1.1. Industry Background

Tesla Motors played a significant role in the comeback of electric vehicles by introducing Tesla Roadster, an all-electric sports car, in 2006 and beginning general production in March 2008. However, the model had a price tag of over $120,000, out of the price range for average buyers. Nissan Leaf ($33,000) and Chevrolet Volt ($41,000) were introduced into the US market in December 2010, marking the beginning of the mass market for EVs.

There are currently two types of EVs: battery electric vehicles (BEVs) which run exclusively on high-capacity batteries (e.g., Nissan Leaf), and plug-in hybrid vehicles (PHEVs) which use batteries to power an electric motor and use another fuel (gasoline) to power a combustion engine (e.g., Chevrolet Volt). As depicted in figure 1, quarterly EV sales increased from less than 2,000 in the first quarter of 2011 to nearly 30,000 in the last quarter of 2013, while the number of public charging stations has increased from about 800 to over 6,000. Nevertheless, the EV market is still very small: EV sales only made up 0.82% (or 113,889) of the total new vehicle sales in the United States in 2015, and there are only about 12,500 public charging stations as of March 2016, compared to over 120,000 gasoline stations.

There are several commonly cited barriers to EV adoption. First, EVs are more expensive than their conventional gasoline vehicle counterparts. The manufacturer’s suggested retail prices (MSRP) for the 2015 model of Nissan Leaf and Chevrolet Volt are $29,010 and $34,345, respectively, while the average price for a comparable conventional vehicle (e.g., Nissan Sentra, Chevrolet Cruze, Ford Focus, and Honda Civic) is between $16,000 and $18,000. A major reason behind the cost differential is the cost of the battery. As battery technology improves, the cost should come down. In addition, lower operating costs of EVs can significantly offset the high initial purchase costs.⁶ A recent study by EPRI (2013) compares the lifetime costs (including purchase cost less incentives, maintenance, and operation) of vehicles of different fuel types and finds that under reasonable assumptions, higher capital costs are well balanced by savings in operation costs: EVs are typically within 10% of comparable hybrid and conventional gasoline vehicles.

The second notable barrier to EV adoption is the limited driving range. BEVs have a shorter range per charge than conventional vehicles have per tank of gas, contributing to consumer anxiety of running out of electricity before reaching a charging station. Nissan Leaf, the most popular BEV in the United States, has an EPA-rated range of 84 miles on a fully charged battery in 2015. Chevrolet Volt has an all-electric range of 38 miles, beyond which it will operate under gasoline mode. This range is sufficient for daily household vehicle trips but may not be enough for longer distance travels.

The third barrier, closely related to the second, is the lack of charging infrastructure. A large network of charging stations can reduce range anxiety and allow PHEVs to operate more under the all-electric mode to save gasoline.⁷ There are two types of public charging stations: 240 volt AC charging (level 2 charging) and 500 volt DC high-current charging (DC fast charging), with the former being the dominant type. The installation of charging stations involves a variety of costs, including charging station hardware, other materials, labor, and permits. A typical level 2 charging station for public use has three to four charging units and costs about $27,000, while a DC fast charging station costs over $50,000.⁸ Charging stations can be found at workplace parking lots, shopping centers, grocery stores, restaurants, dealers, and existing gasoline stations, a point that we will come back to when constructing the instrument for the number of charging stations in the EV demand estimation. Owners of charging stations are often motivated by a variety of considerations such as boosting their sustainability credentials, attracting customers for their main business, and providing a service for employees. Charging stations are often managed by one of the major national operators such as Blink, ChargePoint, and eVgo.

The fourth barrier is the long charging time. It takes much longer to charge EVs than to fill up gasoline vehicles. A BEV may not be able to get fully charged overnight if just using a regular 120 volt electric plug (e.g., it takes 21 hours for a Nissan Leaf to get fully charged). To get faster charging, BEV drivers either need to install a charging station at home or go to public charging stations. It takes six to eight hours to fully charge a Nissan Leaf at a level 2 charging station and only 10–30 minutes at a DC fast charging station.⁹ Unlike BEVs, PHEV batteries can be charged not only by an outside electric power source, but by the internal combustion engine as well. Having the second source of power may alleviate range anxiety, but the shorter electric range limits the fuel cost savings from EVs.

1.2. Government Policy

The diffusion of electric vehicles together with a clean electricity grid can be an effective combination in reducing local air pollution, greenhouse gas emissions, and oil dependency. The EV technology is widely considered as representing the future of passenger vehicles. The International Energy Agency projects that by 2050, EVs have the potential to account for 50% of light duty vehicle sales.¹⁰ Many countries around the world have developed goals to develop the EV market and provide support to promote the diffusion of this technology (Mock and Zhang 2014).¹¹

To reduce the price gap between EVs and their gasoline counterparts, the Energy Improvement and Extension Act of 2008, and later the American Clean Energy and Security Act of 2009, grant a federal income tax credit for new qualified EVs. The minimum credit is $2,500 and the credit may be up to $7,500, based on each vehicle’s battery capacity and the gross vehicle weight rating. Moreover, several states have established additional state-level incentives to further promote EV adoption such as tax exemptions and rebates for EVs and nonmonetary incentives such as high occupancy vehicle (HOV) lane access, toll reduction, and free parking. California, through the Clean Vehicle Rebate Project, offers a $2,500 rebate to BEV buyers and a $1,500 rebate to PHEV buyers. In addition, federal, state, and local governments provide funding to support charging station deployment. For example, the Department of Energy provided ECOtality Inc. a $115 million grant to build residential and public charging stations in 22 US cities in collaboration with local project partners.

