Transit demand.

Since it is difficult to obtain comprehensive public transit ridership data at a national-scale, we use data from the Transit mobile phone app (transitapp.com) as an indicator of changes in daily and hourly transit demand. Transit is a popular mobile phone app providing real-time public transit data and trip planning. The app covers over 200 cities around the world and more than 60 US metro areas (discussed below) with more than five million downloads on Android platform [13] and 73.5 thousand ratings on Apple App store [14]. We accessed daily demand data from Transit via their daily updated webpage: these are change values expressed as a set of percentage of app usage relative to the same date last year, adjusted for annual growth [15]. These estimates use app activity as a proxy for transit ridership demand.

Transit estimates their market share of transit users to be an average of 8% in the United States. To assess Transit app usage data as a measure of transit demand, we compared ridership decrease reports derived from individual transit systems' websites and local news outlets. Most transit systems do not release estimates daily; instead, many report single estimates for a given date. We compared these ridership decrease reports with the corresponding estimates from the Transit app data on the same dates for 40 transit systems with published ridership reports. The average difference between the Transit app estimate and agency reported value is 3.7%; a paired T-test indicates that we cannot reject the null hypothesis that the mean difference is zero (p = 0.14 > 0.05). However, the standard deviation of the differences is 15.96%; this may be due to the varying definitions of normal ridership level among agencies. Although the test suggests the Transit app data are a good approximation of public transit demand, it is important to note that both the Transit app data and transit agencies’ data are inferences. Ridership counts from agencies, especially daily ridership counts (as opposed to monthly) vary in methodology. Moreover, agency counts during COVID-19 are particularly prone to error due to back-door only boarding, fare-free service, and other pandemic-related changes that affect the accuracy of ridership counts. Transit app activity does not include individuals who cannot afford a smart phone and data plan, cannot use the app due to different abilities, or choose not to use it. It is likely that our measures are underestimates of transit demand decline, especially among more disadvantaged social groups and older populations. Compensating for these disadvantages is the large Transit app user base across most transit systems in the United States, allowing comparison across transit systems using a common benchmark.

The daily Transit data include demand decrease estimates for 182 public transit systems across the United States, Canada, Australia, New Zealand, and France. We selected 113 county-level transit systems in 63 metro areas and 28 states across the United States. We excluded 7 state-level or cross-county systems if their ridership could draw from large and geographically diverse areas, such as Pacific Surfliner, which extends to the entire South California coast, and Metro-North Railroad, which crosses multiple counties and states in the Northeast United States. The time period of daily data is from February 15^th to May 17^th. We also use hourly demand decrease estimates for 93 public transit systems across the United States. The time frame for the hourly data is from March 16^th to May 10^th. None of the data collected or accessed in this study in this study contain individual personal information.

COVID-19 case numbers.

We use the daily case numbers for each county from COVID-19 maps and county-level dataset produced by USAFacts [16]. The data include all county-equivalents’ confirmed cases in the US on a daily basis. To find the linkage between the case numbers and the demand decrease, we geocoded each transit system to its corresponding county-equivalent.

Base value.

The parameter B represents the relative decrease from the normal level as demand re-stabilizes at a lower level; we define it as base value as visualized in Fig 2. This represents the ratio of public transit users in the system who still use the transit system regardless of the pandemic. This demand level is not necessarily a persistent state: demand may destabilize and grow again due to external factors, such as re-opening of businesses or stay-at-home fatigue. The base value represents the base level of demand from the initial shock to the system.

We examine relationships between the estimated base values and socioeconomic factors using linear regression analysis. The county-level socioeconomic data are from the latest American Community Survey (ACS) 5-year estimate table (2014–2018). We derived several socio-economic indicators. First is the ratio of population with non-physical occupations. Similar to life fixity [10], this measures the population’s degree of freedom to change the routine of their daily life: it represents how many people can work from home thus avoid regular transit commuting to reduce contagion risk. If a community has higher ratio of non-physical jobs, more workers have the freedom to work from home, meaning that transit demand may decrease more than communities with lower ratios. We use the occupations statistics for employed civilian population 16 years and over from the ACS data. The table contains the number of employed population in different industries. According to the statistics released by US Bureau of Labor Statistics, Information, Financial activities, and professional and business service have the highest proportion who can work from home [5, 17]. Among these occupation categories in the ACS table, we assigned “management, business, and financial operations occupations”, “professional and related occupations”, and “office and administrative support occupations” as the occupations where people can work from home remotely.

Income, population density, and job density are socioeconomic factors that relate to job composition. Also, transit users tend to skew toward lower incomes in the United States; higher incomes also correlate higher population density and job density at the community scale. We use the median income, population density, and employment density data from ACS.

