The Geography of Ridesharing:

A Case Study on New York City ∗

Chungsang Tom Lam† Meng Liu‡ Xiang Hui§

Clemson University Washington University Washington University

Abstract

Despite the popularity of ridesharing, there is limited empirical evidence on how rideshar-
ing activities differ across regions with different levels of accessibility and the implication for
consumers. In this paper, we study the market for rides across New York City neighborhoods.
We construct a novel data set that contains massive API queries on route-specific estimates of
pricing, wait time, and travel time of Uber, Lyft, and the public transit. After linking this data
with actual trip records of taxis, Uber, and Lyft, we document a strong pattern that ridesharing
has a larger market share relative to taxis in neighborhoods with lower accessibility, defined
either in terms of geographic distance to Midtown Manhattan or “economic distance” to job
opportunities. Next, we estimate a discrete-choice model of demand for rides and interpret the
geography of ridesharing through the lens of the model. We find that consumer surplus from
ridesharing varies drastically across geography: passengers that are 5 to 15 miles (resp. more
than 15 miles) from Midtown experience a 60% (resp. 19%) larger consumer surplus relative
to passengers that are within 5 miles from Midtown. Additionally, over half of these gains
comes from reduced wait time. We discuss the implications of the distributional results for
policy makers.

Keywords: Ridesharing; Geography; Transportation; Consumer Surplus

∗
The authors would like to thank Erik Brynjolfsson, Babur De los Santos, Dean Eckles, Chiara Farronato, Andrey
Fradkin, Avi Goldfarb, Daniel Miller, Catherine Tucker, Joel Waldfogel, Patrick Warren for their invaluable comments
and suggestions. We also thank conference and seminar participants at Stanford Workshop at Marketplace Innovation,
Marketing Science Conference, Toulouse Conference on the Digital Economy, AEA, MIT, WashU, Clemson, and
Erasmus University. This study is not sponsored by Uber, Lyft, or any other organizations.
†
Assistant Professor of Economics, Clemson University. Email: [email protected]
‡
Assistant Professor of Marketing, Washington University in St. Louis. Email: [email protected]
§
Assistant Professor of Marketing, Washington University in St. Louis. Email: [email protected]

Electronic copy available at: https://ssrn.com/abstract=2997190

1 Introduction

Ridesharing platforms have gained popularity around the world in recent years and, alongside,

scrutiny from policy makers. Policies and regulations about ridesharing across cities seem to be

predominantly a binary decision: either ridesharing is allowed or it is banned. However, when

policy makers ignore the potential distributional impacts of ridesharing across geographies, a one-

size-fits-all policy will benefit some neighborhoods while being detrimental to others. Studying

the geography of ridesharing and its implication for consumers helps us understand where and

when ridesharing could benefit consumers, therefore creating opportunities for designing nuanced

policies tailored to different geographies.

In this paper, we study how ridesharing affects consumer welfare across neighborhoods in

a metropolitan setting. To achieve this goal, we aim to answer two research questions. First,

how do ridesharing activities differ across neighborhoods with different levels of accessibility?

Second, how might this geography translate into consumer welfare in different neighborhoods?

To answer these questions, we construct a comprehensive data set that describes the market of

rides in New York City (NYC hereafter). Specifically, we gather a massive set of API queries on

route-specific estimates of pricing, wait time, and travel time of Uber, Lyft, as well as the NYC

public transit. We link this data with actual trip records of taxis, Uber, and Lyft published by NYC

Taxi and Limousine Commission (TLC hereafter). We further enrich the data by conducting a field

collection of around 70,000 historical trip records from a sample of Uber and Lyft drivers. The

resulting data set provides a comprehensive picture of the market that allows for analysis at the

route level in real time.

We first summarize and visualize geographic patterns in ridesharing and taxi rides. We find

Electronic copy available at: https://ssrn.com/abstract=2997190

that for both ridesharing and taxis, the total number of pickups is smaller in less accessible regions,

where accessibility is defined either in terms of geographic distance to the city center or “economic

distance”. The geographic distance is calculated as the distance between the pickup location and

Manhattan Midtown, and economic distance is measured by the neighborhood’s total number of

jobs accessible within one hour’s commute by public transit (Kaufman et al. (2014)). However,

the rate at which the number of pickups decreases with respect to distance, which we will refer to

as the trip elasticity of distance henceforth, is smaller than that of taxis. Consequently, the market

share of ridesharing relative to that of taxi is higher in low accessibility neighborhoods under both

definitions of distance. In addition, the geography of ridesharing is also robust to using pickups

by green boro taxis as a benchmark, which follow the same pricing rule as yellow medallion

taxis but are allowed to pick up passengers only outside Manhattan Core1 , i.e., predominantly low

accessibility neighborhoods. We find that the market share of ridesharing remains larger than that

of green taxis in low accessibility neighborhoods.2

The geographical distribution of market shares is an equilibrium outcome, driven by both con-

sumer preference toward various transportation modes and drivers’ decisions on when and where

to provide the service. We provide empirical evidence that is consistent with a high degree of

matching friction between taxi drivers and passengers, which likely gives rise to the low presence

of taxi pickups in low accessibility neighborhoods. First, we show that low accessibility neighbor-

hoods more frequently exhibit excess demand (the incidence of more people needing a ride than

the number of available drivers in the area) instead of insufficient demand for taxis. Second, the

trip elasticity of distance is larger for yellow taxis than for green taxis in low accessibility regions,
1
Manhattan Core is defined as the part of Manhattan below the north edge of Central Park.
2
This finding is consistent with Mitchell L. Ross’ argument: “(green taxis) basically cluster at transit and retail
hubs because they are more likely to find passengers there than if they cruise the streets they are authorized to cruise.”
https://www.nytimes.com/2018/09/03/nyregion/green-cabs-yellow-uber.html

Electronic copy available at: https://ssrn.com/abstract=2997190

even though both taxi services face the same demand due to their identical pricing rules outside

of the Manhattan Core. This difference suggests that the matching friction in low accessibility

neighborhoods relative to the Manhattan Core is even higher for yellow taxis than for green taxis.

The matching friction present among taxis may be alleviated by ridesharing through technology-

enabled matching, incentives created by dynamic pricing, or both. Ridesharing platforms seem to

reduce this information gap in low accessibility neighborhoods mainly through the use of advanced

matching algorithms: According to our data, surge pricing is rare in low accessibility neighbor-

hoods, although it is more frequently activated in Manhattan.

The second part of the paper aims at translating the observed geography to consumer welfare,

leading us one step closer to answering relevant policy questions. We model consumer’s decision

as a discrete choice problem between various competing transportation modes for a given route

that is defined by a particular origin-destination-time combination. We use the logit framework to

map the observed rides into consumer preference toward price, wait time, as well as other observed

and unobserved characteristics.

In the demand estimation, endogeneity arises when the price of a region is correlated with

the unobserved demand condition in the same region, conditional on service type, route, and time

fixed effects. To obtain a consistent estimate of price elasticity, we use an instrumental variable

(IV) that affects the market share of a transportation mode on a given route only through its impact

on price, but it is otherwise uncorrelated with local demand shocks. We construct such an IV using

a design feature of the ridesharing Apps that were present during our sample period: passengers

see and commit to a surge multiplier on Uber and Lyft Apps prior to entering their destination

on their phones.3 Exploiting this feature, we use the surge prices of trips into the focal zone to
3
After our sample period, this old version of Uber was updated to upfront pricing, and now riders need to input

Electronic copy available at: https://ssrn.com/abstract=2997190

instrument for price in this zone. On the one hand, because these ridesharing Apps do not possess

information on where the consumer is going, they cannot adjust their surge multiplier according to

the consumer’s destination, which creates the uncorrelatedness of surge prices at the origin and the

underlying demand condition at the destination. On the other hand, surge pricing at the origin will

affect the price in the focal zone through its effect on the number of available drivers in the focal

zone after their drop-offs.

Using this instrument, we identify heterogeneous consumer preference toward price and wait

time. The parameter estimates show that consumers prefer lower price and less wait time. Also, the

elasticity estimates across regions and different times of the day are consistent with the intuition

that people are less price-sensitive and care more about short wait time during rush hours and in

destinations with more offices. The fact that consumers prefer less wait time in certain locations

during certain times of the day suggests that ridesharing platforms may increase consumer surplus

if the service could effectively reduce wait time in these regions.

Through the lens of the estimated demand model, we translate the geography of ridesharing

to the geography of consumer surplus. Conceptually, we adopt the idea of compensating variation

and estimate how much consumers should be compensated if Uber and Lyft were to be removed

from the market in order to maintain the same level of utility for consumers in different regions.

We find that consumer surplus from ridesharing varies drastically across geography: passengers

that are 5 to 15 miles (resp. more than 15 miles) from Midtown experience a 60% (resp. 19%)

larger consumer surplus relative to passengers that are within 5 miles from Midtown. Additionally,

over half of these gains comes from reduced wait time for getting a ride.

The distributional results on the relative market share and the uneven consumer gains from
their destination to get a fixed price. Lyft also went through a similar design change.

Electronic copy available at: https://ssrn.com/abstract=2997190

ridesharing have important implications for public policy makers. Specifically, the results sug-

gest that the impact of ridesharing can be highly uneven across regions with different levels of

accessibility. It is important to note that metropolitan areas in the United States exhibit significant

geographical disparity in transit access (Tomer et al., 2011; Owen and Murphy, 2018), even for

New York City, which has the nation’s best public transit system. Its neighborhoods at the 90th

percentile in accessibility have access to 18 times more jobs via public transit within an hour than

the neighborhoods at the 10th percentile in accessibility (using data provided by Kaufman et al.

(2014)). This geographical disparity is unlikely to change in the near future, because it depends

on long-run factors such as city planning and re-designing of the public transit network. Taxis are

another important component of metropolitan transit systems, yet we show that they are highly

concentrated in the dense areas, even for taxis that are specifically designed to pick up passengers

in less accessible regions, due to the lack of technology that provides real-time information about

demand.

Based on our findings, urban and transportation planners should recognize the important role

of ridesharing in connecting neighborhoods with low accessibility and disproportionately create

incentives for ridesharing in these neighborhoods, holding the social costs constant. A more inclu-

sive transit network is important because mobility and proximity to markets could affect individu-

als’ labor market outcome and overall well-being, especially those of disadvantaged demographic

groups and individuals without car ownership (e.g., Ihlanfeldt and Sjoquist (1990), Ihlanfeldt and

Sjoquist (1991), Holzer et al. (1994), Hering and Poncet (2010), among other papers in the urban

economics literature).

Electronic copy available at: https://ssrn.com/abstract=2997190

1.1 Related Literature and Contribution

Our paper contributes to three literature strands. First, it contributes to the growing literature on

estimating the impact of ridesharing technology on consumers. The closest paper to ours is Cohen

et al. (2016), which uses internal data from Uber and exploits a series of regression discontinuity

designs (RDD) around the surge multiplier cutoffs and estimate a consumer surplus of $1.6 per

dollar spent on Uber. Our paper is different in two ways. First, our goal is to understand the distri-

butional gain in consumer surplus across regions with different levels of accessibility. Our welfare

estimate may be biased due to data limitation, but we expect the qualitative result on the distri-

butional effect of ridesharing to hold. Second, as the authors admit in the paper, the RDD yields

short-run consumer surplus from Uber, whereas we estimate consumer welfare gain in a long-run

steady state. To do so, we allow for consumer substitution among transportation alternatives, and

study consumers’ substitution pattern towards taxi and public transit when ridesharing platforms

are permanently removed from the market.

