Requests for Ridehailing During an Extreme Weather Event: Exploratory Analysis of New York City

Numerous previous studies have explored the impact of extreme weather events on transportation system performance. These include studies assessing changes in passenger flows and shifts between various transportation modes (e.g., Budnitz et al. 2018; Chang and Nojima 2001; He and Liu 2012; Stamos et al. 2015). One study assessed changes in commute trips after Hurricane Sandy in New York (Kontou et al. 2017). Another noteworthy study examined how travelers access mobility information during an extreme weather event (Brazil et al. 2017). Other specially planned or unplanned events, such as sports games and severe accidents, have also been studied (Cottrill et al. 2017; Pereira et al. 2015).

There is a growing body of literature utilizing new data sources to study the impact of extreme weather events on transportation systems. Three of these studies are notable because they focus on extreme weather events in New York City. One study compared travel behavior in New York during a major snowstorm with normal weather by analyzing taxi Global Positioning System (GPS) pickups and drop-offs in Manhattan (Qing et al. 2015). Another study examined transportation system recovery rates in New York City during two major hurricanes: Hurricane Sandy and Hurricane Irene. Taxi trip data and subway turnstile ridership data were two “big data” sources that were analyzed, and the results showed lower recovery rates for Hurricane Sandy compared with Hurricane Irene (Zhu et al. 2016). A third recent study of New York City analyzed taxi GPS data to evaluate transportation system resilience and recovery patterns during Hurricane Sandy. The methodology involved calculating taxi trip pace during and after the hurricane and comparing this with trip pace in normal weather. Disruptions in traffic conditions lasted for more than 5 days and caused a peak delay of 2 min/mi (Markou et al. 2017).

In summary, there is a growing body of literature that uses new data sources to assess transportation system disruptions, particularly from extreme weather events, and the following analysis aims to contribute to this by focusing on ridehailing.

Ridehailing services provided by transportation network companies (TNCs) such as Uber and Lyft have grown extremely rapidly in the last decade. Recent estimates suggest that there will be approximately 4.2 billion TNC rides in the United States by the end of 2018 (Schaller Consulting 2018). In light of this trend, transportation operators and planners want to understand how travelers use these new services.

There is a small but growing body of related literature pertaining to ridehailing services. Due to a lack of publicly available data from TNCs, many of the leading studies on the use of ridehailing services utilize data from survey-based methods. One of the earliest studies of ridehailing users was published in 2015. A survey of people in San Francisco compared ridehailing with taxi services, and the authors concluded that more than half of ridehailing trips replaced modes other than taxis, including public transit and driving trips (Rayle et al. 2016). A more recent study of ridehailing in California conducted a survey of millennials, and a key finding was that 15% of millennials used ridehailing at least once a month (Circella et al. 2018). Another recent study conducted ridehailing passenger interviews in the Denver region; an important finding was that more ridehailing users are substituting ridehailing for transit than driving alone (Henao 2017). One last recent study included results from an online survey conducted in seven major metropolitan areas: Boston, Chicago, Los Angeles, New York, San Francisco, Seattle, and Washington, DC (Clewlow and Mishra 2017). The results revealed that, in large cities, 21% of adults use ridehailing services, and nearly a quarter (24%) of ridehailing users in cities use ridehailing services on a daily or weekly basis. This study also concluded that parking is the top reason that ridehailing users in cities substitute ridehailing for driving (37%), and avoiding drinking and driving is another reason for choosing ridehailing instead of driving (33%).

In summary, there are a few recent studies examining how ridehailing services such as Uber are used in major metropolitan areas; however, to the best of the authors’ knowledge, none have examined use of ridehailing services during extreme or unusual events.

The remaining literature reviewed includes studies utilizing data from the Transit smartphone application, which is the data set used in this study. First, a comprehensive overview of the Transit app data set —including how users interact with the app and how data are collected and structured—was conducted by Brakewood et al. (2017). Another study performed an exploratory analysis of Transit app usage in New York City and compared self-reported home locations from users with household socioeconomic data from the 2010 Census; the results suggest that travelers use the app Transit regardless of household income, race, or age trends in their neighborhood in New York City (Ghahramani and Brakewood 2016). A third study focused on a specific feature within the app Transit for trip planning and compared the origin–destination pairs typed into the trip planning feature of the app Transit against the Regional Household Transportation Survey from New York (Davidson 2016).

