Heidelberger Kolloquium

Schedule of event:

13/11/2017, 16:15

More event dates:

For this event there are no more dates.


Annette Kopp-Schneider, C060


Dr. Dominik Janzing, Max-Planck-Institut für Intelligente Systeme, Tübingen




Causal inference from non-interventional data -- foundations and recent approaches

Inferring the causal structure between n random variables X_1,...,X_n from passive observations (that is, when no interventions are possible) is a challenging task. Since the 1990s there is an increasing community of researchers that believe that causal inference is possible under appropriate assumptions. They have developed methods that rely on conditional statistical (in)dependences, based on the Causal Markov Condition and Causal Faithfulness [1,2]. More recently, it has been shown that not only statistical independences but also other properties of distributions reveal causal information [3]. We have applied some of the methods successfully in brain research [4]; evaluations of the methods on real data with known causal structure also show encouraging results [5].

The methods can be justified by a principle that states the following asymmetry between cause and effect: P(cause) and P(effect|cause) contain no information about each other, while P(effect) may contain information about P(cause|effect) and vice versa -- in a sense that can be formalized in various ways [3,6]. Motivated by this principle, there are also new approaches to the difficult problem of detecting hidden common causes in multivariate regression models [7].

[1] Spirtes, Glymour, Scheines: Prediction, Causation, and Search, 1993
[2] Pearl: Causality, 2000
[3] Peters, Janzing, Schoelkopf: Elements of causal inference, MIT Press 2017
[4] Grosse-Wentrup, Janzing, Siegel, Schoelkopf: Identification of causal relations in neuroimaging data with latent confounders: An instrumental variable approach, NeuroImage 2016.
[5] Mooij, Peters, Janzing, Zscheischler, Schoelkopf: Distinguishing cause from effect using observational data: methods and benchmarks, JMLR 2016
[6] Janzing, Schoelkopf: Causal inference using the algorithmic Markov condition, IEEE TIT 2010
[7] Janzing, Schoelkopf: Detecting confounding in multivariate linear models via spectral analysis, JCI 2017


K13 Marsilius-Arkaden, Turm Süd

 back to overview
to top