Hostname: page-component-8448b6f56d-c47g7 Total loading time: 0 Render date: 2024-04-23T11:54:11.300Z Has data issue: false hasContentIssue false

Community movement and COVID-19: a global study using Google's Community Mobility Reports

Published online by Cambridge University Press:  13 November 2020

M. Sulyok*
Affiliation:
Institute of Tropical Medicine, Eberhard Karls University, University Clinics Tübingen, Wilhelmstr. 27, 72074, Tübingen, Germany Department of Pathology, Institute of Pathology and Neuropathology, Eberhard Karls University, University Clinics Tübingen, Liebermeisterstr. 8, 72076, Tübingen, Germany
M. Walker
Affiliation:
Department of the Natural and Built Environment, Sheffield Hallam University, Howard Street, S1 1WB, Sheffield, UK
*
Author for correspondence: M. Sulyok, E-mail: mihaly.sulyok@uni-tuebingen.de
Rights & Permissions [Opens in a new window]

Abstract

Google's ‘Community Mobility Reports’ (CMR) detail changes in activity and mobility occurring in response to COVID-19. They thus offer the unique opportunity to examine the relationship between mobility and disease incidence. The objective was to examine whether an association between COVID-19-confirmed case numbers and levels of mobility was apparent, and if so then to examine whether such data enhance disease modelling and prediction. CMR data for countries worldwide were cross-correlated with corresponding COVID-19-confirmed case numbers. Models were fitted to explain case numbers of each country's epidemic. Models using numerical date, contemporaneous and distributed lag CMR data were contrasted using Bayesian Information Criteria. Noticeable were negative correlations between CMR data and case incidence for prominent industrialised countries of Western Europe and the North Americas. Continent-wide examination found a negative correlation for all continents with the exception of South America. When modelling, CMR-expanded models proved superior to the model without CMR. The predictions made with the distributed lag model significantly outperformed all other models. The observed relationship between CMR data and case incidence, and its ability to enhance model quality and prediction suggests data related to community mobility could prove of use in future COVID-19 modelling.

Type
Original Paper
Creative Commons
Creative Common License - CCCreative Common License - BY
This is an Open Access article, distributed under the terms of the Creative Commons Attribution licence (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted re-use, distribution, and reproduction in any medium, provided the original work is properly cited.
Copyright
Copyright © The Author(s), 2020. Published by Cambridge University Press

Introduction

COVID-19 is a highly infectious viral infection, and the main route of transmission is thought to be through respiratory droplets [Reference Chen1, Reference Li2]. The level of COVID-19 transmissibility is greater than for other closely related conditions, such as the SARS virus [Reference Linton3]. Those affected are infectious prior to exhibiting symptoms of illness, or remain unaware of infection because they experience only mild symptoms or are asymptomatic; factors which promote further transmission of the disease [Reference Li2, Reference Rothe4].

Given the highly infectious nature of COVID-19, reducing levels of social interaction and community movement have been seen as key in reducing the rates of COVID-19 transmission [Reference Koo5, 6]. Recommendations have included the practising of social distancing, self-isolation or quarantine, and increasing levels of personal hygiene [Reference Maharaj and Kleczkowski7Reference Wilder-Smith and Freedman9]. Such recommendations were followed by more formal, more stringent and often legally imposed governmental restrictions on personal movement which have included ‘stay at home’ orders, closure of non-essential retail units and schools, and banning of sports and entertainment gatherings [Reference Hale10].

The implementation of such measures in response to infectious disease outbreaks is not new; methods aiming to reduce social contact and limit mobility being used for centuries [Reference Cetron and Simone11Reference Tognotti13]. More recently, measures restricting social interaction and movement have been used in response to the SARS and MERS epidemics which occurred in the last decades [Reference Cetron and Simone11, Reference Anderson14Reference Jang, Jang and Lee16]. That movement affects disease transmission and incidence has been shown in numerous studies [e.g. Reference Mari17, Reference Gog18]. However, although the connection between mobility and disease has been known for centuries, the detailed quantitative study of this relationship has been difficult. Measuring and quantifying the levels of social interaction and mobility, over large geographical areas, and for large populations is often not feasible.

However, over the last 20 years, technological progress has meant that potential new sources of data providing information on population-wide movement patterns have become available. During this time, mobile phone and Internet usage has become almost ubiquitous. Recording of user behaviour, often also including locational information, has provided detailed new sources of data relating to mobility [Reference Velasco19]. Epidemiologists have been eager to utilise such datasets for disease monitoring and surveillance [Reference Bansal20]. Notable studies using such data sources include Wesolowski et al. [Reference Wesolowski21], who modelled the spread of malaria in Kenya using the records from 15 million mobile phone users. Finger et al. were able to use mobile phone records to monitor the effect mass gatherings had on cholera outbreaks [Reference Finger22]. Other diseases similarly studied include cholera [Reference Bengtsson23] and dengue fever [Reference Wesolowski24].

