Have State Renewable Portfolio Standards Really Worked?

Transcription

1 Have State Renewable Portfolio Standards Really Worked? Synthesizing past policy assessments to build an integrated econometric analysis of RPS effectiveness in the U.S. Gireesh Shrimali a, Steffen Jenner b, Felix Groba c, Gabe Chan d, Joe Indvik e a Monterey Institute of International Studies, 460 Pierce St Monterey, CA 93940, gshrimali@miis.edu, Tel.: b Tübingen Univ., Melanchthonstr. 36, Tübingen, Germany, steffen.jenner@uni-tuebingen.de, Tel.: c German Institute of Economic Research, Mohrenstrasse 58, Berlin, Germany, fgroba@diw.de, Tel.: d Gabe Chan, Harvard Univ., 79 Kennedy St, Cambridge, MA 02138, gabe_chan@hksphd.harvard.edu, Tel.: e ICF International, 1725 I Street N.W., Washington, D.C , jindvik@icfi.com, Tel.: Abstract: Renewable portfolio standards (RPS) are the most popular U.S.-state-level policy tools for promoting deployment of electricity from renewable energy (RES-E). Several econometric studies have investigated whether RPS policies have been effective in increasing RES-E deployment. Results are contradictory and require reconciliation. Consequently, there is a strong need to investigate the real impact of RPS policies. We use a panel 50 states over of renewable deployment, policy design elements and market dynamics to represent the heterogeneity in state RPS design and market characteristics. We run a series of time series cross sectional regressions with fixed effects. We trace discrepancies in past findings to differences in data classification and model specification. Outlining the importance of respecting classification changes in publicly available data, we show that the main feature of RPS RPS stringency has no effect or a negative impact on RES-E deployment. We further show that mandatory green power options have a positive effect on RES-E deployment while net metering schemes appear to have a negative impact. We also investigate the impact of individual RPS features. We show that renewable energy certificate (REC) unbundling as well as the presence of RPS in neighboring states makes RPS more effective. We further investigate the impact of RPS on individual renewable technologies and show that the total RES-E deployment is mainly driven by biomass: yet, we do not observe a statistically significant link with solar, geothermal, and wind deployment. 1 Introduction Climate change is accepted as a real problem. There is a wide-ranging consensus that global warming mainly caused by the accumulation of greenhouse gases (GHGs) in the atmosphere is the biggest problem faced by humans (Pew Center, 2008), and that human activity is primarily responsible (IPCC, 2007). Global warming may cause serious, perhaps irreversible damage of the earth's climate, requiring expedited focused effort in controlling GHG concentrations (Stern, 2007). Renewable energy plays a key role in mitigating climate change. In the U.S., power plants are responsible for approximately 40% of the nation s carbon dioxide the primary GHG emissions (Energy Information Administration, 2009b). This has resulted in environmental groups to target the sector (Lyon and Yin, 2010). The growing public concern for the environment and progressively stringent regulation of emissions by the governments has driven the electric power industry to increase the amount of renewable energy in the electricity generation portfolio. Renewable energy can also contribute to energy security and job creation. Policymakers in the U.S. support electricity generation from renewable energy sources (RES-E) to promote green growth, to foster 1

2 energy security and to combat local and global environmental and health problems that arise due to the burning of fossil energy sources (Schmalensee, 2011). Most non-hydroelectric renewable energy is not competitive yet. Generation costs have been declining for renewable energy technologies such as solar photovoltaic (Nemet, 2006). However, power from most conventional energy sources remains cheaper in most places (NREL, 2010). Thus, if policymakers want to increase the fraction of renewable energy capacities they can intervene into the market by means of policies and regulations. Most of the action is at state level. In the U.S., there are many federal level policies in support of renewable energy, such as the production and investment tax credit. However, there is a lack of coordinated and comprehensive action by the federal government, as evident by the absence of a nationwide cap-and-trade program or a clean energy standard. On the other hand, the state and local governments have taken initiative (Engel and Orbach, 2008). Actions have been taken to increase renewable energy capacity and generation, with most of the 50 states enacting policies to encourage the use of renewable energy in their state (DSIRE, 2011). The state-level initiatives are wide-ranging: adopting various renewable energy incentives, enforcing integrated resource planning programs, and initiating cap-and-trade programs (Wasserman, 2010). RPS policies are considered the key policy tool to support the adoption of RES-E to date. Renewable portfolio standards (RPS) are a type of quantity regulation that mandates energy suppliers to reserve a certain fraction of their total electricity sales for electricity generated from renewable energy sources. Though states use multiple policies, such as public benefit funds, required green power option, and net metering, renewable portfolio standards (RPS) are considered as the key policy tool. According to DSIRE (2012): Twenty-one states and the District of Columbia have a mandatory RPS targets effectively in place. Ten other states have adopted RPS schemes but the effective start of the yearly targets lies in future. There are sixteen states with public benefit funds, used to support renewable energy programs. Small additional fees on power consumption usually finance the funds. Seven states have required green power options command policies that require energy suppliers to provide consumers the option to buy electricity from renewable energy sources in place. In forty-two states, net metering mechanisms a policy that allows households that generate their own electricity to only pay for the difference between own generation and own consumption even if generation and consumption do not occur at the same moment support mostly small and distributed RES-E capacities. Ex-post RPS effectiveness analysis is possible. There is considerable experience with RPS policy implementation, including variation in application of different RPS features (DSIRE, 2012). Yet, decentralized policy making, the individual state policies as well as renewable power penetration show a great deal of variance. As a result, renewable deployment varies considerably. While a number of factors might explain both the growth of renewable energy and the disparity in renewable capacity among states, policies adopted by state governments, including changes in the regulatory environment for electricity, are expected to play an important role. Therefore, ex-post effectiveness analysis is possible not only of RPS but also of various RPS features. 2

