Professor David Stephens - Semiparametric Bayesian Inference
Speaker: David Stephens, McGill University
Title: Semiparametric Bayesian inference for optimal dynamic treatment regimes via dynamic marginal structural models
Time: October 15 from 14.00 - 15.00, followed by coffee and cake
Location: University of Copenhagen: CSS 2.0.12, Øster Farimagsgade 5, 1353 København K
Registration: All are welcome, no registration needed.
This presentation is part of the International Seminar Series in Causal Inference. The aim of the seminar series is to bring distinguished causal inference speakers to Copenhagen and to foster new connections among local causal inference researchers across different disciplines and institutions. The seminar is therefore accompanied with two additional opportunities for connections:
- An informal opportunity to stay for coffee and cake after the presentation, where you can chat with the speaker and other participants. We encourage all to participate, no registration needed.
- A possibility to book a one-on-one meeting with David Stephens on October 15. If you are interested in this, please contact Erin Gabriel (erin.gabriel@sund.ku.dk).
To support interaction and community building, we encourage all who have the option to participate in person. If this is not possible for you, you contact Michael Sachs (michael.sachs@sund.ku.dk) to obtain a link for virtual participation.
The seminar is organized by the Pioneer Centre for SMARTBiomed and supported by the Danish Data Science Academy.
Abstract: Identifying dynamic treatment regimes (DTRs) is a key objective in precision medicine. Value search approaches including Bayesian dynamic marginal structural models offer an attractive approach to estimation by mapping candidate regimes to their expected outcome. As parametric models for the expected outcomes may be mis-specified and lead to incorrect conclusions, a grid search over candidate DTRs has been proposed, but this may be computationally prohibitive and also subject to high uncertainty in the estimated value function. These inferential challenges can be addressed using Gaussian Process (GP) optimization methods with estimators for the causal effect of adherence to a specified DTR. We demonstrate how to identify optimal DTRs using this approach in a variety of settings including when the value function is multi-modal and show that the GP modeling approach that recognizes noise in the estimated response surface leads to improved results as compared to a grid search approach. Further, we show that a grid-search may not yield a robust solution and that it often utilizes information less efficiently than a GP approach.