With large-scale empirical data (eg, mobile phone records, GPS data, and location-based social network data) becoming available with increasingly fine spatial and temporal resolution [1], quantitative studies on individual and collective mobility patterns have flourished in the past few years [2-6]. These developments have offered advances with respect to understanding migratory flows, traffic forecasting, urban planning, and epidemic modeling [7-10]. The ongoing COVID-19 pandemic has further intensified discussions on how to optimally use human mobility research to support outbreak responses and nonpharmaceutical interventions (eg, contact tracing) [11-14].

The representativeness of data sets used to infer real-world human mobility, however, has typically not been explicitly incorporated in such analyses. This is potentially troubling as representativeness is known to be especially poor for socially disadvantaged minorities such as low-income groups, women, children, and older people. For example, it has been confirmed that individuals’ probability of travel is not randomly or equally distributed, and there is significant heterogeneity when comparing the travel patterns of different demographic groups [15-17]. For example, women are more localized than men in their movements and visit fewer locations in regions such as Latin America, Bangladesh, and sub-Saharan Africa [18,19]. In the specific case of epidemic outbreaks, low-income individuals are not necessarily able to limit their exposure to a circulating virus by reducing mobility and must continue, for example, commuting behavior to remain employed. Thus, this group is subject to a substantially higher probability of becoming infected in an epidemic than higher-income groups [20]. Further, different contact rates across age groups have been observed in COVID-19 incidence cases [21], and higher COVID-19 infection rates among disadvantaged racial and socioeconomic groups have been observed in multiple studies [22,23]. It has also been argued that including information about demographic heterogeneity in human mobility patterns, for example, by combining demographically stratified travel data with epidemiology research, would make epidemiological models more robust [24].

While it is widely recognized that rich new data sources can provide near–real-time information about human mobility [25] and powerful input into models that estimate imported cases using regional mobility information when modeling pathogen transmission, the state-of-the-art models do not consider data representativeness. This is typically because for privacy concerns, most data sets are not disaggregated demographically. Instead, relevant information on demographic features and social relationships is traditionally collected by censuses and other surveys [26-28].

As we argue below, however, simply considering the population demographics at the origin of a trip does not represent the traveling population.

To systematically evaluate how real-world travel flows differ from census-based estimations, we use an aggregated and anonymized data set collected from 318 million mobile phones. Specifically, we quantified the disparity in the population composition between actual migrations and resident composition according to census data and found significant differences. We then investigated how this nonrepresentativeness impacts epidemiological modeling. The aim of this study is to illustrate the nonrepresentativeness of data on human mobility and the impact of this nonrepresentativeness on modeling dynamics of the COVID-19 pandemic.

By comparing mobility traces from mobile phone users to census data, our study has highlighted a number of striking differences in the demographic composition of those who travel with respect to the overall population. For example, we found that 59% (n=20,067,526) of travelers are young and 36% (n=12,210,565) of them are middle-aged, which is completely different from the composition of the adult population in China, where 36% of people are young and 40% of them are middle-aged. The travel probability and travel distance between men and women were also significantly different. This realization is especially important in the case of epidemic forecasting, and increased awareness of this issue in the scientific community has the potential to improve not only epidemiological models but also our overall understanding of possible biases when inferring human mobility from cell phone data and the representational issues of mobility data. It is important to emphasize that while China is an ideal place to study representativeness, our findings about which fraction of individuals compose the population of travelers are specific to China. The fraction of young, middle-aged, old, male, and female individuals who travel is likely to depend on a range of factors and can be expected to be different in different countries.

Nonetheless, the realization that understanding the representativeness of mobility data is crucial for epidemic monitoring and forecasting is generalizable. Thus, our results imply that when generalizing results from population mobility analysis, these differences should be included in the analysis to avoid potential biases caused by data nonrepresentativeness. For example, in the case of the COVID-19 pandemic, as travelers often have a higher probability of infection, the transmission risk among men and youth could be a promising focus for COVID-19 prevention.

In the recent Omicron waves, imported infections represented the majority of cases in China, and most COVID-19–positive individuals had a travel history to high-risk areas such as Shanghai [34]. As presymptomatic and asymptomatic pathogen carriers can travel to a foreign country and initiate the spread of COVID-19 even when there is no community transmission, human migration behaviors are promising candidates to incorporate into epidemiological models. Our findings emphasize that focusing on the representativeness of mobility data is essential for more sophisticated modeling approaches to capture key mechanisms of epidemic propagation. In future work, we intend to further explore how to accurately quantify the inherent biases related to data nonrepresentativeness for accurate epidemiological surveillance and forecasting.