3 Existing results are contradictory. Much work has been done to assess effectiveness of RPS policies in increasing renewable deployment, including Menz and Vachon (2006), Carley (2009), Yin and Powers (2010), and Shrimali and Kneifel (2011), to name a few. However, results are contradictory, and vary from showing the impact of RPS policies on renewable deployment as positively significant (Menz and Vachon, 2006; Yin and Powers, 2010) to insignificant (Carley, 2009) to even negatively significant (Shrimali and Kneifel, 2011). Thus, there is a strong need to explain why these contradictory results exist, and investigate as to what is the real impact of RPS policies on renewable deployment. We reconcile and explain existing results. The difference in results can be attributed to many factors, including differences in datasets, analytical techniques, and even the period over which the impact of policies is examined. We use latest datasets and show that differences in results are due to usage of different datasets and not due to different methodologies. In particular, we show that one of the key results in recent times that RPS policies drive renewable deployment (Yin and Powers, 2010) is due to the use of an incorrect dataset and not due to the construction of a modified RPS stringency indicator. In this process, we also attempt to reconstruct the results in other papers (Carley, 2009; Shrimali and Kneifel, 2011), and achieve different degrees of success, with the variability potentially due to us using a different set of control variables. We identify the correct state-level dataset. The two widely used datasets that are publicly available are the state-level dataset and the generator-level dataset. We show that the state-level dataset is the correct dataset to use and not the generator-level dataset, because of discrepancies introduced in the latter due to the change in methodology in Using this state-level dataset, we show that the main feature of RPS RPS stringency negatively effects renewable capacity or if Maine is excluded from the sample has no statistically significant effect. We add richness to the results based on the state-level dataset by considering: Other policies such as public benefit funds, mandatory green power options, net metering, etc. We show that mandatory green power options have a positive effect on RES-E capacity while net metering schemes appear to have a negative impact. RPS features such as alternative compliance payments, renewable energy certificate (REC) unbundling and trading, neighborhood effects, trading restrictions, etc. We show that unbundling supports RPS effectiveness. The presence of RPS schemes in neighboring states also has a positive effect. We are unable to detect any statistically significant impact of the other features. Effectiveness of RPS in increasing deployment of different renewable technologies such as wind, biomass, solar, and geothermal. We show that the total RES-E results are mainly determined by biomass. The remaining sections go into the detail of the paper. Section 2 provides a literature review. Section 3 presents our regression model. Section 4 discusses variables as well as different datasets under consideration. Section 5 presents the results and discusses them. Section 6 concludes. 2 Literature Review Samples. There are many quantitative studies that investigate the impact of RPS policies on renewable energy deployment: Alagappan et al. (2011), Carley (2009), Delmas et al. (2011), Dong (2012), Menz and 3

4 Vachon (2006), Shrimali and Kneifel (2011), Yin and Powers (2010). Further, there is an emerging series of studies focusing on other continents. Marques et al. (2010; 2011) and Groba et al. (2011) use the EU sample. Dong (2012) includes both U.S. states and EU member countries in a sample of 53 states but it remains unclear which ones he chose. Salim and Rafiq (2012) investigate major emerging economies. Models. With the exception of the descriptive analysis of Alagappan et al. (2011), all studies use some type of time series cross-section regression model. Menz and Vachon (2006) run OLS regressions without fixed effects for a sample of 37 U.S. states over 5 years. Carley (2009), Dong (2012), Groba et al. (2011), Marques et al. (2010), and Yin and Powers (2010) control for time trends and state level effects. Shrimali and Kneifel (2011) also control for state-specific time trends. Delmas et al. (2011) apply a two-stage regression logit and tobit to also cover public choice variables such as the influence of private interest groups on policymaking. Marques et al. (2011) assess the impact of socio-economic factors on RES-E development by means of a quantile regression. Salim and Rafiq (2012) run modified and dynamic OLS regressions. Policy Covariates. The level of sophistication to capture the impact of policies also varies broadly. Menz and Vachon (2006), Marques et al. (2010), Alagappan et al. (2011), Shrimali and Kneifel (2011), and Dong (2012) use binary variables to represent the existence of support policies. Carley (2009) applies nominal variables to capture more of the heterogeneity that comes from the policy design. Delmas et al. (2011) use the predicted probabilities of RPS adoption from their first stage regression as a covariate in the second stage. Yin and Powers (2010) introduced the incremental share indicator (ISI) of RPS policies. Groba et al. (2011) apply the ISI to the EU sample. Results. Menz and Vachon (2006) reveal a significant positive effect of RPS policies on the development of wind capacity. Since their model does not control for state characteristics and time trends, one can argue that the findings are not accurate enough to actually make a statement about real impact of RPS policies. Menz and Vachon (2006) do not explain why a random effects model is appropriate, e.g. by a Hausman Test. In contrast, almost all other studies including the one at hand showed that state and year effects can be a major bias. Carley (2009) does not find a significant link between the RPS binary and the percentage of RES-E generation. Once she uses absolute generation the RPS trend variable becomes positive and slightly significant. After removing the state effects from the FEVD specification, the link is highly significant which means the state characteristics can be an important driver of absolute RES-E deployment. Delmas et al. (2011) reveal a negative and significant link between the existence of a RPS and absolute RES-E capacity. Dong (2012) also finds a negative and significant connection between the RPS binary and cumulative wind capacity. However, the significance disappears if standard errors are clustered and if the model includes year trends. Groba et al. (2011) do not reveal a significant connection between RPS policies in six EU member countries and wind and solar PV capacity. Shrimali and Kneifel (2011) reveal a negative link between the RPS binary and wind capacity. This effect cannot be approved for total RES- E, biomass, and solar capacity. Geothermal capacity appears to be positively affected by RPS policies. Yin and Powers (2010) do show that RPS binary and RPS trend variables do not connect significantly to the percentage of RES-E capacity. However, they reveal a negative and significant link between the annual RPS fraction and the RES-E ratio. They conclude that a more nuanced covariate, the ISI, is needed 4

