UCSD Compatibility & Investment in the USA Electric Vehicle Market Discussion

  

please prepare a brief half page executive summary of this paper that highlights:

1. the key findings of the paper;

2. major assumptions or limitations of the analysis;

3. issues of relevance for policymaking.

This summary doesn’t require specific analysis or formal report. Just summarizing the points to answer the 3 questions above. You may skim the reading instead of really read it 🙂

Compatibility and Investment in the
U.S. Electric Vehicle Market∗
Jing Li†
January 27, 2019
Abstract
Competing standards often proliferate in early stages of product markets
and may lead to socially inefficient investment. This paper studies the effect of
unifying three incompatible standards for charging electric vehicles in the U.S.
from 2011 to 2015. I develop and estimate a structural model of vehicle demand
and charging network investment to quantify the impact of a uniform charging
standard. Variation in federal and state subsidies identify the demand elasticities. Counterfactual simulations show moving to a uniform charging standard
increases consumer surplus by $500 million; car manufacturers build 2.8% fewer
charging locations and sell 20.8% more electric vehicles.
∗
I thank my PhD advisors Christopher Knittel, Robin Lee, Ariel Pakes, James Stock, and Elie
Tamer for their continual guidance and support. I thank Meghan Busse, Evan Herrnstadt, Gaston Illanes, Sarah Jacobson, Divya Kirti, Shanjun Li, Charles Murry, Daniel Pollmann, Mar Reguant, Frank
Schilbach, Robert Stavins, Che-Lin Su, Richard Sweeney, Thomas Wollmann, Matthew ZaragozaWatkins, Yufei Zhao, Fanyin Zheng and participants in the Harvard Industrial Organization, Labor/PF, and Environmental workshop and lunches for valuable comments. I thank John Smart and
Sera White at the Idaho National Laboratory for information on U.S. federal electric vehicle programs.
Data purchased for this research is generously supported by the Laboratory for Economic Applications
and Policy. I gratefully acknowledge that this material is based upon work supported by the National
Science Foundation Graduate Research Fellowship under Grant No. DGE1144152.
†
MIT Sloan School of Management, lijing@mit.edu.
1
Introduction
When firms invest in incompatible complementary goods or technical standards, should
the government intervene and mandate compatibility? This question generates fierce
policy and antitrust debates in a wide range of industries, from digital markets to manufacturing because of the ambiguous welfare implications from mandating compatibility.
A shift toward compatibility gives consumers access to the combined investments of all
firms, which may benefit consumers by increasing variety, convenience, or other measures of quality. However, benefits to consumers may be offset by a decrease in firms’
investments. Compatibility fundamentally changes the nature of competition among
firms, turning firms’ investments from substitutes that steal business from rivals to
complements that have positive spillovers onto other firms. Therefore, firms may invest too much under incompatibility because private gains from business-stealing do
not contribute to social surplus, but they may invest too little under compatibility
because private incentives do not internalize the positive spillovers. The theoretical
literature shows that private incentives to provide compatibility can be either too high
or too low relative to social incentives. The welfare effect of a compatibility policy is
open to empirical analysis.
This paper empirically assesses the effect of compatibility on market outcomes and
welfare in the U.S. electric vehicle market, which grew ten-fold in the number of models
and annual unit sales from 2011 to 2015. Electric vehicles attract billions of dollars
in government support for the large potential environmental benefits and innovation
spillovers. Similar to other alternative fuel transportation technologies, including hydrogen and natural gas, electric vehicles require a refueling infrastructure for wider
consumer acceptance. Accordingly, car manufacturers have invested heavily in building fast charging stations to refuel electric vehicles. To the chagrin of many consumers,
however, car manufacturers have aligned themselves behind three mutually incompatible standards for fast charging. Incompatibility across three charging standards will
become an increasingly focal policy issue, with the U.S. Department of Transportation’s proposal to build 48 electric vehicle charging corridors on the national highways1
(or similar programs in the future) and many utilities across the nation proposing to
1
The White House Office of the Press Secretary. 2016. “Obama Administration Announces New
Actions to Accelerate The Deployment of Electrical Vehicles and Charging Infrastructure.” Press
Release, November 3. https://obamawhitehouse.archives.gov/the-press-office/2016/11/03/
obama-administration-announces-new-actions-accelerate-deployment
1
build charging stations2 .
I evaluate the effect of a counterfactual compatibility policy in three main steps.
First, I develop a structural model of consumer vehicle purchase behavior and car
manufacturer build-out of charging networks. Second, I estimate the model using
data from the first five years of the U.S. electric vehicle market (2011-2015). Third,
I use the model and parameter estimates to simulate market outcomes when all car
manufacturers adhere to a single standard for recharging electric vehicles. I compare
the status quo to the counterfactual market equilibrium and social planner solution.
The mobility of drivers poses a challenge to specifying the relevant charging stations
for an individual consumer. The most useful refueling stations to a consumer may be
those that are near their driving paths and destinations rather than their home addresses (Houde (2012)). I cast the available charging network as a vehicle characteristic
in the static, discrete-choice framework of Berry et al. (1995). The model captures rich
geographic variation in consumer driving trips relative to charging station locations.
The model also recognizes the importance of the connectivity of stations – how they
are placed relative to each other and to driving origins and destinations – in addition
to the sheer total number of stations that have been built.
I estimate the key parameters of my model using data on market-level vehicle sales
and charging station characteristics, quantities, and locations. I estimate an empirical Bayes posterior mean for market shares to reduce noise and eliminate zero market
shares. To identify the endogenous demand parameters on price and charging stations,
I collect an original panel dataset of federal and state government incentives. Government tax credits and rebates incentivize consumers to purchase electric vehicles and
businesses to install charging stations. Conditional on market and time fixed effects,
within-market changes over time in government subsidies are plausibly exogenous cost
shifters due to idiosyncrasies in policy-making timing. Additionally, a portion of the
charging stations in the dataset were built as part of a program in the American Recovery and Reinvestment Act of 2009 (Recovery Act) that chose recipient cities before
electric vehicles became available for sale. The exact timing of arrival of these Recovery Act charging stations, conditional on market and time fixed effects, are plausibly
exogenous to unobserved product characteristics and contemporaneous local demand
conditions.
I model car manufacturers as competing in static oligopoly. Combining demand
2
Mulkern, Anne. 2016. “PG&E May Build Nation’s Largest Deploymeent of EV Charging Spots.”
E&E News, August 26. https://www.eenews.net/stories/1060042082
2
parameter estimates and the first-order conditions of the profit function, I recover
firms’ markups for vehicles and costs for charging stations, which are in line with
engineer and industry estimates.
Using parameter estimates from the consumer and firm models, I assess the impact
of a counterfactual policy that mandates compatibility in charging stations. A firm’s
charging station build-out problem, conditional on the strategies of their competitors, is
equivalent to a computational problem called “fractional knapsack” which has a simple
and fast algorithmic solution. I find that some firms do have an incentive to cut back by
4.2% on station quantities under compatibility, though consumer surplus improves by
about $500 million and the number of electric vehicles sold increases by about 100,000,
or 20.8%. The reduction in the number of charging locations under compatibility
suggests that firms make excess investments when standards are incompatible. Indeed,
the social planner would build 17.7% fewer locations compared to private investment
choices under incompatibility. Surprisingly, the social planner builds fewer locations
even compared to prive investment choices with compatibility, which means firms do
not invest too little relative to the social optimum. One way to explain this result
is that even under compatible charging standards, vehicles remain differentiated in
battery range. Therefore, the gains from compatibility are not symmetric, and firms
can still benefit from spatial differentiation in their charging locations.
This paper contributes to three different literatures. First, this paper contributes
to the empirical understanding of the impacts of compatibility. Theoretical predictions
of gains from compatibility are ambiguous, as firms’ private incentives to achieve compatibility may be either higher or lower than social incentives (Katz and Shapiro (1985,
1986)). Previous empirical work on the impact of compatibility has found considerable
gains in consumer welfare (Ho (2006)) and producer efficiency (Gross (2016)). However, firms’ strategic responses to compatibility may moderate the gains in consumer
welfare (Ishii (2007), Lee (2013), Knittel and Stango (2008, 2011)). After computing
the counterfactual equilibrium charging network built by firms, I find that compatibility in charging standards improves consumer surplus by about $500 million despite
firms decreasing the number of charging locations that they build. Compatibility does
not fully erode firms’ incentives to invest in complementary goods for their products
when their products remain differentiated.
Second, this paper contributes to the growing literature on endogenous product
positioning by endogenizing charging station investment by car manufacturers. When
3
product varieties are discrete, firms’ product choices can be thought of as entry decisions. A line of literature recovers fixed costs of new product entry in order to compute
welfare or solve for new product introductions (Wollmann (2018), Eizenberg (2014),
Nosko (2014), Sweeting (2013), and Draganska et al. (2009). In some settings, firms
face a continuous choice space, such as in Crawford et al. (2015) and Fan (2013).
Third, a rapidly growing literature investigates different features of the electric
vehicle market. Holland et al. (2016, 2019), Graff Zivin et al. (2014), and Michalek
et al. (2011) evaluate the short-term environmental benefits of electric vehicles. They
find high geographic variation in environmental benefits of electrifying transportation
within the U.S, depending on the fuel mix of electricity production and population
density. A second stream of this literature focuses on the design and impacts of subsidies for electric vehicles and other green technologies (Clinton and Steinberg (2016),
Sheldon et al. (2017), Borenstein and Davis (2015), Holtsmark and Skonhoft (2014)).
This work finds that consumers respond to subsidies in their decisions to adopt electric
vehicles and other green technologies. Using a two-sided market framework, Li et al.
(2017) and Springel (2016) find that car purchases and charging station build-out respond positively to each other and that subsidizing charging station entry is more
cost-effective in increasing electric vehicle sales.3 Recognizing the importance of charging station availability to the growth of the electric vehicle market, car manufacturers
have become involved in building charging stations. This paper differs from prior work
on electric vehicles by using existing subsidies as identifying variation in a structural
model to evaluate a counterfactual policy about charging standard compatibility. It is
the first to study car manufacturer investments in charging stations.
The rest of this paper is organized as follows. Section 2 discusses the growth of
the electric vehicle market, technical details about charging stations and standards,
government policies, and the dataset. Section 3 specifies a model of consumer vehicle choice and car manufacturer investment in charging stations. Section 4 discusses
identification, estimation, and results. Section 5 uses the model estimates to simulate
market outcomes under a compatibility policy. Section 6 concludes.