Researchers have also examined other important welfare topics associated with ridesharing

and the sharing economy in general, such as flexible work arrangements (Chen et al., 2017; Hall

and Krueger, 2016; Hall et al., 2017), routing efficiency and driver moral hazard (Cramer and

Krueger, 2016; Liu et al., 2018; Wang et al., 2019), reduced matching friction (Buchholz, 2015;

Frechette et al., 2016; Zhang et al., 2019), reduced drunk driving (Greenwood and Wattal, 2017),

local consumption patterns (Zhang and Li, 2017), traffic congestion (Li et al., 2016), and the use of

public transit (Hall et al., 2018; Babar and Burtch, 2017). We add to this literature by being among

the first to analyze the role of ridesharing in inclusive mobility.

Second, this paper contributes to the research on geographical impact in transportation in U.S.

Electronic copy available at: https://ssrn.com/abstract=2997190

urban areas. There is a large literature that studies the disparity in transportation in U.S. urban

areas, which finds mobility inequality across U.S. metropolitan areas as well as the correlation

between neighborhood accessibility and measures such as unemployment and economic oppor-

tunities (Tomer et al., 2011; Kaufman et al., 2014; Owen and Murphy, 2018). In addition, these

studies agree that only marginal improvements in public transit systems can be made in light of

deep budget cuts at all levels of government. Building on the measures and insights provided by

these studies, we offer empirical evidence that ridesharing can fit into the economic and social

fabric of metropolitan areas and promote inclusive mobility.

Lastly, researchers have studied the geographical impact of information technology in other

contexts, such as in the area of international trade (e.g., Blum and Goldfarb (2006); Hortaçsu

et al. (2009)), crowdfunding (e.g., (Agrawal et al., 2015)), and in the accommodation market (e.g.,

(Farronato and Fradkin, 2017; Zervas et al., 2017)).

2 Data

Our data set focuses on the five boroughs of NYC, namely the Bronx, Brooklyn, Manhattan, Staten

Island, and Queens from June 1, 2016 to August 31, 2016. The city is divided by TLC into 263

taxi zones that vary in size.4 For example, an average taxi zone in Manhattan is approximately

equivalent to an area of 6 by 6 avenue blocks. Throughout this paper, we treat these taxi zones as

our basic geographical units of analysis, given the nature of the data. Our data set contains data

from four sources:

(1) Real-time Uber and Lyft dynamic pricing and wait time
4
Refer to the following link for a shape file of taxi zones provided by TLC: http://www.nyc.gov/html/
tlc/html/about/trip_record_data.shtml.

Electronic copy available at: https://ssrn.com/abstract=2997190

 The first part of our data is a massive set of API queries on dynamic pricing and wait time

at the geographical centroids of all 263 taxi zones at the minute level, for all service types on

Uber and Lyft in our sample period, namely UberX, UberXL, UberBlack, UberSUV, UberPOOL,

Lyft, LyftLine, and LyftPlus.5 We also queried trip distance, trip duration, and trip cost for a

route between a given pickup zone and a randomly chosen drop-off zone, in any given minute.

Therefore, for each unique route, i.e., a pickup-drop-off pair, we have information on trip distance,

duration, and cost approximately every 4 hours 6 . We use this route-level information mainly in

the demand estimation.

(2) TLC-published Taxi, Uber, and Lyft trip records

NYC TLC publishes taxi, Uber, and Lyft trip records. Taxi trip records (summarized in Ap-

pendix Table 4) contain detailed trip information, such as pickup and drop-off date and time, the

GPS coordinates of the pickup and drop-off locations, number of passengers, trip fares, etc. Uber

and Lyft trips are identified from other For Hire Vehicles (FHV hereafter) trip records using the

dispatching base numbers.7 Far from the level of detail of taxi trip records, Uber and Lyft trip

data only contain pickup date, time, and locations in the form of taxi zones. Restricted by this,

we then map the GPS coordinates of taxi trips into their corresponding taxi zones so that taxi and

ridesharing trips have the same geographical unit.

(3) Field-collected Uber and Lyft trip records

Uber and Lyft trip records provided by TLC lack information on drop-off locations, drop-off
5
UberX and Lyft are the regular and mostly used service types on each platform, respectively. UberXL and LyftPlus
are service types with cars of larger capacity, which usually seat up to six passengers. UberBlack and UberSUV are
the luxury options on Uber. UberPool and LyftLine are the carpool options on Uber and Lyft, respectively.
6
More accurately, if we query on each route in exactly one minute, it takes on average about 263 minutes to query
the same route again.
7
Refer to http://www.nyc.gov/html/tlc/downloads/pdf/find_a_ride.pdf. See the data ap-
pendix Appendix C for how the base numbers are identified.

Electronic copy available at: https://ssrn.com/abstract=2997190

time, and service types. This data limitation hinders our analysis at the service type-route-time

level. To address this problem, we conducted field data collections to acquire 70,277 historical

trip records from 443 Uber and Lyft drivers in NYC. The purpose is to get a representative sample

in order to estimate the empirical distribution of drop-off time and locations. We then use these

probabilities to impute the service-type-route-time trip counts (the detailed procedure in Appendix

D). The representativeness of the field data is supported by a comparison between their pickup

distribution and that of the TLC-published Uber and Lyft trips.

(4) Route-specific subway and bus travel time

We queried Google Maps API to obtain total travel time via public transit (subway and bus).

This data is collected at the level of 263 × 263 routes, by night/day and by weekend, where night is

defined as 0:00am to 6:59am. Specifically, the total travel time for a given route via public transit

includes (1) walking time from the zone centroid to the subway station or bus stop that Google

Maps assigns, (2) subway or bus wait time, (3) subway or bus travel time, and (4) walking time

out of the subway station or from the bus stop to the centroid of the destination zone. Note that

the total public transit time is the best route suggested by Google Maps, which may include solely

subway, solely bus, or a combination of both, depending on specific routes and times.

3 A Description of The Geography of Ridesharing

In this section, we summarize and visualize model-free evidence using our data set. The average

share of trips that experience a price surge is 7% – 8% for Uber non-luxury services and is around

19% for Lyft services. In terms of wait time, both Uber and Lyft’s flagship services, namely UberX

and Lyft, are around 7 minutes. To focus on our research question, we visualize how price and wait

Electronic copy available at: https://ssrn.com/abstract=2997190

 Figure 1: Surge Multiple and Wait Time Change Rapidly across Space and Time
(UberX, Monday, June 6, 2016)
Wait Time (Minutes) Wait Time (Minutes) Wait Time (Minutes) Wait Time (Minutes)

Manhattan
20
15
10
5
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Hour
Queens
30

20

10

0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Hour
Brookyn
20
15
10
5
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Hour
Bronx
30
20
10
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Hour
Surge Multiple
1.0 1.5 2.0 2.5 3.0

Note: This graph plots how UberX minute-level surge multiple and wait time, averaged over all zones of a given NYC
borough, vary across time of the day, on Monday, June 6th 2016. Surge multiples are in varying shades as illustrated
by the legend, and wait time are measured by the bar length.

time vary across geography, and leave the reporting of other summary statistics in Table 3 in the

Appendix.

In Figure 1, we plot the average surge multipliers and wait time across different regions at

different time of the day. Note that conceptually, the same region at different time of the day can

be regarded as a different region, because they could be subject to different underlying market

condition [Tom: conditions]. Two features of the data emerge: First, prices and wait time change

rapidly across geography and time of the day. Second, surge multipliers and wait times are usually

greater in Manhattan than in outer boroughs, and are greater in rush hours than in non-rush hours.

These variations suggest that the prevalence of ridesharing and its implication for consumers could

10

Electronic copy available at: https://ssrn.com/abstract=2997190

 Figure 2: The Geography of Ridesharing and Taxi Rides

vary substantially across space and time.

Next, we plot the number of taxi trips and the share of ridesharing trips out of all trips (taxi plus

ridesharing) in Figure 2. The graph shows that the number of taxi pickups are concentrated in high

accessibility region, i.e., the Manhattan Core, and it decreases as the distance to the Manhattan

core increases. Meanwhile, the share of ridesharing pickups relative to taxi is larger in regions

with low accessibility, namely the outer boroughs.

To see this pattern more clearly, in Figure 3a, we plot the relative share of ridesharing to yellow

taxis against different neighborhoods that differ in the distance to Midtown Manhattan measured

in miles. Consistent with Figure 2, we see that the market share of ridesharing is greater in neigh-

borhoods that are farther away from Midtown Manhattan, suggesting that ridesharing supplements

11

Electronic copy available at: https://ssrn.com/abstract=2997190

taxis in neighborhoods that are underserved by taxis. This substitution seems to happen not only

across space, but also across time: in Figure 11 in the appendix, we find evidence that passengers

substitute toward ridesharing more during the hours when taxis become less available due to shift

change.

In Figure 3b, we performed the same analyses by comparing ridesharing with pickups by green

taxis, which was introduced by TLC in August 2013 to increase taxi coverage in neighborhoods

outside Manhattan. Note that green boro taxis use the pricing structure of yellow medallion taxis,

but they have limited pickup areas.8 We see two interesting patterns. First, although green taxis

are specifically designed to serve outer boroughs, ridesharing still has a larger market share than

green taxis in neighborhoods that are farther away from Midtown Manhattan. Next, comparing

across Figure 3a and Figure 3b, we see that the second graph has a less steep slope, which means

that green taxis have a lower trip elasticity of distance compared to yellow taxis. This difference

reflects different drivers’ cruising strategy, because the two types of taxis essentially face the same

demand due to their same pricing rules outside of the Manhattan Core. The different cruising

strategy is consistent with drivers responding optimally to different matching frictions: Yellow

taxi drivers prefer to cruise towards the Manhattan Core likely because their matching friction of

finding passengers is higher in outer boroughs, relative to that of green taxi drivers, possibly due

to differences in the knowledge of demand in these regions.

Subsequently, we show that the geography of ridesharing activities is robust to the use of “eco-

nomic distance”. Specifically, we follow the literature to measure the accessibility of a neighbor-

hood by the amount of employment and other economic activities it is connected to. We adopt the
8
Specifically, they can only accept street hails anywhere in NYC except south of West 110th Street and East 96th
Street in Manhattan, and they can take pre-arranged trips from JFK and LGA airports. Additionally, green taxis may
drop passengers off anywhere in NYC. By the end of 2015, 7,676 green boro taxi licenses were active in the market.