5 to investigate the stringency of RPS policies. The ISI connects positively to the RES-E capacity ratio in every model specification. We will scrutinize this finding in Sections 4 and 5. 3 Empirical Model In line with earlier studies on the subject, the primary objective of this paper is to determine the effect policy has on the development on renewable energy capacity (or generation) development. We estimate these effects using a panel dataset representing the variables discussed in Section 2, mirroring and considerably enriching and improving data used in other studies on the subject (Yin and Powers, 2010; Carley, 2009; Shrimali and Kneifel, 2011). Following the theoretical approach outlined by Carley (2009) which is based on foundations of public policy finding of associated environmental policy literature (Bennear, 2007; Mazur and Welch, 1999; Ringquist, 1994; Ringquist and Clark, 2002; Sapat, 2004), we estimate the capacity (generation) development for the renewable energy technology i in state s at year t using the following model: X sit = β 0 + β y Y st + β p P sit + α i + γ t + ε it (1) Where X represents the dependent variable in different specifications as will be outlined below, β 0 is a constant, Y represents a matrix of social and economic variables used in other studies that are expected to have an impact on renewable energy development, P is a matrix of policy control variables in order to control for the policy effects and respective policy design elements aimed at encouraging renewable energy dissemination. The proper estimation of this model requires addressing several issues regarding correct econometric specification as standard OLS estimation yields inconsistent or insufficient estimates. First, since we exploit a panel of individual states, unobserved state and year heterogeneity is likely to be present. Not controlling for heterogeneity yields inconsistent estimates as error terms are correlated. Therefore, we use a fixed effects model controlling for state level fixed effects α and time effects γ. While state level fixed effects control for existing differences among states such as renewable energy potential and existing renewable energy capacity, time effects control for exogenous factors such as technological progress and macroeconomic trends affecting all states. A Hausman test on each of the specifications rejects random effects outlining that fixed effects estimation is more appropriate in this context. Controlling for these effects and using clustered robust standard errors on the remaining error ε accounts for potential heteroskedasticity. 4 Variables and Data 4.1 Renewable Energy Supply Quantification Previous econometric studies on the effectiveness of RES-E support policies differ with regard to the dependent variable selection. RES-E supply variables can be characterized along three dimensions. First, renewable energy supply can be measured in terms of capacity (watts) or actual generation (watt-hours). 5

6 Second, state level data can be manually aggregated on the base of EIA annual generator surveys. Or it can directly be downloaded in aggregate for each state-year. Third, renewable energy can be quantified in absolute values or as a percentage of total electricity. The characterization of RES-E dependent variables in previous studies is shown in Table 1. Table 1 Dependent variable selection in previous studies Relative (or %) Absolute Generation Capacity Generator Yin and Powers (2010) State Carley (2009), Shrimali and Kneifel (2011) Marques et al. (2011) Generator Delmas et al. (2011) Carley (2009) Dong (2012) State Groba et al. (2011) Menz and Vachon (2006) Salim and Rafiq (2012) The italic studies investigate EU member countries, while the other studies work with the U.S. sample. Delmas et al. (2011) compiled data for 650 utilities while the other studies use the state as their core unit of analysis. Salim and Rafiq (2012) analyzed RES-E consumption in six major emerging countries Sources The U.S. Energy Information Agency (EIA) provides data for generation and capacity at both the generator level and the state level in the U.S. The EIA forms and documents that collect this data and their web links are shown in Table 2. Table 2 EIA data sources Generator- Level State-Level Generation EIA Form EIA-906, EIA-920, and EIA- 923 Data eia906_920.html EIA Electric Power Annual Capacity EIA Form EIA-860 Annual Electric Generator Reports ml EIA Electric Power Annual State level data Most studies use state level data of the total electric power industry s RES-E generation and capacity that is provided by the EIA Electric Power Annual Aggregated generator level data Yin and Powers (2010) use generator level data. Aggregating data on the basis of EIA generator level data faces two major challenges. First, in 2001, the classification of sources in the EIA generator level data changed for both generation (EIA-906) and capacity (EIA-860) forms. Second, while the forms since 2001 contain both non-utility generators and utility generators, pre-2001 forms contain utility generators only. We worked with EIA to synchronize the databases as much as possible. We find that both reported generation and capacity (EIA-906 and EIA-860) increase abruptly after 2000 (Figure 1). This sharp increase is mainly caused by the exclusion of non-utilities prior to When we add the non-utilities from 1990 to 2000, available in EIA-867 and EIA-906nonu forms, the sharp increase vanishes. The 6

