ABSTRACT
Title of dissertation: ESSAYS ON EMPIRICAL INDUSTRIAL
ORGANIZATION
Yong joon Paek
Doctor of Philosophy, 2018
Dissertation directed by: Professor Andrew Sweeting
Department of Economics
The dissertation focuses on two issues that are broadly related to the urban plan-
ning of Washington, DC. The first two chapters consider the burgeoning food truck
industry in Washington, DC and the third chapter considers public transit rider-
ship and the impact of large maintenance programs that cause temporary but large
decreases in service quality.
In Chapter 1 I build and estimate a tractable model that captures some key
characteristics of the food truck industry. Characteristics of the industry such as
there being many small firms playing an entry game with various dimensions of
heterogeneity (for example, cuisine genre and quality) render regulations and poli-
cies difficult to assess and design, leading to local regulators resorting to ‘ad hoc’
policies to regulate the industry. For example, in Washington, DC scarce parking
spots at popular lunch locations are allocated through a random lottery. This
highlights the importance of a tractable model that captures importance features
of the industry which can be estimated and used to consider counterfactual policies.
In Chapter 2 I consider two counterfactual scenarios. In the first counterfactual
scenario I reduce the reach of the lottery and I find that the lottery allows for the
survival of firms with lower quality and leads to higher prices. Expected utility for
consumers are lower and firm profits are higher in current regime compared to a
couterfactual regime where some locations are not included in the lottery. The net
welfare effect for this counterfactual scenario is an increase in total daily welfare
of $2.294.18. In my second counterfactual scenario where the non-lottery locations
have their parking capacity increased by 2 spaces, I find a positive impact for both
truck owners and consumers with a net increase in total daily welfare of $8,260.26.
Chapter 3 considers the impact of large public transit maintenance programs
on long-run ridership. An agency that manages large transit systems must make
investments to maintain a level of quality to sustain ridership. If consumers face
switching costs when changing their mode of transport, the large and unavoidable
disruptions to services resulting from a large maintenance program may provide
a sufficient negative utility shock for riders to substitute to alternative modes of
transport and not return after the repairs are completed. In this chapter I con-
sider such indirect costs that a transportation agency may incur in the context
of Metrorail, the subway system that stretches through the District of Columbia
(DC), Maryland (MD), and Virginia (VA) operated by the Washington Metropoli-
tan Area Transit Authority (WMATA). I find that there has been a persistent
drop in ridership up to 10 months after the repairs on certain tracks have been
completed.
Essays in Empirical Industrial Organization
by
Yong joon Paek
Dissertation submitted to the Faculty of the Graduate School of the
University of Maryland, College Park in partial fulfillment
of the requirements for the degree of
Doctor of Philosophy
2018
Advisory Committee:
Professor Andrew Sweeting, Chair
Professor Ginger G. Jin
Professor Dan Vincent
Professor Mary Zaki
Professor Joshua Linn
© Copyright by
Yong joon Paek
2018
Dedication
To my parents E.H. Min, C.W. Paek, and grandparents G.S. Min, H.B. Paik, J.S.
Lee, H.W. Kim.
ii
Acknowledgements
This dissertation is a product of a myriad of inputs. Most importantly, the advice
and support from my primary advisor Andrew Sweeting who was ever patient and
generous in sharing his valuable time with me. The advice and conversations I’ve
had with Prof. Sweeting have become the core and foundation of my education at
the University of Maryland, College Park.
I am especially grateful to my dissertation committee for their time. Prof. Gin-
ger Gin’s industrial organization course is what inspired me to finally choose the
field I have ultimately written my dissertation in. Prof. Dan Vincent opened my
eyes to the importance of having a sound theoretical motivation even in empirical
work. I am grateful to Prof. Mary Zaki for meeting with me and her personal
interest in my topic (food trucks). Prof. Joshua Linn, in the short time we have
had correspondence has offered me great feedback on the final chapter of this dis-
sertation.
I could not have completed this dissertation without the great company of the
friends I have made here in Washington, DC. Sean Hector and Thomas Hegland
have been a constant through most of my time here and have allowed me to grow
personally and professionally. Assistance in crucial parts of this dissertation from
Yongjoon Park, a colleague and better friend has also made this dissertation pos-
sible. Silver Yang, has supported me immensely and I could not have pushed on
without her understanding and patience.
iii
Contents
Dedication ii
Acknowledgements iii
Contents iv
List of Tables vi
List of Figures vii
Introduction 1
1 Location Choice and Price Competition with Differentiated Prod-
ucts and Many Firms: An Application to the Mobile Vending
Industry in Washington, DC. 4
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Related Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Industry Details and Data . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.1 The Food Truck Industry . . . . . . . . . . . . . . . . . . . 10
1.3.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.4.1 Demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.4.2 Trucks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.4.3 Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.5 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.5.1 Demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.5.2 Conditional Logit Location Entry Probabilities . . . . . . . . 37
1.5.3 Marginal Costs . . . . . . . . . . . . . . . . . . . . . . . . . 41
1.5.4 Location Entry Costs and Scale Parameter . . . . . . . . . . 43
1.6 Model Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
1.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2 Counterfactual Policy Analysis: Examining the Welfare Impacts
of Washington, DC’s Mobile Roadside Vending License Program 53
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.1.1 Method of Comparison for Counterfactual Scenarios . . . . . 56
2.2 The Impact of the Counterfactual Policies . . . . . . . . . . . . . . 58
2.2.1 Counterfactual I: Reducing the Reach of the Lottery . . . . 58
iv
2.2.2 Counterfactual II: Increasing Parking Capacity at Market
Locations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
2.2.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3 The Impacts of Large Disruptions on Long-Run Public Transit
Ridership: An Analysis of Washington, DC’s Subway Transit Sys-
tem. 73
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.2 Institutional Background . . . . . . . . . . . . . . . . . . . . . . . . 75
3.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.3.1 Data Sources . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.3.2 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . 81
3.4 Estimation and Analysis . . . . . . . . . . . . . . . . . . . . . . . . 83
3.4.1 The Impact of SafeTrack on Ridership . . . . . . . . . . . . 83
3.4.2 The Impact of Service Quality and Fares on Ridership . . . 93
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
Bibliography 98
v
List of Tables
1.1 Location Characteristics from the LODES data . . . . . . . . . . . 19
1.2 Summary Statistic from the Demand Data . . . . . . . . . . . . . . 19
1.3 Comparison of the distribution of cuisine genres within each data set 23
1.4 Nested Logit Estimation Specifications . . . . . . . . . . . . . . . . 34
1.5 Location F.E. from Consumer Demand Estimates . . . . . . . . . . 36
1.6 Truck Fixed-Effects OLS Projection on Individual Characteristics. . 37
1.7 Estimated Entry Costs. . . . . . . . . . . . . . . . . . . . . . . . . . 45
1.8 Estimated Entry Costs. . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.1 Entry Cost Changes Under the Counterfactual Policy . . . . . . . . 59
2.2 Consumer Surplus (CS) per Consumer by location and Implied Total
Change in CS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.3 Change in Entry Costs Under Counterfactual Parking Capacities. . 67
2.4 Consumer Surplus (CS) per Consumer by location and Implied Total
Change in CS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.1 SafeTrack Repairs Schedule. . . . . . . . . . . . . . . . . . . . . . . 78
3.2 OLS Regression of Pre-treatment Growth Rates for Treated and
Control Stations Pairs . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.3 Heterogeneity in the Impact of SafeTrack . . . . . . . . . . . . . . . 87
3.4 Treatment Characteristics and Estimation Results. . . . . . . . . . 88
3.5 The Dynamic Impact of SafeTrack on Ridership. . . . . . . . . . . . 92
3.6 The Impact of Delays and Fare Increases on Ridership. . . . . . . . 94
vi
List of Figures
1.1 Web Search Trends in America for Keywords Gauging the Advent
of the Mobile Vending Industry. . . . . . . . . . . . . . . . . . . . . 12
1.2 The official DC Department of Consumer and Regulatory Affairs
(DCRA) designated lottery locations and number of parking spots
allotted. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3 Map of Downtown Washington, DC and Surrounding Neighbor-
hoods. Markers Indicate Food Truck Locations Revealed by Trucks’
Twitter Feeds. The Triangles and Squares make up 86% of the Ob-
servations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.4 The official DC Department of Consumer and Regulatory Affairs
(DCRA) lottery outcomes. . . . . . . . . . . . . . . . . . . . . . . . 21
1.5 Descriptive Truck Heterogeneity . . . . . . . . . . . . . . . . . . . . 24
1.6 Variation in the Estimated Conditional Location Choice Probabili-
ties by Locations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
1.7 Marginal Costs ($) and Price Cost Margins. Recovered from Model
First Order Conditions. . . . . . . . . . . . . . . . . . . . . . . . . . 42
1.8 Model Fit of Location/Genre Specific Expected Utilities . . . . . . 47
1.9 Model Fit at Competition Level (Second Stage of Model). . . . . . 49
1.10 Model Fit of Location Choice Probabilities . . . . . . . . . . . . . . 51
2.1 Current Regime Model Predicted Prices ($) VS. Counterfactual
Regime Prices ($) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.2 Current Regime Model Predicted Profits ($) VS. Counterfactual
Regime Profits ($) . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.3 Box Plots of Estimated Quality (vk) by Cuisine Genre . . . . . . . . 63
2.4 Current Capacity Model Predicted Prices ($) VS Counterfactual
Capacity Prices ($) . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.5 Current Capacity Model Predicted Profits ($) VS Counterfactual
Capacity Profits ($) . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.1 The Metrorail system and SafeTrack Repair Segments . . . . . . . . 79
3.2 6 Month Moving Average of the Sum of Average Ridership At Every
Origin-Destination Pair. and Average system wise delays. . . . . . . 82
3.3 Graphical Interpretation of Treatment 1. Station Pairs where the
Origin Station is Not on the Inner-Washington” Side of the DC, MD,
VA Area and Hence “Affected” (Triangles) are Defined as Treated,
Along with the Stations Directly Affected by the Repairs (Dots). . . 85
vii
3.4 Graphical Interpretation of Regression Coefficients and Confidence
Intervals. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
viii
Introduction
This dissertation covers two topics. Chapter 1 and 2 deals with competition and
parking space allocation in the food truck industry in Washington, DC, and Chap-
ter 3 deals with public transit ridership and the long-run impacts of a large-scale
maintenance program.
An industry with many firms and various dimensions of heterogeneity is difficult
to analyze and consequently poses challenges for the assessment and implementa-
tion of regulatory policies. One such example is a young but booming food truck
industry. In this industry, hundreds of firms serve different types (‘cuisine genre’)
of products with varying quality at many different locations that have capacity
constraints (parking space is limited). Local regulators have responded to this
emerging industry in various ‘ad hoc’ ways. To obtain a better understanding of
the industry and assess counterfactual policies, in Chapter 1, I build a structural
model of the food truck industry that accounts for important dimensions of het-
erogeneity and competition to assess the welfare effects of counterfactual scenarios.
The model is kept tractable by borrowing ideas from the Oblivious Equilibrium
[Benkard et al., 2008] literature. I apply this model to the food truck industry in
Washington, DC and estimate its key parameters. I find that consumers prefer-
ences are only moderately correlated within cuisine genres and that trucks’ own
price elasticity’s are on average about -1.61. Firms price-cost margin is on average
0.66, which closely resembles industry insiders rule of thumb for a truck that can
operate into the foreseeable future. I find that the choice probabilities for lottery
1
locations are higher than that of the non-lottery locations given both are in a
truck’s choices set. In my model, this implies that given the lottery, entry costs
for the lottery locations are lower than that of the non-lottery locations.
In Chapter 2, I employ the results from chapter one to shed light on the impacts
of counterfactual regulatory regimes. Under the current regime the allocation of
parking spaces for food trucks are determined by a lottery for some locations. My
counterfactual simulations of reducing the reach of the lottery to include fewer lo-
cations suggest that the lottery is dampening competition and decreasing the costs
that firms must incur to enter a market location. As a result, I find that it allows
for the survival of firms with lower quality, higher prices, while lowering expected
utility for consumers, and awarding higher profits to firms. The net welfare effect
for this counterfactual scenario is an increase in total daily welfare of $2.294.18.
I also consider a counterfactual scenario where the non-lottery locations increase
their parking capacity by 2 spaces. I find that this is beneficial for both truck
owners and consumers with a net increase in total daily welfare of $8,260.26.
Chapter 3 considers the impact of large public transit maintenance programs
on long-run ridership. An agency that manages large transit systems must make
investments to maintain a level of quality to sustain ridership. If consumers face
switching costs when changing their mode of transport, the large and unavoidable
disruptions to services resulting from a large maintenance program may provide
a sufficient negative utility shock for riders to substitute to alternative modes
of transport and not return after the repairs are completed. In this chapter I
consider such indirect costs that a transportation agency will incur in the context of
Metrorail, the subway system that stretches through the District of Columbia (DC),
2
Maryland (MD), and Virginia (VA) operated by the Washington Metropolitan
Area Transit Authority (WMATA). I find that there has been a persistent drop
in ridership of about 1.68% up to 10 months after the repairs on certain tracks
have been completed. Also, commuters that are originating from VA seemed to
have substituted more heavily away from using Metrorail compared to their MD
counterparts.
3
Chapter 1 Location Choice and Price Competition with
Differentiated Products and Many Firms: An
Application to the Mobile Vending Industry in
Washington, DC.
1.1 Introduction
In this chapter I build a structural model of the food truck industry that accounts
for important dimensions of heterogeneity, market entry, and competition. Sub-
sequently, I apply this model to the food truck industry in Washington, DC and
estimate its key parameters. To begin thinking about this market as an economic
issue, it is useful to observe the following: there are many small firms, who each
truck chooses to serve a well-defined market based on geographical location (for
example, street parking near a prominent park such as Franklin Square) and time
of day (for example, in DC the “lunch break” shift). There are a limited number of
parking spots at the most popular locations and different consumer demographics
and preferences at each location.
Space is a scarce resource in a congested city and it is important for local
regulators to allocate and manage such resources efficiently. For example, a regu-
lator may want to know the welfare impacts of expanding street parking spaces for
popular food truck lunch locations or they may want to consider how the current
regulatory policy is impacting the industry relative to other policies. To be able
to consider these counterfactual scenarios one needs a tractable but sufficiently
4
realistic model.
Many industries are comprised of a large number of small firms that are het-
erogeneous in various dimensions. These industries are difficult to model and
analyze due to the rapid increase in computational burden for solving equilibria
with many firms and because most of the theoretical literature considers stylized
duopoly models. As a consequence most structural empirical industrial questions
center around oligopolistic industries inhabited by a handful of firms.1 Given these
modelling challenges, the assessment and design of regulatory regimes in these in-
dustries are difficult. For example, in the US food truck industry within a city,
there are hundreds of firms selling products of varying quality and type who must
choose the location to enter and compete in every day. Also, given entry to a
location, each type of firm may have different competitive effects on each other
implying the existence of informational and allocative externalities depending on
where a firm decides to operate in and the market structure that ensues from these
decisions.
The difficulty in modelling this environment comes from the fact there are
many heterogeneous firms. Outside of the food truck industry, any setting where
a firm must secure the right to operate and then compete with other firms who
have also secured this right will be similar. For example, the allocation of types
of shops in a newly developed mall, the allocation of bundles of licenses. In this
paper I build and estimate a model that attempts to account for these important
features of the food truck industry while maintaining tractability. The question
1In a dynamic context, Berry and Pakes [1993] explore the curse of dimensionality and how
quickly computational burden increases as the number of firms increases. My model is static;
however, it captures other dimensions of heterogeneity which increases computational burden
greatly if trying to solve for a standard Nash equilibrium.
5
of designing a parking location allocation mechanism for the industry is interest-
ing. The multiple dimensions of firm and market heterogeneity, and the various
externalities imply a simple auction may not be the ideal mechanism for this allo-
cation problem. However, these interesting features are also the road blocks with
respect to methodology and theory. With this in mind, the goal of this chapter
is not to find and suggest some type of theoretically optimal mechanism, but to
assess and quantify how competing regimes perform. One counterfactual regime I
consider reduces the reach of the currently employed lottery system by 2 locations.
The second counterfactual regime increases non-lottery locations’ parking capacity.