In response to the COVID-19 outbreak, Google released data collated from those accessing its applications using mobile and handheld devices. These Google ‘Community Mobility Reports’ (CMR) [25] show changes in activity and mobility at different location types, compared to before the spread of COVID-19 globally. These datasets are a useful, and global, measure of social activity and movement. Uniquely, they allow comparison between countries. These reports provide the opportunity to study the relationship between social activity and mobility and COVID-19 incidence. In the absence of other global sources of data for these factors, Google's CMR data provide a good indication of the effect health recommendations and governmental restrictions have had on social activity and movement.

The main aim of this study was thus to examine the relationship between mobility and confirmed case numbers for COVID-19 globally, and to ascertain whether cross-country patterns in this relationship were apparent. Such patterns could reflect the range of movement restrictions implemented [Reference Hale10], but could also be due to other cultural or socio-economic differences [Reference Kapitány-Fövény and Sulyok26, Reference Smith27]. Another aim was to integrate CMR data into disease models, to assess whether it could enhance model quality and prediction. The experimental hypothesis is that as COVID-19 case occurrence increases, related reductions in mobility will occur; such patterns are expected internationally due to the increasingly globalised nature of communication channels.

Methods

Google Community Mobility Reports

These were accessed on 23 June 2020 and data for 135 countries downloaded, spanning the period from 15 February 2020 until 19 June 2020 [25]. Google's CMR collates data from those accessing Google applications with smartphones or handheld devices who allow recording of ‘location history’ [Reference Aktay28]. Individual user presence and time spent at specific location categories is collated to indicate activity. Data are categorised into six discrete categories, which can be summarised as ‘retail and recreation’, ‘parks’, ‘groceries and pharmacies’, ‘workplaces’, transport ‘transit’ hubs and ‘residential’ areas. Increases in the categories ‘parks’ and ‘residential’ are indicative of decreased mobility, as they suggest increased activity in locations around the home environment. The other four categories are more indicative of general mobility as they are related to activity around workplaces, retail outlets and use of public transportation.

CMR provides the percentage change in activity at each location category compared to that on baseline days before the advent of COVID-19 (a 5-week period running from 3 January 2020 to 6 February 2020). Daily activity changes are compared to the corresponding baseline figure day, with for example, data on a Monday being compared to corresponding data from the baseline series for a Monday. Baseline day figures are calculated for each day of the week for each country, and are calculated as the median value [25]. The values thus represent the relative change in percentages compared to baseline days, not absolute number of visitors. Missing values were returned if activity was too low upon a specific day and thus failed to achieve the anonymity threshold set by Google.

COVID-19-confirmed cases

Corresponding data on the daily number of confirmed COVID-19 cases were downloaded on 13 July 2020 from the John Hopkins COVID-19 data repository situated on github [Reference Dong, Du and Gardner29].

Cross-correlation analyses

Correlation analyses were performed using Kendall's τ due to the non-parametric nature of the data. Kendall's τ correlations were performed using a ±28-day lag. The τ value representing the strongest negative or positive correlation, and the corresponding lag in days were tabulated and illustrated as heat-map coded world maps for each CMR category.

Expected was a negative correlation between case numbers and activity in those categories indicative of mobility (‘retail and recreation’, ‘grocery and pharmacy’, ‘transit’, ‘workplace’), and a positive correlation for those two categories (‘parks’, ‘residential’) indicative of sedentary behaviour. A ±28-day lag was chosen in order to encompass the incubation period for COVID-19, some studies reporting that it can extend to 15 days [Reference Nie30]. Examining data using such a lag also takes into account that testing for infection often only occurs some time post-symptom onset, and also the delays occurring between testing, confirmation of infection and updating of official figures.

Results were summarised and tabulated on the country (Supplementary material S1) and on the continent-wide level (Tables 1 and 2). Multiple group comparisons were done with the Kruskal–Wallis test, pairwise comparisons with the Dunn test (with Holm correction to multiple testing) on the continent-level data. Model-based clustering was performed on the scaled data excluding missing variables. We used the mclust package [Reference Scrucca31] to select the optimal model based on Bayesian Information Criteria (BIC) for EM algorithm initialised by hierarchical clustering for parameterised Gaussian mixture models.

Table 1. Continent-level summaries of maximum Kendal's τ correlations

Continent-level aggregate summaries of Kendall's τ analyses between confirmed case numbers and Google CMR data. Results show the τ-values corresponding to the strongest correlations. Median values. IQR (in brackets).

Table 2. Continent-level summaries of cross-correlations

Continent-level aggregate summaries of Kendall's τ cross-correlation analyses between confirmed case numbers and Google CMR data. Results show the number of days case numbers were lagged which resulted in the strongest correlations. Median values. IQR (in brackets).

Modelling

CMR data were integrated into models based upon each country's case incidence.

Mixed-effects random intercept generalised additive models were fitted to the data using incident case numbers as the explained variable to all subsets of data using a Tweedie distribution type. Countries were added as random intercepts. This modelling technique was chosen because of the data structure, with their being repeated measurement values, and a high number of grouping factor levels. Similar mixed-effects approaches were used by other authors including Kraemer et al. and Chan et al. [Reference Kraemer32Reference Chan34]. However, in contrast to these studies, here, we used smoothed variables and distributed lag models with a multilevel generalised additive modelling approach.