7 remaining inaccuracy between 2000 and 2001 is most likely caused by the changes in classification that we outlined above. Figure 1 Generator level generation and capacity development Capacity (GW) Generation (TWh) Non-hydro RES-E Capacity (EIA-860) Non-hydro RES-E Capacity (EIA EIA-867) Non-hydro RES-E Generation (EIA-906) Non-hydro RES-E Generation (EIA EIA EIA-906nonu) Consequences on previous findings Yin and Powers (2010) state that they use capacity as the dependent variable but also that they compile capacity data from the EIA-906 forms. This is misleading because EIA-906 forms provide generation data not capacity data, which is provided by the EIA-860 forms. Furthermore, they do not state that they added the non-utility generators prior to We replicate their dataset and find that their estimations might have been affected by the exclusion of the non-utility generators prior to We will discuss this issue in detail in Section Policy and Policy Feature Variables In previous studies, RPS policies have mostly been represented by a dummy variable (RPS Binary) that equals 1 if the policy is effectively in place and 0 for its absence. Yin and Powers (2010) quantified the impact of RPS policies as a count variable (RPS Trend) for the years since policy implementation; and the yearly RPS requirement (RPS Yearly Fraction) as a percentage. Yin and Powers (2010) also introduced a more nuanced instrument, the incremental share indicator (ISI). The ISI represents the mandated increase in renewable generation in terms of the percentage of all generation (Yin and Powers, 2010: 1142). Thus, the ISI is a metric for policy stringency. The ISI is constructed as: ISI it η = RES it *κ * q RES total it it RES qit Q RES it (2) with RES κ it RES η it representing the RPS yearly fraction as a percentage of RES-E to total electricity generation; representing the percentage of RES-E generation capacity that is legally eligible to meet RES ; η it total q it 7

8 representing the annual total electricity generation; and indicating the existing absolute RES-E generation. We built the ISI on the base of the EIA s (2011) generator level data that was presumably used by Yin and Powers (2010), ISI (YP), and on the base of state level data that was used by most other studies, ISI (SL). Besides the five variables measuring the stringency of RPS policies (RPS Binary, RPS Trend, RPS Yearly Fraction, ISI (YP), and ISI (SL)), we also collect binary codes for related RES-E policies: the existence of public benefit funds (PBF), net metering policies (NM), and mandatory green power options (MGPO). We also test the impact of policy design on the effectiveness of RPS policies. In a series of regressions, we account for the RPS features of (i) alternative compliance payments (ACP) in $/MWh and (ii) the maximum effective retail rate increase (MERRI) as a percentage of the average retail rate in the worstcase scenario (Wiser and Barbose 2008). We include binaries for (iii) unbundled REC trading, (iv) allowance of REC trading, and (v) contracting mechanism. We build indices to capture electricity delivery requirements. The (vi) delivery to regions index (DTX) equals 0.5 for broader delivery to the state and 1 for direct transmission. The (vii) delivery from region index (DFX) equals 0.5 for the eligibility of generators anywhere outside the region and 1 in limited areas. In some specifications, we use Yin and Power s (2010) control to capture the effect of the regional REC trading market size. The (viii) RPSMS is constructed as: RES Q it RPSMS!" =!!"#!"#!"!#$!"#!!"!!"!!"!!!"!"#$!!" (3)!"#$!!" with A representing the number of neighboring states to state I, TRADE!" representing a binary code that equals 1/0 if out of state trading is/is not allowed, and SALES!" representing the total electricity sales. In some specifications, we also include (ix) the percentage of neighboring states that have a RPS in place. Chandler (2009) introduced such a control to measure cross-state effects. 4.3 Controls In order to produce comparable results, we use the suite of controls from Yin and Powers (2010). First, State Income captures the median income of a 4-person household in 1,000$. We expect RES-E to increase more rapidly in richer states since they may be in a better position to absorb the additional costs due to shift from conventional to renewable energy production. Second, Electricity Price represents the mean state electricity price in a state in cents/kwh. High electricity prices may increase renewable deployment by lowering market barriers. On the other hand, they may foster reluctance to add further burden to the electricity bills due to RES-E capacity development. We lag this variable once as in Yin and Powers (2010) in order to avoid reverse causality. Third, the electricity Import Ratio controls for the imbalance between domestic sales and out of state power generation. Following Yin and Powers (2010), we quantify the import ratio by the percentage of net imports of electricity and total sales of electricity of the previous year. A high import ratio presumably advances domestic RES-E capacity building. 8