My model of the food truck industry has buyers with nested logit utilities,
where the nests are formed around cuisine type, which is a dimension of hetero-
geneity in the firms. This modelling choice allows seller’s heterogeneity to impact
demand in a meaningful and tractable way. The sellers (trucks) play a two-stage
game where in the first stage given a conjecture about the state of each market
location, they choose prices and choose a location to enter where they incur a entry
cost that is specific to the location. Then in the second stage the trucks explicitly
compete with other firms that have entered the particular market. This ordering of
the seller’s decision where he chooses price then location, is novel in the literature
but is reasonable in my context. It is easily observed that the truck’s prices do not
vary over time but their locations change daily.2 I find that removing a key food
truck location (L’Enfant from the lottery increases daily total welfare by $2,294.18
and that increasing the parking space capacity of current non-lottery locations by
2 spaces increases daily total welfare by $8,260.26
2In Schmalensee [1992] the author briefly discusses motivations for the order in which firm
choices are made. It is argued that the dynamic differences between price choice and product
characteristic choice are not clear, and there is a tradition to think that maybe price is chosen
before other attributes.
6
In Chapter 1, Section 1.2 will discuss where my work fits in the existing litera-
ture. Section 1.3 offers a description of institutional background for the food truck
industry in Washington, DC and describes the data used to estimate the model.
Section 1.4 will develop the structural model and provide an exposition of the equi-
librium. Section 1.5 interprets and discusses the estimation stages and estimation
results. Section 1.5 assesses the model’s performance by comparing the model’s
equilibrium solution outcomes and their estimated counterparts.
1.2 Related Literature
There is a rich body of literature that considers the endogeneity of market struc-
ture. Classic examples include Berry [1992] and Bresnahan and Reiss [1990]. We
learn from these seminal contributions that when a firm decides to enter a market
it is a strategic decision. Favorable states of a market will increase profits, but also
attract more entrants, hence a firm must evaluate these trade-offs before choosing
to enter himself. When it comes to empirically estimating these types of models
the researcher encounters a number of issues due to the large number of configu-
rations that must be checked to find whether a distribution of players constitute
an equilibrium. Such an example is Mazzeo [2002], where model of endogenous
product type choice and market structure is considered under the assumption of
complete-information. Mazzeo [2002] shows that even with two or three product
types, the number of profit inequality constraints that must hold is large and is
quite burdensome to estimate.
A paper that addresses these difficulties is Seim [2006]. Seim [2006] considers
the product characteristic choice decision of firms and applies this framework to
7
the video rental industry. In this industry the product offered itself is relatively
homogenous, but location is a major source of product differentiation and Seim
considers this dimension as an endogenous choice by the firms. However, in Seim
[2006] demand and explicit competition is not modeled, and a reduced form ap-
proach is taken to estimate a firm’s entry probability. For me to estimate the dollar
value of location entry costs and industry welfare I model entry and price com-
petition. There is little work in the literature that models entry and also explicit
competition after entry. For example, Suzuki [2010] models hotel chain entry de-
cisions in a dynamic setting, but again doesn’t model competition explicitly. The
difficulties in modelling competition explicitly come from the lack of data and/or
the computational burden in solving for the equilibrium with many firms.
I overcome these difficulties in modelling the firm’s entry decision and the com-
petition following entry to a location by using self-collected data and by using a
simulation based iterative solution method. My model maintains tractability by
assuming that the average of the realized distribution of firms in the market is
equal to the firms’ long-run conjecture of the state of the market, and I with this
assumption can compute the equilibrium distribution of firm types, quality, and
number of operational firms within the modelled locations. This key assumption
is behaviorally plausible as it is reasonable to argue that individual small firms
use their experiences and observations for a market realization to come up with a
expected long run state of the market and is similar to the assumptions made for
oblivious equilibrium in Benkard et al. [2008].3
One method of dealing with the burden of solving for a complete-information
3For more recent papers using moment-based equilibrium concepts see Xu [2008], Qi [2013],
Saeedi [2014], Sweeting [2015].
8
Nash equilibrium is to assuming incomplete-information to effectively render all po-
tential entrants to be homogenous as in Seim [2006]. This makes a model tractable,
but in the process interesting economic insights related to firm heterogeneity and
types may be lost. My model will allow the trucks to be of different types and
qualities. Allowing for firm types, while considering the entry decisions for firms
allows me to look at the equilibrium distribution of firm types that prevails in
equilibrium. This is important because in many contexts, it is not just the num-
ber of competing firms that matter in an entry decision, but the number of firms
that are more directly in competition to your type. This intuition is related to
Wollmann et al. [2014] where the author considers the importance of product po-
sitioning in the context of commercial vehicles. Wollmann et al. [2014] finds that
product entry has a dramatic impact on prices and purchases. In my context, this
finding emphasizes the importance for researchers to allow the firms not to only
choose whether to enter a market or not, but allowing them to strategically choose
which market to enter given the state of each market as it appears to your own type.
The food truck industry in Washington, DC has been examined by Anenberg
and Kung [2015]. The authors explore the impact of information technology on
food truck growth and suggest that an advantage of food trucks over traditional
brick and mortar establishments is that they can use mobility to capitalize on the
consumer’s taste-for-variety. Obtaining data from DC food trucks’ Twitter feed for
the trucks’ location choice, a logit model is estimated by defining a reduced form
profit function. Their results indicate that there is a negative impact of a truck
visiting the same location in short succession. Relying on strong assumptions on
industry total revenues, they estimate that the loss of choosing the same location
two days in a row results in a $257 loss in the day’s profit, which is about 38% of
9
average daily profits.
However, such reduced form analysis does not provide an opportunity for un-
derstanding the underlying parameters that dictate the consumer’s and vendor’s
market behavior and hence is limited in scope when considering policy simulations
and welfare analysis. Also, Anenberg and Kung [2015] does not deal with the cost
arising from the fact that at each location, parking spots are scarce, which is an
important factor when considering the industry from a regulation and policy per-
spective.
1.3 Industry Details and Data
1.3.1 The Food Truck Industry
Food trucks are very common in large cities and each truck requires a parking spot
to operate. In these cities space is a scarce resource that needs to be carefully
managed. With relatively low start-up costs, mobile vending as opposed to tradi-
tional brick-and-mortar setups buy the entrepreneurs an opportunity to be more
experimental with their products. This coupled with an increase in the import and
export of culinary culture in the last decade, has led to the rapid penetration of
food truck food as standard meals and an increase in variety of the types of cuisine
offered. Accurate numbers depicting the growth of culinary experimentation and
the mobile vending industry are hard to come by, however a report by Mountain
View-based financial software company, Intuit [2012], forecasts that the “rolling
restaurants” are on track to be a $2.7 billion national industry in 2017 and its
market share for meals to jump to 3 or 4 percentage points in the next five years.
10
This is particularly interesting when you consider the market research company,
NPD Group’s [2016] finding that for weekday casual dining and fast casual restau-
rants, food service lunch visits declined by 7% in the quarter ending June 2016
compared to same quarter the year before.
Simple Google trend searches support this phenomenon. In Figure 1.1 I show
web search trends for three keywords; donburi, kimchi and food trucks. It can
clearly be seen that the interests in foreign food items have increased noticeably
in the last decade, particularly since 2010. With interests in food trucks following
the pattern closely.4 In particular, this effect is likely to be much stronger in large
cities, where the growth of the food truck industries are concentrated. I can’t pin
down exactly why the mobile vending industry saw such an explosive growth and
this paper does not attempt to explain this. I will be focusing on the issues that
policy makers must consider now that this industry exists.
4The seasonal component of the interest in food trucks in panel (c) seem to be due to the fact
that warmer temperatures that allow a customer to eat outside garner more interest for food
trucks.
11
(a) Interest in Donburi over time. (b) Interest in Kimchi over time.
(c) Interest in Food Trucks over time.
Figure 1.1: Web Search Trends in America for Keywords Gauging the Advent of
the Mobile Vending Industry.
These policy issues range from hygiene and quality regulation to parking, traffic
and licensing regulations. Out of the broad range of issues, this project will specifi-
cally look at parking allocation at various locations where food trucks agglomerate
at and the competition the food trucks engage in given a tendency for consumers
to have preferences over different cuisine genres as well as the overall quality of
the meals offered. Evidence of policy maker’s concerns can be seen in DC’s food
truck industry. The DC Department of Consumer and Regulatory Affairs (DCRA)
governs the licensing and regulation of mobile vendors in DC. Starting from De-
cember 2013 the DCRA has implemented a monthly lottery5 system that randomly
allocates vendors who register to enter the lottery to a predetermined list of pop-
ular locations. This system was the answer to the widespread problem of trucks
showing up extremely early in the morning to secure the best spots, consequently
congesting the area and inducing many parking violations and pedestrian safety
5The exact lottery mechanism is proprietary so I condition my analysis on the outcome of the
monthly lotteries.
12
concerns. The officially designated lottery locations can be seen in Figure 1.2.6
Figure 1.2: The official DC Department of Consumer and Regulatory Affairs
(DCRA) designated lottery locations and number of parking spots allotted.
Although a simple lottery is easy to implement and reduces the various entry
6The exact page can be found at https://eservices.dcra.dc.gov/VendingLottery/MRVLocations.pdf
13
costs that trucks need to incur to secure a popular spot, it may not be the ideal
mechanism for allocating these scarce parking locations. For example on the 8th
of November, 2016, at Union Station, there were 6 Asian trucks out of 11 trucks
which may indicate a lack of variety. The lottery can generate situations where
there is not an ideal level of variety at different locations, and at these locations
trucks could trade spots for a Pareto improvement. An extreme example would
be, assuming consumers like variety of cuisine genres, if there is no variety with
only Asian trucks at Franklin Square and no variety with only American trucks
at L’Enfant, it would be optimal to change the allocation of the trucks to reflect
more variety at both locations. Also, the marginal value of variety maybe greater
at L’Enfant where there are little outside options for lunch. Another dimension
that must be considered in this allocation problem is truck quality, which opens
up opportunities for more sophisticated mechanisms (for example, auctions) that
may correctly price the value of a scarce spot for different trucks.
A simple motivating example can shed some light onto the subtleties involved
in modelling competition and regulatory policies when consumers have preferences
that are nested within a particular cuisine genre category (for example, Asian cui-
sine and Mediterranean cuisine). Consider a parking lot with 2 food trucks and
no outside option for lunch. These two trucks are of different genres but have
exactly the same quality and suppose that the consumer’s preferences exhibit no
correlations between similar genres. We can model this with a simple logit model
and these two trucks will share the market evenly in equilibrium.
Under these circumstances, if we add a third truck to this location of the same
quality, regardless of genre, the third truck will take equal market share from the
14
incumbent trucks ending up with a third of the market. However, if consumer’s
preferences are correlated within genres this will not be the case. Consider now a
nested logit model, where consumers have preferences with nests formed around
cuisine genres. Now the genre of the third truck impacts the incumbents and the
entrant differently. In particular, both incumbent trucks would prefer the genre of
the third truck to be different from their own, as then they will lose less market
share compared to if the third truck was of the same genre. In other words, as the
correlation within the nests become stronger, trucks and consumers want “more”
variety in the market place. Depending on the correlation of preferences within
cuisine genres, the optimal distribution of quality and types of trucks in the mar-
kets should be different. This simple example shows that when we are allocating
scarce parking spaces to vendors, we must take into account these trade-offs to
correctly measure and analyze welfare.
Allocative externalities where the value of entering a location for a truck de-
pends on the identity of who is going to be at the location makes the setting
theoretically intractable and characterizing and designing an optimal mechanism
in this setting is not the scope of this paper. Knowing these effects exist in the
market, I am attempting to build a model that captures these effects to assess
the current policy and the impact of counterfactual scenarios that need not be
theoretically optimal.
1.3.2 Data
The estimation of the model utilized data from multiple sources. Most of the de-
mand data used to estimate the nested logit demand model is manually collected
and processed in two stages. Firstly, to efficiently collect market share data I
15
investigated the proportions of consumer arrivals at a food truck location during
different times of the consumer’s lunch break. Secondly, using these time dependent
proportions I scaled up the quantity count for each truck (number of consumers
served) during the specific time interval by this proportion to obtain the “total
lunch shift” market shares for each individual truck and the overall food truck
segment market share by comparing it with the number of primary jobs in the
area obtained from the publicly available Census Longitudinal Origin-Destination
Employment Statistics (LODES) data set which is derived from the Longitudinal
Employer-Household Dynamics (LEHD) data set.
In the process of collecting the arrival proportions I visited a total of 10 food
truck locations (lottery designated and non-lottery) from 11am to 1 35pm and
counted the stock of consumers in line at a truck at the location approximately
every 10 minutes. Then assuming an average departure rate of 2 customers per
minute per truck7 I backed out the number of new arrivals and computed the
proportion of arrivals that happened during each 10 minute interval. Effectively,
this is a probability distribution of the consumers arriving to buy lunch during
the aforementioned time frame. I conducted a Kolmogorov–Smirnov test on these
distributions to assess if any statistical differences exist between every pair and
find that any differences between all 10 of the distributions that I counted were
statistically insignificant. There are limitations to this data, as it is manually col-
lected and the sample size is limited, but it has helped me to collect market share
data from multiple locations each day.
The resulting market share data consists of 30 location observations from 12
7This number was obtained from conversations with truck owners and I believe it is reasonable
to believe that the rate at which a truck serves its customers are not time varying during the
shift.
16
Figure 1.3: Map of Downtown Washington, DC and Surrounding Neighborhoods.
Markers Indicate Food Truck Locations Revealed by Trucks’ Twitter Feeds. The
Triangles and Squares make up 86% of the Observations.
17
different locations, of which 9 are lottery and 3 are non-lottery and a total of 331
market share observations from 154 unique trucks (shown in Table 1.1). Location
specific data such as the market size and mean earnings from the LODES data
set8 and truck specific data such as the number of Twitter Followers and their
Yelp reviews are used to supplement the demand data. Figure 1.3 shows the food
truck lunch locations that are observed in the Twitter data colored by lottery (all
modelled), non-lottery but modelled, and not modelled.
I do not observe the average price of each order ticket for each truck, so we must
determine what price to use when we estimate the demand model. For a subset of
the data, I asked the truck owners at the end of the shift what the average price
of their sales item on that day was. For these trucks I use these prices, for others
I make an assumption. For example, I assume that for kabob trucks, that the
menu item of interest is the average price of their pita sandwiches, for taco trucks
I assume that a meal is buying 3 tacos, and for rice bowl trucks I use the median
price of their entrees. If revenue data was available the measure of average price
could be improved, but this was not available. I believe this method is a second
best to thinking about the prices of each vendor. Tables 1.1 and 1.2 show some
summary statistics from the demand data.
8Using the LODES OnTheMap web application (https://onthemap.ces.census.gov/) I drew a
0.15 mile radius circle and defined it as the “market”.
18
Utilized
Locations Lottery Mean Monthly Earnings ($) Market Size
19th & L No 7,266.95 24,583
CNN (First St NE) No 6,725.52 16,365
Farragut Square Yes 7,439.35 23,779
Federal Center / Patriots Plaza Yes 7,149.01 3,659
Franklin Square Yes 7,996.54 14,317
Gallery Place / Chinatown No 6,511.07 10,783
L’Enfant Yes 6,974.15 16,901
Metro Center Yes 7,817.77 23,020
Navy Yard Yes 9,257.72 2,898
NoMa / New York Ave Metro Yes 7,128.24 6,074
State Department Yes 6,550.04 9,358
Union Station Yes 6,512.05 14,124
Note:The Market Size is the total number of primary workers.
Table 1.1: Location Characteristics from the LODES data
Variable Mean Std. Dev. Min. Max.
Truck Characteristics
Quantity 82.98 38.87 13 191
Price 9.48 1.47 7 16
Twitter Followers 1,385.20 2,228.03 0 14,000
Yelp Review 3.70 0.77 1 5
Market Share 0.0064 0.0062 0.0006 0.0614
Location Level Characteristics
Total Consumers Served 1,035.55 346.445 272 1717
Potential Market 15,705.13 4,856.113 2,898 24,583
Outside Option Share 0.93 0.03 0.81 0.98
Mean Daily Earnings 237.74 19.25 217.04 308.59
Note: Outside Option Share is computed as no. of primary jobs−total consumers servedno. of primary jobs
Table 1.2: Summary Statistic from the Demand Data
Another data set that was compiled and used in the estimation of the model is
Twitter data scraped from the DC food truck location aggregator www.foodtruckfiesta.com.
This data set contains individual food trucks location choices that have been an-
nounced via the truck’s Twitter account. In this data set I observe more locations
19
than the ones I have collected demand data from and have included in my model.
