Semiparametric and Joint Modeling of Cancer Screening

Abstract

Introduction of screening for prostate cancer using the prostate-specific antigen (PSA) biomarker of the disease in the late 80ies led to remarkable dynamics of the incidence of the disease and shortly after, cancer mortality showed a decline. Except for fragmentary studies, no comprehensive information exists on the PSA uptake in the European countries that would allow specification of utilization intensity by age and calendar time, which puts forward the problem of estimating PSA utilization patterns from cancer incidence and mortality data. Even in the USA the patterns have been heterogeneous and showed nontrivial dynamics. Capturing the picture by parametric methods has been very challenging. Although prostate cancer mortality rates have fallen dramatically since the widespread adoption of PSA screening in the early 1990s, conclusively establishing screening benefit requires evidence from randomized controlled trials. Former studies did not formally evaluate whether screening efficacy differed between trials when implementation details such as screening patterns are taken into account and conflicting results have been seen between trials.

In the second chapter, we formulate a joint model of cancer progression to symptomatic (clinical) diagnosis and the screening process with the associated detection mode, as both processes interact to produce the observed incidence in the population. The risks of screening and clinical diagnosis are dependent sharing the latent tumor onset and progression processes in the subject, denoted by a common shared frailty term. Intensity of screening and the hazard driving prostate cancer progression are estimated jointly and semiparametrically using the NPMLE method based on the joint model. Asymptotic and finite sample properties of the proposed estimators are studied analytically and by simulations. An application using data from the European cancer registry EUREG is presented.

In the third chapter, we develop a semiparametric joint model of cancer progression to clinical and screening diagnosis based on screening trials data with a mixture of known PSA test schedules per protocol and random unknown schedules before and after implementation of the protocol in both control and screening arm. Ad-hoc screening patterns in both arms before recruitment and after existing the trial, and the hazard driving prostate cancer progression are estimated jointly and semiparametrically. Hypothesis tests comparing the screening risks between the arms and periods are performed to validate if the randomization was contaminated. Applications using the subject-specific incidence data for both control and screening arms from Prostate, Lung, Colorectal, and Ovarian screening trial (PLCO) and cancer incidence data from The Surveillance, Epidemiology, and End Results (SEER) Program are demonstrated.

In the fourth chapter, we derive the lead time to link cancer mortality with cancer incidence and screening efficacy. We use a two-step approach to formally test whether screening efficacy differs between trials using mean lead time as a surrogate of screening intensity. First, the mean lead time is estimated in each trial arm as a proxy for the intensity of screening. Second, the association is quantified between the mean lead time and prostate cancer mortality and tested whether it differs between trials while accounting for differences in screening and diagnosis between arms. We analyze the individual-level data from PLCO jointly with SEER US population data to prove that there is no evidence that screening efficacy differed between trials and screening can significantly reduce the risk of prostate cancer death.