What is overdiagnosis?

Overdiagnosis is the detection of a real disease that would never have caused symptoms or death in a person’s lifetime. The diagnosis is correct: the pathology is present and meets the criteria. But finding it cannot help that patient, and treating it can only do harm. Overdiagnosis is defined at population level, because no test says which individual case was overdiagnosed.

That last clause is the whole difficulty. Overdiagnosis is an unobservable counterfactual in any single patient: you cannot rerun a life without the diagnosis. Everything that follows — the estimation methods, the disputes over magnitude, the reviewer objections — is a consequence of trying to measure a quantity that is only ever visible in aggregate.

Overdiagnosis is not misdiagnosis, and not a false positive

Overdiagnosis and misdiagnosis are opposites in one important respect: overdiagnosis is a correct diagnosis. A pathologist looking at an overdiagnosed papillary thyroid carcinoma sees papillary thyroid carcinoma, and a second pathologist would agree. Nothing was mistaken; the tumour simply was never going to matter.

A false positive is also a different thing. A false positive is an abnormal screening result that is not confirmed on work-up, and it resolves within weeks. An overdiagnosed case is confirmed, entered in the cancer registry, staged, treated, and counted in survival statistics for the rest of the patient’s life.

Overtreatment is the consequence of overdiagnosis, not a synonym for it. Every overdiagnosed cancer that is treated is overtreated, because no treatment of a lesion that would never have surfaced can produce benefit. The reverse does not hold: overtreatment also occurs when a real, progressive disease is treated more aggressively than evidence supports.

The two routes by which a case becomes overdiagnosed

Overdiagnosis arises through two distinct mechanisms, and separating them changes what a paper should report.

Non-progressive or regressive disease. Some lesions meeting histological criteria for malignancy do not grow, or grow so slowly that the doubling time exceeds any plausible remaining lifespan, or regress. The Ryser et al. modelling study in Annals of Internal Medicine (2022) attributed 6.1% of screen-detected breast cancers in a biennial 50–74 programme to detection of genuinely indolent preclinical disease.

Competing mortality. A lesion may be progressive but slow, and the person dies of something else first. The same 2022 analysis attributed 9.3% of screen-detected cases to this route — a larger share than the indolent route. This is why overdiagnosis is a function of the patient, not only of the tumour: the same 4 mm lesion is overdiagnosis in an 82-year-old with heart failure and not overdiagnosis in a 45-year-old.

The competing-mortality route produces a measurable age gradient. In that cohort the overdiagnosis rate rose from 11.5% at the first screen at age 50 to 23.6% at the last screen at age 74. Any paper reporting a single overdiagnosis figure across a wide age range is averaging over a strong gradient, and analyses that ignore competing risks will understate the effect in older cohorts.

The reservoir of subclinical disease

The reservoir is the stock of histologically malignant lesions present in living people who will never be diagnosed, and it is what makes overdiagnosis possible at all. Harach and colleagues sectioned 101 consecutive Finnish autopsy thyroids at 2–3 mm intervals in 1985 and found occult papillary carcinoma in about a third of them, most of the tumours under a millimetre in diameter. None of those people had been diagnosed with thyroid cancer in life.

The reservoir is why diagnostic sensitivity is not an unmixed good. A test that finds smaller lesions reaches deeper into a reservoir whose contents are, by construction, the least likely to progress. This inverts the usual intuition that a more sensitive test is a better test: past a threshold, added sensitivity buys mostly overdiagnosis, because the clinically important lesions were already being found.

Comparable reservoirs are documented for prostate, breast and kidney. The size of the reservoir differs enormously by organ, which is why overdiagnosis estimates for thyroid cancer and for pancreatic cancer are not remotely on the same scale, and why importing a magnitude from one screening literature into another is an error.

The population signature: incidence rises, mortality does not

The characteristic population-level signature of overdiagnosis is a large, sustained rise in incidence with no corresponding fall in disease-specific mortality. Ahn, Kim and Welch reported in the New England Journal of Medicine (2014) that the rate of thyroid-cancer diagnosis in the Republic of Korea in 2011 was 15 times the 1993 rate, following the addition of ultrasound screening to government-funded cancer screening, while thyroid-cancer mortality stayed flat. Virtually all the additional tumours were papillary.