The locations that I model account for approximately 86% of the Twitter data
location choices. The locations I have modelled were chosen to include all of the
lottery, and comparable key non-lottery locations. This means that the locations
that are not modelled are fringe DC locations and also university-oriented locations
such as George Washington University (Foggy Bottom) and Georgetown Univer-
sity (Georgetown). Due to the fact that the researcher does not observe exactly
how the lottery is conducted it is impossible to simulate over different lottery out-
comes, so my model takes the lottery outcomes as given (i.e. the trucks choice sets
are fixed when the trucks are making their location choices). This implies that
to complete this Twitter data, I must construct the choice set for each truck. I
do this by matching the monthly DCRA lottery outcomes and the Twitter data
over 4 months spanning July to October, 2017. This was a combination of a fuzzy
string matching problem9 and deducing the identity of trucks by matching where
a truck has tweeted to be and where the lottery outcome suggests a truck to be.
I have managed to match about 70% of all the trucks that appear in the Twitter
data and these trucks supplemented with the location and truck characteristics
from the above-mentioned data sources make up the final Twitter data set.
9This exercise involves scoring each pair of truck names in the Twitter data and the lottery
outcomes based on similarity of their string names. Once I obtained these scores I investigated
each Twitter data truck’s top 3 similar lottery data truck names to find the correct match.
20
Figure 1.4: The official DC Department of Consumer and Regulatory Affairs
(DCRA) lottery outcomes.
The page from the actual lottery outcome of the month of June, 2017 available
at DCRA’s vending web page10 is shown in Figure 1.4. It lists the site permit
number, the name of the business, and the markets business is allowed to operate
in on each day of the week. To register to enter and operate at these locations there
is a total fixed cost of $175/month which is the sum of a $25 entry fee and a $150
location site permit fee [DCRA, 2013]. A truck can choose to not pay the permit
fee and forgo its allocation, also trucks may trade their location on a given lottery
draw with another truck if the DCRA approves the trade. This may explain some
10https://dcra.dc.gov/mrv
21
of the misfit of my model to the data as the rejection of the allocation post lottery
entry and permit swaps are not modelled. My model assumes that the lottery
outcomes we see are strictly adhered too and there is no convincing evidence that
leads me to believe that a significant portion of the location allocations are traded
or rejected. I observe most of the trucks that are supposed to be at the location in-
ferring from the lottery to be at the location once I arrive to collect the data. From
my demand data I can identify 68% of the trucks I actually observe are supposed
to be there according to the lottery. Also, the non-adherence is typically not in the
form of a different truck being at the location but a designated truck not showing
up (i.e. location has fewer trucks than the lottery allocation suggests) implying
that permit trades don’t compose a significant share of my observations. There is a
total of 139 designated lottery spots across the 9 locations and various measures of
the total number of trucks operating in Washington, DC suggests there are many
more trucks in total. The food truck location aggregator www.foodtruckfiesta.com
lists 245 trucks listed as being permitted to operate in DC. In the lottery outcomes
posted by the DCRA, I consistently count more than 220 individual site permits
allocated through the week.
Given that I don’t observe the whole industry (due to lack of man power and
data collecting resources) I would want the Twitter data and the demand data to
both be representative of the whole industry. Comparing distribution of cuisine
genres which I have divided into 7 categories (dessert trucks have been dropped
from the analysis), American, Asian, Caribbean, Exotic, Indian, Latin American
and Mediterranean, the observations from the Twitter data and the demand data
seem to represent quite a similar sample of the population as shown in Table 1.3.
One exception is in the Indian genre which tends to have less online presence than
22
other genres, hence also Tweeting less.
Demand Data Twitter Data
Genre Freq. % Freq. %
Asian 60 18.13 516 16.14
American 118 35.65 1,073 33.55
Mediterranean 64 19.34 678 21.20
Latin American 38 11.48 387 12.10
Indian 23 6.95 66 2.06
Caribbean 20 6.04 223 6.97
Exotic 8 2.42 67 2.10
Total 331 3,198
Table 1.3: Comparison of the distribution of cuisine genres within each data set
Comparing the two data sets across the number of Twitter followers and Yelp
reviews, the Twitter data seems to be capturing trucks that have on average more
followers and higher Yelp reviews. This may be due to the fact that it is more
likely that a truck with more online presence is more likely to be broadcasting
their daily location through Twitter.
Truck heterogeneity and Quality
Trucks differ in both the quality of the product offered and their cuisine type. The
paper focuses on taking into account these dimensions of heterogeneity in assessing
regulatory policy. Figure 1.5 describes these features in the demand data. From
Table 1.3 we can see that there are overwhelmingly more American genre trucks
which are trucks that serve burgers, sandwiches, pizzas, Barbecue, etc. Mediter-
ranean and Asian trucks are a distant second and third. Now considering this
23
together with the information in Figure 1.5, we can see that while there are an
abundant number of Mediterranean trucks, they don’t seem to be very active on
social media (less Twitter followers on average) and also seem to have a wider
distribution of Yelp reviews compared to the Asian trucks. This kind of hetero-
geneity will imply heterogeneous impacts given a change in regulatory regime. In
particular, we could expect that the low quality metric genres like Indian and
Mediterranean trucks along with abundant genre trucks like American trucks to
be affected most heavily by a regime that does away with the lottery.
(a) Number of Twitter followers by Cuisine Genre (b) Yelp Review Stars by Genre
(Accounts with followers above 5000 excluded)
Figure 1.5: Descriptive Truck Heterogeneity
1.4 Model
1.4.1 Demand
Suppose a location l ∈ {1, ..., L} has entry cost cl, market size Ml, capacity con-
straint µl, and mean income Ȳl. g ∈ {0, .., G} denotes different cuisine types and
g = 0 denotes the outside option. j ∈ {1, ..., J} is an individual truck with quality
to consumers vj and marginal cost mcj. Hence, a triplet (vj, g(j),mcj) defines an
individual vendor and these parameters are all exogenous. The model does not
24
consider cuisine genre choice of the food truck entrepreneur. Finally, let Γg(j)l de-
note the set of all trucks at location l with the same genre as truck j.
Given the firms’ prices and location outcomes (which will be described below),
consumers at the location have nested logit demand with nests formed around the
genre. This implies that the indirect utility of consumer i eating in location l,
consuming truck j’s product is:
Ulij = vj + α ln(Ȳl − pj) + ξlj + ζig(j) + (1− σ)ij (1.1)
and the indirect utility of consuming the outside good is:
Uli0 = α ln(Ȳl) + ζi0(j) + (1− σ)i0 (1.2)
Where  is i.i.d. Type I extreme value across products, ζig(j) is the group spe-
cific taste of the consumer, and σ measures the relative weight of idiosyncratic and
group preferences. Empirically, vj will be parameterized by the truck’s charac-
teristics like Yelp reviews and the number of Twitter followers. Note that in this
chapter including an income effect is important for two reasons. Firstly, it adds
more variation in the variable such that I can estimate α. Secondly, buying lunch
for consumers is a daily decision and is plausibly budgeted given the consumer’s
income.
25
The above setup implies location market shares:
( )( ( ))−σ
vj+α ln(Ȳ −p ) ∑exp l j v∑ (∑ ( exp k
+α ln(Ȳl−pk)
1−σ k∈Γg(j)l 1−
Sjl = ))σ1−σ (1.3)
G exp vk+α ln(Ȳl−pk)g=0 k∈Γgl 1−σ
1.4.2 Trucks
The Truck’s expected profit11 maximization is as follows. Firstly, he chooses a
price, and then given this price and vendor’s conjectures of the cuisine genre and
location specific market shares, he selects a location to enter which I model as a
standard logit discrete choice problem with zero profits if the choice is to not enter
any locations. In other words, profits of vendor j at location l are:
( ) ( )
Πjl pj, p−j, vj,mcj, g(j),Ml, Ȳl = (pj −mcj)S pj, p−j, vj, g(j), Ȳl, σ Ml
− cl + λ(jl − j0) (1.4)
The λ is the scaling parameter12 and the  are Type I Extreme Value. This
ordering of events, where price choice happens before location choice is an inno-
vation to the literature. An additional exogenous state of the model is the set
of locations that each truck can choose from and the number of parking spots at
each location. This location choice set is governed by the policy regime and the
number of parking spots by physical space. In the data, this is the DCRA lottery
outcome and respective parking space allotments as seen previously in Figure 1.2
11Expectations are taken over the truck’s own location choice probabilities and rival’s loca-
tion probabilities through some equilibrium conjecture about cuisine genres and location specific
market shares which will be introduced below.
12Since the “utility” of the truck is in actual dollars, the scale is important here
26
and the number of street parking spaces at the locations on non-lottery locations.
Backwards induction implies the following maximization problem for truck j where
Lj denotes truck j’s location choice set:
∑ ( )Π̂jl(pj ,p−j ,vj ,mcj ,g(j),Ml,Ȳl,cl,σ( )exp∑ λmax ) ...pj Π̂jl(pj ,p−j ,vj ,mcj ,g(j),Ml,Ȳl,cl,σ∈ )l Lj 1 + l∈L exp( j λ )
... ∗ Π̂jl pj, p−j, vj,mcj, g(j),Ml, Ȳl, cl, σ (1.5)
s.t. E[nl|c1, ..., cL] ≤ µl (1.6)
E[nl|c1, ..., cL] denotes the expected number of trucks at location l given the
entry costs of the locations, note that the entry costs are the parameters which in
equilibrium keep the number of trucks at a location feasible. More specifically, the
entry costs affect the location choice probability and the average location choice
probability determines the expected number of trucks at the location. An intu-
itive motivation for the entry costs is that it reflects the value of an additional spot
added to a location. The exogenous µl is effectively the fixed supply of parking
spots, while E[nl|c1, ..., cL] is the demand for parking spots in the market. The
difference between these two variables can be thought of as the excess demand
for parking at a location, and the vector of cl i.e c will be such that the market
in equilibrium clears. With this motivation, the interpretation of the entry costs
become clear and practical. For example, if c1 = 100, this would imply that a
truck is willing to pay up to $100 for an additional free parking spot at location
1. It is the fixed costs of securing a spot at a location relative to entering a fringe
location where capacity doesn’t matter. An alternative way to think about the
27
entry costs is to think of them as some kind of congestion cost arising out of excess
demand. Too many trucks flock to one location inducing costs to the trucks if they
are to secure a spot and do business. The entry costs in equilibrium will settle
on a value such that a feasible number of trucks enter each location in expectation.
Π̂ denotes expected profits at each location. To calculate this entity, you need
to know the probability density for the numerous market configuration outcomes
that can occur. This is intractable with over 200 vendors operating in the DC area,
at many locations, several cuisine genres, and differing qualities. I assume that the
exact price and location choices of the other truck vendors are not observed by
the trucks, however an equilibrium conjecture of the ‘inclusive values’ provides a
sufficient statistic that reflects the number of rival vendors and their genres, rival’s
prices and values. This assumption renders the problem tractable.
1.4.3 Equilibrium
Definition 1. An equilibrium is a G× L matrix of conjectures of Igl denoted Îgl
where: ∑ ( )vk + α ln( Ȳl−pk )Ȳ
Igl = ln exp l 
1− σ
k∈Γgl
and,
∑S
Î = s=1
Igls
gl
S
28
where S is the number of simulated outcomes given a) a set of prices (J × 1) such
that:
∑ ( )Π̂jl(pj ,p−j ,vj ,mcj ,g(j),Ml,Ȳl,cl,σ)exp
max
pj ∑ ( λ ) ...Π̂jl(pj ,p−j ,vj ,mcj ,g(j),Ml,Ȳl,cl,σ∈ )l Lj 1 + exp
( l∈Lj λ )
... ∗ Π̂jl pj, p−j, vj,mcj, g(j),Ml, Ȳl, cl, σ
and is solved and b) entry costs (L× 1) such that location choice probabilities
simulate outcomes which satisfy:
E[nl|c1, ..., cL] ≤ µl
Specifically, the assumptions that I make is that the commonly perceived values
of the inclusive values are formed as the average of actual outcomes of these inclu-
sive values. Each individual firm does not perceive its own impact on the Îgl and
the restriction of the equilibrium concept to focus on the average of the inclusive
values means that the market share (Ŝgl) is an approximation of the nested logit
shares. This greatly simplifies computation. I forward simulate to compute each
Îgl.
The equilibrium is computed with the following iterative process:
On iter1 = 1 and iter2 = 0:
1. Guess a matrix Î iter2=0 of size G × L and vector citer2=0 of size L × 1. Set
piter2=0=0 and set up a grid of v and g.
29
2. Draw S simulations from the random events of the model (i.e. Realizations
of the trucks’ logit location errors), and set tolerances c, I , and p to assess
convergence. Update iter2 = 1.
Then for iter2 = m > 0.
3. Set iter1 = 1 , set Î iter2=m−1 = Î iter1=1 and citer2=m−1 = citer1=1. Now given
Î iter1=1 and citer1=1 solve for the profit maximizing prices (piter2=m) for the
firm using Ŝ, on a grid of genres (g), qualities (v), and marginal costs (mc).
Then for iter1 = k > 0
(a) Given∑Î iter1=k and piter2=m, solve the minimization problem
min Lc l=1 (E[nl|c1, ..., cL]− µ )
2
l . Forward simulate using the S simula-
tion draws for obtaining E[nl|c1, ..., cL]. Denote the solution citer1=k+1
(b) Given the solution citer1=k+1 compute Î iter1=k+1 by forward simulating
S times and taking the average over the simulated Is.
(c) Compute the distance.13 between Î iter1=k+1 and Î iter1=k. Compute the
distance between citer1=k+1 and citer1=k.
(d) If the distances in the previous step are less than I and c respectively,
stop and save (citer1=k+1 = citer2=m and Î iter1=k+1 = Î iter2=m). Other-
wise return to the nested step a) with the updated inclusive values and
costs (citer1=k = citer1=k+1 and Î iter1=k = Î iter1=k+1)
4. Compute the distance between piter2=m−1 and piter2=m, citer2=m−1 and citer2=m,
and Î iter1=m−1 and Î iter1=m).
5. If the distances computed in 4. is less than p, c, I respectively, stop.
13Note that one may choose different updating rules or distance measures for the iteration.
I update the relevant variables in each iteration such that the latest iterations are a weighted
average of the current iterations values and the past iterations values. I do this such that the
model outcomes are not too volatile between iterations.
30
Otherwise update price vector (piter2=m = piter2=m−1). Update the index
iter2 = m+ 1 and return to step 3.
With 224 trucks, 7 cuisine genres, 13 (12 actual locations and 1 outside option)
locations, tolerances set at 0.0001 for prices, entry costs, and the inclusive values
solving the model takes approximately 17 minutes to simulate and solve.
1.5 Estimation
Estimation of the model will proceed in multiple stages:
1. Estimate demand parameters (α, β, σ) using demand data.
2. Estimate a reduced form conditional logit model of location choice using the
Twitter data.
3. Using the estimates from Stage 1 and 2 and the trucks first order conditions
back-out the marginal costs (J × 1 vector mc) for a set of observed trucks
and lottery outcomes.
4. Match model moments derived from estimates in stage 1 and stage 3 to
moments generated from stage 2 to obtain estimates for entry costs and the
scaling parameter (L× 1 vector c, λ).
I will consider each stage in more detail in the following subsections.
1.5.1 Demand
I estimate demand using the realized observations after the firms have entered
a location assuming that this is the equilibrium outcome given the entry game,
more heterogeneity in the consumers was not modelled or estimated, for example,
in the form of random coefficients because the data I am using to estimate the
demand model doesn’t allow me to observe the empirical distribution of consumer
31
characteristics.14 Due to Berry [1994] we know there is a simple inversion for the
nested logit model. Performing the inversion on the indirect utility formulation
outlined in the previous section, the demand model that I estimate is the following:
ln(sjl)−
Ȳl − pj
ln(s0l) = Xjβ − α ln( ) + Ll +Gj + σ ln sj,g(j) + ξjl (1.7)
Ȳl
Xj denotes truck characteristics such as Yelp reviews and the number of Twit-
ter followers. Ȳl denotes the location specific average daily earning and pj is vendor
j’s price as before. Ll is a location dummy variable, Gj is a genre dummy variable
and sj,g(j) denotes vendor j’s share within his genre nest. The endogeneity of price
is a general issue in demand estimation but on top of this I have the issue of there
being very little price variation between and within food trucks. I alleviate these
issues by including the difference between mean daily worker (consumer) earnings
for a certain location and the truck’s price in the utility function and to identify
this earnings-price coefficient. The intuition is that I will compare the relative
market shares of two trucks that are observed to compete in two different loca-
tions that have different mean earnings. This strategy adds variation and also
somewhat deals with the endogeneity of price, as we are looking at the change in
relative market share between two trucks and it is reasonable to assume that the
quality differential between these two trucks shouldn’t change with mean earnings.