3
Greaker and Heggedal (2010) and Pavan (2015) study positive feedback loops between vehicle
demand and refueling infrastructure for hydrogen fuel cell and natural gas cars, respectively.
4
2
The U.S. Electric Vehicle Industry
Institutional details and data availability motivate many aspects of my model. This
section describes the growth of the U.S. electric vehicle market, charging standards
and compatibility policy, the implications of government subsidies and Zero-Emissions
Vehicle (ZEV) regulations for the electric vehicle market and this paper, and lastly,
the data.
2.1
Growth of the U.S. electric vehicle market
Electric vehicles are an increasingly important segment of the U.S. automotive industry, which as a whole accounts for more than 3% of U.S. GDP (U.S. Department of
Commerce (2016)). Since Tesla Motors unveiled the first modern-day electric vehicle in
2006, a luxury sports car priced at more than $100,000, automakers have been selling
models that span a wide range of prices and features.4 Electric vehicles can be classified into two types: (i) battery electric vehicles (BEVs), which only run on electricity,
and (ii) plug-in hybrid electric vehicles (PHEVs), which can take gasoline as a backup
fuel source. Unlike conventional hybrids, plug-in hybrids can be recharged by plugging
into the electric grid. For example, the Toyota Prius first launched as a conventional
hybrid in 2000, and in 2012 also became available as a plug-in hybrid. The unifying
feature across both types of electric vehicles is that they are powered by rechargeable
batteries and can be plugged in for recharging.
The electric vehicle market has expanded since its inception in late 2010 and is
projected by industry analysts to grow much more in the coming decades. The 3
available models in 2011 collectively sold about 14,000 units in U.S. MSAs in 2011.
By 2015, the number of models available and annual units sold both grew about tenfold, to 27 available models and about 140,000 units (Table 1). With fuel efficiency
and environmental regulations becoming increasingly stringent, car manufacturers have
plans to add plug-in technology to most of their car portfolios. Car manufacturers have
developed BEVs with higher battery ranges and lower prices, such as the Tesla Model
3 and Chevrolet Bolt, both launched in 2017. The number of charging locations for
electric vehicles also grew ten-fold, with around 2,000 by the end of 2011 and around
4
Technology for electric vehicles has existed since the 1800s, but gasoline became the dominant fuel
by the 1920s. A confluence of advances in battery technology and tightening environmental regulation
has led to a revival of the electric vehicle market in recent years. See U.S. Department of Energy
(2014) for a detailed account of the history of electric vehicles.
5
20,000 by the end of 2015.
Battery range and charging infrastructure are crucial for electric vehicles, analagous
to tank size and availability of gasoline stations for gasoline cars. Battery range,
the distance that an electric vehicle can travel starting with a fully charged battery,
generally increases with the size of the battery. However, other factors, such as weight,
aerodynamics, and anything else that impacts fuel efficiency also determine range.
All electric vehicles can be recharged by plugging into an ordinary electrical outlet,
so in contrast with cars of other fuel types, such as gasoline, hydrogen, or natural
gas, dedicated refueling infrastructure may not seem obviously necessary. However,
the ordinary outlet is very slow; it may be a reasonable option for overnight charging
at home, but for travel distances that exceed the battery range, drivers need faster
charging options away from home.
There are three speeds of charging options, increasing in power output and fixed
costs of installation. Level 1 are the ordinary wall outlets used by most other consumer
electronic devices. Level 2 charging stations can fully charge an electric vehicle in four
to six hours, which make them suitable for destinations where drivers may park for
a while. In residential homes, they can be attached to the outlet typically dedicated
to laundry dryers and electric ovens. Some employers and owners of shopping malls,
restaurants, and hotels have installed Level 2 charging stations as an amenity to their
employees and customers. However, sessions lasting four to six hours are too long for
many long-distance trips or for drivers who will not otherwise be parked for so long.
The fastest charging stations are called Level 3, or direct-current (DC) fast chargers.
These charging stations work in conjunction with a transformer to deliver high-power,
DC electricity to vehicles. A 30-minute charge session can refuel a battery by 80%.
Level 3 charging stations require the highest fixed costs out of all speeds because of
the transformer and higher permitting, legal, and electrician labor costs.
2.2
Charging standards and compatibility policy
Recognizing the importance of a fast refueling infrastructure for electric vehicle sales
and due to the dearth of available Level 3 charging stations, automakers have entered
the charging network market. Automakers have coalesced around three different Level
3 charging standards, each not compatible, or interoperable, with the others (Figure
1). In contrast, Level 1 and 2 charging standards are uniform across all vehicle brands
6
and have been built by employers, business owners, and government programs.5 A
charging standard has two parts: (i) a set of electronic communications between the
vehicle and the charging station, and (ii) a physical connector.
Car manufacturers only began building the accompanying charging stations after
the launch of fast-charging-capable BEVs, which suggests that firms invest in charging
stations in order to boost vehicle sales (Figure 2). Nissan, in partnership with the Tokyo
Electric Power Company and other Japanese automakers, developed the Chademo
charging standard in 2010, at the same time as the development and release of Nissan’s
BEV, the Leaf. Tesla Motors announced in September 2012 that it would build a
Supercharger network to blanket the U.S., three months after the first delivery of
Tesla’s BEV, the Model S. Meanwhile, other car manufacturers, working through the
Society of Automotive Engineers (SAE), released the specifications of the SAE J1772
Combo6 standard in October 2012. However, no cars were marketed under the Combo
standard until the release of BMW i3 in May 2014. Two months later, BMW announced
that it would build charging stations under the Combo standard.
Incompatibility in fast-charging protocols is a topic of vigorous policy debate and
a potential source of social inefficiency. The European Union Parliament, with the
objective of achieving a single charging protocol, ruled that all stations built after 2018
must at least be compatible with their chosen standard. In other words, multiple standards are allowed on each station via connectors or adapters (European Commission
(2014)). Optimal policy regarding compatibility is an open empirical question.
2.3
Government subsidies and ZEV regulation
Policymakers around the world and across levels of government have been supporting
the growth of the electric vehicle market for a variety of reasons, including environmental benefits and innovation spillovers. Government subsidies and ZEV (zero-emissions
vehicle) mandates have played a crucial role in the growth of the electric vehicle industry.
5
The underlying reason for the lack of entry in building and operating charging stations to sell
electricity for profit remains an important question for future research. One plausible explanation is
that the size of the electric vehicle fleet does not provide enough revenue relative to the fixed costs of
building a charging station.
6
Tesla vehicles may be sold with a J1772 adapter, but “J1772” without the “Combo” modifier is
only the slower, Level 2 portion of the charge port. The “J1772 Combo” has as two DC pins under the
regular J1772 port, hence the “Combo”. See Figure 1 for a depiction of each standard’s connectors.
7
I use panel variation in government subsidies for electric vehicle purchases as instruments for demand estimation. Electric vehicles face two main barriers to higher market
shares: they are more expensive than comparable gasoline cars due to battery manufacturing costs, and they lack adequate refueling infrastructure. Federal income tax
deductions in the U.S. for purchasing an electric vehicle range from $2,500 to $7,500,
depending on the size of the battery. State income tax deductions on top of the federal
incentives range from $250 to $7,500.
I construct two instruments from policy variation to identify the charging station
elasticity of demand. The first instrument is a cost shifter. State governments subsidize charging stations, ranging from 10% to 50% of costs. The state subsidies for
charging stations target businesses such as supermarkets, employers, rest-stops, and
car manufacturers. The second instrument is the number of new stations in a city that
are part of government-funded stimulus projects. In 2009, the American Recovery and
Reinvestment Act (Recovery Act) allotted $100 million to the Department of Energy
to build charging stations. I discuss the identifying assumptions central to the validity
of these instruments in Subsection 4.2.
Zero-emissions vehicle (ZEV) mandates designed by the California Air Resources
Board require a growing percentage of automakers’ overall sales to be zero-emissions.7
Battery electric, plug-in hybrid electric, and hydrogen fuel cell vehicles satisfy ZEV
regulations. The mandate is implemented and enforced through a tradable credit system. An automaker is assigned a credit requirement each year based on total sales
volumes and that year’s ZEV percent requirement. Each qualifying vehicle sold generates credits according to a formula that takes into account the battery range and other
characteristics. For example, plug-in hybrids generate fewer credits than pure battery
electric vehicles. Automakers are allowed to bank any excess credits toward future
years as well as trade credits with other automakers. Although ZEV credit prices began to fluctuate in 2016, the maximum price of $5,000 was binding in the data period
of this paper, 2011-2015. I use $5,000 as the value of each ZEV credit and include the
value of ZEV credits in the firms’ profit function.
7
Under the Clean Air Act, states can choose whether to follow emissions regulations in California.
California implements these mandates along with 9 other states. As of 2016, there are ten states with
ZEV mandates: California, Connecticut, Maine, Maryland, Massachusetts, New Jersey, New York,
Oregon, Rhode Island, and Vermont.
8
2.4
Data and descriptive statistics
My empirical analysis uses a dataset with five main elements. First, market-level
information on consumer demand for cars comes from registrations of new vehicles,
compiled by IHS Automotive (formerly R.L.Polk). These registrations are collected
by each state’s department of motor vehicles and accurately reflect new car purchases.
The dataset reports the number of registrations by car model, geographic area, and
quarter. Each car model is defined as a brand, model name, model year, and fuel
type. I use MSA delineations to define geographic markets. The panel includes 365
MSAs and 20 quarters, from 2011-2015. Second, the car quantities data are merged
into model-level characteristics information from MSN Auto, the Environmental Protection Agency, and Automotive News, including manufacturer-suggested retail price
(MSRP), manufacturer incentives, battery capacity, and fuel efficiency. The price that
enters the firm profit function is the MSRP less manufacturer incentives. The price
facing consumers is MSRP less manufacturer, federal, and state incentives. Third, I
collected panel data on federal and state subsidies described in Section 2.3. Fourth,
charging station investment data, including opening date, location, speed, and standard
are published by the Department of Energy’s Alternative Fuels Data Center. Fifth,
the National Household Travel Survey and the American Community Survey provide
information on consumer heterogeneity in commuting flows and income.
3
3.1
Model
Model overview
My model consists of two main parts: consumer vehicle choice and car manufacturer
profit maximization. The demand model is static in that consumers leave the market
after their product choice and do not purchase again. It takes into account geographic
variation in availability of charging stations and consumer heterogeneity in origin and
destination of driving trips.