9 Finally, the LCV Score is an index created by the League of Conservation Voters (LCV) that tracks the voting behavior of state-level representatives and senators on environmental issues. We expect high LCV Scores to positively correlate with RES-E development. Data for the variables has been compiled from various EIA sources (see Section 4.1), DSIRE (2012), the U.S. Census Bureau (2011), Wiser and Barbose (2008), Wiser et al. (2010), and the League of Conservation Voters (2011). 5 Results Considering the particular attention that Yin and Powers (2010) have received among studies that evaluate the effectiveness of RPS policies in promoting deployment, we first attempt to re-create their results. Table 3 and Table 4 present the regression results on the replicated dataset used by Yin and Powers (2010). Table 5 shows the ISI coefficient under additional specifications of the dependent variable. Table 6 shows the results from our own model that runs on state level capacity data. Table 7 replaces the ISI with alternative covariates that capture additional channels through which RPS policy features may affect RES-E development. We also investigate the effects of additional RPS policy features and models with technology-specific dependent variables for biomass, geothermal, solar thermal and photovoltaic, and wind capacity. 5.1 Recreating Results from Yin and Powers Table 3 Recreation of Yin and Powers (2010) with capacity ratio as dependent variable (1) (2) (3) (4) (5) (6) ISI (YP) 0.325*** 0.311*** 0.134*** 0.160*** 0.160* 0.190*** (0.023) (0.026) (0.024) (0.024) (0.092) (0.042) Public Benefit Fund Binary (0.272) (0.395) (0.260) Net Metering Binary *** (0.214) (0.542) (0.324) Mandatory Green Power Binary 3.588*** 3.588** (0.396) (1.557) (1.297) State Income 0.145*** 0.145*** 0.078*** (0.025) (0.025) (0.022) Electricity Price, lagged * (0.070) (0.309) (0.284) Import Ratio 0.010*** (0.004) (0.014) (0.008) LCV Score 0.018*** 0.018** 0.008* (0.006) (0.008) (0.005) State Effects yes yes yes yes yes Year Effects yes yes yes yes State Clusters (robust) yes yes Time Frame N 1,000 1,000 1,000 1,000 1, R- Squared Standard errors in parentheses. The dependent variable is the percentage of RES-E capacity to total annual electricity capacity on the base of generator level data without correcting for the 2001 inconsistency. * Significant at 10%, ** Significant at 5%, *** Significant at 1%. The ISI (YP) has a positive and significant effect on RES-E capacity across all specifications (Table 3), indicating that the result that ISI positively impacts the share of renewable capacity is fairly robust, 9

10 though the specification with state and time fixed effects and with controls estimate a lower impact of the ISI variable on RES-E capacity (Specifications 2, 3, and 4). However, as the most-comprehensive Specification (5) shows, the estimated coefficient is roughly one third in size of the estimate as presented by Yin and Powers (2010). Under Specification (6), we re-estimate the model as in Specification (5), but limit the data to the time period used in Yin and Powers (2010). Under this specification, the size of the impact marginally increases relative to Specification (5), but remains less than half of the estimate as presented in Yin and Powers (2010) under the same specification. Table 4 Recreation of Yin and Powers (2010) with generation ratio as dependent variable (1) (2) (3) (4) (5) (6) ISI (YP) 0.534*** 0.484*** 0.351*** 0.336*** 0.336** 0.394*** (0.019) (0.022) (0.020) (0.021) (0.137) (0.062) Public Benefit Fund Binary (0.237) (0.394) (0.354) Net Metering Binary (0.186) (0.369) (0.275) Mandatory Green Power Binary 2.112*** 2.112** (0.344) (0.960) (0.904) State Income 0.057*** 0.057* 0.038* (0.022) (0.033) (0.022) Electricity Price, lagged ** (0.061) (0.260) (0.246) Import Ratio *** * (0.003) (0.012) (0.013) LCV Score (0.005) (0.007) (0.006) State Effects yes yes yes yes yes Year Effects yes yes yes yes State Clusters (robust) yes yes Time Frame N 1,000 1,000 1,000 1,000 1, R- Squared Standard errors in parentheses. The dependent variable is the percentage of RES-E generation to total annual electricity generation on the base of generator level data without correcting for the 2001 inconsistency. * Significant at 10%, ** Significant at 5%, *** Significant at 1%. Replacing capacity with generation data (Table 4) yields results that match more closely with the results presented in Yin and Powers (2010), despite the fact that Yin and Powers claim to use capacity data. We estimate for the ISI (YP) coefficient whereas Yin and Powers (2010) estimate the same coefficient at Given that our results for generation data match more closely to the results presented in Yin and Powers (2010), they support the hypothesis that Yin and Powers (2010) used generation data. This hypothesis is corroborated by Yin and Powers (2010) reference to the EIA-906 forms, which contain generation not capacity data. We now show that trying to replicate the Yin and Powers (2010) results may not be a fruitful exercise because the dataset used by the authors does not seem to properly account for the change in EIA classification as described in Recall that the dataset used in Table 3 and Table 4 has inconsistencies introduced during year , primarily by incorporation of non-utility data that was ignored prior to this period. Table 5 shows the coefficient on the ISI covariate for different specifications. The regression model with an identical set of controls as in Specification 6 in Table 3 and Table 4 is run on three different time 10

