Nguyen 1
Faculty of Economics and Business
MSc Economics
Track: Markets and Regulations
Assessment of the EVSE Incentive Program in the U.S.
Electric Vehicle Market: The Case of Missouri
Nhu Nguyen
Student number: 11567783
Nguyen 2
1. Introduction
In the US, petroleum remains the energy source that many industries and people rely on. Particularly, 37% of total energy usage of the country comes from petroleum (EIA, 2016). Importantly, transportation remains the sector that relies the most heavily on petroleum. In 2016, 71% of petroleum is used in this sector alone (EIA, 2016). This extensive use of gasoline, a product of petroleum can create oil dependency issue in the US. The use of petroleum products also aggravates air pollution. Therefore, decreasing the use of petroleum, especially in the transportation sector can significantly mitigate these two issues. Promoting the adoption of plug-in electric vehicles serves as one of the recent popular methods to achieve such a goal. In
achieving this, the US government has launched multiple incentive programs to support the potential consumers of electric cars, both financially and non-financially, directly and indirectly. However, some of these incentives effectiveness have yet to be analyzed, particularly the
incentive to promote the infrastructure.
This paper will concentrate on the indirect incentive programs that aim to increase the number of Electric Vehicle Supply Equipment (EVSE), or so-called the charging stations for electric
vehicles. This is due to the fact that without a sufficient charging station network, the plug-in electric car industry will struggle to boom. In this paper, two methods will be used, whose results will be compared: difference in difference, and synthetic control. These two methods will first assess the effect of the incentive program on the number of charging stations over time. I will then estimate the impact of the number of stations on the sales quantity of electric cars. From there I can estimate how the incentive program influences the sales quantity over time. While
Nguyen 3
such an incentive program is available in a number of states, Missouri emerges as the ideal assessment target due to its unique features, which will be discussed later.
The diff-in-diff approach reports an increase of 82% of stations annually with the availability of an incentive program. The synthetic control approach, however, suggests that the percentage increase will rise over time, with 151% in 2015 and 251% in 2016. Using the figures given by using the synthetic control, I found that the incentive program of Missouri can increase the monthly sales of electric cars by 58.5% in 2016 and 97% in 2017. Nevertheless, the figures may come across as both the result of the incentive program and the positive network externality, rather than the incentive program alone.
2. Literature review and contribution
2.1.Financial incentives
Multiple previous work have studied the impacts of regulations, financial incentive programs and socio-economic factors on the adoption of plug-in electric vehicle (plug-in EV henceforth), with a concentration on financial ones. Lutsey and Sperling (2012) have stressed the importance of gasoline prices and the availability of the Zero Emission Vehicle (ZEV) mandate in some US states in the adoption of EV. Diamond (2009) has also found gasoline price as a significant factor in adopting hybrid electric vehicle (HEV henceforth), where 10% increase in gas price will cause hybrid car share to increase 72-93%. Nevertheless, direct financial incentives have little impact. On the contrary, Jenn et.al. (2013) has found that for each dollar of incentive, HEV sales
increases by 0.0046%, but gas price has little impact. Yet, Jenn et.al. and Diamond’s work are based on HEV rather than plug-in EV, and we thus cannot assume similar results for plug-in EV.
Nguyen 4
Zhang et.al. (2013) have studied the plug-in EV market in China, using survey method and have found direct financial incentives to be a significant factor in the purchase of plug-in EV.
However, among four factors: performance attributes, psychological needs, financial benefits and environmental awareness, financial benefits were ranked the third in terms of the
importance. DeShazo et.al. (2017) have used simulation method to assess the effectiveness of different direct subsidies schemes and found major differences between each scheme. In
particular a progress incentive based on income can increase the sales of plug-in EV much more than the status quo federal tax credit. Overall, most papers have found that direct financial incentives do have an impact on the quantity of plug-in EV, but the scale may not be the largest among all types of incentives.
2.2 Non-financial incentives and socio-economic factors
Besides direct financial incentives, Langbroek et.al. (2016) and Carley et.al. (2013) have also studied other incentive programs and/or socio-economic factors. Their work has indicated a variety of significant factors in the adoption of plug-in EV, such as the availability of HOV lane or free parking fee (non-financial incentives). Socio-economics factors such as age, gender and education also have an effect. Male, younger, and more educated people are more likely to adopt plug-in EV than the counterparts. Helveston et.al. (2015) have compared the effectiveness of purchase subsidies and other car features such as price, national origins or specifications between US and Chinese consumers, using survey method and the logit model. They have found the significant impacts of those features on plug-in EV purchase decision, with the size of the effects differ between consumers of 2 nations. Additionally, without subsidies, plug-in EV compete poorly against HEV and ordinary gasoline vehicles in both nations. Thus, besides financial
Nguyen 5
incentives, certain social features such as education or age, and travelling advantages can also bring consumers closer to the plug-in EV market.
2.3.The role of charging stations
In addition to direct financial incentive schemes and socio-economic factors, several papers have also studied different topics regarding the plug-in EV charging stations. Fetene et.al. (2016) have used theoretical framework to assess the demand and supply of work place charging, and have found that without charging stations subsidies/rebates, there will not be socially optimal supply of work place charging in the US. Sierzchula et.al. (2014) have collected plug-in EV market share and other data from 30 nations to study the impact of various factors on plug-in EV adoption. The result has suggested that education, gasoline price and income do not have much an impact on plug-in EV market share. Meanwhile, both direct financial incentives and the amount of charging infrastructure are significant factors. Importantly, while increasing incentive credit by $1000 only results in 0.06% increase in EV market share, increasing one more charging station per 100,000 persons increases EV share 0.12%, which is doubling the effect of purchase incentive.
Springel (2016) has gone further by using the logit model and indirect network effect model to study the impact of subsidies schemes for purchasing and for constructing charging stations. Her result has indicated that charging stations incentives are more effective than direct purchasing subsidies. In particular, 12.39 million USD funding for charging stations can result in 835 more plug-in EV. The same amount if spent on purchase subsidy only results in 387 more plug-in EV. However, when funding amount increases, the situation reverses. Li et.al. (2014) have also found similar result, using simulation method and vehicle registration data from 382 Metropolitan
Nguyen 6
Statistical Area from 2011 to the end of 2013. Their work has suggested that with a funding amount of 924.2 million USD, purchase subsidies will result in around 168,000 more EV after 10 years, while if the same amount is spent on charging stations, the increase in EV would be
around 245,000 to 373,000. As such, we can observe that any incentive program aims to increase the amount of charging stations has a crucial role in promoting plug-in EV.
This paper aims to assess the effectiveness of current and ongoing incentive programs for building plug-in EV charging stations, using the US data. In the US, the federal government has placed a direct financial/purchasing incentive since 2010. This means that every US state has at least one purchasing incentive. However, subsidies programs for charging stations are only available in a number of states. In this paper, an incentive scheme that comprises of both types is dubbed the “double-incentive”, and a scheme with only the federal purchasing subsidies is called “single-incentive”. The purpose of this paper is studying the effectiveness of the
double-incentive scheme, in compare to the single-double-incentive scheme in promoting plug-in EV adoption.
