Using volunteered geographic information to assess mobility in the COVID-19 pandemic context: cross-city time series analysis of 41 cities in 22 countries from March 2nd to 26th 2020 CURRENT STATUS: UNDER REVIEW

Background: Restricting mobility is a central aim for lowering contact rates and preventing COVID19 transmission. Yet the impact on mobility of different policies of restriction is not well-understood. Methods: Trends were evaluated using Citymapper’s mobility index covering 41 cities worldwide between 2 nd and 26 th March 2020, expressed as percentages of typical usage periods from 0% as the lowest and 100% as normal. China and India were not covered. Multivariate fixed effects models were used to estimate the association of policies restricting movement on mobility before and after their introduction. Policy restrictions were assessed using the Oxford COVID-19 Government Response Stringency Index as well as measures coding the timing and degree of school and workplace closures, transport restrictions, and cancellation of mass gatherings. Results: Mobility declined in all major cities throughout March. Larger declines were seen in European than Asian cities. The COVID-19 Government Response Stringency Index was strongly associated with declines in mobility (r = -0.75, p<0.001). After adjusting for time-trends, we observed that implementing a mobility restriction to the recommended level was associated with a decline of mobility of 10.0% for school closures (95% CI: 4.36% to 15.7%), 15.0% for workplace closures (95% CI: 10.2% to 19.8%), 7.09% for cancelling public events (95% CI: 1.98% to 12.2%), 18.0% for closing public transport (95% CI: 6.74% to 29.2%), 13.3% for restricting internal movements (95% CI: 8.85% to 17.8%) and 5.30% for international travel controls (95% CI: 1.69 to 8.90). In contrast, as expected, there was no association between population mobility changes and fiscal or monetary measures or emergency healthcare investment. Conclusions: Understanding the effect of public policy on mobility is crucial to slowing and reducing COVID-19 transmission. By using Citymapper’s mobility index, this work provides the first evidence about trends in mobility and the impacts of different policy interventions, suggesting that closure of public transport, workplaces and schools are particularly impactful.


Introduction
Even before the advent of germ theory, authorities employed measures to restrict movement to reduce the spread of disease, exemplified by the introduction of quarantine by the Venetian Republic.
The emergence and global spread of COVID-19 has given new prominence and importance to such measures, although in a world that has changed enormously in the past five centuries. On the one hand, the development of motorised transport has vastly increased the opportunities for mobility. On the other, the spread of mobile phones, of which the world's 7.8 billion people now own an estimated 14 billion,(1) has generated a wealth of information about those who are moving. As governments worldwide adopt measures unimaginable even a few months ago to restrict the intermixing of people with the intention to control the pandemic, there is a need to find novel ways to evaluate their effectiveness.
The data collected from mobile phones can include not only whom users call but also users' location.
At first, this was obtained by triangulating data from phone masts, but this has been supplemented by data from the geographical positioning system (GPS) chips included in smartphones. (2) The data collected provide a rich source of information on patterns of travel and are used extensively by software companies, often via apps that users have permitted to track their location, for example so that users of Uber and other taxi companies and their drivers can locate each other, and so that public transport users can use an app to discover when the next bus is due. The volunteered geographic information (VGI) generated is increasingly used by transport planners to develop routes and timetables. (3) The richness of these data and their ability to track movement in real time have attracted the attention of epidemiologists. For example, combining phone-derived mobility data with information on climatic conditions was shown to improve the ability to predict the spread of dengue in Pakistan. (4) Similarly, mobile phone data improved the ability to model schistosomiasis patterns in Senegal (5) and HIV in Côte d'Ivoire. (6) There are, however, a number of limitations, the most important of which is that data may be unavailable due to concerns about privacy, as was the case in west Africa during the Ebola outbreak. (7) Inevitably, epidemiologists are now looking to this source of data to help understand the dynamics of the COVID-19 virus pandemic and, especially, to evaluate the impact of countermeasures, many of which seek to reduce mixing of people. Thus, researchers in Seattle have used data from Facebook Data for Good, a resource developed for tracking mobility during disasters, to show that restrictions on movement were very effective in reducing journeys into the city.(8) The Swedish not for profit organisation Flowminder has been established to facilitate the availability of mobile phone data for those responding to health crises. However, most of these initiatives are, in practice, limited to ad hoc studies in single locations. (9) While these can answer the important question of whether a particular countermeasure was effective in reducing mobility in a particular place, individually they cannot show what happens in others. While it is, in theory, possible to obtain data from multiple locations, this would involve a major logistic exercise to do so from many different telecommunications companies in countries with different legal frameworks.
While recognising the importance of taking context into account, we wanted to know whether countermeasures have been effective across settings. We have identified one source of VGI collected across many different locations, collated by the developers of the Citymapper app. Citymapper gathers large amounts of open source data generated by transport authorities, local transit authorities and users, and processes it to provide users with information on the easiest and fastest way of getting from one place to another. It is used primarily by those making journeys by public transport, walking, cycling, and taxis, with the areas covered reflecting the availability of open data but, generally, covering entire metropolitan areas. Users enter their starting point (manually or linked to the GPS function in their phone) and destination and the pp offers them a choice of routes, including taxis and public transport. Walking instructions are included, using data from OpenStreetMaps. A 2018 report stated that it then had approximately 20 million users but there are no publicly available data on their characteristics. (10) In response to the pandemic, the developers have published data on mobility in cities worldwide, derived from the searches undertaken on the app. Mobility is measured by numbers of journeys planned but CityMapper advise us that a large proportion of planned journeys can be linked to actual journeys made using their GO function and the two are highly correlated, numerically and spatially. It is, as far as we know, the only publicly available data of this type from multiple locations in a large number of countries. Findings from individual cities have been reported in the media(11) but, so far, they have not been related to the timing of implementation of COVID-19 virus countermeasures. We sought to examine the feasibility of using Citymapper data to capture changes in mobility and relate them to the timing of pandemic countermeasures.