Car manufacturers play a series of static stage games. In each period, they first
choose investments in charging stations to arrive in the next period. They next set
prices conditional on the charging stations installed thus far and realized consumer
demand shocks. Each period features the following sequence of events:
0. Station investments from the previous period arrive.
9
Vehicle models from exogenous R&D arrive.
1. Firms choose charging station investment.
2. Consumers realize demand shocks.
3. Firms set prices given demand shocks.
4. Consumers choose a vehicle to purchase.
3.2
Consumer demand
The main purpose of the consumer choice model is to predict the demand response
to alternative quantities and locations of electric vehicle charging stations. I use a
discrete-choice model following the framework of Berry et al. (1995) and Petrin (2002).
Each period, consumers arrive at the market to purchase one of the inside goods, a
plug-in car, or the outside good, a non-plug-in car. The demand model is static in that
consumers choose myopically, without taking into account the future evolution of prices
and other product characteristics, discussed in more detail later in this subsection.
Therefore, the outside good does not include the option value of making the vehicle
purchase decision in the future.
Consumer i chooses a vehicle r in market m and period t. Consumer utility from
choosing one of the inside goods depends on consumer attributes and vehicle characteristics. It is given by:
Uirmt = δrmt + µirmt + εirmt ,
where δrmt is the mean utility common to all consumers within a market and period,
µirmt are mean-zero individual deviations from mean utility, and εirmt are idiosyncratic
tastes assumed to be i.i.d. logit.
Consumers derive mean utility, δrmt , from purchase price, prmt , charging station
access to be detailed later in this section, other characteristics, Xrmt , which includes a
constant, and unobservable characteristics, ξrmt :
δrmt = α log(prmt ) + γ S frm (Gt ) + γ L gr (Gt , br ) + Xrmt β +ξrmt .
| {z }
{z
}
| {z } |
price
charging network quality
other chars.
The individual deviations from mean utility, µirmt , depend on consumer attributes
10
income, yi , and average daily driving distance, di :
µirmt = yi (σyp log(prmt ) + σyS frm (Gt ) + σyL gr (Gt , br ))
+ di (σdp log(prmt ) + σdS frm (Gt ) + σdL gr (Gt , br )).
Demand parameters, θ = (θ1 , θ2 ), can be categorized into the ‘linear’ parameters,
θ1 = (α, γ S , γ L , β), and ‘nonlinear’ parameters, θ2 = (σyp , σyS , σyL , σdp , σdS , σdL ).
Consumers pay a purchase price prmt , which is equal to MSRP less manufacturer
discounts (MD) and state and federal subsidies:
prmt = MSRPrt − MDrt − State Subsidyrmt − Federal Subsidyr .
(1)
MSRP and manufacturer discounts are the same across all markets and only vary
across vehicle models and time. State subsidies vary across vehicle models, markets,
and time, and federal subsidies vary across vehicle models. Sallee (2011) finds that
consumers capture the full federal and state incentives for the conventional hybrid car,
the Toyota Prius, while Busse et al. (2006) find that manufacturer discounts are incompletely passed-through to consumers. Busse et al. (2006) hypothesize that subsidy
pass-through increases with how much consumers know about the subsidies.
I will instrument for price in estimation, to be discussed in more detail in Subsection 4.2. If prices are measured with error, the price coefficient estimates will be valid
if the instruments are uncorrelated with the measurement error. For example, if subsidies are passed through incompletely to consumers due to imperfect competition, then
price coefficient estimates are valid if quarter and market fixed effects control for imperfect competition, or if the instruments are uncorelated with the time-market-varying
component of imperfect competition.
The network size for each standard grows over time, and Nissan’s Chademo network
vastly outnumbers the other two standards (Figure 2). However, a map of charging
locations by standard shows the stark difference in how charging locations are distributed relative to urban centers and highway corridors (Figure 3). Tesla stations
span the U.S. interstate highway system, while in contrast, Chademo and Combo stations cluster near urban areas. I model the value of charging networks as increasing
in the number of charging locations as well as location match quality with consumers’
driving needs.
Consumers in the model take two types of trips: (i) local travel within the con-
11
sumer’s city of residence and (ii) long-distance travel between cities. In each period,
the set of installed charging stations Gt is mapped to utils by the local travel function
frm and inter-city travel function gr .
This paper focuses on the importance of charging infrastructure away from consumers’ homes. Data from the Electric Vehicle Project show that about 22% of charge
events occur away from home.8 However, the percentage of charge events away from
home is not necessarily proportional to its importance for electric vehicle adoption. I
will infer the importance of charging stations by estimating the electric vehicle demand
response to charging station arrivals.
Most drivers conduct all of their away-from-home charging at three or fewer charging locations. Drivers tend to charge at work, near commute destinations, and other
public locations such as grocery stores or shopping malls (Idaho National Laboratory
(2015))9 . Therefore, the local travel function frm assigns a population-weighted count
of the stations in set Gt , where the population weight is based on place of work rather
than place of residence. The function frm further distinguishes between the number
of Level 2 (slower) and Level 3 (fast) charging stations as well as the charging standard that vehicle model r is able to access. Consumers in the model have decreasing
marginal returns from additional stations, captured by the log functional form:
frm (Gt ) =
X

l2
l3
wc log(Nrct
) + log(Nrct
) ,
c∈Cm
where N l2 and N l3 are the number of Level 2 and Level 3 charging stations, respectively.
The population weights wc for each county c in the set of counties Cm in each market
m (counties are wholly contained within MSAs, and geographic borders of counties
align with MSA borders) are defined as:
wc =
Population of market m who drive to county c for work
.
Total population in market m
This simple specification for the local travel function captures realistic and desirable
spatial properties. Consider a city with residential counties surrounding a commercial
core where everyone drives in for work. The commercial core would be the most
8
The Electric Vehicle Project collected data on charging and driving behavior from about 5,800
Nissan Leaf and Chevrolet Volt drivers from 2012 to 2013.
9
See Hardman et al. (2018) for a review of consumer interactions with electric vehicle charging
infrastructure.
12
useful place for a charging location, because home charging is a superior substitute for
charging locations in the residential area. Next, consider how public transit substitutes
for driving. If few people drive to work in an MSA or to a particular county within
an MSA because of ample public transit infrastructure, then the model would assign a
low value to electric vehicle charging locations via a low numerator in the population
weights wc .10
In addition to enabling top-ups for commutes and errands, the set of stations Gt
may also form a network that enables inter-city travel. The inter-city travel function gr
counts the number of city pairs that are connected by paths of Level 3 (fast) charging
stations. Two cities are connected for vehicle r if a set of stations matching vehicle r’s
charging standard traces a path between the cities and if the distance from one station
to the next along the path is less than the battery range br of the vehicle:
gr (Gt , br ) = Nrtcity pairs .
To rule out unreasonable routes, gr only counts a city pair as traversable if the path
of Level 3 charging stations is at most 30% longer than the as-the-crow-flies distance.
In estimation, I normalize gr to be between 0 and 1. Tesla’s charging locations outside
MSAs are advertised as built for the purpose of enabling long-distance driving. The
relationship between the number of connected cities and the cummulative number of
charging locations is remarkably linear, as shown in Figure 4.11
I do not explicitly model the consumer costs of charging station access for three main
reasons. First, Level 3 stations, the focus of this paper, are often free to use for the life
of the car (Tesla) or for the first two to three years after purchase (Nissan and BMW).
For these owners whose vehicles come bundled with free charging, the access price is
zero. Second, Level 2 stations are primarily offered by third-party operators with a wide
variety of nonlinear pricing schemes based on the number of minutes plugged in or the
amount of electricity served. Some employers, retailers, and municipal governments
subsidize access to particular stations on these third-party networks. Without any
detailed data on charging station access pricing, a uniform scaling factor applied to
10
My specification does not incorporate other features of the charging network, such as the dispersion
of charging stations within counties. Constructing measurements of other features is feasible. However,
demand parameters for these other features would be difficult to econometrically identify convincingly,
and they would also make the firms’ charging station investment decision computationally infeasible.
11
The long-distance charging network build-out problem is slightly different from canonical graph
theory problems, because two nodes (MSAs) only count as connected if the path between them is not
“too much” longer than the shortest possible path.
13
all stations to account for an access charge would be absorbed by the constant in the
utility function. Third, non-pecuniary costs of accessing charging stations – such as
search, travel, and hassle costs – are implicitly built into the functions frm and gr .
Firms’ investments in charging stations during each period may be correlated with
the unobserved product characteristic, ξrmt . For example, locally targeted advertising
is unobserved and may be correlated with firms’ investment choices. Moreover, firms
may choose to build in markets with particularly low or high realizations of ξrmt . I
address the endogeneity of price and investment in charging stations by instrumenting
for both variables with a panel dataset of government subsidies, described in Section
2.3. Section 4.2 presents the formal identifying assumptions in more detail.
Modeling the vehicle purchase decision as static may be reasonable due to unique
features of the electric vehicle market from 2011-2015. First, consider the scenario that
consumers wait to purchase because they want the better products that will arrive
in future periods, as in Gowrisankaran and Rysman (2012). Electric vehicle models
on the market did not drastically change between 2011 and 2015. Due to limited
advancements in battery chemistry and manufacturing processes, vehicles by the end
of 2015 had similar range and prices as models released in 2011. Consumers who chose
the outside option, a non-plug-in car, for the option value of a better electric vehicle
in the future would have had to wait until 2017 for significantly longer battery range
at lower prices, such as the Chevrolet Volt, 2018 model-year Nissan Leaf, and the
Tesla Model 3. The most meaningful change in product characteristics, availability of
charging stations, accrued to all electric vehicles that had been purchased. Therefore,
the option value of waiting to buy a plug-in vehicle later may be limited.
A second justification of the static consumer demand model is that consumers cannot easily move homes or change workplaces in the short term. Therefore, purchasing
a vehicle earlier due to anticipated future charging network improvements would imply
an implausibly low weight on present-day driving needs. The static demand assumption can be interpreted as consumers placing predominant weight on the present, rather
than assuming that consumers do not expect the charging network to continue to improve. A formal test of the importance of dynamic considerations will only be possible
after the electric vehicle market has existed for a longer time.
14
3.3
Car manufacturer investment
My model endogenizes firm choices in the quantity and locations of charging stations,
conditional on the standards coalitions that they have joined. These two control variables are part of a dynamic optimization problem that may be driven by firms’ expectations of future periods beyond the time coverage of the dataset available. Therefore,
my model of firms is static. I assume that conditional on the choice of standard and
the charging stations that have already been installed, the static profit function is proportional to the dynamic value function, so that optimization from the static model is
consistent with a long-run dynamic game.