I try a variety of specifications but find that including truck specific fixed effects
give me the most favorable estimates which I choose to use (specification (1) in
Table 1.4).15 The reason for this is as Nevo [2001] describes. I observe the same
14For example, income data from the LODES data set is top coded such that there is no
meaningful variance data do be used.
15Robustness checks including day of the week dummies, controlling for the weather don’t seem
to make a significant difference on the key parameters of interest.
32
truck in multiple locations I can include these fixed effects and because the truck
specific unobservables are not changing as the location of the truck changes the
fixed effect accounts for these unobservables that may be correlated with price.
Subsequently I project these fixed effect estimates on to the observed truck spe-
cific characteristics to obtain estimates for β.
In order to estimate the nesting parameter, I must instrument for the within
group market share. I deal with this by using the DCRA lottery outcome. This
source of randomization of vendors to locations gives me an exogenous number for
how many vendors of a certain cuisine genre may be present at a given location.
In my estimation, I have found this to be a good instrument with first stage
F − stat = 27.27. Staiger and Stock [1997] suggest that instruments be declared
weak if the first-stage F − stat < 10 and the first stage in my estimation clearly
passes this rule of thumb. The demand estimates are shown in Table 1.4 and 1.5.
33
(1) (2) (3)
FE 2SLS 2SLS
ln sj,g(j) 0.241 0.158 0.195
(0.178) (0.127) (0.124)
Ȳ
ln( l
−pj ) 33.16 23.44*** 21.87***
Ȳl
(20.52) (4.457) (4.112)
Twitter followers 4.33e-05*** 4.82e-05***
(1.15e-05) (1.08e-05)
Yelpstars 0.117***
(0.0292)
Constant -4.397*** -5.047*** -5.073***
(0.729) (0.323) (0.252)
Genre FE No Yes Yes
Yelp FE No Yes No
Location FE Yes Yes Yes
Truck FE Yes No No
Observations 331 305 305
R-squared 0.865 0.673 0.681
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Table 1.4: Nested Logit Estimation Specifications
Table 1.5 shows the full specification including the location fixed effects. The
location fixed effect coefficients suggest that Navy Yard, Federal Center, and State
Department are the top three highest demand locations, which reflects the lack of
outside options at these locations. Also, the fixed effects are driven by my assump-
tion of the market size (i.e. using the number of primary jobs at each location for
market size). For example, the high fixed effects in small markets such as Navy
Yard may be due to the specified total market being very small, implying that the
food truck segment of the market will be large. Table 1.6 shows the estimates of
the projection of the fixed effect estimates on individual characteristics such as the
number of Twitter followers, Yelp stars and genre. I will use these estimates in my
34
counterfactual simulations and in the process of backing out the marginal costs.
The average price elasticity implied by the point estimates of α and σ is -1.61
(standard error = 0.7351). The nesting coefficient is estimated to be 0.241 and is
feasible and consistent with utility maximization (i.e. between 0 and 1) however
it is not significantly different from zero (p-value=0.116 ). I suspect with more
observations (especially with the truck level fixed effects) these estimates will have
tighter confidence intervals, but I proceed with these point estimates to the next
stages of estimation. Other observations that can be made is that the Asian trucks
tend to on average have larger market shares than any other genres. The signs on
the number of Twitter followers and Yelp review stars are positive as expected.
35
Preferred
Specification
ln sj,g(j) 0.241
(0.178)
Ȳ −p
ln( l j ) 33.16
Ȳl
(20.52)
19th & L 1.029***
(0.270)
CNN (First St NE) 0.949***
(0.241)
Farragut Square 0.195
(0.281)
Federal Center / Patriots Plaza 1.919***
(0.506)
Franklin Square 0.711***
(0.258)
Gallery Place / Chinatown 1.038***
(0.307)
L’Enfant 0.891***
(0.271)
Metro Center -0.0350
(0.310)
NOMA 1.625***
(0.307)
Navy Yard 2.514***
(0.426)
State Department (20th & Virginia Ave NW) 1.711***
(0.266)
Union Station 1.043***
(0.292)
Constant -4.488***
(0.729)
Observations 331
Number of Truck id 154
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Table 1.5: Location F.E. from Consumer Demand Estimates
36
Projection of Ind. F.E
on Characteristics
Yelpstars=2 0.134
(0.0866)
Yelpstars=3 0.376***
(0.0629)
Yelpstars=4 0.392***
(0.0629)
Yelpstars=5 0.450***
(0.109)
American -0.155***
(0.0516)
Mediterranean -0.624***
(0.0582)
Latin American -0.465***
(0.0653)
Indian -0.682***
(0.0636)
Caribbean -0.544***
(0.0809)
Exotic -0.607**
(0.236)
Twitter followers (000s) 0.478***
(9.87e-06)
Constant -0.0414
(0.0692)
Observations 305
R-squared 0.516
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Table 1.6: Truck Fixed-Effects OLS Projection on Individual Characteristics.
1.5.2 Conditional Logit Location Entry Probabilities
Using with Twitter data, I fit a conditional logit model. Truck j’s utility from
entering location l is modeled as:
37
V = LStage2 +GStage2 × LStage2 +X βStage2 × LStage2jl l j l j l l + jl (1.8)
ij are Type I extreme value errors. Estimating the model above is simply a
conditional logit regression with a dummy variable indicating the location choice
made on the left-hand side with a full set of interactions between truck specific
characteristics (cuisine genre, Twitter followers, and Yelp stars) and location dum-
mies on the right-hand side. In the Twitter data, I see more locations tweeted by
the trucks than I model. I model a total of 12 locations, and these locations account
for approximately 86% of the locations visited by the trucks in the data. Loca-
tions that are not included are mostly in the fringes of central Washington, DC,
one example of such a location is South Capitol (the most southern point in Figure
1.3). Other non-modeled locations include locations such as George Washington
University (Foggy Bottom), Georgetown University (Georgetown) etc. I exclude
these university-based locations as the markets that realized at these locations are
different to the standard ‘lunch shift’ locations on many dimensions. In the empir-
ical estimation, these non-modeled locations make up the outside option and net
profits are normalized to be $500, this helps me pin down the scaling parameter
as it provides a more stable outside option probability in the iterations where I
am matching the moments for this parameter. In terms of interpretation, we can
simply add this constant to the fixed entry costs once they have been estimated,
to interpret the entry costs as a dollar value relative to the outside option being
normalized to $0.
The specification that I proceed with has the full set of interactions as outlined
38
above but has the Yelp stars converted to a dummy variable for whether it is ‘high’
or ‘low’, where the category for ‘high’ is a rating of greater than 3.5 stars. Other
specifications I’ve attempted to estimate faced issues in the estimation with non-
convergence and non-convexity of the maximum-likelihood objective function. The
model has an Pseudo R2 of 0.1859. The coefficients of this estimation are not of
interest directly and merely used to match with model moments in the next stages
of estimation, hence I do not report the table of estimated coefficients. However,
interesting patterns in the location choice probabilities can be observed to get a
clear idea of the data moments that are being matched with the model moments for
estimation of the location entry costs. Figure 1.6 shows a box plot of the estimated
location entry choice probabilities. The figure excludes outliers to being out the
relevant features of the estimates better. The vertical axis of Figure 1.6 shows the
estimated choice probability of the location marked on the horizontal axis being
chosen by a truck, given that the location is in his choice set. The variation in the
mean of these estimates across locations will play a crucial role in the estimation
of the location entry costs in the next estimation stage.
39
Figure 1.6: Variation in the Estimated Conditional Location Choice Probabilities
by Locations.
The choice probabilities of the non-lottery locations (colored in red) are much
lower than the lottery locations, this is because these locations appear in every
truck’s choice set but are not chosen as frequently as the locations that only ap-
pear in a truck’s choice set if allocated by the lottery, which are chosen more
frequently. In other words, trucks on a random day of the week, are more likely to
be visiting a lottery allocated location over a non-lottery location given the lottery
location is in the trucks’ choice set. With respect to the structural model, these
estimates imply that the implied entry costs will be lower in locations we observe
with high choice probabilities and similarly, low choice probability locations will
have high entry costs. This is discussed in more detail below.
40
1.5.3 Marginal Costs
In this stage I back out the constant marginal costs for a set of trucks. I model
224 individual trucks that I see consistently in the lottery outcome data. Given
that www.foodtrukfiesta.com lists the total number of trucks permitted to operate
in DC to be 245 trucks, my model contains the vast majority of the trucks that
are operating consistently in Washington, DC.
Taking the first order conditions of the truck’s maximization from above we
get that the truck j’s marginal costs are:
∑L
− ∑ l SljMlPrj(l)mcj = pj (1.9)L ∂S
l Prj(l)
ljM
∂p lj
Where Ml is the total number of primary jobs at location l and Prj(l) is the
location choice probability of truck j choosing to enter l which I estimate using a
reduced form conditional logit model in Stage 2. Computing the marginal costs
requires the derivatives:
∂Slj −αSlj(1− σSj|g − (1− σ)Slj)
= (1.10)
∂pj (1− σ)(Ȳl − pj)
Which I simulate using the estimated conditional logit choice probabilities given
lottery designated choice sets. The distribution of marginal costs and price-cost
margins16 in the DC food truck industry derived from my model are shown in
16price-cost margin = p−cp
41
Figure 1.6. With these marginal costs, the average price-cost margin is calculated
to be 0.6602. There is little past work on food truck competition and margins
to compare this estimate with. However, to shed some light on its plausibility I
interviewed an acclaimed industry leader. I learned that the general consensus rule
of thumb in the industry was that for a truck truck to successfully operate in the
long-run the food cost component of a menu item should be no higher than 1/3
of the price (i.e. a price-cost margin of 0.66). Interpreting the marginal costs in
my model as the food costs required to make an additional order and assuming
that labor costs are a longer-run decision,17 the average price-cost margins that I
find are strikingly similar to this number. In my estimates however, there is large
variability in the price-cost margins but approximately 95% of the trucks have
margins greater than 0.55.
Figure 1.7: Marginal Costs ($) and Price Cost Margins. Recovered from Model
First Order Conditions.
17It is plausible to think that the number of cooks on a truck are predetermined and they are
paid by hours worked, not numbers of orders produced.
42
1.5.4 Location Entry Costs and Scale Parameter
All the estimates from Stages 1-3 are used to estimate the location specific entry
costs and the scaling parameter. The scale parameter is usually unidentified solely
with discrete choice data, however, in my model the profit function for the firm
is not reduced form and the scale of the firm’s utility is in dollars. Hence it is
important to get the scale/variance of the conditional logit model correct.
The discrete choice model for the vendors can be written by re-scaling the
location specific profit - entry costs (i.e. firm “utility”) terms by λ. If we knew the
entry costs, we could run a simple multinomial logit model on the actual firm utility
and the coefficient on this variable will absorb the scale parameter. However, I
don’t observe the entry costs which also need to be scaled. I estimate the scaling
parameter using the following definition and a fixed-point algorithm:
∑∑
( )
1
j E maxJ l Π̂jl − cl
λ = 1 (1.11)
j E (maxl Vjl + jl − J j0)
This definition follows Anenberg and Kung [2015] and can be interpreted as
defining the scale parameter as the ratio between the average expected maximum
profit in dollar terms (the numerator), and the average expected maximum value
in normalized utility terms (the denominator). More specifically, we this identity
comes from the fact that the reduced form logit model (using reduced form profit
functions) doesn’t separately identify the scale parameter separately but estimates
the product λβ. To identify λ we need to determine the scale of the reduced form
profits.
43
Empirically, the expected maximum profit are the profits that we compute
given the demand and marginal cost estimates. So, for some entry costs cl, I will
approximate the numerator as the average of estimated observed profits − costs
over the trucks I model. The denominator is obtained from calculating the truck’s
inclusive values from the estimation of the conditional logit model above.
Note that the algorithm will take the denominator as data, but in each iteration
re-compute the numerator. The fixed-point algorithm for estimating the entry
costs and the scaling parameter is as follows: Guess an initial value of λiter=0, set
tolerances c and σ.
1. Scale each Π̂ − c by λiter=kjl l .
2. Given the scaled p∑rofit(− costs (i.e. utilit)y) solve the following minimiza-2
tion problem min L PrModelc l=1 l − E[Prl] to obtain a solutions vector of
c which is a L× 1 vector of cls.
3. Now given the solution vector of c, compute the scaling parameter with the
above definition for the scaling parameter and define as λiter=k+1.
4. Compute the distance between λiter=k+1 and λiter=k check if it is less than
the tolerance, if so, stop. otherwise update λiter=k = λiter=k+1 and repeat
from 1.
Where: ∑ ( )J1
PrModell = J ∑
Π̂
exp jl
λ ( )
Π̂jl
j=1 1 + l∈L expj λ
This equation is a part of a system of L+ 1 equations, one each for each mod-
eled location and the outside option. Practically, the outside option that has the
utility being normalized are the non-modelled locations. These include locations
such as university car parks and more of the fringe DC locations marked on the
44
map shown in Figure 1.3. The total of 13 parameters to be estimated (L = 12
location specific entry costs and the scaling parameter) in this routine are the
values of the parameters such that the model moments and the data moments
are matched. Matching the model entry probabilities of firms and the observed
probability of entry is similar to the strategy employed by Seim [2006] to obtain
coefficient estimates on how the number of competitors within a distance band
impact profitability. The difference is that in my model, I also need to consider
the scaling parameter which requires an additional condition to pin down.
The estimates I obtain for the location specific entry costs are shown in Table
1.8. These estimates are taking the demand parameters as data, and have not been
corrected for the imprecision in the demand estimation stage. We can observe the
patterns in Figure 1.6, the lottery locations entry costs are in general lower than
the non-lottery locations and also larger markets exhibit higher entry costs.
Locations Lottery Entry Costs ($) GMM s.e.
19th & L Area No 543.55 14.14
CNN (First St NE) No 861.09 24.75
Farragut Square Yes 122.64 6.96
Federal Center / Patriots Plaza Yes 107.27 4.43
Franklin Square Yes 142.02 7.27
Gallery Place / Chinatown No 412.53 15.14
L’Enfant Yes 76.23 7.53
Metro Center Yes 76.40 6.48
Navy Yard Yes 146.03 4.58
NoMa / New York Ave Metro Yes 556.30 5.85
State Department Yes 145.02 5.92
Union Station Yes 159.95 6.54
Table 1.8: Estimated Entry Costs.
The estimates must be understood with careful consideration of the limitations
45
of the model. For example, for CNN the entry costs are estimated to be very high
($861.09), while this may be in line with the fact that this location is not part of
the lottery, its market size is much smaller than 19th & L which is also not part
of the lottery but has a much lower entry cost. This implies that the Twitter data
predicted choice probabilities for CNN are lower than the what the small market
size in the model would imply. The reason seems to be that the model is not
capturing the decisions of the trucks that are going to CNN well. In my Twitter
data, I observe a few of the same trucks entering this location, maybe due to some
type of dynamic reason such as customer base loyalty. Hence, the estimated choice
probabilities for this location on average are quite low which then the model com-
pensates by boosting the entry cost parameter. This is a issue that is present in
all of the estimates but it seems to be more pronounced with the CNN entry cost
estimate.
Another observation of note is that even among the lottery locations, there is
considerable variation in the estimated entry costs. This suggests that despite the
lottery in practice homogenizing the entry costs for designated locations (recall
the $175 monthly total fee) these locations are still not valued equivalently to the
trucks. There are events in reality where allocations awarded by the lottery are
forgone and not entered, this behavior is unmodeled and inflates the entry cost
parameter. The locations where the this seems to be an issue are the relatively
less populous locations like NoMa.
1.6 Model Performance
Here I use the estimates discussed above and re-solve the model’s equilibrium to
assess how the model performs in explaining the data. Firstly, I consider consumer
46
utilities. Given the nested logit structure of my demand model, from [Train, 2009]
we know that the expected utility for consumer i making a choice among alterna-
tives in Γgl is:
∑ Vij
Iigl = ln exp( ) (1.12)
1− σ
j∈Γgl
Figure 1.8 shows a scatter plot for quantity in equation (1.12) as predicted by
the model and as implied by the location choice probabilities and prices observed
in the data. The model captures the expected utilities quite well.
