Can we believe the DAGs?

Speaker: Odd O. Aalen, Professor, Oslo Centre for Biostatistics and Epidemiology, Dept. of Biostatistics, University of Oslo.

Abstract

Directed acyclic graphs (DAGs) play a large role in the modern approach to causal inference. An elegant theory of causal DAGs has been developed by a number of authors and got its authoritative presentation in Pearl's book on causality. The causal DAGs or causal Bayesian networks are based on a time-discrete network of measurements.

A mechanistic view of causality would focus on the actual process whereby effects come about. This process must be assumed to be continuous in time. How is the relationship between such a mechanistic model and observations taken at discrete times represented in a DAG? We illustrate this through some simple models. These demonstrate a difference between the continuous and discrete model. Typically, discrete observations of a process will obscure the conditional dependencies that are represented in the underlying mechanistic model of the process. 

An example using data from the Swiss HIV Cohort Study is presented. This example demonstrates that the results from the analysis of the data depend strongly on the time resolution of the measurements, and that coarse resolution may give a very incomplete picture of the true mechanistic relationships. Hence, there could be a contradiction between causality as presented in a causal DAG, and that presented in the correct mechanistic model.
 

Reference

Aalen, O. O., Røysland, K., Gran, J. M., Kouyos, R., & Lange, T. (2014). Can we believe the DAGs? A comment on the relationship between causal DAGs and mechanisms. Statistical methods in medical research

Published Jan. 18, 2016 4:25 PM - Last modified Jan. 18, 2016 4:30 PM