12

Electronic copy available at: https://ssrn.com/abstract=2997190

measure proposed by Kaufman et al. (2014) that counts the number of jobs accessible, within an

hour of travel by public transit from a given neighborhood of NYC’s 177 neighborhoods. In con-

structing the measure, the authors drew on Census data and Google Maps API. To fit the measure

into our analyses that use taxi zones as the unit of analysis, we built a crosswalk between NYC

neighborhoods and NYC taxi zones using a geographical software which returns the portion of a

given neighborhood that overlaps with a taxi zone in terms of geographical area. We then computed

the weighted sum of jobs accessible across all neighborhoods that (partially) overlap with a given

taxi zone, using the above-mentioned portions as weights. In addition, we chose to use per-capita

jobs accessible as our preferred measure of accessibility, which is computed by dividing the total

number of jobs accessible by the population of a given zone. The distribution of this accessibility

measure exhibits a clear long tail — the neighborhood at the 90th percentile has access to 18 times

more jobs than the neighborhood at the 10th percentile. In addition, this accessibility measure neg-

atively correlates with the geographical distance measure (with a correlation coefficient at -0.7454,

statistically significant at 1% level).

Using this measure of accessibility, we visualize the correlation between the relative share

of ridesharing with respect to taxis and the level of accessibility of a region. Figure 3c shows a

clear negative correlation between accessibility and the prevalence of ridesharing relative to yellow

taxis. In Figure 3d, we observe a similar pattern for the relative share of ridesharing to green taxis.

Similar to the evidence using geographical distance, we see that ridesharing has a larger market

share than taxis in less accessible neighborhoods. Additionally, comparing across Figure 3c and

Figure 3d, we see again that green taxis are less elastic with respect to economic distance when it

comes to picking up passenger in low accessible region. This suggests that the difference in the

trip elasticity with respect to economic distance is due to differences in drivers’ cruising strategy,

13

Electronic copy available at: https://ssrn.com/abstract=2997190

 Figure 3: Ridesharing and Taxi Rides by Regions’ Accessibility

(a) Yellow Taxis (b) Green Taxis

(c) Yellow Taxis (d) Green Taxis

which is likely driven by different levels of matching friction.

One alternative hypothesis is that the smaller share of taxis in less accessible neighborhoods is

driven by the lack of demand for rides. However, this is not consistent with the evidence that there

is more unfulfilled demand for taxis in less accessible neighborhoods. We use realized pickups that

we observe in the data as a proxy for demand of rides. Specifically, we use July 2016 data and count

the total number of ridesharing and taxi pickups in a given 15-minute period in a given zone. We

use the number of taxi drop-offs in the previous 15-minute period as a proxy for the taxi supply

in the focal period (call this X), and the number of pickups by both taxi and ridesharing in the

14

Electronic copy available at: https://ssrn.com/abstract=2997190

 Figure 4: Enough Demand across Geography

(a) Yellow and Green Taxis (b) Yellow and Green Taxis

focal period as a proxy for demand for rides (call this Y). We define a dummy variable “enough

demand” that equals 1 if Y ≥ X and 0 otherwise. Figure 4a plots each zone’s frequency of

“enough demand”, averaged across all 15-minute intervals of the month. We see that as we move

away from Midtown, the frequency of “enough demand” increases. Similarly, Figure 4b shows

a higher rate of enough demand in less accessible regions. These existence of excess demand,

combined with the finding that yellow and green taxis have different trip elasticity of distance in

outer boroughs, suggest that low demand for taxi is unlikely to be the sole reason for a relatively

small share of taxi pickups in the low accessibility regions. Admittedly, Taxi and ridesharing

services are different products, and therefore the observed differences could in principle be driven

by product differentiation. In Section 4, we explicitly model and estimate consumer preferences

for these attributes and reach a similar conclusion in our counterfactual analyses.

Lastly, we study the role of technology-aided matching and surge pricing in explaining the

geography of ridesharing. Using zone-minute level dynamic pricing on Uber and Lyft from the

data on ridesharing APIs, we plot the relative market share of ridesharing and surge multipliers

15

Electronic copy available at: https://ssrn.com/abstract=2997190

Figure 5: Ridesharing’s Market Share Negatively Correlates with Surge Multipliers and Its Stan-
dard Deviation

(a) (b)

of the taxi zones across regions with different levels of accessibility in Figure 5a. In the graph,

triangles, circles, and cubes represent regions that are within a short, medium, and long distance

to Midtown, respectively (here “short”, “medium”, and “long” are defined by the terciles of the

distance to Midterm). We find that the average share of ridesharing is positively correlated with

the region’s distance to Midtown, but it is negatively correlated with surge multiplier both within

and across each distance bracket. Additionally, if we look at long-distance trips (squares), the

relative share of ridesharing is more than 90% even in regions where the average surge multiplier

is close to 1. In Figure 5b, we repeat the analysis on the standard deviation of surge multipliers,

instead the average. We similarly find that a lower variation in surge multiplier predicts higher

share of ridesharing within a distance bracket. These patterns suggest that, even though dynamic

pricing may play a big role in matching demand and supply in highly accessible regions (Hall et al.

(2015) and Castillo et al. (2017)), it cannot be the main factor that leads to a larger market share

of ridesharing in less accessible regions. In other words, technology-aided matching seems to play

an important role in matching drivers and passengers in less accessible regions.

16

Electronic copy available at: https://ssrn.com/abstract=2997190

4 Translating The Geography of Ridesharing to Consumer Sur-

plus

The descriptive results in Section 3 discover geographic patterns of ridesharing rides relative to taxi

rides across regions with different accessibility, measured in terms of the distance to Manhattan

and in terms of economic distance to job and other economic opportunities. While these results are

novel and interesting on their own rights, one could argue that they are steps away from informing

policy making because it is unclear how to interpret the relative market shares.

In this section, we develop a simple discrete-choice model to estimate consumer preference

and demand for taxis and ridesharing. The purpose of this exercise is to translate the observed

geographic difference in the market share of ridesharing to consumer surplus through a counter-

factual analysis where ridesharing services are hypothetically removed. Additionally, the exercise

also sheds light on relative strength of different mechanisms of changes in consumer surplus, such

as from changes in price or wait time. We should note that the benefits come at the cost of making

more assumptions on the choice model, the data compatibility, and the validity of the instrument

for demand estimation, many of which can only be indirectly tested. Therefore, we view the this

part of our analysis as complementary to the descriptive evidence in Section 3.

4.1 Consumer Demand for Ridesharing

Market conditions constantly change in the market of rides: factors such as prices, wait time, and

conditions of competing alternatives vary at a high frequency across time within a location. Con-

sumers make transit decisions by evaluating these relevant time-varying characteristics. Therefore,

17

Electronic copy available at: https://ssrn.com/abstract=2997190

we set up the demand model at a granular route-time level, which we believe is a reasonable unit

of analysis in this market.

The product in this market is a ride, which is a transportation service that moves a consumer i

from a starting point j to a destination k in time t. We restrict attention to taxis, various ridesharing

service types, as well as the public transit. For a given route jkt, the consumer’s choice of trans-

portation mode depends on price, travel time, wait time, other observed and unobserved service-

specific characteristics, as well as the consumer’s own idiosyncratic taste, which is specified as

0
Usijkt = −αjkt Psjkt + βjkt (T otalojkt − W Tsjt − T Tsjkt ) + Xsjkt Θ + ξsjkt + φjkt + sijkt (1)

where s denotes one of the “inside options”, namely taxis and all service types on Uber and Lyft,

while o stands for the outside option — the public transit. we take the public transit as the outside

option because demand for rides is a “derived” demand — once a trip need (to go from Point A to

Point B at Time T) is identified, the agent has to choose one of the transportation options to make

the trip. We let T otalojkt denote the total travel time of the public transit, or the door-to-door time,

which is the sum of the walking time to the subway station or bus stop, wait time, time en route,

and walking time from the station/stop to the destination. Let W Tsjt denote the wait time of s at

location j and time t, and T Tsjkt is the time en route of s. Therefore, T otalojkt − W Tsjt − T Tsjkt

represents the amount of time s can save compared to the outside option, and βjkt is the marginal

utility of time saved. For example, for a route that is 35-minute long in total via the public transit,

if the UberX wait time is 5 minutes and the travel time is 20 minutes (thus total time is 25 minutes),

then the consumer’s utility of UberX is βjkt × 10 minutes higher than that of the baseline.

We let Psjkt denote the price of the trip, and −αjkt is the marginal utility of price. Let Xsjkt

18

Electronic copy available at: https://ssrn.com/abstract=2997190

represent a vector of observed service-specific characteristics that affect utility, Θ being the asso-

ciated vector of parameters. The other variables are defined as follows: ξsjkt is the unobserved

(to the researcher) route-service-specific utility component; φjkt is the utility difference between

all car options and the public transit, and it measures the comfort of not having to walk to the

subway station/bus stop and being able to sit for the entire trip, as opposed to not being able to sit

down with the public transit sometimes; sijkt is the consumer idiosyncratic error term. We further

assume that the travel time T Tsjkt is the same for all inside service types, since they travel on the

same route at the same time and thus they are subject to the same traffic and road conditions. As

a result, the subscript s in T Tsjkt is removed from here on. The purpose of regulating a common

travel duration will become clear later. Finally, we normalize the utilities of all inside options

against the utility of the outside option, i.e., the mean utility of the public transit is set at 0, i.e.,

Uoijkt = oijkt .9

4.2 Empirical Framework

Usually, the preference parameters of discrete choice models are obtained by estimating a mapping

from the characteristics space to realized market shares of options (e.g. Berry (1994), Berry et al.

(1995), Nevo (2000, 2001), and Petrin (2002)). However, market shares of service types cannot be

computed because we do not observe route-level (or, jkt-specific) public transit ridership. With a

market jkt defined so finely, we do not have a reasonable benchmark of market size to impute the

market shares like authors who study other settings (e.g. Berry et al. (1995) use population size of

a geographical area to proxy the market for automobiles). Nonetheless, Chevalier and Goolsbee
9
Liu et al. (2017) study the welfare impact of DiDi in China. They assume that consumers make choices at the
origin level, as opposed to the origin-destination pair level in our model.

19

Electronic copy available at: https://ssrn.com/abstract=2997190

(2009) demonstrate that odds ratios of the logit framework can be used to establish identification.

We follow this approach to avoid the problem due to unavailable market shares, while at the same

time flexibly allow for preference heterogeneity across markets. Specifically, we assume a Type

1 extreme value distribution of the error term, which amounts to a standard logit at each jkt cell.

Then the market share of s in the market jkt, though not observable, has to satisfy the following:

exp(δsjkt )
M arketSharesjkt = PS , (2)
1 + n=1 exp(δsjkt )

0
where δsjkt = −αjkt Psjkt + βjkt (T otalojkt − W Tsjt − T Tjkt ) + Xsjkt Θ + φjkt + ξsjkt is the mean

conditional utility of service s at jkt. To ease illustration, let taxi cabs be denoted as c. Then taking

logs of the predicted odds ratios of taxis’ share and the share of any ridesharing service type yields

Dcjkt
log( ) = αjkt (Psjkt − Pcjkt ) + βjkt (W Tsjt − W Tcjt ) (3)
Dsjkt

+ (Xcjkt − Xsjkt )0 Θ + ξcjkt − ξsjkt ,

where Dcjkt and Dsjkt are trip counts of taxi and service type s, respectively. Note that the terms

T otalojkt and T Tjkt are canceled out in the algebra, because they are common across service times

within the same route. Equation 3 indicates that the relative market shares of taxi and service type

s at the route-time level should be affected by their differences in price, wait time, as well as other

observed and unobserved characteristics.