Methods
We extracted mobility from Citymapper's mobility index for the period 2nd to 26th March 2020. (12) We included all 41 cities in 22 countries worldwide covered by CityMapper but we note that they do not have any cities in China or India (Figure 1). The mobility index, published since the onset of the COVID-19 pandemic, reports the volume of trips planned using the app on a particular day compared with a "recent typical usage period". In most cities this was the four weeks between 6 th January and 2 nd February 2020, but in Paris it was 3 rd February-1 st March and Hong Kong/Singapore 2 nd -22 nd December 2019. Thus, the index can go above 100%, reflecting above average mobility, to below 100%, corresponding to lower than average mobility.
Data on the social restrictions were taken from the Oxford COVID-19 response dataset.(13) This dataset codes restrictions on movements including school and workplace closures, transport restrictions, and cancellation of mass gatherings, in practically all countries since 1 January 2020.
Each intervention is coded as 0 for no measures, 1 for measures that are recommended and 2 for those that are mandatory (although coding varies among variables, reflecting the nature of different measures). In addition, the Oxford team has compiled a composite index termed the COVID-19 Government Response Stringency Index. Most of the analyses are undertaken at the level of cities, applying the national data on countermeasures to all cities in the country.

Statistical modelling
First, we present descriptive time trends for mobility over time. Then we evaluate city-specific slopes ('fixed effects' modelling) comparing the changes in mobility before and after the introduction of alternative mobility restrictions. Furthermore, we adjust statistical models for potential secular timetrends, pre-existing mobility rates and other time-invariant factors that could account for betweencity differences. Robust standards errors were clustered by city to adjust for spatial correlation within nations. All analyses were conducted using Stata v15.1.
Patient and public involvement: There was no patient or public involvement in this study.
Data sharing: All data are available from the Citymapper website Transparency statement: The lead author (the manuscript's guarantor) affirms that the manuscript is an honest, accurate and transparent account of the study being reported, and that no important aspects of the study have been omitted.
Role of the finding source: No additional funding was required. Open access fees will be paid by the London School of Hygiene & Tropical Medicine.
Dissemination to participants and related patient and public communities: The results will be provided to UK and European COVID-19 bodies that the authors are advising. where the pandemic started, all cities were already exhibiting reduced mobility by the beginning of the period but further declines were relatively small compared to those seen in the rest of the world.

Trends in mobility during the COVID-19 pandemic
The two Australian cities and Sao Paolo followed similar trajectories to those seen in Europe.
We can see the importance of differential timing of mobility restrictions, including potential anticipatory effects (Figure 2). Eleven municipalities in the Lombardy region, whose capital is Milan, were placed in lockdown on 21 st February. This was extended nationwide on 9 th March. Consistent with this, mobility had already fallen in Milan but the reduction in Rome was somewhat later. In contrast, in the United Kingdom, restrictions were introduced nationwide and all three British cities moved in the same way. Notably, in both countries, mobility was falling well before the formal implementation of restrictions.
Next, we evaluated the link between stronger government restrictions on movement and the Citymapper Mobility Index.

Associations with mobility of alternative policy restrictions on movements
In univariate analyses, all of the measures reported were associated with reduced mobility (see Appendix 3). However, when including them jointly in the model (Table 1), several were no longer associated and in other cases the magnitude of association diminished slightly (** p < 0.01). Clearly, there is a risk of multicollinearity as many measures were implemented simultaneously, but diagnostic tests revealed a variance inflation factor of 1.85, indicating that this was not a major problem.

