is there a limit to1 adoption of dynamic …from june 2014 to june 2015, uber’s manhattan monthly...

0

IS THERE A LIMIT TO ADOPTION OF DYNAMIC RIDESHARING SYSTEMS? 1

EVIDENCE FROM ANALYSIS OF UBER DEMAND DATA FROM NEW YORK CITY 2 3 4

5 Raymond Gerte (Corresponding Author) 6 Department of Civil and Environmental Engineering 7 University of Connecticut 8 261 Glenbrook Road Unit 3037, Storrs, CT 06269 9

Tel: 203-258-3545 Email: [email protected] 10

11 Karthik C. Konduri 12 Department of Civil and Environmental Engineering 13

University of Connecticut 14 261 Glenbrook Road Unit 3037, Storrs, CT 06269 15

Tel: 860-486-2733 Email: [email protected] 16 17

Naveen Eluru 18 Department of Civil, Environmental and Construction Engineering 19 University of Central Florida 20

12800 Pegasus Drive, Room 301D, Orlando, FL 32816, USA 21 Tel: 407-823-4815 Email: [email protected] 22

23 24 25

26

27 28 29

30 31

32 33

34 35 Presented for Presentation and Publication to 36 ABJ70: Committee on Artificial Intelligence and Advanced Computing Applications 37 Submitted on August 1, 2017 38

39 40

41 42 43 Revised word count: 7,157 words + 2 Figures/Tables = 7,657 words 44 45 46

mailto:[email protected]



1 Gerte, Konduri, and Eluru

ABSTRACT 1 Recent technological advances have paved the way for new mobility alternatives within 2 established transportation networks, including on-demand ride hailing/sharing (e.g. Uber, Lyft) 3 and citywide bike sharing systems. One aspect that is common across these innovative modes is 4

a lack of direct ownership by the user; in each of these mobility offerings, a resource not owned 5 by the end users’ is shared for fulfilling travel needs. This concept has flourished and is being 6 hailed as a potential option for autonomous vehicle operation moving forward. However, 7 substantial investigation into how shared mode offerings impact travel behaviors and integrate 8 into existing transportation networks is lacking. This paper explores if the growth in the adoption 9

and usage of these modes is unbounded, or if there is a limit to the adoption of shared mode 10 offerings. Recent trends and shifts in Uber demand usage from New York City are investigated 11 to explore the hypothesis. Using publicly available data about Uber trips, temporal trends in the 12 weekly demand for Uber are explored on the Borough of Manhattan. A panel based random 13

effects model accounting for both heteroscedasticity and autocorrelation effects is estimated 14 wherein weekly demand is expressed as a function of a variety of demographic, land use, and 15

environmental factors. It is observed that demand appeared to increase initially after the 16 introduction of Uber, but it seems to have stagnated and weaned over time in heavily residential 17

portions of the island, contradicting the observed macroscopic unbounded growth. The 18 implications extend beyond already existing fully shared systems, but also affect the planning of 19 future mobility offerings. 20

21 22

23 24 25

Keywords: Dynamic Ridesharing, Uber, Shared Systems, Open-Source Data, Linear Panel Model 26

27


1. INTRODUCTION 1 Technology has long been the driver for advances in our everyday means of transport. The 2 development of data-enabled and location aware mobile phones, combined with adoption of 3 sharing economies (also referred to as collaborative economies or peer-to-peer networks), has 4

opened the door for novel modal offerings like Uber, Lyft, and citywide bike sharing systems. 5 These specific modal offerings are referred to generally as dynamic ridesharing or ridesourcing 6 systems. They capitalize on the growth of mobile phone adoption and integrated global 7 positioning (GPS) capabilities to actively link riders to drivers in an on-demand setting. 8 Historically, ridesharing has filled unique niches in the transportation landscape (carpooling, 9

preplanned trips, etc.) and has consistently been encouraged for its ability to minimize user costs 10 and reduce congestion. However, the widespread adoption of conventional ridesharing has been 11 linked with an inability to capture impromptu trips (1). Due to the fact that almost all trips need 12 to be prearranged, conventional ridesharing has substantial limitations and users that recognize 13

its benefits are unwilling to make the switch because of inconvenience. With the growth of 14 dynamic ridesharing, the ability to travel on demand in a manner similar to taxi trips opens a 15

wealth of mobility options. In this paper, the temporal trends in one ridesharing offering, namely 16 Uber, are explored. 17

As the current industry leader, Uber’s growth is readily apparent. At the close of 2014, 18 Uber operated in 230 cities worldwide, and by June 2015, Uber grew to operate in 300 cities (2, 19 3). In only six years of existence (2009 to 2015), Uber reported facilitating over one billion trips. 20

As of June 2016, Uber reported surpassing two billion total trips in 450 cities (4). The popularity 21 of Uber can be associated with its inherent flexibility and user-perceived efficiency of use. When 22

using the Uber app, users are able to hail a ride remotely from anywhere within the service city, 23 given that a driver is available. This is an improvement over the call ahead or on street hailing 24 methods of its most similar competitor, the taxi, while also offering service in locations that may 25

not have many viable alternatives. In a major metropolitan hub like New York City, the growth 26

of Uber is substantial. From June 2014 to June 2015, Uber’s Manhattan monthly June pickups 27 increased by over 250% with over one million more trips made using the dynamic ridesharing 28 app (5). 29