3. Methodology
For the purpose of assessing the effectiveness of the double-incentive scheme on the adoption of plug-in EV, the diff-in-diff approach will first be used. Then the synthetic control method will be used to compare with the results given by the diff-in-diff approach. Treatment and control
entities are US states. Notably, the charging station incentive programs, either in the form of tax credit or rebate do not directly affect the quantity of plug-in EV but rather the amount of
charging stations. Thus, the diff-in-diff and synthetic control method will exploit the direct effect of the incentive programs on the quantity of charging stations. It is hypothesized that states with
Nguyen 7
charging station incentive program enacted will see a larger increase in the quantity of charging stations compare to states without. For the treatment states, it is ideal to obtain states with the incentive program starting at the beginning of the year, so that the policy can have full effect in that year. Additionally, treatment state should not have more than one charging station incentive program or a state purchase incentive going on, either a state governmental program or a
utility/private program, or the station quantity may be overestimated. Only state governmental incentive program rather than a private incentive program will be chosen, due to the accessibility to the program for every business entity and/or resident of the state. Also, the dataset will consist of states with programs that do not end before 12/2016 and do not have any gap during the enacted period of the programs, so that the effect of the policy in all years after the enacted date can be assessed appropriately.
For those reasons, Missouri emerges as the perfect candidate to be the treatment state. The control states will include 14 states that do not have either charging station incentive or state purchasing incentive program during the assessment period, resulting in an overall 15 states being assessed. The station incentive program of Missouri started in January 2015. The
assessment period extends from 2011 to 2016. This is because by 2011, every state has had the federal purchasing incentive program available. Additionally, data is obtainable until 2016. However, due to the small sample size of treatment state, there might be biased estimation of the policy effect. Therefore, I will employ synthetic control method constructed by Abadie and Gardeazabal (2003) and Abadie et.al.(2010) in addition to the conventional diff-in-diff method, since the synthetic control method can work well for small sample size of treatment units. Also,
Nguyen 8
the method can allow for time-varying state-specific heterogeneity (McClelland and Gault, 2017).
In the next section of the paper, the effect of the quantity of stations on the quantity of plug-in EV will be estimated. This step is done in order to estimate the effect of the station incentive programs on the quantity of plug-in EV. It is assume that more stations will lead to more plug-in EV adoption. Importantly, there exists positive network effects between the number of charging stations and the quantity of plug-in EV (Springel, 2016; Li et.al., 2014). This means that the variable of plug-in EV amount and charging station amount will determine each other simultaneously, and thus creating the issue of reverse causality. A lagged value of charging station amount and the current plug-in EV amount can also determine each other, as an
expectation of more plug-in EV purchasing in the near future may induce more construction of charging stations now. In order to solve the endogeneity issue, an instrumental variable will employed (IV henceforth). This IV will be a dummy variable for the availability of charging station incentive programs. It is expected that a state with a charging station incentive program will tend to have more stations. Ideally, this IV will not be directly correlated with the quantity of plug-in EV, as discussed earlier. This assumption exists to ascertain the exogeneity of the IV. With the size of the effect of charging station amount on plug-in EV amount in hand, I can then estimate the size of the effect of the charging station subsidy on plug-in EV adoption.
Due to the unavailability of information regarding the exact enacted date and end date (if they end before 12/2016) of charging station incentive programs (the dummy IV), the sample size will include 40 states, which will be listed in the below data gathering section.
Nguyen 9
4. Data gathering
The dataset consists of two sets of data, one for the diff-in-diff and synthetic control section, and one for the plug-in EV quantity model.
4.1.Diff-in-diff dataset
The data for the open dates and locations (state and address) of all charging stations in the US is obtained from the US Department of Energy. I then calculate the quantity of stations per 100,000 persons in each year using this data, which is my dependent variable. The information regarding the availability of state charging station incentive program of 14 control states is extracted from the US Department of Energy. The availability, timeframe and end date of the incentive program in Missouri was collected on the US Department of Energy and the State of Missouri Revisor of Statues. I gathered the data for state yearly income from the US Bureau of Economic Analysis (BEA). Finally, the data for yearly state population and state land area (for calculation of
population density) is acquired from the US Census Bureau. For the diff-in-diff section, the panel data with yearly observations of 15 states from 2011 to 2016 will be employed. A variation of the same panel data will be used when exercising the synthetic control approach. Particularly, instead of consisting yearly observations, the dataset will include quarterly observations.
4.2.Plug-in EV model dataset
The data for monthly sales of plug-in EV, both PHEV (plug-in hybrid EV) and BEV (battery EV) is collected from Auto Alliance website, beginning from 1/2011 to 12/2016. The data for the monthly quantity of stations is gathered from the same source with the diff-in-diff and synthetic
Nguyen 10
control section (Department of Energy). Income per state varies little during a year, and thus was not reported by the BEA. Therefore, the yearly data for the income from BEA is still used. Similarly, the plug-in EV model still uses the dataset with yearly population density, due to the lack of report of the Census Bureau in terms of monthly state population, assuming that there are small variances in the population during months in a year. The yearly data for gasoline price was collected from the US Department of Energy, due to also the lack of monthly report for gasoline price. The information for the state purchasing rebate/tax credit and charging station incentive program (for the IV) was gathered from the same sources as with the information for station incentive program in the diff-in-diff section (Department of Energy). However, due to the comprehensiveness of the information, I collected many other sources for these information. These programs can be state run or private run. Below in table (1) the list of sources from which the availability, enacted date and/or end date of incentive programs were collected from. Note that information on enacted date of these programs is collected for the period before 12/2016 only. Additionally, the information regarding the enacted and end date (if applicable) for some states’ programs is not available on the websites in the list. Rather, it was achieved by calling a staff/employee using the phone number available on the websites. One such example is
Nguyen 11
Table (1). Sources of the availability and enacted date of incentive programs
State Purchasing incentive Station incentive State Purchasing incentive Station incentive New York NY Dept. of Taxation and Finance Dept. of Energy (none) West Virginia Dept. of Energy Dept. of Energy (none) Colorado CO Dept. of Energy Co Energy Office South Carolina Greg Montgomery (Abundant Power); Dept of Energy Dept. of Energy (none) Massachusetts MA State website MA State website Wyoming Dept. of Energy Dept. of Energy (none) Texas Lori Clark (TX Clean City Coordinator) Dept. of Energy Montana Dept. of Energy Dept. of Energy (none) Delaware DE Division of Energy and Climate DE Division of Energy and Climate Idaho Dept. of Energy Dept. of Energy (none) Maryland MD Energy Administratio -n Dept. of Energy Nebraska Dept. of Energy Dept. of Energy (none) Florida FL State Legislature (Statute 163.08) JEA and
OUC Utility Kansas
KCP&L Utility Dept. of Energy Georgia Don Francis (Clean City Coordinator) Don Francis (Clean City Coordinator) Louisiana Dept of Energy (none) Dept. of Energy (none) Connecticut Jenn Reilly (CT Bureau of Air Management) Jenn Reilly (CT Bureau of Air Management Illinois IL Dept. of Commerce IL State Legislature (Statute 120-30) Wisconsin Dept. of Energy (none) Dept. of Energy (none) Michigan Dept. of Energy; Indiana Michigan Power Dept. of Energy (none) Iowa Dept. of Energy (none) Dept. of Energy (none) Indiana Dept. of Energy; Indiana Michigan Power Dept. of Energy (none)
Nguyen 12 Arkansas Dept. of Energy (none) Dept. of Energy (none) Vermont VT Clean City Coalition Plug-in America Nevada NV Governor’s Office of Energy; South West Energy Dept. of Energy (none) Arizona AZ State Legislature (Statute 43-1176) Dept. of Energy (none) Washington WA State Legislature (Code 47.38.060) Dept. of Energy Oregon Ross Elizabeth (OR Dept. of Energy) Dept. of Energy (none) Utah UT Dept. of Environmenta -l quality UT State Legislature (Code 19-1-403.3) Kentucky Dept. of Energy (none) Dept. of Energy (none) Ohio Ohio Laws and Rules (Code 122.075) None Alabama Dept. of Energy (none) Dept. of Energy (none) North Carolina Dept. of Energy Dept. of Energy (none) South Dakota Dept. of Energy (none) Dept. of Energy (none) Missouri MO Revisor of Statute (Statute 135.710) Dept. of Energy (none) New Mexico Dept. of Energy (none) Dept. of Energy (none) Pennsylvania Michele Ferguson (PA Dept. of Environmenta -l Protection PA Dept. of Environment al Protection Mississippi Dept. of Energy (none) Dept. of Energy (none)
New Jersey Dept. of
Energy Dept. of Energy (none) North Dakota Dept. of Energy (none) Dept. of Energy (none)
Nguyen 13
Some notes for the information regarding incentive schemes
Although many states have purchasing and/or station incentive programs, they are most the time not the same. First, Both charging station and purchasing incentive programs can take the form of either direct rebate, or tax credit, or grant. Also, The amount of rebate/tax credit/grant differs from state to state. For example, Colorado offers 80% rebate of all costs of building stations, while New York offers a tax credit of 50% costs of building stations. Missouri’s tax credit amounts to 20% of the costs. The same applies for purchasing incentives. Importantly, A state may have both station and purchasing incentive programs going on, at the same or at different timeframes.