Robustness checks
We performed a series of sensitivity and robustness checks. First, we assessed the external validity of Citymapper data against other data on mobility. In Appendix 1, we plot the changes in the Citymapper Index in London with changes presented by the British Cabinet Office in journeys in London, collected by the Department of Transport,(14) finding that it follows a very similar trajectory.
The data for Seattle also closely matched what was reported in the study cited above that used Facebook Data for Good, (8) although in that case there were only four data points for comparison.
When indexed on 5 th March, the first day in common in both datasets, values for three of the four subsequent dates were within one percentage point of each other.

Principal findings
There is a critical need to understand the impact of measures intended to restrict mobility during epidemics. Our analyses used one publicly available measure, from Citymapper, to capture changes in mobility in 41 cities worldwide and relate them to the imposition of pandemic countermeasures.
Our findings demonstrate that several policy restrictions, notably closures of public transport, workplaces and schools, had a substantial impact on reducing population mobility. Our finding of no such impact of fiscal or monetary measures (which were hypothesised not to have mobility impacts) adds to the specificity and therefore the plausibility of our findings. (15)

Strengths and weaknesses of the study
Our study has several important limitations. Most importantly, the data published by Citymapper are generated from within what is essentially a black box. The company has, however, given us some additional information and, following our request, published some of this on their website. Our robustness checks did, however, show that, at least in two cities, the changes observed correlated with those found using other data sources.
Second, there may be specific local factors that would need to be considered in interpreting the data.
For example, mobility patterns in Hong Kong may have been reduced over the past year due to the pro-democracy demonstrations. Third, Citymapper has only published its Mobility Index since the beginning of the pandemic and we have no data from before 2 nd March. Fourth, although the data include cities from across the world, there are some notable gaps, such as mainland China and India.
Fifth, as the data are user-generated, there may be selection bias due to the characteristics of those using the app. Thus, another study using mobile phone data showed how, in New York City, there was a substantially greater reduction in mobility among residents of wealthy areas than poor ones, reflecting the greater ability of the former to work from home. (16) Finally, while mobility is a reasonable proxy for contact rates, our study has not attempted to model the subsequent impact of mobility restrictions on incidence or transmission of COVID-19. There is some evidence, for example, that in the case of COVID-19 and other novel corona viruses, school closures may have smaller impacts on transmission than is the case with established influenza virus infections.(17) Future research is needed to operationalise mobility measures as a parameter of contact rates in standard susceptible-exposed-infected-recovered (SEIR) or SEAIR (susceptible, exposed, asymptomatic, infectious, removed) epidemiological models of COVID-19 spread, with findings feeding into models of the impact of non-pharmaceutical interventions, such as that developed by Imperial College. (18) Citymapper is only one source of and, as the pandemic progresses, new sources are coming online, although often not in ways that make it easy to conduct analsyes such as the ones we conducted.
Thus, Google has published a series of COVID-19 Community Mobility Reports but the graphs are in pdf format and it is not clear that the data are available. (19) The Citymapper data seem, as far as we can ascertain, to be unique in being in the public domain and covering cities worldwide. This makes it possible to do the analyses we have performed and thus understand the consequences of different policies. This is important given the high social costs of these policies, which will inevitably weigh upon politicians called on to make difficult choices.

Meaning of the study
Turning to the results, our findings such that restriction policies are essential for rapid and dramatic reductions in population mobility, evidenced for example in the steep decline in mobility in the UK after implementation of restricctions compared with the more modest previous reductions following government advised but not mandated behaviour change. Our data cover the period of initial implementation of restriction policies; it is uncertain whether coercive policies are sustainable given the possibility of social unrest particularly if measures are perceived as socio-economically inequitable. (20) Our analyses suggest that closure of public transport, workplaces and schools achieved substantial reductions in mobility. However, further work is needed to examine how and if these policies could generate unintended consequences including: undermining health systems and other essential services, due to staff's loss of transport or childcare; economic hardship arising from loss of earning; and increasing children's social contact with grandparents.(21) Given these considerations, it is likely that politicians will look to pragmatic policies, such as: reducing but not closing public transport services; workplaces and schools remaining open in the case, respectively, of essential services and the children of key workers; (22) and rapid and generous income maintenance programmes.

Future Research
Unsurprisingly, there is relatively little other research with which to compare these findings.
Exceptions include recent reviews of the effect of school closures on disease transmission, finding that this measure has some but not a great impact, (17)

Supplementary Files
This is a list of supplementary files associated with this preprint. Click to download. Appendix.pdf