Three models were established. First, the smoothed numerical date was used as the explanatory variable. Second, smoothed data for five of the CMR location categories were additionally added as explanatory variables. Change in ‘parks’ mobility was omitted from modelling; increases in ‘park’ activity was expected for northern hemisphere countries during this spring period. Third, the smoothed numerical date was applied as the explanatory variable, but instead of adding data for the five CMR location categories, a spline-described lag of ±14 days for each category was created and used in the distributed lag model.

Model comparison and validation

Model performance was compared using the BIC. Then each model was compared to each other (Supplementary material S2). To validate the models, we made predictions of the incident case numbers for the time interval from 19 June 2020 to 08 September 2020. Data were obtained on 13 September 2020 as described previously. To compare the predictive performance of the models, we calculated root-mean-square-errors (RMSE) of the different predictions by comparing the predicted values with the reported ones. Comparisons of predictions were also made by using the two-sided Diebold–Mariano test.

Observations with missing values were omitted from calculations. All calculations were performed in R version 3.6.3 using the dlnm, mgcv and lme4 packages. The statistical code together with the data is provided in Supplementary material S1.

Results

Cross-correlation analyses

Figure 1 shows the strongest level of correlation, as Kendall's τ, that was found between the COVID-19 case incidence for each country and daily percentage change in activity for each location category. A positive correlation indicates that as COVID-19 case numbers increased, there was an increase in activity indicated by data on that location category. A negative correlation indicates that while COVID-19 case numbers rose, there was a decline in activity in that location category or vica versa.

Fig. 1. Maximum absolute τ values. Results of Kendall's τ cross-correlation between COVID-19-confirmed case number and measures of community activity. Strong continent-wide regional patterns are apparent. Generally for the four categories indicative of mobility (‘retail and recreation’, ‘grocery and pharmacy’, ‘workplace’ and ‘transit’) strong negative correlations were observed across countries of North America, Russia, Australia, India and Western Europe. Positive relationships are seen in the South Americas, Eastern Europe, India and Southern Africa. For ‘residential’ activity, which is indicative of increased sedentary behaviour, the opposite was generally observed. For ‘parks’ the picture was mixed, possibly reflecting the difference nature of legal restrictions on a country by country basis; some countries implemented lockdown while others did not, some permitted outdoors exercise, others not [Reference Hale10].

Particularly strong negative correlations are apparent for prominent countries of Western Europe, across the North Americas, Russia and Australia, for location categories ‘retail and recreation’, ‘grocery and pharmacy’, ‘workplace’ and ‘transit’ activity. This indicates that as disease incidence rose, activity levels declined. However, weaker, or even positive correlations, can be observed for these categories for countries across South America, in Eastern Europe and for India. The reverse pattern was observed for ‘parks’ and ‘residential’ categories; at these locations, an increase in activity would be expected if there was an increase in time spent close to habitation.

Distinct geographical patterns are noticeable from the map, with broad trends being apparent across large geographical areas. Thus, as well as on a country-wide basis, results were collated and examined on a continent-wide basis, as provided in Tables 1 and 2. As illustrated in the accompanying boxplots, when aggregating data this way, negative correlations between ‘retail and recreation’, ‘grocery and pharmacy’, ‘workplace’ and ‘transit’ activity with disease incidence occurred across all continents, except South America. Again, the opposite was observed for ‘parks’ and ‘residential’ activity.

Figure 2 shows the number of days of lagging, which resulted in the strongest correlation for each location category, for each individual country. Considerable negative time lags result in the strongest correlations for prominent countries across all categories, including the USA, Canada and Russia, and countries of Western Europe. The exception to this pattern is for ‘parks’, where a positive lag results in the strongest correlation for Russia.

Fig. 2. Lags to maximum correlations. Amount of time lagging in days resulting in the maximum Kendall's τ between COVID-19 for confirmed case number and measures of community activity (colour online only). Interesting are that the strongest correlations were when case numbers were negatively lagged by amounts of −20 days or greater for large areas across North America, Western Europe, Central Asia and Russia for the four categories indicative of mobility. This suggests that reductions in mobility in such areas occurred substantially prior to corresponding increases in COVID-19 case numbers. This is thus likely to have been substantially prior to formal legislation imposing movement restrictions coming into place. This indicated that personal behavioural choices and perceived risk perception may have played a greater role in driving movement patterns than legal restrictions.

Conversely, the strongest correlations are obtained with positive time lags for countries of South America and for Australia across all categories. The exception is ‘parks’ activity for countries of South America where negative lagging results in the strongest correlations; this is apparent from the accompanying boxplots which aggregate data continent wise. Significant differences were found for all six parameters (the lags in days producing the strongest correlation, and the τ value representing the correlation for all six CMR categories (Table 1 and 2)).

Clustering analysis was used to group countries on the basis of maximum absolute τ values and the corresponding lags. This identified four groups; the first where low lags resulted in greatest correlation (mainly Asian and Eastern European; e.g. India, Pakistan, Afghanistan, Russia, Belarus; but Brazil and Sweden also belong to this group). These countries showed weak levels of correlation, obtained with lags of around 0 or only weakly positive (median lags 3–10 days).