It should be noted that a state may have more than one station incentive program going on, at the same or different timeframe. The same applies for purchasing incentive program. The reason is that different departments and/or agencies can run their own incentive programs. These states may have both state and utility/private station and/or purchasing incentive program, so
simultaneously run programs may not be operated solely by the state government. Additionally, In a few states (for example Florida or California), some state incentive programs are only available in certain counties/cities/metro areas, which are offered by a particular municipal. Finally, in some states, incentive programs end before 12/2016, and then may or may not be revised after a period of time. If revised, they can have similar or different terms and conditions.
Nguyen 14
5. Diff-in-diff model
5.1.The model
The model will take the form:
𝑌𝑖𝑡 = 𝑐𝑜𝑛𝑠𝑡 + 𝛼1𝑍𝑖+ 𝛼2𝐷𝑡+ 𝛼3𝑠𝑢𝑏𝑀𝑂𝑖𝑡 + 𝛼4𝑊𝑖𝑡+ 𝜀𝑖𝑡
Where 𝑌𝑖𝑡 is the quantity of charging stations per 100,000 persons in each year, from 2011 to 2016. This is done so as to control for the population of each state, as more populous states will be likely to have more charging stations, sometimes even without incentive programs. 𝑍𝑖 is the set of state dummies, which will control for state fixed effects. 𝐷𝑡 is the set of year dummies , which will control for time fixed effects. 𝑠𝑢𝑏𝑀𝑂𝑖𝑡 is the dummy variable for Missouri, which will take the value of 1 when the state is Missouri, and the year is 2015 and onward. 𝑊𝑖𝑡 is the
set of control variables, which includes the population density (persons per square mile), and log of income. 𝜀𝑖𝑡 is the error term. The regression is run using robust option to control for
heteroskedasticity and autocorrelation.
5.2. Data description and growth trend
Table (2) describes the summary of statistic for the diff-in-diff regression model. Pop is the population, dens is the population density, and inc is income of a state. We can observe that large variation exists between different states for every statistic. The difference in population and density among states is likely to cause the significant difference in the number of charging stations. States that are more densely populated are likely to have more urban areas. Normally, charging station owners are companies (as perks for employees), businesses, especially shopping
Nguyen 15
malls, or auto corporations such as Tesla or Nissan. Urban areas tend to have more malls, auto corporations and businesses, and thus more stations.
Table (2). Statistic summary of diff-in-diff
(variable station is the number of stations per 100,000 persons) Variable Obs Mean Std. Dev. Min Max Total stations 96 42 66.5 0 443 Station 96 1.37 1.53 0 8.37 (per 100,000) Population (million) 96 2.72 1.70 0.57 6.09 Density (sq.mile) 96 47.88 36.31 5.81 109.78 Income ($, approx) 96 43,860 9430 31,200 71,000
This section also examines the trend of the growth in the number of charging stations in some states and of the US overall. Figure (1) shows the change in the total number of charging
stations of Iowa, North Dakota, Wisconsin, Alabama, Missouri and the average of all states over time (line avgstat). Notably, most states have close, small number of charging stations in 2011. This can be due to the fact that the plug-in EV market started to blossom only after 2010, when the federal tax credit became available. Also, none of the states included in the sample has any incentive program for plug-in EV adoption, either direct or indirect in and before 2011. There is a clear trend of increasing in the number of charging stations over time for all states except for North Dakota. This growth reflects the natural growth of plug-in EV sales over time, when more and more models introduced (3 models in 2011, and more than 30 models in 2016) with
affordable price after the federal tax credit. While all other states’ number of charging stations demonstrate steady growth over time, that of Missouri shows a large jump between 2014 and 2015. This jump can be due to the tax credit incentive program for charging station in 2015, or changes in the state unrelated to the incentive, and we therefore need to test for it.
Nguyen 16
Figure (1). Change in total number of stations of selected states over year
5.3. The result
The diff-in-diff regression result can be found in figure (2). The coefficient for subMo (dummy variable for Missouri after the incentive program is enacted) is 3.27, significant at 0.01 level. This suggests that the incentive program does have a significant impact on the number of charging stations in Missouri. With the incentive program, Missouri has on average 3.27 more stations per 100,000 persons than the control states. This is the average effect of the program over the course of 2015 and 2016. Therefore, without the incentive program, Missouri would have around 244 stations in 2016 (counterfactual value), instead of 443, given the state population of around 6.091 millions in 2016, which is an increase of around 82%. As a result, using the diff-in-diff method, we can conclude that a tax credit for the costs of building a
charging station results in around 82% more charging stations in a state than without. In the case of Missouri, the tax credit amounts to 20% of all the costs of constructing a charging station site,
Nguyen 17
with a capped amount of $10,000. Nevertheless, there exists one fundamental issue with this estimate. Using the same interpretation, I can calculate that without the incentive program, the number of stations would be around 27 stations in 2015 (counterfactual value, with a population of around 6.072 millions). This is unrealistic, since the total number of stations in Missouri in 2014 was 66. Thus, the number of stations in 2015 must be greater than 66. I therefore suspect that either the diff-in-diff estimate might be biased, or the effect of the incentive program differs widely from year to year, and thus we may not count on the average effect of all the years.
Interestingly, the coefficient of log of density is negative. However, it is not significant at any level. This means that the population density has little impact on the number of charging stations per 100,000 persons. The coefficient of log of income is positive. This suggests that the richer a state is, the more charging stations per 100,000 persons that state have, which is quite intuitive. However, the coefficient is not significant at any level, indicating that income of a state also does not have much of an effect on the number of stations. The coefficients for all years 2012, 2013 2014, 2015 and 2016 are positive and significant, at 0.05 and 0.01 level depending on the year. Moreover, the value of those coefficients increase over time. This indicates that the number of charging stations grows steadily in most states from 2012 onwards. It may be seen as a sign that the plug-in EV industry started to grow beginning in 2012. Nevertheless, as explained and mentioned earlier, this result may be biased, due to the small sample size of treatment state (only Missouri). As such, I will employ the synthetic control method in addition to the diff-in-diff method, which is often regarded as a more robust version of the diff-in-diff method.
