The outbreak of SARS-CoV-2, a coronavirus that causes the disease COVID-19, has caused a pandemic on a scale unseen in a generation. Without an available vaccine to reduce transmission of the virus, public health organizations and elected officials have called on the public to practice social distancing. Social distancing is a set of practices in which individuals maintain a physical distance to reduce the number of physical contacts they encounter [1,2]. These practices include maintaining a distance of at least six feet from other people and avoiding large gatherings [3]. At the time of this writing, in the United States, nearly every state had implemented statewide “stay-at-home” orders to enforce social distancing practices [4].

Social distancing is an important tool in the fight against COVID-19; however, its implementation by the general public can vary widely. Although a state governor may issue an order for the practice, individuals in different states may respond to this order in different ways. Courtemanche et al [5] showed that social distancing policies in the United States reduced the daily growth rate of COVID-19 cases. However, if we only consider the social distancing policy duration and daily confirmed cases, it is difficult to rule out potential confounders, including additional policies for wearing masks and improving hygiene as well as other social norms. Therefore, understanding actual reductions in travel and social contacts is critical to measuring the effectiveness of such policies. Using mobile phone data, Badr et al [6] found that mobility patterns were strongly correlated with decreased rates of COVID-19 case growth for the 25 most affected counties in the United States. These social distancing policies may remain in effect for an extended period of time. Thus, the public may begin to relax their practices, making additional policies necessary. Researchers showed the effectiveness of strict social distancing followed by testing and contact tracing by modeling mobility data from Cuebiq Inc in the Boston metropolitan area [7]. Additionally, epidemiologists have already modeled the impact of social distancing policies on the course of disease outbreaks [8-10]. These models may be more effective when incorporating actual measures of social distancing rather than assuming that official policies are implemented in practice.

It can be challenging to obtain data on the efficacy of social distancing practices, especially during an ongoing pandemic. In a recent Gallup poll that surveyed Americans, it was found that many adults are taking precautions to maintain distance from others [11]. However, while polling can provide insights, it cannot provide a solution. Polling is relatively expensive; thus, it is a poor choice for ongoing population surveillance practices and providing data on specific geographic locales (ie, US states and major cities) [12]. Additionally, polling around public health issues suffers from response bias, as individuals may overstate their compliance with established public health recommendations [13].

Over the past decade, analyses of social media and web data have been widely adopted to support public health objectives [14]. In this vein, several efforts have emerged over the past few months to track social distancing practices using these data sources. Google has released COVID-19 Community Mobility Reports [15] that use Google data to “chart movement trends over time by geography, across different categories of places such as retail and recreation, groceries and pharmacies, parks, transit stations, workplaces, and residential.” The Unacast Social Distancing Scoreboard uses data collected from 127 million monthly active users to measure the implementation of social distancing practices [16]. Researchers at the Institute for Disease Modeling have used data from Facebook’s Data for Good program to model the decline in mobility in the greater Seattle area and its effect on the spread of COVID-19 [17]. Using mobile phone data, the New York Times completed an analysis that showed that stay-at-home orders dramatically reduced travel; however, it was found that in states where such orders were not quickly enacted, residents continued to travel widely [18].

Identifying and isolating individuals who have potentially been exposed to a virus can blunt the spread of a pandemic. Contact tracing involves finding people who have had contact with an infected individual during the time the individual was contagious. In the current pandemic, several efforts have been made to develop digital contact tracing tools. Google and Apple announced a joint effort to build a Bluetooth-based contact tracing platform, which enhances the interoperability between Android and IOS devices using apps from public health authorities [19]. Singapore [20] and Australia [21] released similar apps that use Bluetooth to exchange “digital handshakes” to establish contacts. Many countries have developed their own contact tracing responses [22]. Li and Guo [23] presented a review of the development of contact-tracing apps for COVID-19. These efforts provide new and important opportunities to study social distancing and contact tracing in real time.

We present the Twitter Social Mobility Index, a measure of social distancing and travel patterns derived from public Twitter data. We used public geolocated Twitter data to measure how much a user travels in a given week. We computed a metric based on the standard deviation of a user’s geolocated tweets each week, and we aggregated these data over an entire population to produce a metric for the United States as a whole, for individual states, and for some US cities. We found that in the United States as a whole, there was a dramatic drop in travel in the later weeks of the study period, with travel between March 16 and April 27, 2020, showing the lowest amount since January 1, 2019, the start of our data set. Additionally, we found that travel reductions were not uniform across the United States but varied from state to state. However, there was no clear correlation between social mobility and confirmed COVID-19 cases at the state level. A key advantage of our approach is that unlike the other travel and social distancing analyses referenced above, we rely on entirely public data, which enables others to replicate our findings and explore different aspects of these data. Additionally, because Twitter contains user-generated content in addition to location information, future analyses can correlate users’ attitudes, beliefs, and behaviors with changes in social mobility.