We assume that consumer preference for price and wait time is the same within a route jkt. We

make this assumption because we have defined routes at a very granular level, and consumers with

the same pickup and dropoff location within a narrow time window likely have similar preferences.

20

Electronic copy available at: https://ssrn.com/abstract=2997190

On the other hand, consumer preference heterogeneity is allowed across jkt’s:

0
αjkt = Yjkt ΘA + α (4)

0
βjkt = Yjkt ΘB + β , (5)

where Yjkt is a row vector of dummy variables that contain various combinations of pickup areas

and time blocks, and ΘA and ΘB are the vectors of the corresponding coefficients for price and

time, respectively. Specifically, the areas include Lower Manhattan (dummy), Midtown Manhattan

(dummy), Upper East and West Manhattan (dummy), and Non-Manhattan Core (dummy). The

time blocks include morning rush (weekdays 7 a.m. - 9 a.m.), evening rush (weekdays 4 p.m. - 7

p.m.), weekday day time (weekdays 10 a.m. - 3 p.m.), weekday night (weekdays 8 p.m. - 11 p.m.),

weekday late night (weekdays midnight - 6 a.m.), weekend day time (weekends 5 a.m. - 5 p.m.),

weekend night (Friday 8 p.m. - 11 p.m. and weekends 6 p.m. - 11 p.m.), and weekend late night

(weekends midnight - 4 a.m.).

Note that consumers do not know the precise taxi wait time at a given location at a given time,

unlike in the case of ridesharing. We assume that consumers act according to their expectations of

W Tcjt for taxi time when choosing transportation modes, and −βjkt W Tcjt can be absorbed by the

many location and time fixed effects in the regression. The parameter for consumers’ preference

towards wait time, βjkt , is identified by the variation in the wait time of Uber and Lyft, namely

W Tsjt .

We choose to use the simple logit framework primarily due to its tractability in generating

analytical solutions that allow for identification in the absence of market shares. The simple logit

21

Electronic copy available at: https://ssrn.com/abstract=2997190

is well known for the independence of irrelevant alternatives problem, namely that the odds ratio

of two options need to stay the same regardless of the availability of an alternative. In principle,

we could estimate a model that allows for more flexible substitution pattern, such as a nested logit

model. However, this would require an instrumental variable that shifts demand in one nest but not

in the other, which is difficult to find. We will discuss the validity of our current instrument under

the simple logit framework in Section 4.3.

Additionally, we need to make assumption on the data compatibility. Specifically, Uber and

Lyft trip records published by TLC lack drop-off information and trip service types, which prevents

us from getting the trip counts of various Uber and Lyft service types at the jkt route, or Dsjkt . To

address this issue, we use the surveyed 70,277 Uber and Lyft historical trips in the same time period

to construct proxies. This sample consists of a random subsample and a convenience subsample.

In Appendix D, we show that the convenience subsample resembles the random subsample closely,

so we can rely on the randomness of the full sample to infer Dsjkt . To infer Dsjkt , we first estimate

a probit function to predict the probability of a certain trip that takes place in a certain jkt cell,

using the full sample. We then impute Dsjkt by distributing the total Uber/Lyft pickups at a jt to

various service types and destinations, based on these empirical probabilities (detailed in Appendix

E).

The level of analysis is at origin-destination-15 minute level. We choose a 15-minute time

window because it can approximate the real-time setting, while not dividing the time too finely

such that consumer information collection and decision making are split into different time periods.

Accordingly, all the explanatory variables are collapsed into the 15-minute averages, while trip

counts are sums over 15 minutes. We set the origin and destination at the taxi zone level and

PUMA (Public Use Microdata Areas) level, respectively, where taxi zones (263 in total) are a

22

Electronic copy available at: https://ssrn.com/abstract=2997190

more granular geographical unit than PUMA (55 in total). The reason for choosing such a level

of analysis is that prices and wait times vary across taxi zones so pickup locations need to be at

the taxi zone level to fully exploit the richness of the API data. On the other hand, having 263

destination zones means studying 263×263 markets, which is a stretch for imputation of Dsjkt .

Therefore, we study 263×55 markets at the 15 minute interval.

4.3 Identification

Our estimation, like other demand estimation studies, is subject to price and wait time endogene-

ity, given that Uber and Lyft pricing algorithms can potentially take into account many factors that

affect demand, both observed and unobserved to the researchers. Then a simple OLS estimation of

Equation 3 would lead to biased estimates of variables of interest. To deal with this potential en-

dogeneity problem, note that we control for service type, route, and time fixed effects at a granular

level. Next, we implement an IV strategy, and the IV should affect the market share of a service

type on a given route only through its impact on price, but it is otherwise uncorrelated with other

demand shocks.

We construct such an IV using a design feature of the ridesharing Apps that were present during

our sample period: passengers see and commit to a surge multiplier on Uber and Lyft Apps prior to

entering their destination on their phones. Exploiting this feature, we instrument for Psjkt − Pcjkt

using the average surge prices of trips into the focal zone in the previous time period. Because the

ridesharing Apps do not possess information on where the consumer is going, they cannot adjust

their surge multiplier according to the consumer’s destination, which creates the uncorrelatedness

of surge prices at the origin and the underlying demand condition at the destination.10 The exclu-
10
This may not be the case anymore given that Uber and Lyft now practice “upfront pricing”, which requires rider

23

Electronic copy available at: https://ssrn.com/abstract=2997190

sion assumption can be violated if platforms can somehow infer the distribution of destinations

for trip requests at a given origin. We rely on the many location and time fixed effects to help

alleviate this concern. On the other hand, surge pricing at the origin will affect the price in the

focal zone through its effect on the number of available drivers in the focal zone after they drop off

their consumers.

To deal with the endogeneity of W Tsjt , we use the total number of drop-offs at all neighboring

zones of the focal zone as the instrument. This essentially measures the stock and availability of

cars close to the focal zone, which affects the wait time at the focal zone. The exclusion restriction

assumption that we make is that the availability of cars in the neighboring zones is not correlated

with demand shocks in the focal zone.

We perform first-stage regressions to test the strength of the instruments. Since we have dozens

of endogenous variables (because of interactions of time-location dummies with price and wait

time), the first-stage tests entail an equal number of regressions, each of which regresses an en-

dogenous variable on all regressors. F statistics and partial R2 ’s are reported in Appendix Table 6.

Given that all F-statistics are bigger than 10 (Staiger and Stock, 1997), and that the partial R2 ’s are

relatively high (Shea, 1997), our instruments are not weak instruments.

The stochastic unobserved utility component ξ is assumed to be normally distributed with mean

zero and independent across both service types and jkt’s. The error term ξcjkt − ξsjkt in Equation

(3), however, creates a within-jkt correlation among the observations, as these observations share

a common part ξcjkt in the error. Specifically, COV (ξcjkt − ξsjkt , ξcjkt − ξs0 jkt ) 6= 0 for two on-

demand service types s and s0 in the same jkt. Given that we use instrumental variables that are not
destinations before showing a fixed price for the trip. In this case, origin price can be affected by the destination de-
mand shock, if the ridesharing platforms practice dynamic programming and discriminate riders based on destinations.
For example, the demand shock at the destination may lead the platform to willingly offer a low price at the origin,
just to have the driver relocated to the destination.

24

Electronic copy available at: https://ssrn.com/abstract=2997190

correlated with ξcjkt , a random effects estimator would be appropriate to deal with the correlation

in errors. One complication to the problem, however, is that the analysis sample is unbalanced —

as illustrated in Section 4.1, there are varying number of observations within a given jkt due to the

sub-sampling filters applied (Dcjkt > 0 and Dsjkt ≥ 0.1). We choose to follow the general method

proposed in Balestra and Varadharajan-Krishnakumar (1987) to estimate Equation (3) by Feasible

Generalized Two-Stage Least Squares. Details of the application to generate our estimator are

reported in Appendix F.

In the estimation, we drop airport pickups from the sample because only very few trips end in

neighboring zones of airports, and the correlation between these trips and airport wait time is very

weak to justify the use of the wait time IV. To-airport trips, however, are kept in the sample. We

include the dummy variable “to-airport” in the vector Yjkt to allow consumers on trips to the airport

to have additional price and time sensitivities. Similarly, we include the dummy variable “rain”

to allow for extra dis-utility of wait time. The vector of observed characteristics Xsjkt includes

luxury and capacity. Luxury measures the units of luxury service provided by the trip, which is a

dummy variable on UberBlack and UberSUV, multiplied by the duration of the jkt trip. Capacity

is defined similarly, except for service types UberXL, UberSUV, and LyftPlus. Finally, the set of

fixed effects includes pickup zone, drop-off PUMA, pickup hour by weekend (dummy), pickup

PUMA by time block, pickup PUMA by drop-off borough, and drop-off PUMA by time block,

dummies of trips at various trip duration levels (e.g. 10-minute trip, 15-minute trip, etc.).

25

Electronic copy available at: https://ssrn.com/abstract=2997190

 Table 1: Heterogeneous Consumer Preference

Price(α), Time(β) Lower Midtown UE & UW Non Core

Morning rush α 0.199*** 0.058* 0.136*** 0.242***
(0.044) (0.031) (0.016) (0.023)
β 0.554*** 0.655*** 0.479*** 0.501***
(0.091) (0.090) (0.046) (0.080)
Evening rush α 0.026** 0.094*** 0.127*** 0.099***
(0.012) (0.016) (0.013) (0.020)
β 0.807*** 0.676*** 0.487*** 0.644***
(0.053) (0.038) (0.042) (0.083)
Weekday night α 0.124*** 0.168*** 0.335*** 0.233***
(0.011) (0.018) (0.041) (0.016)
β 0.570*** 0.512*** -0.173 0.320***
(0.037) (0.031) (0.127) (0.077)
Weekday late night α 0.170*** 0.149*** 0.457*** 0.094***
(0.019) (0.015) (0.141) (0.021)
β 0.421*** 0.447*** -0.642 0.840***
(0.067) (0.050) (0.544) (0.095)
Weekday day time α 0.113*** 0.062*** 0.266*** 0.210***
(0.012) (0.012) (0.018) (0.014)
β 0.663*** 0.830*** 0.102 0.561***
(0.058) (0.064) (0.074) (0.117)
Weekend night α 0.209*** 0.274*** 0.294*** 0.210***
(0.019) (0.025) (0.022) (0.009)
β 0.272*** 0.115** -0.014 0.471***
(0.046) (0.054) (0.057) (0.046)
Weekend late night α 0.180*** 0.137*** 0.383*** 0.237***
(0.011) (0.014) (0.043) (0.010)
β 0.587*** 0.551*** -0.484*** 0.519***
(0.044) (0.032) (0.152) (0.059)
Weekend day time α 0.173*** 0.135*** 0.176*** 0.222***
(0.018) (0.014) (0.013) (0.015)
β 0.442*** 0.533*** 0.458*** 0.512***
(0.054) (0.056) (0.071) (0.069)
Airport α -0.080***
(0.013)
β 0.668***
(0.128)
Rain β 0.474***
(0.027)
Luxury (per service minute) -0.104***
(0.009)
Capacity (per service minute) -0.110***
(0.010)
N 14,464,715

Note: This table presents the demand estimation results. Throughout the table, α indicates a row of price
sensitivity estimates and β indicates a row of time sensitivity estimates. “Lower” stands for lower Manhat-
tan; “UE & UW” stands for Upper East and Upper West; “Non Core” stands for non Manhattan core. “Air-
port” is a dummy for to-airport trips. The time blocks are defined as: morning rush (weekdays 7am-9am),
evening rush (weekdays 4pm-7pm), weekday daytime (weekdays 10am-3pm), weekday night (weekdays
8pm-11pm), weekday late night (weekdays0am-6am), weekend day time (weekends 5am-5pm), weekend
night (Friday 8pm-11pm and weekends6pm-11pm), and weekend late night (weekends 0am-4am). Stan-
dard errors are in parentheses. * represents statistical26
significance at 10% level, ** 5%, and *** 1%.