The second cluster contained mainly industrialised westernised countries such as the UK, Germany, Italy, Spain, the USA, Canada and Australia. This cluster had strong negative τ values, obtained using lags of only a few days (median −2 to 3.5). Group three (typically African and Asian countries; e.g. South Africa, Nigeria, Saudi Arabia, but also Mexico) had strong negative correlations obtained with strong negative lags (median −27 to −28). The final group (Eastern European and South American countries, such as Ukraine, Poland, Venezuela, Colombia) had strong positive correlations, and a high number of positive lags (27–28 days median). Detailed country-level values and cluster-level summary are provided in Supplementary material S2 and S3.

Pairwise comparisons with multiple testing correction found significant differences between Africa and South America in the lag producing the maximum correlation for ‘workplace’ data (Z: −4.74; adjusted P-value: <0.001) and between the Kendall's τ value for ‘retail and recreation’ between Africa and EU (Z: 4.049, adjusted P-value: 0.0045), the Kendall's τ value for ‘transit’ stations between EU and South America (Z:−4.253, adjusted P-value: 0.0019), and the Kendall's τ value for ‘residential’ data between the EU and South America (Z: 3.55, adjusted P-value: <0.001).

Modelling: The contemporaneous CMR-expanded model proved superior (BIC: 79096.29) to the numerical date-only model (BIC: 79802.47) and expanded model with distributed lag CMR (BIC: 80286.14). In the contemporaneous CMR-expanded model, the significant independent fixed-effects covariates were the numerical date and ‘retail and recreation’ (negative estimate) change compared with baseline mobility data. Summaries of the contemporaneous CMR-expanded model are shown in Table 3.

Table 3. Model summaries

Model summary statistics using contemporaneous CMR data.

When we validated the models, the results showed a somewhat different hierarchy of the models than based on the previous quality measures: the best performing model was the one with all variables with distributed lags (RMSE:6690.44). This was followed by the model without distributed lags (RMSE:6794.99). The worst predictive performance was identified with the model without the mobility data (RMSE:6840.16). Difference in predictive performance was significant between all models (pairwise two-sided Diebold–Mariano test, in all comparisons P < 0.001).

Discussion

Here, clear correlations between COVID-19 case incidence and levels of mobility, as represented by Google's CMR data, were found. Distinct patterns were discernible over broad geographical areas, with clear negative relationships between disease occurrence and mobility being particularly apparent across North America, Western Europe, Russia and Australia. Reductions in those CMR categories indicating levels of social activity and mobility, ‘retail and recreation’, ‘grocery and pharmacy’, ‘workplace’ and ‘transit’ were apparent in these areas as COVID-19 case incidence rose. ‘Parks’ and ‘residential’ activity increased in line with COVID-19 incidence, suggesting increased time spent in a location close to home as case numbers rose. Thus, as COVID-19 epidemics developed, levels of mobility and movement declined.

That the relationship is clear for these areas may be related to the progression of the COVID-19 epidemic globally. COVID-19 was first identified in China, before spreading to Western European and North America; it is in these areas where the expected negative correlations are strongest and most apparent [Reference Sohrabi35]. The results from clustering analysis, with the identification of distinct groups of countries, illustrate that similar patterns are observed by countries close together, or situated on the same continent. This lends support to the idea that patterns in disease progression occurred on a continent-wide level. The broad similarities apparent over continents also support this idea.

Another explanation for these patterns is that country-specific socio-economic or cultural factors may be influencing mobility levels in response to the disease outbreak. Although a generalisation, those countries where strongly negative correlations were most apparent, are those often categorised as being ‘developed’ nations; they possess well-established communication and media outlets, effective governmental control and host populations possibly more compliant with governmental restrictions. Knowledge and understanding about COVID-19 may have been more widely disseminated and accessible in these countries, influencing personal perceptions of risk, and thus personal movement decisions.

The results from the cross-correlation analyses suggest that mobility reductions occurred sometime prior to corresponding increases in COVID-19 cases in many countries where negative correlations were observed. This may be because of delays in disease identification and reporting which meant that data on confirmed cases did not reflect actual disease incidence. It may be related to the spread of COVID-19 globally, and similar patterns may become more apparent for other areas of the globe once the full progression of the COVID-19 epidemic is understood; further examination at a later date may prove productive. However, these patterns in lagging could also be related to population attributes of those countries. Other studies have found that behavioural characteristics shared by those in nation states influence attitude to risk and subsequent behaviour [Reference Chan34]. Those populations at where the strongest negative correlations are observed may be more risk averse.

Although legally enforced restrictions on mobility are the most obvious factor causing a reduction in population mobility, potentially more important are personal behavioural choices made in response to the threat posed by infectious disease [Reference Lau36, Reference Sadique37]. For example, surveying of Americans found they avoided public gatherings, such as sporting events, malls and public transport in response to fears of H1N1 [Reference SteelFisher38]. Following the SARS epidemic of 2002, large reductions in travel into and out of Hong Kong are reported [Reference Ferguson15]. Surveys found a reluctance to travel and engage in social activities in communities where there was a perceived risk from SARS infection [Reference Lau36].

