Toward a Data Processing Pipeline for Mobile-Phone Tracking Data

TL;DR

This paper introduces a statistical framework for processing mobile-phone tracking data, transforming raw position data into accurate mobility trajectories using a probabilistic model and particle Gibbs smoothing, improving upon existing algorithms.

Contribution

It proposes a formal statistical approach, including a probability model and particle Gibbs inference, to enhance trajectory estimation from noisy mobile tracking data.

Findings

The framework improves trajectory estimation accuracy over the binning algorithm.

It enables formal smoothing of noisy location data.

The method is suitable as a default processing tool for future studies.

Abstract

As mobile phones become ubiquitous, high-frequency smartphone positioning data are increasingly being used by researchers studying the mobility patterns of individuals as they go about their daily routines and the consequences of these patterns for health, behavioral, and other outcomes. A complex data pipeline underlies empirical research leveraging mobile phone tracking data. A key component of this pipeline is transforming raw, time-stamped positions into analysis-ready data objects, typically space-time "trajectories." In this paper, we break down a key portion of the data analysis pipeline underlying the Adolescent Health and Development in Context (AHDC) Study, a large-scale, longitudinal study of youth residing in the Columbus, OH metropolitan area. Recognizing that the bespoke "binning algorithm" used by AHDC researchers resembles a time-series filtering algorithm, we propose a statistical framework - a formal probability model and computational approach to inference - inspired by the binning algorithm for transforming noisy, time-stamped geographic positioning observations into mobility trajectories that capture periods of travel and stability. Our framework, unlike the binning algorithm, allows for formal smoothing via a particle Gibbs algorithm, improving estimation of trajectories as compared to the original binning algorithm. We argue that our framework can be used as a default data processing tool for future mobile-phone tracking studies.