Electronic copy available at: https://ssrn.com/abstract=2997190

4.4 Consumer Preference for Price and Wait Time

The estimation results are shown in Table 1. First of all, our results identify consumer preference

toward lower prices and shorter wait time — across almost all location-time-block combinations,

the price effects on utility are estimated to be significantly negative (−αjkt ), and the marginal utility

of time (βjkt ) is estimated positive and strong. These estimates directly indicate that consumers

are price-sensitive and value less wait time. In addition, these estimates suggest that consumers

are willing to make price-time trade-offs, and these economic fundamentals rationalize the use of

dynamic pricing by ridesharing companies.

Second, we find sensible heterogeneity in consumer price and time preference across locations

and times. For example, consumers in Midtown Manhattan during morning rush hours tend to be

more time-sensitive and less price-sensitive, compared to consumers in most other location-times.

This may reflect the preference of relatively high-income workers who rush into their workplaces

on weekday mornings. A very similar pattern appears in Lower Manhattan during evening rush

hours, which may be driven by Wall Street workers. Also, consumers going to the airport value

time more and are additionally less sensitive to price. Similarly, extra disutility of wait time is

found on consumers who hail a ride during the rain. These heterogeneities are consistent with

intuition and provide confidence on the identification. In addition, the coefficient estimates of

luxury and capacity are sizable, indicating that NYC consumers value these features that are made

conveniently available on ridesharings compared to the offline options.

27

Electronic copy available at: https://ssrn.com/abstract=2997190

4.5 Taxi Supply

To study consumer welfare gain by a counterfactual with no ridesharings, it is important to under-

stand how taxi supply would respond in this hypothetical situation. In this section, we propose a

model of taxi market equilibrium. The uniqueness of the taxi market, well articulated in Orr (1969),

is the difference between operating hours and passenger service units supplied: taxi drivers’ costs

depend on the hours of driving and searching for passengers, while their revenues depend on the

fare multiplied by the number of passenger service units supplied. This is due to the nature of the

taxi matching technology. Here, we modify the framework used in Orr (1969) to fit our purpose.

Let the taxi demand be specified as,

D = f (F a , q; M, ΘD )

where D is the number of passenger service units demanded; F a is the fixed administered fare; q

is the total operating hours of taxi drivers; M is the number of potential consumers; and ΘD is a

∂D
vector of demand parameters. Also, D is continuous in F a and ∂F a
< 0; D is continuous in q, and

∂D
∂q
> 0, because as more taxi hours are provided, consumers are more quickly matched to drivers

∂D ∂2D
and the average consumer wait time decreases; D(q = 0, F a ; M, ΘD ) = 0, ∂q q=0
>> 0, ∂q∂q
< 0,

∂2D
and ∂q∂M
> 0.

The taxi market is a market with a rather elastic supply of labor; only modest skills are required

to operate taxi cabs, and the practice of daily lease of medallions to the drivers imposes quite low

entry and exit costs.11 Cab drivers respond positively to wage increases by working longer hours
11
Farber (2015) documents “a fair amount of entry, exit, and reentry among taxi drivers”. Hall et al. (2017) demon-
strate the horizontal labor supply curve for Uber drivers, which may as well be the case for cab drivers.

28

Electronic copy available at: https://ssrn.com/abstract=2997190

(Chen and Sheldon, 2016; Farber, 2015; Hall et al., 2017). Under competitive conditions, the

market equilibrium is characterized as a steady state where marginal cost and average revenue are

equalized:
F a ∗ D(F a , q; M, ΘD )
= M C(q) (6)
q

∂M C
where M C(q) is the marginal cost of operating a taxi, M C(q) > 0, M C(q) << ∞, ∂q
≥ 0,

∂2M C
and ∂q∂q
≥ 0. The fixed medallions impose a hard constraint on the number of operating hours

available in the market, that is, the maximal amount of daily operating hours is the number of

medallions multiplied by 24 hours. Let this maximum be q̄. Then the algebra leads to that the

12
equilibrium operating hours supplied is concave in D up till q̄. Let wait time for taxi passengers

be defined as,

W T = f (q, D; ΘW T )

where WT is a continuous function, twice differentiable in q and D, with parameters denoted by

∂W T ∂2W T
the vector ΘW T . Particularly, ∂q
< 0 and ∂q∂q
> 0: holding other things constant, more taxi

service supplied leads to less wait time for consumers, yet this decrease in wait time diminishes

∂W T ∂2W T
as service units increase. ∂D
> 0 and ∂D∂D
> 0: holding other things constant, more trips

demanded lead to longer consumer wait time, and this increase in wait time is greater as more trips

are demanded.

A graphical characterization of the taxi market equilibrium is presented in Figure 6a, where the

equilibrium path is the combination of the part of q ∗ before q̄ and the part of vertical line above

q ∗ (the curve in red). An immediate implication on the equilibrium service units and equilibrium
12 ∂q Fa ∂2q
Differentiating both sides of Equation 6 with respect to D leads to ∂D = M C(q)+ ∂M C > 0. Then ∂D∂D =
∂q q
∂q
F a ∂D [2M C 0 (q)+M C 00 (q)q]
− [M C(q)+ ∂M C 2 < 0.
∂q q]

29

Electronic copy available at: https://ssrn.com/abstract=2997190

taxi wait time is depicted qualitatively in Figure 6b — wait time increases as equilibrium service

units increase after the capacity constraint. This is because supply cannot further adjust after the

∂W T ∂W T ∂q∗
capacity constraint. Therefore, ∂D∗
= ∂q∗ ∂D∗
+ ∂W
∂D
T
= 0 + ∂W
∂D
T
> 0.13 The intuition is simple:

as demand increases and maximal taxi capacity is reached, more and more consumers compete

with each other to get matched to a fixed number of operating taxi cabs, which leads to longer

average consumer wait time.14

Figure 6: Taxi Market Clearing

(a) Taxi Market Equilibrium (b) Model Prediction of Equilibrium Service Units
and Wait Time

One direct way to test the model prediction is to do a scatter plot of taxi pickups with taxi

wait time, and check whether the empirical pattern resembles Figure 6b. Unfortunately, in this

study we can not observe, estimate, or simulate the actual taxi wait time. However, Frechette et al.

(2016) are able to simulate taxi wait time from observed taxi cabs, taxi search time, and exogenous

time-varying factors, combined with a simulated matching function (Figure 6 of their paper). We
T ∂q∗
13
Before the capacity constraint q̄, the term ∂W ∂W T
∂q∗ ∂D∗ is negative, so the sign of ∂D∗ is undetermined. We use a
flat line in Figure 6b to describe the relationship between q and D before q̄, but it should be noted that this relationship
can be either positive, zero, or negative.
14
Note that the market equilibrium proposed here abstracts away from the spatial equilibrium models such as Lagos
(2000), Lagos (2003), and Buchholz (2015). It is possible that when there is an exogenous shock of demand, taxi cabs
relocate spatially and form a new spatial equilibrium, which results in a different average wait time than implied by
our model.

30

Electronic copy available at: https://ssrn.com/abstract=2997190

contrast their wait time estimates with UberTaxi wait time from our API queries in Figure 7, and

find that UberTaxi wait time follows a similar trend as their estimates across hours of day, although

UberTaxi wait time is less volatile. We believe that UberTaxi is a reasonable proxy for taxi wait

time and use it in the test of the taxi market equilibrium.

Figure 7: UberTaxi Wait Time and Taxi Wait Time from Frechette et al. (2016)

Figure 8: Taxi Trips and Taxi Wait Time

(a) Model Prediction of Equilibrium Service Units (b) Hourly Taxi Wait time and Taxi Trips in Manhat-
and Wait Time tan Core, Weekdays

The data strongly support the model prediction, as shown in Figure 8b. Overall, there is a pos-

itive correlation between taxi trips and wait time. In particular, the average wait time is relatively

low below some trip quantity threshold but becomes much higher after the threshold with a greater

variation. In addition, there is a sharp contrast between rush hours and non-rush hours: first, rush

31

Electronic copy available at: https://ssrn.com/abstract=2997190

hour wait time is on average higher; and second, the correlation between trip volume and wait time

after the threshold at 10,000 trips, is greater during rush hours (correlation coefficient is 0.42) than

non-rush hours (correlation coefficient is 0.19). This is likely due to the certain spatial distribu-

tion of commuting routes during rush hours, which exacerbates the within-location imbalance of

demand and supply. We leverage these empirical correlations in the counterfactual analysis.

4.6 Consumer Surplus

To evaluate the change in consumer surplus due to ridesharing, we adopt the concept of com-

pensating variation – how much consumers should be compensated if Uber and Lyft were to be

removed from the market such that the consumers can maintain the same level of utility? In the

counterfactual, we assume that taxis and the public transit are the only viable options. We also

assume that the public transit remains the same operation, without capacity constraint when more

riders substitute toward it; the taxi system remains the current fixed number of medallions and

administered fares.

There are two major reasons why consumers could be worse off in the absence of ridesharing

services: first, existing ridesharing users would lose all the amenities from ridesharing services

which they value more than other alternatives, due to revealed preference (that is, they would not

have used ridesharing services had these services not provided the users with the highest utility);

second, as shown in Section 4.5, when more consumers willingly substitute toward taxis, the equi-

librium average wait time increases, which makes existing taxi users worse off. We then take the

following procedure to compute consumer surplus:

1. First, we estimate φjkt , the utility difference between all car service types and the public

32

Electronic copy available at: https://ssrn.com/abstract=2997190

transit, which measures the comfort of not having to walk to the subway station/bus stop and being

able to sit for the entire trip. We allow for this utility in rush hours to differ from the utility in

non-rush hours to account for the extra disutility of riding the public transit during rush hours,

which is denoted as φr . In the search for φ and φr . We rely on the fact that once Uber and Lyft

are removed from the market, the number of taxi trips will increase because of sheer substitution.