Figure 1.8: Model Fit of Location/Genre Specific Expected Utilities
Checking the model fit for prices (Figure 1.9 Panel (a)), we can see that the
observed prices tend to be higher, also we see that the prices observed in the
market are set in a almost discrete way (for example, 8.99, 9.99, etc.), and this
bunching seen in reality throws the fit of the model off. It appears that the model
is consistently predicting prices that are lower than we observe in the data. Given
the bunching it is plausible that in reality the sellers round their price up to the
47
nearest 0.99 and this explains the consistent under-prediction. Given these features
of the data, I believe the model fit is quite good. The market shares for the trucks
at each location seem to be predicted well by the model too (Figure 1.9 Panel (b).
48
(a) Model Fit of Prices.
(b) Model Fit of Market Shares.
Figure 1.9: Model Fit at Competition Level (Second Stage of Model).
49
Figure 1.10 shows the panel of location choice fit checks for each location, with
the conditional logit estimated choice probabilities from the Twitter data on the
horizontal axis and the model predicted choice probabilities on the vertical axis.
The model seems to do fairly well in some locations and less well in others. Recall,
the issue with the CNN entry costs. This is reflected in the CNN panel of Figure
1.10. Firstly, the choice probabilities are small in general (all less than 0.03 per-
cent) and also the conditional logit model predicts a near zero choice probability
for a lot of trucks. The other location with predictions that don’t seem to be doing
very well is Gallery Place. Here for a lot of the trucks, the Twitter data implies
a much higher choice probability than the model predicts. Further examination
shows that the trucks that the model is not performing well for are disproportion-
ately Asian and Latin American trucks. There must be some unmodeled reason
that these genres of trucks enter non-lottery but very active locations with a higher
probability than the other genres.
50
Figure 1.10: Model Fit of Location Choice Probabilities
1.7 Conclusion
In this chapter I have collected data on and estimated a model of the food truck
industry in Washington, DC. Due to the novelty of this industry, regulators in prac-
tice fall back on ad hoc policy regimes without understanding the consequences of
potential regulations and/or regulations already implemented. Theoretically, it is
also a challenging landscape to model and analyze due to the large number of firms,
locations, and externalities arising out of who the locations are the allocated to.
Building on a large and well-developed literature dealing with endogenous product
positioning and firm entry, chapter 1 offers a structural model as an example of
how, despite the challenges, this industry can be analyzed and scrutinized in a
51
rigorous way.
I find that consumers preferences are only moderately correlated within cuisine
genres (σ̂ = 0.241) and that trucks’ own price elasticity’s are on average about
-1.61. Firms price-cost margin is on average 0.66, which closely resembles industry
insiders rule of thumb for a truck that can operate into the foreseeable future. I
find that the choice probabilities for lottery locations are higher than that of the
non-lottery locations given both are in a truck’s choices set. In my model, this
implies that given the lottery, entry costs for the lottery locations are lower that
of the non-lottery locations. However, the heterogeneity of each lottery location is
still reflected in my estimates.
My model seems to perform well in predicting post-entry competition but is
worse at predicting the location choice and entry stage of the model. This is due
to the myriad of factors that the model fails to capture completely such as truck
specific location choice decision rules, lottery allocations forgone and swapped by
trucks, and any other general non-adherence to the allocation mechanism.
In chapter two I will use these estimates and the to consider two counterfactual
policy regimes. Firstly, a scenario in which some locations are removed from the
lottery and open to entry to all of the trucks and secondly, a counterfactual policy
where the parking capacity of the non-lottery locations are increased. This coun-
terfactual policy will be compared with the model prediction to assess the welfare
effects of going from the current regime to another.
52
Chapter 2 Counterfactual Policy Analysis: Examining the
Welfare Impacts of Washington, DC’s Mobile
Roadside Vending License Program
2.1 Introduction
The food truck industry in Washington, DC relied on a propriety lottery run by the
DC Department of Consumer and Regulatory Affairs (DCRA), namely the Mobile
Roadside Vending License Program. The program allocates registered trucks to a
set list of locations (see Figure 1.2). I assess the welfare impacts of two alternative
regulatory policies in the Washington, DC food truck industry. Firstly, a counter-
factual policy where some locations are removed from the lottery, such that trucks
are allowed to enter more locations without being subject to the lottery, up to
the location specific capacity constraint being satisfied in expectation. Secondly,
I will consider a counterfactual policy where the number of parking spaces are
increased by 2 spots in each non-lottery location (CNN, L’Enfant, Gallery Place
in my model). In Chapter 1 I have introduced and estimated the model that I
will utilize in this counterfactual analysis. The importance of these counterfactual
policy questions are not limited to academic curiosity, it is also practical for policy
makers in many cities all over America. This dissertation chapter will assess real
policies that can be implemented by a regulatory body. Such motivation is quite
novel in the literature.
The initial motivation for the implementation of the current lottery system, was
53
both to protect brick-and-mortar stores and to stop trucks from engaging showing
up earlier and earlier to secure a space at favorable locations. In 2018 it has now
been 5 years since the inception of this policy and some negative side-effects seemed
to have emerged. Not only are trucks continuing to show up early to secure loca-
tions in the non-lottery spots anyway [MacFarlane et al., 2018] while the DCRA
is slow to respond by adding such locations to the official lottery. There seems to
be a slowing down of the industry with truck owners reporting declining revenues.
Anecdotal evidence suggests this is due to a lack of innovation and market shares
being taken by food trucks that are similar and low quality but are able to stay in
business anyway by simply obtaining the right to operate in the lottery designated
locations, which are the most popular locations in DC [Hayes, 2018].
This anecdotal consensus combines both long-run and short-run arguments. In
the long-run, allowing the market to be overrun by similar and low quality trucks
is bad for the entire industry as the appeal of food trucks in general may decrease.
In the short-run this suggests a benefit to some of the food trucks, as it is allowing
the non-differentiated low quality trucks to survive. In my counterfactual analysis
I focus on relatively marginal changes to policies. This is because my model is
geared to predict short-run states of the market, as key important long-run as-
pects of the industry are not modelled. For example, I don’t model the stage of a
food truck entrepreneurs decision to enter the industry in general and his choice
of cuisine genre and quality. I take these long-run aspects as exogenous and solely
focus on the pricing and location choice aspects of the industry. That said, with
my model, by looking at how the lottery is impacting consumer surplus and the
distribution of trucks operating in the modelled locations, we can get an idea of
the direction of the long-run implications.
54
The first counterfactual policy will open up some of the currently lottery lo-
cations, which are in general the more popular locations to entry, by more trucks
which in the model should increase the entry costs to these locations and decrease
the entry costs to other locations. My second counterfactual policy will slacken
up the capacity constraint for each non-lottery location while keeping the set of
lottery locations unchanged. This will reduce entry costs for the trucks but make
competition fiercer at these locations (more firms competing for the same mar-
ket). As the parking spot allocation mechanisms and capacity constraints change,
there are various trade-offs that must be considered. In the current regime, the
various costs that a truck may experience to secure a spot for the most popular
locations are artificially decreased by removing these locations from the choice sets
of many trucks and reducing the number of potential entrants. From the truck’s
perspective this is a good thing as long as the entry costs to the location without
the lottery is higher than the costs to register for the lottery. But the lottery
may have an adverse impact for the truck’s too. For example, if you are a high
quality and popular truck that has high market shares, your potential profits may
be decreased due to the fact that the lottery is not allowing you to enter only the
most populous locations every day. Effectively the lottery may be propping up
lower quality trucks in two ways; by decreasing the entry costs to a location which
would be unaffordable for a low quality truck without the lottery, and also allowing
them to face a less competitive market place once they have entered the popular
location by distributing these high profit parking spots to arbitrary trucks that
are not the most competitive. In this scenario the lottery would be benefiting low
quality trucks and hurting high quality trucks. In other words, the lottery’s and
capacity constraint’s implications on industry profits and utilities are ambiguous.
55
In terms of prices, the lottery should be reducing the downwards pressure on
prices by restricting the intensity of competition. This again may allow lower qual-
ity trucks that otherwise wouldn’t be able to compete to survive resulting in higher
prices and lower average quality for the consumers. There are also implications of
the parking space allocations on the distribution of truck types. If the distribu-
tion of the cuisine genre is not even across genres (like in reality), a lottery that
randomly allocates a set of trucks to a location may result in a distribution that
is not preferred by the trucks and/or by the consumers (for example, if demand
resembles nested logit utilities). In this case getting rid of the lottery will help
even out the variety of trucks across the different locations.
2.1.1 Method of Comparison for Counterfactual Scenarios
The counterfactual scenarios will be compared to the status quo across several
dimensions. Below I outline them before heading into the comparison itself.
Supply Side: Trucks
I will assess how the counterfactual scenario has impacted the supply side of the
model by looking at the entry costs, prices, and expected profits. The entry costs
are calculated by matching the location average choice probability implied by the
locations capacity constraint and the analogous model moments. For example if
there are 200 trucks active, then the average choice probability for Union Station
would be PrUnion tation = 14/200 because 14 is the capacity constraint at UnionS
Station. In other words, the entry costs will adjust the model’s location choice
probabilities such that on average, the capacity constraint will be satisfied for
56
every location.1 Looking at the change in prices will help me gauge the impact
of the change in competitive structure due to the counterfactual scenario and the
changes in expected profits will assess the net effect of the changes on the trucks.
Demand Side: Consumers
To calculate the net welfare effect of the policy change we must also calculate the
net effect on consumer surplus. Given the nested logit demand structure, following
[Train, 2009] I can calculate the expected utility of a whole location in the standard
‘inclusive value’ formulation, for each l ∈ 1, ..., L and denoting the outside option
as g = 0:
∑∑ ( )1−σG vk + α ln(Ȳl − pk)Il = ln exp   (2.1)
1− σ
g=0 k∈Γgl
Where an approximation of the inner sum is obtained in solving the model
(i.e. Îgl) I can then obtain the dollar representation of the expected utility by
multiplying Il by the inverse of the marginal utility of income. In my model, the
marginal utility of income is:
α
MUlY = (2.2)
(Yl − p)
In my calculation I simply use the average price of all the trucks that are
operating in the modelled locations in the equilibrium.
1The standard deviation of the number of trucks at each location during the simulation is
reasonable. The location with the most spread is L’Enfant with a standard deviation of 4.19.
This implies that a capacity constraint violation of more than 4 trucks is quite unlikely.
57
2.2 The Impact of the Counterfactual Policies
2.2.1 Counterfactual I: Reducing the Reach of the Lottery
Supply Side: Trucks
The counterfactual policy I consider first is for the DCRA to reduce the reach of
the lottery by removing some key locations as a lottery designated location. Dur-
ing estimation, the truck’s choice sets were restricted and the model took what was
determined by the publicly available DCRA lottery outcomes as given. In contrast,
in this counterfactual scenario some lottery locations are added to every trucks’
choice set. Given these expanded choice sets, the trucks make entry decisions and
choose prices such that their profits are maximized as outlined in Chapter 1.
I solve the model for two scenarios for this counterfactual analysis. Firstly,
releasing L’Enfant from the lottery and secondly, releasing L’Enfant and Franklin
Square. These locations have been chosen because they large and prominent loca-
tions in the DC food truck industry.
The changes to the entry costs can be seen in Table 2.1. We can observe the
entry cost increases in the hypothetically excluded locations. Column (2) which
show the entry costs when only L’Enfant is excluded shows that entry costs will
increase from $123 to $282.62. On the other hand, all the other locations’ entry
costs decrease. This is expected, as L’Enfant is now accessible to all trucks, there is
more pressure on L’Enfant ’s capacity constraint whereas there is now less pressure
on the capacity constraint of the other locations. From column (3), which shows
the entry costs if Franklin Square is also excluded from the lottery we can see a
similar change. The entry costs for Franklin Square increases while the entry costs
58
for the other locations fall. To assess the impact of the counterfactual policies on
trucks, we will need to look at the equilibrium prices and profits.
Entry Costs ($)
L’Enfant L’Enfant + Franklin
Lottery Current Excluded Excluded
(1) (2) (3)
19th & L No 499.95 493.76 492.76
CNN (First St NE) No 657.15 652.17 650.39
Farragut Square Yes 101.24 84.10 84.56
Federal Center / Patriots Plaza Yes 82.38 60.21 57.41
Franklin Square Yes 141.93 129.97 287.74
Gallery Place / Chinatown No 400.54 396.40 394.50
L’Enfant Yes 123.00 282.62 281.08
Metro Center Yes 68.50 64.29 54.44
NOMA Yes 155.67 146.52 144.96
Navy Yard Yes 440.17 425.31 416.85
State Department Yes 167.56 159.83 148.22
Union Station Yes 128.37 128.28 115.42
Table 2.1: Entry Cost Changes Under the Counterfactual Policy
Figure 2.1 shows the prices in the scenarios I am considering. Prices do not
seem to have changed too much in the counterfactual scenario.
59
(a) L’Enfant Excluded (b) L’Enfant + Franklin Sq. Excluded
Figure 2.1: Current Regime Model Predicted Prices ($) VS. Counterfactual Regime
Prices ($)
While prices do not change much, profits for the trucks decrease as the reach
of the lottery is reduced (Figure 2.1). which shows a scatter plot of the simulated
trucks’ profits under the current and counterfactual scenarios. Current regime
profits are higher than the counterfactual profits and we can see that as more lo-
cations are taken out of the lottery, profits fall lower. How much a truck’s profit is
impacted in the counterfactual scenario is dependent on many factors such as your
quality, genre, and marginal costs but on average it seems like it is a net negative
impact on profits. Taking a closer look at how the counterfactual regime impacts
on profits, I find that having a lot of trucks in your own genre has large impacts
on your profits. The trucks that are impacted the most by the removal of these
locations from the lottery were American, the most common genre.
60
(a) L’Enfant Excluded (b) L’Enfant + Franklin Sq. Excluded
Figure 2.2: Current Regime Model Predicted Profits ($) VS. Counterfactual
Regime Profits ($)
There are some trucks that stop operating in the modelled locations and this
impacts the distribution of genres in these locations (i.e. no feasible price to the
profit maximization problem). In particular as it is the more common cuisine gen-
res that will be affected the most by the lack of a lottery, the model would suggest
that the distribution of genres become more even distributed. Another force that
will impact whether a truck continues operating or not is the truck’s quality i.e.
given a distribution of genre types, lower quality trucks will be the quickest to stop
operating as entry costs and competition increases. If there is some strong correla-
tions between a particular genre and quality, the lottery may not necessarily make
the distribution of trucks more even. This is what I observe in my counterfactual
analysis.
Removing one location from the lottery results in one truck not operating in the
modelled locations. In particular this truck is a low quality Indian truck and in the
scenario with two locations removed from the lottery we see an additional Ameri-
can truck cease to operate in the modelled locations. These observations suggest
61
that Indian trucks must on average be low quality trucks and this is a stronger
force than the more intense competition that American trucks’ experience from
other trucks of their genre. In other words, Indian truck qualities should be much
lower than the American trucks. We can confirm this is indeed the case, by look-
ing at the estimated quality measures (i.e. the empirical parameterization of vj
estimated from the demand estimates). Figure 2.2 shows that Indian trucks have
low quality, compared to the American trucks. Given that I only find the nesting
parameter in the nested logit model to be 0.241 it is reasonable that the impact of
there being a lot of trucks in your own genre isn’t a strong driver of who exits or
stays in the market but despite this I observe an American truck which on average
has relatively high quality choosing the outside option location.
62
Figure 2.3: Box Plots of Estimated Quality (vk) by Cuisine Genre
With all these considerations, the decrease in total industry profits from the
counterfactual policy amount to $2,698.06 and $5,289.85 respectively for the two
scenarios.
Demand Side: Consumers
Table 2.2 shows the computed average consumer surplus for each location. The
interpretation of column (1), (2) or (3) is the dollar value of the expected utility
of having lunch at each location. The counterfactual scenarios change consumer
surplus due to various reasons and the effect shown in the table is the net impact
of a combination of lower prices, higher value trucks, and consumers demand being
63
factored in with less constraint when firms decide which location to enter. The im-
plied increase in consumer surplus after getting is $4,992.24 and $5,435.83 each day
respectively for the two scenarios. To separate the effect of prices and other factors
that influence consumer surplus, I calculate the change in consumer surplus under
the counterfactual equilibrium but holding prices at the status quo equilibrium.
