Allo cating Emissions Permits in Cap-and-trade Programs: Theory and Evidence Preliminary. Please do not cite without p ermission Comments very welcome February 2010 Meredith Fowliey Abst ract The allo cation of emis sion s p ermits in ontentiou s p olicy d esign is su e. Re ce ncy an d distributional imp lic ations of a ermits in d etail. This p ap er te sts wh e the s tan dard theory. I de velop a s imp oretical re lationsh ips b e twee n p e rm uction de cision s. Data gathe red f rom us ed to analyze the se relation ship s e f or explicit e nvironm ental compliance c emissions ) in their short-ru n s up ply d resp ondin g to the le ss s alient pro d uc e rmit allo cation u p dating. I th an k Step hen Holland , E rin Mans ur, and iversity of California, Berkeley, th e Un ive rsity of M elpfu l comme nts an d d is cus sion. I grate fu lly a th rou gh grant SE S-0922401, the UC En ergy Ins University of Michigan. Erik John son provide d e x yDepartment of Agric ultural and Resou rce Econ o Berkeley CA, 94720. em ail: fowlie@b erkeley.ed u 1 Intro duction 1Billions of dollars worth of tradable emissions p ermits are allo cated each year to U.S. industrial pro ducers regulated under emissions "cap-and-trade" programs.In theory, how these p ermits are allo cated can have signi cant implications for who will b ear the costs and how e¢ ciently the mandated emissions reductions will b e achieved. Permit allo cation has thus emerged as one of the more contentious issues in p ermit market design. Regulatory agencies have b een allo cating tradable emissions p ermits under the auspices of lo cal, regional, and nationwide cap-and-trade programs for over a decade. Over this time p erio d, theoretical analyses of the e¢ ciency and distributional implications of p ermit market design choices have grown increasingly sophisticated. However, we know relatively little ab out how p ermit allocation design a/ects rm decision-making in real world settings. This pap er brings evidence to b ear on a rst-order empirical question: are rms resp onding to p ermit allo cation incentives as standard theory predicts? Traditionally, p olicy makers have chosen b etween two general approaches to allo cating emissions p ermits: auctioning and grandfathering. Under an auction regime, emissions p ermits are sold to the highest bidder. In contrast, "grandfathered" p ermits are freely distributed to regulated sources based on pre-determined, rm-sp eci c characteristics. In the absence of other market failures, this choice b etween grandfathering and auctioning should have no b earing on p ermit market e¢ ciency in the short-run (Montgomery, 1974). 2Many economists favor auctioning on the grounds that revenues can b e used to o/set distortionary taxes (Crampton and Kerr, 2002; Goulder et al.,1999). 3However, in practice, p olicy makers have routinely chosen to forego auction revenues in favor of handing p ermits out for free to regulated entities.The ability to make concessions to adversely impacted and p olitically p ow1The value of p e rmits allo cate d to emittin g f ac ilitie s und er the NOx Bud get P rogram and the Acid Rain Program each year is rough ly $1.4 B and $4.5B, resp ectively. The U.S . Environ mental Prote ction Agenc y (EPA) has estimated th at th e value of allowan ces allo cated an nu ally un der prop ose d Fe deral climate le gislation would exce ed $200 b illion (US E PA, 2008) . 2Oth er e ¢ cien cy-re lated argume nts in f avor of auc tioning p e rtain to the mitigation of p re -existin g regu latory distortions an d distrib ution al con cerns . For examp le, Dinan and Rogers (2002) an d Parry (2004)? emp hasize th e p otential distribu tion al implic ations of th e allo cation d esign choice, de mons tratin g that high income ind ividu als are like ly to gain more f rom f re ely allo cate d allowanc es th an are low inc ome ind ividu als, 3A ma jority of p e rmits are distrib uted f ree ly to re gu late d e ntitie s in Sou th ern California s RE CLAIM program, 1 erful stakeholders via grandfathering has arguably b een as imp ortant a factor in the widespread adoption of emissions trading programs as the promise of cost minimization and gains from trade. More recently, a third design alternative has emerged. Under a "contingent allo cation" regime, up dating rules established ex ante are used to determine how a rm s p ermit allo cations will b e p erio dically up dated over the course of the trading program. Allo cation up dating is typically based on a rm s pro duction choices (such as output levels or fuel inputs). The incentives created by contingent allo cation rules are quite di/erent from those asso ciated with grandfathering or auctioning b ecause up dating creates an incentive to increase whatever activity determines future emissions p ermit allo cations. 