Individual causal effect estimation accounting for latent disease state modification among bipolar participants in mobile health studies

TL;DR

This paper introduces a novel N-of-1 causal inference method using an autoregressive hidden Markov model to account for latent disease states in bipolar disorder, improving effect estimation from mobile health data.

Contribution

It proposes an adapted autoregressive hidden Markov model for causal effect estimation that accounts for unobserved disease states in longitudinal mobile health data.

Findings

The method accurately identifies latent disease states.

It improves causal effect estimates over naive approaches.

Application to real data reveals state-dependent effects.

Abstract

Individuals with bipolar disorder tend to cycle through disease states such as depression and mania. The heterogeneous nature of disease across states complicates the evaluation of interventions for bipolar disorder patients, as varied interventional success is observed within and across individuals. In fact, we hypothesize that disease state acts as an effect modifier for the causal effect of a given intervention on health outcomes. To address this dilemma, we propose an N-of-1 approach using an adapted autoregressive hidden Markov model, applied to longitudinal mobile health data collected from individuals with bipolar disorder. This method allows us to identify a latent variable from mobile health data to be treated as an effect modifier between the exposure and outcome of interest while allowing for missing data in the outcome. A counterfactual approach is employed for causal inference and to obtain a g-formula estimator to recover said effect. The performance of the proposed method is compared with a naive approach across extensive simulations and application to a multi-year smartphone study of bipolar patients, evaluating the individual effect of digital social activity on sleep duration across different latent disease states.