Tutorial: An introduction to the different causal frameworks in neuroimaging

Neuroimaging techniques allow to measure brain functions and to establish associations with cognitive functions. Yet, associations per se do not inform about the causal mechanisms that govern how brain functions give rise to cognitive functions. To advance our understanding of the human brain, it is essential to move beyond associational studies and identify the causal mechanisms that form the basis of cognition.

This tutorial introduces the participant to the most common notions of causality and enables her to form an opinion about the strengths and weaknesses of each framework. We will discuss both the potential and current limitations for causal inference in neuroimaging, which is crucial for informing the design of subsequent studies. Furthermore, we will present methods that can|readily and with little or no extra work|enrich a standard (associational) neuroimaging analysis with warranted causal hypotheses.

Moritz Grosse-Wentrup and Sebastian Weichwald
MPI for Intelligent Systems, Tubingen, Germany

Program Outline
This tutorial will be two hours in duration, presented on Wednesday, 21 June 2017.

1. The foundations of the different causal frameworks
We will discuss the theoretical foundations of causal inference and the problems one faces when conceptualising a consistent notion of causation. In particular, we give a thorough introduction to the following causal frameworks commonly used in neuroimaging and discuss their (untestable) assumptions and respective pros and cons: (a) Potential Outcomes, (b) Granger Causality, (c) Dynamic Causal Modelling, and (d) Causal Bayesian Networks.

2. Causal inference in practice
In this part of the tutorial we demonstrate how the frameworks above, with all their subtle theoretical caveats, can fruitfully be applied in practice. While the STATS101 slogan ‘correlation does not imply causation’ is true in its generality, the goal of this tutorial is to give concrete examples how cause-effect relationships
can be inferred from neuroimaging data. We encourage people not to be scared away by the intricacies of causal inference by providing ready-to-use methods (and explicitly discussing the required assumptions) that can enrich standard (correlational) neuroimaging studies by warranted causal statements.

Suggested Prerequisites and Materials