For decades, understanding how a city breathes or how a population flees a disaster was a matter of "best guesses" and infrequent census surveys. Today, we use "digital breadcrumbs"—the pings from smartphone GPS, call detail records (CDRs), and app logs—to map human movement in near real-time.
While these digital proxies offer unprecedented granularity, they come with a fundamental flaw: they do not see everyone. Across 157 papers in this field, a clear consensus emerges. While researchers like Linus Bengtsson and Serina Chang have demonstrated that mobility data can predict disease outbreaks and disaster recovery, they also warn that the "digitally invisible"—the elderly, the impoverished, and the very young—are systematically excluded from the models meant to protect them.
The Mechanism of Digital Proxies
The core idea of modern mobility research is to treat human movement as a high-resolution signal. Instead of waiting for a ten-year census, researchers like Bita Sadeghinasr and Armin Akhavan show in Estimating Commuting Patterns from High Resolution Phone GPS Data that app-based GPS data can replicate city-scale commuting patterns with higher accuracy than traditional cellular tower methods.
This signal is processed through increasingly sophisticated architectures:
- Tensor Factorization: Jingyuan Wang developed the NR-cNTF model, which treats a city as a multi-dimensional cube (Origin $\times$ Destination $\times$ Time) to reveal hidden rhythms in urban development.
- Foundation Models: In Pretrained Mobility Transformer: A Foundation Model for Human Mobility, Haoyu He applies the "language model" approach to movement, training transformers on millions of trajectories to "read" the city and predict future locations.
- Collaborative Filtering: When data is sparse (e.g., a user only pings a tower when making a call), zipei fan uses spatio-temporal filtering to "fill in the blanks" of a user's path.
The Digital Divide as Selection Bias
The intuition behind the "digital divide" in mobility data is simple: a model is only as good as its training set. If the training set is "smartphone users," the model inherently reflects the behaviors of the affluent and the tech-savvy.
Samuel P. Fraiberger and colleagues highlighted this in Uncovering socioeconomic gaps in mobility reduction during the COVID-19 pandemic using location data. They found that in developing nations, the most impoverished individuals often lack mobile devices entirely. Consequently, the "poorest" group visible in the data is actually more affluent than the absolute poorest segment of the population.
This selection bias isn't just a statistical nuisance; it creates a "blind spot" for policy. Esteban Moro and Dan Calacci argue in Mobility patterns are associated with experienced income segregation in large US cities that relying on smartphone data overlooks informal gathering spaces and the movement of those with lower digital footprints, potentially reinforcing urban segregation.
Why Representation Matters in Crisis
The stakes of this bias are highest during disasters and outbreaks. Linus Bengtsson demonstrated in Improved Response to Disasters and Outbreaks by Tracking Population Movements with Mobile Phone Network Data that while SIM card tracking is vastly superior to "best-guess" relief methods, it risks missing the elderly and children—the very groups most vulnerable to cholera or displacement.
Similarly, Takahiro Yabe and Naoya Fujiwara discovered universal scaling laws for how populations return home after disasters. However, they caution that these "universal" laws are derived from mobile users. If the elderly return home more slowly due to limited mobility or lack of resources, a model biased toward younger smartphone users will overestimate a community's recovery speed.
Toward a Human-Centered Resilience
To fix this, researchers are moving toward "functional criticality." Ali Mostafavi and Chenyue Liu propose in Quantifying Functional Criticality of Lifelines that we should measure infrastructure not by its physical size, but by how much people rely on it.
However, they emphasize that this reliance is demographic-specific. A low-income resident might rely on a specific bus route that doesn't show up as "critical" in a dataset dominated by taxi trajectories or private vehicle GPS. As Mohamed F. Mokbel argues in his vision paper, the future of the field lies in "Mobility Data Science"—a specialized pipeline that explicitly accounts for geographic and demographic biases to ensure that urban planning and disaster response do not leave the digitally invisible behind.