Correcting socioeconomic bias in mobile phone mobility estimates using multilevel regression and poststratification
Leo Ferres, Lætitia Gauvin · 2026
Read Correcting socioeconomic bias in mobile phone mobility estimates using multilevel regression and poststratification by Leo Ferres, Lætitia Gauvin (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-16193.txt.
Introduction
Ferres and Gauvin correct socioeconomic bias in mobility estimates drawn from mobile-phone call records in Santiago, Chile. Because one carrier's users skew by social group, simple averages get distorted. One such average is travel radius. Fitting a Bayesian multilevel model by socioeconomic status, gender, and place, then reweighting to census demographics, cuts the naive average radius estimate by about 17 percent. Even a geography-only version captures much of the bias.
Catalog record
| CREATOR | Leo Ferres, Lætitia Gauvin |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 39976 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.16193
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Correcting socioeconomic bias in mobile phone mobility estimates using multilevel regression and poststratification Leo Ferres, Lætitia Gauvin 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. Keywords: call detail records, human mobility, multilevel regression and poststratification, socioeconomic bias, radius of gyration, Bayesian hierarchical models 1 Introduction Mobile phone data, particularly Call Detail Records (CDR), have transformed the study of human mobility González et al. 2008 ; Barbosa et al. 2018 . Passive observation of millions of users over extended periods enables measurement of mobility patterns at unprecedented spatial and temporal resolution, with applications ranging from urban planning Lenormand et al. 2015 to epidemic modelling Wesolowski et al. 2012 ; Tizzoni et al. 2014 ; Gozzi et al. 2021 and poverty estimation Blumenstock et al. 2015 . Despite these advances, a fundamental limitation of CDR data is that the user base of any single mobile carrier is not necessarily a random sample of the population Ricciato et al. 2017 ; Vanhoof et al. 2018 . Prior work has documented this primarily in low-penetration settings, where mobile phone ownership itself is socioeconomically stratified: phone owners in Rwanda and Kenya are wealthier, better educated, and more predominantly male than the general population Blumenstock and Eagle 2010 ; Wesolowski et al. 2013a , and resulting biases propagate into modelled outbreak dynamics Chin et al. 2025 . However, in high-penetration markets such as Chile, where mobile subscriptions exceed one per capita Pew Research Center 2016 , the ownership channel is largely closed. The relevant bias instead concerns carrier-specific representation : the user base of any single operator is a non-random subset of the near-universally phone-owning population, selected by carrier preference, contract type, and voice-call usage patterns. The direction and magnitude of this bias is an empirical question that depends on the local market structure. The development of methods to correct for representativeness bias in mobility metrics research has received limited attention. Although the presence of bias in mobile phone data has been recognized since the early use of Call Detail Records (CDRs), and has been extensively quantified and documented Wesolowski et al. 2013b ; Zhao et al. 2016 ; Liu et al. 2024 ; Cabrera and Rowe 2025 , relatively few studies propose concrete methods to correct these biases. Existing approaches often rely on simple rescaling of observations to improve representativeness Wesolowski et al. 2013b ; Meppelink et al. 2020 , but methodological developments specifically tailored to CDR data remain limited. More advanced approaches have been developed in related contexts. For example, some studies estimate bias and infer relative population changes over time using difference-in-differences frameworks without requiring ground truth data Pestre et al. 2020 . However, these methods are primarily designed to capture temporal dynamics rather than to correct representativeness bias in mobility metrics. Similarly, work addressing temporal bias in GPS-based mobility data has been proposed Sanchez et al. 2026 , but with a different objective. In other domains, bias correction methods such as Multilevel Regression with Post-stratification (MRP) have been widely applied Wang et al. 2015 and offer a promising framework for addressing representativeness issues. However, such approaches have not yet been adapted to CDR-based mobility data. Some studies have applied post-stratification techniques to adjust indicators derived from mobile phone surveys, such as total fertility rate Sánchez-Páez et al. 2023 , but these approaches do not incorporate a regression component. Recent studies on mobile phone-based mobility data continue to highlight bias as a persistent issue, emphasizing the need for new methodological developments Li et al. 2024 ; Gauvin 2024 . This is particularly important given the increasing use of human mobility data in policy and research. Uncorrected biases may lead to misleading conclusions and suboptimal policy decisions. If mobility patterns differ systematically across the demographic groups that are differentially represented, the bias in aggregate estimates can be substantial. In this paper, we make three contributions. First, we quantify the representativeness bias in CDR data from a major Chilean carrier by comparing the demographic composition of CDR users with the 2017 national census for the Santiago Metropolitan Region. Second, we propose multilevel regression and poststratification (MRP) as a principled statistical correction. MRP, widely used in political science for small-area estimation from non-representative surveys Park et al. 2004 ; Lax and Phillips 2009 ; Gelman and Little 1997 , combines a multilevel model that relates the outcome to demographic and geographic predictors with poststratification weights derived from census population counts. Third, we apply this framework to the radius of gyration, a standard mobility metric, and show that MRP-corrected estimates differ meaningfully from naive CDR averages, particularly at disaggregated levels. 2 Data and Context 2.1 Call detail records We use CDR data from the former Telefónica Chile mobile network (Movistar), one of the country’s three major mobile carriers, covering May to July 2016. At that time, Telefónica held approximately 31.9% of Chile’s mobile market. Mobile penetration was about 127 subscriptions per 100 inhabitants, implying that basic mobile access was effectively universal Subsecretaría de Telecomunicaciones de Chile 2016 . The main representativeness concern is therefore not phone ownership, but which users are captured in this carrier’s voice-call records. The original dataset contains anonymised voice-call records for users in three Chilean metropolitan regions. We restrict the analysis to the Santiago Metropolitan Region (Región Metropolitana), where census and socioeconomic data are most complete. For each user, we know the antenna location of each call, the time, and call metadata (placed or received, duration) as well as the socioeconomic group (GSE) and gender. We restrict the sample to Chilean nationals with a valid GSE classification and at least one call per day on average, following standard filtering practices to ensure reliable location estimates Pappalardo et al. 2015 . This activity threshold is relatively aggressive and may itself introduce socioeconomic bias if call frequency correlates with GSE; we return to this point in Section 7 . After filtering, the dataset contains approximately 118,400 user-month-weekday observations. 2.2 Construction of mobility metrics and spatial assignment for MRP Home locations are assigned