While Uber’s growth is noteworthy, there has also been substantial controversy in 30 regards to the ridesharing giant. Established taxi groups and regulators have filed numerous 31

lawsuits against Uber and other dynamic ridesharing corporations for infringing on market share 32 without the burden of for hire vehicle (FHV) licensure and regulation (6). There has also been 33

considerable debate on the ethics of the dynamic pricing incorporated into Uber’s app (7). 34 Furthermore, there is ongoing discussion regarding Uber’s safety, its policies concerning drivers, 35 and how these policies compare to similar regulations governing other FHV companies (8). The 36 growing impact of dynamic ridesharing is becoming an important piece in the planning of cities 37 and regions around the U.S. and elsewhere. These systems are not only expanding and growing 38

in the short term, but also have substantial implications for the future, namely in regards to the 39 deployment of autonomous vehicles (AVs). AV offerings are projected to perform in a shared, 40

on-demand manner, similar to ridesharing offerings. Also, shared AVs have been identified as 41 one of the key pieces in realizing the complete suite of AV benefits (9). 42

As such, understanding ridesharing can provide insight into how these dynamic systems 43 are currently performing and being received by the public, while also offering a possible window 44 into the impact of future modal offerings employing the same concept. Substantial investigation 45 into how dynamic ridesharing impacts travel behaviors and integrates into existing transportation 46


networks is lacking. To address these limitations, this study focuses on quantifying and 1

understanding temporal trends in Uber demand within a major metropolitan center. In particular, 2 the study explores if growth in the adoption and usage of these modes is unbounded or if there is 3 a limit to the adoption of dynamic ridesharing. In order to pursue this research, trip pickup data 4

made available by Uber through a Freedom of Information Law (FOIL) request by the analytics 5 group FiveThirtyEight, is utilized. Weekly demand, aggregated spatially to a taxi zone level, is 6 analyzed using a panel based random effects model while also accounting for potential 7 heteroscedasticity and autocorrelation issues. The demand is expressed as a function of various 8 demographic, land use, and environmental factors to further explore the trends. 9

The remainder of this paper is organized as follows: Section 2 summarizes the existing 10 research and positions the current study, Section 3 describes the data and presents the approach 11 to data preparation, Section 4 includes a descriptive analysis of the data, Section 5 outlines the 12 model used, Section 6 discusses the results, and finally, Section 7 provides concluding remarks 13

and offers some directions for future research. 14 15

2. EXISTING RESEARCH 16 The existing research on dynamic ridesharing demand, and analysis of demand trends in 17

particular, is quite limited. These limitations can be attributed to the lack of readily available data 18 from the providers of these modal offerings, and limitations in data that is publicly available. 19

The first vein of research helps to gather a more complete understanding of dynamic 20

ridesharing system performance in relation to available supply and demand. Chen et al. and Guo 21 et al. (7, 10, 11) explore the characteristics of Uber’s surge pricing and assess its impact on 22

supply and demand. The authors collected four weeks of comprehensive data on all Uber vehicle 23 types using the Uber App API and Client App API. The data gathered included information on 24 the supply of vehicles, the demand for vehicles, and the associated surge price multiplier in both 25

downtown San Francisco and Midtown Manhattan. They found that the majority of vehicles in 26

their study areas were concentrated in commercial and tourist areas within each study zone. 27 Further, they observed that surge pricing did in fact negatively impact demand while 28 encouraging supply at the study scale. Correa et al. (12) used the same data to quantify changes 29

in Uber and taxi demand over space and time and found substantial significant growth in demand 30 over time with high correlation between taxi and Uber modes in the core of the city. Increases in 31

roadway length, income, and job opportunities, as well as decreases in transit access time and 32 vehicle ownership, were observed with increased demand for Uber. 33

Another major line of research is the comparison between dynamic ridesharing and 34 traditional mode offerings such as taxi and transit. Rayle et al. (13) found that dynamic 35 ridesharing users are typically younger and more educated than those that use taxis. Also, while 36 ridesharing does remove potential trips from taxis, there is substantial evidence that users opted 37 to switch from available transit modes citing travel time as the reason. Survey respondents using 38

dynamic ridesharing were skewed towards no vehicle ownership with multiple users traveling 39 together. Researchers Cramer and Krueger (14) compiled a direct comparison of Uber’s 40

utilization rate (percentage of time drivers are serving passengers) against that of taxis in five 41 major U.S. cities. Their results showed that as a whole, Uber is the more efficient mode of 42 transportation compared to taxis. Poulsen et al. (15) explored the relationship between Uber and 43 green cabs in New York City and found that while demand for both modes is growing, Uber’s is 44 growing at a much faster rate and green cabs perform better in poor neighborhoods. These 45 studies highlight the fact that Uber is in fact similar to existing taxi infrastructure, however, they 46


highlight key differences that warrant further investigation into how people are using dynamic 1

ridesharing on a citywide scale. 2 The final set of literature pertaining to dynamic ridesharing and demand is informal in the 3

form of blog posts and news articles. These are mostly based on Uber demand information for all 4

of New York City made openly available by researchers at FiveThiryEight; the same data also 5 used in this paper. FiveThirtyEight researchers have published numerous articles on the topic by 6 analyzing the data. Fischer-Baum and Bialik (16) address how Uber’s growth should be 7 classified. They found that in the urban core of Manhattan, Uber trips seem to come almost 8 entirely from taxis, with the sum of total trips across both modes remaining the same across both 9

study years. Schneider (5) explored the trends in Uber demand and found that Uber pickups in 10 the urban core are growing, and notes that these pickups increase more substantially in times of 11 inclement weather compared to increases in taxi pickups. 12