4In a theoretical, " rst-b est" setting, it is straightforward to demonstrate that p erio dically up dating rms future p ermit allo cations based on present pro duction choices will undermine the e¢ ciency of p ermit market outcomes b ecause the implicit subsidy conferred by allo cation up dating encourages rms to increase output to economically ine¢ cient levels. (Bohringer and Lange, 2005; Sterner and Muller, 2008).However, contingent up dating can welfare dominate more standard p ermit allo cation approaches when there are additional, pre-existing distortions to contend with. For example, the theory literature has explored how allo cation up dating can b e used reduce ine¢ ciencies resulting from the exercise of market p ower (Fischer, 2003; Gerbasch and Requate, 2004; Neuho/, Martinez, and Sato, 2006), tax interaction e/ects (Fischer and Fox, 2007), and emissions leakage (Bernard et al., 2007; Quirion and Demailly, 2006). Allo cation up dating can also b e used to address p olitical concerns ab out the incidence of a cap-and-trade programs (Jensen and Rasmussen, 2000). 5Political supp ort for contingent allo cation up dating is increasing. Industry groups endorse it as a "common sense way to promote e¢ ciency, fairness, and environmental protection".Policy exp erts concede that it may o/er the most pragmatic approach to mitigating the adverse e/ects of environmental regulation on domestic industry comp etitiveness when emissions regulations are the Eu rop e an Un ion s Emiss ion s Trading Program, th e nation -wide Ac id Rain Program, and the re gion al NOx Bu dget Trad in g Program. 4Here, " rst-b es t" ref ers to a regulatory environ ment in wh ich th e only market d is tortion or imp e rf ection is th e environme ntal exte rn ality that the emissions regulation is de signed to internalize. 5"Advantages of Allo cating E missions Credits Based on E ¢ cienc y." Th e Clean E nergy Group . May 15, 2009. http ://www.thec le an en ergygrou p.com/lsgbrie n gs .as p 2 6incomplete.7Federal climate legislation passed in June 2009 by the House of Representatives includes provisions for allo cation up dating as a means of comp ensating trade exp osed emitters for compliance costs incurred.8In California, legislators are considering allo cation up dating as a means of mitigating impacts on consumer prices and reducing emissions leakage to unregulated entities (CPUC-CEC, 2008). 9Can p ermit allo cation design b e e/ectively used to achieve these kinds of p olicy ob jectives? This dep ends in large part on whether rms resp ond to p ermit market incentives as standard theory predicts. In p olicy debates, stakeholders questioned the extent to which the implicit subsidy conferred by up dating will b e factored into rms "real world" pro duction decisions.Others contend that these subsidy e/ects will b e mitigated by pre-existing regulatory and market constraints (NCEP, 2008). Finally, past research has demonstrated that gain-loss asymmetries, asymmetric pass-through of op erating costs, and heuristic approaches to dealing with cognitive constraints can all a/ect how rms p ercieve and resp ond to di/erent kinds of pro duction incentives (see, for example, Duxbury and Summers, 2004; Hirshliefer, 2001 ; Zachmann and von Hirschhausen, 2008). If these kinds of institutional factors and/or b ehavioral phenomena a/ect rms environmental compliance decisions, p ermit allo cation rules may not have the intended e/ect. This study makes three contributions to b oth the academic literature and the ongoing p olicy discourse. First, a static analytical mo del is used to illustrate the essential short-run implications of the p ermit allo cation design. In an e/ort to clarify the terms of the p olicy debate, these implications are sometimes over-simpli ed to the p oint of misrepresentation. For example, a recent and in uential rep ort that seeks to "clear up