The Quiet Minority: Who Slips Out of the Record When Patients Refuse NHS Data Sharing

Every dataset has a hole in it. For England's national health data estate, the most discussed hole is the one created deliberately: the national data opt-out, which allows patients to prevent their confidential patient information being used for research and planning beyond their direct care. The proportion who use it is modest — broadly one record in twenty, though it varies sharply by place and practice. The effect on what the data can be trusted to say is anything but modest.

What opting out actually changes

The national data opt-out applies in England and operates at the point where data leaves the settings in which it was collected. It does not stop a patient's information being used to treat them, and it does not remove them from their GP's own record. It prevents that information being disclosed for purposes beyond individual care, subject to limited exceptions. Scotland, Wales and Northern Ireland run separate arrangements, which is itself a source of confusion for anyone comparing figures across the UK.

Separately, a smaller number of patients have registered a type 1 opt-out with their own practice, restricting extraction at source. The two mechanisms overlap imperfectly, and the practical result is a patchwork in which the same individual may be included in one national collection and excluded from another.

Missing, but not at random

Statisticians have a phrase for the assumption that missing data can be safely ignored: missing completely at random. It almost never holds in health records, and the opt-out is no exception. What evidence exists suggests a consistent pattern. People who opt out skew younger than the patient population as a whole, are more likely to live in urban and affluent areas, and are more likely to be digitally engaged — people who have read about data sharing, formed a view, and acted on it. General practice staff report that opt-outs cluster among articulate, working-age patients who are also, on average, healthier than the cohort that never engages with the question at all.

There is a second and more troubling pattern. Communities with a history of being poorly served by medical research — including some migrant and minority ethnic groups, and people with experience of past data controversies — are more likely to withhold. The reasons are not irrational. They are rooted in a reasonable assessment of risk and of whether the benefits of research have historically been shared.

The consequence is a dataset that is not a random sample of the population. It is a sample with a slight but structured bias towards older, sicker and more settled patients.

What bias in the base does to prevalence

For common conditions the effect is usually small and partly correctable. For rarer conditions, and for anything estimated at small-area level, it can be substantial. A prevalence estimate built on a denominator that has lost five per cent of the population in a non-random way is not simply five per cent less complete; it is systematically skewed, and the direction of the skew differs by condition.

Conditions associated with younger adulthood — certain mental health diagnoses, some autoimmune disease, sexual health presentations — are likely to be underestimated. Conditions concentrated in older age may be modestly overestimated, since the remaining denominator is older than the true population. Researchers working on rare disease registries are candid that opt-out patterns compound the existing difficulty of small numbers.

A patient advocate involved in research oversight put the tension plainly: the right to refuse is not an obstacle to be managed, it is the point. The engineering problem is to make the answer honest when part of the population has declined to be counted.

The counterfactual problem

There is a deeper methodological difficulty. You cannot directly measure the health of people who have removed themselves from the data, because that is precisely what they have asked you not to do. Some studies attempt to characterise the opted-out population in aggregate, using practice-level counts rather than individual records, and a handful have compared linked sources to infer direction of bias. These are valuable but partial. In practice, most analyses simply document the proportion opting out and note it as a limitation.

That is honest but thin. A limitations paragraph does not tell a commissioning body whether its prevalence figure is five per cent or fifteen per cent wrong.

What researchers can do instead

The most promising responses are structural rather than statistical. Secure data environments, in which accredited researchers analyse data inside a controlled setting and never receive record-level extracts, address some of the public concern that drives opting out in the first place. Transparency registers listing approved projects, plain-English explanations of what linkage actually involves, and public panels with genuine influence over decisions all appear to matter more than reassurance campaigns.

Where opt-out rates are known and patterned, weighting and sensitivity analyses can quantify how much a conclusion would have to shift to be overturned. Several 2026 analyses now report estimates under plausible assumptions about the opted-out population rather than a single point figure — a practice worth encouraging.

Clinicians have a part too, though a delicate one. A GP in a practice with an unusually high opt-out rate described the conversation as one of the hardest in the consulting room: patients want to know who will see their data, what happens if they refuse, and whether refusal will affect their care. The honest answer to the last question is no, and saying so clearly, repeatedly and without pressure is probably the most effective thing the system can do.

Trust as infrastructure

The opt-out is sometimes described in policy discussion as a risk to data quality. It is better understood as a signal. A rising opt-out rate is a measurement of public confidence, and a falling one is not automatically an improvement if it has been achieved by making refusal harder to find.

The UK's health data estate remains one of the most valuable assets in medical research precisely because it covers a whole population rather than a self-selected slice of it. Preserving that coverage depends less on technical fixes than on a straightforward bargain: people are asked to contribute their data, told honestly what it will be used for, and given a real and easy way to say no. The records that disappear when they do are a permanent reminder of what that bargain costs — and of why it has to be kept.