Background: In cancer studies progression-free survival (PFS) is becoming a very important endpoint in the development of new therapeutic agents. Two methods of determining progression are typically used: (1) the local radiologist evaluates scans and (2) scans are reviewed by an independent blinded (central) reviewer. The second method is considered to be the reference standard but is expensive, time consuming, and logistically difficult. The first method has measurement error associated with it, but, it is less expensive and easier to obtain.
Purpose: This article explores a new method for analyzing PFS data.
Methods: When PFS data using the test with measurement error are analyzed, inferences about covariate effects may be invalid due to bias. A sampling strategy is evaluated where data are collected on a subset of subjects using the reference test and on all subjects using the test that has error. The strategy is designed to maintain valid inferences while requiring the more expensive or difficult test on a small proportion of patients. In the analysis of the data we incorporate subject-specific and time-dependent covariates into the diagnostic errors (sensitivity and specificity) of the tests. We also propose a modeling formulation that accounts for unobserved covariate affects on diagnostic error through a shared random effect. We explore the effect of different diagnostic test properties on inference via simulation and use the methodology to analyze a renal cancer example.
Results: The simulations show inference is correct when a subset of measurements without error are collected.
Limitations: When the sensitivity and specificity of the local review is low a large fraction of centrally reviewed tests are needed to have high efficiency.
Conclusions: When designing a study where PFS is the primary endpoint collecting centrally reviewed data on a subset of patients may provide a valid an more feasible approach than collecting centrally reviewed data on all patients.