Two studies using Transit app data are particularly relevant to this analysis. The first study examined the Uber feature of the app Transit, which allows users to “request” an Uber, and compared this data to one of the few publicly available Uber trip data sets, which is from the New York City Taxi and Limousine Commission (TLC). The results revealed that the rate of Uber requests via the app Transit were higher within 250 ft of subway stations compared with Uber pickups from the larger TLC data set (Davidson et al. 2017). However, this study of Uber requests did not consider usage during extreme or unusual events.

The second study analyzed user interaction data from the app Transit to assess overall utilization during a severe weather event—Winter Storm Jonas—which is also the focus of this analysis (Remy et al. 2018). Overall app usage was compared between three cities: New York City, Washington, DC, and Philadelphia. Although this paper focused on overall app use during extreme events, the authors briefly noted that the app users continued to request Uber services via the app Transit during the snowstorm, despite travel bans. This surprising finding implies that ridehailing use during extreme events could be a fruitful area for additional analysis, and this study aimed to begin to fill this gap in the literature by conducting an exploratory analysis of the Uber requests during Winter Storm Jonas.

A major snowstorm hit the northeastern United States from January 22 to January 24, 2016. Several metropolitan regions along the eastern seaboard experienced record amounts of snowfall, including New York City where approximately 66 cm (26 in.) of snow accumulated (National Oceanic and Atmospheric Administration 2016). Because of the severe nature of the storm, officials in the New York metropolitan region closed the aboveground subway lines and bus network and banned travel on the roads in New York City on Saturday, January 23. Fig. 1 summarizes the critical changes to the transit network and Uber services during the snowstorm in the New York City region (New York State 2016; New York City Metropolitan Transportation Authority 2016a, b, c; CBS New York 2016; NJ Real-Time News 2016).

The time frame for this study consists of 6 weeks from January 1 to February 11, 2016. This includes 3 weeks before and 2 weeks after Winter Storm Jonas (as well as the week of the storm from January 22 to 29, 2016). This period of analysis includes two other special events: New Year's Day and Martin Luther King, Jr. Day.

The New York metropolitan area was selected as the area of analysis because it has high usage of the app Transit, high levels of transit ridership (McKenzie and Rapino 2011), and substantial use of Uber services (Schaller Consulting 2018). The study area was defined using the counties served by the New York Metropolitan Transportation Council (NYMTC) and the North Jersey Transportation Planning Authority (NJTPA) (NYMTC 2016; NJTPA 2016). Boundaries of this region were obtained by the maximum and minimum coordinates (latitude and longitude) of the region, and a bounding box was drawn around that area. The New York metropolitan area and the corresponding bounding box are shown in Fig. 2.

This study focuses on a free smartphone application known simply as “Transit.” The app Transit was originally developed in 2012 in Montreal, Canada, to offer nearby transit information for iPhone users. Later, an Android version of the app Transit was created, and the app was extended to several other cities. In 2019, the app Transit was available in more than 175 regions across 12 counties, and in most cities, it provides real-time transit information and trip planning while also integrating shared mobility modes such as Uber.

This analysis focuses on the Uber feature of the app Transit, which allows users to “request an Uber” by clicking on an icon within the app. Users are then directed to Uber's app for booking, which would need to be installed on the user's smartphone. It is important to emphasize that booking data (i.e., completed Uber trips) are not included in the Transit app data set used in this paper. In other words, the act of clicking on the “request an Uber” feature in the app Transit does not guarantee that the Uber trip was completed. Instead, this data more likely represents a desire or intention to use Uber. In addition, the “request an Uber” feature displays the estimated wait time for an Uber vehicle based on the Transit app user' data set does not include destinations for most records.

Last, it should be noted that the Uber feature in the app Transit was recently enhanced so that users can now book an Uber ride directly within the app Transit; booking data such as trip origin and destination from the updated Uber feature was not available to the authors.

The raw data set for this study was provided by the app developers directly to the authors. This data set is not publicly available; however, Transit does share data with other organizations. For example, Transit has a partnership with the Massachusetts Bay Transportation Authority in which the agency promotes Transit as its preferred app, and as part of the agreement, the app developers provide data to the transit authority to be used for planning purposes (Metro Magazine 2016).

The raw data set contained files in Java Script Object Notation (JSON) format. The files were then converted into Comma Separated Values (CSV) format. Two CSV files were the focus of this analysis. The first file collected information on the interactions of all users who opened the app Transit regardless of the app features that they used. This file includes coordinates (latitude/longitude) based on the location services in the user's smartphone, a unique identifier (ID) associated with the smartphone (called the “Device ID”), a unique ID associated with the app interaction (called a “Session ID”), and the date and time of the interaction. The second file used in this analysis is a subset of the first one, which pertains specifically to interactions with the “request an Uber” feature. When a user opened the app, if s/he clicked on “request an Uber,” an interaction was stored in this file. It should be noted that there were some minor issues with storing data for January 30 and 31; therefore, all of the data associated with these 2 days were removed from the data set.

