Overdiagnosis

Why can finding far more cancers early lengthen measured survival without saving a single life?

A screening programme starts. Within a few years the figures look wonderful: far more cancers found, found earlier, and five-year survival among those diagnosed has climbed sharply. Everyone involved is doing exactly what they set out to do.

Now the awkward part. Over the same years, the number of people dying of that cancer per hundred thousand of the population has not moved. Both statements can be true at once, and the fact that they can is not a paradox to be explained away — it is a property of what "survival" measures. So: what exactly is being counted, and what would we have to count instead?

Take the arithmetic first, because it is not subtle and it is not a bias in the statistical sense. Survival is measured from the date of diagnosis. Suppose a person's cancer would have become symptomatic in 2030 and killed them in 2033 — three years of measured survival. Screening finds it in 2027 instead. If the treatment changes nothing whatever about the disease, the death still falls in 2033, and measured survival is now six years. The clock started earlier; nothing else happened. That is lead-time bias, and it is an identity: advance the diagnosis by t and survival rises by exactly t, with the date of death untouched. A five-year survival threshold is especially easy to cross this way.

That alone would be enough to break survival as evidence. But there is a second effect, and it is stranger. Screening does not sample cancers at random. It samples the population at intervals — yearly, biennially — and catches whatever is detectable at that instant. Ask which cancers are likely to be sitting there detectable when the scan happens. A tumour that grows slowly is detectable for a long span, so almost any screen will catch it. A tumour that goes from undetectable to symptomatic in eight months has a good chance of arriving and declaring itself entirely between two screens — an interval cancer.

So the screen-detected group is systematically enriched for the slow ones and depleted of the fast ones, before any treatment is given. Of course they survive longer. They were a different population from the start. That is length bias.

Now push length bias to its limit and ask the question that names this piece. What about disease so slow, or so stable, that it would never have produced a symptom in that person's remaining lifetime? Under a microscope it meets every criterion for cancer. It is not a false positive — the pathologist is right. But it was never going to matter, and the person is now a patient, treated, and counted for the rest of their life as a cancer survivor. Every such case inflates survival, inflates incidence, and provides a benefit to no one.

So what is the honest test? Not survival, and not cases found. Two things. First, disease-specific mortality in the whole population at risk, which no shifting of diagnosis dates can flatter. Second, an incidence test: if screening genuinely pulls diagnoses forward in time, an early rise in incidence should be followed by a compensating fall, because those cases have been removed from the future. Persistent excess incidence, with late-stage incidence and mortality both unmoved, is the signature of cases that were added rather than advanced.

Two programmes made this visible. Thyroid cancer detection in South Korea rose many-fold after ultrasound screening spread widely, while mortality from thyroid cancer stayed essentially flat — the incidence never came back down. And infant screening for neuroblastoma, trialled in several countries, roughly doubled the number of cases found while producing no reduction in deaths; the programmes were discontinued.

What would refute the account. If screen-detected excess incidence were reliably followed by an equal later deficit, and population mortality fell in step, then everything found was real and advanced rather than added, and overdiagnosis would not be occurring in that programme. Either observation on its own would be enough to abandon the claim for that disease.

Where honesty is required. How much overdiagnosis a given programme produces is genuinely contested: estimates depend on the modelling method, the comparison group, and the follow-up length, and for the same programme they range widely enough that quoting a single percentage would misrepresent the evidence. I am declining to give one. More importantly, none of this tells anyone whether to be screened. Overdiagnosis is invisible in the individual — no one can be told which case theirs is — and weighing it against the chance of benefit is a decision for a person and a clinician who can examine them, not something a mechanism can settle.

The neighbouring piece on screening false positives asks a different question, and the difference matters. There, the test is wrong: healthy people are flagged because rare disease plus imperfect specificity means most positives are errors. Here the test is right. The tissue is abnormal, the diagnosis defensible, the pathologist vindicated — and the finding is still useless or harmful. Overdiagnosis is not a failure of accuracy. It is a mismatch between what the disease definition captures and what would ever have harmed the person.

The second correction is about direction of inference. Rising survival is compatible with a programme that saves many lives and with one that saves none, so it cannot distinguish them. It is a proxy that responds to the act of measuring. Mortality in the population is not.

• Why lowering a diagnostic threshold expands disease definitions, and how the same arithmetic appears in blood pressure, bone density, and gestational diabetes.
• How randomised screening trials attempt to sidestep all three effects, and why their long follow-up makes them so rare.

Every term the collection defines is gathered in the glossary.