Government intervention in this market could be justified from the following perspectives. First, indirect network effects in the EV market represent a source of market failure since marginal consumers/investors only consider the private benefit in their decision, and the network size on both sides is less than optimal (Liebowitz and Margolis 1995; Church, Gandal, and Krause 2002). In addition, given the nature of the market, each side of the market is unlikely to internalize the external effect on the other side through market transactions. If EVs are produced by one automaker, the automaker would have an incentive to offer a charging station network to increase EV adoption. Nissan and GM are the two early producers of EVs, but more and more auto makers are entering the competition. Nissan is a large owner of charging stations but GM is not.¹²

Second, the external costs from gasoline consumption in the United States and many countries around the world are not properly reflected by the gasoline tax (Parry and Small 2005; Parry et al. 2014). Compared to conventional gasoline vehicles, EVs offer environmental benefits when the electricity comes from clean generation such as renewables. In regions that depend heavily on coal or oil for electricity generation, EVs may not demonstrate an environmental advantage over gasoline vehicles.¹³ Electricity generation continues to become cleaner around the world due to the adoption of abatement technologies (e.g., scrubbers), the deployment of renewable generation, and the switch from coal to natural gas. In addition, technologies are being developed to integrate EVs and renewable electricity generation such as solar and wind. The integration of the intermittent energy source with EV charging not only can help EVs fully realize its environmental benefits but also can leverage EV batteries as a storage facility to address the issue of intermittency and serve as an energy buffer (Lund and Kempton 2008).

Third, technology spillovers among firms often exist, especially in the early stage of new technology diffusion (Stoneman and Diederen 1994). The development of EV technology requires significant costs, but the technology know-how once developed can spread through many channels, including worker migration and the product market. Bloom, Schankerman, and van Reenen (2013) estimate that the social returns to R&D are larger than the private returns due to positive technology spillovers, implying underinvestment in R&D. In addition, the social returns to R&D by larger firms are larger due to stronger spillovers.

1.3. Data

We construct a panel data set consisting of quarterly EV sales by vehicle model and the number of charging stations available at 353 MSAs from 2011 to 2013. Table 1 presents summary statistics of the variables used in our regression analysis. Data on quarterly vehicle sales of each EV model in each MSA are purchased from IHS Automotive. The sales data include 17 EV models: 10 BEVs and 7 PHEVs. Due to different introduction schedules, there were two vehicle models in our 2011 data: Nissan Leaf and Chevrolet Volt. The 2012 data include four more vehicle models: Ford Focus EV, Mitsubishi i-MiEV, Fisker Karma, and Toyota Prius Plug-in. The 2013 data include 11 additional models: Honda Accord Plug-in, Ford C-Max Energi, Cadillac ELR, Honda Fit EV, Fiat 500E, Smart ForTwo Electric Drive, Tesla Model S, Porsche Panamera, Toyota RAV4, Chevrolet Spark EV, and Ford Transit Connect EV. In 2013, the top four EV models are Nissan Leaf, Chevrolet Volt, Tesla Model S, and Toyota Prius plug-in with market shares (sales) of 25.8% (22,610), 24.4% (23,094), 17.4% (18,650), and 9.4% (12,088), respectively.

For our analysis, we focus on the 353 MSAs (out of 381 MSAs in total) for which observations are available in all three years, and their EV sales accounted for 83% of the national EV sales during our data period. Panel A of figure 2 depicts the spatial pattern of EV ownership (the number of EVs per million people) in the last quarter of 2013. It shows that large urban areas have a higher concentration of EVs. The MSA with the highest concentration is San Jose–Sunnyvale–Santa Clara, CA, with 5,608 EVs per million people by the end of 2013. The next two MSAs are both nearby: San Francisco–Oakland–Fremont and Santa Cruz–Watsonville. The MSA with the lowest concentration is Laredo, TX, with only 36 EVs per million people (nine EVs with a population of a quarter of a million).

We obtain detailed information on locations and open dates of all charging stations from the Alternative Fuel Data Center (AFDC) of the Department of Energy. By matching the ZIP code of each charging station to an MSA and using the station open date, we construct the total number of public charging stations available in each quarter for each MSA. Panel B of figure 2 shows the spatial distribution of charging stations (the number of charging stations per million people). The pattern is very similar to what we observe in panel A for EV ownership. The correlation coefficient between the two variables is 0.63, partly reflecting the interdependence of EVs and charging stations. The top three MSAs with the most charging stations per million people are Corvallis, OR, Olympia, VA, and Napa, CA, with 210, 170, and 117 public charging stations per million people, respectively. These three MSAs are the number eleventh, fifth, and sixth in terms of the EV concentration in panel A.

We collect data on state-level incentives such as tax credits and rebates for both electric vehicles and charging stations from AFDC. From the American Chamber of Commerce cost-of-living index database, we collect quarterly gasoline prices for each MSA from 2008 to 2013. Household demographics are collected from the American Community Survey.