misp erceptions, common among many stakeholders, ab out how allo cation decisions do and do not a/ect the way an emissions trading program works in 6Se e, f or example, th e testimony of Richard Morge ns te rn of Re sources f or the Future. Competitiveness and Cl imate Pol icy : Avoiding Leakage of Jobs and Emissions : Hearings before the Committee on Energy and Commerce U.S. House of Representatives. March 18, 2009 7The ju sti cation for contin ge nt u p dating in this c ontext re sts on th e con cern th at rms would dive rt n ew inve stments and pro d uc tion to manuf acturin g facilities lo c ated in c ountries withou t commen surate regulations . Continge nt allo cation up datin g, inten de d as a s top -gap me as ure, comp en sates rms for the c omplian ce c osts incu rred so as to mitigate adverse comp e titive nes s imp ac ts. 8The d esign re commen dation s of b oth the Calif orn ia Pu blic Utilitie s Comm is sion an d th e WCI in clud e th e minimization of the impacts of carb on re gu lations on cons ume rs and the mitigation of leakage as ob jec tives of the allo cation pro ces s. In th is context, allo c ation u p dating h as eme rge d as a p olitically palatable option. 9Se e, for example, RGGI, 2004. 3 practice" asserts that [a]lo cation a/ects the distribution of b ene ts and burdens among rms and industry sectors; it do es not change program results or overall costs (NCEP, 2007)." In fact, this conventional wisdom do es not hold when future p ermit allo cations are contingent up on current pro duction choices. The simple mo del presented in this pap er intuitively demonstrates how p ermit allo cation design can signi cantly a/ect rms short-run pro duction choices (and thus aggregate so cial costs). Second, the study provides insight into how di/erent p ermit market incentives are a/ecting rms short-run pro duction decisions. An emerging literature examines the relationships b etween p ermit price series and wholesale electricity price series (Bunn and Fezzi, 2007; Fell, 2009; Sijm et al., 2006 ; ). Given the complexity of interactions b etween p ermit markets and wholesale electricity markets, it has b een di¢ cult to deduce rm-level b ehaviors from wholesale market price dynamics. This pap er examines rm-level resp onses to changing p ermit market conditions in unprecedented detail. 1 0Finally, the pap er o/ers some of the rst empirical evidence of how contingent allo cation up dating is working in practice. Previous attempts to study the e/ects of contingent p ermit allocation up dating, and revenue recycling more generally, have b een unable to convincingly separate the e/ects of the implicit subsidy from the overall e/ect of the environmental regulation.I exploit an unusual p olicy setting in which the implicit subsidies conferred by a regional emissions trading program vary systematically across pro ducers and across seasons. Much of this variation is exogenous- in an econometric sense- to rms short-run pro duction decisions. The empirical analysis b egins by demonstrating that statistical relationships in the data are generally consistent with standard theoretical predictions regarding pro ducers resp onse to p ermit allo cation incentives. I then derive a tractable reduced form mo del of the underlying data generating pro cess that can b e implemented empirically using a discrete choice framework. Conditional on 1 0Sterne r and Is akss on (2006) ? we re the rs t to e mpirically inves tigate th e e/e cts of reve nu e re cyc lin g in the context of marke t-base d emiss ions regu lation. Th ey an alyze a Swe dish program in which e missions charge s are refu nd ed to p ollu ting rms in p rop ortion to outp ut. Becau se reb ate s do not vary across rms or acros s time , the authors c an not sep arate the e/e ct of the tax fro th e e/ect of th e re cycled revenue s. In a more rec ent p ap er, Sj im et al. (2006) en counter s im ilar d i¢ c ultie s in their analysis of how se qu entially gran df ath ere d p e rmits in the E U ETS imp ac ts electric ity market outcomes . The lack of clarity s urroun ding how cu rrent p ro du ction d ecisions will in uen ce fu tu re p ermit allo c ations in the Eu rop e an Un ion s E mis sions Tradin g System comp lic ates th eir analysis of how th is implic it u p dating h as a/ec ted rm de cision makin g. 4 the assumptions of the mo del, this