Nonrepresentativeness of Human Mobility Data and its Impact on Modeling Dynamics of the COVID-19 Pandemic: Systematic Evaluation

Abstract

Background: In recent years, a range of novel smartphone-derived data streams about human mobility have become available on a near-real-time basis. These data have been used, for example, to perform traffic forecasting and epidemic modeling. During the COVID-19 pandemic in particular, human travel behavior has been considered a key component of epidemiological modeling to provide more reliable estimates about the volumes of the pandemic's importation and transmission routes, or to identify hot spots. However, nearly universally in the literature, the representativeness of these data, how they relate to the underlying real-world human mobility, has been overlooked. This disconnect between data and reality is especially relevant in the case of socially disadvantaged minorities.

Objective: The objective of this study is to illustrate the nonrepresentativeness of data on human mobility and the impact of this nonrepresentativeness on modeling dynamics of the epidemic. This study systematically evaluates how real-world travel flows differ from census-based estimations, especially in the case of socially disadvantaged minorities, such as older adults and women, and further measures biases introduced by this difference in epidemiological studies.

Methods: To understand the demographic composition of population movements, a nationwide mobility data set from 318 million mobile phone users in China from January 1 to February 29, 2020, was curated. Specifically, we quantified the disparity in the population composition between actual migrations and resident composition according to census data, and shows how this nonrepresentativeness impacts epidemiological modeling by constructing an age-structured SEIR (Susceptible-Exposed-Infected- Recovered) model of COVID-19 transmission.

Results: We found a significant difference in the demographic composition between those who travel and the overall population. In the population flows, 59% (n=20,067,526) of travelers are young and 36% (n=12,210,565) of them are middle-aged (P<.001), which is completely different from the overall adult population composition of China (where 36% of individuals are young and 40% of them are middle-aged). This difference would introduce a striking bias in epidemiological studies: the estimation of maximum daily infections differs nearly 3 times, and the peak time has a large gap of 46 days.

Conclusions: The difference between actual migrations and resident composition strongly impacts outcomes of epidemiological forecasts, which typically assume that flows represent underlying demographics. Our findings imply that it is necessary to measure and quantify the inherent biases related to nonrepresentativeness for accurate epidemiological surveillance and forecasting.

Keywords: COVID-19; data representativeness; epidemiological modeling; human mobility; population composition.

Figure 1

Profiles of the intercity movements extracted from mobile phone data between January 1 and February 29, 2020, in China. (A) and (B) show the daily number of travelers for different age and gender groups; (C) and (D) show the respective ratios. Dashed horizontal lines denote the composition of respective groups in the latest 7th census. As children generally do not have mobile phones, the proportions of young (20-39 years), middle-aged (40-59 years), and older people (≥60 years) add up to 100%.

Figure 2

Dynamics of the incidence rates among different groups predicted by the age-structured model. Solid lines indicate the incidence rates of different age groups from census data, and dashed lines indicate the incidence rates by using traveling data from mobile phones.