Learning Markov Processes with Latent Variables From Longitudinal Data

Abstract

We present a constructive proof of (nonparametric) identication of the parameters of a bivariate Markov chain when only one of the two random variables is observable. This setup generalizes the hidden Markov model in various useful directions, allowing for state dependence in the observables and allowing the transition kernel of the hidden Markov chain to depend on past observables. We give conditions under which the transition kernel and the distribution of the initial condition are both identied (up to a permutation of the latent states) from the joint distribution of four (or more) time-series observations.

Keywords

JEL codes

Reference

Ayden Higgins, and Koen Jochmans, “Learning Markov Processes with Latent Variables From Longitudinal Data”, TSE Working Paper, n. 22-1366, September 2022, revised May 2024.