Nguyen 18
Figure (2). Diff-in-diff result
(dep. variable station is the number of stations per 100,000 persons) (1)
VARIABLES Station (per
100,000) 2012.year 0.378** (0.181) 2013.year 0.577*** (0.176) 2014.year 0.972*** (0.187) 2015.year 1.573*** (0.258) 2016.year 2.489*** (0.361) subMO 3.270*** Ln_density (1.096) -3.690 (7.316) Ln_income 0.610 (3.725) Constant 10.093 (41.290) Observations 90 R-squared 0.849
Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
6. Synthetic control approach
6.1. The model
In this section, I will apply the synthetic control model constructed and used by Abadie and Gardeazabal (2003) and Abadie et.al. (2010) to evaluate the effectiveness of the station incentive program on the number of stations over time.
Nguyen 19
In the sample, there are 15 states over t = 2011, 2012, 2013, 2014, 2015 and 2016 periods, with Missouri (which I call state 1, for ease of notations) is the only treatment state, and other 14 states are control states (state 2 to 15). The charging station incentive program took place in January 2015 affected only Missouri. Let 𝑌𝑖,𝑡𝑁𝑇 be the number of charging stations per 100,000 persons in state i in year t that can be observed without the availability of the program. Similarly, let 𝑌𝑖,𝑡𝑊𝑇 be the number of charging stations per 100,000 persons in state i in year t that can be observed if the program is in place. Before the program officially enacted, it has no effect on the number of chargers (there is no action taking place as an anticipation of the program), since the tax credit only applies to chargers being built from January 2015 onwards. Thus:
𝑌𝑖,𝑡𝑁𝑇= 𝑌𝑖,𝑡𝑊𝑇 ∀𝑖, 𝑡 ∈ {2011, . . . ,2014}
The goal is to estimate the causal effect of the incentive program on the number of chargers in Missouri, which is:
𝛼1,𝑡 = 𝑌1,𝑡𝑊𝑇 − 𝑌1,𝑡𝑁𝑇 , 𝑡 ∈ {2015, 2016}
Where 𝑌1,𝑡𝑊𝑇 is observable. 𝑌1,𝑡𝑁𝑇, the number of charging stations per 100,000 of Missouri from 2015 onwards without the program is not observable, and thus is a counterfactual result. Therefore, in order to obtain 𝛼1,𝑡, it is necessary to construct 𝑌1,𝑡𝑁𝑇.
Similar to the diff-in-diff model, 𝑌𝑖,𝑡𝑁𝑇 follows a factor model ∀𝑖: 𝑌𝑖,𝑡𝑁𝑇= 𝐷𝑡+ 𝛼1𝐺𝑖𝑡+ 𝛼2𝑍𝑖,𝑡+ 𝜀𝑖,𝑡
Where 𝐺𝑖𝑡 is the vector of observed covariates unaffected by the incentive program. In this case it is again the log of income and log of population density. It also includes the values of the number of charging stations per 100,000 before 2015, which will the elaborated later. 𝑍𝑖,𝑡 is the
Nguyen 20
In order to estimate 𝑌1,𝑡𝑁𝑇, a synthetic version of Missouri by combining the data of 𝐺
𝑖𝑡 of the
control states will be created. In order to achieve this, I will reweight the control states in a way such that the 𝐺𝑖𝑡 vector of the synthetic Missouri converges to that of Missouri. This will also
result in an automatically matched 𝑍𝑖,𝑡. This whole process will allow for heterogeneity of different states, and thus acting as a fixed effect option.
Let W = (𝑤2, 𝑤3, . . . . , 𝑤15) be the vector of weights of the control states, with 𝑤𝑖 ≥ 0 ∀𝑖 For a given W , the outcome of the combination of the control states, or the synthetic cohort at time t is: 𝑌𝑤,𝑡 = ∑ 𝑤𝑖𝑌𝑖,𝑡 15 𝑖=2 = 𝐷𝑡+ 𝛼1(∑ 𝑤𝑖𝐺𝑖,𝑡 15 𝑖=2 ) + 𝛼2(∑ 𝑤𝑖𝑍𝑖,𝑡 15 𝑖=2 ) + (∑ 𝑤𝑖𝜀𝑖,𝑡 15 𝑖=2 )
Suppose that ∃𝑊∗ = (𝑤2∗, 𝑤3∗, . . . . , 𝑤15∗ ) such that the synthetic cohort matches Missouri in the years before the incentive program enacted, then:
∑ 𝑤𝑖∗𝑌𝑖,𝑡 = 𝑌1,𝑡 ∀𝑡 ∈ {2011, . . . ,2014} 15 𝑖=2 ∑ 𝑤𝑖∗𝐺𝑖,𝑡 = 𝐺1,𝑡 15 𝑖=2 ∀𝑡 ∈ {2011, . . . ,2014}
Then for all 𝑡 > 2014, we have:
𝔼 [𝑌1,𝑡𝑁𝑇− ∑ 𝑤𝑖∗
15 𝑖=2
𝑌𝑖,𝑡] → 0 As the number of pre-treatment periods grows large
Therefore, ∑15𝑖=2𝑤𝑖∗𝑌𝑖,𝑡 essentially acts as counterfactual 𝑌1,𝑡𝑁𝑇 that needs to be constructed. Thus
Nguyen 21
α̂1,t =
𝑌
1,𝑡𝑊𝑇−
∑ 𝑤𝑖∗ 15 𝑖=2𝑌𝑖,𝑡 ≈
𝑌
1,𝑡𝑊𝑇−
𝑌1,𝑡𝑁𝑇Which is the difference between the observed number of charging stations per 100,000 and the synthetic number from 2015 onwards.
Nonetheless, a 𝑊∗ that matches exactly the synthetic cohort with 𝑌1,𝑡𝑁𝑇 rarely exists. McClelland and Gault (2017) have suggested that as long as the assumption that the synthetic cohort
approximates 𝑌1,𝑡𝑁𝑇, then the implementation of the synthetic control is valid. A set of predictors
of the outcome variable is necessary to construct the data for synthetic Missouri. In this paper, such outcome variable is the number of charging stations per 100,000 persons. McClelland and Gault (2017) have indicated the importance of the lagged values of number of stations for some pre-treatment periods, which also belong to 𝐺𝑖,𝑡. This practice has also been used in the analysis of California tobacco policy analysis of Abadie et.al. (2010). However, Kaul et.al. (2016) have suggested not to use the outcome variable of all pre-treatment periods, since doing so can disregard the effects of all other covariates. Usually, some of the covariates are important in constructing accurate the post-treatment values of the outcome variables. Therefore, eliminating the effects of other covariates may produce biased post-treatment results for the synthetic cohort.
6.2. Application of the model to the data set
As McClelland and Gault (2017) have advocated for a dataset with more pre-treatment periods, I will employ the synthetic control on the same dataset, albeit breaking a year into four quarters. This effectively generates more pre-treatment periods. Thus, instead of having four pre-treatment periods (2011 to 2014), I now have sixteen pre-treatment periods (from 2011 quarter 1 to 2014 quarter 4), and eight post-treatment periods (from 2015 quarter 1 to 2016 quarter 4). It should be
Nguyen 22
noted that although the quarterly data for the number of charging stations is available, only yearly data for income and population density per state are available. However, the changes in income and population density during the year should be sufficiently insignificant for the
accuracy of the estimation. The following predictors are included: log of income, log of density, number of charging stations per 100,000 persons in 2011 quarter 4, 2012 quarter 4, 2013 quarter 4 and 2014 quarter 4, which results in a total of 6 predictors. I include the last quarters of each year since they will best represent the natural growth of new stations built over time.