I also assume that vehicle characteristics other than price and charging stations
evolve according to an exogenous product development process. This is a reasonable
assumption given that the data period of 5 years from 2011 to 2015 is short relative
to the product development cycle in the automotive industry overall and particularly
in the electric vehicle segment. Blonigen et al. (2019) show that over vehicles of all
fuel types, 70% of models are redesigned every 4 to 7 years, and an entirely new model
takes even longer. The earliest electric vehicle models released in the 2011 model year
came out with major updates only in late 2016, which is after the end of the data
period.
The profit πjt (Gt ) of firm j in period t from its electric vehicle models r ∈ Jjt is
the sum over markups from cars sold minus the cost of charging station investment
c(ajt ), given by


X X

(prt − mcrt + ZEVrmt ) srmt (Gt , prmt ; Xrmt , ξrmt , θ) Nmt  − c(ajt ),
{z
}|
{z
} |{z}
|
m r∈Jjt
per-car markup and ZEV credit
mkt share
(2)
mkt size
where mcrt denotes the marginal cost of producing car r in t.
In each period, firms simultaneously choose charging infrastructure investment ajt
that will arrive at the beginning of the next period, incurring cost c(ajt ) and conditional
on all the stations that have already been installed. Then, firms set prices prt to
maximize profits. I model firms setting one price for the country for each model r and
period t, prt , which is the MSRP minus manufacturer discounts. The firm’s price prt
does not include government subsidies that are contained in the consumer-facing price
prmt , defined in Equation 1. I do not observe and therefore do not take into account
the region-specific discounts and dealer-specific nonlinear picing contracts.
15
Firms choose the number of stations to allocate across 365 local networks and the
inter-city network, so the infrastructure investment choice ajt is a vector with 366
elements. The cost of ajt is assumed to be linear in the total number of stations, given
by,
c(ajt ) = κ|ajt | + ωjt .
I model firms as maximizing profits from plug-in vehicles, ignoring externalities
on other products in a firm’s portfolio, due to data availability and computational
constraints. Two main arguments can justify this assumption. First, many car manufacturers have set up divisions dedicated to new electric models with unique physical styling. The organizational economics literature has studied when decentralized
decision-making is optimal for multi-product firms with asymmetric product divisions
(Rantakari (2008) and Roberts and Saloner (2012)). Second, institutional details suggest that the three firms explicitly modeled for charging station investment, BMW,
Nissan, and Tesla, had zero or very little cannibalization in other segments over the
time period of this study.12 The remaining firms had not built charging stations by
the end of 2015.
4
Estimation, Identification, and Results
In this section, I describe the estimation of the demand and cost parameters and how
they are identified. I address the problem of zero market shares by shrinking the data
toward an empirical Bayes prior formed over similar markets. This procedure pulls
the market shares away from zero, which is important in order to apply the estimation
framework of Berry (1994) and Berry et al. (1995). Readers who are not interested in
the technical details of the empirical Bayes procedure can skip directly to Subsection
4.2 for how the demand parameters are identified and Subsection 4.3 for demand and
supply estimation results.
12
Tesla Motors sells only electric vehicles and invests heavily in its network of charging stations.
BMW, the active firm in the SAE Combo standard, stated in a press release after selling the i3
for two years that more than 80% of worldwide i3 customers were new to the BMW Group (2015.
https://www.press.bmwgroup.com/global/article/attachment/T0242822EN/337735). Nissan began designing the Leaf in 2006 to leapfrog the conventional hybrid car, a segment in which they were
not competitive (Burgelman and Schifrin (2011)). Since then, Nissan has prioritized being the industry leader in zero-emissions and electric vehicles by championing the Chademo standard and investing
in charging infrastructure worldwide (Nissan Motor Corporation (2012)).
16
4.1
Zero market shares
This paper studies the U.S. electric vehicle industry from its inception, when new car
models initially sold zero quantities in some local markets. The dataset covers all
new vehicle registrations for each market and period, so any observed zeros are not
due to sampling error, such as from disaggregating a national sample or survey to the
local level. As described in McFadden (1974) and Berry et al. (1995), each consumer’s
choice is an independent draw from a multinomial distribution with a set of purchase
probabilities. The observed market share aggregates over the sampled consumers’
multinomial draws. Each market is finite even when the consumer sample is the full
population, and coupled with small purchase probabilities, the observed market shares
include zeros. In my sample, 35.7% of market shares in any given model-market-quarter
combination are 0, ranging from 15.5% (2011) to 45.5% (2015), as shown in Table 2.13
The true purchase probabilities underlying the observed market shares are unknown. Common practice in demand estimation is to use the observed market shares
in place of the true purchase probabilities, which is implicitly using the maximum
likelihood estimator (MLE). Zero market shares are censored at zero and therefore
mask information about the true underlying purchase probabilities. They also make
the inversion step impossible in the Berry (1994) and Berry et al. (1995) estimation
framework. I instead use a parametric empirical Bayes or shrinkage estimator, which
generates strictly positive posterior estimates of the true purchase probabilities from
information in other markets. This is similar to the transformation in Gandhi et al.
(2017). To preserve important heterogeneity across markets, each market’s empirical
Bayes prior is formed using similar markets. I define the set of similar markets to be
the the 50 markets closest in income per capita, or about 13.7% of the 365 total number
of markets.
I model the quantities purchased of each vehicle in each market, Krm , as a draw
from a binomial distribution with Nm trials and purchase probability s0rm . The time
subscripts t have been suppressed throughout this subsection for simplicity. The purchase probabilities s0rm are different for each vehicle and market and are drawn from
a Beta prior distribution with hyperparameters λ1rm and λ2rm . The total number of
vehicles purchased is Nm . I choose this Beta-Binomial model of market shares for
13
The number of zeros increases over time because the number of available plug-in models increases.
17
simplicity, though it can be generalized to a Dirichlet-Multinomial:
Krm ∼ Binomial(Nm , s0rm ),
s0rm ∼ Beta(λ1rm , λ2rm ).
The posterior distribution of the purchase probability is also a Beta distribution,
srm ∼ Beta(λ1rm + Krm , λ2rm + Nm − Krm ),
with posterior mean given by,
ŝrm =
λ1rm + Krm
.
Nm + λ1rm + λ2rm
The observed shares, which are the MLE, are,
LE
ŝM
=
rm
Krm
.
Nm
The strictly positive posterior mean, ŝrm , replaces the MLE, which contains zeros. In
large samples, the empirical Bayes posterior would be very similar to the observed
shares because the data would be informative enough to dominate the prior from other
markets.
For each car r in market m, the Beta prior are formed using the 50 markets closest
in per capita income, l ∈ Bm , where l is a market from the set of similar markets Bm .
The parameters of the Beta prior, λ1rm and λ2rm , are estimated from maximizing the
log of the likelihood over the outcomes in the markets that form the priors,
Y Krl Γ(λ1rm + λ2rm )Γ(λ1rm + Krl )Γ(Nl − Krl + λ2rm )
f (Krl , l ∈ Bm |λ1rm , λ2rm ) =
.
Nl
Γ(λ1rm )Γ(λ2rm )Γ(Nl + λ1rm + λ2rm )
l∈B
m
I estimate a pair of hyperparameters λ̂1rm and λ̂2rm for each vehicle, market, and
period, and construct the posterior mean estimate of purchase probabilities, ŝrm =
λ̂1rm +Krm
. As expected, the posterior estimates of market shares have lower variNm +λ̂1rm +λ̂2rm
ance, and all shares are strictly positive, as shown in the bottom panel of Table 2.
Reassuringly, the means of the observed and empirical Bayes posterior market shares
are quite similar, .00085 and .00082, respectively. Observed zero market shares have
posterior mean estimates ranging from 5.5e-9 to .00162, as depicted in Figure 5(b).
18
Berry et al. (2004) provide conditions on the number of consumers relative to the
number of products for consistency and asymptotic normality of the demand estimates
when using the MLE estimator as true purchase probabilities. I assume that the same
conditions hold when using the empirical Bayes estimator. Appendix A discusses the
advantages of the empirical Bayes estimator over other common methods.
4.2
Identification
Firm investments in charging stations each period may be correlated with unobserved
product characteristics. Therefore, additional instruments are required to identify the
demand parameters compared to the usual instruments for price. I maintain the standard assumption that other product characteristics besides price and charging network
are exogenous. I first discuss how I identify the price coefficient using variation from
government subsidies. The identifying assumption is that the vector of instruments
Z price is orthogonal to unobserved characteristics ξ(θ2 ),
E[Z price ξ(θ2 )] = 0.
(3)
I use three sets of instruments for price that are plausibly uncorrelated with unobservable characteristics ξrmt . The first two sets of instruments are federal and state
subsidies. Since prices that consumers pay also include time-varying manufacturer discounts, the federal and state subsidies used as instruments are not the sole sources of
variation in price. The third set of price instruments, the average characteristics of
other products in the market (BLP instruments), are relevant because they affect the
markups that firms can charge. The BLP instruments are uncorrelated with ξrmt given
the assumption that the other product characteristics arrive as part of an exogenous
development process.
Market and time fixed effects are included for all specifications. The instruments
vary within market, over time, and across vehicle models. Market fixed effects control
for local factors that do not vary much from 2011 to 2015, such as local inclinations to
be green, the proportion of housing stock with off-street parking (and enable at-home
charging), the types of electrical wiring in the housing stock, and quality of public
transit. Time fixed effects control for national factors that do not vary across markets,
such as national macroeconomic trends and global fuel price shocks.
Federal plug-in vehicle subsidies vary by car model and are determined by a piece-
19
wise linear function of the battery size. This instrument provides identification from the
functional form, which I argue is uncorrelated with unobservables ξrmt after conditioning on battery size. It is reasonable to assume that policymakers set federal subsidies
independently from ξrmt , because Congress approved and determined the structure of
these plug-in vehicle subsidies as part of a stimulus package in 2009. I also assume
that firms do not determine components of ξrmt based on the subsidy functional form.
State plug-in vehicle subsidies vary by state and vehicle model and change over
time. With market and time fixed effects, the identifying assumption is that changes
in state subsidies over time and differences in subsidies across states or across car
models within states are uncorrelated with product unobservables ξrmt . Anecdotes
of the state legislative process support the identifying assumption that the timing of
subsidy changes are plausibly random. For example, some states enact laws that are
effective immediately, others enact laws that are effective for the next tax (calendar)
year beginning in January, while others enact laws that are effective for the next fiscal
year beginning in July. The structure of state subsidies are also plausibly exogenous
after controlling for characteristics that these subsidies condition on, such as battery
size. Appendix B shows the variation in state subsidies for electric vehicles and charging
stations over markets, time, vehicle models, and charging speeds.