The same signature appears outside cancer. Wiener and colleagues, analysing United States national data from 1998 to 2006, reported that pulmonary embolism incidence rose from 62.1 to 112.3 per 100,000 adults after computed tomography pulmonary angiography came into routine use, while pulmonary-embolism mortality moved only from 12.3 to 11.9 per 100,000. Cutaneous melanoma shows a similar divergence: diagnoses several-fold higher than four decades ago with broadly stable mortality.

This signature is suggestive, not conclusive. A genuine rise in disease incidence, a real improvement in treatment, or a change in death certification could each produce part of the pattern. The signature carries weight when the incidence rise is confined to small, early-stage lesions, when advanced-stage incidence is unchanged, and when the rise tracks the introduction of a test rather than a change in exposure.

Neuroblastoma screening: the closest thing to a controlled demonstration

Neuroblastoma screening in infancy produced the cleanest evidence of overdiagnosis available, because two large population studies with concurrent comparison populations reported in the same issue of the New England Journal of Medicine in April 2002. The Quebec Neuroblastoma Screening Project screened infants across a whole province and compared them with unscreened North American populations; the German neuroblastoma screening study offered screening to about 2.6 million children, of whom roughly 1.5 million were screened, and compared them with a control area of about 2.1 million.

The German study found no fall in stage 4 disease (3.7 versus 3.8 cases per 100,000) and no fall in mortality (1.3 versus 1.2 deaths per 100,000), with substantial overdiagnosis estimated at 7 cases per 100,000. The Quebec study found no reduction in neuroblastoma mortality against any of its four comparison populations. Earlier reports from the Quebec project had already established the other half of the signature: screening roughly doubled incidence (standardised incidence ratio 2.17, 95% CI 1.79 to 2.57; Woods et al., Lancet 1996), with the excess concentrated below one year of age and no offsetting reduction in cases diagnosed later (Woods et al., European Journal of Cancer 1997).

Japan halted its national infant neuroblastoma screening programme in 2004. The neuroblastoma case matters methodologically because the untreated natural history was independently known — spontaneous regression of infant neuroblastoma is documented — so the excess cases could be identified as overdiagnosis rather than argued about indefinitely. Screening programmes for adult cancers rarely offer that.

How overdiagnosis is estimated, and what each method assumes

Three families of methods exist for estimating overdiagnosis, and each carries an assumption that determines whether the resulting number means anything.

Excess cumulative incidence in a randomised trial. Compare cumulative incidence in screened and unscreened arms after screening has stopped and enough time has passed for the screened arm’s lead time to be exhausted. Zackrisson and colleagues reported in the BMJ (2006) that in Malmö women aged 55–69 at randomisation, the excess persisting 15 years after the trial ended corresponded to about 10% of cancers diagnosed in the screened arm — substantially smaller than the excess visible while screening was still running, which had not yet been eroded by the control arm catching up. The assumption is that the control arm was never screened — which is why contaminated control arms make this design uninterpretable.

Microsimulation modelling. Fit a natural-history model with a preclinical sojourn distribution to observed screening and incidence data, then compute the fraction of detected cases whose clinical surfacing would never have occurred. Ryser et al. (2022) estimated 15.4% of screen-detected breast cancers overdiagnosed under biennial screening at ages 50–74, with an uncertainty interval of 9.4% to 26.5%. The assumption is the model’s structure, and different structures fitted to the same data give different answers.

Follow-up of untreated or actively surveilled cohorts. Observe lesions that were diagnosed but not treated. This bounds progression directly, and its limitation is that patients on active surveillance are selected for low-risk features.

Why published estimates disagree so violently

Published overdiagnosis estimates for breast screening range from under 5% to over 50%, and a substantial part of that spread is denominator choice rather than genuine disagreement about biology. The UK independent panel chaired by Michael Marmot (2012) reported both numbers from the same data: about 11% of breast cancers diagnosed over an invited woman’s remaining lifetime, and about 19% of cancers diagnosed during the screening period. Neither figure is wrong; they answer different questions.

Follow-up duration matters at least as much. In the National Lung Screening Trial, Patz and colleagues estimated in JAMA Internal Medicine (2014) that 18.5% of low-dose CT-detected lung cancers were overdiagnosed overall, but that the fraction differed enormously by histology — close to four-fifths for bronchioloalveolar carcinoma and far lower for the aggressive subtypes. Later analyses with several more years of follow-up put the overall figure much lower still, because the control arm continued to accrue the cancers that screening had merely found early. The excess in the screened arm peaks around the end of screening and then decays as the control arm catches up — so a short-follow-up estimate is inflated by residual lead time, and an estimate reported at the end of the screening phase is not an overdiagnosis estimate at all.