Additionally, 2014 quarter 4 is the last pre-treatment period. It thus is expected to be a good predictor of the outcome value in 2015 quarter 1. The values of the predictors are averaged over the years before the incentive program enacted, which is 2015.
Table (3) shows the weight of 14 control states in the synthetic control. Arkansas is weighted the most heavily in the synthetic outcome, with 31.1%. Alabama contributes 1%, Kentucky 17.1%, North Dakota 11.7%, Nebraska 13.2%, Wisconsin 14% and West Virginia 12.7%. The rest 7 control states have no weight in the synthetic outcome. Table (4) shows the comparison between Missouri and its synthetic cohort, computed using quarterly data. The “Missouri” column shows the average values of log density and log income over the years 2011 to 2014, and the number of charging stations per 100,000 persons in 2011, 2012, 2013 and 2014 quarter 4 for Missouri. The “synthetic” column shows these values for the synthetic version of Missouri. The Root Mean Squared Prediction Error for the estimation of the synthetic Missouri reported by Stata is 0.029.
Nguyen 23
Table (3). Weight of each state in the synthetic control (Quarterly)
State | Weight State | Weight
AL | .001 AR | .311 IA | 0 ID | 0 KY | .171 MS | 0 MT | 0 ND | .117 NE | .132 NM | 0 SD | 0 WI | .14 WV | .127 WY | 0
Table (4). Comparison between Missouri and Synthetic Missouri (quarterly) (variable station is the number of stations per 100,000 persons) Predictors Missouri Synthetic
Ln_density 4.475 3.957 Ln_income 10.648 10.638 station (2011q4) .299 .299 (per 100,000) station (2012q4) .598 .597 (per 100,000) station (2013q4) .794 .794 (per 100,000) station (2014q4) 1.089 1.088 (per 100,000) 6.3. The results
Figure (3) shows the graphical comparison between the number of charging stations per 100,000 persons of Missouri and its synthetic cohort. In the figure, period 204 represents 2011 quarter 1, period 220 represents 2015 quarter 1, and so on. We can observe that the synthetic cohort fits Missouri very well for years before 2015, which indicates a good fit and validation of the
Nguyen 24
synthetic control. This may well suggest that any noticeable difference between Missouri and its synthetic cohort is not generated coincidentally by the synthetic control. We can also observe that Missouri detaches from its synthetic cohort beginning in 2015 rather than 2014, which is intuitive and reasonable.
Figure (3). Quarterly trends in number of charging stations in Missouri and its synthetic cohort
This is due to the breaking up of the effect of the incentive program in Missouri in 2015 into four quarters. Since planning and building a station often takes several months, the number of stations in the first quarter of 2015 is not significantly different from that of the last quarter of 2014. We therefore witness a difference from Missouri and it synthetic cohort beginning from 2015 quarter 2. From the graph, it is apparent that the gap between Missouri and its synthetic cohort is
noticeably large, and increases in size over time. This may be due to the fact that during the first year of the program (2015), Missouri enjoys large jump in the number of stations. This leads a jump in plug-in EV adoption. Since there exists positive network effects (Springel, 2016; Li et.al., 2014), the jump in plug-in EV quantity triggers even more construction of new stations the
Nguyen 25
following years, especially when the program is in place. Businesses, especially shopping malls or supermarkets wish to install charging stations when many people decide to drive electric, since they can attract those plug-in EV adopters to shop in their mall/market while charging. Vice versa, consumers will decide to purchase plug-in EV only when the state can provide them with sufficient charging infrastructure.
Also, businesses and other non-profit entities may not be aware of the program at first, and prompted to take advantage of it later. The effect of the incentive program is computed as the difference between the number of charging stations per 100,000 persons of Missouri and its synthetic cohort. This difference overtime is reported over eight quarters from 2015, which is summarized in table (5) below. The counterfactual number of charging stations of Missouri is also computed for each quarter, which can also be found in table (5).
Table (5). Difference between Missouri and its synthetic cohort, in numbers Period Difference Counterfactual Real values
2015Q1 0.089 70 75 2015Q2 0.536 78 111 2015Q3 0.766 83 129 2015Q4 2.24 90 226 2016Q1 2.61 101 260 2016Q2 3.20 112 307 2016Q3 4.02 116 361 2016Q4 5.22 125 443
Nguyen 26
The counterfactual column in the table presents the total number charging stations available in Missouri in a given quarter without the incentive program. Based on the synthetic control, the counterfactual in 2015 quarter 4 is 90, while the real amount of stations reached 226, which made up a difference of around 151%. The counterfactual for 2016 quarter 4 is 124, and the observed amount of stations was 443, which results in a difference of around 251%. On average, the difference between those two entities over the course of eight quarters is 2.33 charging stations per 100,000 persons, a much smaller amount than the yearly version and the diff-in-diff approach. Overall, the synthetic control reported the result that the incentive program increases the total number of charging stations by 151% by the end of 2015, and 251% by the end of 2016 in Missouri. Importantly, it may well be the case that the magnifying effect of the program is obtained thanks both to the incentive program and the network effects between the number of charging stations and the number of plug-in EV in a state.
I also performed a comparison between the results using the diff-in-diff method and those using the synthetic control approach. Therefore a synthetic control, also using yearly dataset instead of quarterly dataset was conducted. The procedure and results can be found in the Appendix.
6.4. Testing significance by using placebo test
The synthetic control method does not produce inferences on confidence intervals or standard errors of estimated values. Thus, the method did not equip me with the significance level of the effect of the incentive program as in the diff-in-diff approach. In order to test the significance of the estimates, McClelland and Gault (2017) and Abadie et.al. (2010) have suggested the use of the placebo test. The test is performed as follow. First, the synthetic control method is applied to
Nguyen 27
every of the 14 control states, as if each of them was the state that has the incentive program enacted in 2015. The donor states for each of the control states are 14 other states in the sample, including Missouri.
Then, the post-treatment gap between each control state and its synthetic cohort is compared to that of Missouri. Since those control states do not have the incentive program, their number of charging stations per 100,000 persons should be close to that of their synthetic cohort in both pre and post-treatment periods. In other words, if the gap between a state and its synthetic cohort created by the synthetic control is the largest for Missouri, then there is evidence that the program is effective. On the contrary, if the gap for every control state is random and similar to that of Missouri, then the post-treatment gap between Missouri and its synthetic cohort is randomly and coincidentally generated by the synthetic control. In this case, the estimates may be insignificant. McClelland and Gault (2017) have suggested that a significant estimates should be interpreted as the acceptance of the alternative hypothesis (the treatment does have an effect on the outcome variable), rather than a rejection of the null hypothesis.
Figure (4) summarizes the gap between 15 states in the sample and their synthetic cohort overtime. As in the graph, the gap between Missouri and its synthetic cohort is nearly zero during the pre-treatment period (before 2015, or period 220), representing the fine matching of the two. The gaps for other 14 control states were also sufficiently small for the pre-treatment period, indicating good fit of all the synthetic cohorts with the observed values. The gap for all states began to widen beginning in the first quarter of 2015 (period 220). The graph shows that the difference can be either negative or positive, which means that some states have a larger
Nguyen 28
amount of charging stations than their synthetic cohorts, while some have a smaller amount. It is clear that the gap between Missouri and its synthetic cohort is positive, and being the largest among all 15 states. This suggests that the incentive program does have a positive effect on the number of charging stations in Missouri.