Additional assumptions on the distribution of unobservable characteristics ξrmt and
the instruments are necessary to identify the charging station coefficients. I assume that
unobserved product characteristics ξrmt evolve according to a first-order autoregressive
process,
ξrm,t (θ2 ) = ρξrm,t−1 (θ2 ) + νrm,t (θ2 ),
that νrm,t are mean zero, independent across vehicle models r, markets m, and time
periods t, and that
E[Z station ν(θ2 )] = 0
(4)
for a vector of instruments Z station .
I use three sets of instruments for charging stations that are plausibly uncorrelated with innovations in demand unobservables, νrmt . First, similarly to the vehicle
price subsidy instrument, state subsidies for charging stations are cost shifters that are
uncorrelated with demand shocks conditional on market and time fixed effects.
The second set of charging station instruments are the number of new stations that
are funded by the Recovery Act of 2009. As described in Section 2.3, recipient cities
20
were chosen before electric vehicles arrived to the U.S. market. Each city received the
same number of stations predetermined by program funding availability, independent
of the realized evolution of the electric vehicle market in each city. Regulators may
have chosen recipient cities where they expected the highest growth rates or marginal
impacts in local electric vehicle adoption. However, the exact timing of stations arriving
in each recipient city could be due to idiosyncratic permitting and construction lags that
are plausibly uncorrelated with νrmt . In the data, Recovery Act stations arrive between
the 2nd quarter of 2011 and the 2nd quarter of 2014. Similarly, car manufacturers
may invest differently in response to the Recovery Act charging stations. If so, the
assumption that stations take at least one period to be built means that the arrival of
car manufacturers’ stations are uncorrelated with contemporaneous νrmt .
The third set of charging station instruments are the one-period lags of the charging
station quality variables. The stations arriving in the beginning of period t were chosen
by car manufacturers based on ξrm,t−1 , before νrm,t were realized. Therefore, new
stations arriving in period t are uncorrelated with νrm,t .14
4.3
Estimation Results
Demand parameters are estimated using a GMM framework with moment conditions
in Equations 3 and 4. Table 3 reports results from the logit model (Columns 1 and 2)
and a random-coefficients logit model (Columns 3 to 5). For both logit and randomcoefficients logit demand, the coefficients are positive for battery range, capacity, horsepower, and all-wheel drive, as expected. The coefficient for the BEV dummy variable
is negative, indicating that BEVs are less preferable than PHEVs. A plausible reason
is that consumers like having gasoline as a backup fuel source.
There are seven endogenous regressors: price and interactions of three measures of
charging network quality (Local Levels 2 and 3 and # of City pairs) and two vehicle
fuel types (PHEV and BEV). Column 1 shows OLS results, and Column 2 shows
results for the IV specification with instruments as described in Subsection 4.2. The
first-stage minimum eigenvalue statistic, the analog to the first-stage F-statistic for
multiple endogenous regressors, has a value of 59.42, indicating strong instruments
(Stock and Yogo (2005)).
The price coefficient can be directly interpreted as a price elasticity due to the log
14
The single event of Tesla introducing an adapter to Chademo stations occurred in April 2015.
Without any panel variation in adapter availability, this event is absorbed by the time fixed effect.
21
specification. The OLS and IV specifications yield similar price elasticities (Columns 1
and 2). A price elasticity of -2.7 from the IV specification is in line with prior literature
on the automobile industry (Berry (1994), Berry et al. (1995), and Goldberg (1995)).15
Station locations and quantities are endogenously chosen by firms, so OLS estimates
of the parameters on charging network quality may be biased. Instrumenting for the
endogenous regressors increases the precision and magnitude of the coefficients.
Comparing the PHEV and BEV interactions shows that the availability of fast
charging stations (Level 3) for both local driving and inter-city travel matters more
for BEVs. In contrast, Level 2 charging seems to be more important for PHEVs. A
plausible explanation is that for PHEV drivers, gas stations are superior substitutes
for Level 3 stations for fast refueling. Therefore, PHEV demand may not respond to
Level 3 charging station availability as much as BEV demand. PHEV consumers may
see the Level 2 local charging network as a way to opportunistically top up, while BEV
consumers see the Level 2 network as poor insurance for running out of electricity
because they are so slow. Consumers may suffer from so much range anxiety that they
disregard the Level 2 network and only purchase a BEV if they are certain the battery
range can cover the majority of their day-to-day needs. Therefore, PHEV demand may
be more responsive to Level 2 charging station availability than BEV demand.
Using the price elasticity constructed from the price coefficient and random coefficients (Columns 3 to 5), I compute the markups and marginal costs of vehicles implied
by the first-order condition of the firm profit function (Equation 2). Table 4 reports
the resulting estimates. I estimate that vehicle markups range from about $7,500 at
the 10th percentile to about $26,000 at the 90th percentile, in line with gross margins
reported by car manufacturers in financial filings. Vehicle marginal costs range from
about $17,000 at the 10th percentile to about $57,000 per vehicle at the 90th percentile.
I recover charging station costs from the first-order condition of the firms’ profit
function with respect to charging stations. My cost estimate also includes the discounted present value of the electricity that car manufacturers often offer for free for
the first few years of ownership or for the life of the car, but I cannot separate the
capital cost and electricity cost components due to data constraints. I estimate that
Level 3 charging stations cost about $10,000 per year on average, implying an upfront,
15
Li et al. (2017) find a much smaller price elasticity of .61 and a charging station elasticity of .84
with a specification that includes product fixed effects. The difference in our estimates is driven by
differences in the conditional price variance. The instruments I use to address price endogeneity only
require market and time fixed effects to be valid, so I do not include product fixed effects.
22
discounted present value of $143,000 per station using a 7% weighted average cost of
capital. This estimate of charging station fixed costs is in line with engineering estimates and rumors in the electric vehicle industry that a Level 3 station costs range
from $50,000 to $150,000.1617
5
Uniform Charging Standard
I evaluate market outcomes and welfare in the counterfactual policy regime of a unified
charging standard for electric vehicles. In each period, firms play a simultaneous-move
game as described in detail in Section 3. Firms choose where and how many new charging stations to build, which take one period to complete. Consumers take into account
the available charging network when they choose between plug-in vehicles and the outside option to maximize utility. To find an equilibrium of the firms’ simultaneous-move
game, I simulate firms playing iterated best-response until no firm has any profitable
deviations. In each iteration, a firm conditions on the charging station investments
built by its competitors and itself in previous iterations to solve its charging station
investment problem, which is equivalent to the “fractional knapsack” problem and can
be solved with a fast, greedy algorithm (Subsection 5.2).
I present the counterfactual results in three parts in order to build intuition. Readers who are not interested in technical details of the knapsack problem can skip to the
final results in Subsection 5.3. First, I compute only demand response to a single charging standard, with the number and locations of charging stations held fixed. Second,
taking into account vehicle demand responses to charging stations, firms re-optimize
the geographic placement of stations when there is a single standard, given a fixed
number of stations. Third, firms optimize over the number of stations in each period,
taking into account geographic placement decisions, demand response, and competitors’ responses. Throughout the counterfactual analysis, I assume that automakers do
not change other vehicle characteristics, including price, whether cars are capable of
fast charging at all, and the battery range of each car. Therefore, vehicles are still
differentiated products in the counterfactual, with different charging networks depend16
Etherington, Darrell.
2013.
“Inside Tesla’s Supercharger Partnere Program: The
Costs And Commitments Of Electrifying Road Transport.”
Tech Crunch, July 26.
https://techcrunch.com/2013/07/26/inside-teslas-supercharger-partner-program-thecosts-and-commitments-of-electrifying-road-transport/
17
Holland, Ben. 2014. “Pulling Back the Veil on EV Charging Station Costs.” Rocky Mountain
Institute Blog, April 29. https://rmi.org/pulling-back-veil-ev-charging-station-costs/
23
ing on whether they are capable of fast charging and whether their range covers the
distance between pairs of charging stations.
I remain agnostic about how to achieve compatibility. One policy option is to
mandate or subsidize R&D for adapters for physical connectors and interoperability
of the communication protocols. A stronger policy is the European Union’s rule that
any new stations must at least contain a particular standard.18 All the policy options
besides creating adapters would require retrofits of existing stations. My estimated
welfare impacts serve as an upper bound on the coordination, R&D, and retrofit costs
that society would be willing to pay to achieve compatibility in electric vehicle charging
in the U.S., in the confines of my model and its assumptions.19
As shown by Small and Rosen (1981) and Williams (1977), the change in consumer
surplus in any counterfactual scenario from a comparison scenario is given by:
Z
∆CS =
i
1
dui /dyi
”
ln
J
X
!
exp(δr1 + µ1ir )
j=1
−
ln
J
X
!#
exp(δr0 + µ0ir )
dF (yi , li ), (5)
j=1
where dui /dyi is the marginal utility of income. Social welfare is the sum of consumer
welfare and producer profits.
5.1
Compatible stations with fixed quantities and locations
This subsection presents two results from simulating access to stations of other standards, holding the quantities and locations of charging stations fixed at the status
quo. First, I calculate the increase in consumer surplus from giving Tesla vehicles
access to Chademo stations, which can be compared to the retail price of a one-way
adapter developed by Tesla. Second, I present the model’s predicted demand response
to compatibility across all standards.
Chademo is the dominant and de facto standard in Japan. Tesla developed a oneway adapter to give Tesla vehicles access to Chademo stations, likely for the Japanese
market, but in March 2015 also released the adapter in the U.S. market. I infer
based on a conversation with a Tesla engineer that the adapter took at least two years
of development. In the simulation, I assign Tesla vehicles a new charging network
consisting of all the existing Tesla stations as well as the Chademo stations. The
18
19
See Ferwerda et al. (2018) for details on the evolution of charging standards in Europe.
See Simcoe and Farrell (2012) for a discussion of paths toward compatibility.
24
charging network quality variables for all other vehicles are fixed to the status quo,
as are the arrival locations and times of all stations. I then calculate the change in
consumer surplus from the change in network access (Equation 5). The retail price
of $450 for a Tesla-to-Chademo adapter is very similar to the model’s prediction of
$426.49 increase in average consumer welfare. Interestingly, Tesla initially launched the
adapter for pre-order at $1000, and then quickly adjusted the retail price to $450. The
retail price need not equal the average consumer surplus change, but this comparison
shows that the demand model and parameter estimates predict sensible magnitudes for
welfare relative to actual market prices for a limited version of compatibility. Moreover,
the retail price for a one-way adapter can give us a sense of the order of magnitude in
value we might expect from compatibility.