provides a basis for investigating causal relationships b etween p ermit allo cation design features and rms short-run pro duction decisions. I provide evidence to suggest that rms are accounting for the explicit compliance costs (i.e. the costs of holding p ermits to o/set emissions) in their short-run pro duction decisions. The data also indicate that pro ducers in allo cation up dating regimes are resp onding to the implicit subsidy. I fail to reject the hyp othesis that pro ducers are equally attentive to the p ollution disincentive and pro duction incentives conferred by allo cation up dating. The pap er pro ceeds as follows. Section 2 develops a simple theoretical mo del that is used to intuitively demonstrate the rst order implications of alternative p ermit allo cation designs. Section 3 intro duces the NOx Budget Program. Section 4 discusses the data collected from this program. Section 5 discusses the underlying data generating pro cess and lays out an empirical strategy. Section 6 summarize the estimation results. Section 7 concludes. 2 Allo cating emissions p ermits: Theory The analytical framework intro duced in this section serves two purp oses. The mo del is rst used to demonstrate the most essential partial equilibrium implications of the p ermit allo cation design choice and then to motivate the empirical analysis. The mo del is intentionally simple. Many of the institutional details and market imp erfections captured by mo dels found elsewhere in the literature (such as pre-existing tax distortions, the exercise of market p ower, or incomplete regulation) have b een stripp ed away. Eliminating some of the complexities of real p olicy settings helps to highlight the most basic trade-o/s b etween static pro duction e¢ ciency, static allo cative e¢ ciency, and distributional concerns. The analysis will fo cus on short-run relationships exclusively. To the extent that allo cation up dating is seen as a way to smo oth the transition to auctioning regimes, these short-run relationships will b e very imp ortant. Also, a clear characterization of short-run interactions is an essential rst step towards understanding longer-run outcomes and implications. 5 2.1 A static partial equilibrium framework Pro duction of a homogeneous go o d generates harmful p ollution. Industry pro duction in time t is denoted Qt. Let qitdenote the quantity pro duced by rm i in time t. Pro ducers are characterized by increasing marginal cost technologies; Ci(qi) and cdenotes the unit-sp eci c total cost function and marginal cost, resp ectively. E¢ cient factor markets are assumed. Marginal op erating costs thus re ect the true opp ortunity cost of allo cating inputs to pro duction in this industry. Emissions rates eiiqitare constant p er unit of output. Demand is characterized by an a¢ ne inverse demand function Pt= a bQt. To keep the mo del transparent and tractable, only two price taking rms are represented. A more general mo del with N > 2 is easily formulated but more di¢ cult to intuitively interpret. The two rms are indexed c (denoting the relatively "clean" pro ducer) and d (denoting the relatively "dirty" pro ducer). For the purp ose of this example, I assume emissions rates are negatively correlated with op erating costs: ; ec< ed, cc> cd1 1: Industry emissions are regulated under a cap-and-trade program; aggregate emissions in p erio d t cannot exceed an exogenously determined cap E: The time path of p ermitted emissions (i.e. E1; E2t; :::) is set by the regulator ex ante. To comply with the program, rms must o/set uncontrolled emissions with p ermits. These p ermits are tradable in an emissions p ermit market. There are no spatial or temp oral restrictions on p ermit trading. I assume that rms acts as price takers in b oth the p ermit and pro duct markets. This short-run analysis conditions on existing pro duction technology and op erating characteristics; emissions rates and op erating costs are exogenously determined and xed. Emissions reductions can thus b e achieved in two ways: through increasing the share of the market served by the relatively clean pro ducer or reducing the quantity consumed. In existing and planned cap-and-trade programs, p ermits are allo cated via auctioning, grandfathering, symmetric "output-based" up dating, or asymmetric up dating. Stylized representations of all four approaches are considered b elow. 1 1This assu mption n ds emp irical sup p ort in the e missions market I cons id er he re. Un it-leve l f ue l op erating c osts and NOx emiss ion s rates are n egatively corre lated in the data analyz ed in the sub seque nt s ection. However, the re are ce rtainly examples of low-e mitting f acilities with re latively low op erating costs, and high-e mitting f acilities with relatively high op e rating cos ts . 