11 frames: the full period for which data is available ( ), the period after the change in classification in the generator level data ( ), and the period used by Yin and Powers ( ). Thus, only when we correct for the bias, we are able to replicate the sign of the ISI coefficient that Yin and Powers estimate. All other dependent variable specifications yield ISI coefficient estimates that are negative, although some specifications yield statistically significant coefficients and some do not. Table 5 The coefficient on ISI under different specifications of the dependent variable Generation GL- T* ** *** GL- T * SL- T 0.116*** *** SL- U Capacity GL- T* * *** GL- T 0.106*** *** SL- T 0.105*** *** SL- U 0.153** 0.036* 0.134** GL: generator level; SL: state level; T*: sample with bias presumably used by Yin and Powers (2010); T: total electric power industry; U: electric utilities only. Sources: Generation data: (GL-T*) from EIA-906; (GL-T) from EIA-906, EIA-867, and EIA-906nonu; (SL-T) from EIA Electric Power Annual; (SL-U) from EIA Electric Power Annual. Capacity data: (GL-T*) from EIA-860A; (GL-T) from EIA-860A and EIA-867; (SL-T) from EIA Electric Power Annual; (SL-U) from EIA Electric Power Annual. The remaining control variables are the same as in Table 3 and Table 4, Specification (6) (Public Benefit Fund, Mandatory Green Power, State Income, Electricity Price, Import Ration, LCV Score). The sign of their coefficients remain the same throughout and not reported here. Results can be received upon request. The key insight from this table is that if and only if the erroneous change in classification that exists in the GL data remains in the dependent variable (i.e. the GL-T* variable in the and time frame) does the coefficient on the ISI variable turn out to be positive. When this major data coding error is corrected, the generator-level dataset gives results very similar to the ones derived from the state-level datasets. Therefore, going forward, we follow Carley (2009), Dong (2012), Groba et al. (2011), Marques et al. (2011), Menz and Vachon (2006), Salim and Rafiq (2012), and Shrimali and Kneifel (2011) in using state-level datasets only. 5.2 Our Model Table 6 introduces the results of our own model that runs on state-level data and measures the impact of the ISI variable in a model with and without state and year fixed effects and various controls. Specification (7) drops Maine from the sample because it appeared to be an outlier in the scatter plot. Table 6 Full model results I with capacity ratio as dependent variable ISI (SL) Public Benefit Fund Binary Net Metering Binary Mandatory Green Power Binary (1) (2) (3) (4) (5) (6) (7) 0.411*** * *** 0.297*** * *** (0.038) (0.025) (0.023) (0.070) (0.044) (0.036) (0.069) (1.070) (0.446) (0.486) (0.450) ** ** (0.689) (0.365) (0.491) (0.500) 4.767*** 4.760*** 3.882*** 4.071*** (1.698) (1.613) (1.434) (1.530) 11

12 State Income * 0.118** 0.121** (0.067) (0.038) (0.049) (0.050) Electricity Price, lagged ** * * (0.167) (0.135) (0.142) (0.149) Import Ratio 0.015** 0.036*** 0.031*** 0.032*** (0.007) (0.012) (0.010) (0.011) LCV Score *** 0.019** 0.018** (0.017) (0.009) (0.009) (0.009) State Effects Yes yes Yes yes yes Year Effects yes yes yes State Clusters (robust) Yes Yes yes yes Time Frame N 1,000 1,000 1,000 1,000 1,000 1, R- Square Standard errors in parentheses. The dependent variable is the percentage of RES-E capacity to total annual electricity capacity on the base of state level data. * Significant at 10%, ** Significant at 5%, *** Significant at 1%. We estimate a negative and strongly significant coefficient for the ISI (SL) covariate in the full model Specification (6). Given that RPS policies are one of the most popular policy tools to support RES-E development at the state level, this finding is surprising. However, Carley (2009), Delmas et al. (2011), Dong (2012), and Shrimali and Kneifel (2011) also find negative effects of RPS policies on RES-E development. Though we further discuss the policy implications of this finding in the conclusion, it is clear that, across the sample, RPS policies do not show the expected positive effect on RES-E capacity. It may well be that some RPS policy features, such as alternative compliance payment or regional trading make these policies in more effective in some states. Another important takeaway is the importance of fixed effects. When we do not account for the statespecific or year-specific effects (Specification (4)), the ISI coefficient is actually positive and significant. This is similar to the result in Menz and Vachon (2006), which do not account for fixed effects. However, as we show, neglecting the fixed effects produces inconsistent results (Section 3). In Specification (7) we exclude the state of Maine because it is an outlier in terms of total added electricity capacity. By the end of 1999, Maine added roughly 1,500 MW of natural gas capacity to its total capacity of roughly 3,000 MW by installing five gas-burning plants supplied by the newly built Maritimes and Northeast gas pipeline. As a result, the RES-E capacity ratio sharply decreased from 27% in 1999 to 16% in 2000 as the natural gas capacity ratio sharply increased. We exclude Maine in order to test the robustness of our full model with and without this outlier. Comparing the results from Specification (6) and (7) shows that the statistical significance of the ISI coefficient has been driven by the inclusion of Maine, the single state dropped in Specification (7). Most strikingly, the ISI coefficient is not significant anymore after dropping Maine. The other coefficients remain robust and we interpret them next. Public benefit funds are employed in 16 states: CA, CT, DE, HI, IL, MA, ME, MI, MN, MT, NJ, OR, PA, RI, VT, and WI. Conditional on the four model specifications that include a dummy variable for the presence of a public benefit fund (4 7), we do not find a significant impact of public benefit funds. This is in line with the result found in some of the previous studies (Menz and Vachon 2006, Yin and Powers 2010). 12