Finally, it is important to emphasize that this data set does not contain any personal information, such as email address or phone number, to protect the privacy of Transit app users. In addition, no demographic or personal information is required nor stored by Transit. Last, all of the following analyses presented in maps, graphs, and tables in this paper were conducted at an aggregate level (either spatially or temporally) to protect users' privacy.

In this section, requests for Uber were examined visually from January 1 to February 11, 2016.

To further explore temporal trends, the rate of use of the “request an Uber” feature in the app Transit was calculated by dividing hourly usage to mean hourly usage obtained in normal weather conditions. This calculation iswhere R(t) = rate of use of the “request an Uber” feature per hour; Q(t) = mean Uber requests per hour during normal weather conditions from January 1 to January 21 and from January 29 to February 11; and U(t) = Uber requests for each hour during the analysis period (from January 1 to February 11). The results of the rate of Uber usage calculation are shown in Fig. 6. In this graph, the hourly rate of use R(t) that was calculated from Eq. (1) is shown in black, and the mean is a red line (i.e., 100% of the baseline). The vertical yellow and blue lines indicate the snowstorm and the transit shutdown periods, respectively.

As shown in Fig. 6, the rate of Uber requests, R(t), was distorted during the snowstorm and a few days after it. On the morning of the second stormy day, Saturday, January 23, the rate of usage is almost four times the mean value; then, when the travel ban was imposed, it decreased below the mean. A huge spike in Uber requests occurred on Sunday morning, and the number of Uber requests in the app Transit reached nearly 1,000% compared with normal weather. This large increase is probably because the travel ban was lifted but subway services were still partially closed on Sunday morning. On Monday, January 25, the first weekday after the snowstorm, the rate of Uber requests during the morning commuting period reached almost 500% of normal usage. The rate of Uber requests remained above 100% until Tuesday, January 26, and then it returned to normal on Wednesday, January 27 with some small fluctuations. Another noteworthy increase in Uber requests was observed after the stroke of midnight on New Year's Eve; however, no noteworthy anomalies were shown on the Martin Luther King, Jr. holiday.

In this part, Uber requests per hour are decomposed to assess temporal trends. Decomposition is a statistical method that deconstructs time series data into several components, and each component represents an underlying pattern. Time series data are commonly decomposed into the seasonal, trend-cycle, and remainder/error/noise components.

The specific decomposition method used for the following analysis is called seasonal trend decomposition using Loess (STL). Loess is a method for estimating nonlinear relationships. Notably, STL is robust to outliers so unusual observations do not affect the seasonal and trend components (Hyndman and Athanasopoulos 2018). The following STL decomposition was done using the open source statistical programming language R.

Additive decomposition was selected for this analysis, and Eq. (2) shows the formulationwhere y_t = number of Uber requests via the app Transit per hour, t; S_t = seasonal component; T_t = trend-cycle component; and R_t = remainder component.

Fig. 7 shows the decomposition of Uber requests via the app Transit from January 1 to February 11, 2016, by week. In this figure, the top graph shows the original data, the three graphs below this show each component (seasonal, trend, and remainder, respectively), and the gray vertical bar on the right-hand side of each graph depicts the scale. As can be seen in the second graph depicting seasonality, weekly and daily patterns can be observed that are similar throughout the study period. The trend-cycle graph shows a general increasing and then decreasing pattern around and after the storm. The remainder graph at the bottom of Fig. 7 illustrates the fluctuations after removing the seasonal and trend-cycle patterns. As shown in Fig. 7, the peaks that were observed in the original Uber requests data (top graph)—such as on New Year's Day and the last days of the snowstorm—are also observed in the remainder graph (bottom graph). Last, the variations shown in the remainder graph (bottom graph) are similar to the variation in the original data (top graph) because the scale bars for these two graphs are similar in length; this implies that the spikes in the number of Uber requests on New Year's Day and around the snowstorm period are not seasonal spikes.

To further investigate the strength of each of the three components comprising the time series, a statistical summary for each component was extracted, and the results are shown in Table 1. The ranges (min–max) in Table 1 show that the seasonal and trend-cycle patterns have smaller ranges compared with the remainder component; this implies that the remainder component has a substantial influence on the data. The remainder is largely the result of two peaks on January 1 (corresponding to the New Year's holiday) and January 24 (corresponding to the extreme weather event). This is consistent with the previously discussed comparison of the scales in Fig. 7.