6 2.2 The b enchmark case Outcomes under alternative allo cation rules will b e compared against a " rst b est" b enchmark that maximizes total economic surplus S (Q) sub ject to technology op erating constraints and the constraint that aggregate emissions do not exceed the exogenously set cap: : S (Qt) = Z P (Qt)dQt Cc(qct) Cd(qdt) (1) Qt 0 max qct;qdt s:t: ecqct + edqdt qc + qdt = Qt: = Et t Note that the welfare measure S (Qt) re ects the utility asso ciated with total consumption less pro duction costs but do es not capture the b ene ts asso ciated with emissions reductions. Because aggregate emissions are held constant at E across all scenarios, changes in S (Q t) across scenarios will re ect changes in absolute welfare vis a vis this b enchmark. The rst order conditions for this maximization problem imply: P (Q t) tei = 0; i = c; d: (2) ciq it P cdP is the shadow value of the emissions constraint at time t. ccqq d cmaximize economic surplus sub ject to the constraint. where t Rearranging these rst order conditions (and omitting clarity) yields: = edec: (3) edFigure 1 helps to illustrate this result. The downward sloping line, representing the emissions constraint, connects all allo cations of pro duction across the two rms that exactly satisfy the emissions cap. The slop e of this line isec: The economic surplus function S (Q) is also pro jected into this space. The level sets of the surplus function app ear as concentric iso-surplus curves. The slop e of an iso-surplus curve measures the rate at which pro duction at the clean rm can b e substituted for pro duction at the dirty rm while holding total economic surplus constant. The 7 so cially optimal allo cation of pro duction o ccurs at the p oint where the emissions constraint is just tangent to an iso-surplus curve. All other p oints that exactly satisfy the emissions constraint are asso ciated with lower levels of economic surplus. Two e¢ ciency prop erties of this equilibrium are worth highlighting: Property 1 : Marginal abatement costs (measured in terms of foregone pro ts p er unit of emissions reduction) are set equal across pro ducers: P =P : (4) ceddq d ceccq c This assures that abatement activities have b een e¢ ciently allo cated among pro ducers. Given pro duction level Q , the cost of meeting the emissions constraint E is minimized. Property 2 : Emissions abatement activities are allo cated e¢ ciently across the supply and demandside of the pro duct market: @ S @ccq c cd : (5) E= ed ecq d Intuitively, the derivative of the welfare function with resp ect to the emissions constraint captures the marginal cost of reducing emissions via conservation measures on the demand side (i.e. through a reduction in consumption). The marginal abatement cost on the supply side is the cost of reallo cating pro duction from the low cost, high emitting pro ducer to the high cost, low emitting pro ducer so as to incrementally reduce emissions. Equation [5] implies that an optimal balance is struck b etween the two short-run abatement options. This result is derived in App endix 1. Returning to Figure 1, the broken line connects all p oints asso ciated with an aggregate production level Q : Pro duction allo cations on the emissions constraint lying strictly ab ove (b elow) the optimal outcome are asso ciated with more (less) consumption than is consistent with compliance constrained economic surplus maximization, thus requiring higher (lower) levels of supply-side abatement to achieve compliance. 