Studies examining personal risk perception and its relationship with movement during the COVID-19 outbreak show that personal behavioural and psychological traits may be important in influencing the levels of mobility. Chan et al. studied nationwide personality traits along with Google CMR data, finding that countries with agreeable and conscientious personality traits showed greater levels of mobility reduction than countries exhibiting more openness [Reference Chan34]. Another study examined survey data on nationwide risk-taking attitudes, finding that this affected mobility; it also reported a clear effect on the mobility of the WHO pandemic announcement [Reference Chan33]. In another work, the importance of freedom of assembly and association was identified as the most important predictor of COVID-19 doubling time among cultural norms [Reference Kapitány-Fövény and Sulyok26]. The factors affecting personal movement decisions are complex and inter-related; including the personal perceived risk of infection, socio-economic factors and peer group behaviour [Reference Smith27].

The examination of mobility and disease occurrence, and its use in disease modelling has been examined for recent disease epidemics. The 2009 H1N1 influenza pandemic provides a good recent parallel for the current COVID-19 pandemic. Much of the H1N1 infection was spread globally through international air travel [Reference Khan39, Reference Apolloni, Poletto and Colizza40]. Modelling studies have shown how human movement and mobility patterns influenced the geographic spread and timing of this epidemic [Reference Colizza41, Reference Merler and Ajelli42]. Bajardi et al. modelled disease spread using travel data, concluding that restrictions on international travel played little role in controlling the spread of this epidemic globally [Reference Bajardi43].

The potential use of mobility data, and its potential use in COVID-19 modelling, was identified early [Reference Buckee44]. Oliver et al. review the potential use of mobile phone data in COVID-19 modelling [Reference Oliver45]. Studies using such data are rapidly appearing. Notable studies include Jia et al. who used data collated from mobile phone records showing population outflows from Wuhan, to assess the impact quarantining had on mobility, and to predict the frequency and distribution of COVID-19 infections across China [Reference Jia46]. Another recent study modelled the spread of COVID-19 using mobile phone data, modelling the effect of different movement control measures on COVID-19 incidence [Reference Zhou47]. Kraemer et al. used travel-related data from the website Baidu to show patterns in COVID-19 establishment; a relationship between the frequency of travel out of Wuhan accounted for patterns in disease across China [Reference Kraemer32]. Non-Chinese-based studies include that of Badr et al., which modelled the relationship between mobility and confirmed case numbers for individual US counties, and related these to state-wide restrictions on movement [Reference Badr48]. In Italy, Bonaccorsi et al. examined the economic effect of mobility restrictions finding that the mobility effects were greater in areas with greater fiscal capacity. In contrast to other studies [Reference Bonaccorsi49], Chinazzi et al., who studied the travel restrictions in China, concluded that they played little effect in halting the spread of the infection [Reference Chinazzi50]. Many of these studies examine mobility patterns in individual countries; however, Google's CMR is also being used to model COVID-19 across broader geographical ranges. Zhu et al. have used CMR to gauge future case numbers and the reproductive number of COVID-19 across South American countries [Reference Zhu51].

Epidemiologists are constantly seeking new sources of data sources with the potential to enhance existing disease forecasting and modelling. The initial attempts made here to use data from Google's CMR in disease modelling are promising. Integrating contemporaneous mobility data into models of case incidence resulted in models providing better quality measures than those utilising lagged-distributed data. Other similar datasets, such as Apple's Mobility Trends data [52], could be examined and compared to see if similar relationships as found here are apparent. Particularly interesting is that Google's CMR could also offer the potential to examine movement at the local scale rather, meaning more precise and locally based understanding of disease dynamics could be gained.

A disadvantage of using Google's CMR is that the data do not directly equate to some COVID-19 control measures. For example, ‘social distancing’ has been widely promoted as a measure to reduce the transmission of COVID-19 [Reference Kissler53]. Valenti et al. [Reference Valenti54] used CMR data as an estimate of social distancing in the modelling of deaths in Brazil. However, CMR data indicate only general activity at specific location types, and provide no direct indication of adherence to such rules. Another disadvantage is that users of Google technology may not be representative of a country's population as a whole. Demographic groups particularly affected by COVID-19, such as the elderly, who may be more cautious, may be underrepresented in such data. Age, economic and sexual differences in the make-up of Google users may occur between countries, and thus affect country comparisons. One of the disadvantages of this study is that such country-specific reasons are not considered; such research requires further study and consideration of other socio-economic, behavioural and psychological factors. The method used here, examining associations between mobility and case number, provides no insight into the causative factors driving such patterns.

A strength of this study was that it examined broader global trends using CMR. Distinct patterns were observed by examining data on a continent-level scale; most studies examine mobility data only at a national scale, meaning such patterns may be missed. Despite political wish thinking, the spread of infectious disease occurs regardless of notional national boundaries, meaning such examination is pertinent. The format of data provided by CMR means easy cross-country comparison is possible. Another strength of the study was that the modelling method used extended that of previous work; smoothing of variables was applied with GAM modelling to deal with non-linearities. Lagging was made by applying spline-described values to achieve a distributed lag model instead of adding each and every lagged value to reduce complexity and to avoid overfitting.

