture, and the design should match the strength of the claim.