The gains from compatibility are asymmetric because of differences in battery range.
Stations are more useful for vehicles that have the battery range to traverse the gaps
between them, as illustrated in Figure 6. Tesla cars, which have about 200 miles of
battery range, benefit from access to Combo and Chademo stations because they can
easily traverse the distance between all stations. However, other firms’ electric vehicles
have at about 80 miles of electric range and cannot traverse the distance between Tesla
stations, which are placed 100 to 150 miles apart.20 Comparing Combo and Chademo
networks, cars on the Combo standard gain more because they can access the much
larger Chademo network in this counterfactual.
Simulating demand response to a uniform charging standard while holding fixed
the quantities and locations of stations shows that sales of plug-in vehicles with fastcharging capability, or those that can use Level 3 stations, would increase by about
26,000 units (17.1%, Table 5). Car brands in the Tesla and Combo coalitions increase
sales, but sales of Chademo car brands decrease. The intuition for this result is that
the Chademo charging network is an important factor in generating the market shares
observed in the data, but in the counterfactual these cars lose the advantage of having
more than three times as many charging stations as the cars on other standards (Figure
2). Fast-charge-capable cars steal market share from other plug-in vehicles that cannot
fast-charge, though the majority of gains in market share for fast-charge vehicles come
from stealing market share from the outside good, the non-plug-in vehicles.
20
Even in Subsection 5.3, when firms re-optimize quantities and locations of charging stations,
firms benefit asymmetrically from placing stations closer together. Tesla has no incentive to place
their stations closer together for the benefit of its competitors.
25
5.2
Compatible stations with adjustment in locations
This subsection solves the firm location problem for charging station placement for any
given quantity of stations. First, I will show that each firm’s charging station allocation
problem, conditional on stations already installed by itself and its competitors, maps
to a computational problem called fractional knapsack. Therefore, a greedy algorithm
that chooses locations for stations in order of highest marginal profit gives the optimal
solution. Second, I will describe the equilibrium outcomes with three firms locating
stations in a static oligopoly game. The equilibrium outcomes are found from firms
moving in a predetermined order to play iterated best-response. An equilibrium is
reached when no firm has a profitable unilateral deviation.
A single firm’s discrete choice problem of allocating a given N stations across L
locations is computationally infeasible to solve by enumeration. For example, with

L = 366 locations and N = 285 stations to allocate, there are N +L−1
≈ 1.058 × 10192
N
possible arrangements.
Placing N stations across L independent locations is equivalent to the fractional
knapsack problem. In the knapsack problem, a thief robbing a vault finds n items.
Each item has a value and a weight, both integers. The thief wants to maximize the
value of his loot, but he can only carry W pounds in his knapsack. In the charging
station placement problem, firms maximize profits over their station location choices,
subject to the constraint of building at most N stations. The equivalent of an item
is the location and how many new stations will be built at that location. The total
number of possible items is n = L × N , because there are L total locations and each
location can receive up to N new stations. The ‘knapsack capacity’ of the charging
station allocation problem is W = N . The value of each station is its marginal profit
from car sales. The regular, or 0-1, knapsack problem requires that the thief take whole
items, while the fractional knapsack problem allows the thief to take parts of items.
Both versions of the knapsack problem can be solved in pseudo-polynomial time with a
dynamic programming algorithm. The key to mapping the charging station allocation
problem to fractional knapsack is that location profits are independent and that stations
have uniform weight of 1.
Independence in profits across local markets is given by how charging stations enter
the consumer utility function. The demand model specifies that consumers derive
utility only from stations inside their own market and from stations that help with
inter-city travel. In the model, consumers do not derive utility from any stations
26
within MSAs that are not in their home city, which could be a reasonable assumption
for a variety of reasons, such as not requiring any transit at the destination, avoiding
hassle from driving in an unfamiliar city, and using public transit or riding with others.
If only charging stations within a consumer’s home market affect utility, then building
stations in one market would not affect a firm’s profits in other markets. Therefore,
profits from new stations in local networks are independent across markets.
The inter-city network enters the marginal utility of stations in local networks.
An additional pair of connected cities changes consumer utility and firm profits from
an additional station in every local network. Therefore, to be precise, the charging
station allocation problem is equivalent to fractional knapsack conditional on the total
number of stations N and the number of stations allocated to the inter-city network,
N inter . Firms solve the location problem N + 1 times, once for each possible value of
N inter ∈ {0, · · · , N }.
The fractional knapsack problem has the greedy-choice property (Cormen et al.
(2009)) and can be solved with a greedy solution. Choosing items in order of highest
value-to-weight ratio yields the maximum-value knapsack. The greedy solution in the
charging station placement problem is, for each possible number N inter of stations
devoted to the inter-city network, rank the market-quantity combinations in order of
decreasing marginal profit and choose the N − N inter highest. Record total profits
for each N inter . The N inter and corresponding allocation that gives highest profits is
the optimal solution. Allocating stations across local markets is the “inner loop,” and
finding the number of stations to allocate to local markets and the inter-city network
is the “outer loop.”
Under compatibility, firms locate stations with higher dispersion across markets.
This result carries an intuitive interpretation. Consumers derive decreasing marginal
utility from additional charging stations of each type. When stations are incompatible,
each firm faces a separate decreasing marginal value curve. The first station that a
firm builds in a market carries high value, and firms tend to build stations in the same
high-profit markets. However, under compatibility, additional stations are worth less
if other firms have already built nearby. Therefore, firms build in more markets but
fewer stations in each market (Figure 7). The number of markets that have at least
one charging station increases from 179 to 339 out of a total of 365 markets, indicating
increased spatial dispersion of stations.
27
5.3
Full compatibility counterfactual
The full optimization problem to endogenize both quantities and locations can be solved
by nesting the optimal location problem within the optimal quantity problem. Given
a number of stations Nj to allocate, solve the location problem and record the profits.
Repeat for every Nj ∈ {0, . . . , N̄j }, where N̄j is a resource constraint, and choose the
Nj∗ and the associated location solution. The resource constraint bounds the search
space for computational ease and also reflects managerial or capital constraints that
firms may face. I choose the resource constraint to equal the number of stations that I
observe firms building in the data. Each period, firms move in iterated best-response
to choose optimal quantities and locations until an equilibrium is reached.
Table 6 presents and compares the counterfactual outcomes from each of the three
regimes: (1) Incompatible charging standards, (2) Compatible charging standards with
private charging investment, and (3) Compatible charging standards with Social Planner charging investment. Over 2011-15, private investment under compatibility results
in about $500 million higher consumer surplus than under incompatibility; the Social
Planner achieves about $3 billion higher consumer surplus than private charging investment. With a compatible charging network, both private investment and social
planner investment achieve higher consumer surplus and higher overall producer surplus. However, not all firms are better off with compatibility. Firms originally on the
Chademo standard, such as Nissan and Mitsubishi, lose profits under compatibility because they lose one dimension of product differentiation, which is a car with a smaller
battery coupled with ample local charging availability. Nissan loses about $200 million
from compatibility.
Compatibility changes the nature of competition among firms, turning investments
in charging stations from demand substitutes to demand complements. When a firm
builds a station under compatibility, it improves the product quality of its competitors.
Therefore, firms have less business-stealing motive to invest in charging infrastructure.
The social planner fully internalizes the business stealing effects, and builds about
330 fewer stations (a 17.7% decrease). Firms build about 50 fewer stations (a 2.8%
decrease) under compatibility. Notably, all the decrease in private investment comes
from Nissan. BMW and Tesla build up to their resource constraint in both regimes.
The number of electric vehicles sold increases by about 100,000 units under compatibility and private investment compared to under incompatible charging standards.
About 85,000 more electric vehicles would be sold if the Social Planner were to make
28
the investment choices instead. The vehicles that have fast-charging capability gain
market share overall by stealing market share from the outside option (vehicles of other
fuel types) and from other electric vehicles that do not have fast-charging capabilities
(“Other Plug-In” in the last row of Table 6). The environmental impacts of a compatibility policy would depend on population density and the electricity fuel mix where
consumers substitute to electric from other fuel types. My finding that some manufacturers would lose profits from moving to compatibility may explain why the automotive
industry has not standardized electric vehicle charging on its own.
6
Conclusion
This paper studies how firms compete in product markets by investing in complementary goods and how firms’ investment incentives change when previously incompatible complementary goods become compatible. The electric vehicle market itself is
an important market to understand because it could become a larger presence in the
automotive industry and carry large potential environmental benefits.
This paper presents and estimates a structural model of consumer vehicle demand
with utility over the electric vehicle charging network. Consumers have tastes over the
local usefulness of the charging network relative to their commuting patterns as well
as over national traversability. The demand parameters are combined with a model
of oligopolistic car manufacturers to recover vehicle markups and charging station
costs. The simulated counterfactual results show that, under compatibility, firms would
reduce investments in charging stations. Yet, the size of the electric vehicle market
would still expand since consumers can access all stations. A compatibility policy
would improve social welfare despite the cutback in car manufacturer charging station
investment.
This paper motivates two lines of future work. First, the dynamic incentives in
investment intended to influence the equilibrium number of standards remain unexplored. Second, and more generally, a deeper understanding of industries’ ability and
willingness to self-organize into a uniform standard or to make joint investments would
inform antitrust and innovation policy.
Finally, this paper contributes to understanding the role of directed technological
change in climate change policy. Although a market price on environmental damages from emissions and pollution may be part of the first-best solution, Acemoglu
29
et al. (2016) develop an endogenous growth model to show that the optimal climate
policy path includes both carbon taxes and research subsidies for clean technologies.
Aghion et al. (2016) show that firms in the automobile industry respond to higher
tax-inclusive fuel prices by innovating more in alternative fuel (electric, hybrid, and
hydrogen) technologies. This paper’s findings supports the argument that in addition
to market failures in the upstream innovation stage, other inefficiencies and market
failures in downstream product markets can hinder technological change.
References
Acemoglu, D., U. Akcigit, D. Hanley, and W. Kerr (2016). Transition to Clean Technology. Journal of Political Economy 124 (1), 52–104.
Ackerberg, D. A. and M. Rysman (2005). Unobserved Product Differentiation in
Discrete-Choice Models: Estimating Price Elasticities and Welfare Effects. RAND
Journal of Economics 36 (4), 771–788.
