Time series causality inference
using the Phase Slope Index
Florin
Popescu and Guido Nolte, Fraunhofer Institute FIRST, Berlin,
Germany
A method recently introduced by Nolte et. al (Phys Rev Lett 100:23401,
2008)
estimates the causal direction of interactions robustly with respect to
instantaneous mixtures of independent sources with arbitrary spectral
content, i.e. in observations which are dominated by non-white
spatially
correlated noise and in which dynamic structural interaction plays
little
part. The method, named Phase Slope Index (PSI), is unlikely to assign
causality in the case of lack of dynamic interaction among time series,
unlike Granger causality for linear systems. Results show that PSI does
not
yield false positives even in the case of nonlinear interactions. The
meaning of instaneous noise mixtures in different data domains will be
discussed in the context of correct correlation vs. causation
inference, and
the theoretical relationship of PSI to other time-series causality
inference
methods will be expanded upon.
[
NIPS 2009 Causality and Time Series
Mini-Symposium]