Biostatistics Seminar Series: Daniel S. Malinsky

Abstract: To draw reliable conclusions about treatment and policy decisions we rely on assumptions about the background causal processes which generated our data. In highly-controlled settings, such as clinical trials, the requisite causal assumptions are typically satisfied by design. However, when dealing with complex but less structured data – such as that from electronic health records, biological processes that are not well-understood, and some observational studies – we may have little reason to trust convenient causal assumptions a priori. An important challenge in these settings is to select causal models on the basis of the data itself as much as possible. By representing causal models with graphs (DAGs or more complicated structures), a growing body of research has produced algorithms that exploit the statistical implications of candidate graphical structures to perform principled causal model selection: distinct graphs will imply (sometimes) distinct patterns of conditional independence and dependence, and under some assumptions we can perform a kind of "pattern matching" to select a set of candidate graphs consistent with observations. Subsequently, the selected models can be used to support identification of causal effects and (if they are identified) efficient estimation. This talk will include a brief introduction to graphical causal model selection and focus on a novel method which has been developed to make model selection possible from nonstationary time series data. Time-permitting, I will also discuss related work on respecting (causal) fairness constraints in automated decision-making systems and efficient inference in settings with data missing not-at-random.