2.3 Grandfathering and auctioning regimes 8 I now consider a p erfectly comp etitive industry sub ject to a market-based emissions cap-and-trade program. Let Aitrepresent the rm s p ermit allo cation in p erio d t. Under grandfathering, the numb er of p ermits the rm receives (free of charge) from the regulator each p erio d is determined at the outset of the program. Under auctioning, A it= 0 8 i: Under either scenario, rms future p ermit allo cations are indep endent of their pro duction decisions going forward. Let trepresent the p ermit price (an endogenously determined parameter). The rm s compliance costs (i.e. the cost of holding p ermits to o/set uncontrolled emissions) are t(Ait eiqit): The pro t maximization problem faced by price taking rm i in time p erio d t is thus: max : it = Ptqit C (qit) + t(Ait eiqit); i = c; d: (6) Assuming price-taking b ehavior in b oth the p ermit and pro duct markets, the rst order conditions for this pro t maximization problem are given by [2]. Thus, the e¢ ciency prop erties [4] and [5] are satis ed under b oth grandfathering and auctioning. 2.4 Output-based up dating Under a contingent up dating regime, the total quantity of emissions p ermits to b e allo cated in each p erio d, and the rules sp ecifying how rms pro duction decisions will determine future p ermit allo cations, are determined at the outset of the program. An output-based up dating regime de nes a rm s p ermit allo cation in p erio d t as a function of the rms pro duction decisions in the preceding p erio d or p erio ds. I consider the simplest of output-based allo cation rules: a rm s p ermit allo cation in p erio d t + 1 is determined by its market share in p erio d t: Ait+1 = qit stqit: (7) Et+1Qt The size of the subsidy conferred by output-based up dating will dep end on the total numb er of p ermits allo cated in the future p erio d E, how the rm discounts future revenue streams i(t), the future p ermit price t+1t+1, and total industry pro duction Qt: Contingent-up dating adds 9 an additional argument to the rm s pro t function.: max : it = Pitqit C (qit) teiqit + i(1) t+1stqit: (8) Some additional assumptions further simplify the analysis. Unrestricted banking and b orrowing of p ermits, rational exp ectations, and zero arbitrage together imply that p ermit prices are constant in present value terms. If all rms discount future p erio d gains at the market rate and take total sector output Qas given, the implicit subsidy p er unit of pro duction in p erio d t simpli es to st1 2.tOmitting t subscripts for simplicity, the rst order conditions for pro t maximization under output-based up dating are: 0P (Q) (ei s ) = 0; i = c; d (9) ciq0 i The subscript denotes equilibrium outcomes under non-contingent p ermit up dating. Rearranging [9] yields: ed s ec s P cdP = ccqq0 d0 c : This equilibrium outcome { q0 c, q0 d 13 12 t tAt+1qt Q2 t tim e t + 1 p er s hare of ou tput Qt 13 10 g lies strictly ab ove the optimal outcome on the emissions constraint line and is asso ciated with a level of total economic surplus that is strictly less than S (Q) (see gure 1): Intuitively, the implicit pro duction subsidy increases each rm s willingness to supply vis a vis grandfathering or auctioning. At higher levels of pro duction, more supply-side abatement is required to comply with the emissions cap. Consequently, the market share of the relatively clean pro ducer increases relative to the rst b est b enchmark. App endix 2 demonstrates that the equilibrium p ermit price will b e unambiguously higher than , re ecting higher supply-side marginal abatement costs. Consumption levels exceed that asso ciated with optimal compliance, so there is a transfer of surplus from pro ducers to consumers vis a vis grandfathering. Alte rn atively, rms could take into accou nt how the ir own p ro du ction dec is ion s a/ec t aggre gate pro d uction levels (an d thus th e s iz e of the imp licit s ubs id y). Th is would re duc e th e p e rce ive d p ro du ction su bsid y by: Intu itively, if the rm in crementally in crease s pro duc tion in time t , it d ecrease s the nu mb er of p e rmits allo cated in , alth ough it is n ow entitled to an ad dition al s hare . Making s imilar distribu tion al comparis ons with auc tionin g is c om plic ate d by the f ac t that con sum er and pro- 2.5 Asymmetric contingent allo cation In practice, the implicit pro duction subsidy intro duced by contingent up dating is often asymmetric. For p olitical and practical reasons, the s parameter often varies across units based on technology typ e, fuel e¢ ciency, or other observable op erating characteristics. To accommo date asymmetric up dating, the rm s ob jective function is mo di ed slightly to allow the implicit pro duction subsidy