As already highlighted, a limitation here is that the reasons underlying the patterns observed in mobility were not examined. No attempt to examine the sociological issues affecting mobility was made. Chan et al. [Reference Chan33] note that reductions in mobility coincided with the WHO announcement of a global pandemic in January 2020; further modelling could integrate this into future work. Individual country correlations were not compared directly with each other. Further work could also examine the timing of disease progression in each country, relating it directly to actual changes in activity data. We speculated that economic development status may account for patterns observed here; examination of GDP data, which reflects such status, and its relationship with mobility may be of interest.

In summary, a relationship between levels of social activity and community movement with disease incidence was found for countries globally using Google's CMR. As COVID-19 became established globally, levels of mobility declined, this may be either because of government recommendations and imposed legal restrictions, or through personal behavioural changes resulting from fear of disease. Google's CMR illustrates the effect these measures had on community movement. Interesting is that reductions in mobility appeared to occur substantially prior to the implementation of legal restrictions on movement in many countries, suggesting the importance of personal perceived risk of infection and personal behavioural modifications rather than government edict. Further study to ascertain how countries differ in their adherence to such measures, and whether this is apparent in CMR, would be most interesting. An understanding of the cultural, social and economic factors possibly accounting for some of the differences observed between countries could be productive.

In conclusion, the observed relationship between CMR data and case incidence, and its ability to enhance model quality and prediction suggests data related to community mobility could prove of use in future COVID-19 modelling.

Supplementary material

The supplementary material for this article can be found at https://doi.org/10.1017/S0950268820002757.

Financial support

This research did not receive any specific grant from funding agencies in the public, commercial or not-for-profit sectors.

Conflict of interest

None.

Data

Publicly available sources of data were used. Google Community Mobility Reports: https://www.google.com/covid19/mobility/. Data on case incidence: John Hopkins University, Center for Systems Science and Engineering (CSSE), accessed via github: https://github.com/datasets/covid-19. Data are also available under the following link: https://github.com/msulyok/GoogleMobilityDataCOVID