2.1. Model Setup and Properties

We assume that EV sales q_t(N_t, p_t, x_t) depends on the number of public charging stations in the market (N_t), the price of the EV (p_t), and other product characteristics combined (x_t) that affect consumers’ choice, such as the fuel cost.¹⁴ The installed base of EVs is the cumulative sum of EV sales minus scrappage by the time t, denoted by $Q_t={\displaystyle {\sum }_{h=1}^t}{q}_h*s_{t,h}$, where s_t,_h is the survival rate at time t for EVs sold in time h. The number of charging stations that have been built N_t(Q_t, z_t) depends on the EV market size Q_t and other variables combined z_t that might affect the fixed cost of investment. To facilitate the illustration, we specify the following functions for EV demand and charging station deployment:

The parameters β₁ and γ₁ capture the magnitude of the indirect network effects on the two sides. Feedback loops (or two-way feedback) arise if both β₁ and γ₁ are nonzero. Intuitively, a shock to the system, for example, an increase in x_t, would change EV sales q_t, which would in turn affect the installed base Q_{t plus; 1}. This would then lead to changes in the number of charging stations N_{t + 1} and hence affect q_{t + 1}. The impact would circle back and forth between these two equations. If both β₁ and γ₁ are positive, positive feedback loops would arise, and they can amplify the shocks (either positive or negative) in either side of the market, such as a tax credit for EV purchases or subsidy on charging station investment. The parameter β₂ (negative) is the price elasticity of demand and captures consumer price sensitivity.

To understand the property of the system such as the existence of the steady state and its property, we assume that the survival rate s_t, _h is δ^{t − h}, where δ < 1. Further assume $p_t=p$, x_t = x, and z_t = z. Substituting equation (2) into equation (1), we have:

The right-hand side ${\beta }_1{\gamma }_{2}z+{\beta }_2\ln (p)+{\beta }_{3}x$ is constant with respect to q_t, and we denote it by c. During period 1, Q_{t − 1} = 0. Solving the equation, we obtain $q_1=\exp [c/(1-{\beta }_1{\gamma }_1)]$. To solve for the steady state solution (q*, N*), we use the steady state condition $q_t=q_{t+1}=q^{*}$ and find:

The stock of EVs in the steady state $Q^{*}=q^{*}/(1-\delta )$ where the outflow of EVs due to scrappage is equal to the inflow from new EV sales.¹⁵ To examine the stability of the steady state, we write $Q_{t-1}=q_{t-1}+\delta Q_{t-2}$ and substitute it into equation (3): $\ln (q_t)-{\beta }_1{\gamma }_{1}ln(q_t+\delta q_{t-1}+{\delta }^2{Q}_{t-2})=c$. This defines an implicit function of $q_t=G(q_{t-1})$. When ${\beta }_1{\gamma }_1<1$, it can be shown that $G(0)>0,G^{\prime }()>0$, and $G^{″}()<0$. Therefore, the steady state solution is stable as shown in figure 3. In our following policy analysis, we take ${\beta }_1{\gamma }_1<1$, which is also confirmed in our empirical analysis.

The partial effect of vehicle price p on EV sales in the steady state is:

2.2. Implications on Policy Choices

Now we conduct simulations to understand how feedback loops magnify policy shocks and their implications on policy choices. We fix p_t, x_t, and z_t in equations (1) and (2) and assume certain values for model parameters as reported in table 2. We then solve for q_t, Q_t, and N_t sequentially for each period. Because of the positive feedback loops (by assuming both β₁ and γ₁ being positive), EV sales and the number of charging stations will keep growing naturally until they reach the steady state where the inflow of new vehicles equals the outflow of vehicles due to scrappage. To examine how positive feedback loops could amplify a policy shock, we simulate a scenario where all EV buyers are provided with a $7,500 subsidy for the first five periods and no more subsidy is offered afterward.

As shown in panel A in figure 4, due to both the price effect (captured by β₂) and the indirect network effects (captured by β₁ and γ₁), the subsidy increases EV sales substantially compared with the no-policy case during the first five periods. When the subsidy terminates, EV sales continue to increase through feedback loops but with a smaller magnitude. The sales increase due to subsidy gets smaller as feedback loops diminish and the two growth paths eventually overlap. In both cases, the path of EV sales converges to the same steady state but the policy shock makes the system converge to the steady state more quickly: indirect network effects expedite this process through positive feedback loops.

Figure 4, panel B, depicts a similar pattern in the dynamic path of charging station deployment. With the positive policy shock on the EV purchase side, the stock of charging stations increases quickly for the first five periods and continues to grow at a decreasing rate after the policy. It eventually converges to the same steady state as in the no-policy scenario. These two graphs demonstrate that feedback loops from indirect network effects magnify a shock to any side of the system and alter the convergence process on both sides.

The existence of indirect network effects on both sides of the market could have important policy implications. To foster the development of the EV market, policy makers can choose to subsidize consumers for EV purchase directly (policy 1) or to subsidize charging station investment (policy 2). We conduct simulations to examine the relative cost effectiveness of these two policy options. Policy 1 provides EV buyers with a subsidy of $7,500 per EV in the first five periods. Policy 2 uses the same account of total funding as in policy 1 to build charging stations. We compare the cumulative sales increase over time due to these two policies (with a 5% annual discount rate).¹⁶

To examine the implication of the relative strength of indirect network effects on policy choices, we vary the ratio of β₁/γ₁ by holding β₁ constant while changing γ₁. Figure 5 depicts, for any given price sensitivity β₂ (say −1.5), as β₁/γ₁ increases (i.e., indirect network effects in EV demand become relatively stronger), the second policy (subsidy on charging stations) becomes more and more effective measured by the increase in cumulative sales over time. The two policies are equivalent when β₁/γ₁ is 1 given the price elasticity of −1.5.

