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

TL;DR

This paper applies multilevel regression and poststratification (MRP) to correct socioeconomic bias in mobile phone mobility data, improving the accuracy of population mobility estimates from biased CDR samples.

Contribution

It introduces the use of MRP for CDR-based mobility studies, demonstrating bias correction using socioeconomic and geographic data in a real-world Chilean dataset.

Findings

MRP reduces the radius of gyration estimate by about 17%.

A geographic-only model still captures much of the socioeconomic bias.

MRP provides a principled correction for non-representative CDR data.

Abstract

Call detail records (CDR) from mobile phone networks are widely used to study human mobility however CDR data from a single mobile operator are inherently biased because the observed users do not mirror the population distribution. Using data from a major Chilean carrier in Santiago, we observe the user base is skewed by socioeconomic group, so aggregate metrics like radius of gyration are distorted by the population that is actually observed. To correct this sampling bias, we apply multilevel regression and poststratification (MRP), a method that is not yet standard for CDR-based mobility studies. We fit a Bayesian multilevel model for individual mobility using socioeconomic status, gender, and geography, with partial pooling across comunas, and then poststratify the predictions to match census demographics. This approach reduces the naive CDR estimate of average radius of gyration by about 17%. Importantly, a version of the model that uses only geographic information still captures much of the bias, showing that MRP can be useful even when the socioeconomic composition of users is not fully known, as long as spatial patterns of socioeconomic groups exist. This example demonstrates how MRP can provide a principled correction for non-representative CDR-derived mobility estimates, rather than treating the carrier sample as if it were a random population sample.