A third set of indicators is the ratios of minority and female population, including African American and Hispanic populations. Many studies demonstrate the disadvantaged status of African American, female, and Hispanic populations in mobility and job accessibility across different metro areas in the United States [18–20]. Therefore, we investigate the relationship between base value and each vulnerable group’s ratio. We derive the sex, race, and Hispanic population data from the ACS data.

A fourth socioeconomic indicator is community age structure. Older individuals are more at risk of hospitalization and death from COVID-19 [21]. We use 45 years as a threshold to identify high-risk populations. We measure the ratio of people with age over 45 from the ACS 5-year estimates (2014–2018) data.

Moreover, we use measures of awareness. If local residents are aware and concerned about COVID-19, the base value may be lower because more people will avoid public transit trips. To test this, we use the Google search trend index to represent the awareness of the local people [22–24]. We collected the average Google search trend data for different designated market area that each transit system locates in for 90 days from January 18^th to April 17^th 2020 [25], the latter being the latest day that any system experienced further significant decline. We use “Coronavirus” over “COVID19” as the search keyword for its popularity among the public [26].

Finally, transit dependency is also a potential factor affecting the transit system's usage rate [27] and base value. If an area has more people depending on transit, the usage rate of transit during the pandemic is supposedly higher. We derive the ratio of people who transit to work and the percentage of house units with no vehicle access from ACS data to infer transit dependency.

To supplement the analysis, we also refer to the user survey results conducted by Transit app, Inc. about the demography of the passengers during the pandemic [28, 29]. The survey was conducted in early April, 2020 across the United States (n = 15000) and Canada (n = 10000) via the Transit app interface. The survey investigates the age, race (including Spanish speakers), gender, trip purpose, occupation composition of the passengers who remained transit system users during the pandemic [15, 28, 29].

Cliff/base points and decay rate.

The cliff point t_c and base point t_b are time points when demand decline started and when it re-stabilized as shown in Fig 2. We calculated these from the fitted logistic curves using confidence interval theory rather than the observed data to provide more stable estimates. We derived these measures by first constructing the probability density function of the normalized logistic function F(x): (2) (3) (4) where P is the probability density function of the normalized logistic function; to normalize the logistic function, we subtract the normal value b and divide the result by B to construct the P function so that . α is the confidence threshold. From Eqs (2)-(4), we can see that ; α = 0.05 ensures that 95% of the decline falls between the cliff and base points. From the formula, the cliff and base points are: (5) (6) The cliff point is the first day when the demand curve began to diverge from normal—i.e., when transit users start to avoid the transit. The base point is the day when decline slows and transit demand has re-stabilized. We expect the cliff point to be impacted by policies and actions in local communities. We therefore consider the date when each state declared a state of emergency due to COVID-19 and conduct correlation analysis between the emergency declaration date and the cliff point.

Parameter k in Eq (1) represents the rate of transit demand decline; we therefore define it as decay rate. It indicates the speed of response from users who have the ability to stay at home or not use public transit. We conduct correlation analysis between decay rate and cliff/base points to examine the relationship between de-stabilization, re-stabilization and the rate of demand decline.

Response intervals with incubation lags.

Response intervals compare the time of community spread with the initiation (cliff point) and conclusion (base point) of transit demand decline in each system. This measures the responsiveness of transit demand to the pandemic. Although declines in transit demand are not welcome from a revenue perspective, lower demand means fewer people potentially exposed on transit; it also means the remaining dependent riders are less exposed and can practice social distancing more easily. Ideally, a transit system initiates and finishes its demand decline before there is community spread.

Meanwhile, the date of first reported community spread is not necessarily the first date of actual spread due to the incubation period for the disease. The median of incubation period is five days and can be as long as 14 days [30]; the virus can also spread asymptomatically [31–33]. Additionally, the availability of testing kits and response times from local authorities can lead to an underestimation of total cases [34, 35]. Since the incubation period and case load accuracy both vary, we introduce incubation lag as a parameter of the response intervals relative to the cliff and base points: (7) (8) where t_s is the date of first confirmed case in the county of the transit system; l is the incubation lag; t_c and t_b are the cliff and base point; and r_c and r_b are the response intervals relative to the base point and cliff point, respectively. A positive response interval means that the transit users responded earlier than the epidemic spread; the larger the value is, the earlier the transit users responded relative to the onset of the spread of the virus. The cliff response interval indicates transit users’ awareness while the base version indicates the response time. For example, a negative base response interval that is large in magnitude indicates that public transit reached its minimum level long after the first confirmed case in the county.