In addition to the relative strength of indirect network effects on both sides, the policy comparison also depends on the price elasticity of EV demand. When consumers are more sensitive to prices (e.g., going from −1.5 to −1.6), the policy of subsidizing charging stations becomes relatively less effective for a given β₁/γ₁. This finding is illustrated by the outward shift of the curve when the price elasticity changes to −1.4 and −1.6. The result is intuitive: if consumers are less price sensitive, it would take a larger subsidy on EV purchases in order to push consumers to buy EVs, hindering the effectiveness of the policy.

To summarize, the policy of subsidizing charging stations becomes more effective relative to the policy of subsiding EV purchases when indirect network effects on EV demand become stronger (holding network effects on the charging station side constant) or when consumers are less sensitive to price. These findings offer a theoretical foundation for the policy comparison after our empirical analysis.

3.1. EV Demand

To describe the empirical demand model of EVs, let k index an EV model such as Nissan Leaf and Chevrolet Volt, m index a market (MSA), and t index a year-quarter. We estimate the following equation:

We also include a full set of year-quarter (e.g., the first quarter of 2011) fixed effects and MSA model (e.g., Nissan Leaf in San Francisco) fixed effects in equation (4). Year-quarter fixed effects T_t control for national demand shock for EVs common across MSAs such as consumer awareness. MSA-model fixed effects δ_km not only control for time-invariant product attributes such as quality and brand loyalty that could affect vehicle demand but also control for time-invariant local preference for green products (Kahn 2007; Kahn and Vaughn 2009) and demand shocks for each model (e.g., a stronger preference or dealer presence for Nissan Leaf in San Francisco). The term ε_kmt is the unobserved demand shocks that are time varying and market specific (for example, unobserved local government subsidy for purchasing EVs or market-specific promotions for a vehicle model that vary over time).

It is well documented in the vehicle demand literature that failing to control for unobserved product attributes could lead to downward bias in the price coefficient estimates (for example, Berry, Levinsohn, and Pakes 1995; Beresteanu and Li 2011). MSA-model fixed effects absorbs both observed and unobserved vehicle attributes variations that are time invariant, and what is left is the variation of vehicle attributes over time. Since most of the EV models in our sample appear for only one year and there is little variation of the observed attributes for the models that appear for more than one year, we believe that using MSA-model fixed effects could control for unobserved product attributes and alleviate the need to use the methodology developed in Berry et al. (1995) to deal with price endogeneity where they only have national-level sales data (i.e., one market).¹⁹ The price coefficient is identified from the fact that effective EV prices vary across markets and over time due to state-level subsidies and temporal price variations.

Although we include a rich set of control variables, the charging station variable is still endogenous due to simultaneity: the unobserved time-varying and market-specific demand shocks could affect charging station investment decisions and hence the stock of charging stations. To deal with the endogeneity, we use the IV strategy, and a valid IV needs to be correlated with the number of charging stations in an MSA (the endogenous variable) but not correlated with the unobserved shocks to EV demand. The IV we employ is the interaction term between the number of grocery stores and supermarkets in an MSA in 2012 with the number of charging stations in all MSAs other than the MSA corresponding to a given observation (lagged for one quarter). Grocery stores and supermarkets are a major owner of charging stations, and they build charging stations to attract customers and boost green credentials, among other reasons. These places could be good sites for public charging stations because EV drivers can charge their vehicles while shopping. Nissan has been actively partnering with grocery store owners to build charging stations. Kroger, the country’s largest grocery store owner, has installed about 300 charging stations in their stores across the country. Our data show that the number of grocery stores in an MSA is positively correlated with the number of charging stations.

However, the number of grocery stores does not vary with time in our sample period, and it is therefore absorbed by the MSA fixed effects. To introduce temporal variation, we multiply it with the lagged number of existing charging stations in all MSAs other than the MSA corresponding to a given observation, which captures the national-level trend in charging station investment due to aggregate shocks such as temporal variations in costs, investor confidence, and federal incentive programs. The construction of this IV is similar in spirit to the Bartik instrument used in the labor literature to isolate local labor demand changes (Bartik 1991). The intuition for the IV is that national shocks to charging station investment (captured by the lagged number of charging station in all MSAs other than own) have disproportional effects on charging station investment across MSAs: MSAs with a larger number of grocery stores and supermarkets (hence better endowment of good sites for charging stations) will be affected by these national shocks more than others, leading to variations in charging stations across MSAs. Our first-stage results in table 4 show that the interaction term has a positive and highly statistically significant impact on charging station investments.

We argue that this instrument should satisfy the exogeneity assumption. The number of grocery stores and supermarkets is unlikely to affect EV sales directly. There might be common unobservables that influence both the EV sales and the number of grocery stores, especially at the cross-sectional level. However, our model controls for MSA fixed effects and should capture these time-invariant unobservables. At the temporal dimension, EV sales vary from year to year but the number of grocery stores is very stable given the maturity of the industry. In fact, the number of grocery stores is measured at the end of 2012. The temporal variation in the IV comes from the total (lagged) number of charging stations in all MSAs other than the own city. Time fixed effect would control for time-varying common shocks across MSAs. Excluding the home city’s charging stations also removes the concern that one MSA’s installation of a large number of charging stations could overly influence the estimation results.