That is, the values of φ and φr must be such that the corresponding counterfactual taxi ridership

is greater than or equal to the current ridership. It is important to note that using this boundary

equality (counterfactual ridership is equal to current ridership) leads to conservative estimates of

φ and φr , given the monotonic relationship between these values and consumers’ preference for

taxis, which then leads to a conservative estimate of taxi wait time change and a lower bound for

the welfare estimate.

2. We use taxi wait time estimates of Frechette et al. (2016) as the counterfactual taxi wait

times, and directly compute the mean conditional utilities of the public transit and taxis in the

counterfactual. Comparing these mean conditional utilities yields the counterfactual taxi trips.

3. We then infer current taxi wait times. We use the empirical relation between UberTaxi wait

time and taxi trips in Figure 8b to approximate the true relation between taxi wait times and trip

volumes, where we estimate the following OLS regression on day-hours when taxi trips exceed the

capacity threshold 10,000:

Taxi wait time = π0 + π1 Taxi trips (1,000s) (7)

where we estimate the equation separately for rush hours and non-rush hours; π1 is estimated at

0.155 for rush hours (N=233, t=7.09), and at 0.057 for non-rush hours (N=559, t = 4.51). The

33

Electronic copy available at: https://ssrn.com/abstract=2997190

estimated coefficients, counterfactual wait time (Frechette et al. (2016)), counterfactual taxi trips

(from Step (2)), and current taxi trips (directly from data) help us compute the current taxi wait

times. For example, for rush hours,

current wait time = 0.155*(current taxi trips - counterfactual taxi trips) + counterfactual wait time

(8)

4. With the inferred current taxi wait time, we compute the utility difference of current rideshar-

ing as well as taxi users by comparing consumers’ current options with counterfactual best options.

It is important to note that taxi wait time estimates in Frechette et al. (2016) are for Manhattan core

only. As a result, our inferred key values in Step 2, 3, and 4 so far, namely counterfactual taxi

trips and current taxi wait times, are for Manhattan core only. Therefore, we make assumptions on

Non-Manhattan core taxi wait times and computer consumer surplus accordingly. We take three

alternative values, namely 10 minute, 15 minutes, and 20 minutes, as various approximation of the

true outer boroughs taxi wait times.

We present the counterfactual results in Table 2. The consumer surplus of ridesharing is about

72 cents per dollar, or $14 for an average trip, when the taxi wait time in non-Manhattan core is a

conservative 10 minutes. As the taxi wait time increases, the value of ridesharing also increases as

the incumbent options becomes even less appealing. We find the majority of the consumer surplus

comes from shortened wait time, compared to taxis and the public transit. Luxury cars are also

valued significantly. Similarly, consumers value the utility of sitting in a car (thus not having to

walk and possibly stand in the public transit), as 13% of welfare comes from “comfort”. A very

thin share of the consumer welfare increase comes from price, yet this is not surprising since in

NYC, Uber and Lyft prices overall compare to taxi prices at the base-price level. In addition, taxi

34

Electronic copy available at: https://ssrn.com/abstract=2997190

 Table 2: Consumer Surplus Due to Entry of Ridesharing Platforms (Unit: Dollar)

Non-Manhattan Core Taxi Wwait Time

10 min 15 min 20 min
Consumer Surplus of Ridesharing Users
Per dollar spent on RS platforms 0.72 0.86 0.96
Per RS trip 14.05 16.79 18.67
Per RS service minute 0.92 1.10 1.22

Ridesharing Welfare Sources

Time 56.4% 58.2% 57.9%
Price 8.3% 6.5% 5.4%
Luxury 18.8% 16.0% 14.4%
Capacity 3.5% 3.0% 2.7%
Comfort 13.0% 16.2 % 19.5%

Consumer Surplus of Taxi Users

Per dollar spent on taxis 0.16 0.21 0.21

riders gain 16 cents per dollar spent, because taxi wait time becomes shorter as a result of consumer

substitution toward ridesharings.15

15
Similar benefits to users of incumbent goods or services due to lower prices (as a result of competition) are
documented in the literature, such as consumers of automobiles (Petrin, 2002), cable TV (Goolsbee and Petrin, 2004),
and hotels (Farronato and Fradkin, 2017).

35

Electronic copy available at: https://ssrn.com/abstract=2997190

 Figure 9: Per-dollar Consumer Surplus of Ridesharing Varies across Neighborhoods

Note: In the graph, the density of taxi and ridesharing trips across distance to midtown is plotted by lowess smoothing
against the left axis. The per-dollar consumer surplus across distance is plotted by lowess smoothing against the right
axis.

Given that the goal of this paper is to understand the geography of ridesharing and its implica-

tion on consumers, we compare the per-dollar consumer surplus across neighborhoods represented

by their distance to Midtown Manhattan (recall that this distance measure is highly correlated with

the accessibility measure). In Figure 9, we first plot the density of taxi and ridesharing trips across

distance to Midtown. Consistent with the previous results, we see that although both taxi and

ridesharing pickups are concentrated in regions that are closer to Midtown Manhattan, the level of

concentration is smaller for ridesharing trips. Next, we plot the geography of consumer surplus

from ridesharing. We find that it varies drastically across geography and exhibits a U-shape: pas-

sengers that are 5 to 15 miles from Midtown experience 60% larger consumer surplus (average

per-dollar consumer surplus is $1.15) relative to passengers that are within 5 miles from Mid-

town (average per-dollar consumer surplus is $0.72). For passengers that are more than 15 miles

from Midtown, their surplus ($0.86) is 19% larger than trip that start within 5 miles from Mid-

36

Electronic copy available at: https://ssrn.com/abstract=2997190

town. These patterns are intuitive given that alternative options are more conveniently available in

Manhattan than in outer boroughs, which makes ridesharing less valuable (in percentage terms) in

Manhattan than in outer boroughs. For consumers that are in the most remote areas (greater than

15 miles from Midtown Manhattan), the surplus becomes smaller relative to medium-range outer

boroughs (5 to 15 miles from Midtown Manhattan), likely because the alternative options (public

transit and taxis) are not as inaccessible as they are in medium-range outer boroughs.

Figure 10: The Geography of Consumer Surplus by Utility Component

Note: This graph plots the per-dollar consumer surplus against distance to midtown, separately for each of the five
utility components. All curves are plotted by lowess smoothing.

Furthermore, we decompose the geography of consumer surplus of ridesharing into distinct

utility components. As shown in Figure 10, the U-shaped pattern observed in the total consumer

surplus in Figure 9 appears to be mainly driven by the consumer surplus from saved time for

consumers in the medium-range outer boroughs. The geographical distribution of price-induced

consumer surplus is relatively uniform. Consumer surplus from luxury ridesharing options ap-

pears greater in Manhattan than in outer boroughs, which is consistent with the business-oriented

37

Electronic copy available at: https://ssrn.com/abstract=2997190

positioning of luxury options. Similar to price, the consumer surplus due to capacity and comfort

seems to be evenly distributed across geographies.

Taken together, our calculation of consumer surplus points to a highly uneven distributional

benefit of ridesharing across geographies. Neighborhoods that are within a medium distance to

Midtown Manhattan (i.e., 5 to 15 miles from Midtown) benefit the most from the presence of

ridesharing, and the major welfare source is the reduced time waste as compared to other trans-

portation modes.

5 Discussion

Mobility inequality correlates with economic inequality. Yet achieving inclusive mobility is a

difficult task, primarily due to the lack of incentives for public transit expansion. Taxis are a

good way of point-to-point transportation, but they tend to cluster in dense areas and leave non-

dense areas underserved, because this is an equilibrium response to the geographical distribution of

passengers and lack of real-time demand information. Therefore, the proposition that aims to make

taxi services more evenly distributed across space by expanding the medallion capacity constraint,

restricting pickup areas, or both is unlikely to produce satisfactory outcomes, as exemplified by the

case of NYC green taxis.

In this paper, we find that ridesharing could promote inclusive mobility through tech-aided

matching, which makes the ride-hailing service conveniently available for consumers in areas

that are underserved by other transportation modes. Specifically, we use data from actual trips

to present evidence that ridesharing coverage dominates that of taxis in less accessible areas. We

then provide evidence that the advanced driver-passenger matching of ridesharing companies could

38

Electronic copy available at: https://ssrn.com/abstract=2997190

mitigate the lack of information about demand in these areas. Finally, we interpret the geography

of ridesharing through the lens of a choice model, and the estimation results suggest a dispropor-

tionate larger gain in consumer surplus of ridesharing services in low accessibility areas.

The literature that explores the economics behind ridesharing platforms has primarily focused

on their average effects, i.e., how ridesharing platforms integrate massive real-time market infor-

mation to enhance capacity utilization, in a generic dense metropolitan market. Our findings shed

light on the importance of the underexplored distributional effects, i.e., the ability of ridesharing

to connect and serve various metropolitan neighborhoods with inferior access to other mobility

resources. Our findings suggest that technology can play a key role in mitigating geographical

disparity in transportation, and this calls for impact studies in other domains to explore other ways

to foster a more inclusive society.

References
Agrawal, A., C. Catalini, and A. Goldfarb (2015). Crowdfunding: Geography, social networks,
and the timing of investment decisions. Journal of Economics & Management Strategy 24(2),
253–274.
Babar, Y. and G. Burtch (2017). Examining the impact of ridehailing services on public transit use.
Balestra, P. and J. Varadharajan-Krishnakumar (1987). Full information estimations of a system of
simultaneous, equations with error component structure. Econometrics Theory 3(2), 223–246.
Berry, S., J. Levinsohn, and A. Pakes (1995). Automobile prices in market equilibrium. Econo-
metrica: Journal of the Econometric Society, 841–890.
Berry, S. T. (1994). Estimating discrete-choice models of product differentiation. The RAND
Journal of Economics, 242–262.
Blum, B. S. and A. Goldfarb (2006). Does the internet defy the law of gravity? Journal of
International Economics 70(2), 384–405.
Buchholz, N. (2015). Spatial equilibrium, search frictions and efficient regulation in the taxi in-
dustry. Technical report, Working paper.
Castillo, J. C., D. Knoepfle, and G. Weyl (2017). Surge pricing solves the wild goose chase. In
Proceedings of the 2017 ACM Conference on Economics and Computation, pp. 241–242. ACM.

39

Electronic copy available at: https://ssrn.com/abstract=2997190

Chen, M. K., J. A. Chevalier, P. E. Rossi, and E. Oehlsen (2017). The value of flexible work:
Evidence from uber drivers. Technical report, National Bureau of Economic Research.

Chen, M. K. and M. Sheldon (2016). Dynamic pricing in a labor market: Surge pricing and flexible
work on the uber platform. In EC, pp. 455.

Chevalier, J. and A. Goolsbee (2009). Are durable goods consumers forward-looking? evidence
from college textbooks. The Quarterly Journal of Economics 124(4), 1853–1884.

