The spillover impacts of urbanization and energy usage on CO2 emissions: A regional analysis in the United States

This study endeavors to identify how urbanization and energy factors (use, prices, and policies) have influenced CO2 emission patterns per capita among the 48 U.S. states and the District of Columbia from 2000 to 2015. To examine interconnections between state-level emissions, three different spatial panel data models are estimated: the spatial autoregressive model, spatial error model, and the spatial Durbin model models. This study makes contributions to the literature by applying a spatial panel data approach and including policies on urbanization and renewable standards. Urbanization is found to have a statistically significant, negative direct effect on emissions, yet a negative spillover effect such that a 1.0% increase in state i's urbanization level leads to a 0.30% decrease in its own state's emissions, but a 0.012% increase in an adjacent state j's emissions. Although state-level emissions are increased by higher energy use and coal consumption, both energy prices and renewable portfolio standards decrease emissions both within their own and adjacent states.


Introduction
Due to concerns about global climate change, the public has an interest in policies that advocate limiting the rise of CO 2 emissions. Studies of urban environmental transition emphasize that urbanization can have both negative and positive effects on the environment making it difficult to determine the net effect a priori (Cole and Neumayer, 2004;Liddle and Lung, 2010). Examining the linkages among CO 2 emissions, urbanization, and other factors are significant as the United States is the second largest aggregate emitter globally from 2020 to 2050, based upon its share of global energy-related CO 2 emissions and global roadmap for building and construction (International Energy Agency [IEA], 2020). The availability of state-level data from EIA and the Carbon Dioxide Information Analysis Center (CDIAC) is the basic information required to study CO 2 emissions, urbanization, energy consumption, and socioeconomic variables in the United States ( Figure 1). This research investigates how urbanization and other variables reflecting energy use impact carbon emissions in the United States.
Furthermore, a controversial discourse has recently advanced about the subject of whether CO 2 emissions measurement must be based upon space-related consumption or production method (Davis et al., 2011;Peters et al., 2011). Production-based accounting (PBA) of CO 2 emissions for fossil fuel consumption estimates the greenhouse gas emissions from all the coal, crude oil, and gas consumed in a state by power plants, residential, and industrial production of goods and services. While the PBA has several drawbacks. First, it excludes CO 2 emissions driving from international trade and transportation. Since such CO 2 emissions do not occur within a specific state, its impact on specific territories is complicated. Second, industries that are energy-intensive in states with CO 2 emission regulations and taxes may move into other states with fewer regulations and lower costs of energy. However, the goods and services produced in the less regulated states could then be exported to the more regulated states. Hence, reducing CO 2 emissions in one state might be directly correlated with rising CO 2 emissions in other states. Third, the CO 2 emission leakage from the outsource production of carbon-intensive goods to the United States is a reallocation of CO 2 emissions to states. Thus, one state's production can also be driven by consumption in other states. Consumption-based accounting (CBA) is an approach to consider these issues. It subtracts all CO 2 emissions that are incorporated in exported goods and services from states, including CO 2 emissions of transportation, and the contained CO 2 emissions in the inventories of the importing states (Aldy, 2005;Auffhammer and Carson, 2008;Peters et al., 2011;Tawfeeq et al., 2020). Hence, with respect to production-based inventories, low-emission states may be less clean in the CBA approach and high CO 2 emissions states could produce goods for the standard of living of low CO 2 emission states. Thus, previous research has found that CO2 emissions are primarily driven by city size, population, GDP, and an economy's energy consumption.
We utilized several complementary methods to investigate the problem in this study, arguing that the contradiction between urbanization and CO 2 emissions arises not only from different scales of analysis (i.e., state versus national) but also from limitations in how previous studies theorized and estimated the factors of CO2 emissions and urbanization. To address this shortcoming, we synthesize insights from nonspatial and spatial econometric methods to develop a framework for understanding how urbanization and energy factors affect CO 2 emissions across 48 states and District of Columbia in the United States. Hence, the spatial linkage is examined by determining spatial panel data models that are control for spatial impacts over time and space (LeSage and Pace, 2009;Aldy, 2005). Indeed, estimated parameters might be biased when spatial correlation is not considered.
