4.2.2 Parameter Estimates: Factors Influencing Ridesourcing Usage During Pandemic
Weather and weekend
The results of RENB models exhibited that snowfall and weekend variables had significant positive relationships with ridesourcing trips. At the same time, temperature showed a significant negative relationship in both models, which was consistent with the findings of previous studies (Ghaffar, Mitra, and Hyland 2020). These indicated that regardless of the pandemic, ridesourcing was adhered to weekend and increased snowfall days but appeared as a less lucrative modal option with increasing temperature. However, while precipitation produced fewer ridesourcing trips during the normal time, it produced more trips during the pandemic. Moreover, the results of Models III-V suggested that increased precipitation, temperature, and snowfall led to a more negative change, i.e., greater decline of ridesourcing trips and total daily trip duration during the pandemic than regular time.
Such drop in ridesourcing usage could partly be explained by the share of non-work trips by ridesourcing services, in general. The 2017 National Household Travel Survey shows that of all the ridesourcing trips, non-work trips constitute the biggest share. Being flexible and discretionary in nature, non-work trips are more likely to be dropped by any extreme weather condition, be it increased precipitation or increased maximum temperature or increased snow.
Table 2
Parameter estimates of RENB models
Variables | Dependent Variables |
I. Total Daily Trip (2019) | II. Total Daily Trip (2020) |
Weather | | |
Precipitation | -0.031*** | 0.117*** |
Max. temperature | -0.002*** | -0.013*** |
Average Snowfall | 0.013*** | 0.051*** |
Weekend vs. Weekday | | |
Binary 1 if the day is weekend, otherwise 0 | 0.170*** | 0.031*** |
COVID-19 Features | | |
No. of daily new cases in county | - | -0.025*** |
No. of daily cumulative cases in county | - | -0.004*** |
COVID-19 Policy | | |
Pre-COVID: Binary 1, if the date corresponds to pre-COVID period, otherwise 0 | - | -0.102*** |
Emergency: Binary 1, if the date corresponds to state of emergency order, otherwise 0 | - | -0.431*** |
Stay at home: Binary 1, if the date corresponds to stay at home period | - | -1.614*** |
Socio-Demographic Characteristics | | |
Median age | -0.005*** | -0.005*** |
Income (baseline: Low median income) | | |
Medium median income in census tract | 0.110*** | -0.063*** |
High median income in census tract | -0.01 | -0.399*** |
Car-ownership and Commute Characteristics | | |
No. of zero car households in census tract (in 1000s) | 659.011*** | 522.573*** |
% workers who commute by public transit in census tract | 0.007*** | 0.005*** |
% workers who commute by carpool in census tract | 0.002** | -0.005*** |
Built Environment Variables | | |
Ln (Population density) (count/sq. ft) | 0.286*** | 0.159*** |
Ln (Employment density) (count/sq. ft) | 0.062*** | 0.019*** |
Land-use mix | 0.138*** | 0.091*** |
Number of restaurants in census tract | 0.020*** | 0.017*** |
Parking Feature | | |
Parking rate ($/hour) in census tract | 0.127*** | 0.188*** |
Transit Supply Features | | |
No. of CTA bus stops in census tracts | 0.009*** | 0.025*** |
Number of CTA rail stations (baseline: zero rail station) | | |
Binary 1: if census tract has 1 rail station | -0.006 | -0.053*** |
Binary 1: if census tract has 2 rail stations | 0.003 | -0.152*** |
Binary 1: if census tract has 3 rail stations | -0.746*** | -0.559*** |
Binary 1: if census tract has 4 or more rail stations | -2.054*** | -2.135*** |
Crime Records | | |
Number of homicides (baseline: zero homicide) | | |
Binary 1: if number of homicide record is between 1–5 | 0.019** | -0.017 |
Binary 2: if number of homicide record is more than 5 | 0.270*** | 0.240*** |
Constant | -2.000*** | -0.018 |
Model Fits and Diagnostics | | |
Wald Chi2 (27) | 26987.43*** | 6378.68*** |
LR test vs. Pooled (Chi squared) | 4.8E05*** | 3.1E05*** |
Ln_r | -0.286*** | -0.251*** |
Ln_s | 1.438*** | 1.262*** |
Wooldridge test for serial autocorrelation F (1,790) | 447.39*** | 496.23*** |
Hausman test for Fixed vs Random Model1, Chi-square (9, 4) | 4.68 | 7.38 |
Breusch and Pagan Lagrangian multiplier test for random effects Chi-square (01) | 1.1E06*** | 2.6E07*** |
N (Number of observations) | 278,478 |
n (Number of groups) | 792 |
T (Number of days) | 364 |
Note: * Significance at 10%, **Significance at 5%, *** Significance at 1%. 1. insignificant Hausman test statistic means a random effect is preferable to a fixed effect model.