Figure (4). Number of stations per 100,000 persons gaps in Missouri and placebo gaps in 15 control states
The state of Arkansas possesses the second largest gap, with its real number of charging stations being lower than its synthetic cohort. The difference between them is around -3.7 stations per 100,000 persons in 2016, quarter 4. It is suspected that this relatively large gap may be created by poor fit of the predictors. I therefore have employed different sets of predictors, either adding more or taking out some lagged values of the outcome variable (number of stations per 100,000 persons). Almost all sets of predictors produce very similar results, except for one: the same predictor set as before, but without the value of the outcome variable in 2012, quarter 4.
Nguyen 29
The gaps of all the states using this set of predictor is summarized in figure (5). The graph shows that the gap for Arkansas has shrank substantially, which further emphasizes the largest gap for Missouri. Importantly, the gap for Missouri remains nearly unchanged in both pre and post-treatment in figure (5) when compared to that in figure (4). However, the gaps for other states vary more in figure (5), mostly in the pre-treatment periods. This indicates that the set of predictors without the lagged value in 2012, quarter 4 is not necessarily more superior than that with the lagged value. Nonetheless, the gap for Missouri is the largest among all states in all set of predictors that were tested. Therefore, it can be concluded that the synthetic control estimates are significant. The tax credit program does have an increasing/magnifying effect on the number of charging stations, using the results from both the diff-in-diff approach and the synthetic control method. However, due to the possible biasness of the diff-in-diff results, the estimates computed using the synthetic control method will remains the main method of the paper to estimate the effectiveness of the incentive program on plug-in EV adoption.
Figure (5). Number of stations per 100,000 persons gaps in Missouri and placebo gaps in 15 control states (without 2012 quarter 4 as predictor)
Nguyen 30
7. Plug-in EV adoption and charging stations model
Any incentive program that promotes constructing of charging stations is expected to affect the number of plug-in EV only through the effect it has on the number of stations. It is therefore necessary to also obtain the effect that the number of charging stations has on the number of plug-in EV and its size. For example, a charging station incentive program increases the number of stations 50% in a given year. Additionally, each 1% increase in the number of charging station will result in a 1.2% increase in the number of plug-in EV. The effect of the incentive program on the number of plug-in EV would then be 1.2 * 50, which is a total of 60%.
As discussed earlier, the number of plug-in EV and of charging stations will determine each other. This results in the structural model with two regressions:
𝑙𝑛𝑃𝐸𝑉𝑖𝑡 = 𝑐𝑜𝑛𝑠𝑡 + 𝛽1𝑆𝑇𝐴𝑇𝐼𝑂𝑁𝑖,𝑡−𝑝+ 𝛽2𝑙𝑛_𝑖𝑛𝑐𝑜𝑚𝑒𝑖,𝑡 + 𝛽3𝑙𝑛_𝑝𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑖,𝑡 + 𝛽4𝑙𝑛_𝑔𝑎𝑠𝑖,𝑡−𝑞 + 𝛽5𝑠𝑢𝑏𝑝𝑐ℎ𝑖,𝑡+ 𝛽6𝑤𝑖𝑛𝑡𝑒𝑟𝑖𝑡+ 𝜖𝑖𝑡 (1)
𝑆𝑇𝐴𝑇𝐼𝑂𝑁𝑖𝑡 = 𝑐𝑜𝑛𝑠𝑡 + 𝛾1𝑙𝑛𝑃𝐸𝑉𝑖,𝑡−𝑧 + 𝛾2𝑠𝑢𝑏𝑠𝑡𝑖𝑡+ 𝜇𝑖𝑡 (2)
Where 𝑙𝑛𝑃𝐸𝑉𝑖𝑡 is log of the monthly sales of plug-in EV of state i at time (month) t.
𝑙𝑛_𝑖𝑛𝑐𝑜𝑚𝑒𝑖,𝑡 is log of yearly income of a state. Similarly, 𝑙𝑛_𝑝𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑖,𝑡 is log of population.
𝑠𝑢𝑏𝑝𝑐ℎ𝑖,𝑡 is the dummy variable for state purchasing incentive availability, which will equal 1 if
Nguyen 31
is the lag of log of gasoline price. I included the lag of gasoline price since it is likely that people will not respond to a change in gasoline price by buying a plug-in EV right away. In this
analysis, q = 6, since US consumers are often well aware of the volatility of oil price. Hence, gasoline price will likely affect their decision only when they observe a clear and relatively long run upward or downward trend of oil price. Also, consumers may not purchase a plug-in EV right after they decide to purchase one, partially due to financial status. It may take several months to save or to complete an auto loan. 𝑆𝑇𝐴𝑇𝐼𝑂𝑁𝑖,𝑡−𝑝 is the lag of log of charging stations quantity. This variable can be the log of cumulative/total number of stations available in a given month and year (variable STATION), or the log of number of newly built stations in a given month and year (variable STATION1). Consumers’ decision to purchase plug-in EV will likely depend partially on both the total number of charging stations available and the change in the number of new charging stations being built. If they observe an insignificant but rapidly growing number of stations, they will also be more eager to purchase a plug-in EV.
In addition, consumers will not be likely to respond to an increase in the number of stations quantity by purchasing plug-in EV immediately. Purchasing a car is costly and requires a lot of paperwork in the US. Thus, similar to a change in gasoline price, people may decide to buy a plug-in EV after observing a desirable number of charging stations available, but they may need to save for the purchase, to file auto loan paperwork, to do research about the car models, or to arrange accommodate for the plug-in EV in their garage. Therefore, it is likely that the past value of the number of charging stations rather than the current value that has a significant effect on current plug-in EV sales. I chose p = 12 in order to take into account the amount of time necessary for the preparation and purchase of a plug-in EV.
Nguyen 32
To take into account to some degree of time fixed effect, the dataset includes the variable winter, which equals 1 if it is during the winter (December, January and March), and 0 otherwise. Finally, the variable subst in (2) is the dummy variable for the availability of an
incentive/subsidy program for charging stations. It equals 1 if there is an incentive program in the state, either a governmental or utility/private one, and 0 otherwise. IVs are needed for both the number of charging stations in (1) and the number of plug-in EV in (2). Subst will be the IV for 𝑆𝑇𝐴𝑇𝐼𝑂𝑁𝑖,𝑡−𝑝 in (1), and subpch will be the IV for 𝑙𝑛𝑃𝐸𝑉𝑖,𝑡−𝑧 in (2). I chose I chose z = 6, as businesses and other organizations/entities may respond to the increase in plug-in EV quantity much more quickly than consumers respond to the increase in charging stations quantity, since building charging stations may be indirectly profitable. However, the construction and
paperwork procedures may take several months. Therefore, a period of around six months was employed for new stations to be available for public. State income per capita, population, gasoline price and the availability of purchasing subsidy are included as control variables due to the possible effect of them on plug-in EV quantity, as discussed in previous literature.
The model comes with possible limitations. First, the model examines the effect of the amount of charging stations in the past on the current sales quantity of plug-in EV, and vice versa.
However, the amount of current charging stations (plug-in EV) can also be affected by the future quantity of plug-in EV (charging stations). This results from the positive expectation of
consumers of an increasingly growing plug-in EV market. This factor is not incorporated in the model due to technical limitation. Also, there may be omitted variables in both (1) and (2), due to lack of data.
Nguyen 33
Additionally, this study is not able to address the continuous effect of the number of plug-in EV on the number of charging stations, and vice versa resulted by the network externality and the incentive program simultaneously. Besides, annual data for the control variables were employed rather than the monthly data, due to the availability of data. Therefore, there might be small variations in the final results of the coefficients of the number of newly stations in a month, the total number of stations, and log of plug-in EV monthly sales. Finally, the data that I collected and performed regressions upon in the study is state level. Therefore, data on micro-level might also give different results.