Using spatial econometric models is an adequate approach to study the effects of urbanization, per-capita gross state product (GSP), energy use, energy prices, and coal consumption on CO 2 emissions by state level from 2000 to 2015. Additionally, the spatial reliance signifies that an energy policy implemented in one state could have spillover impacts in neighboring states (Kindle and Shawhan, 2011). The estimate of such spillovers is significant to determine the direct and indirect impacts of state-level policies adopted in the United States which affect the level of CO 2 emissions. Hence, this study attempts to control spatial correlation in estimating the impact of urbanization and other driving forces on CO 2 emissions.
In this context, the objective of this study is to examine the spillover impacts of changes in energy use and urbanization growth on CO 2 emissions in the US. More importantly, the main goal of this study is to provide a basic reference for inverters and policy makers to make proper decisions and set CO 2 emission reduction targets and policies.
This research makes novel contributions to literature through its use of a spatial panel data approach. More specifically, we have taken into consideration urbanization and energy-related factors on CO 2 emissions when examining state-level panel data for 2000-2015. This time period captures recent developments in state-level economic growth and energy use policies, like state-level passage of renewable portfolio standards (RPS). Moreover, this period covered prerecession and post-recession of 2008 in the U.S. economy.

Theoretical model
Urban areas are the center of energy consumption and the consumption of fossil fuels in transportation, electricity, and industrial goods, all of which generate CO 2 emissions. However, many power plants and industrial firms are in rural regions and they burn fossil fuels-emitting a high amount of CO 2 and polluting the entire environment. It has been theoretically shown that an increase in energy use occurs in urban areas relative to higher economic activities and income levels (Canadell et al., 2007). Moreover, Itkonen (2012) assumes that energy consumption and income are linearly associated. Itkonen (2012) modifies the EEO model by examining the effect of income on CO 2 emissions when the consumption of energy has a positive linear relationship with income. Also, Jaforullah and King (2017) model CO 2 emissions, using the EEO approach to identify determinants of CO 2 emissions. The EEO model, although, indicates that energy consumption is a function of income; energy use and income are two critical exogenous variables of CO 2 emissions. However, based upon their model, they concluded that the consumption of energy was not an independent determinant of CO 2 emissions (Jaforullah and King, 2017).
We argue that energy consumption can be used as an independent variable to explain CO 2 emissions. First, most of the empirical studies on the relationships between energy use, GDP, and CO 2 emissions have included energy consumption as a main independent variable (Soytas et al., 2007;Martinez-Zarzoso and Maruotti, 2011;Wang, et al., 2011;Chuai et al., 2012;Tawfeeq et al., 2019;Sadorsky, 2014). This literature has focused on the linkages between CO 2 emissions, income, and energy consumption by utilizing energy use variables as explanatory variables. Second, a single study by Jaforullah and King (2017) may not be generalizable to ignore the energy consumption effect on CO 2 emissions. Lastly, the spatial spillover effects of both income and energy consumption on CO 2 emissions in other states will be explored in this research; hence, both variables are included in the model below.
The idea of spatial spillovers among economic activities is related to the concept of economic distance, which indicates that the closer two locations are to one another in a geographic distance, the more possibility that their economies are interconnected (Conley and Ligon, 2002). Spatial linkages suggest that policies adopted in one region would impact whether policies are implemented in neighboring regions. Thus, a U.S. state is more likely to adopt a law and or policy if its neighboring states have already done so (Mooney, 2001). In fact, geographical location has been identified as a critical factor of cross-region economic growth due to indicators like the diffusion of technology (Tawfeeq et al., 2019). We could argue that CO 2 emissions may decrease with technological development, then the diffusion of technology would likely enhance the conditions of neighboring environment.
Furthermore, it is considered that spatial panel data models have been divided into two categories. One is nondynamic, which has been utilized in the context of forecasting recently (Elhorst, 2009;Baltagi et al., 2012). Another spatial panel is dynamic which controls for either time-invariant heterogeneity across geographical areas or spatial autocorrelation between areas (Anselin, 2002;Anselin et al., 2008). Since the main empirical aim of this study is to precisely model the main drivers of CO 2 emissions at the U.S. state level, we formulate dynamic spatial panel data models.
