Estimating Causal Effects from Panel Data with Dynamic Multivariate Panel Models

Abstract

Panel data are ubiquitous in scientific domains such as sociology and econometrics. Various modeling approaches have been presented for the causal inference based on such data, including Markov models, cross-lagged panel models, and their extensions. Existing panel data modeling approaches typically impose some restrictive assumptions on the data-generating process, such as Gaussian responses, effects that are constant in time, and ability to consider only short-term causal effects. With the goal of surmounting these restrictions, we present the dynamic multivariate panel model (DMPM) that supports time-varying, time-invariant, and individual-specific effects, multiple simultaneous responses across a wide variety of distributions, and arbitrary dependency structures of lagged responses of any order. We take a Bayesian approach and leverage state-of-the-art Markov chain Monte Carlo methods for the estimation of the posterior distributions of the model parameters and causal effects of interest. We demonstrate the use of DMPM by applying the model to both real and synthetic data.