Data from smartphones and wearable devices provide rich longitudinal information on participants and allow for causal inference for daily exposures and outcomes. However, informative missingness, latent variables, unmeasured confounding, and uneven data collection rates are common in mobile health studies and may introduce bias. In addition, there are likely violations of stationarity, a key assumption for traditional longitudinal methods.
To overcome these challenges, we first propose an expectation maximization algorithm to adapt the conventional test for unit root non-stationarity to a context with missing data, and develop a sensitivity analysis for data missing not at random. Using our method, we identify a patient with bipolar spectrum disorder who has a unit root in their daily negative mood score data. We hypothesize the non-stationarity may result from the underlying latent disease states such as mania or depression, and thus we additionally develop a model to identify and control for latent modification and confounding.
Specifically, we propose a hidden Markov model for individual causal inference which handles missing data in the outcome through marginalization and multiple imputation. We compare the performance of our proposed model with a frequentist and a Bayesian implementation to a naive approach in a simulation and application to a multi-year smartphone study of bipolar patients. We employ the approach to evaluate the individual effect of digital social activity on sleep duration across different latent disease states.
Lastly, we employ functional data analysis methods to summarize overnight wrist actigraph data, to evaluate the role of sleep as a mediator between stress and positive mood. We demonstrate that functional principal component analysis identifies key information about sleep that is otherwise lost using a scalar representation of sleep duration.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/ht5n-y607 |
| Date | January 2024 |
| Creators | Fowler, Charlotte Rachel |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0019 seconds