|
Abstract:
|
The FDA clinical outcomes assessment (COA) guidance outlines the need to evaluate meaningful change via stratified CDFs for change scores. Stratification is made across treatment arms or proxy effect groups, referred to in the literature as “Anchor groups”. FDA guidance requires descriptive assessments of the difference between stratified CDFs. This omits the strength of inference. Moreover, in the context of certain trial designs (e.g. progression-free survival) subject attrition can arise from biomarker deterioration. In this context missing data satisfies the MAR assumptions, inducing bias in descriptive statistics characterizing change scores. An event/trial binomial model is presented permitting inference on the magnitude of differences between stratified CDFs. In addition, this procedure can incorporate covariates such as biomarkers driving missing data, to produce unbiased estimates of meaningful change. Simulated data are used to demonstrate the bias of descriptive estimates in the presence of MAR data and the robust nature of model-based estimates. This offers a way to satisfy regulatory requirements while taking advantage of modern statistical methods.
|
