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

TL;DR

This study tackles the challenge of deriving dynamic population stocks for disaster response by leveraging Facebook location data in the Philippines while explicitly addressing privacy-induced censoring and sampling biases. A two-stage Bayesian framework first imputes censored tile-level counts using a censored Poisson hierarchical model, then links imputed Facebook uptake to true population via a sequence of increasingly sophisticated models, culminating in a spatially explicit beta-binomial regression with Hilbert space Gaussian Process components. The approach demonstrates robust out-of-sample performance, with errors down to approximately 18% in urban areas and 24% in rural areas, and highlights the importance of accounting for overdispersion and spatial autocorrelation to obtain reliable credible intervals. The workflow yields a dynamic, scalable complement to census data, enabling near-real-time population estimates to inform humanitarian response, while acknowledging limitations related to data access, privacy constraints, and regional transferability across different digital ecosystems.

Abstract

Accurate and timely population data are essential for disaster response and humanitarian planning, but traditional censuses often cannot capture rapid demographic changes. Social media data offer a promising alternative for dynamic population monitoring, but their representativeness remains poorly understood and stringent privacy requirements limit their reliability. Here, we address these limitations in the context of the Philippines by calibrating Facebook user counts with the country's 2020 census figures. First, we find that differential privacy techniques commonly applied to social media-based population datasets disproportionately mask low-population areas. To address this, we propose a Bayesian imputation approach to recover missing values, restoring data coverage for $5.5\%$ of rural areas. Further, using the imputed social media data and leveraging predictors such as urbanisation level, demographic composition, and socio-economic status, we develop a statistical model for the proportion of Facebook users in each municipality, which links observed Facebook user numbers to the true population levels. Out-of-sample validation demonstrates strong result generalisability, with errors as low as ${\approx}18\%$ and ${\approx}24\%$ for urban and rural Facebook user proportions, respectively. We further demonstrate that accounting for overdispersion and spatial correlations in the data is crucial to obtain accurate estimates and appropriate credible intervals. Crucially, as predictors change over time, the models can be used to regularly update the population predictions, providing a dynamic complement to census-based estimates. These results have direct implications for humanitarian response in disaster-prone regions and offer a general framework for using biased social media signals to generate reliable and timely population data.

Figures (26)

• Figure 1: Illustration of the analysis workflow. The first step targets the missingness in the Facebook data at the tile level. A Bayesian statistical approach is used to impute the user count at the locations for which figures are not reported at the time point of interest (4 May 2020). Secondly, the imputed data is aggregated at the administrative area level and the proportion of facebook users is calculated using existing census data. This proportion is the target of a second model that includes a series of geolocated predictors (pictured on the right end side as stacked choropleth maps) that are used to explained the observed social media uptake.
• Figure 2: (a) Facebook user density in three different municipalities showing different temporal trends in March-December, 2020. Data for the available timestamps (00:00, 08:00, 16:00) are plotted separately to highlight daily differences. The sudden drop in user density in November 2020 in Bulacan (and to a lesser extent in Zambales) reflects the effects of population evacuation in response to Typhoon Vamco. (b) Normalised region-level user density distribution for weekday and weekend time windows. Different user behaviour is clearly present in different regions, with Bulacan and the Manila areas witnessing respectively an increase and a decrease in user counts during the weekend (July and October data). On the other hand, user counts in Zambales are much less variable, reflecting its rural nature. Interestingly, in the month of May, the week-to-weekend shifts are strongly dampened. This reflects the impact of the restricted mobility imposed by the COVID-19 lockdowns, and suggests that the observed user counts can be closely matched against the census results (which reflect a person’s place of residence).
• Figure 3: (a) Histograms of all administrative areas by population size calculated before (orange) and after (blue) the imputation process is implemented. The three panels refer to the different degrees of urbanisation. Most of the areas recovered through the imputation process are Rural. (b) Scatter plots of the population density against the area of each administrative unit, with the markers' shape distinguishing between units with (circles) and without (squares) signal. Only a sample of the areas with signal is mapped for ease of visualisation. The two panels show the situation before (top) and after (bottom) the imputation process is implemented. Data missingness is biased towards small or sparsely populated areas.
• Figure 4: Some statistical properties of the user counts for the tiles included in the imputation process (those with missing data on 4 May 2020). (a) Histogram of the number of Facebook users recorded throughout all the available 2020 weekly and nighttime data for a sample tile. The clearly truncated distribution highlights the effect of the censoring process. (b) Histogram (in percentage) of the number of times each tile occurs in the data. Most tiles have 151 entries. (c) Histogram of the proportion of censored entries per tile. For the large majority of the tiles used in the imputation process the signal is missing every day. Spikes in data availability are visible around 25%, 50%, and 75% of censored entries.
• Figure 5: Histograms of user count predictions generated by the imputation model for six sample tiles. For all tiles with partially observed data, the model-inferred user counts median is lower than that of the observed values.