Cohen, P., R. Hahn, J. Hall, S. Levitt, and R. Metcalfe (2016). Using big data to estimate consumer
surplus: The case of uber. Technical report, National Bureau of Economic Research.

Cramer, J. and A. B. Krueger (2016). Disruptive change in the taxi business: The case of uber. The
American Economic Review 106(5), 177–182.

Farber, H. S. (2015). Why you cant find a taxi in the rain and other labor supply lessons from cab
drivers. The Quarterly Journal of Economics 130(4), 1975–2026.

Farronato, C. and A. Fradkin (2017). The welfare effects of peer entry in the accommodations
market: The case of airbnb. Technical report, Working paper.

Frechette, G. R., A. Lizzeri, and T. Salz (2016). Frictions in a competitive, regulated market
evidence from taxis.

Goolsbee, A. and A. Petrin (2004). The consumer gains from direct broadcast satellites and the
competition with cable tv. Econometrica 72(2), 351–381.

Greenwood, B. N. and S. Wattal (2017). Show me the way to go home: An empirical investigation
of ride-sharing and alcohol related motor vehicle fatalitie. MIS Quarterly 41(1).

Hall, J., J. Horton, and D. Knoepfle (2017). Labor market equilibration: Evidence from uber.

Hall, J., C. Kendrick, and C. Nosko (2015). The effects of ubers surge pricing: A case study. The
University of Chicago Booth School of Business.

Hall, J. D., C. Palsson, and J. Price (2018). Is uber a substitute or complement for public transit?
Journal of Urban Economics 108, 36–50.

Hall, J. V. and A. B. Krueger (2016). An analysis of the labor market for ubers driver-partners in
the united states. Technical report, National Bureau of Economic Research.

Hering, L. and S. Poncet (2010). Market access and individual wages: Evidence from china. The
Review of Economics and Statistics 92(1), 145–159.

Holzer, H. J., K. R. Ihlanfeldt, and D. L. Sjoquist (1994). Work, search, and travel among white
and black youth. Journal of Urban Economics 35(3), 320–345.

Hortaçsu, A., F. A. Martı́nez-Jerez, and J. Douglas (2009). The geography of trade in online trans-
actions: Evidence from ebay and mercadolibre. American Economic Journal: Microeconomics,
53–74.

40

Electronic copy available at: https://ssrn.com/abstract=2997190

Ihlanfeldt, K. R. and D. L. Sjoquist (1990). Job accessibility and racial differences in youth em-
ployment rates. The American economic review 80(1), 267–276.

Ihlanfeldt, K. R. and D. L. Sjoquist (1991). The effect of job access on black and white youth
employment: A cross-sectional analysis. Urban Studies 28(2), 255–265.

Kaufman, S., M. L. Moss, J. Tyndall, and J. Hernandez (2014). Mobility, economic opportunity
and new york city neighborhoods.

Lagos, R. (2000). An alternative approach to search frictions. Journal of Political Economy 108(5),
851–873.

Lagos, R. (2003). An analysis of the market for taxicab rides in new york city. International
Economic Review 44(2), 423–434.

Li, Z., Y. Hong, and Z. Zhang (2016). An empirical analysis of on-demand ride sharing and traffic
congestion.

Liu, M., E. Brynjolfsson, and J. Dowlatabadi (2018). Do digital platforms reduce moral hazard?
the case of uber and taxis. Technical report, National Bureau of Economic Research.

Liu, M., T. Tunca, Y. Xu, and W. Zhu (2017). An empirical analysis of price formation, utilization,
and value generation in ride sharing services.

Nevo, A. (2000). Mergers with differentiated products: The case of the ready-to-eat cereal industry.
The RAND Journal of Economics, 395–421.

Nevo, A. (2001). Measuring market power in the ready-to-eat cereal industry. Econometrica 69(2),
307–342.

Orr, D. (1969). The” taxicab problem”: A proposed solution. Journal of Political Economy 77(1),
141–147.

Owen, A. and B. Murphy (2018). Access across america: Transit 2017.

Petrin, A. (2002). Quantifying the benefits of new products: The case of the minivan. Journal of
political Economy 110(4), 705–729.

Shea, J. (1997). Instrument Relevance in Multivariate Linear Models : A Simple Measure. The
Review of Economics and Statistics 79(2), 348–352.

Staiger, D. and J. H. Stock (1997). Instrumental Variables Regression with Weak Instruments.
Econometrica 65(3), 557.

Tomer, A., E. Kneebone, R. Puentes, and A. Berube (2011). Missed opportunity: Transit and jobs
in metropolitan america.

Wang, Y., C. Wu, and T. Zhu (2019). Mobile hailing technology and taxi driving behaviors.
Marketing Science 38(5), 734–755.

41

Electronic copy available at: https://ssrn.com/abstract=2997190

Zervas, G., D. Proserpio, and J. W. Byers (2017). The rise of the sharing economy: Estimating the
impact of airbnb on the hotel industry. Journal of Marketing Research 54(5), 687–705.

Zhang, K., H. Chen, S. Yao, L. Xu, J. Ge, X. Liu, and M. Nie (2019). An efficiency paradox of
uberization. Available at SSRN 3462912.

Zhang, Z. and B. Li (2017). A quasi-experimental estimate of the impact of p2p transportation

platforms on urban consumer patterns. In Proceedings of the 23rd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, pp. 1683–1692. ACM.

42

Electronic copy available at: https://ssrn.com/abstract=2997190

Appendix A Summary Statistics

Table 3: Uber and Lyft Dynamic Pricing and Wait Time

Variable Mean Std. Dev. Min Max N

Surge Frequency
UberX 0.073 0.260 0 1 32398537
UberXL 0.072 0.259 0 1 32398537
UberBlack 0.012 0.112 0 1 32359662
UberSUV 0.013 0.114 0 1 32421149
UberPool 0.080 0.272 0 1 32359652
Lyft 0.180 0.384 0 1 31257800
LyftLine 0.180 0.384 0 1 31257800
LyftPlus 0.180 0.384 0 1 31257800
Surge Multiple
UberX 1.037 0.1622 1 4.2 32398537
UberXL 1.037 0.1619 1 4.2 32398537
UberBlack 1.008 0.0856 1 2.9 32359662
UberSUV 1.008 0.0884 1 2.9 32421149
UberPool 1.022 0.1028 1 3.4 32359652
Lyft 1.100 0.2727 1 5.0 31257800
LyftLine 1.100 0.2727 1 5.0 31257800
LyftPlus 1.100 0.2727 1 5.0 31257800
Wait Time (minutes)
UberX 6.949 8.912 1 45 32421149
UberXL 13.509 14.794 1 45 32421149
UberBlack 8.821 10.176 1 45 32421149
UberSUV 14.059 15.324 1 45 32421149
UberPool 7.496 9.870 1 45 32421149
Lyft 7.034 9.882 1 45 33133051
LyftLine 6.884 9.869 1 45 33133051
LyftPlus 9.510 9.993 1 45 33133051
Note: The data for this table come from Uber and Lyft API queries, June-August 2016, where both price
and wait time of 263 NYC zones are queried in approximately one-minute intervals. The small variation in
variable sizes reflects rare cases of missing values and/or duplicated queries. There is no variation in surge
frequency or surge multiples across service types within Lyft platform.

43

Electronic copy available at: https://ssrn.com/abstract=2997190

 Table 4: NYC Taxi Trips Records

Variable Mean Std. Dev. Min Max N

Trip duration (minutes)† 14.378 11.436 1 180 47483432
Trip distance (miles) 2.985 3.515 0.01 199 47483432
Base fare 12.858 10.039 0 314 47483432
Extra fee 0.337 0.435 0 21.5 47483432
MTA tax 0.498 0.023 0 2.34 47483432
Tip 1.685 2.267 0 300 47483432
Tolls 0.270 1.249 0 111.65 47483432
Improvement fee 0.299 0.011 0 0.6 47483432
Total fare 15.955 12.293 2.54 360.34 47483432
Passenger count 1.631 1.283 0 9 47483432
Yellow taxi 0.887 0.315 0 1 47483432
Manhattan pickup 0.850 0.356 0 1 47483432
Queens pickup 0.084 0.277 0 1 47483432
Bronx pickup 0.005 0.074 0 1 47483432
Brooklyn pickup 0.060 0.237 0 1 47483432
Staten Island pickup 0.000 0.006 0 1 47483432
Manhattan dropoff 0.827 0.377 0 1 47483432
Queens dropoff 0.076 0.266 0 1 47483432
Bronx dropoff 0.012 0.110 0 1 47483432
Brooklyn dropoff 0.082 0.275 0 1 47483432
Staten Island dropoff 0.000 0.014 0 1 47483432
Note: Trip duration is calculated as the difference between the pickup time and the drop-off time. Unreasonable
trips from the raw data are dropped using the following filters: trips with any negative cost components (cost com-
ponents include base fare, extra fee, MTA tax, tip, tolls, improvement fee), trips with negative distance, trips with
negative duration, trips greater than 200 miles, trips longer than 180 minutes. In total, less than 0.5% of the raw
sample are dropped by these filters.

44

Electronic copy available at: https://ssrn.com/abstract=2997190

Appendix B Ridesharing and Taxi Rides Across Hour of the Day
Besides the substitutability between ridesharing and taxi across geography observed in Figure 2,
we can also see this across time. Figure 11 shows that ridesharing gain larger market shares around
4 a.m. and 4 p.m., likely due to the morning and afternoon taxi shift changes16 that create an outflux
of taxi cabs from Manhattan17 . This figure suggests that consumers substitute toward ridesharing
more during the hours when taxis become less available.

Figure 11: More ridesharing Trips during Taxi Shift-change Hours

(a) NYC taxi, Uber, and Lyft trips by time of an (b) Shares of Uber and Lyft trips by time of an
average weekday average weekday

Appendix C Identifying Uber and Lyft Trips from TLC FHV Trip Records
Data
The FHV trip data does not specifically indicate the company name of each trip, instead it shows the
trip’s dispatching base number. Using the official TLC list of FHV bases 18 , we are able to identify
Uber and Lyft trips by the correspondence between base numbers and company names. Specif-
ically, the base numbers associated with Uber are B02512, B02395, B02617, B02682, B02764,
B02765, B02835, B02836, B02864, B02865, B02866, B02867, B02869, B02870, B02871, B02872,
B02875, B02876, B02877, B02878, B02879, B02880, B02882, B02883, B02884, B02887, B02888,
and B02889. The base numbers associated with Lyft are B02510 and B02844.

