|
Abstract:
|
Given a treatment and a clinical outcome, it is often of interest to seek a biomarker which can be measured shortly after the treatment is administered and used to predict the clinical outcome reliably. The principal stratification approach to assessment of these "surrogate" endpoints yields estimands which are generally not statistically identifiable from observed data. In this talk, we present two causal estimands in the context of a large randomized vaccine trial, where subjects may be infected before their immune response is measured. We describe how different sets of assumptions affect their identifiability and how the two estimands may be relevant to assessing the surrogate value of a biomarker. We also suggest a sensitivity analysis for estimation and testing of surrogate value, and show that such analyses may have low power, even when the sample size is very large.
|