7.1.Plug-in EV adoption model
The statistic summary is presented in table (6). Plug-in EV includes two types: PHEV (plug-in hybrid EV) and BEV (battery EV). PHEV and BEV represent monthly sales of each type. Monthly EV is the monthly sales of both plug-in hybrid EV and battery EV. Cumulative EV is cumulative sales of plug-in EV. New station is the number of new stations being built at time t. Total station is the total number of stations available at time t. Population denotes state
population, and gas price is gasoline price, in dollar per gallon. There is substantial difference between states for all variables but gasoline price. Besides the differences in these features, states also vary greatly in terms of unobservable characteristics, such as culture, rules and laws, or political perspective (which might also affect plug-in EV adopting decision), which will be taken into account using fixed effect option.
Nguyen 34
Table (6). Statistic summary, plug-in EV model
Variable Obs Mean Std. Dev. Min Max PHEV 2,880 44 88 0 3347 BEV 2,880 42 95 0 1344 Monthly EV 2,880 86 145 0 3375 Cumulative EV 2,880 2370 4045 0 25616 New station 2,880 4.4 12.5 0 506 Total station 2,880 129 179 0 1167 Population (million) 2,880 6.21 5.69 0.567 27.9 Income ($,approx) 2,880 47,000 9050 31,200 71,000 Gas price ($/gallon) 2,880 3.05 .571 1.96 3.79 Density (sq.mile) 2,880 194.33 257.67 5.85 1220.89
In order to obtain better information regarding the quantity of plug-in hybrid EV and battery EV, this section also studies the trend of total monthly sales of 40 states of plug-in hybrid EV and battery EV, which is shown in figure (6). From the figure, we can observe that the sales of both plug-in hybrid EV and battery EV have increased rapidly from 2011 to the first half of 2013. However, the rapid increase has ceased afterwards. Since 6/2013, sales of both types have experienced intermittent downward and upward trends, with the sales of each ranging from around 1,800 to 4,000 from 6/2013 to 12/2016. In addition, sales of plug-in hybrid EV in general experiences less volatility than that of battery EV. This may indicate that consumers need more time to make purchasing decision for a battery EV than for a plug-in hybrid EV. Importantly, in contrary to a common belief in the US that consumers prefer plug-in hybrid EV to battery EV due to plug-in hybrid EV’s convenience in terms of fuel refilling, the figure shows that total monthly sales of plug-in hybrid EV does not generally far exceed that of battery EV, except for 2012 and some months in 2014. This may be due to the increasing in the number of charging
Nguyen 35
stations over time, the higher (generally) federal and state subsidy amount for battery EV, and/or a sense of satisfactory when experiencing a new fuel.
Figure (6). Total monthly sales of plug-in hybrid EV and battery EV
It is necessary to test the relevance of station subsidy availability as an IV for the number of charging stations. To achieve this, the first stage is performed by regressing the log of total number of charging stations available at time t (variable STATION) on the dummy for station subsidy availability and included control variables in (1) was performed. Similarly, I regressed the log of number of newly built stations at time t/in a month (variable STATION1) on the dummy for station subsidy availability and the control variables.
Figure (7) reports the results of these two first stage regressions. In both regressions, the coefficients of station subsidy availability are positive and significant at 0.01 level. This is an expected result, as states with charging station incentive programs will generally experience higher growth rate of the number of stations and ultimately have more stations than states
Nguyen 36
without (as proven earlier). The result therefore confirms the relevance of station subsidy availability as an IV for the total number of stations and the number of newly built stations in a month. Regarding the exogeneity of station subsidy availability, it is expected that an incentive program designed to promote plug-in EV sales only by supporting the construction of charging stations.
Figure (7). IV relevance (first stage)
(1) (2)
VARIABLES total station new station
subst 1.151*** 0.684*** (0.0605) (0.0441) Constant -11.52*** -10.47*** (1.412) (1.031) Observations 2,880 2,880 R-squared 0.614 0.348
Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
Using station subsidy availability as an IV for the total number of stations and the number of newly built stations in a month, two versions of (1) were estimated: the first version studies the effect of the 12th lag of total number of stations on the plug-in EV monthly sales. Meanwhile, the second version estimated the effect of the 12th lag of the number of newly built stations in a
month on the plug-in EV monthly sales, using the same set of control variables. For each version, I estimated the model using both OLS and 2SLS method. The regression results can be found in figure (8). Column (1) and (2) reports the results for OLS and 2SLS estimates for the first version, respectively. Column (3) and (4) reports the OLS and 2SLS estimates for the
Nguyen 37
second version, respectively. Column (5) includes the OLS estimate for the regression of log of plug-in EV monthly sales (𝑙𝑛𝑃𝐸𝑉𝑖𝑡) on both the 12th lag of log of total number of stations
(𝑆𝑇𝐴𝑇𝐼𝑂𝑁𝑖,𝑡−12) and the 12th lag of log of the number of newly built stations in a month
(𝑆𝑇𝐴𝑇𝐼𝑂𝑁1𝑖,𝑡−12). This OLS regression was included to have better examination on the coefficients of the control variables.
Figure (8). Plug-in EV regression result
(1) (2) (3) (4) (5)
VARIABLES total station
OLS total station 2SLS new station OLS new station 2SLS OLS STATION (12th lag) 0.165*** 0.388*** 0.161*** (0.0108) (0.0839) (0.0109) Ln_income -2.253*** -6.487*** 0.819 0.390 -2.210*** (0.724) (1.765) (0.726) (0.849) (0.724) Ln_population 6.160*** 2.010 9.023*** 7.654*** 6.141*** (1.000) (1.891) (1.024) (1.245) (0.999) Ln_gas (6th lag) -0.150** 0.222 -0.380*** -0.0826 -0.137* (0.0728) (0.160) (0.0745) (0.119) (0.0731) subpch 0.214*** 0.156*** 0.260*** 0.278*** 0.217*** (0.0418) (0.0504) (0.0436) (0.0507) (0.0418) winter -0.254*** -0.219*** -0.284*** -0.312*** -0.256*** (0.0224) (0.0276) (0.0233) (0.0281) (0.0224) STATION1 (12th lag) 0.0564*** 0.418*** 0.0233* (0.0127) (0.101) (0.0124) Constant -66.39*** 41.17 -142.2*** -117.4*** -66.59*** (15.16) (43.38) (14.90) (18.58) (15.15) Observations 2,400 2,400 2,400 2,400 2,400 R-squared 0.278 0.213 0.279 Number of state1 40 40 40 40 40
Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
Nguyen 38
For the first version, the coefficient of the 12th lag of log of total number of stations is positive
and significant at 0.01 level for both OLS and 2SLS. However, the magnitude given by the OLS estimate is much smaller than that of the 2SLS, suggesting that endogeneity issue of the total number of stations may well exist. The value of the coefficient of the 12th lag of log of total number of stations using 2SLS is 0.388. This indicates that if the total number of stations increases by 1%, the sales of plug-in EV twelve months later will increase by 0.388 percent. Therefore, a station incentive program such as one in Missouri that increases the number of stations by 151% at the end of 2015 can result in an increase of plug-in EV monthly sales by around 58.5% at the end of 2016. If the program increases the number of stations by 251% in 2016 then plug-in EV monthly sales will expect to increase around 97% at the end of 2017.
