How Overdiagnosis Is Actually Measured

Overdiagnosis, the detection of a cancer that would never have caused symptoms or death, cannot be observed in an individual, because once a cancer is treated its natural fate is unknown. It can only be estimated across a population, and researchers use four broad methods: long-term follow-up of a randomized trial, pathology or imaging studies, natural-history modeling, and excess-incidence cohort or ecological studies. The estimates vary widely because they depend on the definition, the denominator, and how lead time is handled.

A number you cannot read off one patient

Overdiagnosis means finding a cancer that would never have gone on to cause symptoms or death. The hard part is that you can never point to a single treated patient and say this one was overdiagnosed, because once a cancer is removed you cannot observe the future in which it was left alone. Overdiagnosis is therefore always a population estimate, and the method you use to estimate it shapes the answer.

The randomized trial approach

The lowest-bias method follows a randomized screening trial for a long time after screening ends. If screening only advances the timing of diagnosis, the unscreened arm should eventually catch up in total cases. Any lasting excess of cancers in the screened arm, after the control arm has had time to catch up, is an estimate of overdiagnosis. It is the cleanest design, but it is slow, may not generalize, and cannot be used to monitor a program in real time.

Pathology, imaging, and modeling

A second approach studies the biology of screen-detected cancers, inferring from their features how many were unlikely to progress. It is simple but rests on the shaky assumption that those features reliably predict a cancer's future. A third approach builds mathematical models of the natural history of disease, estimating lead time and back-calculating the overdiagnosed fraction. Modeling is faster and more flexible, but its output depends on assumptions about the very natural history that no one can observe directly.

Excess incidence and its traps

The most common real-world method compares cancer incidence in screened and unscreened populations and calls the excess overdiagnosis. Done carelessly it inflates the number, because incidence also rises simply from pulling future diagnoses forward in time, and from underlying trends unrelated to screening. Credible excess-incidence studies adjust for lead time and for background incidence, and a systematic review of methods concluded that well-conducted cohort and ecological studies are the most practical way to monitor overdiagnosis over time.

Why the estimates disagree so much

The same cancer can carry very different overdiagnosis figures depending on the denominator, whether it is all screen-detected cancers, all cancers in screened people, or all cancers in the whole population. Estimates from randomized breast screening place a meaningful fraction of screen-detected cases in the overdiagnosed category, and modeling estimates for prostate cancer detected by prostate-specific antigen run higher still. A careful reader treats any single percentage as one method's answer under one definition.

The questions to ask

Before you accept an overdiagnosis number, ask which method produced it, what the denominator was, and whether excess incidence was adjusted for lead time and background trends. Ask whether the follow-up was long enough for an unscreened group to catch up. The figure is real and important, but it is an estimate with a method attached, and the method is where most of the disagreement lives.