One concern regarding the mining of social media data is user privacy. Unlike the data used by the companies described above, all the data we used is publicly available. Users choose to post their location data to Twitter publicly; therefore, these data are accessible to all users. However, while the location data are public, the potential remains for violating user privacy and producing unintended consequences for users, such as highlighting users who are failing to social distance. To ensure privacy in our index, we aggregated all mobility metrics to produce population-level analyses. None of our work considers the identity of individual users, and we removed identifiable user information from the distributed data aggregations. Furthermore, we caution others who pursue work similar to ours to consider privacy ramifications for users when collecting new data and conducting similar analyses.

There is widespread recognition that real-time tweets from millions of users can yield insights into a variety of population-level trends. Our study follows a tradition of using this insight to develop population-level indices and measures from Twitter data. Previous work includes tracking population-level sentiment as an economic indicator that can track stock price [24], political indices that reflect the popular opinion on major socioeconomic issues [25] or opinions about political candidates [26,27], and measures of pop culture such as reception of entertainment programs [28]. The Twitter Social Mobility Index is a measure of this kind, aggregating Twitter data from millions of people to produce real-time measurements of social distancing.

There is a long line of work on geolocation prediction for Twitter, which requires inferring a location for a specific tweet or user [29-32]. This includes work on patterns and trends in geotagged Twitter data [33]. Although most of these works focus on inferences of users’ current locations and thus are not suitable for tracking user movements, there may be opportunities to combine these methods with our approach.

Many studies have analyzed Twitter geolocation data to study population movements. Hawelka et al [34] demonstrated a method for computing global travel patterns from Twitter, and Dredze et al [35] adapted this method to support efforts in combating the Zika virus epidemic. Several studies have used human mobility patterns from Twitter data [36-39]. These studies include analyses of urban mobility patterns [40-42]. Finally, some of these analyses considered mobility patterns around mass events [43].

Our findings are presented on a website [44], and we will continue to update our analysis during the COVID-19 pandemic.

We present the Twitter Social Mobility Index, a measure of social mobility based on public geolocated tweets. Our analysis shows that there was a large drop in mobility overall in the United States. However, the drop was inconsistent and varied significantly by state. It appears that states that were early adopters of social distancing practices experienced more significant drops than states that had not yet implemented these practices.

Several limitations of using geo-tagged tweets as the subject of our study must be kept in mind. First, users on Twitter and other social media platforms are not representative of the general population. Their demographics, such as age, race, ethnicity, education level, income, and political affiliation, do not perfectly mirror the larger population. In the United States, Twitter users are younger, more educated, have higher incomes, and are more likely to identify as Democrats than the general public [53,54]. Therefore, while our sample of users is large, it is highly biased.

Second, not all users are equally likely to use geotagging features on Twitter, and they may use the features in different ways. For example, in a previous study [32], demographic differences were found in the groups of people who used the two different types of geolocation information (ie, coordinates and place). GPS-tagged tweets are posted more often by young people and by women compared to tweets with self-reported locations.

Third, while we obtained access to millions of geotagged tweets, this is still a relatively small proportion of the total number of nongeotagged tweets on the platform, and it is also small compared to private measures of social mobility computed by companies such as Google and Apple.

Fourth, a small proportion of geotagged tweets report fake geolocation information. However, we believe that this is a negligible problem, as previous work found the rate of fake geolocation to be around 0.22% on social media in general [55] and even lower on Twitter. In our preliminary analysis, we considered mobility data based on GPS from mobile devices alone while excluding place information, as this method has greater precision. However, our results with these limited data were similar to our results with the full data set, except that they were less stable. Therefore, we decided to include all location data.

Despite these limitations, our results produced metrics that align with expected trends given national social distancing guidelines and related statewide policies. This suggests that there is sufficient information in our data to overcome these limitations. Additionally, the public nature of Twitter data has advantages over proprietary and private data sources. More work is needed to compare our mobility trends with those of other data sources.

Our work on this data is ongoing, and there are several directions that warrant further study. First, as states begin to reopen and some states maintain restrictions, tracking changes in population behaviors will be helpful in making policy decisions. Second, we focused on the United States; however, Twitter data provides sufficient coverage to replicate our analysis for many countries. Third, tweet content exists for each user in the data set; this content can reflect the user's attitudes, beliefs, and behaviors. Studying these factors together with users’ mobility reduction could yield further insights. Our findings are presented on a website [44], and we will continue to update our analysis during the pandemic.