Given the theoretical concepts cited in the literature above, we can formulate an equation as follows: where Y i represents per capita CO 2 emissions, X i is an n x k matrix of state-level characteristics including urbanization rate, GSP, the consumption of energy, coal consumption, and energy prices in state i and time-period t. RPS it includes a dummy variable to describe a renewable energy policy proxy at the state level during each time period. A Hausman test (Hausman and Taylor, 1981) is conducted to select panel data specification between fixed and random effects: The linkages between energy use, urbanization, and CO 2 emissions will be examined while controlling for possible spatial impacts in the panel data. To account for this insight, a term for spatial linkage considering the association between states, arguing that there exist possible spatial linkages between state-level economic activities and energy use, which in turn produce CO 2 emissions (Auffhammer and Carson, 2008). Spatial linkages can be constructed in various ways, depending on the relationship between the dependent variable and the explanatory variables. When specifying interaction between spatial units, the model may contain a spatially lagged dependent variable or a spatial autoregressive process for the error term, including the spatial autoregressive model (SAR), spatial lag model (SLM), or a spatial error model (SEM) (Anselin et al., 2008). The spatial Durbin model (SDM) contains a spatially lagged dependent variable and spatially lagged independent variables (LeSage and Pace, 2009). All spatial models have a weight matrix (W), which quantifies the spillover between regions. Elhorst and Elhorst (2014) names the weight matrix as a tool to explain the spatial linkages of the geographical units in the sample. There is a variety of units of measurement for spatial dependency such as neighbors, distance, and links (Getis, 2007). In this study based upon the nature of the research, we conduct and apply a variety of weight matrices.
The main factors which lead to spillover impacts across states allow us to include spatial lags. First, there exist more than 3200 fossil fuel-fired power plants within all states in the United States (EIA, 2016(EIA, , 2020, and these power plants are likely to spillover emissions to neighboring states. The second factor giving rise to spatial lags is urbanization and its impact on the own state and neighboring state. The third factor is interstate goods transportation which leads to consider spillover effects of pollution across states for energy consumption. These would suggest that state-level urbanization and energy use would lead to state spillover impacts of CO 2 emissions by spatial lags. The spatial dependence method provides the theoretical basis for the literature on interstate level impacts of CO 2 emissions. Anselin (2002) states two important motivations for considering spatial impacts in regression models from a theory-driven as well as data-driven perspective. A theory-driven framework follows from the formal specification of spatial interaction, which are interacting agents and social interaction in an econometric model, where an explicit interest in the spatial interaction of a particular variable conquers and a theoretical model creates the model specification.
If we have two regions i and j that are spatially correlated and supposed error terms have normality, then: where the dependent variable in neighbor i influences the dependent variable in neighbor j and vice versa. In nonspatial models, each observation has a mean of x i β and a random component ε i where the observation i represents a location or point in space at one location and is considered to be independent of observations in other locations. In other words, statistically independent observations imply that E(ε i ε j ) = E(ε i )E(ε l ) = 0. This assumption of independence greatly simplifies models. However, a spatial approach implies that each observation corresponds to a location or region (LeSage and Pace, 2009). Consistent with the intuition of correlation between state CO 2 emissions through commerce (Carson, 2010), we integrate a term for spatial spillover effects to account for the relationship between states that might have spatial linkages within states and among neighboring states for CO 2 emissions. In other words, we argue that there is a potential spatial linkage between state-level urbanization rate, energy prices, renewable energy policies, economic activities, and state-level coal and energy consumptions which in turn generate CO 2 emissions. Accordingly, there are five types of spatial models that might consider investigating the spatial linkage between CO 2 emissions, energy and coal consumption, urbanization, energy prices, renewable energy policies, and GSP. The first type is an SLM, in which the dependent variable, CO 2 emissions, in state i is affected by the CO 2 emissions in state j. There are five different spatial models. The first one is the spatial autoregressive lag model where the dependent variable in neighbor j influences the dependent variable in neighbor i and vice versa. The second is an SEM (assumes dependency in the error term). The third is a model or spatial lag of control variables that assumes that only control variables play a direct role in determining dependent variables. Lastly, there are two other models which are SDM and Spatial Error Durbin Model that include spatial lags of the control variables as well as the dependent variable and a spatial lag of the control variables (WX), as well as spatially dependent disturbances.
