Social Media Data for Population Mapping: A Bayesian Approach to Address Representativeness and Privacy Challenges

Technical Summary: Social Media Data for Population Mapping: A Bayesian Approach to Address Representativeness and Privacy Challenges

Problem Statement
Accurate and timely population data are critical for disaster response and humanitarian planning, yet traditional census data often fail to capture rapid demographic changes or short-term spatial mobility. While social media data (specifically Facebook user counts) offers a promising alternative for dynamic population monitoring, its utility is constrained by two primary challenges:

1. Representativeness: Social media users constitute a skewed subset of the total population, and the relationship between observed user counts and true population stocks varies significantly across socio-economic and geographical contexts.
2. Privacy-Induced Missingness: To protect user privacy, Meta (Facebook) applies differential privacy techniques that censor low-count data. Specifically, user counts below a threshold (10 users) are masked or censored. The authors find that this censoring disproportionately affects low-population and rural areas, leading to significant data gaps (approximately 1.7% of municipalities have no counts, and >6.6% have >80% missing tiles) and introducing bias that standard smoothing or interpolation methods fail to address.

Methodology
The study employs a Bayesian framework to address these challenges using the Philippines as a case study, leveraging 2020 census data and Facebook user counts from May 4, 2020 (selected to align with census reference dates during COVID-19 lockdowns). The workflow consists of three main stages:

1. Bayesian Imputation for Censored Data:
   To recover missing values caused by privacy censoring, the authors develop a Bayesian Poisson count hierarchical model. Unlike simple imputation methods, this approach treats missing values as informative rather than arbitrary. The model utilizes the full distribution of user counts over time (March–December 2020) for each tile. It distinguishes between observed counts and censored counts (those below the threshold of 10) using a likelihood function that incorporates the Poisson cumulative distribution function for censored observations. Tiles with only censored values are grouped by their frequency of appearance in the dataset to infer a shared distribution, allowing for the recovery of missing counts in rural areas.

2. Statistical Modeling of User Proportions:
   Using the imputed data, the authors model the proportion of Facebook users in each municipality (Bayan) as a function of several predictors:

   • Degree of Urbanisation (DUC): Categorized as Rural, Peri-urban, and Urban.
   • Working-age population proportion (15–65 years).
   • Nighttime light intensity (satellite radiance).
   • (Note: Network usage density was analyzed but excluded from the final models due to high collinearity with other predictors and data sparsity).

   Three model variants were developed and compared:

   • $M_{bin}$: A hierarchical binomial model.
   • $M_{betabin}$: A beta-binomial regression model introducing an overdispersion parameter ($\rho$) to account for variance exceeding the binomial assumption.
   • $M_{full}$: The full model incorporating the overdispersion of $M_{betabin}$ plus a spatially explicit component using Hilbert space approximate Gaussian Processes (HSGPs) to capture spatial correlations between neighboring administrative areas not explained by the predictors.

3. Validation:
   The models were trained on 80% of administrative areas (stratified by DUC) and tested on the remaining 20%. Performance was evaluated using Absolute Error of the Median (AEMed), Squared Error of the Mean (SEMean), and Continuous Ranked Probability Score (CRPS).

Key Results

• Imputation Efficacy: The Bayesian imputation approach successfully recovered data coverage for 5.5% of rural municipalities (323 Bayans) that previously had no user counts. This process corrected the bias where low-density areas were systematically underrepresented.
• Model Performance: The full model ($M_{full}$), which accounts for both overdispersion and spatial correlation, demonstrated the best generalizability.
  • Urban Areas: Achieved an AEMed of ~24% and CRPS of ~18%.
  • Rural Areas: Achieved an AEMed of ~39% and CRPS of ~27%.
  • The inclusion of overdispersion and spatial components significantly improved the calibration of credible intervals, particularly for rural areas where user rates are low and variance is high.
• Predictor Relationships: The analysis confirmed that higher urbanization, a larger working-age population, and higher nighttime light intensity are positively associated with Facebook user proportions. However, in rural areas, higher population density was associated with lower user proportions, likely reflecting socio-economic disparities (poverty) in high-density rural settlements.
• Spatial Correlation: The spatial component in $M_{full}$ absorbed variability previously attributed to predictors, resulting in lower uncertainty (standard deviation) in the posterior predictive distributions.

Significance and Claims
The paper claims to provide a general framework for converting biased, privacy-censored social media signals into reliable, timely population estimates. Its primary contributions are:

1. Addressing Privacy Bias: It explicitly models the differential privacy censoring mechanism rather than ignoring it, demonstrating that failing to do so leads to the systematic exclusion of rural populations.
2. Dynamic Estimation: The proposed models allow for regular updates of population predictions as new social media data becomes available (every 8 hours), offering a dynamic complement to static census-based estimates.
3. Humanitarian Utility: By demonstrating the ability to estimate population stocks at fine spatial resolutions (municipality level) in a disaster-prone region like the Philippines, the work offers a tool for more targeted humanitarian response.
4. Methodological Rigor: The study highlights the necessity of accounting for overdispersion and spatial correlation to obtain accurate estimates and appropriate uncertainty intervals, moving beyond simple point estimates.

The authors maintain a modest stance, acknowledging that the models rely on specific assumptions (e.g., stationarity of user behavior, the accuracy of census data as ground truth) and that the approach is most applicable to regions with significant social media adoption. They note that while the current models use specific predictors, the framework is extensible to include other real-time data sources as they become available.