Our IV strategy leverages the interaction term between a national-level variable with only temporal variation and an MSA-level variable with only spatial variation. The rationale behind the IV is that different MSAs have different preexisting conditions/ability to absorb national shocks to charging station investment such as changes in macro-economic conditions and costs. One might be concerned that different MSAs may have a different susceptibility to unobservable demand shocks at the national level and the number of grocery stores could be correlated with this susceptibility for some reason. To address this concern, we include a variety of MSA-level controls interacting with the time trend. We use the sales of hybrid vehicles in 2007 (several years before EVs entered the market) to proxy for preference heterogeneity for greener vehicles or environmental friendliness. We also include personal income, the share of college graduates among residents, the share of commuters driving to work, the share of commuters using public transport to work, and the share of white residents. We use the interactions of these variables with the time trend to control for potential heterogeneity in the diffusion path of EVs across MSAs. Our results are robust to the inclusion of these controls, providing further support that our IV is a valid exclusion restriction.

In some of the robustness checks, we use local policy variables such as subsidies on charging stations as additional IVs and obtain similar results. We do not use them in our benchmark specifications due to the concern that local policies whether subsidizing charging stations or EV purchases could be a response to local unobserved demand shocks and hence be endogenous.

3.2. Charging Station Deployment

We derive the empirical model of charging station investment from an entry model presented in the appendix where the profit depends on both the installed base of EVs and the total number of charging stations in a market. Under certain functional form assumptions, the total number of charging stations in a free-entry equilibrium is given by the following equation:

We also include a full set of time and MSA fixed effects. The term T_t denotes year-quarter fixed effects to control for time-varying common shocks to charging station investment across MSAs such as macro-economic conditions. Market fixed effects φ_m control for time-invariant and MSA-specific preferences for charging stations. For example, some MSAs may be “greener” than others and invest more on alternative fuel infrastructure. Similarly, MSAs with a higher population density and a limited private installment of charging stations may have more public charging stations. The term ζ_mt is the unobserved shock to charging station investment, for instance, the unobserved local policies to support the charging station building. In the estimation, we add one to N_mt and $Q_{mt}^{EV}$ to deal with zero values for some of the observations. We obtain similar results by dropping these observations and use ln(N_mt) and $\ln (Q_{mt}^{EV})$ instead.

The issue of endogeneity due to simultaneity also arises in this equation. Both N_mt and $Q_{mt}^{EV}$ are stock variables, but the inflows to each variable are determined at the same time. As a result, time-varying and MSA-specific shocks to investment decisions (the error term in the equation) could be correlated with current EV sales, which are part of the installed base. The instrument variables emerge more naturally in this equation. In particular, we instrument for the installed base of EVs with a set of current and past gasoline price variables. The fuel cost savings from driving EVs depend on the price difference between gasoline and electricity, which varies across locations. In MSAs with higher gasoline prices, consumers may have a stronger incentive to purchase EVs.²⁰ Because the installed base of EVs is the cumulative sales of EVs, we include gasoline prices not only in the current quarter but annual gasoline prices in the past three years as instruments. For example, for the installed base of EVs in the second quarter in 2013, we use the gasoline price in the second quarter in 2013, the average gasoline price in 2012, the average gasoline price in 2011, and the average gasoline price in 2010 as instrumental variables.

These gasoline price variables (including current and past gasoline prices) should affect the installed base, which is confirmed in the first-stage regression in table 8. But they are unlikely to affect investment decisions directly (i.e., other than through the installed base). Since we include both time and MSA fixed effect, the remaining variation in gasoline prices is largely driven by how time-varying crude oil prices interact with market conditions that are likely time-invariant during our data period (e.g., market structure in wholesale and retail gasoline markets and distance to refineries). These interactions lead to time-varying and MSA-specific differences in gasoline prices, which are unlikely to be correlated with charging station investment directly. The decision of charging station investment hinges on, among other things, the EV market potential (proxied by the installed base of EVs) and the fixed costs of investment. Fixed costs of charging station investment include the cost of equipment (chargers) and labor cost, neither of which is likely to be correlated with gasoline price variations (after controlling for MSA and time fixed effects). The operating costs of the charger largely depend on electricity prices. There is no direct link between electricity and gasoline prices (after controlling for common shocks such as national economic conditions using time fixed effects).

4.1. Regression Results for EV Demand

Columns a–e in table 3 report the ordinary least squares (OLS) estimation results for five different specifications where we add more control variables successively (first-stage results reported in table 4). Column a includes only six explanatory variables. Column b adds in year-quarter fixed effects to control for time-varying common unobservables across MSAs. Column c further adds vehicle model fixed effects to control for unobserved product attributes such as quality and brand loyalty that affect consumer demand. Column d includes rich MSA-model fixed effects to control for both unobserved product attributes and MSA-specific demand shocks for different EV models. Column e adds two additional variables to control for potential heterogeneity in the diffusion pattern across MSAs. The first variable is the interaction term between the sales of hybrid vehicles in 2007 (to proxy for preference for green vehicles) and the time trend, and the second variable is the interaction between average personal income and the time trend. Column f implements a GMM estimation strategy and uses the interaction term of number of grocery stores and supermarkets in an MSA in 2012 with the lagged number of charging stations in all the other MSAs as the instrument for the number of charging stations.

Given the log-log specification, the coefficient estimates can be interpreted as elasticities. All the specifications provide intuitive and statistically significant coefficients on the key variables of interests: EV demand increases with a larger network of charging stations, a lower vehicle price, and more home charging stations funded by the EV project supported by the DOE, and higher income. The coefficient on the charging station variable captures indirect network effects from charging station investment on EV demand. The GMM results show that a 10% increase in charging stations would result in an 8.4% increase in EV sales, which is higher than all the OLS estimates (ranging from 1.8% to 5%). This suggests that the number of charging stations is negatively correlated with the unobserved shocks to EV demand, leading to downward bias in OLS. One example of unobserved shocks is local EV incentives that local governments provide to compensate for the lack of public charging stations. Another example is the home charging incentives from local electric utilities. Many local utilities offer a rebate for installing a home charging station and a discounted rate for home EV charging as part of the demand side management program. Local governments often partner with local utilities to provide more generous home charging incentives when there is a lack of private investment in public charging stations.

