Methods & Meta-science

A gentle introduction to the Partial Information Decomposition

In many cases in neuroimaging or data analysis we evaluate a statistical dependence in many variables, and find effects in more than one. For example, we might find an effect between a presented stimulus and neural responses in different spatial regions, time periods or frequency bands. A natural question in such situations is to what extent is the statistical relationship in the two responses common, or overlapping, and to what extent is there a unique effect in one response that is not related to the other response. An information theoretic framework called the Partial Information Decomposition (PID) has been proposed to address these questions. The first part of the session will be a gentle introduction to information theoretic quantification of statistical interactions, introducing co-information, redundancy, synergy and the basic theory of the PID, as well as introducing some applications (including, interactions between neural responses, interactions between multi-modal stimulus features in speech, interactions between neural responses and behaviour and predictive model comparison). The second part of the session will go into a bit more detail into the details of the implementation of the PID including the theory and computation of the Iccs redundancy measure, and more discussion of issues such a misinformation (negative unique information), applications etc. There will be a break between the two parts to give people the chance to opt-out of the more technical second part.