In this study, we applied three types of spatial models. The first type is an SLM, in which the dependent variable, CO 2 emissions, in state i is impacted by the CO 2 emissions in state j. This specification captures spatial spillover effects of the CO 2 emissions in one state expected to increase the likelihood of similar impacts in neighboring states. The second type of spatial model involves a spatial lag of the dependent variable and a spatial lag of the explanatory variables; this is referred to as an SDM. In this model, it is assumed that there is not only spatial dependence within the dependent variable but the determinants of demand for energy such as energy prices and GSP in one state are directly impacted by neighboring states. The third type of spatial dependence is SEM in which the error terms across different spatial units are correlated. With spatial error in an ordinary least squares regression, the assumption of uncorrelated (independently distributed) errors is violated, and as a result, the estimates are inefficient. We will compare the model specifications of each of these spatial models against the traditional, nonspatial panel data model.

Empirical model
In this study, we examine the linkages between energy factors, urbanization, and CO 2 emissions while controlling for possible spatial impacts in the panel data. To test and estimate the proper spatial model, this study selects an ordinary panel model, spatial lag model (SAR), SEM and the SDM. In addition, the SDM is utilized to determine direct and indirect effects which contain a spatially lagged dependent variable and spatially lagged independent variables (LeSage and Pace, 2009). According to the variables that are described, we can have a main empirical model to estimate which is given by the following: ln(c it ) = β 0 + β 1 ln(ur it ) + β 2 ln(ec it ) + β 3 ln(cc it ) + β 4 ln(gsp it ) + β 5 ln(ep it ) + β 6 rps it + μ i + η t + ε it where c it represents to per capita CO 2 emissions, ur it is urbanization rates, ec it represents to total energy consumption, cc it represents the consumption of coal, gsp it is per capita GSP, ep it denotes energy prices, and rps it is dummy variable for renewable energy policy which is RPS for state i over time t. β 1 , β 2 , β 3 , β 4 , β 5 , β 6 are coefficients of explanatory variables. Moreover, all variables are converted to natural logarithms to interpret the coefficients as elasticities. 1 The parameter μ i denotes the entity effect (or heterogeneity) for each state and η t denotes a common time fixed effect. The entity effect is fixed meaning that it is assumed that this variable is correlated with the explanatory variables and approximately fixed over time for each state within the sample. If the empirical model estimates with no controlling for the entity effect, then estimation might result in omitted variable bias, while the fixed effect is correlated with the explanatory variable. The entity effect can be interpreted as characteristics within states that do not change over time such as unobservable geographic characteristics such as lakes and rivers. The time period impacts control for time-specific shocks that influences all states in a given period of time, for example, renewable energy standard policy (RPS) that impact CO 2 emissions in the most states in the U.S. boundary. Following the previous discussion, there are three main types of spatial panel specifications to explore spatial relationships between variables. Consistent with the insight of a simple spatial method, the Spatial Autoregressive Regression (SAR) model is: where y it is a (n * 1) vector of CO 2 emissions, N j=1 W ij Y jt is the prespecified n * n matrix of spatial interaction impacts, ρ is a spatial autocorrelation coefficient, and X is a matrix of explanatory variables including urbanization rate, energy prices, per capita GSP, RPS, the consumption of coal, and energy use. β is a 1 * k vector of parameters to be estimated, and ϵ is a vector of errors (Anselin, 2002). The SEM mode is: where λ is a spatial parameter like ρ in SAR model and all other notations are as described previously. The estimate results in the spatial models (SAR and SEM) show that the spatial coefficients (λ and ρ) are statistically significant with contiguity-based matrix, justifying the use of spatial econometric panel data models. This specification says that the error for CO 2 emissions depends on the average of the errors in neighboring observations and its idiosyncratic component, implying that the unobserved errors u and ϵ are entirely predictable from neighboring error Wϵ. Furthermore, when the endogenous variable can be predicted as a function of spatially lagged exogenous variables, we can utilize SDM. The SDM is given by: where parameter γ is a(k * 1) vector of spatial autocorrelation coefficients on the exogenous variables, and ρ denotes a scalar spatial autocorrelation coefficient of endogenous variable. The SDM can be utilized to control if the model can be simplified to an SLM or an SEM since the models nest dependence in both the disturbances and the dependent variable (LeSage and Pace, 2009).