13 Mandatory green power options are employed in 8 states: CO, IA, ME, MT, NM, OR, VA and WA. We estimate a robust positive and significant coefficient across all model specifications for this variable. This is in line with almost all of the previous studies i.e., Delmas et al. (2011), Menz and Vachon (2006), Shrimali and Kneifel (2011), and Yin and Powers (2010). In particular, the presence of this policy seems to have a very significant impact on the renewable share approximately 4%, as indicated by Specifications (4)-(6). Net metering is implemented in almost every U.S. state. Its dummy variable appears to have a negative impact and sizable impact the presence of this policy is correlated with a ~1% reduction in renewable share in Specifications (6) and (7). This finding is counter-intuitive in the first place. However, net metering schemes are mostly capped at levels of 10 to 250kW and/or at around 1% of peak, whereas the EIA state-level data only contains capacities of at least 1 MW. It could thus be that net metering schemes work effectively but the increase in absolute capacity that can be attributed to this policy is just too small, resulting renewable share going down as the total electricity capacity increases in the state. Among the controls, we observe a positive impact of state income on RES-E development. This supports that hypothesis that wealthier states are more amenable to incorporate renewables in their energy portfolio. Carley (2009) detects a similar link for RES-E in the US; Groba et al. (2011) find a positive impact of GDP per capita on wind generation in Europe. The coefficient for electricity price is negative and significant at the 10% level in Specifications (6) and (7). This is consistent with high electricity prices making policymakers and utilities reluctant in incorporating expensive renewables that would further increase costs. Carley (2009), Delmas et al. (2011), and Yin and Powers (2010) find similar results, but Shrimali and Kneifel (2011) do not. We also find a positive and significant link between the electricity import ratio and RES-E capacity ratio. A potential interpretation is that producers may believe that switching from imported electricity to domestically generated renewable electricity will be more lucrative than continuing to buy power from other states. Marques et al. (2011) and Yin and Powers (2010) also present positive and significant links for this covariate. The LCV Score has a small but robust positive and significant impact on RES-E capacity development. Our interpretation is that a public and political environment that is supportive of environmental and renewable energy policy may be more willing to contribute to the adoption of RES-E capacity. Carley (2009) also revealed a positive link between the LCV Score and RES-E generation. 5.3 Replication of Previous Studies So far, we have only looked at one RPS indicator ISI. Given that ISI measures the strength of the policy over time, we believe this to be the most relevant indicator for measuring RPS policy stringency. However, we also examine other RPS indicators used previously (Carley, 2009; Shrimali and Kneifel, 2011). These include: RPS Binary, a dummy variable indicating the presence of RPS RPS Trend, a variable representing the number of years RPS has been in place 13

14 RPS Yearly Fraction, a variable that represents the RPS yearly targets, in terms of the fraction of electricity generation/capacity that must be supplied by renewable sources as specifically defined by each state. Table 7 presents models with these alternative measures, with two specifications per RPS variables to include the time period in corresponding studies. Table 7 Full model results II with capacity ratio as dependent variable (1) (2) (3) (4) (5) (6) RPS Binary * (0.583) (0.948) RPS Trend ** *** (0.147) (0.209) RPS Yearly Fraction ** ** (0.040) (0.043) Public Benefit Fund Binary (0.469) (0.364) (0.465) (0.281) (0.484) (0.413) Net Metering Binary ** ** ** (0.481) (0.262) (0.479) (0.271) (0.491) (0.343) Mandatory Green Power Binary 3.867*** 1.850*** 3.893*** 1.787*** 3.898*** 2.929*** (1.426) (0.569) (1.427) (0.580) (1.434) (0.784) State Income 0.120** ** ** 0.060** (0.048) (0.033) (0.049) (0.037) (0.049) (0.028) Electricity Price, lagged * * * (0.142) (0.147) (0.123) (0.132) (0.145) (0.159) Import Ratio 0.033*** 0.020*** 0.031*** 0.016*** 0.032*** 0.017*** (0.009) (0.006) (0.009) (0.004) (0.010) (0.004) LCV Score 0.021** ** ** 0.012** (0.009) (0.007) (0.009) (0.008) (0.009) (0.006) State Effects yes Yes Yes yes Yes yes Year Effects yes Yes Yes yes Yes yes State Clusters (robust) yes Yes Yes yes Yes yes Time Frame N 1, , , R- Square Standard errors in parentheses. The dependent variable is the percentage of RES- E capacity to total annual electricity capacity on the base of state level data. * Significant at 10%, ** Significant at 5%, *** Significant at 1%. The interpretation of the coefficients in Specifications (1), (3) and (5) in Table 7 using the complete time period ( ) is as follows. If a state without an RPS enacts an RPS, the ratio of RES-E capacity to total electricity capacity decreases by 0.98% (Specification 1). The negative coefficient on the RPS Trend variable indicates that the ratio of RES-E to total electricity continues to decline by an additional 0.38% per year after an RPS is enacted (Specification 3). Finally, the coefficient on RPS Yearly Fraction provides evidence that RPS policies that require one additional percent of required renewable deployment decrease the ratio of RPS capacity to total electricity generation by 0.1% (Specification 5). Specifications (2), (4), and (6) can be used to re-examine many of the findings in the previous literature that use state-level data in particular, Carley (2009) and Shrimali and Kneifel (2011), by using corresponding time-periods. We caution the reader that, given that we kept the suite of independent variables from our main analysis (Section 5.1 and 5.2), the suite of independent variables other than the RPS variables and the dependent variable specification do not match exactly. Future research is needed to synergize the datasets and model specification more carefully with these studies. For now, the reader should focus on the signs and statistical significance of the main RPS coefficients instead of the 14