When you read or write an overdiagnosis number, three qualifiers determine its meaning: the denominator, the follow-up horizon after screening stopped, and the age range. A figure without all three is not comparable to any other figure.

Overdiagnosis beyond cancer screening

Overdiagnosis is not confined to oncology, and the non-cancer forms are usually driven by threshold changes rather than by imaging. Lowering the diagnostic threshold for a continuously distributed marker converts a population of people previously called healthy into patients overnight, and whether that is beneficial depends entirely on whether treatment at the new threshold reduces hard outcomes.

Three mechanisms recur outside cancer. Threshold shifts in conditions defined by a cut-point on a continuous variable — blood pressure, glycaemia in pregnancy, estimated glomerular filtration rate, bone mineral density. Incidental findings from imaging performed for an unrelated indication, where the incidentaloma is real, benign in the overwhelming majority, and triggers a surveillance cascade. Expanded criteria for syndromic diagnoses, where broadened definitions capture milder presentations whose prognosis differs from the population in which the original evidence was generated.

The pulmonary embolism example is instructive because no threshold changed: sensitivity did. Computed tomography pulmonary angiography detects subsegmental emboli that older ventilation-perfusion scanning did not, and the clinical significance of an isolated subsegmental embolus in a stable patient remains contested.

How overdiagnosis is detected in a real manuscript

Overdiagnosis rarely announces itself in a manuscript. It shows up as a set of reporting choices, and the following are the ones that most reliably indicate the concept has been mishandled.

Survival used as the outcome of a screening comparison. Five-year survival rises mechanically under overdiagnosis, because overdiagnosed cases enter the denominator and never die of the disease. A screening paper that reports improved survival without disease-specific mortality has reported an artefact. This is the highest-yield single check on any screening manuscript.

Stage shift presented as evidence of benefit. A larger proportion of early-stage disease is compatible with lives saved and equally compatible with pure overdiagnosis. The discriminating quantity is the absolute incidence of advanced-stage disease, which must fall if early detection is intercepting progressive cancers. Papers that report the proportion of early-stage cases, rather than the rate of late-stage cases, cannot distinguish the two.

Incidence reported without a post-screening follow-up window. If cumulative incidence is compared while the screened arm is still accruing lead time, any excess is uninterpretable.

Comparison of screen-detected with symptom-detected cases. Screen-detected cases are enriched for slowly growing disease by length-time bias before any overdiagnosis is considered, so their better outcomes are expected under the null.

In-situ and microcarcinoma cases pooled with invasive disease. Ductal carcinoma in situ and sub-centimetre papillary thyroid carcinoma carry the highest overdiagnosis fractions, and pooling them into a single “cancers detected” count hides the group that drives the estimate.

No competing-risk framing in an older cohort. A screening benefit claimed in a cohort with a median age above 75 and no accounting for other-cause mortality is not addressing the dominant mechanism of overdiagnosis in that population.

Registry incidence treated as disease occurrence. Cancer registries count diagnoses, and registry data have limitations that make a diagnosis count a measure of diagnostic activity as much as of disease.

Two related time-related artefacts travel with these: immortal time bias in cohort analyses of screened populations, and confounding by the healthy-screenee effect, since people who attend screening differ systematically from those who do not.

What a peer reviewer says when overdiagnosis is mishandled

“The authors report improved five-year survival; survival is not an appropriate endpoint for evaluating screening, and disease-specific mortality should be presented.” “The observed stage shift is compatible with overdiagnosis; please report the absolute incidence of advanced-stage disease in both groups.” “Follow-up after the cessation of screening appears insufficient to exclude residual lead time, so the excess incidence cannot be interpreted as overdiagnosis.” “The overdiagnosis estimate is reported without specifying its denominator.” “The comparison of screen-detected with clinically detected cases is subject to length bias and does not support the causal claim made in the discussion.” “The rising incidence reported here is attributed to increasing disease burden without considering changes in diagnostic intensity over the same period.”

The last of those is the objection that most often decides a paper, because it points at the study’s central interpretation rather than at a statistical detail. An incidence trend paper that does not address diagnostic intensity has an alternative explanation sitting in plain view, and reviewers who work in screening will raise it. Anticipating it in the manuscript is far stronger than answering it in a response letter.

How to report overdiagnosis honestly

Report cumulative incidence in both arms with a stated post-screening follow-up horizon, and give the year at which the curves were compared. Pre-specify the denominator and name it in every sentence that carries the estimate. Report disease-specific mortality as the primary outcome of any screening evaluation, with all-cause mortality alongside it, since a screening programme that shifts deaths between causes has not helped anyone.