The price coefficients change from −0.470 to −0.817 from columns b–c after vehicle model fixed effects are included. This is consistent with our discussion above: vehicle model fixed effects control for unobserved product attributes which could be positively correlated with prices. Ignoring unobserved product attributes will bias the price coefficient toward zero. Going from columns c–d where MSA-model fixed effects are included, the EV demand function changes from being inelastic with a price elasticity of −0.817 to being elastic with a price elasticity of −1.378. MSA-model fixed effects control for MSA-specific time-invariant demand shocks (such as environmental preference), and these demand shocks could affect state-level tax incentives. For example, higher incentives are used to counter negative demand shocks. Hence MSA fixed effects could control for the potential endogeneity in state-level tax incentives. The GMM results provide a price elasticity of −1.288. Although this is at the lower end of the price elasticity estimates in the literature on automobiles, we believe that the magnitude is reasonable compared with the literature for two reasons.²¹ First, EV buyers are more affluent and hence less price sensitive compared with average vehicle buyers. Second and perhaps more importantly, EV buyers can be characterized as early adopters, and one can argue that many of them choose EVs out of their strong environmental concerns and/or making a statement by driving an EV as has been documented in the case of hybrid vehicles (e.g., Kahn 2007; Sexton and Sexton 2014).²²

Lower fuel cost is one of the major benefits of EVs, and higher gasoline prices will have a positive impact on EV adoption by increasing future fuel cost savings from driving EVs in place of conventional vehicles. To capture the heterogeneous impact of gasoline price on the demand of the two types of EVs, two interaction terms of quarterly gasoline prices with BEV and PHEV dummy variables are included. The results from all the specifications find a positive and statistically significant effect of gasoline price on BEV purchase. While the interaction term with PHEV is positive and significant in columns a–c of table 3, columns d–f do not find a significant impact of gasoline price on PHEV demand when MSA-model fixed effects are included. Intuitively, BEV drivers could be more sensitive to gasoline prices than PHEV buyers given that BEVs run exclusively on electricity while PHEVs run mostly on gasoline for long-distance driving given its short range of battery charge. That is, PHEV drivers do not make a long-term commitment to an alternative fuel to the extent that BEV drivers do. In addition to gasoline prices, we included electricity prices in previous analysis and the coefficient estimate was small in magnitude and statistically insignificant in all specifications. This could be because (1) the operating cost from using electricity is a small portion of vehicle lifetime cost for EVs and (2) there is not much MSA-specific temporal variation in electricity prices. The coefficient on the interaction term between hybrid sales in 2007 and time trend is positive and statistically significant, implying that MSAs that had more sales of hybrid vehicles in 2007 (proxy for preference for greener products) have a faster diffusion of EVs.

The results from these regressions imply that the increased availability of public charging stations has a statistically and economically significant impact on EV adoption decisions. Our estimation results confirm that even if most EV drivers can charge vehicles at home, better access to public charging facilities elsewhere is still an important demand factor by, for example, alleviating range anxiety.²³ Based on the parameter estimates on charging station and price variables, a back-of-the-envelope calculation shows that the demand effect from having one more charging station from the sample average of 22.6 is equivalent to that from a reduction of EV price by $961 (the average price is $33,127). When the number of charging stations increases to 27.3 (the sample average in 2013), the equivalent price reduction is $795. At the sample maximum of 320 charging stations, one more charging station is only equivalent to $68 price reduction, showing the diminishing effect implied by the log-log functional form.

4.2. Alternative Specifications for EV Demand

We take the estimates in column f in table 3 as our baseline specification. To check the robustness of the results, we estimate a variety of different specifications, and the results are reported in table 5. Column a includes the interaction term between the number of charging stations and average commute time to work in the MSA to capture the heterogeneous impact of charging stations across cities with different commuting patterns.²⁴ The positive coefficient estimate on the interaction term suggests that the availability of charging stations has a larger impact on EV demand in the MSAs with a longer commute. This is intuitive since in MSAs where people have a longer commute, range anxiety would be more of an issue. Across MSAs, the elasticity of charging stations with respect to the EV sales ranges from 0.27 to 1.05 depending on the average commute time.

Column b adds the quadratic terms of the time trend variables to our baseline specification to allow more flexible time effect. The coefficient estimates on the key parameters are almost intact. Column c includes the interaction term of charging stations with a BEV dummy to capture the different impact of charging stations on BEVs and PHEVs. The coefficient estimate on the interaction term is positive while not statistically significant. Column d includes two more instruments: the tax credits and the availability of public funding for building charging stations, both of which appear in the charging station equation. We did not include them in our baseline specification because the exogeneity assumption may not hold for the subsidies as they could be a response to unobserved EV demand shocks. Column e removes the price variable in the regression to deal with the concern that the price variable, especially state-level incentives, could be endogenous. Column f adds the interaction terms between various demographic variables and the time trend to further control for MSA-level heterogeneity in the diffusion pattern. Across these specifications, the estimated effects of charging station availability on EV demand as well as other parameter estimates are similar to those from the baseline specification in table 3. Column g reports the demand estimation using the logit model as in Berry (1994), and it produces almost identical elasticity estimates as our baseline specification.

