Experiments on leveraging mobile network data for official statistics

Abstract

This paper assesses the feasibility of producing official statistics from macro–level Mobile Network Operator (MNO) data using a quasi–randomization (QR) approach, traditionally applied to draw inferences from non–probabilistic surveys. Using data from a large sample survey as a proof–of–concept, we emulate MNO macro aggregates and quantify the selection error that arises when estimates are based solely on mobile phone users rather than on a representative sample of the target population of interest. Results show that estimates derived from MNO–like data are affected by notable selection bias. The QR approach substantially reduces this bias through poststratification using auxiliary variables; however, it cannot address two additional issues inherent to real MNO data: user ambiguity (mismatch between subscriber and actual device user) and device duplication (individuals carrying multiple devices). Simulation studies demonstrate that user ambiguity can introduce unpredictable distortions in the estimates, while even low levels of device duplication can artificially inflate estimates. We conclude that QR is effective in mitigating selection error, while fully addressing user ambiguity and device duplication requires additional auxiliary information or improvements in the quality and structure of MNO input data.