This change should capture the consumer surplus gains from non-price impacts
such as variety and quality. I find the in both scenarios the consumer surplus
increase from non-price effects are approximately $3,938 out of the total aforemen-
tioned change.
Considering the changes in consumer surplus with the change in truck profits,
in the first scenario the net impact on welfare is an increase of $2,294.18 and in
the second scenario an increase of $145.98.
64
65
L’Enfant L’Enfant + Franklin
Current Excluded Excluded Market Size Implied Change in CS ($)
(1) (2) (3) (4) ((2)-(1)) × (4) ((3)-(1)) × (4)
19th & L 0.289 0.288 0.291 24,583 -20.12 59.38
CNN (First St NE) 0.086 0.051 0.051 16,365 -576.07 -569.17
Farragut Square 0.309 0.363 0.366 23,779 1,280.82 1,358.19
Federal Center / Patriots Plaza 0.087 0.133 0.136 3,659 168.72 178.07
Franklin Square 0.355 0.410 0.422 14,317 799.73 968.36
Gallery Place / Chinatown 0.173 0.128 0.129 10,783 -491.58 -471.18
L’Enfant 0.267 0.367 0.372 16,901 1,684.53 1,777.21
Metro Center 0.279 0.338 0.338 23,020 1,362.39 1,359.18
NOMA 0.265 0.190 0.192 6,074 -453.40 -443.10
Navy Yard 0.193 0.192 0.192 2,898 -2.50 -4.28
State Department 0.150 0.214 0.216 9,358 589.88 609.06
Union Station 0.200 0.241 0.244 14,124 579.84 614.11
Total 4,922.24 5,435.83
Table 2.2: Consumer Surplus (CS) per Consumer by location and Implied Total Change in CS
2.2.2 Counterfactual II: Increasing Parking Capacity at Market Locations
In this counterfactual scenario I increase the parking space constraints at non-
lottery modelled locations. This will have impacts on entry costs to both lottery
and non-lottery locations. The locations directly impacted are 19th & L, CNN, and
Gallery Place, where I increase the number of parking spaces here by 2 spaces. We
would expect to see entry costs to these locations decrease directly as a function
of the capacity constraint not binding as tightly. Also, as more trucks can enter
these locations, the demand for the other locations will decrease and we should see
a indirect decrease in entry costs to the locations where the capacity remains the
same. This is a benefit to the trucks, as a crucial input (parking space) to doing
business is less scarce and consequently costs to entry have reduced. However,
there is a trade-off. With more parking spaces at these locations there are more
trucks and hence competition within the location is fiercer. The simulations of this
counterfactual scenario will shed light on the magnitude of such effects.
Supply Side: Trucks
Table 2.3 shows the changes in entry costs under the counterfactual scenario. We
can see that the entry cost decreases due to the direct effect of a location obtain-
ing more parking spaces is larger than the indirect effect that the other locations
experience. CNN especially experiences a large entry cost change as it is a very
constrained location to begin with (only has 4 spaces where trucks can park), while
19th & L and Gallery Place see changes in the magnitude of about $40-$50. The
lottery locations only see entry cost changes of approximately $15-$17.
66
Entry Costs ($)
Current Counterfactual Difference
(1) (2) (1)-(2)
19th & L 499.95 457.20 42.75
CNN (First St NE) 657.15 537.44 119.71
Farragut Square 101.24 84.69 16.55
Federal Center / Patriots Plaza 82.38 65.40 16.98
Franklin Square 141.93 125.37 16.55
Gallery Place / Chinatown 400.54 350.93 49.61
L’Enfant 123.00 107.28 15.72
Metro Center 68.50 51.85 16.66
NOMA 155.67 138.00 17.67
Navy Yard 440.17 422.93 17.24
State Department 167.56 151.81 15.75
Union Station 128.37 111.99 16.38
Table 2.3: Change in Entry Costs Under Counterfactual Parking Capacities.
Figure 2.4 shows the change in prices as the additional parking spaces become
available. Prices fall considerably for a moderate number of trucks while for most
trucks prices aren’t changing. For most trucks, their choice probabilities of enter-
ing these non-lottery locations are quite low hence it makes sense that the impact
of prices aren’t ubiquitous across all trucks.
67
Figure 2.4: Current Capacity Model Predicted Prices ($) VS Counterfactual Ca-
pacity Prices ($)
Investigating how expected profits have changed in Figure 2.5, profits have on
average increased under the counterfactual scenario. The figure suggests that the
impact of increased competition due to the increased capacity does not erode away
the benefits of the reduced entry costs (the average profit increase is $8.08) and
results in a net profit increase to the trucks. The total increase in profits in the
industry is $1,810.22.
68
Figure 2.5: Current Capacity Model Predicted Profits ($) VS Counterfactual Ca-
pacity Profits ($)
Demand Side: Consumers
I calculate consumer surplus in the same way as above and these calculations
are shown in Table 2.4. I also plot the equilibrium prices under the current and
counterfactual capacity. With the counterfactual parking capacities, prices fall and
consumer surplus increases.
69
70
Current Counterfactual Market Size Implied Change in CS ($)
(1) (2) (3) ((2)-(1)) ×(3)
19th & L 0.29 0.32 24,583 755.26
CNN (First St NE) 0.09 0.09 16,365 11.39
Farragut Square 0.31 0.36 23,779 1,276.56
Federal Center / Patriots Plaza 0.09 0.13 3,659 171.76
Franklin Square 0.35 0.41 14,317 852.88
Gallery Place / Chinatown 0.17 0.15 10,783 -212.45
L’Enfant 0.27 0.35 16,901 1,368.49
Metro Center 0.28 0.34 23,020 1,432.57
Navy Yard 0.26 0.20 6,074 -396.55
NOMA 0.19 0.19 2,898 -1.94
State Department 0.15 0.21 9,358 591.58
Union Station 0.20 0.24 14,124 600.47
Total 6,450.04
Table 2.4: Consumer Surplus (CS) per Consumer by location and Implied Total Change in CS
Again, keeping the prices at the status quo equilibrium to calculate utilities
to gauge the impact of the non-price effects of the counterfactual scenario I find
that $6,450.04 of the $5,112.67 consumer surplus increase is due to such non-price
changes. In contrast to the prior counterfactual scenario, this counterfactual gives
rise to an unambiguous welfare increase of $8,260.26.
2.2.3 Conclusion
I find that the net change in welfare going from the current policy to counter-
factual policies I consider are positive. I consider counterfactual scenarios where
fewer locations are a part of the lottery and where the capacity of food truck lunch
locations parking spaces is increased. Note that this doesn’t account for a variety
of other factors that may be considered in a broader definition of welfare. For
example, how the DCRA is spending the revenue generated from the lottery. It
may be that the enforcement of the parking capacities and increasing the safety of
passers-by in these parking locations generates considerable welfare not quantified
in my analysis.
The short-run implications of the lottery are clear from the results of my first
counterfactual analysis. The lottery is facilitating the survival of trucks that would
otherwise not enter any of the modelled locations (recall, that I model most of the
most popular and largest locations). My results agree with the anecdotal consensus
of veteran truck owners in DC as interviewed in [Hayes, 2018]. The article paints
a picture of deeply worried food truck vendors that have been operating in DC
for much longer than the lottery has been around, for example, Kirk Francis who
co-owns the Captain Cookie & the Milk Man food trucks has been quoted to claim:
“If you went to Franklin Square five years ago during lunch, 12 out
71
of 15 trucks would be “great” and “chef-driven” such as Cap Mac,
TaKorean and Dangerously Delicious Pies. Only three would be ...
“budget trucks.” Now it’s the reverse”.
This is exactly what I find the effect of the lottery to be. The lottery is depressing
the quality and variety of the trucks in the market to the extent that quality is cor-
related with genre. The long-run implications of these results on consumer surplus
and the existing trucks are beyond the scope of this chapter, but as the expected
utility from the food truck segment decreases due to the lottery the market may
not be able to support the large number of heterogeneous trucks in the market
going into the future.
My second counterfactual analysis considering the expansion of parking fa-
cilities for food trucks at non-lottery locations suggests that this is a Pareto-
improvement. Both profits and consumer surplus increases. The total benefit
that arises from lower entry costs and greater competition generates a surplus to
both sides of the market. The surplus increase sum to $8,260.26. However, to cap-
ture these benefits, Washington, DC will need to add a total of 6 parking spaces
across 3 already very congested locations.
My findings suggest that reducing the reach of the lottery to include less lo-
cations and adding additional parking spaces, will be beneficial for the food truck
industry of Washington, DC.
72
Chapter 3 The Impacts of Large Disruptions on Long-Run
Public Transit Ridership: An Analysis of Wash-
ington, DC’s Subway Transit System.
3.1 Introduction
Public transit systems are very important in congested cities for many reasons.
The positive externalities of a well-functioning public transit system are primarily
based on improved efficiencies with respect to; congestion, environmental harm,
and labor mobility. Hence, public transit is of direct interest to many parties, in-
cluding but not limited to; governments, engineers, environmental scientists, and
economists, consequently there is a large body of interdisciplinary work attempting
to model and understand public transit. Modeling demand for transport is compre-
hensively outlined in Domencich and McFadden [1975], and ever since, countless
authors in various locale and continents have applied discrete choice models to
understand the behavioral and economic aspects of transport mode choice.1 Sim-
ilarly, there are abundance of research on the costs and benefits of a successful
public transit system.2 This chapter does not attempt to model the decision mak-
ing process of a consumer of transportation and/or attempt to uncover the various
costs and benefits of a public transit system.
This chapter explores more specific questions that must be considered for agen-
1Some examples include; Beirão and Cabral [2007], Paulley et al. [2006], dell’Olio et al. [2011],
Paulley et al. [2004].
2Some examples include;Litman [2017], Rissel et al. [2012], Geurs and van Wee [2004].
73
cies that manage large metropolitan subway systems using reduced form methods.
Transportation agencies need to decide whether to invest in quality improvements,
such as making repairs to improve reliability and reduce delays. There are two
types of costs that the agency must consider. Firstly, the direct cost of the re-
pairs and maintenance and secondly, any reductions in ridership that occur before,
during, and after the repairs. This chapter focuses on the latter type of costs an
agency must consider. I will be exploring impacts on ridership of a large-scale
system repair program geared to increase quality. Lastly, I will also quantify the
magnitude of these impacts in terms of price changes using an observed fare in-
crease. The background for this analysis is the DC Metrorail system, operated
by the Washington Metropolitan Area Transit Authority (WMATA) and their
year-long system repair program SafeTrack which lasted between June, 2016 to
June, 2017. SafeTrack systematically closed down and “single-tracked”3 segments
of the subway system. There are policy implications regarding fare increases and
large-scale maintenance that arise out of the analysis that may be of value when
designing future large repair programs and fare hikes.
My analysis focuses on the “AM Peak” riders to focus on the morning com-
muter market. I find that the price hike implemented at the end of June, 2017
of $0.10 on every ‘tier’ of AM Peak fares decreased average monthly ridership by
2.05%. I also find that the disruptions caused by SafeTrack have a dynamic im-
pact on ridership. There is initially an anticipatory effect as scheduled repair dates
loom. This is expected as WMATA was encouraging people to look for alternative
modes of commuting well before the actual beginning of the repairs. One month
prior to the repairs on a segment, the original-destination station pairs that are
3Where only one side of the rail tracks are used between stations for trains going in either
direction. This induces major delays and unpredictability in train schedules.
74
directly affected, see an average decrease of 2% in ridership. During the month
of the repairs the station pairs that are directly affected see an average of 9.11%
decrease in ridership and I find that there is a persistent decrease to ridership even
after the repairs have been complete. For about 10 months after the repairs have
been finished, ridership doesn’t fully recover to pre-repair levels, persistently being
about 1.68% lower than pre-repair periods.
This chapter will be organized as follows. In section 3.2 I will discuss the insti-
tutional background. Section 3.3 will describe the data and illustrate descriptive
patterns in the data. Sections 3.4 will discuss my estimation methods and esti-
mates, and Section 3.5 will conclude.
3.2 Institutional Background
Metrorail, a subway system whose network spreads across DC, MD, and VA has
seen a large decline in ridership in the last 4-5 years, and the cause4 of this exodus
of metro riders seems to be the persistent degradation of service quality (including
fatal accidents [Jansen, 2016]). The National Transportation Safety Board (NTSB)
has voiced concern over the state of Metrorail and SafeTrack was WMATA’s re-
sponse to the situation. WMATA describes SafeTrack in the following way on their
website:5
What is SafeTrack?
SafeTrack is an accelerated track work plan to address safety
4Simultaneously, with the system’s degrading quality over time, Washington, DC embraced
the advent of the car sharing industry [Moylan and Graves, 2015] which worsened the situation.
However, the impact of car sharing on public transit ridership is not the focus of the chapter.
5https://www.wmata.com/service/SafeTrack.cfm
75
recommendations and rehabilitate the Metrorail system to
improve safety and reliability. Through SafeTrack, Metro will com-
plete approximately three years’ worth of work into approximately one
year. The plan significantly expands maintenance time on weeknights,
weekends and midday hours and includes 16 “Safety Surges” - long
duration track outages for major projects in key parts of the
system.
Why is SafeTrack necessary?
Metrorail is currently open 135 out of 168 hours per week, leaving insuf-
ficient time for maintenance and other necessary track work. By closing
the system at midnight on weekends and expanding weekday main-
tenance opportunities, SafeTrack addresses FTA and NTSB safety
recommendations and deferred maintenance backlogs while restoring
track infrastructure to good health. In addition the 16 “Safety Surges”
will utilize long-duration track outages through around-the-
clock single tracking or line-segment shutdowns that will im-
pact rush hour commutes.
How will SafeTrack impact my commute?
Due to reduced capacity and expected longer travel times, Metrorail
riders are encouraged to consider using alternate travel op-
tions while safety surge work is scheduled on their line. Trains
and platforms may be extremely crowded during peak periods and cus-
tomers may experience extended delays. Review the links below for
more information about each surge project and potential impacts on
76
your commute.
I have emphasized the parts of WMATA’s SafeTrack description that is directly
relevant to this chapter. SafeTrack was undoubtedly an important and ambitious
undertaking. I will be exploring how ridership has been evolving before, during,
and after SafeTrack to better understand the impact and results of this program.
WMATA explicitly states to metro users that commute times will be longer dur-
ing SafeTrack impacted stations and that the consideration of alternative travel
options are encouraged. This potentially gives rise to dynamic impacts, which I
explore.6
Table 3.1 and Figure 3.1 shows the schedule of repairs and the network segments
that fell under SafeTrack. We can see that not all of the repairs are of the same
magnitude, some are much larger than others. There are two types of repairs.
Repairs can either lead to trains on a segment shutting down or to single tracking.
During the repairs that require a full segment shut down, a free shuttle service is
provided between the closed stations. Also, on June 25th, 2017 a fare increase of
$0.10 across all tier of peak time fares was implemented.
6My analysis ignores the indirect impacts of a disruption. The subway system is a large
network where there is a persistent disruption through the entire system even if only a small
segment of the system was disrupted. These impacts are ignored.
77
78
Date Duration (Days) Line Type of Impact Segment Affected
1 Jun 4 - 16, 2016 13 Orange Single Track East Falls Church to Ballston
2 Jun 18 - July 3, 2016 16 Orange Shutdown Eastern Market to Minnesota Ave & Benning Road
3 Jul 5 - 11, 2016 7 Yellow Shutdown National Airport to Braddock Road
4 Jul 12 - 18, 2016 7 Yellow Shutdown Pentagon City to National Airport
5 Jul 20 - 31, 2016 12 Orange Single Track East Falls Church to Ballston
6 Aug 1 - 7, 2016 7 Red Single Track Takoma to Silver Spring
7 Aug 9 - 21, 2016 13 Red Single Track Shady Grove to Twinbrook
8 Aug 27 - Sep 11, 2016 16 Yellow Single Track Franconia-Springfield to Van Dorn Street
9 Sep 15 - Oct 26, 2016 42 Orange Single Track Vienna to West Falls Church
10 Oct 29 - Nov 22, 2016 25 Red Shutdown Fort Totten to NoMa
11 Nov 28 - Dec 20, 2016 23 Orange Single Track East Falls Church to West Falls Church
12 Feb 11 - 28, 2017 18 Blue Shutdown Rosslyn to Pentagon
13 Mar 4 - Apr 12, 2017 40 Blue Single Track Braddock Rd to Huntington/Van Dorn St
14 Apr 15-May 14, 2017 30 Green Shutdown Greenbelt to College Park/Prince George’s Plaza
15 May 16 - Jun 15 31, 2017 Orange Shutdown New Carrollton to Stadium-Armory
16 Jun 17 - 25 9, 2017 Red Shutdown Shady Grove to Twinbrook
Table 3.1: SafeTrack Repairs Schedule.