Some states, such as California and Oregon, have adopted a Zero Emission Vehicle (ZEV) program, which requires a certain part of automakers’ sales to be clean fuel vehicles, and some automakers have introduced EV models in those regions only to comply with the regulations. To control for more intense competition in those markets due to more EV models introduced, column h in table 6 includes ZEV-specific time fixed effects, and the results are similar to previous results with a modest increase in the coefficient estimate on charging stations. Column i uses only BEV sales, and the estimates are not systematically different from the estimates using the full sample with BEVs and PHEVs. Columns j and k increase the lag of the total number of charging stations in all other MSAs to two and three-quarters when constructing the instrumental variable and there is no substantial change of the estimates except that the coefficient of the charging stations decreased slightly, primarily due to loss of observations. However, increasing the lag to more than three-quarters leads to weak IV.

As shown in column g in table 6, our demand specifications would yield nearly identical results as the Berry logit model. With only EV models in our data, our analysis treats all other non-EV models to be in one category (i.e., the outside good). Limiting the choice set and the substitution pattern across choices could potentially affect our estimate of the price elasticity and our policy simulations. EV models represent a different technology that is dramatically different from conventional gasoline vehicles; therefore, consumers are likely to consider them as a separate category in making purchase decisions, especially given that the EV buyers in our data period are often motivated by strong environmental concerns according to the California Plug-in Electric Vehicle Owner Survey (2014). Nevertheless, some PHEVs do have conventional hybrid counterparts. For example, the Toyota Prius plug in has a hybrid version, Toyota Prius. Considering most of PHEVs have limited electric range (11–38 miles), some consumers may compare PHEVs with hybrid vehicles. Recognizing this, column i in table 6 only include BEV models in the regression. The results are very similar to those obtained using the full sample, suggesting that the limitation in our choice set and modeling framework may not have a large impact on the key parameters of interest.²⁵

4.3. Regression Results for Charging Station Deployment

Columns a–d in table 7 report the OLS regression results for the charging station equation (5) (first-stage results reported in table 8). In column a, only the four explanatory variables of interest are included. Column b includes year-quarter fixed effects to control for time trends that are common to all MSAs such as federal subsidies for building charging stations that occur during a specific period of time. Column c further includes MSA fixed effects to control for time-invariant MSA-level baseline differences in charging station investment. Column d adds in the interaction term between the hybrid vehicle sales in 2007 (proxy for environmental friendliness) and time trend to control for heterogeneity in the diffusion pattern of charging stations.

All OLS regressions find a positive and statistically significant coefficient for the installed EV base. The estimate results in column d suggest that a 10% increase in the EV fleet size would lead to a 1.2% increase in the number of public charging stations. The GMM results in column e show that a 10% increase in EV fleet size would result in a 6.1% increase in charging stations. In column f, we add the EV incentives (tax credits and rebates at the federal and state levels) as an additional instrument and the coefficient for charging stations increases from 0.613 to 0.659. We take column e as our baseline IV specification due to the concern that the EV incentives (especially those at the state level) could be endogenous as they could be a response to the unobserved shocks to the deployment of charging stations.

The GMM coefficient estimates are higher than all the OLS estimates, suggesting that the installed base of EVs is negatively related to the unobserved shocks to charging station investment, leading to downward bias in OLS. An example of the unobserved shocks is the unobserved local policies: policy makers may design policies to support charging station investment to counteract negative EV demand shocks.

The results in column e show that tax credits and the availability of public funding for charging stations have positive but statistically insignificant coefficients. The tax credits and public funding are both at the state level. The dependent variable in the charging station equation, however, only includes publicly accessible charging stations, which are mainly subsidized by the federal government directly through federal projects such as the EV Project and ChargePoint Project. Although state-level tax credits and funding also apply to public charging stations, they mostly support the installation of charging stations at workplace and multifamily dwellings, which are usually privately accessible and are excluded from our analysis. The interaction term of grocery stores with the lagged number of stations in all MSAs other than own has a positive and statistically significant coefficient, consistent with our argument for using it as a relevant instrument in the EV demand equation.

5.1. Impact of Income Tax Credits

The federal government has adopted several policies to support the EV industry, including providing federal income tax credits for EV purchase, R&D support for battery development, and funding for expanding charging infrastructure. The CBO (2012) estimates that the total budgetary cost for those policies will be about $7.5 billion through 2017. The tax credits for EV buyers account for about one-fourth of the budgetary cost and are likely to have the greatest impact on vehicle sales. Under the tax credits policy, EVs purchased in or after 2010 are eligible for a federal income tax credit up to $7,500. Most popular EV models on the market are eligible for the full amount.²⁶ The credit will expire once 200,000 qualified EVs have been sold by each manufacturer.

In order to examine the effectiveness of the income tax credit policy in terms of stimulating EV sales, we use our parameters estimates from the two baseline GMM regressions to stimulate the counterfactual sales of EVs that would arise in the absence of the $924.2 million worth of tax credits to EV buyers from 2011 to 2013. The impact of the policy depends not only on the price elasticity of EV demand in the EV demand equation but also on the magnitude of indirect network effects captured in both equations.

We assume in our simulations that the MSRPs will not be affected, implying that consumers previously captured all the subsidies. We believe that this is a reasonable assumption in the EV launch stage when automakers produce EVs likely at a loss since the production level is far below the efficient production scale.²⁷ While more and more states are providing subsidy programs to encourage the adoption of EVs, the retail prices for electric vehicles have actually been decreasing (for the same model) during our sample period likely due to decreasing production cost and the increasing competition.

