Ruth Keogh - International Seminar Series in Causal Inference

Speaker: Ruth Keogh, London School of Hygiene & Tropical Health

Title: Estimating the effects of longitudinal treatment strategies with grace periods on time-to-event outcomes: going beyond clone-censor-weight

Time: September 5 from 10.00 – 11.00.

Location: University of Copenhagen: CSS 22-0-19, Bartholinsgade 2, København K

This presentation is part of the International Seminar Series in Causal Inference. The seminar is organized by the Pioneer Centre for SMARTbiomed and supported by the Danish Data Science Academy.

Abstract:  
This work focuses on longitudinal treatment strategies with ‘grace periods’, meaning strategies that require individuals to sustain a treatment over time, but allow them to have a gaps in treatment or delays before treatment starts. The motivating example is a study of individuals with a thoracic aortic aneurysm, where interest lies in the effects of receiving surgery within 12 months or not on overall and cause-specific mortality. In this example there is a grace period of 12 months within which individuals may receive surgery, and a challenge for estimation is that an individual’s observed treatment history can be consistent with more than one treatment strategy at some points in time. A popular approach to estimating the effects of treatment strategies with grace periods on time-to-event-outcomes using observational data is the clone-censor-weight approach. This involves making clones of the data corresponding to each treatment strategy, censoring individuals when they deviate from the strategy of interest, and applying inverse probability weights to account for the dependent censoring. The clone-censor-weight approach has the limitation that it relies on correct specification of the weights models. This work explores alternative doubly-robust approaches, which are more robust to modelling assumptions and which accommodate use of machine learning methods. I will describe recently proposed AIPW and TMLE approaches, and illustrate these for the motivating example, considering both overall mortality and competing events.

Webinar link: https://ucph-ku.zoom.us/j/65842758350