15 remaining control coefficient values. In particular, Specification (6) attempts to recreate the results in Shrimali and Kneifel (2011) and finds similar results on the impact of RPS yearly fraction a result that pretty much remains the same when the underlying period is changed from to On the other hand, Specifications (2) and (4) attempt to recreate the results in Carley (2009) and do not find exactly the same results on the impact of the RPS dummy as well as RPS trends. For example, Specification (2), similar to Carley (2009) finds that the impact of RPS binary is negative, though statistically insignificant. The results match well with Carley (2009) when the period is extended from to in Specification (1). However, Specification (3) and Specification (4) are not able to replicate the positive coefficient of the RPS Trend variable in Carley (2009), perhaps due to the fact that Carley (2009) used not only a different set of independent variables but also logged RES-E generation ratio as the dependent variable. In summary, we estimate that for all of the tested independent variables that previous studies have used as a proxy for RPS policy stringency a dummy, a year count, and the percentage of the RPS requirement there is a negative and significant (at the 10% level) coefficient on an RPS covariate over the timeframe. The p-values on these RPS variables increase after Maine is dropped from the sample: for the binary and fraction variables, the coefficient on the RPS variable is not significant at the 10% level without Maine in the sample; however, the trend variable remains negative and significant at the 5% level. 5.4 RPS Features In this section, we examine the impact of individual RPS features. Due to lack of space we do not present the table with statistical results and discuss the results directly instead. The alternative compliance payments (ACP), which essentially capture the amount of financial penalties levied on responsible parties in case on non-compliance, appears to be negatively correlated with renewable share. A potential hypothesis for this may be that the amount of penalties may be causing states to buy from outside the state, which may have a negative impact on the renewable share in the state; however, we have not been able to verify this hypothesis. The MERRI variable does not have a significant impact. After dropping Maine, its coefficient is negative and significant, thus emphasizing our finding on the ACP feature. Higher pressure on energy suppliers to comply with the RPS that is induced by means of a higher $/MWh penalty and/or as a higher percentage of the average retail rate in the worst-case scenario does not increase in-state RES-E capacity development. The presence of REC unbundling, which essentially captures whether the renewable energy attributes of electricity are unbundled, appears to be positively correlated with renewable share. In fact, the magnitude is large greater than 2% indicating that the flexibility provided by unbundling has a large positive impact on the renewable share and should be promoted as much as possible. It must be pointed out, however, that REC unbundling may be increasing renewable deployment overall, and not only local deployment. The presence of REC trading, which essentially captures whether the unbundled RECs are traded in a market, does not appear to have a significant impact. Further, it is highly correlated with other covariates. 15

16 The REC contracting variable, which essentially captures whether higher flexibility is allowed on how long REC credits can be contracted for, does not seem to have a significant impact on renewable share. The REC market size variable essentially captures the size of the market that the RECs generated in a state can be sold into. The impact of this variable is negative, similar to the ISI variable. Thus, given our original result negative impact of ISI this result is not surprising. Yin and Powers (2010) also used a similar variable, and also found a negative insignificant link. It seems if an RPS scheme does not stimulate much additional RES-E capacity in one state, it does not help to include the RPS schemes from other states. The Neighbors with RPS variable essentially captures the percentage of neighboring states that have a RPS policy in place. The impact of this variable is statistically significant and positive, indicating the importance of network effects. If a state s neighbors have implemented RPS policies, it creates a better environment for renewable development in the region, which should result in better RES-E development. Chandler (2009) elaborates on this diffusion effect. She argues that the adoption of policies in one state can affect the public and political debate in neighboring states. The delivery-to-region variable, which essentially captures whether increasingly stricter requirement are placed on where the underlying electricity is delivered to, does not seem to have a statistically significant impact on renewable share. However, the delivery-from-region variable, which essentially captures whether increasingly stricter requirement are placed on where the underlying electricity is delivered from, seems to have a negative impact on renewable share. Placing increasing restrictions on where the electricity is eligible to be delivered from apparently harms total RES-E deployment. 5.5 Technology Analyses In this section, we examine the impact RPS and other polices on the capacity of biomass, geothermal, solar, and wind. Due to lack of space we do not present the table with statistical results and discuss the results directly instead. In general, the regressions with biomass turn out to be similar to the full model results in Section 5.2. Since biomass capacity is by far the largest among all RES-E capacities, we argue that biomass deployment drives the overall results. ISI has a statistically significant negative effect on biomass capacity development, while no significant link could be found between the ISI and any of geothermal, solar, or wind development. That is, the negative impact of ISI on total renewable share is driven by the impact on biomass. However, after excluding Maine the significance disappears for biomass. Again, the outlier seems to bias the coefficient of the full sample. The presence of a public benefit fund has a statistically significant positive impact on biomass capacity development. This finding is consistent with the hypothesis that biomass plants have been the principal beneficiaries of this policy. On the other hand, no statistically significant link could be found between public funds and the deployment of other renewable technologies. We estimate a significant negative coefficient on the existence of net metering on wind development and insignificant coefficients on biomass and solar capacity. Recalling our interpretation in Section 5.2, we 16