The coefficient of log of population is positive, but not significant at any level in the 1st version 2SLS result (column 2).The result indicates that more populated states might have slightly higher monthly sales of plug-in EV. This is quite counter-intuitive, as states with a larger population tend to have more charging stations, ceteris paribus. However, the coefficient is positive and significant at 0.01 level in the 2nd version 2SLS result (column 4) and a magnitude of 7.654. Additionally, the coefficient is also positive and significant in every OLS regression (column 1, 3 and 5) with magnitude at least 6.16. Therefore, it is likely that the result reported for the 2SLS regression of the 2nd version for log of population is more suitable than that of 1st version.
The coefficient of the 6th lag of log of gasoline price is positive and not significant at any level in the 2SLS regression for version 1. It is negative yet also insignificant in the 2SLS regression for
Nguyen 39
version 2. This implies that a change in gasoline price has little impact on plug-in EV sales, and gasoline price is an unimportant factor in deciding whether to purchase a plug-in EV. Plug-in EV sales includes both plug-in hybrid EV and battery EV sales. Also, gasoline still remains the main fuel for most plug-in hybrid EV models. Hence, it may be the case that consumers generally purchase plug-in hybrid EV as a substitute for an ordinary all-gasoline vehicle and fuel the car mostly with gasoline to take advantage of purchase subsidies, which are not granted to all-gasoline vehicles. On the contrary, battery EV adopters fuel their vehicle with only electricity, and gasoline price may be an important deciding factor them. I will test this hypothesis later.
The coefficient of purchasing subsidy availability (subpch) is positive as expected, and
significant at 0.01 level in both OLS and 2SLS regressions for both versions. The value for the coefficient in the 2SLS regression of the first version is 0.156, while that of the second version is 0.278. This means that the availability of a state purchase incentive program (either state run or utility/private run) on average results in an increase of 15.6-27.8% of plug-in EV monthly sales. Interestingly, the coefficient of log of state income per capita is negative and significant at 0.01 level in both regressions for version 1. However, it is positive yet not significant at any level for version 2. Overall, the result for this variable is mixed for five regressions. Therefore, the income per capita of a state may ambiguous effect on plug-in EV sales.
A possible explanation for the negative coefficient is the change in the perception of consumers over time. Contrary to early plug-in EV adopters, who adopt plug-in EV mainly to make a statement or to prove their environmental awareness (Li et.al., 2017), late adopters are likely to
Nguyen 40
be more sensitive to price and financially bound. This can be prove partially through the positive and significant coefficient of purchasing subsidy availability. Also, the state cultures and
subsidies may explain for the insignificance of the coefficient in column (3) and (4). Although a particular state may be richer than another, but that state may not have an “EV culture”. One example could be North Dakota, where the income per capita in 2016 was more than $64,000, yet the state has only 157 plug-in EV sold from 2011 to 2016. On the contrary, one state may be relatively poor. However, the wealthy individuals of that state may be interested in plug-in EV, or there may be generous state purchase subsidies, which can boost the sales of plug-in EV.
Similar to the coefficient of the 12th lag of log of total number of stations, the coefficient of the 12th lag of log of number of newly built stations in a month is positive and significant in both OLS and 2SLS regressions, with large difference in the magnitude between two regressions. The magnitude of the coefficient in the 2SLS regression is 0.418. This means that as the number of newly built stations increases 1%, the monthly sales of plug-in EV will increase 0.418% twelve months later. This further solidifies the hypothesis that consumers react to both the stock and flow of charging stations. The coefficient of the dummy for winter months is negative in all regressions, and significant at 0.01 level also in all regressions. This suggests that consumers overall purchase less plug-in EV in winter months than in other months. The magnitude of the coefficient for the 2SLS regressions of both versions are 0.291 and 0.312. This indicates that monthly sales of plug-in EV in those months decrease from 29.1% to 31.2%, which is a sharp drop. One possible explanation for this is that plug-in EV in general suffer from major range reduction when being operated in cold weather (Howard, 2013). Also, consumers may have an incentive to purchase early, so that paperwork can be completed on time and before the tax credit
Nguyen 41
application, which starts at the beginning of February each year.
In addition to the plug-in EV regression model, the section also includes a similar regression with the dependent variable being the log of monthly sales of only battery EV (variable lnBEV) to study the possible different effect of gasoline price, purchase subsidy and the number of charging stations on battery EV sales than on plug-in hybrid EV sales. The regression results can be found in figure (9). For this regression model, only the 2SLS regression was performed, due to acknowledgement of the endogeneity of the total number of stations and the number of newly built stations in a month. Column (1) shows the result for the regression using the 12th lag of log of total number of stations as independent variable, while column (2) shows the result for the regression using the 12th lag of log of the number of newly built stations in a month as independent variable. Similarly to the plug-in EV regression model, an OLS regression that includes both station variables as independent variables was included (column 3).
Nguyen 42
Figure (9). Battery EV regression result
(1) (2) (3)
VARIABLES OLS BEV and total
station (2SLS)
BEV and new station (2SLS) STATION1 (12th lag) -0.0180 0.735*** (0.0159) (0.153) STATION (12th lag) 0.335*** 0.683*** (0.0140) (0.112) Ln_income -0.402 -7.041*** 5.069*** (0.927) (2.347) (1.290) Ln_gas (6th lag) 0.259*** 0.855*** 0.319* (0.0937) (0.213) (0.181) Ln_population 7.808*** 1.252 11.19*** (1.280) (2.516) (1.891) subpch 0.299*** 0.209*** 0.424*** (0.0535) (0.0670) (0.0771) winter -0.138*** -0.0851** -0.249*** (0.0287) (0.0367) (0.0427) Constant -113.5*** 55.87 -223.3*** (19.41) (57.70) (28.22) Observations 2,400 2,400 2,400 R-squared 0.371 Number of state1 40 40 40
Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
The coefficients of both the 12th lag of log of total number of stations and the 12th lag of log of number of newly built stations in a month in the 2SLS regressions are positive and significant at 0.01 level. Their magnitude are 0.683 and 0.735, respectively. This means that an increase of 1% in the total number of charging stations results in an increase of 0.683% in monthly sales of battery EV twelve months , and an increase of 1% in the number of newly built stations results in an increase of 0.735% in monthly sales of battery EV. Apparently, these values are much higher than those of the plug-in EV regression model. It shows that the sales and ultimately the number
Nguyen 43
of battery EV depend much more heavily on the number of charging stations than do plug-in hybrid EV, which is reasonable and intuitive.
The coefficient of the 6th lag of log of gasoline price are positive for both regressions and significant at 0.01 level (for column (1)) and 0.1 level (for column(2)). The magnitude for each regression is 0.855 and 0.319, which means that an increase of 1% in gasoline price leads to an increase of 0.319-0.855% monthly sales of battery EV. This helps demonstrate the sensitiveness of battery EV adopters in compared to plug-in hybrid EV adopters. As suggested above, late adopters of battery EV may concentrate more on financial restrictions than first movers.
Therefore, long run monetary saving earned from using electricity can be an important incentive for them to purchase battery EV.
The coefficient of purchasing subsidy availability is positive and significant at 0.01 level for both regressions. Their magnitude is 0.209 for column (1), and 0.424 for column (2). These values are larger than those of the plug-in EV model, which further suggests that financial subsidies affect battery EV sales to a greater extent than they do for plug-in hybrid EV, and that battery EV adopters stress the importance of financial restrictions and subsidies more than plug-in hybrid EV adopters. Similar to the plug-in EV model, the coefficient of the dummy for winter months is negative in both regressions, with significance level of 0.05 (column (1)) and 0.01 (column (2)). However, their magnitude (0.0851 in column (1) and 0.249 in column (2)) are smaller than those of the plug-in EV model. This suggests that plug-in hybrid EV sales drops more than that of battery EV during winter months.
