|
Abstract:
|
The statistical challenge of evaluating potential surrogates in a randomized trial is how to estimate the true relationship between biomarker and clinical endpoint since both biomarker and clinical endpoint are outcome variables. In this paper, we use direct and indirect effect concept (Pearl, 2002) to formalize the question, and propose a two-stage estimation approach to generalize the traditional instrumental variables method (Bowden, 1984). A simulation shows that our approach provides more consistent estimation than the proportion of treatment effect explained approach (Freedman, 1992). Our model assumptions are weaker than the sequential ignorability required by Freedman's approach. We also show the connection of our approach to the meta-analysis approach (Daniels, 1997) and the principal stratification method (Frangakis, 2002).
|