Table 3
Parameter estimates for random effects panel regression models
Variables | Dependent Variables1 |
III. Δ Total Daily Trip | IV. Δ Total Daily trip Mile | V. Δ Total Daily Trip Duration |
Weather | | | |
Δ Precipitation | -2.136*** | 0.071 | -0.539*** |
Δ Max. temperature | -0.532*** | -0.024*** | -0.071*** |
Δ Snowfall | -11.557*** | -0.684*** | -2.207*** |
Weekend vs Weekday | | | |
Binary 1 if either of the corresponding date of 2019 and 2020 is weekend | -83.038*** | -3.270*** | -11.960*** |
COVID-19 Features | | | |
No. of daily new cases in county | -0.015*** | -0.001*** | -0.003*** |
No. of daily cumulative cases in county | -0.314*** | -0.015*** | -0.043*** |
COVID-19 Policy | | | |
Pre-COVID: Binary 1, if the date corresponds to pre-COVID period | 79.375*** | 3.964*** | 15.094*** |
Emergency: Binary 1, if the date corresponds to state of emergency order | -56.917*** | -2.430*** | -9.377*** |
Stay at home: Binary 1, if the date corresponds to stay at home period | -61.771*** | -2.647*** | -7.259*** |
Socio-Demographic Characteristics | | | |
Income (baseline: Low median income) | | | |
Medium median income in census tract | -1.090 | 1.607 | 2.715 |
High median income in census tract | -225.254*** | -3.38 | -24.920*** |
Car-ownership and Commute Characteristics | | | |
No. of zero car households in census tract (in 1000s) | -5.3e + 05*** | -2.2e + 04*** | -8.4e + 04*** |
% workers who commute by carpool in census tract | 7.554*** | 0.347*** | 1.231*** |
% workers who commute by public transit in census tract | 5.143*** | 0.088 | 0.538* |
Built Environment and Land-use Variables | | | |
Ln (Population density) | 58.089** | 13.804** | 30.179** |
Ln (Employment density) | -37.175*** | -2.913*** | -8.665*** |
Land-use mix | 99.113 | 24.820*** | 53.904** |
Number of restaurants in census tract | -17.448*** | -1.470*** | -4.133*** |
Parking Feature | | | |
Parking rate ($/hour) in census tract | -351.052*** | -13.824*** | -53.386*** |
Transit Supply Features | | | |
No. of CTA bus stops in census tracts | 1.638 | 0.604* | 1.221 |
Number of CTA rail stations (baseline: zero rail station) | | | |
Binary 1: if census tract has 1 rail station | -33.076 | -0.628 | -2.856 |
Binary 1: if census tract has 2 rail stations | -322.841 | -1.436 | -28.153 |
Binary 1: if census tract has 3 rail stations | -521.328 | -16.751 | -81.949 |
Binary 1: if census tract has 4 or more rail stations | -4.7e + 03*** | -163.635*** | -735.024*** |
Crime Records | | | |
Number of homicides (baseline: zero homicide) | | | |
Binary 1: if number of homicide record is between 1–5 | 36.031 | 2.088 | 7.154 |
Binary 2: if number of homicide record is more than 5 | 34.474* | 2.863* | 8.007* |
Constant | -183.294 | -117.491* | -231.314* |
Model Fits and Diagnostics | | | |
Overall R Square | 0.56 | 0.48 | 0.54 |
sigma_u | 244.89 | 15.96 | 46.05 |
sigma_e | 348.87 | 19.14 | 60.11 |
rho | 0.33 | 0.41 | 0.37 |
Wald Chi square (27) | 18803.84*** | 12942.24*** | 15330.98*** |
Breusch-Pagan Lagrangian multiplier test for random effects | 5.5e + 06*** | 8.4e + 06*** | 6.8e + 06*** |
Hausman test for Fixed vs Random Model2, Chi-square (9, 4) | 11.60 | 14.04 | 13.10 |
Test for heteroskedasticity: Chi square (791) | 1.4e + 12*** | 7.7e + 07*** | 5.2e + 07*** |
Wooldridge test for autocorrelation in panel data: F (1, 790) | 708.91*** | 23.77*** | 54.23*** |
N (Number of observations) | 278,478 | | |
n (Number of groups) | 792 | | |
T (Number of days) | 364 | | |
Note: * Significance at 10%, **Significance at 5%, *** Significance at 1%. 1. Since the change value could be negative, a negative coefficient indicates that a larger independent variable leads to a lower (more negative) relative change, i.e., a greater decrease. 2. insignificant Hausman test statistic means a random effect model is preferable to a fixed effect model.