The research highlighted above provides interesting insights into the demand for dynamic 13

ridesharing, but there are substantial gaps. One gap this work aims to fill is if growth in demand 14 for these systems is subject to saturation, or if it is unbounded. To this end, the study explores 15

demand over time, while also capturing the impact of demographic, built environment, and 16 environmental variables. In order to explore this idea quantitatively, this study develops a panel-17

based random effects model that is similar in nature to the works by Faghih-Imani et al. (17). 18 The authors study the relationship between NYC Citi Bike ridership, land use, and urban form 19 amongst other studies focused on Citi Bike and Taxi travel time. Their research focuses on three 20

main types of variables: those related to the environment, temporal variables, and land use 21 variables. A similar structure is also observed in an analysis of bus ridership with the 22

incorporation of real-time location information by Brakewood et al. (18), and work by Campbell 23 and Brakewood (19). These studies, along with some of the variables described in the work 24 above, served to motivate the modeling aspect of this research and helped guide selection of 25

potential demographic, land use, and environmental factors that may influence Uber demand. 26

27

3. DATA 28 As noted in the limitations regarding dynamic ridesharing research, restrictions in data 29

availability and quality have had a substantial impact on the volume of research. Recent adoption 30 of open-data initiatives (wherein data is made publicly available) and the practice of making data 31

available in response to public requests (through FOIL and other similar mechanisms) have made 32 much of this work possible. The Uber trip information was made available by Uber after a FOIL 33

request and was gathered from FiveThirtyEight’s GitHub repository (20). While these data are 34 the only source of information on Uber demand, they also have several limitations that dictated 35 the data preparation and approach to analysis. First, Uber data is only available for six months in 36 2014 (April through September) and six months in 2015 (January through June). Within this, trip 37 pickup location is recorded at a latitude/longitude level in 2014 and aggregated up to predefined 38

taxi zone level in 2015, including pickup date and time information. Due to the spatially 39 aggregated nature of 2015 data, taxi zone was chosen as the spatial unit of analysis. 40

Subsequently, all remaining data (demographic, land use, and environmental factors) were 41 spatially aggregated to the taxi zone level. Taxi zone delineations are defined by the NYC Taxi 42 and Limousine Commission and line up with neighborhood boundaries in New York City. This 43 analysis focused only on the taxi zones in the Borough of Manhattan. There are 69 unique taxi 44 zones in Manhattan which includes three zones for ferries to Ellis Island, Liberty Island, and 45 Governors Island, and four zones for major parks, namely, Central Park, Battery Park, Inwood 46


Hill Park, and Highbridge Park. To add context, there are 288 census tracts located in the same 1

study area. The boundaries of census tracts and taxi zones are coincident. However, taxi zones 2 are relatively larger in spatial extent compared to tracts and on average each taxi zone is 3 comprised of about four census tracts. Built environment variables were collected at the tax lot 4

level and gathered from the NYC Planning Departments website (21). Information on transit 5 stations/stops and bike share infrastructure was accessed through NYC open data hub (22). 6 Environmental variables like temperature and precipitation were gathered at the day level from 7 2014 through 2015 in Central Park from the National Climate Data Center hosted by the 8 National Oceanic and Atmospheric Administration (23). All demographic data was collected at 9

the census tract level from the American Community Survey’s five year estimates for 2014 and 10 2015 and aggregated to the taxi zone level. 11 Substantial data processing and data preparation was performed prior to the analysis. 12 First, the taxi zone aggregated Uber pickups were summed to the week level because of the high 13

level of variability at the day level. From the 12 months of initial data, there were 54 total weeks, 14 however, weeks that fell at the beginning and end of the dataset were incomplete. As a result, 15

these incomplete weeks were dropped leaving 49 total complete weeks of demand data. For the 16 remainder of the paper, the weekly pickup counts will be used to constitute the demand and will 17

be referred to as such. Daily weather information was treated similarly. Built environment 18 variables were assumed to not vary across the investigation period and the demographics of each 19 taxi zone were assumed to be represented by the 2014 and 2015 ACS estimates, respectively. 20

21

4. DESCRIPTIVE ANALYSIS 22 The growth in Uber overall has been well documented, in both the literature presented above and 23 in its emergent prevalence in the societal lexicon. When analyzing the data at the borough level, 24 substantial growth in Uber demand over time is easily observed. The total number of trips made 25

in Manhattan from April 2014 through September 2014 was just under 140 thousand per week. 26

In 2015, that number rose to just over 430 thousand trips per week from January through June. 27 However, it is important to note that this growth is not evenly distributed. At the taxi zone level, 28 the top five taxi zones in both 2015 and 2016 are as follows: Midtown Center, Union Square, 29

Tribeca/Civic Center, and both East and West Village. All of these zones are located either in 30 midtown or lower Manhattan and together amount for about 20 percent of the total trips in 2014 31

and 2015, respectively. In reviewing the shifts in demand by each taxi zone individually, 32 differences were noted in zones that were more residential. These differences were typically 33

manifested as lower overall demand each week, but the growth of that demand appeared lower 34 than that in commercial zones. This observation is reasonable because residential areas typically 35 have less tourism or visitor traffic. 36

To better understand overall demand and the temporal trends, the taxi zones were 37 classified into five groups based on a residential built environment measure derived from the 38

land use dataset. The first group, residential taxi zones, are those zones that have a higher ratio of 39 residential land use than commercial land use. The second group, commercial zones, are those 40

that have a higher ratio of commercial land use than residential. These two groups were then 41 further divided into those found in Northern Manhattan (north of 86

th street on the east side of 42

the island and north of Central Park) and those not. Taxi zones that were exclusively parks were 43 treated as the fifth group. Figure 1 highlights these trends with respect to four of the land use 44 areas classified above, excluding parks. From the map inset, it is shown that residential zones are 45 surrounding Central Park and the northern portion of the island, as well as in the south east. The 46


line shading for the demand plots range from dark (less residential/commercial) to light (heavily 1

residential/commercial). Additionally, residential and commercial dominant zones are shaded in 2 blue and orange hues respectively. Therefore, more mixed-use areas i.e. places with an even split 3 of residential and commercial land, are represented by the darker hues. Weekly trends show that 4

there is general increase in demand for all zones from 2014 to 2015. Looking only at the 5 commercial dominant taxi zones, a few notable trends are apparent. First, a banding effect can be 6 observed wherein adjacent zones experience similar levels of demand. The upper bands with the 7 most demand correspond to Midtown and surrounding taxi zones, while the remaining zones are 8 distributed further downtown. Almost all commercial zones located in central and southern 9