In order to specify nonspatial panel models against the spatial models (i.e. SAR and SEM), we utilize several Lagrange Multiplier (LM) tests. These tests examine whether the spatial models approach offers a proper specification. Moreover, we explore the joint significance of state fixed effects (μ i ) and time period fixed effects (η t ) (Elhorst, 2012). The null hypothesis tests are: The LR tests are employed to investigate those null hypotheses. If the p < 0.05, then we can reject the null hypothesis of joint insignificance (Elhorst, 2012). If we fail to reject the spatial model, then the next step will be to explore whether the SDM model can be simplified to the SAR or SEM model. The hypothesis tests are: where H 0 :γ = 0 verifies whether the SDM can be simplified to the SAR, and H 0 :γ + ρβ = 0 proves whether it can be simplified to the SEM (Elhorst, 2009;LeSage and Pace, 2009). All tests follow a χ 2 distribution. If we reject both hypotheses, suggesting that the SDM is the best fit for the spatial panel data. Inversely, if we cannot reject the hypotheses, it suggests that the SAR and SEM provide the best fit for the panel data.

Data description
A panel dataset of 48 states and the District of Columbia is used from 2000 to 2015. Data for CO 2 emissions, energy consumption, coal consumption, and energy prices are obtained from the U.S EIA 2 (2016) and the CDIAC. 3 Per capita CO 2 is computed by total state CO 2 emissions divided by state population. Emissions are estimated through the emissions-energy-output model (Itkonen, 2012): where α x are thermal conversion factors for coal, crude oil, natural gas, and cement into CO 2 emissions. Energy consumption represents per capita total energy use in each state on a million BTU basis. The consumption of coal is determined by the amount of coal used by electric power plants in billion BTU. Prices are represented by state-level price indices over fossil fuel energy sources. Finally, urbanization rate represents the percentage of urban areas in each state according to the USCB 4 . The GSP variable is based on state-level economic output converted to real values with a base year of 2009. RPS is an energy policy instrument that aims to stimulate the supply of renewable energy in the U.S. electricity markets. This variable takes a value of one starting in the year that a state has passed the RPS law and zero otherwise using DSIRE 5 . Descriptive statistics for all variables are shown in Table 1, and all variables are converted to logarithms for the analyses.

Results and discussion
The Hausman test results show that the calculated χ 2 is 32.86 (Table 2), which denotes that the estimated parameters are biased under the specification of random effects. Based upon the LR tests in Table 2 for nonspatial models, we can conclude that both the state and time period fixed effects should be added to the model. As noted earlier, the Wald tests show that the SDM is the best spatial panel model (Table 2). Table 3 shows the results of the SDM. The ρ sdm coefficient is positive and statistically significant, justifying the use of spatial econometric models by indicating the presence of a spatial autoregressive impact. This coefficient shows that an increase in CO 2 emissions in neighboring states leads to an increase of approximately 0.2 times emissions in adjacent states. This result implies that if spatial dependence is ignored, interpreting estimated impacts with OLS will be incorrect.
Statistically significant and negative direct effects include urbanization rate, energy prices, RPS, and per capita GSP, while energy use and coal consumption have statistically significant positive direct effects (Table 3). These results indicate that increased urbanization, energy prices, and economic growth along with the passage of an RPS will decrease state-level CO 2 emissions, while increasing coal consumption and energy use will lead to increased emissions. Coal consumption dramatically increases CO 2 emissions with a 1% increase leading to an own state increase of 1.31% in emissions and a total effect of more than double at 2.54%. The size of the total effect for coal consumption indicates that encouraging the replacement of coal with renewable or natural gas energy plays a critical role in decreasing CO 2 emissions.  Table 4 reports the estimation results for direct and indirect effects from the SDM model. Among the direct effects, each independent variable has a statistically significant impact at a 1% level, except for coal consumption which is significant at the 5% level. Table 4 illustrates direct and indirect impacts of predictors on CO 2 emissions. Indirect effects are statistically significant apart from   Note: The following models presented are: Column (1) pooled OLS (no fixed or time-period effects), Column (2), fixed effects only (individual fixed effects), Column (3) time-period effects only, and Column (4) both individual fixed effects and time-period effects, respectively. All variables are measured as natural logs and *p < 0.10, **p < 0.05, ***p < 0.01. Standard errors are in parenthesis.