Appendix D Field Collection of Uber and Lyft Trip Records

We conducted two rounds of field surveys of Uber and Lyft drivers. The first round took place in
January 2017 and the second route took place in March 2017. In both rounds, we requested for
drivers’ historical trip records from June to August 2016 so that all of our data sources are from
the same time period.
We employed two sampling strategies in the data collection: a random sample and a conve-
nience sample. In the collection of the random sample, the research team split into 4 groups,
16
The majority of yellow medallion cabs are operated two shifts per day.
17
Most of taxi leasing garages are located outside of Manhattan.
18
See http://www.nyc.gov/html/tlc/downloads/pdf/find_a_ride.pdf

45

Electronic copy available at: https://ssrn.com/abstract=2997190

where two groups started around 9am at 2 locations in Manhattan, and the other two groups started
at 2 locations in Brooklyn. Each group requested a trip to a randomly-selected borough out of
Manhattan, Bronx, Brooklyn, and Queens. Upon arrival and before the driver answered a subse-
quent trip, the group made request to the driver for voluntary data disclosure. If the request was
declined, they then offered a small sum of money in exchange for the data. The group collected
as many trips as possible when the driver willingly accepted the request, either voluntarily or with
a small sum of money, and the group chose to walk away when the monetary offer was rejected.
The group repeated the same process throughout the day untill 9pm. In total, the research team
collected 10,333 trips from 56 drivers out of 76 attempted.
For the convenience sample, we approached Uber and Lyft drivers at places where they either
were taking a break and/or were between trips. These places include restaurants, coffee shops,
street corners, and parking lots. We followed the same data request procedure as for the random
sample. Because we had more time interacting with drivers and recording data this way, the con-
venience sample is multiple times as large as the random sample.
The validity of this paper’s estimation approach crucially relies on how representative the field-
collected ridesharing trips are of the population of ridesharing trips. We find support of the rep-
resentativeness by demonstrating that the distribution of field-collected ridesharing pickups aligns
with that of the TLC ridesharing pickups, across times and geographies (Table 5).

46

Electronic copy available at: https://ssrn.com/abstract=2997190

Table 5: Comparison between TLC Ridesharing Pickups and Field-collected Ridesharing Pickups

TLC data Field

All Time
LM 0.1884 0.1749
Midtown 0.2667 0.2863
UE & UW 0.0998 0.1448
NMC 0.4451 0.3940
Morning rush
LM 0.1606 0.1487
Midtown 0.2331 0.2770
UE & UW 0.1482 0.1831
NMC 0.4580 0.3912
Evening rush
LM 0.1921 0.1681
Midtown 0.3239 0.2875
UE & UW 0.1099 0.1538
NMC 0.3742 0.3906
Weekday night
LM 0.2171 0.2096
Midtown 0.3051 0.3283
UE & UW 0.0804 0.1402
NMC 0.3973 0.3219
Weekday late night
LM 0.1839 0.2009
Midtown 0.2303 0.3327
UE & UW 0.0717 0.1257
NMC 0.5142 0.3407
Weekday day time
LM 0.1692 0.1863
Midtown 0.2736 0.2946
UE & UW 0.1155 0.1677
NMC 0.4417 0.3598
Weekend night
LM 0.2002 0.1549
Midtown 0.2537 0.2296
UE & UW 0.0862 0.1213
NMC 0.4599 0.4942
Weekend late night
LM 0.2356 0.1686
Midtown 0.2241 0.2584
UE & UW 0.0485 0.0705
NMC 0.4918 0.5437
Weekend day time
LM 0.1671 0.1349
Midtown 0.2282 0.2348
UE & UW 0.1052 0.1381
NMC 0.4995 0.4923

47

Electronic copy available at: https://ssrn.com/abstract=2997190

Appendix E Inference of Dsjkt
Due to the data limitation of Uber and Lyft trips published by TLC, the finest level one can get is
U ber Lyf t U ber
Djt and Djt , where Djt measures the total trip counts of all 5 service types on Uber at the
Lyf t
pickup location j in time t, and Djt is the total trip counts of all 3 service types on Lyft at the
pickup location j in time t. To infer trip counts at a finer sjkt (type-pickup-dropoff-time) level,
we exploit the field-collected sample of 70,277 Uber and Lyft trip records using the following
procedure:
Step 1: Construct a vector of zeroes, whose length is s × j × k × t, i.e., at the type-month-day-
time-hour-15min level. Fill any sjkt cell with 1, if a trip is observed in that particular cell from
the field-collected sample. 19 Then the vector contains 0’s and 1’s.
Step 2: Estimate a probit model of the vector in Step 1 to predict the probability of a trip in
sjkt by a number of location-time fixed effects:

P r(1 trip in sjkt) =f (pickup zone, dropoff puma, service type, pickup hour,
pickup borough × pickup hour, pickup borough × dropoff puma,
pickup borough × service type, dropoff puma × service type)

U ber Lyf t
Step 3: Calculate Dsjkt by distributing Djt and Djt into service types and drop-off lo-
cations. This requires constructing weights using the estimated psjkt in Step 2, and applying the
following formulas (Note that s = 1, 2, 3, 4, 5, for 5 service types on Uber, and s = 6, 7, 8, for 3
service types on Lyft):
For Uber:
psjkt U ber
Dsjkt = P263 P5 Djt
k=1 s=1 psjkt
For Lyft:
psjkt Lyf t
Dsjkt = P263 P8 Djt
k=1 s=6 psjkt
U ber Lyf t
These weights ensure that the inferred Dsjkt ’s return the value of Djt and Djt , when summed
over service types and drop-off locations.

Appendix F Feasible Generalized Two-Stage Least Square Estimator

Denote esjkt = ξcjkt − ξsjkt , where the variances of ξcjkt and ξsjkt are σc2 and σs2 , respectively.
Let Sjkt be the set of ridesharing service types available at a particular jkt; Tjk is the set of time
periods for a particular route jk; Kj is the set of destinations for a particular pickup location j; J
is the set of all available pickup locations. Further denote Csjkt as the number of unique sjkt’s and
19
In several rare cases we observe two trips within the same sjkt cell. In these cases, we randomly drop one of the
two trips.

48

Electronic copy available at: https://ssrn.com/abstract=2997190

Cjkt as the number of unique jkt’s, which are calculated by summing the cardinality the relevant
sets
XX
Cjkt = |Tjk | (9)
J Kj
XXX
Csjkt = |Sjkt | (10)
J Kj Tjk

Then the estimator can be constructed by the following procedure:

Step 1: Estimate Equation (3) by a simple two-stage least squares regression without account-
\
ing for correlated errors, which leads to the composite residual êsjkt = ξcjkt − ξsjkt .
Step 2: Decompose the composite residual
1 X
ξˆcjkt = êsjkt (11)
|Sjkt | S
jkt

and

ξˆsjkt = êsjkt − ξˆcjkt (12)

Step 3: Compute the variance estimates σ̂c2 and σ̂s2 , using the decomposed residuals ξˆcjkt and
ξˆsjkt
1 XXX ˆ 1 XXX ˆ
σ̂c2 = (ξcjkt − ξcjkt )2 (13)
Cjkt J K T Cjkt J K T
j jk j jk

and
1 XXXX 1 XXXX
σ̂s2 = (ξˆsjkt − ξˆsjkt )2 (14)
Csjkt J Kj Tjk Sjkt
Csjkt J Kj Tjk Sjkt

Step 4: Construct a Csjkt × Csjkt block diagonal matrix with Cjkt number of blocks
..
 
.
O O
− 21
  
Ω̂ =  O Q̂jkt |Sjkt |×|Sjkt | O 

 (15)
...
O O Csjkt ×Csjkt
 
where each block element ( Q̂jkt |S |×|S | ) is a square matrix with diagonal elements equal to
jkt jkt
1 1 1−|Sjkt | 1 1 1
(
|Sjkt | (|S |σ̂ 2 +σ̂ 2 ) 12
− σ̂s
) and off-diagonal elements equal to (
|Sjkt | (|S 2 2
1 − σ̂s
) . O are
jkt c s jkt |σ̂c +σ̂s ) 2
matrices of zeros.
Step 5: The random effects estimator can be constructed explicitly as
0 0 0 0 0 0
(α̂ β̂ Θ̂)0 = (X∗ Z∗ (Z∗ Z∗ )−1 Z∗ X∗ )−1 X∗ Z∗ (Z∗ Z∗ )−1 Z∗ D∗ (16)
where Z denotes the matrix that contains all the instruments, and

49

Electronic copy available at: https://ssrn.com/abstract=2997190

  .. .. .. 
. . .
X = Psjkt − Pcjkt W Tsjt − W Tcjt (Xcjkt − Xsjkt )0 
 
.. .. ..
. . .
..
 
.
 Dcjkt 
D = log( Dsjkt )


..
.

1
X∗ = Ω̂− 2 X
1
D∗ = Ω̂− 2 D
1
Z∗ = Ω̂− 2 Z

Appendix G First Stage Tests for Instrumental Variables

Table 6 reports first-stage regressions to test the strengths of the instruments. Each number rep-
resents either the F-statistic or the partial R2 associated with the regression of an endogenous
variables on all regressors. For example, the fist number in the table, 431.85, represents the F
statistic of all instruments with respect to prices of Lower Manhattan in morning rush hours. The
partial R2 , 0.097, measures how much of the variation in prices of Lower Manhattan in morning
rush hours is explained by all instruments. Given that all F-statistics are bigger than 10 (Staiger
and Stock, 1997), and the partial R2 ’s are relatively high (Shea, 1997), our instruments are not
weak instruments.

50

Electronic copy available at: https://ssrn.com/abstract=2997190

 Table 6: First Stage F-test and Partial R2

Price Wait Time

F-stat Partial R2 F-stat Partial R2
Morning rush
LM 431.85 0.097 5039.903 0.035
Midtown 389.582 0.155 3394.364 0.083
UE & UW 3952.628 0.118 2682.597 0.062
NMC 33.927 0.080 12.481 0.044
Evening rush
LM 1113.876 0.146 1538.263 0.096
Midtown 586.41 0.168 1691.842 0.114
UE & UW 3062.924 0.123 64973.491 0.139
NMC 28.592 0.092 16.06 0.078
Weekday night
LM 709.163 0.132 7716.904 0.115
Midtown 1601.706 0.140 2400.526 0.126
UE & UW 4963.373 0.143 130478.994 0.147
NMC 36.802 0.095 42.819 0.102
Weekday late night
LM 2758.778 0.146 3040.974 0.074
Midtown 13701.841 0.178 4484.22 0.091
UE & UW 2965.271 0.163 6476.689 0.103
NMC 33.965 0.110 26.895 0.06
Weekday day time
LM 1342.167 0.200 1055.851 0.077
Midtown 1942.77 0.244 1425.041 0.111
UE & UW 3207.152 0.185 4107.147 0.074
NMC 39.43 0.076 10.827 0.038
Weekend night
LM 3511.998 0.147 5324.336 0.136
Midtown 2488.762 0.158 1489.467 0.143
UE & UW 20343.801 0.144 170436.431 0.145
NMC 47.637 0.123 46.362 0.094
Weekend late night
LM 734.557 0.124 3788.657 0.147
Midtown 7870.996 0.174 6663.631 0.147
UE & UW 44411.786 0.163 12545.38 0.188
NMC 58.275 0.132 30.523 0.169
Weekend day time
LM 1815.033 0.147 4033.533 0.094
Midtown 7080.772 0.181 3072.913 0.106
UE & UW 4326.934 0.142 14718.355 0.100
NMC 55.469 0.096 27.511 0.078
To Airport 165.552 0.225 100.863 0.077
Rain 1128.828 0.045

51

Electronic copy available at: https://ssrn.com/abstract=2997190