Our simulation results in table 9 show that EV sales would have been 56,690 less (or 40.44% of the total sales) from 2011 to 2013 without the $924.2 million worth of income tax credit to EV buyers. If we shut down feedback loops, the sales contribution from the tax credit policy would only have been 33,949 (24.2% of the total sales). This implies that feedback loops magnify the policy shock and explain 40% of the sales increase from the policy. The results suggest that feedback loops have a multiplier effect of 1.67. The CBO finds the policy impact to be 30% of the total EV sales while their study only considers the price effect of the tax credit but not the role of indirect network effects in amplifying the policy effect (CBO 2012). DeShazo et al. (2014) study the California Clean Vehicle Rebate Projects for EVs and find a 7% increase in EV sales from the rebate of $1,838 on average. Neither of these studies takes into account indirect network effects, and their estimates likely provide the lower bounds of subsidy impacts.

5.2. Policy Comparison

Our stylized model suggests that feedback loops from indirect network effects on both sides of the market have important policy implications. A policy shock on one side of the market would affect the other side. To promote EV adoption based on a variety of rationales as discussed in section 1.2, policy makers face a problem of optimal policy design in that the tax revenue can be used to subsidize one or both sides of the market. We compare the subsidy policy on EV purchase with an alternative policy of subsidizing charging station investment. The alternative policy uses the same budget of $924.2 million evenly in each quarter during 2011–13 to install charging stations in all MSAs (proportional to population). As a lower bound estimate of the investment cost of charging stations, we assume the government is only responsible for the purchase and installation of the charging hardware and the charging station company will then operate and maintain the charging stations, as in the case of the EV Project and ChargePoint Project, the two federal charging station support programs. As a robustness check, we also estimate an upper bound investment cost for charging stations assuming the government will also need to maintain and operate those charging stations. The lower bound and the upper bound of the charging station investment cost are $27,000 and $50,244, respectively.²⁸

The policy comparison between these two policies is provided in table 10. We assume the policy period runs from 2011 to 2013 (i.e., no subsidy available in either policy after that). The existing tax credit policy (policy 1) has led to 56,690 more EVs from 2011 to 2013, amounting to $16,303 for one additional EV. The policy effect will continue to exist until the feedback loops die out in 2055.²⁹ The impact on EV sales from this policy in the long term would be 184,049, amounting to $5,022 per policy-induced EV purchase. If instead, the government had spent the $924.2 million subsidizing charging stations by purchasing and installing charging infrastructure (policy 2 with lower station cost), EV sales would have increased by 124,904 during these 3 years. The cumulative impact on EV sales from this policy until year 2055 would be 403,558, amounting to $2,290 per induced EV, only 46% of the unit cost under the existing policy. If the government were also responsible for maintaining and operating those stations (policy 2 with higher station cost), EV sales would have increased by 75,199 during these three years and 267,741 in the long term, amounting to $3,452 per induced EV, still preferable to policy 1.

As depicted in figure 6, policy 2, which subsidizes charging station investment, demonstrates a dominant advantage in stimulating EV sales in the early stage of the EV market. The $924.2 million spending during the 3 years can install about 18,395 to 34,231 charging stations depending on the actual investment cost. This is more than one-eighth to one-fourth of the total number of gasoline stations in the country and almost one and half to three times of the current total number of public charging stations in the whole country. This large amount of public charging stations should dramatically alleviate or even eliminate range anxiety for potential EV buyers. Our results indicate that building charging stations is a more effective way to boost EV sales in the EV launch stage. As shown in our regression results, indirect network effects on the EV demand side are much stronger than those on the charging station side (coefficient ratio being 1.4) and consumers are not very sensitive to prices (price elasticity being −1.3). As a result, the policy that builds charging stations stimulates EV sales at a much faster pace, consistent with our findings in the model section.

The long-run simulations are based on a variety of assumptions and are meant to be illustrative. As the technology improves, the EV driving range is likely to increase, weakening indirect network effects from charging stations to EV demand. In addition, in the longer term, as EVs become more of a serious choice for average vehicle buyers, consumer price sensitivity among EV buyers could increase. Both of these changes would affect the policy outcomes and weaken the effectiveness of policy 2 relative to policy 1.

As discussed in section 4.2, the network size of charging stations has heterogeneous impacts on the EV demand across locations with different average commute time. This implies heterogeneity in the relative strength of indirect network effects on the two sides of the EV market and hence heterogeneity in policy comparison. That is, policy 2 may not be always preferred as previous analysis suggests. In MSAs where indirect network effects on EV demand are not strong since drivers have shorter commutes and home charging is enough to ensure their daily trips, policy 1, which subsidies EV purchase, could be more effective. Figure 7 depicts the relative effectiveness of the two policies in promoting EV adoption. The MSA with policy 2 being most effective is New York–New Jersey–Long Island (NY-NJ-PA) where the average commute time is longest, while the MSA with policy 1 being most effective is Grand Forks (ND-MN) where the average commute time is shortest. This suggests that a more elaborate and cost-effective policy design would be to subsidize charging stations in areas with long commute (e.g., large MSAs) but to subsidize EV purchases in areas with short commute (e.g., small MSAs). This type of regionally differentiated policy could be implemented at the federal level but perhaps more feasibly at the state and local levels. For example, in states or cities where the average commute is longer, state and local governments should focus on building charging station infrastructure while subsidies on EV adoption, for example, through rebate and HOV lane usage, can be implemented in states and cities where the average commute is shorter.