79
Aug/01 (7 days) ●
● Apr/15 (30 days)
●
●
●
Oct/29 (25 days)
●
●
● ●
Sep/15 (42 days) ● Jun/18 (16 days) May/16 (31● days)
● ●
● ● ●
●
Nov/28 (23 days)
● Jul/12 (7 days) ● Feb/11 (18 days)
● ● ●
●
● ●
●
Jun/04 (13 days) and Jul/20 (12 days) ● ●
●
● ●
●
● ● ●
Jul/05 (7 days)
Mar/04 (40 days)
● 0km2km4km 0km2km4km
Aug/27 (16 days)
blue orange yellow silver blue orange yellow silver
Metro Line Metro Line
green red silver green red silver
(a) 2016 SafeTrack Repairs. Repair Segments Highlighted Red and Blue (b) 2017 SafeTrack Repairs.
for Clarity.
Figure 3.1: The Metrorail system and SafeTrack Repair Segments
3.3 Data
3.3.1 Data Sources
The data I use for the analysis is from two sources. One is a data set obtained
directly from WMATA of monthly average ridership between an origin-destination
pair, by time of day divided into 5 time-intervals and by service type consisting
of Weekday, Weekends and various others (for example public holidays). I only
use the weekday data to focus on commuters. There are a couple of reasons for
this. Firstly, AM and PM peak ridership constitutes on average 56.2% of the rid-
ers while only taking 44.7% of operating time of the system. Secondly, to capture
the persistent impacts of a repairs program such as SafeTrack, I want to focus
on the segment of the market that is actively and repeatedly making a transport
mode choice. The daily time intervals in the data are AM Peak, Midday, PM Peak,
Evening, and Late Night Peak. This monthly data goes from June, 2013 to October,
2017. Out of these I only consider the AM Peak ridership, because the PM Peak
ridership seems to be strongly correlated with the AM Peak ridership. To verify
this, I run a regression of log monthly average ridership on each station pair’s cor-
responding AM-PM commute origin-destination match (for example, regress AM
ridership of Crystal City-Dupont on PM ridership of Dupont-Crystal City). The
coefficient on this regression is 0.97 and is not statistically significantly different
from 1 where standard errors are calculated using pairwise clustering on origin-
destination. Given this, I carry out the rest of the analysis using only the AM
Peak data. In the analysis below, the treatments I focus on are repairs 1-13 in
Table 3.1. I avoid the others because the data set ends not long after these events
which means that I am unable to track the long-run impacts of these treatments.
80
The second data set is from a publicly available data set, that contains data on
the length of every delay that the Metrorail system experienced from September,
2012 to September, 2016 categorized by cause of the delay. Unfortunately, this data
set only encompasses the very beginning of SafeTrack so I am unable to observe
actual quality changes after SafeTrack is complete. However, this data can shed
light on the relationship between delay (a primary quality measure for consumers)
and ridership. I use this data to control for the effect of delays on ridership when
I estimate how the price increase implemented at the end of SafeTrack impacted
ridership.
3.3.2 Descriptive Statistics
From the monthly ridership data, it can clearly be observed that ridership has
been falling for the last 4-5 years. Figure 3.2 shows the sum over all station pairs’
monthly average daily ridership and the average daily system wide delays over
time, the latter number is calculated by adding all the reported service delay min-
utes over a given month and dividing by 60 × 30 = 1800 to obtain average daily
delays in hours. The sum over all station pairs’ ridership approximately represents
the average system wide daily ridership for a given month. The raw data is very
jagged, because there are strong seasonal effects in the data (people ride the metro
more during the warmer months as opposed to winter, also during the holiday sea-
son AM Peak ridership is limited) so I plot the 6 month moving average to bring
out the more important features of the data. Some features of note are the grad-
ual decline in total ridership from late 2013 until the beginning of SafeTrack that
and the clear dip in ridership during SafeTrack starting June, 2016 and somewhat
recovering by March, 2017, by when most of the repairs have been executed.
81
Figure 3.2: 6 Month Moving Average of the Sum of Average Ridership At Every
Origin-Destination Pair. and Average system wise delays.
The increase in total daily delays the whole system accrues can be seen to
sharply increase around March, 2014. This coincides with when the declining
ridership trend begins. Wait time and timeliness seems to be a strong determinant
in consumers commute mode choice. The causes listed for the increase in delays
are primarily due to increased train and infrastructure related issues. In light
of these descriptive figures, it seems that a large-scale repair program seems to
have been the right solution. It makes sense to drive delays down, to decrease the
rate of ridership decline or even increase ridership to 2013-2014 levels. However,
it is difficult to eyeball these effects. The next section will assess the impacts of
SafeTrack and the following fare increase in a more systematic way.
82
3.4 Estimation and Analysis
3.4.1 The Impact of SafeTrack on Ridership
The Magnitude of Impact and SafeTrack Characteristics
We have seen on Table 3.1 that not all SafeTrack repairs are born equal. There is
heterogeneity in the treatment. For example, the treatment for repairs that last 7
days on continuous single tracking are likely to have different impacts on ridership
to repairs that last 25 days with complete segment shut downs. I explore the re-
lationship between these heterogeneities and the magnitude of impact on ridership.
Firstly, I estimate a regression equation with SafeTrack repairs 1-13 in Table
3.1 each coded as separate treatments.7 The corresponding coefficients on these
treatment dummies can be interpreted as the impact of each repair treatment on
the corresponding treated stations pairs and will uncover the differences in mag-
nitude of the effect of each treatment. Secondly, I will consider each treatment’s
affect in light of the treatment’s characteristics to discern any relationships be-
tween the magnitude of impact at treatment characteristics.
The regression equation estimated is:
I×∑(I−1) ∑T
ln(Ridershipijt) = α + O.D. FEij + Time FEt
∑1 112
+ βkSafeTrack Repair kijt + ijt (3.1)
k=1
7Repairs 3 and 4 are combined into one treatment because they are chronologically and
geographically contiguous. This gives me 12 treatments to consider.
83
Subscript ij denotes a original-destination (O.D.) pair respectively for i 6= j and
i, j ≤ I. SafeTrack Repair kijt is the treatment variable, a dummy variable that
is equal to 1 if the station pair ij in period t is “affected” by scheduled SafeTrack
repair k where k indexes the treatment number (i.e. SafeTrack repair). I define
“affected” observations as station pairs where a segment scheduled for SafeTrack
repair lies in between the origin and destination that defines the observation and if
the origin station is not a inner-Washington station. This is because I am focusing
on the AM commuters that are commuting into Washington. How the treatment
variable is formulated is illustrated in Figure 3.3. I control for the time trend not at
the pair specific level (there are too many pairs) but by sample groups categorized
by level of ridership. Pairs of stations are categorized as; HIGH if there are more
than or equal to 400 riders, MEDIUM if there are between 3 and 400 and LOW if
there are 2 or less riders. These cutoffs were determined by eyeballing the data and
it is reasonable to think there are different time trends for stations with less than 3
riders vs. a station with more than 400 riders. Each group LOW, MEDIUM, and
HIGH make up approximately 55%, 44%, and 1% of the stations pairs respectively.
The identification strategy to estimate equation (3.1) is a difference-in-differences
which requires a common pre-treatment trend assumption.
84
● ●
Metro Line orange silver Treatment ● Under Repairs Affected
Figure 3.3: Graphical Interpretation of Treatment 1. Station Pairs where the
Origin Station is Not on the Inner-Washington” Side of the DC, MD, VA Area
and Hence “Affected” (Triangles) are Defined as Treated, Along with the Stations
Directly Affected by the Repairs (Dots).
85
OLS
Treated 0.00127
(0.00133)
Constant -0.0296***
(0.000991)
Observations 154,000
R-squared 0.000
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Table 3.2: OLS Regression of Pre-treatment Growth Rates for Treated and Control
Stations Pairs
Another important note is how we should calculate the standard errors. As-
suming each station pair is independent is not a very good assumption because
there may be arbitrary correlation between pairs if the same origin or destination
is part of another observation. Given this, I cluster observations across two dimen-
sions, original and destination following Cameron et al. [2006], to account for the
fact that each observation falls into both of these dimensions. In Table 3.2 I show
the results of a regression on the growth rates8 on a treatment dummy variable
to investigate whether the common pre-existing trends assumptions are met for
identification of the above equation. I find that there is no significant difference
in the pre-treatment trends. Table 3.3 shows the estimation results for the above
equation. Controls and fixed effects coefficients are not reported.
8Calculated as the first difference of log ridership.
86
(1) (2) (3)
Treatment Quadratic Linear No
k= Time Trend Time Trend Time Trend
1 -0.101*** -0.101*** -0.115***
(0.0260) (0.0260) (0.0313)
2 -0.0553 -0.0553 -0.0615
(0.0403) (0.0402) (0.0488)
3 -0.0572 -0.0572 -0.0616
(0.0362) (0.0361) (0.0445)
5 -0.0637** -0.0636** -0.0702**
(0.0301) (0.0301) (0.0350)
6 -0.0138 -0.0139 -0.0175
(0.0134) (0.0134) (0.0164)
7 -0.0293 -0.0294 -0.0322
(0.0642) (0.0643) (0.0739)
8 -0.0224 -0.0224 -0.0223
(0.0254) (0.0255) (0.0264)
9 -0.429*** -0.429*** -0.518***
(0.131) (0.131) (0.156)
10 -0.0411 -0.0412 -0.0559
(0.0348) (0.0348) (0.0421)
11 -0.0891** -0.0892** -0.102**
(0.0392) (0.0391) (0.0462)
12 -0.0623*** -0.0621*** -0.0640**
(0.0209) (0.0210) (0.0253)
13 -0.195*** -0.196*** -0.238***
(0.0416) (0.0415) (0.0510)
Observations 361,199 361,199 361,199
R-squared 0.235 0.235 0.094
Number of OD id 8,183 8,183 8,183
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Note:Repair k = 3 combines repair 3 and 4 in Table 3.1.
Table 3.3: Heterogeneity in the Impact of SafeTrack
The results are robust to different specifications, and we can clearly observe
the heterogeneity in each repair. There are some obvious relationships that can
be drawn. Firstly, as expected the treatments with the largest negative impacts
87
are the ones with the longest duration. For example, treatment 9 and 13 are the
repairs that lasted 42 and 40 days respectively. Also, treatments with line segment
shut downs seemed to have non-significant effects, with the exception of treatment
12. This may seem counter-intuitive, but it could simply reflect the fact that the
shuttle service between closed segments are quite effective in moving the passen-
gers between the segment that is shut down. Table 3.4 summarizes some of these
comparisons.
Treatment Index to % Change Statistically Shutdown? Duration State
k= Table 3.1 Significant?
1 1 -10.10% Yes No 13 VA
2 2 -5.53% No Yes 16 MD
3 3+4 -5.72% No Yes 7 VA
4 5 -6.36% Yes No 12 VA
5 6 -1.39% No No 7 MD
6 7 -2.94% No No 13 MD
7 8 -2.24% No No 16 VA
8 9 -42.87% Yes No 42 VA
9 10 -4.12% No Yes 25 MD
10 11 -8.92% Yes No 23 VA
11 12 -6.21% Yes Yes 18 VA
12 13 -19.56% Yes No 40 VA
Results from Specification (2) in Table 3.2 are used to construct this table.
Table 3.4: Treatment Characteristics and Estimation Results.
However, the most striking and robust relationship between treatment charac-
teristic and magnitude of effect is geographic in nature. It seems that all of the
statistically significant decreases are treatments that are on the VA side of the
Metrorail system. Only one of the VA treatments (treatment 3) is insignificant.
Treatment 3 is found to have an insignificant effect, but in all 3 specifications the
t− stat is very close from being large enough for significance despite the fact that
this treatment only lasts 14 days. My findings suggest that commuters commuting
88
in from VA substituted the most heavily away from riding the subway.
The Dynamic Impact of SafeTrack on Ridership
There is reason to believe that SafeTrack has an impact on ridership that follows a
dynamic path that starts and continues, before and after the actual repairs. Firstly,
there is potential for an anticipatory effect. SafeTrack repairs did not start sud-
denly with no warning. WMATA announced SafeTrack one month prior to the first
set of repairs beginning and encouraged commuters to look for alternative modes
of transport. Secondly, during the repairs we would see a big decrease in ridership,
both due to the fact that there is literally less trains moving less passengers be-
tween origin and destination, and also because less people are choosing this mode
of transport for their daily commute as there are more delays.9 Thirdly, we could
expect to see a persistent impact on ridership for the stations that went through
the repairs. For example, if there are significant switching and re-optimization
costs to researching to switch to another form of transport for the daily commute,
then we would expect a large negative shock to utility from programs like Safe-
Track would push these people to actually switch to another mode of commuting
and not return to the metro even if the delays have shortened and quality has in-
creased. If this is indeed the case, simply assessing the impact of SafeTrack with a
treatment dummy variable as in equation (3.1) will fail to capture these dynamics.
To capture these effects, I estimate the following regression equation:
9WMATA’s data driven blog documents big surges in Metrobus utilization during repairs
[Catherine, 2016]
89
I×∑(I−1) ∑T
ln(Ridershipijt) = α + O.D. FEij + Time FEt + β0Repair Monthijt
∑1 1−1
+ βlRepairs to begin l periods laterijt
∑l≥LK
+ βkRepairs completed k periods agoijt + ijt (3.2)
k≥1
The subscripts for station pairs and time are as in equation (3.1). With equa-
tion (3.2), I am tracing out the full path of how a treatment of getting a repair
done impacts the station pair’s ridership. My choice of L and K are -2 and 10 re-
spectively. By looking at these before, during, and after repair dummies we can get
a better understanding of whether the aforementioned dynamics are present in the
data. The time fixed effects will capture unobserved month-year specific variables
that impacted the entire metro system, such as seasonal effects and the impact of
delays. Later, when I explore the impact of delays and the post-SafeTrack fare in-
crease, I use these month-year fixed effect coefficients as the left-hand side variable.
Table 3.2 shows the set of estimations (fixed effects are not reported) and Fig-
ure 3.2 visually depicts the estimates from the linear time trend estimates. The
variable on the left-hand side is a log transformation and the coefficients are on
dummy variables so can be interpreted as a percentage change given the dummy
is turned on, for example, β−1 = −0.0251 can be interpreted as, there being on
average 2.5% less riders one month before a station pair is scheduled for SafeTrack
repairs.
90
The estimates are robust to different ways of controlling for the time trend and
we can see the dynamic path of ridership as a station pair experiences a SafeTrack
repair. Ridership fell about 2.5% a month before the actual repairs (i.e. anticipa-
tory effect) and it takes about 2 months for ridership to recover from the shock.
I observe a consistently negative point estimate of the post-repair coefficients al-
though some of the coefficient aren’t significantly different from 0. However, an
F-test for joint significance for all the coefficients for k > 2 suggest that jointly,
the coefficients are all indeed significant (F −Statistic = 31.97). There is indeed a
long-run impact from the disruptions caused by SafeTrack. About 1.68% percent
of former Metrorail commuters seem to have switched to another mode of trans-
port for good.
Figure 3.4: Graphical Interpretation of Regression Coefficients and Confidence
Intervals.