COVID-19 features and COVID-19 policy
As expected, the number of daily new COVID-19 cases and cumulative cases in the county, imposition of the state of emergency, and stay-at-home orders all presented a significant negative relationship in Model II – V, indicating that the spread of COVID-19 across the counties and the subsequent COVID-19 policies contributed to the decline of ridesourcing trips during the pandemic. Moreover, the results of Models III-V suggested that the state of emergency and stay-at-home policy triggered a more negative change, i.e., a greater decline in total daily ridesourcing trips, trip miles, and trip durations during the pandemic than regular time. However, contrast existed for the pre-COVID period. This exhibited a significant negative sign in Model II, potentially implying people being cautious about ridesourcing usage in early 2020, even before the first COVID-19 case was reported in Chicago. However, these variables were positive in Model III – V, indicating a lesser decline in total daily ridesourcing trips, trip miles, and trip durations during the pre-COVID period. Such findings were intuitive since rising daily new cases, and cumulative cases probably made the ridesourcing users concerned about the possible spread of the COVID-19 virus, as perceived by a higher risk of contagion in areas with more cases. Once the state of emergency and stay-at-home orders were in place, individual and collective movements was reduced to the minimum, contributing to lower supply and demand for ridesourcing services.
Socio-demographic characteristics
The model results (Model I and II) indicate a negative relationship between the median age of the residents across the census tracts and ridesourcing usage. This result complies with the findings (Dias et al. 2019a; Alemi et al. 2019) of the existing literature on ridesourcing usage that refers to younger travelers being more likely to make frequent ridesourcing trips as opposed to older travelers.
Compared to census tracts with low median income, census tracts with medium income presented a significant positive sign in Model-I, while they exhibited a significant negative sign in Model II. Though census tracts with high income did not appear significant in Model I, but portrayed a significant negative relationship in Model II. Additionally, census tracts with high median income were more likely to experience a greater decrease in ridesourcing trips and trip durations. These results indicated that regions with low-income residents generated more ridesourcing trips than the rest during the pandemic, and the opposite was true for the regular period. There are two plausible explanations for these opposite results between the regular period and the pandemic. First, as low-income residents tend to rely more on public transportation (Pucher and Renne 2003) and during the pandemic, many public transportation services reduced or closed their services (Qi et al. 2021) low-income residents might use more ridesourcing services as an alternative. Second, higher-income ridesourcing users of the pre-covid period might switch to personal cars during the pandemic as high-income residents disproportionately switched from other modes to cars during the pandemic (Brough, Freedman, and Phillips 2021). However, the results of ridesourcing usage among different income groups during the regular period were in line with the findings of previous studies (e.g., Ghaffar, Mitra, and Hyland 2020). It is to be noted that we could not include other socio-demographic variables, such as educational attainment, because of multicollinearity issues.
Car-ownership and commute characteristics
The number of households with no car and the percentage of workers commuting by public transit presented a significant positive relationship in Model I and II, indicating that regardless of the pandemic, regions with a higher number of households with no car and a higher share of public transit commuters continued using ridesourcing services. This finding aligns with the conventional expectation of higher ridesourcing demand in areas with more zero-car households (Blumenberg, Paul, and Pierce 2021; Ghaffar, Mitra, and Hyland 2020). There are two possible explanations for this result: first, the ridesourcing services might continue working as a complementary mode to public transit during the pandemic, as previous studies reported that zero-car households with no or lower access to transit are more likely to rely on ridesourcing to access transit (Sikder 2019). Second, ridesourcing services might be used as a substitute mode for public transit during the pandemic. Since during the pandemic, many public transportation services were closed or reduced their services, those who used to commute by public transit switched to ridesourcing services. However, a closer look at the coefficients for the share of workers commuting by public transit in the 2020 model indicated a slight drop in the popularity of ridesourcing in those areas during the pandemic. The possible reason behind the lower magnitude of the coefficients in the pandemic model might be associated with reduced commuting trips due to the work-from-home policy during the pandemic.