Population with non-physical occupations.

The results confirm the hypothesis that greater decreases in transit demand are associated with a higher percentage of people with non-physical occupations. People who can work at home avoid public transit; people who cannot work at home and rely on public transit continue to use it.

Although we did not include the Hispanic population indicator due to multicollinearity, a significant negative correlation between ratio of Hispanic population and ratio of population with non-physical occupations suggests vulnerability of Hispanic populations during this health crisis: if a community has a higher Hispanic population, it is likely for the community to have a higher base value, which means more people use transit during the pandemic (presumably for work). This is also consistent with the occupation statistics: according to the labor force characteristics survey made by the US Bureau of Labor Statistics, the Hispanic population has the lowest percent (22%) of management, professional, and related occupations compared with White (41%), African American (31%), and Asian people (54%) in 2018 [37]. The correlation result also applies to the median income, population density, and employment density, which are all correlated with the ratio of population with non-physical occupations.

User surveys conducted by Transit app in April 2020 support these results. According to the survey, 92% of respondents who still use public transit report that they still use it to commute to work [28]. Meanwhile, the top-four occupation categories that are most likely to still using transit are production, installation, maintenance and repair, food preparation and service, and protective service. Although the categorizations of the survey and the ACS data are different, this is generally consistent with the occupation categories we derive from the ACS data.

The Transit survey also indicates that Spanish speakers are more likely to continue using the Transit app for trip planning purposes: English-language users dropped 71% from early February while Spanish-language users dropped by 50% over the same time period [29]. The income correlation is also confirmed by the survey results: compared with the survey results conducted by American Public Transportation Association (APTA) in 2017 [38], active users skew towards lower income brackets during the pandemic, especially for those whose annual income is less than $15000 [29]. The survey results provide first-hand evidence and reaffirm the correlation results about the vulnerability of Hispanic population and low-income population.

Age.

The ratio of the population over 45 years old is associated with higher base values; older people in a community indicate higher levels of continued transit use during the pandemic. The Transit user survey also supports this result. By comparing the users’ age composition in surveys conducted in September 2019 and April 2020, Transit found a drop in young people under 18 and between 25 to 44 years old; meanwhile, the relative ratio of people between 45 to 64 years old doubled [29].

African American and female.

The regression analyses also indicates the dependence of African Americans on public transit, even during a pandemic. It is the most influential among the factors in Table 1 based on model fit (R-squared). There is a strong positive correlation between the ratio of African Americans in the population and the base value. These results are also consistent with the results of the Transit user survey. During the pandemic, African American people have the greatest share (>35%) of riders compared with other races in the US, while Caucasians were the majority (>40%) of the rider before the pandemic based on the 2017 APTA survey [28, 38]. The disproportionately small decrease of the African American population’s transit demand supports the conclusion that cities with more African Americans are more likely to have a higher base value.

Higher base values are also highly correlated with larger ratios of females in the population; however, we also did not include it due to multicollinearity with the ratio of African Americans in the population. A higher ratio of females in the population is also correlated with lower income and a lower ratio of people with non-physical occupations. The Transit user survey supports these results in a dramatic manner. Among all the US users surveyed, the male and female proportions were roughly equal before the COVID-19 pandemic; during the pandemic, 56% are females while only 40% are males [28]. For some cities such as Philadelphia, more than 68% of riders are women. Meanwhile, Transit app users of color are also more likely to be females during the pandemic; more than 70% of the African-American riders during the pandemic are female [28].

Awareness.

The Google search trend index for the keyword “Coronavirus” is significantly associated with the base value; cities with a higher search index tend to have a lower base value, as a higher search index means “Coronavirus” has a higher ratio among all Google searches in that region. This indicates the effects of people’s awareness and concern on avoiding non-essential public transit trips. However, in Table 1, Google search trend index is not highly influential: it has the lowest R-squared improvement. This suggest that ethnicity, occupation, and age composition outweigh the awareness or preference when it comes to whether people will stay at home.

Transit dependency.

The Transit user survey found 85% of users do not have access to a car [28], supporting that transit passengers during the pandemic are mostly captive passengers. Surprisingly, the ACS data on people using transit to commute and households with no vehicles do not have a significant correlation with base values. This suggests that the ACS dataset may not be a good metric for transit dependency during an emergency such as a pandemic. The presence of a vehicle in a household does not mean it is reliable, affordable to operate, or available to a given household member. Also, transit dependency is heterogeneous in many US cities: while most residents are not transit dependent, there are neighborhoods with concentrated poverty and transit dependence [39, 40]. The user survey shows that access to private vehicles is highly heterogeneous for different household income levels [28].