|
Abstract:
|
We consider estimation of an optimal dynamic treatment regimen for treating patients in response to their measured past, based on observing a sample of patients over time w.r.t. to their time-dependent treatment, time-dependent covariates, and final outcome of interest. We showcase a super-learner of the optimal dynamic treatment regimen, and a TMLE of the mean counterfactual outcome under the optimal dynamic treatment. We also develop a novel online estimator that allows for statistical inference in the difficult case that there is a positive probability on no treatment effect, in which case the target parameter is not regular. We discuss the analogue for optimal dynamic treatments under resource constraints.
Finally, we discuss some of our applications to learn the optimal rule for assigning blood-transfusions for trauma patients, involving a collaboration with UCSF trauma surgeons, and for learning the optimal rule for keeping HIV patients in care in Africa, involving a collaboration with UCSF HIV scientists. We also demonstrate the practical performance of the methods with some simulations.
Joint work with Alex Luedtke
|