References

Chen, N et al. (2020) Epidemiological and clinical characteristics of 99 cases of 2019 novel coronavirus pneumonia in Wuhan, China: a descriptive study. The Lancet 395, 507513.CrossRefGoogle ScholarPubMed
Li, Q et al. (2020) Early transmission dynamics in Wuhan, China, of novel coronavirus-infected pneumonia. New England Journal of Medicine 382, 11991207.CrossRefGoogle ScholarPubMed
Linton, NM et al. (2020) Incubation period and other epidemiological characteristics of 2019 novel coronavirus infections with right truncation: a statistical analysis of publicly available case data. Journal of Clinical Medicine 9, 538.CrossRefGoogle ScholarPubMed
Rothe, C et al. (2020) Transmission of 2019-nCoV infection from an asymptomatic contact in Germany. New England Journal of Medicine 382, 970971.CrossRefGoogle ScholarPubMed
Koo, JR et al. (2020) Interventions to mitigate early spread of SARS-CoV-2 in Singapore: a modelling study. The Lancet Infectious Diseases 20, 678688.CrossRefGoogle ScholarPubMed
Imperial College London. Report 9 – Impact of non-pharmaceutical interventions (NPIs) to reduce COVID-19 mortality and healthcare demand. Available at http://www.imperial.ac.uk/medicine/departments/school-public-health/infectious-disease-epidemiology/mrc-global-infectious-disease-analysis/covid-19/report-9-impact-of-npis-on-covid-19/ (Accessed 21 September 2020).Google Scholar
Maharaj, S and Kleczkowski, A (2012) Controlling epidemic spread by social distancing: do it well or not at all. BMC Public Health 12, 679.CrossRefGoogle ScholarPubMed
Hellewell, J et al. (2020) Feasibility of controlling COVID-19 outbreaks by isolation of cases and contacts. The Lancet Global Health 8, e488e496.CrossRefGoogle ScholarPubMed
Wilder-Smith, A and Freedman, DO (2020) Isolation, quarantine, social distancing and community containment: pivotal role for old-style public health measures in the novel coronavirus (2019-nCoV) outbreak. Journal of Travel Medicine 27, taaa020.CrossRefGoogle ScholarPubMed
Hale, T et al. (2020) Variation in government responses to COVID-19. Blavatnik school of government working paper, 31. Available at https://www.bsg.ox.ac.uk/research/publications/variation-government-responses-covid-19 (Accessed 22 July 2020).Google Scholar
Cetron, M and Simone, P (2004) Battling 21st-century scourges with a 14th-century toolbox. Emerging Infectious Diseases 10, 20532054.CrossRefGoogle ScholarPubMed
Gensini, GF, Yacoub, MH and Conti, AA (2004) The concept of quarantine in history: from plague to SARS. The Journal of Infection 49, 257261.CrossRefGoogle ScholarPubMed
Tognotti, E (2013) Lessons from the history of quarantine, from plague to influenza A. Emerging Infectious Diseases 19, 254259.CrossRefGoogle ScholarPubMed
Anderson, RM et al. (2004) Epidemiology, transmission dynamics and control of SARS: the 2002–2003 epidemic. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 359, 10911105.CrossRefGoogle Scholar
Ferguson, NM et al. (2006) Strategies for mitigating an influenza pandemic. Nature 442, 448452.CrossRefGoogle ScholarPubMed
Jang, WM, Jang, DH and Lee, JY (2020) Social distancing and transmission-reducing practices during the 2019 coronavirus disease and 2015 Middle East respiratory syndrome coronavirus outbreaks in Korea. Journal of Korean Medical Science 35, e220.CrossRefGoogle ScholarPubMed
Mari, L et al. (2012) Modelling cholera epidemics: the role of waterways, human mobility and sanitation. Journal of the Royal Society, Interface 9, 376388.CrossRefGoogle ScholarPubMed
Gog, JR et al. (2014) Spatial transmission of 2009 pandemic influenza in the US. PLoS Computational Biology 10, e1003635.CrossRefGoogle ScholarPubMed
Velasco, E et al. (2014) Social media and internet-based data in global systems for public health surveillance: a systematic review. The Milbank Quarterly 92, 733.CrossRefGoogle ScholarPubMed
Bansal, S et al. (2016) Big data for infectious disease surveillance and modeling. Journal of Infectious Diseases 214, S375S379.CrossRefGoogle ScholarPubMed
Wesolowski, A et al. (2012) Quantifying the impact of human mobility on malaria. Science 338, 267270.CrossRefGoogle ScholarPubMed
Finger, F et al. (2016) Mobile phone data highlights the role of mass gatherings in the spreading of cholera outbreaks. Proceedings of the National Academy of Sciences 113, 64216426.CrossRefGoogle ScholarPubMed
Bengtsson, L et al. (2015) Using mobile phone data to predict the spatial spread of cholera. Scientific Reports 5, 8923.CrossRefGoogle ScholarPubMed
Wesolowski, A et al. (2015) Impact of human mobility on the emergence of dengue epidemics in Pakistan. Proceedings of the National Academy of Sciences of the USA 112, 1188711892.CrossRefGoogle ScholarPubMed
Google. COVID-19 Community Mobility Report. Available at https://www.google.com/covid19/mobility?hl=en (Accessed 12 September 2020).Google Scholar
Kapitány-Fövény, M and Sulyok, M (2020) Social markers of a pandemic: modeling the association between cultural norms and COVID-19 spread data. Humanities and Social Sciences Communications 7, 97.CrossRefGoogle Scholar
Smith, RD (2006) Responding to global infectious disease outbreaks: lessons from SARS on the role of risk perception, communication and management. Social Science & Medicine 63, 31133123.CrossRefGoogle ScholarPubMed
Aktay, A et al. (2020) Google COVID-19 Community Mobility Reports: Anonymization Process Description (version 1.0). arXiv:2004.04145 [cs]; Published online: 9 April 2020.Google Scholar
Dong, E, Du, H and Gardner, L (2020) An interactive web-based dashboard to track COVID-19 in real time. The Lancet Infectious Diseases 20, 533534.CrossRefGoogle Scholar
Nie, X et al. (2020) Epidemiological characteristics and incubation period of 7015 confirmed cases with coronavirus disease 2019 outside Hubei province in China. The Journal of Infectious Diseases 222, 2633.CrossRefGoogle ScholarPubMed
Scrucca, L et al. (2016) Mclust 5: clustering, classification and density estimation using Gaussian finite mixture models. The R Journal 8, 289317.CrossRefGoogle ScholarPubMed
Kraemer, MUG et al. (2020) The effect of human mobility and control measures on the COVID-19 epidemic in China. Science 368, 493497.CrossRefGoogle ScholarPubMed
Chan, HF et al. (2020) Risk attitudes and human mobility during the COVID-19 pandemic. arXiv:2006.06078 [econ, q-fin]; Published online: 10 June 2020.Google Scholar
Chan, HF et al. (2020) Can psychological traits explain mobility behavior during the COVID-19 pandemic? CREMA Working Paper Series. Center for Research in Economics, Management and the Arts (CREMA), June Report No.: 2020–08.Google Scholar
Sohrabi, C et al. (2020) World Health Organization declares global emergency: a review of the 2019 novel coronavirus (COVID-19). International Journal of Surgery 76, 7176.CrossRefGoogle Scholar
Lau, JTF (2003) Monitoring community responses to the SARS epidemic in Hong Kong: from day 10 to day 62. Journal of Epidemiology & Community Health 57, 864870.CrossRefGoogle Scholar
Sadique, MZ et al. (2007) Precautionary behavior in response to perceived threat of pandemic influenza. Emerging Infectious Diseases 13, 13071313.CrossRefGoogle ScholarPubMed
SteelFisher, GK et al. (2010) The public's response to the 2009 H1N1 influenza pandemic. New England Journal of Medicine 362, e65.CrossRefGoogle ScholarPubMed
Khan, K et al. (2009) Spread of a novel influenza A (H1N1) virus via global airline transportation. New England Journal of Medicine 361, 212214.CrossRefGoogle ScholarPubMed
Apolloni, A, Poletto, C and Colizza, V (2013) Age-specific contacts and travel patterns in the spatial spread of 2009 H1N1 influenza pandemic. BMC Infectious Diseases 13, 176.CrossRefGoogle ScholarPubMed
Colizza, V et al. (2009) Estimate of novel influenza A/H1N1 cases in Mexico at the early stage of the pandemic with a spatially structured epidemic model. PLoS Currents 1, RRN1129.CrossRefGoogle ScholarPubMed
Merler, S and Ajelli, M (2010) Human mobility and population heterogeneity in the spread of an epidemic. Procedia Computer Science 1, 22372244.CrossRefGoogle Scholar
Bajardi, P et al. (2011) Human mobility networks, travel restrictions, and the global spread of 2009 H1N1 pandemic. PLoS ONE 6, e16591.CrossRefGoogle ScholarPubMed
Buckee, CO et al. (2020) Aggregated mobility data could help fight COVID-19. Science 368, 145146.Google ScholarPubMed
Oliver, N et al. (2020) Mobile phone data and COVID-19: missing an opportunity? arXiv:2003.12347 [cs]; Published online: 27 March 2020.Google Scholar
Jia, JS et al. (2020) Population flow drives spatio-temporal distribution of COVID-19 in China. Nature 582, 389394.CrossRefGoogle ScholarPubMed
Zhou, Y et al. (2020) Effects of human mobility restrictions on the spread of COVID-19 in Shenzhen, China: a modelling study using mobile phone data. The Lancet Digital Health 2, e417e424.CrossRefGoogle ScholarPubMed
Badr, HS et al. (2020) Association between mobility patterns and COVID-19 transmission in the USA: a mathematical modelling study. The Lancet Infectious Diseases 20, 12471254.CrossRefGoogle ScholarPubMed
Bonaccorsi, G et al. (2020) Economic and social consequences of human mobility restrictions under COVID-19. Proceedings of the National Academy of Sciences 117, 1553015535.CrossRefGoogle ScholarPubMed
Chinazzi, M et al. (2020) The effect of travel restrictions on the spread of the 2019 novel coronavirus (COVID-19) outbreak. Science 2020, eaba9757.Google Scholar
Zhu, D et al. (2020) Social distancing in Latin America during the COVID-19 pandemic: an analysis using the Stringency Index and Google Community Mobility Reports. Journal of Travel Medicine, taaa125.CrossRefGoogle ScholarPubMed
Apple Inc. Mobility Trend reports. Available at https://covid19.apple.com/mobility (Accessed 21 September 2020).Google Scholar
Kissler, S et al. (2020) Social distancing strategies for curbing the COVID-19 epidemic. Published online: March 2020.CrossRefGoogle Scholar
Valenti, VE et al. (2020) Social distancing measures may have reduced the estimated deaths related to Covid-19 in Brazil. Journal of Human Growth and Development 30, 164169.CrossRefGoogle Scholar
Figure 0

