A. Garcia Hernandez, T. Perez Perez, D. Rizopoulos, M. C. Pardo Llorente
In clinical trials, when the response is not observed after treatment discontinuation, we face a missing not at random problem that can be addressed using reference-based imputation or delta adjustment. The most popular methodology uses multiple imputation.
We propose a new formulation to tackle this problem fully analytically. We model the process of interest (Z) using its association with a) time to drug discontinuation T and b) the longitudinal response in a hypothetical world without drug discontinuation Y. We use the equation Z(t) = Y(t) + Δ(t)·I(t > T) where Δ(t) is used to quantify the unobserved difference between Y(t) and Z(t) after T. We translate six popular imputation rules into a Δ(t) and build close-equations estimators for the marginal means of Z.
Our model allows implementing a broad range of rules. It is faster and more efficient than multiple imputation and avoids standard error over-estimation − known issue for the solution based on multiple imputation.
Keywords: Missing not at random, reference-based imputation, delta adjustment.
Scheduled
Biostatistics I
June 9, 2022 10:10 AM
A13