Correcting socioeconomic bias in mobile phone mobility estimates using multilevel regression and poststratification

Call detail records (CDR) from mobile phone networks are widely used to study human mobility, but the user population of any single carrier is not representative of the general population. Using data from a major Chilean carrier for the Santiago Metropolitan Region (47 comunas, 730 census zones) covering approximately 118,400 user-month-weekday observations in 2016, we document non-uniform penetration rates across socioeconomic groups: the highest-income group (ABC1) has the lowest penetration (2.2%), while middle-income groups C2 (6.9%) and D (5.9%) are overrepresented. Because mobility patterns vary with socioeconomic status, this compositional bias distorts aggregate mobility metrics such as the radius of gyration. We use multilevel regression and poststratification (MRP) as a correction for sample bias. Specifically, we fit a Bayesian multilevel model of individual-level mobility as a function of socioeconomic group, gender, and geography, allowing for partial pooling across comunas. We then poststratify the model predictions to align with the demographic composition reported in the census. The resulting MRP-adjusted estimate of the average radius of gyration is 20.2km, $17%$ lower than the naive CDR-based estimate of 24.2km. Cross-validation on held-out comunas yields a root mean squared error of 0.71 on the log scale, and sensitivity analyses indicate that the comuna-level random intercept is a key component of the model. This application illustrates how MRP can be used to correct CDR-derived mobility estimates for non-representative user populations.