Manhattan have an upward trend even after noting seasonal decreases. Only one commercial 10 dominant taxi zone is found in the northern part of the island and that is Morningside Heights. 11 Home to Colombia University, this zone exhibits atypical behavior when compared to the other 12 commercial and residential zones. Interestingly, the substantial increases and decreases observed 13

in 2015 line up with typical semester beginning and ending points. 14 On the other hand, focusing in on the residential areas in central and lower Manhattan, 15

darker lines indicate the lesser residential of the residential areas. These zones have the largest 16 increase in demand of the residential zones and have higher volatility and clear seasonal shifts. 17

The lighter color lines indicate heavily residential zones and have fewer weekly trips and less 18 overall variability at the week level. Perhaps the most interesting observed trend within these 19 zones is that demand appears to stagnate over time in the heavily residential areas. The results 20

are similar in the residential zones in northern Manhattan. The more mixed zones have the 21 highest demand and growth while the most residential of the group fall within the middle of the 22

graph. Here too, the lines below the 1000 trip mark appear to decrease. The observations of 23 flattening demand lend credence to the hypothesis that growth in Uber demand may not be 24 unbounded and that it may be subject to a limit. We further explore this hypothesis using a 25

random effects panel model below. 26


1 The demand lines are shaded by the ratio of residential floor area to total floor area for the residential dominant zones, and by ratio of

commercial floor area to total floor area for the commercial dominant zones. Darker lines indicate a less intense relationship, while

lighter lines indicate a zone is almost entirely residential or commercial, respectively.

1 Figure 1: Distribution of Residential and Commercial Taxi Zones in Manhattan Along with Trends in Uber’s Weekly Pickup

Demand by Zone.1


5. PANEL MODEL 1 The nature of the weekly Uber demand data (i.e. panel) lends itself to a linear panel based 2 analysis (24). The demand data are available temporally by week for 25 complete weeks in 2014, 3 and 24 complete weeks in 2015. However, not all elements of the demographic, land use, and 4

environmental data used to explain the demand are time variant. As a result, a random effects 5 structure was employed that allows consistent estimation of coefficients associated with time 6 invariant variables. 7

Following the exposition of linear panel models in Cameron and Trivedi (24), an outline 8

of the model structure and assumptions are presented below. Let 𝑧 = 1, 2, … , 𝑍 represent an 9

index for the unique taxi zones and 𝑡 = 1, 2, … , 𝑇 represent an index for the week. The 10 dependent variable, weekly pickup demand, is then modeled using a linear panel structure 11 similar to the one below: 12

13

𝑦𝑧𝑡 = 𝛼𝑧 + 𝛽′𝑋𝑧𝑡 + 𝜀𝑧𝑡 14

Where 𝑦𝑧𝑡 represents the weekly pickup demand in taxi zone z at week t, 𝛼𝑧 captures the zone 15

specific unobserved heterogeneity, 𝑋𝑧𝑡 represents the vector of independent variables and time 16

variables, and 𝛽 represents the regression coefficients associated with the independent variables, 17

𝜀𝑧𝑡 represents the errors within each taxi zone 𝑧 for time period 𝑡. 𝜀𝑧𝑡 is assumed to be strictly 18

exogenous and normally distributed as 𝑁(0, 𝜎𝜀2). Further, specific assumptions on 𝛼𝑧 result in 19

the fixed effect and random effect variants. In a random effects model, used in the analysis, 𝛼𝑧 is 20

assumed to be random and distributed normally as 𝑁(𝛼, 𝜎𝛼2). Also, 𝜀𝑧𝑡 and 𝛼𝑧 are assumed to be 21

independent. Further, heteroscedasticity and autocorrelation are allowed in specification of 𝜀𝑧𝑡. 22 The model was estimated using Stata software using the feasible generalized least 23

squared routine (i.e. xtgls). In order to test the autocorrelation and heteroscedasticity 24 specification of the error term, separate hypothesis tests were run. First, to test heteroscedasticity, 25

a likelihood ratio test proposed by Poi and Wiggins (25) was used that compares heteroskedastic 26 treated panel model against the same model assumed to be homoscedastic. Second, to test 27

autocorrelation, the procedure outlined in Wooldridge (26) was adopted. 28

6. DISCUSSION OF RESULTS 29 As noted earlier, a linear panel model was estimated using the generalized least squares approach 30 in Stata software. The model allowed heteroscedasticity and autocorrelation in the specification 31 of the disturbance term. Due to the presence of autocorrelation and gaps in the data (i.e. missing 32 weekly data between 2014 and 2015), standard procedure of maximum likelihood estimation was 33

not applicable. The final panel contained 61 taxi zones over 49 weeks for a total of 2,989 34 observations. Parks and zones with missing independent variables were dropped from the model, 35 including Ellis Island, Liberty Island, Governors Island, Randall’s Island, Central Park, Battery 36

Park, Inwood Hill Park, and Highbridge Park. Before running the panel model, the tests for 37 heteroscedasticity and autocorrelation identified above were performed. The results indicated 38 that both autocorrelation (with an F-test statistic of 331 with degrees of freedom of 1 and 60 and 39 p-value of 0.00), and heteroscedasticity (with a chi squared statistic of 2242.47 with degrees of 40

freedom of 60 a p-value of 0.00) were present and as such the error term specification needed to 41 accommodate these issues. In the interest of space, the complete listing of autocorrelation and 42 heteroscedasticity parameters is not presented here. 43

