Ascertaining cause of death among men in the Prostate Cancer Intervention Versus Observation Trial

Abstract

The Prostate Cancer Intervention Versus Observation Trial (PIVOT) randomized 731 men with localized prostate cancer to radical prostatectomy or observation. We describe the methods and results for cause-of-death assignments in PIVOT, and compare them to alternative strategies for ascertaining prostate cancer-specific mortality, as well as to the methods and results in the similar Scandinavian Prostate Cancer Group Study 4 (SPCG-4) trial. Three PIVOT Endpoints Committee members, blinded to randomized treatment assignments, reviewed medical records and death certificates when available to assign a cause of death using a primary and a secondary adjudication question. Initial disagreements were resolved through discussion. The level of initial agreement among committee members was examined, as well as guesses at randomized treatment assignments for a convenience sample of cases. Final cause of death determinations were compared to death certificates. Complete agreement on cause of death by all three committee members before any discussion was achieved in 200/354 (56%) cases on the primary and 209/354 (59%) cases on the secondary. However, complete agreement on the primary rose to 306/354 (86%) when 'definite' and 'probably' categories were collapsed, as planned a priori. The three committee members' proportions of correct guesses of randomized treatment assignment were 82/121 (68%), 113/148 (76%), and 99/134 (74%). Using the committee's final adjudications as a gold standard, death certificates had suboptimal sensitivities, specificities, or predictive values depending on how they were used to determine cause of death. There was no separate 'gold standard' by which to judge the accuracy of the final endpoints committee adjudications, and useful death certificates could not be obtained on about a third of PIVOT participants who died. The low level of initial agreement on cause of death among endpoint committee members and the potential for biased determinations due to partial unblinding to treatment assignment raise methodologic concerns about using prostate cancer mortality as an endpoint in clinical trials like PIVOT.