Table 1. Continent-level summaries of maximum Kendal's τ correlations

Figure 1

Table 2. Continent-level summaries of cross-correlations

Figure 2

Fig. 1. Maximum absolute τ values. Results of Kendall's τ cross-correlation between COVID-19-confirmed case number and measures of community activity. Strong continent-wide regional patterns are apparent. Generally for the four categories indicative of mobility (‘retail and recreation’, ‘grocery and pharmacy’, ‘workplace’ and ‘transit’) strong negative correlations were observed across countries of North America, Russia, Australia, India and Western Europe. Positive relationships are seen in the South Americas, Eastern Europe, India and Southern Africa. For ‘residential’ activity, which is indicative of increased sedentary behaviour, the opposite was generally observed. For ‘parks’ the picture was mixed, possibly reflecting the difference nature of legal restrictions on a country by country basis; some countries implemented lockdown while others did not, some permitted outdoors exercise, others not [10].

Figure 3

Fig. 2. Lags to maximum correlations. Amount of time lagging in days resulting in the maximum Kendall's τ between COVID-19 for confirmed case number and measures of community activity (colour online only). Interesting are that the strongest correlations were when case numbers were negatively lagged by amounts of −20 days or greater for large areas across North America, Western Europe, Central Asia and Russia for the four categories indicative of mobility. This suggests that reductions in mobility in such areas occurred substantially prior to corresponding increases in COVID-19 case numbers. This is thus likely to have been substantially prior to formal legislation imposing movement restrictions coming into place. This indicated that personal behavioural choices and perceived risk perception may have played a greater role in driving movement patterns than legal restrictions.

Figure 4

Table 3. Model summaries

Supplementary material: File

Sulyok and Walker supplementary material

Sulyok and Walker supplementary material

Download Sulyok and Walker supplementary material(File)
File 475.4 KB