A Statistical Framework for Harnessing Human Activity Data to Understand Behavior, Health, and Well-Being

Abstract

As smartphone and other sensor-based tools for obtaining human activity and ecological momentary assessment (EMA) information become more common, the volume of resulting data will increase exponentially. At the same time, statistical methodology for analyzing these data lag significantly behind data collection methods. This dissertation concerns the development of an overarching statistical framework for analysis of smartphone- and other sensor-collected human activity and behavior data. Within this framework, powerful statistical tools are developed and presented for (1) intelligently borrowing information from other smartphone users to obtain high-precision personalized inference via an extension of multi-source exchangeability models, a method from the clinical trials literature; (2) analysis of human activity patterns by exploiting the sequential and temporal features in smartphone data via a sequence distance framework borrowed from bioinformatics; and (3) Monte Carlo simulation, imputation, and prediction of human activity patterns. Each method is rigorously evaluated and applied to real data obtained from Daynamica, a smartphone application for fusing sensor-based activity tracking with smartphone-enabled EMA. Numerous avenues for future research are presented to further advance this nascent subdomain of statistics.