What Is a Clinical Trial? Finding Out Whether a Treatment Actually Works

The design of a trial is a sequence of defences, each against a specific way of being fooled:

• •A control group, receiving placebo or the current standard treatment, which removes the effects of time, of regression to the mean and of attention
• •Randomisation, allocating participants by chance, which balances known and unknown differences between groups and is the feature that permits a causal claim
• •Blinding of participants, which removes the effect of expectation, and of assessors, which removes biased measurement, with double-blind meaning both
• •A pre-registered protocol specifying the primary outcome before the data exist, which prevents choosing the measure that happened to work after seeing the results
• •Adequate sample size, calculated in advance, since a study too small to detect a real effect wastes the participants' time and produces a misleading negative
• •Intention-to-treat analysis, counting participants in the group they were assigned to regardless of what they actually did, which prevents the dropouts from quietly reshaping the comparison

Drug development runs through stages with different questions. Preclinical work in cells and animals establishes plausibility and initial safety. Phase one enrols a small number of usually healthy volunteers and asks about safety, dosing and how the body handles the compound. Phase two enrols a few hundred patients with the condition and asks whether there is any signal of benefit at a tolerable dose. Phase three enrols hundreds to thousands, is randomised and controlled against the best available alternative, and is the study regulators require. Phase four is post-marketing surveillance after approval, which matters because rare harms appear only once a drug reaches large populations, and several drugs have been withdrawn at this stage. The attrition is severe, with something under ten percent of compounds entering phase one eventually approved, and most failures occurring at phase two, which is why development costs are quoted in hundreds of millions per approved drug and why the figure is disputed.

The logic developed in stages over two centuries. James Lind ran a controlled comparison of six treatments for scurvy aboard a naval ship in 1747 and identified citrus, in what is usually cited as the first clinical trial, though he did not randomise and the Navy took decades to act. Placebo controls and blinding appear in the nineteenth century. The modern design is generally dated to a British trial of streptomycin for tuberculosis published in 1948, designed by Austin Bradford Hill, which used random number allocation concealed from the recruiting physicians and is the standard example of the method. Ethical standards developed later and in response to abuses: the Nuremberg Code of 1947 followed the trials of Nazi doctors, the Declaration of Helsinki was adopted in 1964, and American regulation was reformed after the Tuskegee study, in which hundreds of Black men with syphilis were observed without treatment for forty years including after penicillin became available, was exposed in 1972.

The failures are well catalogued and mostly structural rather than fraudulent. Publication bias means positive results are far more likely to be published than negative ones, which distorts the literature systematically and was the reason trial registration became mandatory, so that a study's existence is recorded before its result is known. Selective outcome reporting changes the primary endpoint after the fact. Surrogate endpoints measure something easier than what matters, such as a laboratory value rather than survival, and a treatment can improve the surrogate and worsen the outcome, which has happened repeatedly. Unrepresentative participants, historically skewed toward white men, limit what the results say about everyone else, and the exclusion of pregnant women from trials means that prescribing in pregnancy rests on far weaker evidence than elsewhere. Industry funding is associated with results favouring the sponsor, an association documented across many analyses. And comparison against placebo when an effective treatment exists tells a clinician nothing about which to use.

A single trial is evidence rather than proof, which is why the apparatus above it exists. Systematic reviews locate every study on a question, including unpublished ones where possible, assess each for risk of bias and combine them, and meta-analysis pools the results statistically to produce an estimate with more precision than any single study. That process regularly overturns conclusions drawn from individual trials and has established the scale of publication bias by comparing registered studies with published ones. Newer designs address specific limitations: adaptive trials modify allocation as results accumulate so that fewer participants receive the worse option, platform trials test several treatments against a shared control, and pragmatic trials run in ordinary clinical settings rather than in controlled conditions to see whether a treatment works in practice as well as in principle. The underlying problem does not change, which is that human beings are extremely good at seeing an effect where there is none.

A randomised controlled trial separates a treatment's effect from the effects of time, expectation, attention and selective memory, using a control group, random allocation, blinding of participants and assessors, a pre-registered outcome, an adequate sample and analysis by assigned group. Drug development runs through phases with heavy attrition, the modern design dates to a 1948 tuberculosis trial, and the recurring failures are publication bias, changed endpoints, surrogate measures and unrepresentative participants, which systematic review and meta-analysis exist to correct.