Speaker: Prof Mats Julius Stensrud, Chair of Biostatistics at EPFL and Director of the Doctoral School in Mathematics at EPFL
Abstract: Individualizing treatments is justifiable if it results in better (expected) utility in the population. Thus, specifying utility functions -- the mapping of a individual's outcomes and other features to some number -- forms a basis for determining preferences among treatment options. However, explicit consideration of utilities is often avoided in the statistics and causal inference literature, perhaps because of its philosophical and subjective reputation. Yet, decision makers and methodologists will usually agree, at least superficially, with the classical Hippocratic maxim: ``First, do no harm''. An emerging literature offers an operationalization of ``harm'' that promises to possibly `provide the key for next-generation personalized decision making' (Mueller and Pearl, 2023), see also Ben Michael et al. (2022). Their methods propose strategies for supplementing experimental data with confounded non-experimental data that are often disregarded in classical approaches. We demonstrate that their proposals take for granted a particular operationalization of ``harm'', and we contrast it with notions of harm defined by alternative parameters, which can be identified with assumptions that are empirically testable. To bring clarity, we formally articulate two approaches, which we refer to as counterfactual and interventionist, respectively. These two approaches are distinguished by utility functions of fundamentally different natures, with concrete implications for practical decision making. In particular, they each correspond to different notions of ``harm'' and to different interpretations of the Hippocratic principle. We apply our results to study two previous examples from the optimal regimes literature, illustrating that the different approaches lead to different decisions.
This will be a free hybrid seminar. To register to attend remotely, please click here: https://cam-ac-uk.zoom.us/meeting/register/9TMh_9qJQUme4H_ljLQwog