Multivariate point process treatments: Causal inference via instrumental variables

Neural recordings reveal rich patterns of dependence across brain regions, but identifying causal effects in the presence of unmeasured confounding remains challenging. In this talk, I will present an instrumental variable framework for studying the causal effects of multivariate point process treatments. Motivated by neuroscience experiments with external stimuli, we use the potential outcomes framework to define component-specific causal effects and establish conditions for their nonparametric identification through a system of convolution equations. We develop a nonparametric estimation procedure with a one-step correction and use robust covariance estimation for hypothesis testing and the construction of simultaneous confidence bands. An application to Neuropixels recordings from a visual discrimination task in mice examines how activity in visual cortical and thalamic regions affects hippocampal activity.