per capita GSP. The negative direct effect and positive indirect effect of urbanization rate suggest that growth in urbanization will reduce own state CO 2 emissions while at the same time increasing emissions of neighboring states. These results imply that a 1% increase in urbanization is linked with a 0.3% decrease in own state per capita CO 2 emissions and an 0.012% increase in per capita emissions in neighboring states. The intuition here is that the total effect of urbanization advances reductions in CO 2 emissions so growing cities across the United States has a total effect of decreasing CO 2 emissions. Energy and coal consumption have positive direct and indirect effects on CO 2 emissions, implying that an increase in either one leads to emission increases, both own state and neighboring states. In fact, coal and energy consumption as electrical energy in power plants generate CO 2 emissions. The statistically significant effect of per capita GSP is negative within own state emission, but the indirect effect is not statistically significant. This result suggests that if the own state per capita GSP increases, it will lead to own state CO 2 emissions reductions, but not reductions in neighboring states. The estimated coefficients on the GSP are all highly significant and consistent with the inverted U-shaped relationship illustrated by the environmental Kuznets curve (EKC) hypothesis. As shown in Table 3, the negative direct and indirect effects of energy prices suggest that if statelevel energy prices increase, this effect will not only decrease in-state CO 2 emissions but also CO 2 emissions from neighboring states by an even larger amount. The estimate shows that energy price linkage to CO 2 emissions is inelastic in the short run. This indicates that the result is consistent with expectations as in the short run consumers do not have the possibility to change their quantity of energy and can only alter the behavior of consumers (Bhattacharyya, 2011). In other words, if one state increases its energy prices then it has a negative, short-run impact on its own CO 2 emissions. In accordance with the law of demand it is expected that all coefficients on the prices are negative (Table 3).
Finally, with negative direct and indirect effects, if an RPS is implemented in one state, it will lead to reductions in own state CO 2 emissions as well as an equally large reduction of emissions in neighboring states. This effect is indicative of interstate electricity transport and renewable energy certificates for electricity generation that can be traded across state lines.  Note: All variables are estimated as natural logs and *p < 0.10, **p < 0.05, ***p < 0.01. Standard errors are in parenthesis.
The estimated results for spatial models are reported in Table 5. The results show that at a 1% statistical significance level, CO 2 emissions are a decreasing function of explanatory variables for urbanization rate, energy prices, RPS, and per capita GSP. Table 5 also reveals that at a 1% statistical significance level, CO 2 emissions are increasing functions of energy use and coal consumption. The ρ sdm coefficient in the SDM model is statistically significant with contiguity-based matrix, justifying the use of spatial econometric models by indicating the presence of a spatial autoregressive impact. Therefore, dynamic spillover long-run linkages between fossil fuel energy use and urbanization are of importance to energy policy makers in the United States. Policy toward lessening coal consumption might have a significant impact on the reduction of CO2 emissions in the states. The implications of this study for policy are as follows. First, U.S. utilities are using more clean energy and substituting natural gas for coal, as this substitution is affected by both efficiency and the price of natural gas. Since the coal industry has been in crisis, it is proper policy to promote exporting coal to other countries might be a solution to overcome the industry's crisis. Second, since there are not statistically significant, long-run impacts from any variables on coal consumption in the post shale gas time period, the expansion of shale gas production has substantially changed the factors that determine coal consumption. On the one hand, this substitution has resulted in important shifts in the U.S. energy market and reducing CO 2 emissions. On the other hand, this substitution implies that another proper policy could be to encourage using coal for other uses (i.e. manufacturing or steel industry).
