Improving Evaluations of Cancer Screening through Better Methods of Estimating Preclinical Distributions

In a randomized controlled cancer screening trial, the screen-detected cancers present a length-biased sample of all preclinical durations, as cases that progress more slowly prior to diagnosis are more likely to be caught by regular screening than their faster-paced counterparts. This leads to an overestimate of the life-extending benefits of the screening test being evaluated. Previous research has shown that the severity of the length-biased sampling effect depends on the joint distribution of preclinical and clinical durations, but these simulation studies did not make data-driven choices for the distributions of these cancer growth periods.
We discovered that a mixture of two exponential distributions fit the clinical durations from three historic screening trials well after developing a simple exploratory procedure to estimate the parameters of such a mixture model. Furthermore, we found that, when simulating preclinical durations from the same type of mixture distribution, the parameters could be inferred from the pattern of diagnoses in the trial, a key finding given that preclinical durations are unobservable in a real trial. We used these simulated results to train a predictive model to estimate the distribution of preclinical durations from observable trial outcomes. Finally, we calculated the mean length-biased sampling effect under a variety of preclinical duration distributions, screening test sensitivity models, and screening programs. Using our approach to predict preclinical duration distribution parameters from trial outcomes could allow for the length-biased sampling effect to be more accurately specified for a real cancer screening trial.