Graphical Causal Models and Inferences to Mechanisms from Brain Imaging: Possibilities and Limitations

Clark Glymour
Carnegie-Mellon University

Many recent studies have applied statistical techniques to imaging data to attempt to infer causal cascades and feedbacks among regions
of the brain in participants performing simple cognitive or motor tasks. Statistical procedures include regression, specification of linear structural equation models, Granger causal models from time series, and other methods. In parallel, work in computer science and statistics since the 1980s has developed an abstract formalism--graphical causal models, sometimes called "causal Bayes nets"--that represents non-parametrically the constraints on
probability distributions implied by causal claims imbedded in any of a wide class of statistical models and methods, including those above.
The graphical causal model representations lend themselves to a number of principled, automated search methods developed in the computer
science literature in the last two decades, few of which have been applied to imaging data. An understanding of the abstract formalism permits an analysis of what can and cannot be learned from
fMRI and other imaging data by available methods. This talk will provide an overview of all of the above.

Audio (MP3 File, Podcast Ready) Presentation (PowerPoint File)

Back to Summer School: Mathematics in Brain Imaging