The Impact of Time Series Length and Discretization on Longitudinal Causal Estimation Methods
share
The use of observational time series data to assess the impact of multi-time point interventions is becoming increasingly common as more health and activity data are collected and digitized via wearables, social media, and electronic health records. Such time series may involve hundreds or thousands of irregularly sampled observations. One common analysis approach is to simplify such time series by first discretizing them into sequences before applying a discrete-time estimation method that adjusts for time-dependent confounding. In certain settings, this discretization results in sequences with many time points; however, the empirical properties of longitudinal causal estimators have not been systematically compared on long sequences. We compare three representative longitudinal causal estimation methods on simulated and real clinical data. Our simulations and analyses assume a Markov structure and that longitudinal treatments/exposures are binary-valued and have at most a single jump point. We identify sources of bias that arise from temporally discretizing the data and provide practical guidance for discretizing data and choosing between methods when working with long sequences. Additionally, we compare these estimators on real electronic health record data, evaluating the impact of early treatment for patients with a life-threatening complication of infection called sepsis.
READ FULL TEXT