A variety of explanatory factors, including demographic, land use, and environmental, 44 were explored. Additionally, temporal variables were explored to capture the impact of 45


unobserved time varying factors. The final model is presented in Table 1. All the variables 1

included are significant at least at the 10 percent level of significance. Below, the influence of 2 the various explanatory variables on the weekly Uber pickup demand are discussed. 3 4

Table 1. GLS Random Effects Panel Model Results 5

Parameter Random Effects Model

Coefficient Standard Error

Constant -19056.180 -6274.937 ***

Time Variables

Summer Dummy Variable -278.133 46.857 ***

Winter Dummy Variable 89.647 50.472 *

Difference in weeks from initial week 93.323 6.446 ***

Environmental Variables

Weekly Precipitation (in.) 25.892 6.262 ***

Weekly Average Max Temperature (oF) -10.143 1.174 ***

Built Environment Variables (Taxi Zone Level)

Count of Citi Bike Stations 242.850 29.211 ***

Total Built Area (10,000 sq. ft.) 0.936 0.071 ***

Percent Residential by Floor Area 64.247 7.003 ***

Percent Retail by Floor Area 151.426 33.043 ***

Demographic Variables (Taxi Zone Level)

Percent African American -58.925 7.517 ***

Percent Male 60.847 17.042 ***

Percent Age Under 19 220.657 23.752 ***

Percent Age Over 65 -0.213 0.047 ***

Percent of Households with No Vehicle 126.400 60.253 **

Percent of Households with One Vehicle 131.103 62.971 **

Average Household Size -2991.058 363.236 ***

Percent of Adults with at least Some College 17.922 9.481 *

Interaction Variables

(Percent Residential by Floor Area) x (Difference in weeks) -0.909 0.089 ***

(Percent Retail by Floor Area) x (Difference in weeks) 9.581 0.653 ***

(* p < 0.10, ** p < 0.05, *** p < 0.01) 6 7 6.1.Influence of Temporal Variables 8 A variety of temporal variables were explored to capture the influence of unobserved time 9

varying and seasonal factors. As the scope of the data was limited to only two six month periods, 10 the extent of seasonal coverage was also limited to only include winter, spring, and summer. In 11 the model, seasonal effects are reported with reference to spring. Spring was chosen because it 12 was the only season present in both years of the data used in the analysis. It appears that Uber 13 demand at the taxi zone level decreases during the summer as noted by the negative coefficient. 14


This observation is also confirmed by the descriptive analysis wherein a clear drop was observed 1

during the summer months of 2014. This may be attributed to individuals’ preference to use non-2 motorized modes, such as walking or biking, when the weather is warmer compared to other 3 seasons. While reviewing research on seasonal influences on motorized and non-motorized 4

travel, Guo et al. (27) found that during the summer, people make more discretionary trips via 5 non-motorized modes. On the other hand, an opposite reasoning can be offered to explain the 6 positive effect of the winter dummy variable on Uber demand: given the extreme weather that 7 New York City experiences during the winter months, people may seek the comfort of a climate 8 controlled vehicle in order to escape cold conditions. 9

Additionally, the difference in week from initial week variable is defined as the number 10 of weeks past the first complete week (i.e. week ending on April 6

th, 2014) of demand data. 11

Here, a positive influence is apparent and can be interpreted to mean that as time progresses, the 12 pickup demand within taxi zones is increasing. This finding is in line with both the observed 13

trends at the descriptive level, as well as research and reports presented earlier on the growth of 14 Uber overall. 15

One of the primary hypotheses was to explore potential tapering off of Uber demand with 16 time. To this end, the impact of higher order versions of the difference in week variable were 17

explored after controlling for the other explanatory variables. For example, a negative coefficient 18 associated with a second order influence will lend evidence in support of the hypothesis. It was 19 found that higher order versions of the difference in week variable were not significant. 20

However, this contradicts the observations from the trend analysis by taxi zones that showed that 21 there is evidence of a tapering effect for residential zones. To further explore the tapering off 22

effect, interaction variables between difference in week and other explanatory factors, namely 23 demographic and land use variables were explored. The results for the interaction variables are 24 discussed at the end of this section. 25

26

6.2.Influence of Environmental Variables 27 The environmental variables are included to extend interpretation beyond just seasonality and 28 quantify how weekly weather trends influence zonal level Uber demand. The positive influence 29

related to total weekly precipitation was expected and is intuitive. As the overall total of weekly 30 precipitation increases, demand for Uber also increases. This is plausible because rain/sleet/snow 31

move people away from other non-automobile modes and towards trip making in the comfort of 32 an on-demand vehicle. A substantial amount of users in Manhattan use non-automobile modes 33

like Citi Bike, walk, and transit among others (17, 19). These riders may potentially opt for an 34 Uber trip if it means avoiding exposure to the elements. As the average high temperature by 35 week increases the demand for Uber also increases. Increased temperatures, especially those 36 outside of the normal weather condition, may be prompting individuals to trend towards non-37 motorized travel in a similar manner to the summer influence. 38

