A Principled Approach to Multiple Causal Inference

Abstract

Causal inference was born out of randomized control trials. But randomized experiments are often impractical, if not expensive or untenable, and fail to make use of large observational studies. Modern machine learning methods stand diametrically opposite: they find idiosyncratic correlations between features in data but do not reveal causal effects. Any hybrid approach to blend the two worlds rests on strong assumptions. Traditionally, causal inference has focused on single causes. Here, however, we develop and apply a principled approach to multiple causal inference. We first acknowledge recent developments (Wang et al., 2019; Ranganath et al., 2018) in multiple causal inference by distilling their assumptions and contributions. This review supports other expositions found throughout the thesis. But prior work uses directed mutual information exogenously. We borrow from the gambling literature to show how directed information emerges as the solution to an operational story instead of an ad hoc demand. This theoretical contribution justifies the methodology that follows. Next, recent progress focuses on causal estimation. We develop a testing counterpart to the literature that identifies significant causes while preserving false discovery rates. Finally, we show a novel application of the toolkit for understanding and intervening on the subtle physical mechanisms that cause tropospheric ozone peaks.