PREDICTED MORTALITY AS A SURROGATE ENDPOINT IN THE PLCO TRIAL
Our objective is to use estimate predicted mortality, which is derived by weighting cancer cases diagnosed during a screening trial (including screen-detected cases and those diagnosed following other pathways) by the cancer- and stage-specific survival rate from diagnosis to a pre-specified time since randomisation. A fixed follow-up time from randomisation is used to account for lead-time bias in screen-detected cases. The assumed survival rates are retrieved from data sources available before the trial is completed, external from the trial. Differences in predicted mortality between trial arms might be useful as an early indicator of whether the studied screening intervention is likely to result in a cancer-specific mortality reduction. This information would enable policy-makers to plan for potential implementation and start pilot programmes and implementation research whilst awaiting observed mortality outcomes.
Predicted mortality has been previously applied to study the likely outcomes of breast (e.g., Day and Duffy, J R Statist Soc A 1996; Moss et al. BJC 2005) and colorectal cancer screening trials (Cuzick et al. J Med Screen 2007), with good correlation between the predicted and the subsequently observed mortality estimates. However, this potentially useful surrogate endpoint has not been assessed in the PLCO trial.
We aim to consider a role for predicted mortality as a surrogate (early) outcome in cancer screening trials, using data from the PLCO trial.
The individual-level data requested from the available PLCO datasets for all four cancers include trial arm allocation and year of enrolment, basic sociodemographic statistics and variables representing the inclusion/exclusion criteria in the published trial analyses, screening participation, cancer diagnoses including the relevant clinical variables such as stage and other prognostic markers, mode of detection, time of diagnosis in days from randomisation, and reason (including deaths) and time of censoring of the follow-up.
Assumed cancer-specific survival by stage (or similar) at diagnosis will be derived from the SEER data base and will be based on information that would have been available to the trialists at the time of diagnosis.
We will compare the calculated predicted mortality based on data to different follow-up times ≤13y, with the published 13-year differences in the observed cumulative mortality rates between the trial’s intervention vs. control arms (Andriole et al. JNCI 2012 for prostate, Oken et al. JAMA 2011 for lung, Schoen et al. NEJM 2012 for colorectal, and Buys et al. JAMA 2011 for ovarian cancers).
Separate analyses will be undertaken to study the effect of different assumptions of mortality risk differences between trial participants and the general population, and, in the intervention arm, between those who complied with screening and those who remained unscreened (see, e.g., Cuzick et al. J Med Screen 2007).
Dr Matejka Rebolj, Ms Nefeli Kouppa, Prof Peter Sasieni, and Prof Sarah Pinder, King’s College London, UK
Prof Sian Taylor-Phillips and Prof Keith Abrams, Warwick University, UK
Dr Adam Brentnall and Prof Stephen Duffy, Queen Mary University of London, UK
Prof Sam Janes, University College London, UK
Prof Robert Smith, American Cancer Society, USA
Prof Ruth Etzioni, Fred Hutchinson Cancer Research Center, USA