39 6.3.Influence of Built Environment Variables 40

Unfortunately, in the final model formulation, the influence of number of bus stops and subway 41 stops at the taxi zone level was found to be statistically insignificant. However, the large positive 42 influence of Citi Bike stations at the taxi zone level is an interesting result. At a high level, the 43 positive influence makes sense as Citi Bike stations are primarily located in lower Manhattan and 44 these taxi zones correspond to the areas with the highest trip attraction. However, upon closer 45 inspection, it can be seen that even with the inclusion of other built factors like total floor area 46


which are more direct determinants of trip attraction, Citi Bike stations have a significant and 1

sizeable impact. One plausible reason for this is that the users of the on-demand shared Citi Bike 2 system may also be inclined to use shared on-demand services provided by Uber. It is also 3 plausible that Uber is used by such individuals to complete trips that Citi Bike could not. This is 4

consistent with research by Faghih-Imani et al. (17) – in their analysis of Citi Bike and Taxi 5 modes, the authors found that for trips with substantial distance taxi modes proved to be more 6 suitable option – a similar scenario could be at work here. 7

The total built area is assumed to capture the overall intensity of each taxi zone, and the 8 positive relation to Uber demand follows from our expected results. With more built area, there 9

is more to see, do, and places to live and this falls within other findings that Uber demand is 10 highest in the core of Manhattan and in tourism driven taxi zones. The percent of residential 11 floor area has a positive influence on Uber pickup demand. Residents, just like visitors and 12 tourist, make trips and there are substantial amounts of demand in these residential zones – 13

referring to Figure 1, it can be seen that residential zones like East and West village have some 14 of the highest demand of all observed taxi zones. 15

Variables representing other land use, namely, percentages of overall commercial, retail, 16 garage (parking), and office areas were explored. These remaining land use percentages were 17

also included in the model but the only other type that proved significant was the percentage of 18 retail by floor area. This strong positive interaction shines light on previously noted trends 19 regarding Uber’s propensity to serve tourism and central business districts. Within Manhattan, 20

these locations coincide with higher proportions of retail area and as such higher Uber demand. 21 22

6.4.Influence of Demographic Variables 23 Understanding how demographic makeup of taxi zones influence Uber demand can provide 24 valuable insights into how different groups are adopting dynamic ridesharing services. When 25

exploring the link between Uber demand and percentage of the population that is African 26

American in a taxi zone, a negative relationship was observed. This observation raises questions 27 on equity. A better understanding of the factors influencing (or lack thereof) African American 28 users to Uber is needed. 29

The second demographic factor explored is that of gender on Uber demand. Here the 30 percentage of males within the population was found to have a positive relationship with Uber 31

pickup demand. A possible explanation for this can be safety. As mentioned previously (8), Uber 32 has been dealing with issues regarding safety. As a result, male riders may feel more comfortable 33

using the dynamic ridesharing system than their female counterparts. 34 Age has been an important topic in regard to Uber adoption, and previous research 35 indicates that Uber users tend to skew younger (13). When holding the typical adult age range as 36 the constant categorical factor and exploring the effects on individuals under 19 and over 65, our 37 findings fall in line with existing research. The percentage of the population under 19 years of 38

age is positively associated with demand for Uber than their older counter parts. On the other 39 hand, senior individuals exhibit a lower demand for Uber. Propensity for technology and existing 40

mode bias are potential explanations for this relationship. 41 Vehicle ownership was investigated over conventional income variables due to findings 42 in the Uber literature. In San Francisco, survey respondents that utilized Uber identified as 43 owning few vehicles (13). As such, we investigated the influence of household vehicle 44 ownership on Uber demand. It was observed that both no vehicle and one vehicle households 45 exhibit heightened Uber demand compared to households with two or more vehicles. Overall 46


vehicle deficiency can explain this, as households still need to make trips to fulfill their activity 1

needs. 2 The modeled effects of household size in regard to demand indicate that larger 3 households have a substantially lower demand for Uber than smaller ones. As household size 4

increases, traveling in general typically becomes more difficult and potentially also expensive 5 especially for a mode like Uber that has limited capacity per vehicle. 6

The final significant demographic variable explored was aimed at quantifying the 7

relationship of education on Uber demand. While only marginally significant, the findings 8

indicate that individuals with higher education exhibit higher demand for Uber. Once again in 9

line with preliminary survey findings, this can be can be attributed to a higher propensity for 10

adopting and using technology by those associated with higher levels of education. 11

12

6.5.Influence of Interaction Variables 13 As noted earlier, the main hypothesis was to explore if there is a potential tailing off effect in the 14

growth of Uber demand with time. As noted before, higher order versions of the weekly 15 difference variable were found to be insignificant despite the evidence from the descriptive 16

analysis. In an effort to further explore the tailing off effect, interaction variables between the 17 weekly difference variable and different demographic, and land use variables were investigated 18 to see if there are any potential negative shifts in the demand for specific tax zones and/or 19

population groups over time. This exploration provided substantial evidence in support of our 20 proposed hypothesis. We found that the interaction between percent retail and weekly difference 21

had a substantial positive impact on Uber demand, suggesting that Uber demand in high retail 22 areas is growing with time. As Uber becomes more readily adopted, a greater proportion of users 23 are making trips from and to areas with higher retail using the service. However, when looking at 24

the interaction of the percentage of residential area by weekly difference there is a clear negative 25

impact. While the magnitude of the relationship is small, the negative coefficient indicates that 26 residential areas over time have a waning or decline in overall Uber demand. There are a couple 27 of possible reasons for this observation. First, some residents may have preference towards 28

existing modes (and bias against Uber) and as such are not going to adopt new systems. Second, 29 if Uber only replace SOV trips, and/or borrows from other non-automobile modes, it will not 30

improve congestion on roadways. As a result, those who are conscious to travel time 31

performance and are price conscious may prefer the cheaper alternative with similar travel time 32 performance (e.g. subway, bus). 33

34 7. CONCLUSION AND FUTURE RESEARCH 35 With the substantial growth in dynamic ride sharing offerings over the last few years, a 36

comprehensive investigation into understanding the underlying factors impacting demand is of 37

interest. Using multiyear data from an existing ride sharing service, namely, Uber from New 38

