EMPIRICAL ESTIMATION OF OVERDIAGNOSIS IN TRIALS AND POPULATION STUDIES

   Purpose: To demonstrate why many empirical estimates of overdiagnosis based on excess incidence under screening are wrong and establish general conditions for unbiased estimation in randomized clinical trials and population settings.

   Methods: We develop a deterministic framework to project the effects of screening on the incidence of disease. Using the framework, we show that incidence depends on the preclinical detectable period, test sensitivity, screening patterns, and the true frequency of overdiagnosis, and we project expected incidence in common trial designs and population settings. In each setting, we compare the excess incidence in a screened group relative to an unscreened group with the true frequency of overdiagnosis to determine if and when unbiased empirical estimation is possible.

   Results: In trials with no control arm contamination, only cumulative excess incidence is unbiased if the screen arm stops screening and only annual excess incidence is unbiased if the screen arm continues screening (see Figure). In both designs, under imperfect test sensitivity, unbiased estimation requires waiting until screening stabilizes plus the maximum preclinical period. In trials with continued screen arm screening and in settings in which screening is adopted by the population, cumulative excess incidence is persistently biased.

    Figure: Hypothetical annual (upper panels) and cumulative (lower panels) incidence if the screen arm stops screening (left panels) and if the screen arm continues screening (right panels). In each design, 50,000 individuals are randomized to each arm, preclinical disease develops at an annual rate of 100 per 100,000 individuals, the range of preclinical periods is 0–6 years, episode sensitivity is 50%, and 25% of cases detected by screening are overdiagnosed (shaded strips). Black arrows indicate minimum follow-up for unbiased estimation of overdiagnosis using excess incidence in the screen relative to the control arm.

   Conclusions: No trial or population setting automatically permits unbiased empirical estimation of overdiagnosis. Sufficient follow-up and appropriate analysis remain crucial.