John Williams, Department of Marketing

Structural Causal Modelling: A revolution?

According to Turing Award-winning computer scientist Judea Pearl, a revolution is happening in quantitative data analysis, due to achievements in mathematical theory pertaining to structural causal models (SCMs). A SCM encodes a researcher's theoretical causal model into a directed acyclic graph (DAG). This enables the analysis of the logical implications of the model, in particular, showing which relationships are identified (i.e. able to be calculated with a unique solution) and which are not. This capability leads to the solution of the problems of (a) confounding; and (b) spurious correlation. Even more, Pearl has shown that under some mild assumptions, (c) analysis of observational, cross-sectional data can warrant causal inference at the same level as analysis of data generated from a randomized controlled trial (RCT).

To most data analysts, these features are very exciting, or even unbelievable. When I first read about this framework I was very sceptical, however I have become a cautious convert (but not an expert). So in this seminar I will explain the logic behind these claims and show a few examples that I hope will motivate you to learn more.

Note: this framework is not about analysing data, but analysing causal models, i.e. DAGs. The causal models are a given, i.e. specified by the researcher, analyst or theorist. No special software is needed to analyse data that (presumably) reflects the model. Most importantly, the framework does not allow "proving" causality. So leave your philosophical arguments citing David Hume and Karl Pearson at the door please.

Last modified:

This page is maintained by the seminar list administrator.