91
(1) (2) (3)
Months to/from Quadratic Linear No
Repair time trend time trend time trend
-2 -0.00221 -0.00223 -0.00342
(0.00825) (0.00826) (0.00992)
-1 -0.0251*** -0.0251*** -0.0305***
(0.00793) (0.00793) (0.00936)
0 -0.0911*** -0.0911*** -0.108***
(0.0185) (0.0185) (0.0221)
1 -0.0441*** -0.0441*** -0.0545***
(0.0102) (0.0102) (0.0121)
2 -0.0230*** -0.0230*** -0.0316***
(0.00742) (0.00741) (0.00793)
3 -0.0151 -0.0151 -0.0183
(0.0139) (0.0139) (0.0160)
4 -0.0130 -0.0130 -0.0169
(0.00992) (0.00989) (0.0113)
5 -0.0141* -0.0141* -0.0161*
(0.00832) (0.00829) (0.00933)
6 -0.00740 -0.00741 -0.0110
(0.00754) (0.00749) (0.00845)
7 -0.0213*** -0.0213*** -0.0236***
(0.00708) (0.00703) (0.00801)
8 -0.0294*** -0.0294*** -0.0364***
(0.00682) (0.00678) (0.00778)
9 -0.0158** -0.0157** -0.0189**
(0.00714) (0.00710) (0.00762)
10 -0.0124 -0.0123 -0.0148
(0.00906) (0.00902) (0.00994)
Observations 361,199 361,199 361,199
R-squared 0.234 0.234 0.093
Number of OD id 8,183 8,183 8,183
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Table 3.5: The Dynamic Impact of SafeTrack on Ridership.
92
3.4.2 The Impact of Service Quality and Fares on Ridership
In this section I look at how service quality and fares impact ridership. Due to
the absence of delay data during and after SafeTrack, it is not possible to assess
exactly how much of an impact SafeTrack has had on actual level of delays. I
replace the missing data, with zeros and then account for the missing time periods
with a dummy variable. This should allow me to capture the impact of delays
on ridership for the parts where data is not missing, and then when it is missing,
account for it in levels.
The left-hand side variable I use are the estimated the month-year dummy
coefficients from the previous regression estimation (in particular, specification
(1)). These coefficients capture the unobserved characteristics (characteristics not
included in the previous regression) of the monthly average ridership in a particular
month-year. Recall, the regression in the previous section is in log scale, which
means that the interpretation of the coefficient estimates in this section are in
percentage changes per unit change in the right-hand side variable. Firstly, I
estimate the following equation:
Time FEt = Month FEt + α + β1Delayt + β2Post fare increaset (3.3)
+ β3SafeTrack + β4Missing + +β5time+ t
The month fixed effects should capture any seasonal fluctuations that are a
function on month of year. The effect of delay on ridership is captured by the
β1 (delays are a continuous measure). β2 captures the impact of the fare increase
on ridership. This should capture solely the fare increase because firstly, the time
93
fixed effect has a lot of the variation coming from other relevant dependent vari-
ables stripped out from it already. Secondly, the delays and month fixed effects
are controlling for any variation in the time fixed effects that don’t vary within
month-year but do vary across month-year. These variations would have all been
pushed into the time fixed effects we are using in the estimation of equation (3.1).
β3 captures the impact of SafeTrack at the system level. Note that the treatments
in equation (3.1) are determined at the origin-destination pair level so indirect,
system wide impacts of SafeTrack are not fully extracted from the time fixed ef-
fects. β4 captures the levels difference of ridership while delays are missing. Lastly,
β5 controls for the time trend.
OLS
Fare Increase -0.0205
(0.0176)
Delays -0.001916
(0.0091)
Safetrack -0.0298847**
(0.0110)
Missing Delays Dummy 0.0118
(0.0290)
Constant 1.8486***
(0.2569 )
Time (months) -0.0030***
(0.0004)
Observations 46
R-squared 0.947
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Table 3.6: The Impact of Delays and Fare Increases on Ridership.
Table 3.6 shows the estimated coefficients for equation (3.3), because of the
94
correlation between delays and the time trend variable, we see that delays are
not found to be statistically significant in the estimated equation. The estimates
suggest that per year, ridership has been falling on average 3.6%. I have also
examined the impact of lagged delays on ridership, however, lagged delays do not
affect ridership at the time interval the data comes in (i.e. one month intervals).
This suggests that most commuter transport mode substitutions happen within a
month. The fare’s impact on ridership is insignificant. The point estimate suggests
that post-fare increase, ridership is on average 2.05% lower than before the fare
increase. The absolute value of the change in fares is $0.1 for any destination
originating from a peak time entry.10 The actual average percentage change in
fare is more convoluted to calculate because the fares are calculated in a non-
linear way, as a function of distance travelled on the track and the coordinate
distance between two stations. However, the price change in percentages ranges
from being a 4.56% increase ($2.15 to $2.25) to a 1.89% increase ($5.90 to $6.00)
and in 2013 the average fare collected by WMATA was $2.90 [Duggan, 2013]. This
would imply an average effective fare increase of 3.45% which suggests ridership
elasticity with respect to fare is -0.59. This is in line with the literature’s prior
findings. Paulley et al. [2004] and Litman [2017] both offer a comprehensive survey
on transit elasticities. The authors in both surveys suggest that metro ridership’s
short-run elasticities are in the range of -0.3 to -0.6, while long-run elasticities are
close to 1.
To compare these estimates with the persistent impacts of SafeTrack, taking
the back of the envelope elasticity calculation above, the estimates suggest that
10Price changes on weekly and monthly passes are different. I couldn’t get data on the market
share for these pass holders and single trip pay-as-you-go card holders, but it is reported at
https://planitmetro.com/?s=smartbene that 87% of Federal Metrorail customers are pay-
ing for the usage using a pay-as-you-go standard fare payment method. This may suggest that
the aforementioned fare increase is relevant to the vast majority of Metrorail users.
95
the impact that SafeTrack had on ridership in the long-run is similar to an average
fare increase of 2.85%.11 The average daily ridership across all treated origin-
destination pairs for the time periods before SafeTrack (June, 2013 to May, 2017)
is 135,767.4. My dynamic coefficients then imply there is about 2,279 (135, 767.4×
0.0168 = 2, 279.36) less AM Peak commuters riding Metrorail to get into DC in the
morning. Assuming that an average commuter pays $3.00 for his/her commute,
this a fare revenue loss of about 204, 840 per month from the AM commute market.
Assuming that the commuters who substituted away from the subway may also
not be riding in the PM Peak time period, the total daily impact is possibly closer
to double the amount above.
3.5 Conclusion
I find that the commuter origin-destination pairs that suffered the largest losses
during SafeTrack are ones that are originating from Virginia, quantifying why this
might be the case goes beyond the scope of the current chapter. There is an indirect
cost to investing in a large maintenance program such as SafeTrack in the form of
persistent rider losses. If consumers face switching costs, the disruptions caused
by maintenance may provide a sufficiently large negative utility shock such that
riders consider other options for transportation. This chapter explored this indirect
cost to assess the persistent impact that SafeTrack had on Metrorail ridership in
the DC, Maryland, Virginia (DMV) area. Considering how delays were increasing
and ridership was decreasing consistently since 2013-2014 a big push to increase
reliability and timeliness seems to have been a good idea. However, I find evidence
of persistent ridership losses. Ridership for the subway system in the DMV area
does not fully recover to pre-SafeTrack levels even up to 10 months after a segment
∑10
11 k=2 β̂k/9
0.59 = 2.85. Using estimates from specification (2) in Table 3.5.
96
has been worked on. My estimates suggest that 2 to 10 months after the repairs
ridership is on average 1.68% lower compared to 2 months before the repairs.
This level of persistent ridership loss translates approximately to losses of about
$410,000 per month. I also find that ridership is inelastic with respect to fares.
This is in line with the literature’s findings.
97
Bibliography
E. Anenberg and E. Kung. Information technology and product variety in the city:
The case of food trucks. Journal of Urban Economics, 90(C):60–78, 2015. URL
http://EconPapers.repec.org/RePEc:eee:juecon:v:90:y:2015:i:c:p:60-78.
G. Beirão and J. S. Cabral. Understanding attitudes towards public transport
and private car: A qualitative study. Transport Policy, 14(6):478 – 489, 2007.
ISSN 0967-070X. doi: https://doi.org/10.1016/j.tranpol.2007.04.009. URL
http://www.sciencedirect.com/science/article/pii/S0967070X07000522.
C. L. Benkard, P. Jeziorski, B. V. Roy, and G. Y. Weintraub. Nonstationary obliv-
ious equilibrium. Economics Working Paper Archive 568, The Johns Hopkins
University,Department of Economics, Aug. 2008. URL https://ideas.repec.org/
p/jhu/papers/568.html.
S. Berry and A. Pakes. Some applications and limitations of recent advances in
empirical industrial organization: Merger analysis. The American Economic
Review, 83(2):247–252, 1993. ISSN 00028282. URL http://www.jstor.org/
stable/2117672.
S. T. Berry. Estimation of a Model of Entry in the Airline Industry. Economet-
rica, 60(4):889–917, July 1992. URL https://ideas.repec.org/a/ecm/emetrp/
v60y1992i4p889-917.html.
S. T. Berry. Estimating Discrete-Choice Models of Product Differentiation. RAND
Journal of Economics, 25(2):242–262, Summer 1994. URL https://ideas.repec.
org/a/rje/randje/v25y1994isummerp242-262.html.
98
T. F. Bresnahan and P. C. Reiss. Entry in monopoly markets. The Review of
Economic Studies, 57(4):531–553, 1990. ISSN 00346527, 1467937X. URL http:
//www.jstor.org/stable/2298085.
A. C. Cameron, J. B. Gelbach, and D. L. Miller. Robust inference with multi-
way clustering. Working Paper 327, National Bureau of Economic Research,
September 2006. URL http://www.nber.org/papers/t0327.
Catherine. Tens of thousands of customers relied on metrobus during safetrack
surges 3 and 4. Blog Post 48, PlanItMetro, 2016. URL https://planitmetro.
com/2016/09/12/tens-of-thousands-of-customers-relied-on-metrobus-during-
safetrack-surges-3-and-4/.
DCRA. Vending handbook. 2013. URL https://dcra.dc.gov/sites/default/files/
dc/sites/dcra/publication/attachments/VendingHandbook.pdf.
L. dell’Olio, A. Ibeas, and P. Cecin. The quality of service desired by public
transport users. Transport Policy, 18(1):217 – 227, 2011. ISSN 0967-070X. doi:
https://doi.org/10.1016/j.tranpol.2010.08.005. URL http://www.sciencedirect.
com/science/article/pii/S0967070X10001009.
T. Domencich and D. McFadden. Urban travel demand: a behavioral analysis.
Contributions to economic analysis. NORTH-HOLLAND, 1975. URL https:
//books.google.com/books?id=Nx4pAQAAMAAJ.
P. Duggan. Proposed hikes in metro rail fares would make a pricey sub-
way system even pricier. 2013. URL https://www.washingtonpost.com/
local/trafficandcommuting/proposed-hikes- in-metro- rail- fares-would-make-
a-pricey-subway-system-even-pricier/2013/12/07/be6e35ba-5e86-11e3-bc56-
c6ca94801fac story.html?utm term=.b3178762ea38.
99
K. T. Geurs and B. van Wee. Accessibility evaluation of land-use and trans-
port strategies: review and research directions. Journal of Transport Geogra-
phy, 12(2):127 – 140, 2004. ISSN 0966-6923. doi: https://doi.org/10.1016/j.
jtrangeo.2003.10.005. URL http://www.sciencedirect.com/science/article/pii/
S0966692303000607.
L. Hayes. No longer trendy, food trucks facing declining revenue find ways to
survive. 2018. URL https://www.washingtoncitypaper.com/food/article/
20986600/no-longer-trendy-food-trucks-facing-declining-revenue-find-ways-to-
survive.
Intuit. Food trucks motor into the mainstream. Report, Intuit., 2012. URL http://
gourmetstreets.com/wp-content/uploads/2013/12/Free-Food-Truck-Industry-
Report-3.pdf.
B. Jansen. Investigators: D.c. metro ignored safety provisions for years. 2016. URL
https://www.usatoday.com/story/news/2016/05/03/washington-metropolitan-
area-transit-authority-subway-death/83864546/.
T. A. Litman. Evaluating Public Transit Benefits and Costs: Best Practices Guide-
book. Victoria Transport Policy Institute, 2017. URL http://www.vtpi.org/
tranben.pdf?b81542c0?db0c3fd8.
S. MacFarlane, R. Yarborough, S. Jones, and C. Decker. Group of food trucks
squatting on desirable parking spaces in southwest d.c. 2018. URL https://
www.nbcwashington.com/investigations/Group-of-Food-Trucks-Squatting-on-
Desirable-Parking-Spaces-in-Southwest-DC-475185583.html.
M. J. Mazzeo. Product choice and oligopoly market structure. RAND Journal of
100
Economics, 33(2):221–242, 2002. URL http://EconPapers.repec.org/RePEc:rje:
randje:v:33:y:2002:i:summer:p:221-242.
A. Moylan and Z. Graves. Ridescore 2015: Hired driver rules in u.s. cities. Re-
port 48, R Street Institute, 2015. URL http://www.rstreet.org/wp-content/
uploads/2015/12/RSTREET48.pdf.
A. Nevo. Measuring market power in the ready-to-eat cereal industry. Economet-
rica, 69(2):307–342, 2001. ISSN 1468-0262. doi: 10.1111/1468-0262.00194. URL
http://dx.doi.org/10.1111/1468-0262.00194.
NPD Group’s. Recent menu price hikes, more telecommuters, and shopping
online contribute to foodservice lunch visit declines, September 2016. URL
https://www.npd.com/wps/portal/npd/us/news/press-releases/2016/recent-
menu- price - hikes - more - telecommuters - and- shopping- online - contribute - to -
foodservice-lunch-visit-declines/. [Online; posted 13-September-2016].
N. Paulley, R. Balcombe, R. Mackett, H. Titheridge, J. Preston, M. Wardman,
and J. Shires. The demand for public transport: a practical guide. Report 593,
Transportation Research Laboratory: London, UK, 2004. URL http://discovery.
ucl.ac.uk/1349/.
N. Paulley, R. Balcombe, R. Mackett, H. Titheridge, J. Preston, M. Wardman,
J. Shires, and P. White. The demand for public transport: The effects of fares,
quality of service, income and car ownership. Transport Policy, 13(4):295 – 306,
2006. ISSN 0967-070X. doi: https://doi.org/10.1016/j.tranpol.2005.12.004.
URL http://www.sciencedirect.com/science/article/pii/S0967070X05001587.
Innovation and Integration in Urban Transport Policy.
101
S. Qi. The impact of advertising regulation on industry: the cigarette advertising
ban of 1971. The RAND Journal of Economics, 44(2):215–248, 2013. ISSN 1756-
2171. doi: 10.1111/1756-2171.12018. URL http://dx.doi.org/10.1111/1756-
2171.12018.
C. Rissel, N. Curac, M. Greenaway, and A. Bauman. Physical activity asso-
ciated with public transport use—a review and modelling of potential bene-
fits. International Journal of Environmental Research and Public Health, 9
(7):2454–2478, 2012. ISSN 1660-4601. doi: 10.3390/ijerph9072454. URL
http://www.mdpi.com/1660-4601/9/7/2454.
M. Saeedi. Reputation and adverse selection: Theory and evidence from ebay.
Discussion paper, The Ohio State University, 2014.
R. Schmalensee. Sunk costs and market structure: A review article. The Journal
of Industrial Economics, 40(2):125–134, 1992. ISSN 00221821, 14676451. URL
http://www.jstor.org/stable/2950504.
K. Seim. An empirical model of firm entry with endogenous product-type choices.
The RAND Journal of Economics, 37(3):619–640, 2006. ISSN 07416261. URL
http://www.jstor.org/stable/25046263.
D. Staiger and J. H. Stock. Instrumental variables regression with weak instru-
ments. Econometrica, 65(3):557–586, 1997. ISSN 00129682, 14680262. URL
http://www.jstor.org/stable/2171753.
J. Suzuki. Land Use Regulation as a Barrier to Entry: Evidence from the Texas
Lodging Industry. Working Papers tecipa-412, University of Toronto, Depart-
ment of Economics, Oct. 2010. URL https://ideas.repec.org/p/tor/tecipa/
tecipa-412.html.
102
A. Sweeting. A model of non-stationary dynamic price competition with an appli-
cation to platform design. Working Paper 15-03, NET Institute, 2015.
K. E. Train. Discrete choice methods with simulation. Cambridge university press,
2009.
T. Wollmann, D. Pollmann, M. Shepard, M. Sinkinson, H. Tabakovic, and
E. Tamer. Trucks without bailouts: Equilibrium product characteristics for
commercial vehicles job market paper. 2014.
Y. D. Xu. A structural empirical model of r&d, firm heterogeneity , and industry
evolution. 2008 Meeting Papers 744, Society for Economic Dynamics, 2008. URL
https://ideas.repec.org/p/red/sed008/744.html.
103