Abstract
Excessive consumption of fast food (FF) is a primary driver of chronic disease, yet population-level dietary intake is notoriously difficult to measure. Traditional research relies on static home-based environment mapping or self-reported surveys, both of which are prone to significant bias. This study investigates whether mobility data from smartphones can serve as a more accurate indicator of FF intake, obesity, and diabetes. By analyzing 14.5 million food outlet visits from 243,644 anonymous smartphone users in Los Angeles County (LAC) and comparing them with the 2011 Los Angeles County Health Survey (LACHS), we demonstrate that mobility-based metrics are not only significantly associated with self-reported intake but are also superior predictors of obesity and diabetes compared to traditional self-reports.
Introduction
Fast food accounts for approximately 16% of the total caloric intake in the United States. Despite its impact on public health, measuring actual consumption at scale remains a challenge. Current methodologies often suffer from two major limitations:
- Static Environmental Assumptions: Researchers often assume individuals only eat near their homes, ignoring the "activity space" of commutes and workplaces.
- Recall Bias: Self-reported dietary data is frequently inaccurate due to under-reporting and memory errors.
This research explores the utility of "big data"—specifically passive smartphone geolocation pings—to bridge these gaps by providing a dynamic, objective measure of food environment exposure and behavioral patterns.
Methodology
To transform raw GPS data into health indicators, the researchers developed a multi-stage processing pipeline:
Data Sources
- Cuebiq Mobility Data (2016-2017): Anonymized GPS data from over 243,000 users in LAC.
- Foursquare API: Used to categorize over 53,000 food outlets, including 4,151 specific fast-food locations.
- LACHS (2011): A representative survey of 8,036 LAC residents providing ground-truth data on self-reported FF intake, obesity, and diabetes.
Feature Engineering
The study identified "stays" of at least five minutes, filtering out transient pings like traffic stops. Two primary metrics were calculated and aggregated at the census tract level (scaled 0–10):
- FF visits/time: The frequency of daily time periods (morning, midday, evening) where a user visited at least one FF outlet.
- FF visits/food: The ratio of FF visits relative to all food-related visits (e.g., grocery stores, full-service restaurants).
Statistical Analysis
Multinomial and binary logistic regressions were used to associate mobility features with health outcomes. Model performance was evaluated using Akaike weights (derived from the Akaike Information Criterion) to determine which indicators best fit the observed prevalence of obesity and diabetes.
Key Results
Association with Self-Reported Intake
Mobility-based metrics showed a strong, significant correlation with how residents described their own eating habits. For every 10% increase in the FF visits/time metric, the odds of a resident reporting "frequent" fast food intake increased by 35% (OR, 1.35; 95% CI, 1.28-1.42).
Predicting Chronic Disease
Mobility data proved to be a robust indicator of diet-related disease. A 10% increase in FF visits/time was associated with:
- 16% greater odds of obesity (AOR, 1.16; 95% CI, 1.12-1.21).
- 15% greater odds of diabetes (AOR, 1.15; 95% CI, 1.09-1.21).
Superiority Over Self-Reporting
Crucially, models utilizing mobility data outperformed those using self-reported intake when predicting health outcomes. For obesity, the probability that the mobility-based model was the "best fit" was 0.90, compared to a near-zero probability ($3.6 \times 10^{-7}$) for the self-report model. Similar results were observed for diabetes, suggesting that passive mobility data captures a "cleaner" signal of health-harming behaviors than human memory.
Discussion
The "Social Signal" of Mobility
Aggregated mobility data may capture the "social signal" of a neighborhood—the collective behavioral norms and environmental pressures that influence individual health. Unlike static maps, this approach accounts for the "dynamic" food environment, capturing consumption that occurs during commutes or near workplaces.
Overcoming Measurement Error
By relying on passive data collection, this method avoids the inherent biases of dietary surveys. While individual privacy is maintained through neighborhood-level aggregation, the data remains granular enough to provide actionable insights for public health surveillance.
Limitations
- Temporal Gap: The mobility data (2016-17) and health survey (2011) were collected several years apart. However, sensitivity analyses suggest the underlying associations remain robust.
- Ecological Fallacy: Neighborhood-level averages may not perfectly represent every individual within that tract.
- Demographic Bias: Smartphone users often skew younger and higher-income, though post-stratification weighting was used to mitigate this effect.
Conclusion
This study demonstrates that population-scale mobility data provides a powerful, passive indicator of fast food intake and diet-related disease. These metrics offer a scalable complement to traditional dietary surveillance, allowing researchers and policymakers to better understand how movement through the urban environment shapes public health outcomes. Future work may utilize these methods to redefine "food deserts" based on actual travel patterns rather than simple proximity to home.
Summary updated