York City, the temporal trends in demand for this service are explored. Existing research 39 supports the growing adoption of Uber, but here it is hypothesized that while macroscopic trends 40 may suggest unbounded increases in the demand, meso/microscopic levels may offer a different 41 picture. It is posited that the demand for ride sharing services has a limit. To explore these 42 underlying factors and test our hypothesis, a random effects panel model was developed. 43

The findings of our analysis indicate that there is a significant growth over time, but that 44 seasonality in demand exists with an increase in demand in winter months and decrease in 45 demand in summer months. Weather was also noted to have a statistically significant impact, 46


with increased precipitation positively influencing demand and warmer temperatures decreasing 1

the weekly demand. At a built environment level, the presence of transit was not determined to 2 have significant impact, but bike share infrastructure had a positive influence. Similarly, the 3 more built area a taxi zone had, as well as the higher percentage of the floor area that is 4

residential and retail based land use, indicated a higher demand for Uber service. The influence 5 of zonal level demographics was also investigated. The percentage of population that is male, 6 under 19 years of age, has some college education, and has one or no household vehicles showed 7 positive impacts on weekly Uber demand. The percentage of the population that is African 8 American, over 65, and larger households all had negative impacts on Uber demand. Within the 9

model, the weekly difference had a positive effect on Uber demand, it was included at higher 10 order to capture the observed tailing over time, however the results were insignificant. 11 Interaction variables were then included to test our hypothesis for observed declines in demand 12 over time with respect to various built environment variables. A positive influence of retail area 13

across time was determined. Perhaps the most substantial finding, a negative relationship was 14 observed when exploring the influence of percent residential area across time. This negative 15

relationship lines up with the decreasing trends observed in the descriptive analysis. Together, 16 these results offer substantial evidence in support of a trailing in Uber demand over time in 17

residential dominant areas. 18 These last findings have substantial impact beyond the scope of just Uber demand 19 analysis and have potentially important policy implications for the overall adoption of dynamic 20

shared modes. The significant tapering off within residential zones indicates that even as the 21 demand grows at a citywide scale, residents’ inertial modal biases and familiarity with existing 22

modes may limit the dynamic shared ride adoption by all individuals within individual subareas 23 (28). A second policy implication that can be drawn here is that users are typically motivated to 24 change modes if the alternative provides improvement over existing offerings (e.g. lower costs, 25

or shorter travel times). In the current operational framework, Uber and other dynamic shared 26

ride providers most often replace single occupancy (SOV) auto-modes further contributing to the 27 effects of congestion and hindering overall adoption (13, 16). Further investigation into these 28 effects is warranted. 29

This research presents a substantial first step in understanding dynamic ridesharing in a 30 real world context, but it is important to recognize some of the limitations of this work. First, 31

holding Uber pickups as a stand in for the demand can be problematic as it inherently omits any 32 demand by users who attempted to request rides but were not able to have those requests filled. It 33

also inherently omits any user without a smartphone or knowledge of the Uber interface, which 34 may be disproportionately represented by elderly and low income groups. Second, with data only 35 on the pickup event of each trip, a rich understanding of how these trips are integrating into the 36 existing network is not possible. Third, the spatially aggregated nature of the data limited this 37 analysis to the neighborhood level offered by taxi zones. The data quality and structure issues 38

noted with the Uber data motivates an important discussion. As open data becomes more readily 39 available, it is vital for open data providers to adopt a standard of practice for quality and 40

structure that can in turn promote additional research endeavors. 41 Additionally, the panel model used in this research assumes strong exogenity which may 42

not be supported by the data and alternate model forms must be explored. A comprehensive 43 spatial analysis will help to further explore and characterize the spatial relationships to Uber 44 demand. Also, to supplement and confirm the findings presented here, we anticipate further 45 exploration through a primary survey which will collect disaggregate level information on users 46


and their preferences. Lastly, a comprehensive study of the temporal trends in all modes 1

including taxis, subways, buses, and other dynamic rideshare providers from the region can help 2 understand the tradeoffs between on-demand modes, like Uber, and existing shared ride services, 3 and will be explored in future work. 4


REFERENCES 1 1. Furuhata, M., Dessouky, M., Ordóñez, F., Brunet, M., Wang, X., & Koenig, S. (2013). 2

Ridesharing: The state-of-the-art and future directions. Transportation Research Part B: 3 Methodological, 57, 28-46. doi:10.1016/j.trb.2013.08.012 4

2. Harding, S., Kandlikar, M., Gulati, S. Taxi apps, regulation, and the market for taxi journeys, 5 Transportation Research Part A: Policy and Practice, Volume 88, 2016, Pages 15-25, 6 ISSN 0965-8564, http://dx.doi.org/10.1016/j.tra.2016.03.009. 7

3. Lazo, L. (2015, June 4). Uber turns 5, reaches 1 million drivers and 300 cities worldwide. 8 Now what? Retrieved July, 2017, from https://www.washingtonpost.com/news/dr-9

gridlock/wp/2015/06/04/uber-turns-5-reaches-1-million-drivers-and-300-cities-10 worldwide-now-what/?utm_term=.b30a6f5dc40d 11

4. Tepper, F. (2016, July 18). Uber has completed 2 billion rides. Retrieved July, 2017, from 12 https://techcrunch.com/2016/07/18/uber-has-completed-2-billion-rides/ 13

5. Schneider, T. (2015, November 17). Analyzing 1.1 Billion NYC Taxi and Uber Trips, with a 14 Vengeance. Retrieved July, 2017, from http://toddwschneider.com/posts/analyzing-1-1-15

billion-nyc-taxi-and-uber-trips-with-a-vengeance/ 16 6. Sun, C., & Edara, P. (2015). Is Getting an Uber-Lyft from a Sidecar Different from Hailing a 17

