What Clinically Proven Does and Does Not Mean

The phrase "clinically proven" appears on packaging, websites and advertisements. It sounds final. It does not say which study was done, in whom, with what result, or how large the effect was. Here we set out what regulators actually ask for, how often promising work fails, and how to read a result for yourself.

What the phrase can and cannot carry

The US Federal Trade Commission uses the phrase in its examples, including an advertisement claiming that software is "clinically proven to improve memory". We did not find a definition of it in the FTC Health Products Compliance Guidance, issued on 20 December 2022. The guidance does set the standard that matters. Health claims require "competent and reliable scientific evidence", and randomized, controlled human clinical trials are described as the most reliable form of evidence.

The same guidance makes three points a reader can use. Evidence must be relevant to the specific product, dose and population being advertised. Studies cannot be considered in isolation, and conflicting results raise serious questions. Animal and laboratory studies may be useful background but are not sufficient without confirmation in human trials. It also says that "clinically tested ingredient" implies the test results prove a benefit, and that replication in an independently conducted study adds to the weight of evidence.

What the FDA asks for

For drugs and biologics, the FDA's draft guidance on substantial evidence of effectiveness, issued in December 2019 and nonbinding, describes "adequate and well-controlled" investigations as the primary basis. The FDA generally requires at least two such trials, each convincing on its own, with exceptions such as one large trial or one trial plus confirmatory evidence. The effect shown must be clinically meaningful in the FDA's judgment. The guidance also notes that poor execution, such as extensive dropout, can make a trial of any design inadequate.

Two lessons follow. A single positive study is a starting point. And the size of an effect matters as much as its existence.

How often promising work fails

The FDA's description of clinical research, with content marked current as of 01/04/2018, puts the share of drugs moving on from Phase 1 at about 70 percent, from Phase 2 at about 33 percent and from Phase 3 at about 25 to 30 percent. Multiplying the stages gives 0.70 x 0.33 x 0.25 = 5.8 percent at the low end and 0.70 x 0.33 x 0.30 = 6.9 percent at the high end, before regulatory review.

Other estimates are higher. A study by Wong, Siah and Lo, summarised by MIT Sloan on 31 January 2018, found that nearly 14 percent of drugs in clinical trials eventually win approval. The published abstract covers 406,038 trial entries and over 21,000 compounds from January 2000 to October 2015. Success varied by area, from 33.4 percent for infectious-disease vaccines to 3.4 percent for cancer. The two sets of figures differ because they use different data and methods. The honest summary is that most candidates that enter trials do not reach approval, and the exact share depends on the field and the dataset.

A worked example

The figures here are illustrative, not real data. Imagine a trial of an unnamed intervention measured on a 0 to 100 symptom score. There are 1,000 participants in each arm. The average difference between arms is 1.5 points, and the standard deviation is 12 points.

The standard error of the difference is 12 x square root of (2 / 1000) = 0.54. The test statistic is 1.5 / 0.54 = 2.8, which gives a p-value of about 0.005. The 95 percent confidence interval runs from about 0.4 to 2.6 points. The result is statistically significant. Now suppose experts judge that a change of at least 5 points matters to patients. Even the top of the interval is below that. The FTC guidance addresses this directly: statistically significant results must translate to a benefit that is clinically meaningful.

Reverse the case. With the same 1.5 point difference and only 50 participants per arm, the standard error is 2.4, the statistic is 0.63 and the p-value is about 0.53. That is not significant, but it does not show there is no effect. It shows the trial was too small to tell.

Going deeper

The arithmetic above uses these formulas:

• Standard error of a difference in means = SD x square root of (2 / n), for two arms of n participants with equal spread
• Test statistic z = Difference in means / Standard error
• 95 percent confidence interval = Difference +/- 1.96 x Standard error
• Chance of reaching approval from Phase 1 = P(Phase 1) x P(Phase 2) x P(Phase 3), assuming each stage is independent of the others

A related idea is the surrogate endpoint. The FDA's table of surrogate endpoints, last updated 29 April 2026, quotes the statutory definition: a marker, such as a laboratory measurement, radiographic image or physical sign, that is not itself a direct measure of benefit to patients. It is used either because it predicts clinical benefit well enough to support traditional approval, or because it is reasonably likely to, which can support accelerated approval. When you read a result, ask whether the measure was a marker or something patients feel, and whether the population matches yours.

We have linked each source so the figures can be checked. If you find an error, tell us.

Informational only, not medical advice.