Aghion, P., A. Dechezleprêtre, D. Hémous, R. Martin, and J. Van Reenen (2016).
Carbon Taxes, Path Dependency, and Directed Technical Change: Evidence from
the Auto Industry. Journal of Political Economy 124 (1), 1–51.
Berry, S. (1994). Estimating Discrete-Choice Models of Product Differentiation. RAND
Journal of Economics 25 (2), 242–262.
Berry, S., J. Levinsohn, and A. Pakes (1995). Automobile Prices in Market Equilibrium.
Econometrica 63 (4), 841–890.
Berry, S., O. B. Linton, and A. Pakes (2004). Limit Theorems for Estimating the
Parameters of Differentiated Product Demand Systems. Review of Economic Studies 71, 613–654.
Blonigen, B. A., C. R. Knittel, and A. Soderbery (2019). Keeping it Fresh: Strategic Product Redesigns and Welfare. International Journal of Industrial Organization Forthcoming.
Borenstein, S. and L. W. Davis (2015). The Distributional Effects of U.S. Clean Energy Tax Credits. In J. R. Brown (Ed.), Tax Policy and the Economy, Volume 30,
Chapter 6, pp. 191–234. University of Chicago Press.
30
Burgelman, R. and D. Schifrin (2011). Nissan’s Electric Vehicle Strategy in 2011;
Leading the Way Toward Zero-Emission. Stanford GSB Case SM189.
Busse, M., J. Silva-Risso, and F. Zettelmeyer (2006, September). $1000 Cash Back: The
Pass-Through of Auto Manufacturer Promotions. American Economic Review 96 (4),
1253–1270.
Clinton, B. and D. Steinberg (2016). Providing the Spark: Impact of Financial Incentives on Battery Electric Vehicle Adoption. Unpublished.
Cormen, T. H., C. E. Leiserson, R. L. Rivest, and C. Stein (2009). Introduction to
Algorithms (3 ed.). Cambridge, Massachusetts: MIT Press.
Crawford, G. S., O. Shcherbakov, and M. Shum (2015). The Welfare Effects of Endogenous Quality Choice in Cable Television Markets. CEPR Discussion Paper No.
10793.
Draganska, M., M. Mazzeo, and K. Seim (2009). Beyond Plain Vanilla: Modeling
Joint Product Assortment and Pricing Decisions. Quantitative Marketing and Economics 7, 105–146.
Eizenberg, A. (2014). Upstream Innovation and Product Variety in the US Home PC
Market. Review of Economic Studies 81, 1003–1045.
European Commission (2014). Directive 2014/94/EU of the European Parliament
and of the Council of 22 October 2014 on the Deployment of Alternative Fuels Infrastructure. https://eur-lex.europa.eu/legal-content/en/TXT/?uri=CELEX%
3A32014L0094.
Fan, Y. (2013). Ownership Consolidation and Product Characteristics: A Study of the
US Daily Newspaper Market. American Economic Review 103, 1598–1628.
Ferwerda, R., M. Bayings, M. Van derKam, and R. Bekkers (2018). Advancing ERoaming in Europe: Towards a Single “Language” for the European Charging Infrastructure. World Electric Vehicle Journal 9 (4).
Gandhi, A., Z. Lu, and X. Shi (2017). Estimating Demand for Differentiated Products
with Zeroes in Market Share Data. https://www.ssc.wisc.edu/~xshi/research/
gandhi lu shi.pdf.
31
Goldberg, P. K. (1995). Product Differentiation and Oligopoly. Econometrica 63 (4),
891–951.
Gowrisankaran, G. and M. Rysman (2012). Dynamics of Consumer Demand for New
Durable Goods. Journal of Political Economy 120 (6), 1173–1219.
Graff Zivin, J. S., M. J. Kotchen, and E. T. Mansur (2014). Spatial and Temporal Heterogeneity of Marginal Emissions: Implications for Electric Cars and Other
Electricity-Shifting Policies. Journal of Economic Behavior & Organization 107,
248–268.
Greaker, M. and T.-R. Heggedal (2010). Lock-In and the Transition to Hydrogen
Cars: Should Governments Intervene? The B.E. Journal of Economic Analysis and
Policy 10 (1), 1–30.
Gross, D. P. (2016). The Ties that Bind: Railroad Gauge Standards and Internal Trade
in the 19th Century U.S. Harvard Business School Working Paper 17-044.
Hardman, S., A. Jenn, G. Tal, J. Axsen, G. Beard, N. Daina, E. Figenbaum, N. Jakobsson, P. Jochem, N. Kinnear, P. Plötz, J. Pontes, N. Refa, F. Sprei, T. Turrentine,
and B. Witkamp (2018). A Review of Consumer Preferences of and Interactions
with Electric Vehicle Charging Infrastructure. Transportation Research Part D 62,
508–523.
Ho, K. (2006). The Welfare Effects of Restricted Hospital Choice in the US Medical
Care Market. Journal of Applied Econometrics 21, 1039–1079.
Holland, S. P., E. T. Mansur, N. Z. Muller, and A. J. Yates (2016). Are There Environmental Beneftis from Driving Electric Vehicles? The Importance of Local Factors.
American Economic Review 106 (12), 3700–3729.
Holland, S. P., E. T. Mansur, N. Z. Muller, and A. J. Yates (2019). Distributional
Effects of Air Pollution from Electric Vehicle Adoption. Journal of the Association
of Environmental and Resource Economists Forthcoming.
Holtsmark, B. and A. Skonhoft (2014). The Norwegian Support and Subsidy Policy of
Electric Cars. Should it be Adopted by Other Countries? Environmental Science &
Policy 42, 160–168.
32
Hortaçsu, A. and J. Joo (2018). Semiparametric Estimation of CES Demand System
with Observed and Unobserved Product Characteristics. arXiv:1511.05600v8.
Houde, J.-F. (2012). Spatial Differentiation and Vertical Mergers in Retail Markets for
Gasoline. American Economic Review 102 (5), 2147–2182.
Idaho National Laboratory (2015).
Plugged In:
How Americans Charge
Their Electric Vehicles. https://avt.inl.gov/sites/default/files/pdf/arra/
PluggedInSummaryReport.pdf.
Ishii, J. (2007). Compatibility, Competition, and Investment in Network Industries:
ATM Networks in the Banking Industry. Unpublished.
Katz, M. L. and C. Shapiro (1985). Network Externalities, Competition, and Compatibility. American Economic Review 75 (3), 424–440.
Katz, M. L. and C. Shapiro (1986). Product Compatibility Choice in a Market with
Technological Progress. Oxford Economic Papers 38, 146–165.
Knittel, C. R. and V. Stango (2008). Incompatibility, Product Attributes and Consumer Welfare: Evidence from ATMs. The B.E. Journal of Economic Analysis &
Policy 8 (1), Article 1.
Knittel, C. R. and V. Stango (2011). Strategic Incompatibility in ATM Markets.
Journal of Banking & Finance 35, 2627–2636.
Lee, R. S. (2013). Vertical Integration and Exclusivity in Platform and Two-Sided
Markets. American Economic Review 103 (7), 2960–3000.
Li, S., L. Tong, J. Xing, and Y. Zhou (2017). The Market for Electric Vehicles: Indirect
Network Effects and Policy Design. Journal of the Association of Environmental and
Resource Economists 4 (1), 89–133.
McFadden, D. L. (1974). Conditional Logit Analysis of Qualitative Choice Behavior.
In P. Zarembka (Ed.), Frontiers in Econometrics, pp. 105–142. New York: Academic
Press.
Michalek, J. J., M. Chester, P. Jaramillo, and C. Samaras (2011). Valuation of Plug-In
Vehicle Life-Cycle Air Emissions and Oil Displacement Benefits. Proceedings of the
National Academy of Sciences 108 (40), 16554–16558.
33
Nissan Motor Corporation (2012).
Nissan Sustainability Report.
www.nissan-global.com/EN/DOCUMENT/PDF/SR/2012/SR12 E All.pdf.
https://
Nosko, C. (2014). Competition and Quality Choice in the CPU Market. Unpublished.
Pavan, G. (2015). Green Car Adoption and the Supply of Alternative Fuels. Toulouse
School of Economics Working Paper 17-875.
Petrin, A. (2002). Quantifying the Benefits of New Products: The Case of the Minivan.
Journal of Political Economy 110 (4), 705–729.
Quan, T. W. and K. R. Williams (2018). Product Variety, Across-Market Demand Heterogeneity, and the Value of Online Retail. The RAND Journal of Economics 49 (4),
877–913.
Rantakari, H. (2008). Governing adaptation. The Review of Economic Studies 75,
1257–1285.
Roberts, J. and G. Saloner (2012). Strategy and organization. In R. Gibbons and
J. Roberts (Eds.), Handbook of Organizational Economics. Princeton University
Press.
Sallee, J. M. (2011). The Surprising Incidence of Tax Credits for the Toyota Prius.
American Economic Journal: Economic Policy 3, 189–219.
Sheldon, T. L., J. DeShazo, and R. T. Carson (2017). Designing Policy Incentives
for Cleaner Technologies: Lessons from California’s Plug-in Electric Vehicle Rebate
Program. Journal of Environmental Economics and Management 84, 18–43.
Simcoe, T. and J. Farrell (2012). Four Paths to Compatibility. In M. Peitz and
J. Waldfogel (Eds.), The Oxford Handbook of the Digital Economy. Oxford: Oxford
University Press.
Small, K. A. and H. S. Rosen (1981). Applied Welfare Economics with Discrete Choice
Models. Econometrica 49 (1), 105–130.
Springel, K. (2016). Network Externality and Subsidy Structure in Two-Sided Markets: Evidence from Electric Vehicle Incentives. https://www.dropbox.com/s/
t6iboy3erwdumdp/kspringel ev.pdf?dl=0.
34
Stock, J. and M. Yogo (2005). Testing for Weak Instruments in Linear IV Regression. In
Andrews, Donald W.K. (Ed.), Identification and Inference for Econometric Models,
pp. 80–108. Cambridge University Press.
Sweeting, A. (2013). Dynamic Product Positioning in Differentiated Product Industries: The Effect of Fees for Musical Performance Rights on the Commercial Radio
Industry. Econometrica 81 (5), 1763–1803.
U.S. Department of Commerce (2016).
Automotive Spotlight.
www.selectusa.gov/automotive-industry-united-states.
