TL;DR
Researchers investigated whether Facebook’s targeted advertising reach estimates can serve as a "digital census" for tracking global lifestyle diseases like obesity, diabetes, and smoking. While the data proves too volatile and confounded for reliable cross-country comparisons, it excels at revealing demographic health trends within individual countries that closely mirror traditional "gold standard" statistics.
What Problem Does This Paper Address?
Non-communicable "lifestyle" diseases—such as heart disease, stroke, and diabetes—are the world's leading causes of death, yet tracking their prevalence is a slow, expensive, and logistically complex task for organizations like the WHO.
Social media offers a tantalizing alternative: a real-time, demographically rich data source. However, most previous research has focused on single countries (like the US) and hasn't accounted for the "black box" nature of social media algorithms. This paper asks: Can we use the Facebook Marketing API to build a global health surveillance system, and more importantly, can we trust the data it provides?
Methods
The authors utilized the Facebook Marketing API, which allows advertisers to see "reach estimates" (the number of monthly active users matching specific criteria) before they ever spend a dime.
- Interest Selection: They defined "marker interests" directly related to health (e.g., "Obesity awareness," "Insulin resistance," "Smoking cessation").
- Placebo Controls: Crucially, they introduced "placebo interests"—generic topics like "Reading," "Technology," or "Facebook"—to act as a baseline. This helped determine if a correlation was truly about health or just a reflection of general social media activity levels.
- Normalization: To account for varying Facebook penetration and activity levels, they calculated the fraction of users with a specific interest and used z-scores to compare demographic subgroups (age and gender) within countries.
- Validation: They compared these estimates against "ground truth" data from the World Health Organization (WHO) across 47 countries.
- Temporal Analysis: They re-collected data five months later to test the stability of Facebook’s estimates over time.
Datasets
- Facebook Ads Audience Estimates: Aggregated data for 47 countries, segmented by gender and eight age groups (18–59).
- WHO Global Health Observatory: Ground truth statistics for diabetes (fasting blood glucose), smoking prevalence, and obesity (BMI ≥ 30).
- pySocialWatcher: The authors developed and used this generic Python package to automate the collection of audience estimates from the Facebook API.
Key Insights and Ideas
- The Placebo Effect: One of the most striking findings was that "placebo" interests often outperformed targeted health interests in cross-country models. For example, general interest in "Fitness & Wellness" correlated more strongly with obesity rates than specific "Obesity awareness" tags. This suggests that Facebook's interest categories are heavily influenced by latent socioeconomic factors.
- Within-Country Strength: While the API struggled to compare absolute disease rates between, say, Brazil and Sweden, it was highly effective at showing relative trends within a country.
- Demographic Mirroring: The data accurately reflected known global health disparities. For instance, Facebook data showed men have a much higher interest in tobacco-related topics than women, which aligns with WHO data showing that roughly 40% of men smoke compared to 9% of women.
- The "Black Box" Danger: The authors highlight that Facebook’s internal algorithms for defining "interests" are opaque and prone to sudden shifts. An interest group might grow not because more people are interested in it, but because Facebook updated its Natural Language Processing (NLP) pipeline.
Key Results
- Weak Cross-Country Correlation: Only a few targeted interests, like "Dieting" ($r = 0.314$), showed significant correlation with WHO obesity reports.
- High Temporal Volatility: When comparing data from October 2016 to March 2017, the researchers found significant instability. For 45% of the age-group comparisons, the correlation between the two time periods fell below 0.70, and in 17% of cases, the ordering of age groups actually flipped.
- Age-Related Trends: The data successfully captured the transition of health interests over a lifespan. Younger users showed higher interest in "Hookah" and "Cigarettes," while older users shifted toward "Lung cancer awareness" and "Smoking cessation."
Limitations and Future Work
The primary limitation is the opacity of the Facebook Marketing API. Because it is a proprietary system, researchers cannot account for changes in how Facebook classifies users. Furthermore, the API does not provide historical data, making it difficult to distinguish between seasonal trends and actual changes in public health awareness.
The authors suggest that future work should involve triangulating this data with direct surveys run through the advertising platform to validate if the "reached" audience actually possesses the interests Facebook attributes to them. Despite the limitations, they argue that for within-country demographic analysis, Facebook Ads remain a powerful, low-latency tool for public health researchers.
Summary updated