Taxi? Transportation Research Record: Journal of the Transportation Research Board, 18 2536, 60-66. doi:10.3141/2536-08 19

7. Chen, L., Mislove, A., & Wilson, C. (2015). Peeking Beneath the Hood of Uber. Proceedings 20

of the 2015 ACM Conference on Internet Measurement Conference - IMC 15, 495-508. 21 doi:10.1145/2815675.2815681 22

8. Picchi, A. (2014, December 11). The rising safety issues that could throttle Uber. Retrieved 23 July, 2017, from http://www.cbsnews.com/news/the-rising-safety-issues-that-could-24 throttle-uber/ 25

9. Fagnant, D. J., & Kockelman, K. (2015). Preparing a nation for autonomous vehicles: 26

opportunities, barriers and policy recommendations. Transportation Research Part A: 27 Policy and Practice, 77, 167-181. doi:10.1016/j.tra.2015.04.003 28

10. Chen, M. K., & Sheldon, M. (2015). Dynamic pricing in a labor market: Surge pricing and 29

the supply of Uber driver-partners. University of California (Los Angeles) Working Paper 30 URL http://citeseerx. ist. psu. edu/viewdoc/download. 31

11. Guo, S., Liu, Y., Xu, K., & Chiu, D. M. (2017, March). Understanding ride-on-demand 32 service: Demand and dynamic pricing. In Pervasive Computing and Communications 33

Workshops (PerCom Workshops), 2017 IEEE International Conference on (pp. 509-514). 34 IEEE. 35

12. Correa, D., Xie, K., & Ozbay, K. (2017). Exploring the Taxi and Uber Demand in New York 36 City: An Empirical Analysis and Spatial Modeling. Transportation Research Board 96th 37 Annual Meeting (No. 17-05660). 38

13. Rayle, L., Dai, D., Chan, N., Cervero, R., & Shaheen, S. (2016). Just a better taxi? A survey-39 based comparison of taxis, transit, and ridesourcing services in San Francisco. Transport 40

Policy, 45, 168-178. doi:10.1016/j.tranpol.2015.10.004 41 14. Cramer, J., & Krueger, A. (2016). Disruptive Change in the Taxi Business: The Case of 42

Uber. American Economic Review, 106(5), 177-182. doi:10.3386/w22083 43 15. Poulsen, L. K., Dekkers, D., Wagenaar, N., Snijders, W., Lewinsky, B., Mukkamala, R. R., 44

& Vatrapu, R. (2016, June). Green Cabs vs. Uber in New York City. In Big Data 45 (BigData Congress), 2016 IEEE International Congress on (pp. 222-229). IEEE. 46


16. Fischer-Baum, R., & Bialik, C. (2013, October 13). Uber Is Taking Millions Of Manhattan 1

Rides Away From Taxis. Retrieved July, 2017, from 2 https://fivethirtyeight.com/features/uber-is-taking-millions-of-manhattan-rides-away-3 from-taxis/ 4

17. Faghih-Imani, A., Anowar, S., Miller, E. J., & Eluru, N. (2017b). Hail a cab or ride a bike? A 5 travel time comparison of taxi and bicycle-sharing systems in New York City. 6 Transportation Research Part A: Policy and Practice, 101, 11-21. 7 doi:10.1016/j.tra.2017.05.006 8

18. Brakewood, C., Macfarlane, G. S., & Watkins, K. (2015). The impact of real-time 9

information on bus ridership in New York City. Transportation Research Part C: 10 Emerging Technologies, 53, 59-75. doi:10.1016/j.trc.2015.01.021 11

19. Campbell, K. B., & Brakewood, C. (2017). Sharing riders: How bikesharing impacts bus 12 ridership in New York City. Transportation Research Part A: Policy and Practice, 100, 13

264-282. doi:10.1016/j.tra.2017.04.017 14 20. FiveThirtyEight, (2015) Uber Trip Data. [Data Files]. Retrieved July, 2017, from 15

https://github.com/fivethirtyeight/uber-tlc-foil-response/tree/master/uber-trip-data 16 21. NYC Planning (2016). PLUTO. [Data File]. Retrieved July, 2017, from 17

https://www1.nyc.gov/site/planning/data-maps/open-data/dwn-pluto-mappluto.page 18 22. NYC Open Data (2016) [Data File]. Retrieved July, 2017, From 19

https://opendata.cityofnewyork.us/ 20

23. National Centers for Environmental Information. NOAA. [Data File]. Retrieved July, 2017, 21 from https://www.ncdc.noaa.gov/data-access 22

24. Cameron, A. C., & Trivedi, P. K. (2005). Microeconometrics: methods and applications. 23 Cambridge university press. 24

25. Poi, B., & Wiggins, V. (2003). Testing for panel-level heteroskedasticity and autocorrelation. 25

StataCorp. 26

26. Wooldridge, J. M. (2010). Econometric analysis of cross section and panel data. MIT press. 27 27. Guo, J., Bhat, C., & Copperman, R. (2007). Effect of the built environment on motorized and 28

nonmotorized trip making: substitutive, complementary, or synergistic?. Transportation 29

Research Record: Journal of the Transportation Research Board, (2010), 1-11. 30 28. Lee, J. H., & Goulias, K. G. (2017). A Decade of Dynamics of Residential Location, Car 31

Ownership, Activity, Travel and Land Use in the Seattle Metropolitan 32 Region. Transportation Research Procedia, 23, 420-439. 33

is there a limit to1 adoption of dynamic …from june 2014 to june 2015, uber’s manhattan monthly...

Documents