U.S. Department of Energy (2014). The History of the Electric Car.
energy.gov/articles/history-electric-car.
https://
http://
Williams, H. C. W. L. (1977). On the Formation of Travel Demand Models and
Economic Evaluation Measures of User Benefit. Environment and Planning 9, 285–
344.
Wollmann, T. G. (2018). Trucks Without Bailouts: Equilibrium Product Characteristics for Commercial Vehicles. American Economic Review 108 (6), 1364–1406.
35
Figure 1: Types of Level 3 (DC, Fast) Charging Standards
Level 3 (DC, Fast) Charging Standards
SAE J1772 Combo
Chademo
Tesla
BMW: i3
GM: Bolt, Spark EV
Volkswagen: e-Golf
Ford
Chrysler
Daimler
Nissan: LEAF
Mitsubishi: i-MiEV
Kia: Soul EV
Toyota
Peugeot
Citroën
Tesla: Model S, X
Depiction of plug shapes from Alternative Fuel Data Center
Notes: This figure depicts the connectors of each Level 3 electric vehicle charging
standard, along with the make and model of cars compatible with each standard as of
the end of 2015. Not all electric vehicles on the market are capable of DC fast-charging.
The automakers that are listed without vehicle models had pledged support for a
particular standard, but did not yet sell vehicles that have fast-charging capabilities as
of the end of 2015.
36
Nissan LEAF
Tesla Model S
BMW i3
0
Number of Charging Locations
500
1000
1500
Figure 2: Charging Network Size over Time, by Standard
2010
2011
2012
2013
Chademo (Nissan)
2014
Tesla
2015
2016
Combo (BMW)
Data Source: Alternative Fuel Data Center
Notes: This figure shows the number of charging locations available over time for
each charging standard. Vertical bars mark when the first cars compatible with each
standard became available in the U.S. market.
1. Nissan began deliveries of the Leaf in December of 2010 and began building Chademo
stations at the same time.
2. Tesla began deliveries of the Model S in June of 2012 and announced the Tesla
Supercharger program three months later, in September of 2012.
3. BMW began deliveries of the i3 in May of 2014 and announced a program to build
stations under the Combo standard two months later, in July of 2014.
37
Figure 3: Level 3 (DC, Fast) Charging Locations Plotted on a U.S. Map
38
Notes: This figure shows the Level 3 charging locations for each standard as of September 2015, using data from the
Alternative Fuels Data Center of the Department of Energy.
Number of Origin−Destination Pairs (Normalized 0−100)
0
20
40
60
80
Figure 4: The Number of Connected City Pairs on the Tesla Network
coeff = .28536971
0
50
100
150
200
Number of Tesla Charging Locations Outside MSAs
250
Notes: This figure shows the number of city (MSA) pairs that a Tesla car could travel
using Tesla’s network (y-axis), normalized to between 0 and 100, plotted against the
number of charging locations available as they were built over time from 2011-2015
(x-axis).
1. In an OLS regression of the number of connected city pairs on the number of charging
locations available, the coefficient is about .285, which means an additional pair of
cities was connected for about every 3.5 charging locations placed by Tesla.
2. The solid line shows the predicted values from the OLS regression.
39
Figure 5: Empirical Bayes Posterior Mean vs. Observed Market Shares
(b) Zooming in on Observed Shares Ranging from 0 to .0006
0
Empirical Bayes Posterior Mean Market Shares
.05
.1
.15
0
40
Empirical Bayes Posterior Mean Market Shares
.0005
.001
.0015
.002
(a) All Market Shares
0
.05
.1
Observed Market Shares (MLE)
.15
0
.0002
.0004
Observed Market Shares (MLE)
Notes: This figure plots the empirical Bayes posterior means against the observed market shares. Posterior mean estimates
may be larger or smaller than the original observed market shares, represented in the scatter plots as being above or below
the 45-degree line. Subfigure (a) shows all data points. Subfigure (b) zooms into the smallest market shares.
.0006
Figure 6: Counterfactual Traversability of the National Charging Network under Compatibility
(a) Tesla
2012q1
2013q1
2014q1
2015q1
2016q1
2011q1
.02
0
0
National Traversability, Combo
.005
.01
.015
National Traversability, Chademo
.005
.01
.015
.8
National Traversability, Tesla
.2
.4
.6
0
41
2011q1
(c) SAE 1772 Combo
.02
(b) Chademo
2012q1
2013q1
2014q1
2015q1
2016q1
2011q1
2012q1
2013q1
2014q1
2015q1
2016q1
Notes: This figure shows the counterfactual traversability (number of city pairs that vehicles can drive) for each of the
three standards, holding locations and quantities of stations fixed.
1. Traversability is normalized by the total number of city pairs to between 0 and 1.
2. The solid lines represent traversability in the status quo with 3 incompatible standards, and the dashed lines are the
counterfactual traversability with a uniform standard.
Figure 7: Markets with Charging Station Presence from Each Standard (Fixed Number of Stations)
(a) Incompatibility (Status Quo)
(b) Compatibility (Counterfactual)
42
Notes: This figure shows the number of markets that have stations of each standard using Venn diagrams. The left figure
depicts the status quo. The right figure depicts compatibility. Under compatibility, all standards are interoperable or the
same, so the standards labels merely reference the firms’ affiliations in the status quo. In the counterfactual simulations,
firms re-optimize locations of stations, holding fixed the number of stations. Under compatibility and with fixed number
of stations, firms build in more markets and fewer stations in each market.
Table 1: Evolution of Key Variables, 2011-2015
2011
Number of markets (MSA)
Number of EV models
MSRP of EV models (min)
MSRP of EV models (max)
EV unit sales
Battery range (min)
Battery range (max)
2012
2013
354
356
347
3
6
15
32,780 29,125 22,995
109,000 116,000 102,000
13,542 41,643 93,734
35
11
11
245
76
139
2014
2015
346
22
22,995
135,700
127,699
11
208
346
27
22,995
140,700
140,320
11
238
Notes: This table shows key variables of the U.S. electric vehicle market from 2011 to
2015, using vehicle registration data from IHS Automotive and vehicle characteristics
data from MSN Auto.
43
Table 2: Unit Sales, Market Shares, and Empirical Bayes Posterior Market Shares
Variable
All vehicle sales
44
Plug-in sales
2011 plug-in
2012 plug-in
2013 plug-in
2014 plug-in
2015 plug-in
sales
sales
sales
sales
sales
Observed market share
Posterior mean share
Mean
Std. Dev.
Min
10%
Median
90%
% Zeros
N
13,798.7
28,488.5
202
1,140
3,973.5
37,471
0
40,200
20.4
9.5
10.7
11.8
12.0
8.4
50.2
35.0
49.1
47.8
61.1
42.8
0
0
0
0
0
0
0
0
0
0
0
0
1
2
2
1
1
1
16
18
17
20
20
13
35.7
15.5
23.0
30.8
33.2
45.5
40,200
1,424
3,910
7,966
11,694
15,206
.00085
.00082
.0019
.0015
0
2.78e-9
0
.000027
.00024
.00035
.0023
.0020
35.7
0
40,200
40,200
Notes: This table shows summary statistics of vehicle sales, observed market shares, and estimates of empirical Bayes
posterior mean market shares. Each observation corresponds to outcomes for an available vehicle model, market (MSA),
and quarter, based on data from IHS Automotive from 2011 to 2015.
1. The top panel shows unit sales of all fuel types in the first row, followed by unit sales of plug-in vehicles by year.
2. The bottom panel depicts observed market shares and estimates of empirical Bayes posterior mean market shares.
Table 3: Demand System Estimates
Logit
Logit with Random Coefficients
OLS
(1)
IV
(2)
RC IV
(3)
×Income
(4)
×Long Distance
(5)
log(Price)
-2.316***
(0.0787)
-2.732***
(0.625)
-2.909***
(0.0248)
-0.166***
(0.0260)
-0.382***
(0.0411)
log(Local Level 2) × PHEV
0.0931***
(0.0177)
0.129***
(0.0295)
0.227***
(.00826)
-0.0818***
(.00771)
0.00136
(.00857)
log(Local Level 2) × BEV
0.0614**
(0.0245)
0.0912***
(0.0339)
0.0848***
(0.0105)
-0.00874
(.00968)
-0.0145
(0.0115)
log(Local Level 3) × PHEV
-0.00300
(0.00904)
0.0236**
(0.0114)
0.0224*
(0.0123)
0.0215
(0.0133)
-0.0139
(0.0133)
log(Local Level 3) × BEV
0.0580***
(0.00776)
0.0671***
(0.00867)
0.0619***
(0.0158)
0.153***
(0.0106)
-0.0460***
(0.0137)
-0.234
(0.300)
-0.902*
(0.509)
0.014
(0.0447)
0.00128
(0.0109)
-0.00885
(0.00932)
0.00552***
(0.00155)
0.00524**
(0.00267)
-0.0546***
(0.00584)
0.0224***
(0.00393)
0.0410***
(0.00700)
BEV dummy
-1.889***
(0.133)
-2.276***
(0.217)
-2.180***
(0.0243)
–
–
Battery range
0.00760***
(0.00104)
0.00915***
(0.00174)
0.00952***
(0.000153)
–
–
Battery size
0.0302***
(0.00322)
0.0288***
(0.00415)
0.0293***
(0.000219)
–
–
Horsepower
0.00608***
(0.000437)
0.00749***
(0.00265)
0.00796***
(.0000946)
–
–
All-wheel drive dummy
0.631***
(0.0980)
0.989***
(0.231)
1.054***
(0.00549)
–
–
Electricity price
0.00499
(0.0107)
-0.00456
(0.0116)
-0.00909***
(0.000230)
–
–
MSA inc. per cap. ($1000)
0.0365
(0.0828)
0.215*
(0.115)
0.0627***
(.00848)
–
–
–
–
0.858***
(0.000230)
–
–
40,200
0.376
X
X
–
35,418
0.388
X
X
59.42
35,418
X
X
–
VARIABLES
# City pairs × PHEV
# City pairs × BEV
Autocorelation of ξ (ρ)
Observations
R-squared
Market FE
Time FE
Min. Eigvalue Stat (IV F-stat)
Notes: Logit in (1) and (2) are from linear regressions; random-coefficients logit in (3) through (5) are from
GMM. (4) shows interactions between product characteristics and household income, and (5) shows interactions
with an indicator for whether a household drives more than 60 miles per day. A unit of observation is an
available vehicle model, market, and quarter. For (1) and (2), robust standard errors in parentheses. For (3)
– (5), standard errors are bootstrapped. *** p
Purchase answer to see full
attachment