Separate in-situ and minimal-risk histology from invasive disease in every table. Present the estimate by age stratum rather than as a single pooled figure. Where the estimate depends on a natural-history model, report results under at least two model structures, because structural uncertainty typically exceeds the sampling uncertainty that the confidence interval describes.

State the limitation positively. “Overdiagnosis could not be estimated in this cohort because follow-up after screening ended was 4 years” is a specification a reader can act on. “Overdiagnosis may be present” is not.

Contested points worth stating plainly

The magnitude of breast screening overdiagnosis remains genuinely disputed, and a page that pretends otherwise is misleading. Estimates cluster near 10–20% of screen-detected cancers in modelling studies and near or above 30% in some trial-based analyses, and the disagreement turns on control-arm contamination, follow-up length and model structure rather than on any single resolvable fact.

Whether particular entities should be called cancer at all is contested. Proposals to reclassify low-risk lesions — non-invasive follicular thyroid neoplasm with papillary-like nuclear features was renamed in 2016 to remove the word carcinoma — change measured incidence without changing biology, which means incidence trends across such a reclassification are not comparable.

The direction of screening policy also responds to this evidence. Thyroid operations in South Korea fell by roughly a third in the year after the overdiagnosis argument received sustained public and professional attention — a rare instance of a measured reduction in diagnostic activity following publication, rather than the usual ratchet in the other direction.

Finally, overdiagnosis is not an argument against screening. Prostate-specific antigen screening in the European Randomized Study of Screening for Prostate Cancer required, at 16 years, 570 men invited and 18 men diagnosed to prevent one prostate-cancer death; both benefit and overdiagnosis are real, and the policy question is how they trade off for a given person. A manuscript that treats overdiagnosis as a knock-down argument is making the mirror image of the error it is criticising.

Related

Lead-time bias · Length-time bias · Confounding · Collider bias

Checked before submission by the epidemiology reviewer and the overclaim check, which flag survival endpoints in screening comparisons and incidence trends interpreted without reference to diagnostic intensity.

Review my manuscript

Frequently asked questions

What does overdiagnosis mean in medicine?

Overdiagnosis means correctly identifying a disease that would never have produced symptoms or caused death if it had gone undetected. The pathology is real and the diagnostic criteria are met. The problem is that detection cannot benefit the patient, while treatment carries the usual risks.

What is the definition of overdiagnosis in epidemiology?

Overdiagnosis in epidemiology is defined as the diagnosis of a condition that would not have become clinically apparent during a person’s lifetime, arising either because the lesion is non-progressive or because the person dies of another cause first. It is quantified as a fraction of diagnosed cases in a population.

How is overdiagnosis different from misdiagnosis?

Overdiagnosis is a correct diagnosis of a disease that did not need finding; misdiagnosis is an incorrect diagnosis. A second pathologist reviewing an overdiagnosed thyroid carcinoma would confirm the cancer. Overdiagnosis is also distinct from a false positive, which is an abnormal test result not confirmed on work-up.

What is overdiagnosis in cancer screening?

Overdiagnosis in cancer screening is the detection of a histologically malignant tumour that would never have surfaced clinically. Its population signature is a sustained rise in incidence without a matching fall in disease-specific mortality, as observed in South Korean thyroid cancer, where diagnoses in 2011 were 15 times the 1993 rate with mortality unchanged.

Can you give an example of overdiagnosis?

Infant neuroblastoma screening provides the clearest example. Two 2002 studies in Quebec and Germany roughly doubled the number of neuroblastomas diagnosed, with no reduction in advanced-stage disease and no reduction in mortality. Japan ended its national screening programme in 2004. The extra cases were tumours that would have regressed unnoticed.

Why does overdiagnosis matter for my manuscript?

Overdiagnosis determines which outcomes a screening or incidence paper may legitimately report. Improved five-year survival, a favourable stage distribution, and rising incidence are each fully compatible with overdiagnosis and prove nothing about benefit. Reviewers in screening research raise this objection routinely and it commonly decides the paper.

How do researchers measure overdiagnosis?

Researchers measure overdiagnosis through excess cumulative incidence in randomised trials followed long after screening stops, through microsimulation models of preclinical natural history, and through follow-up of untreated cohorts. Estimates vary widely because they depend on the denominator chosen, the follow-up horizon, and the age range studied.