Causal Inference in Neural Time Series

Nonparametric statistical frameworks for discovering causal structure in high-resolution neural recordings

Understanding how neural activity propagates across brain regions requires going beyond correlation — we need to identify directed, causal relationships from neural recordings.

The challenge. Neural time series (fMRI, EEG, ECoG, calcium imaging) are high-dimensional, non-stationary, and exhibit complex lag structures. Classical causal discovery methods make strong parametric assumptions that are often violated in neural data.

Our approach. I develop nonparametric causal inference frameworks tailored to neural time series:

  • CITS (Causal Inference in Time Series): A nonparametric statistical framework for high-resolution neural recordings (in review, Nature Communications). CITS enables principled discovery of directed functional circuitry without restrictive parametric assumptions, scaling to the thousands of neurons recorded by modern probes.

  • TPC (Time-Aware PC Algorithm): A consistent causal discovery method that explicitly models temporal structure in neural data, with formal statistical guarantees (PLoS Computational Biology, 2022; Statistics and Computing, 2023).

Applications. These methods have been applied to map directional functional connectivity differences in Alzheimer’s disease and Mild Cognitive Impairment from fMRI (Frontiers in Computational Neuroscience, 2023), and to characterize directed network dynamics in the mouse visual cortex from Allen Institute electrophysiology datasets.

References