Feedback-augmented Non-homogeneous Hidden Markov Models for Longitudinal Causal Inference

Hidden Markov models are widely used for modeling sequential data but typically have limited applicability in observational causal inference due to their strong conditional independence assumptions. I introduce feedback-augmented non-homogeneous hidden Markov model (FAN-HMM), which incorporate time-varying covariates and feedback mechanisms from past observations to latent states and future responses. Integrating these models with the structural causal model framework allows flexible causal inference in longitudinal data with time-varying unobserved heterogeneity and multiple causal pathways. I show how, in a common case of categorical response variables, long-term causal effects can be estimated efficiently without the need for simulating counterfactual trajectories. Using simulation experiments, I study the performance of FAN-HMM under the common misspecification of the number of latent states, and finally apply the proposed approach to estimate the effect of the 2013 parental leave reform on fathers' paternal leave uptake in Finnish workplaces.