School of Mathematics and Statistics - Theses
Permanent URI for this collection
Search Results
1 - 1 of 1
-
ItemA novel method for the G-estimation of Structural Nested Accelerated Failure Time Models with time-varying treatment effects( 2024-02)Conventional methods for the analysis of time-to-event data with right censoring may fail to produce unbiased estimates of the causal effect of always-treated versus never-treated in the target population represented by data collected from a sample of study participants who frequently switch treatments. One promising framework, based on the potential outcomes approach to causal inference, seeks to overcome these deficiencies by specifying a structural nested accelerated failure time model (SNAFTM) and estimating the acceleration factor using an algorithm known as Gestimation. These models assume an accelerated failure time relationship between an individual’s observed time-to-event and their potential outcome, the time-to-event that they would have experienced under the never-treated regimen. This thesis provides a thorough derivation of the SNAFTM proposed by Robins 1992, used in Witteman et al. 1998 and explored more recently in Joffe et al. 2012, Vock, Durheim, et al. 2016, Vansteelandt and Joffe 2014, Picciotto, Ljungman, et al. 2016 and Mertens and Vansteelandt 2018. A novel estimation routine for SNAFTM’s, based on the Newton-Raphson algorithm, is proposed. This approach can accommodate the need for artificial censoring and treatment effects that vary with calendar time, time-since-diagnosis and time-on-treatment. The frequentist properties of estimates produced by this implementation of G-estimation are explored. The new method is applied in a comparison of modes of haemodialysis for the treatment of end stage kidney disease using data from the Australian and New Zealand Dialysis and Transplant Registry (ANZDATA).