Department Seminar by Marzieh Mussavi Rizi

Please Note: This seminar will be given online.

Dynamic Treatment Regimes and Interference

Identifying interventions that are optimally tailored to each subject is often of interest in areas such as precision medicine, economics, and political science. DTRs are sequences of decision rules that take individual patient information as input and output treatment recommendations. The analyst assesses the counterfactual effect of several candidate treatment policies and selects the one with the best expected outcome, under a specific set of stated assumptions. One of these usually specifies that individuals’ treatment do not impact the outcome of other individuals, which is the so-called no interference assumption. However, the propensity for humans to form intricate social networks makes this assumption questionable in many  applications. Consequently, imposing this assumption when it, in fact, does not hold, may impact the validity of our resulting DTR. In this talk, we demonstrate that even when conventional models are correctly specified and unbiased estimators are obtained, ignoring the nature of interference—even in the simple case of dyadic, or pairwise interference—can lead to non-optimal or even detrimental DTRs. Finally, we develop a methodology that overcomes this shortcoming and demonstrates an ability to find optimal DTRs in this context.