The percentage of workers commuting by carpool exhibited a statistically significant positive sign in Model I but switched to a significant negative in Model II, meaning that areas with a higher share of workers commuting by carpool were less likely to use ridesourcing services during the pandemic. The pre-pandemic model results corroborate prior findings (e.g., Ghaffar, Mitra, and Hyland 2020). The possible reason for the negative sign in the 2020 model might be the overall reduction of commuting trips or modal switch to personal cars by the carpool commuters. The results of Model III-V indicated that a smaller relative decline in ridesourcing usage was attributed to a higher percentage of workers commuting by carpool or public transit compared to a greater decrease in areas with a higher number of households with no car. One possible explanation for this finding could be linked to the overall mobility drop during the pandemic. Since usual mobility plummeted to a minimum, many households with no car drastically curtailed their ridesourcing usage for discretionary trips, or they probably shifted to other transportation modes perceived as safer than ridesourcing services (Boston and Colias 2020).
Built environment
All the built environment and land use variables appeared significant and positive in Model I-II, implying higher population and employment density, land use mix and number of restaurants in the census tracts generate higher ridesourcing usage, irrespective of the pandemic. Results of Model III-V suggest that areas with higher employment density and more restaurants were attributed to a greater decrease in ridesourcing usage, whereas a smaller decrease in ridesourcing usage was attributed to higher population density and land use mix. Findings on employment density sounded plausible and intuitive because of the working-from-home policy during the pandemic. For example, compared to 2019, Chicago experienced a 90% hike in remote workers in service sectors, in the second quarter of 2020 (Federal Reserve Bank of Chicago, 2020), implying places with higher employment density could be tied to a higher share of workers working from home during the pandemic. The explanation for the greater decrease in ridesourcing usage in areas with more restaurants could be attributed to the state of emergency and stay-at-home orders when restaurants remained closed for several months.
Parking feature
As expected, the parking rate presented a positive and statistically significant relationship in Models I-II, while presenting a negative relationship in Model III-V, indicating that census tracts with higher parking rates are more likely to encourage ridesourcing usage, irrespective of the pandemic. However, a greater decrease in ridesourcing trips being originated in areas with higher parking rates was more likely during the pandemic. This parking finding has important implications for shorter trips being made with ridesourcing services. Of all the short ridesourcing trips (i.e., trips made within census tract) made in 2019, census tracts around downtown Chicago constituted the biggest share (Du, Rakha, and Breuer 2021). Hence, shorter non-work trips originating from the downtown areas where parking rates were higher, were more likely to be affected during the pandemic, resulting in a greater decline in ridesourcing usage.
Transit supply features
Number of bus stops and rail stations in tandem refers to transit infrastructure. The results of these two variables exhibited the same statistical significance with opposite signs. Regardless of the pandemic, census tracts with three or more CTA rail stations were negatively associated with ridesourcing demand, while the number of bus stops was positively associated with ridesourcing usage. The results of the pre-pandemic period are consistent with Ghaffar, Mitra, and Hyland (2020). The results in Model III-V exhibited a significant and negative relationship for census tracts with four or more rail stations, indicating a greater decline in resourcing trips, trip distances, and trip durations for regions with four or more rail stations. Though the number of bus stops did not appear significant in Model III and Model V, it portrayed a significant and positive relationship in Model IV, indicating a smaller decline in ridesourcing trip distances for regions with bus stops.
Crime records
The results of Model I-II indicated that census tracts with higher homicide records generated more ridesourcing trips both during the normal period and the pandemic. Travelers in places with higher homicide records and no personal vehicle are less likely to feel safe for shorter trips to and from grocery or transit stations that could otherwise be traveled by walk or biking. They might find ridesourcing as a safer alternative for such trips in places with higher homicide records compared to places with fewer homicide events. However, the results in Model III-V suggest that census tracts with homicide records of more than 5 contributed to a smaller decline in ridesourcing usage during the pandemic.