These empirical results not only contribute to advancing the current literature but also deserve certain attention from energy policy makers in the U.S. market. Since more than 90% coal have been used in power plants in the United States, as capacities of power plants are decreasing and new capacities are being established to replace them, power firms are promoting to use of more efficient and cost-effective energy sources (i.e. natural gas) to build new power plants over coal-fired Table 5. Estimation results of spatial panel data models  plants. Given the coal issue of the energy market in the United States, more study needs to be undertaken to offer a comprehensive resource for policymakers pursuing actual solutions to alleged problems surrounding this industry. Thus, a U.S. state is more likely to adopt a law and or policy if its neighboring states have already done so (Mooney, 2001). In fact, geographical location has been identified as a critical factor of cross-region economic growth due to indicators like the diffusion of technology (Keller, 2004). We could argue that CO 2 emissions may decrease with technological development, then the diffusion of technology would likely enhance conditions of neighboring environment. It is important to develop federal and state policy to manage CO 2 emissions in the cities. Moreover, other policies to decrease CO 2 emissions would promote developing urban areas and implement a renewable energy policy.

Conclusions
The empirical results suggest that urbanization rate, energy prices, RPS, and per-capita GSP are inversely related to CO 2 emissions in their own states. These results imply that increased urban development, economic growth, and the implementation of RPS all reduce CO 2 emissions at the state level. The SDM also shows that while more urbanization reduces the own state CO 2 emissions, it increases CO 2 emissions of neighboring states by about 1/3 amount of the own state reduction. One possible reason for this positive indirect effect of urbanization stems from interstate transportation of goods and services. As urban areas grow, there is increased importation of goods and products from other states, which create energy-related CO 2 emissions from the production and transportation of these goods. The direct and indirect effect results of the energy use and coal consumption variables imply that higher coal consumption and energy use lead to higher CO 2 emissions not only within a state but also on a regional basis. Hence, reducing fossil fuel use in an economy is an effective choice for decreasing CO 2 emissions. Coal consumption has the highest impact on CO 2 emissions of any variable, which suggests that an appropriate policy for reducing CO 2 emissions is to use more efficient energy sources and renewable energy.
Our results contribute to the literature by quantifying opposite direct and indirect effects of urbanization on state-level CO 2 emissions along with the negative effects of income and RPS on state-level emissions. The income effects provide additional support for the EKC, such that income growth reduces per capita CO 2 emissions (Aldy, 2005;Tawfeeq et al., 2019). For RPS, the indirect effects of CO 2 emission reductions are as large as the own state direct effects, indicating that cross-border impacts are important. This result is important due to the prominence of state and local government actions in climate change mitigation now that the U.S. government has withdrawn from the Paris Climate Accords.
However, this study suffers from several limitations. First, the problem of measurement error of CO 2 emissions that is consistent with the rest of literature since CO 2 emission is based on the measurement of emissions from the burning of fossil fuels and not real ambient CO 2 emissions. Second, there are other factors that contribute to CO 2 emissions that include land use conversions from natural areas or farm to urban which are necessary to create urbanization. These factors are not accounted for in this analysis. Third, the spatial panel data process could suffer from problems of endogeneities within the independent variables and issue of theoretical framework.
The impacts of factors and independent variables from Table 4 are essentially consistent with the previous literature and theoretical expectations offered in literature review and the theoretical model (Newman and Kenworthy, 1999;Cole and Neumayer, 2004;Andrews, 2008;Gonzalez, 2009;Jaforullah and King, 2017). More specifically, our results are in line with the findings of Liddle and Lung (2010), Poumanyvong and Kaneko (2010), Chuai et al. (2012), and Burnett et al. (2013) since they had evidence favorable to the effects of urbanization, GDP, and the energy factors (use, prices, and policies) on CO 2 emissions. However, our study contrasts with some of these studies since our findings show that urbanization has negative direct effect and positive indirect effects on CO 2 emissions at the state level in the United States. Finally, this study only provides a primary interpretation of the effect of urbanization on CO 2 emissions based upon state-level data. Hence, it excludes more explicit factors examining city size, green technologies implemented, substitutions of renewable energy for fossil fuels, and structural changes among various economic sectors in the U.S. urban areas. Future work could be to more directly investigate the roles of economic activities, sectors of economy in urban areas, renewable energy sources, and other excluded factors in explaining the effect of urbanization on environment and economic development. We look forward to future study along with these lines.

Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.