History-dependence shapes causal inference of brain-behaviour relationships

TL;DR

This paper explores how history-dependence in neural and behavioral data affects causal inference, emphasizing the inherent uncertainty in determining true causal relationships due to open system dynamics and confounding variables.

Contribution

It demonstrates that uncertainty about history-dependence is fundamental and cannot be fully resolved, impacting how causality should be assessed in neuroscience.

Findings

Uncertainty in history-dependence is inherent in open systems.

Assumptions about history-dependence can cause spurious correlations.

Interventions require knowledge of latent dynamics for effective causal inference.

Abstract

Behavioural and neural time series are often correlated with the past. This history-dependence may represent a fundamental property of the measured variables, or may arise from how confounding variables change over time. Here we argue that undecidability about the ground-truth of history-dependence is a general computational property of systems that exchange information with its environment, and show that the resulting uncertainty has a direct impact on causal inference. We first argue that uncertainty in the ground truth of history-dependence is an inherent property of open systems that cannot be explicitly falsified. Simple model systems are then simulated to show how different assumptions about history-dependence can lead to spurious correlations and statistical properties of data distributions that are typically unaccounted for. We then consider this problem from an interventionist perspective, showing that interventions can only be guaranteed to remedy the spurious correlation problem when the latent dynamics between the intervention and measured processes are known a priori, and the effect of the intervention is invariant at the chosen level of analysis. We conclude that uncertainty about history-dependence is a fundamental property of the study of neural systems, and in light of this discuss how causality should be assessed in neuroscience.