to vary across rms. Omitting the t subscripts for simplicity, the rst order conditions for pro t maximization are: 00P (Q) (ei + si) = 0; i = c; d (10) ciq00 i I assume that the up dating parameters are de ned such that the total numb er of p ermits allo cated through up dating do es not exceed the total cap. This implies that the average up dating parameter cannot exceed the average emissions rate. In this asymmetric case, the equilibrium p ermit price is no longer equal to the economic cost of redispatching pro duction activity in order to incrementally reduce emissions. If the implicit pro duction subsidy favors the relatively dirty (clean) rm, the p ermit price must rise ab ove (fall b elow) the true marginal abatement cost in order to counteract this asymmetry in compliance incentives. The equilibrium p ermit price is: = cedcq qd cd + (sc ) : (11) ecc sd 0 0 The e¢ ciency and distributional implications of asymmetric up dating vis a vis rst b est will dep end on the ratio of the up dating parameters scand sdsd. If the implicit subsidy p er unit of p ollution is exactly equal across rms (i.e. implying thatsc=edec), the rst b est quantities and pro duct price are achieved (see App endix 3). If the subsidy p er unit of emissions is larger for the relatively clean rm (as is the case with output based up dating), the outcome will o ccur at a p oint on the emissions constraint ab ove the optimal outcome. This would result in to o little conservation on the demand side and to o much abatement on the supply side. Conversely, if the subsidy p er unit of emissions is larger for the more p olluting rm, consumer prices will increase and consumption levels will fall relative to rst b est. A larger share of the mandated emissions du cer s urplus und er auc tioning will de p end on how au ction re venues are allo cated. 11 reductions will b e achieved on the demand-side. 2.6 Motivating the empirical exercise sConditional on the assumptions outlined ab ove, contingent allo cation up dating will distort outcomes away from the e¢ cient short-run equilibrium in a rst b est setting (except in the very sp ecial case wheresij=eeij; 8 i 6 =; j ). Abatement e/orts will b e ine¢ ciently allo cated across producers and consumers, and across rms with di/erent pro duction technologies and cost structures. General equilibrium mo dels calibrated to sp eci c p olicy contexts have b een used to quantitatively estimate the p otential magnitude of these distortions. In several instances, ine¢ ciencies induced by allo cation up dating have b een found to b e economically signi cant (Burtraw et al. (2005), Jensen and Rasmussen (2000), Neuho/ et al. (2005)). Much of the theory literature is devoted to extending this kind of analytical exercise to more complicated, second-b est settings. In cases where the implicit subsidy can b e used to mitigate one or more pre-existing distortions or imp erfections (such as the exercise of market p ower in the pro duct market, or incomplete emissions regulation) contingent allo cation up dating may welfare dominate grandfathering and auctioning. This entire literature is predicated on the assumption that compliance cost minimizing rms in cap-and-trade programs accurately account for all p ermit market incentives in their supply decisions However, there are several reasons why this assumption might not hold in practice. First, the b ehavioral nance literature o/ers evidence to suggest that rms may fo cus on information that is more readily accessible and easy to understand at the exp ense of information that is more opaque or that requires more resources to pro cess (Hirshleifer, 2001; Sarin and Web er, 1993). It is arguably much easier for plant managers to translate p ermit prices into compliance costs p er unit of pro duction, versus the exp ected future subsidy p er unit of current pro duction. Particularly when allo cation up dating rules are convoluted or confusing. Second, researchers have found evidence of gain loss asymmetry (whereby agents place more emphasis on minimizing losses versus maximizing gains) in private sector decision-making (Fiegenbaum, 1990). A recent study do cuments asymmetric cost transmission in p ermit markets, whereby 12 rms pass op erating cost increases through to customers at a di/erent rate than they pass any cost reductions (Zachmann and von Hirschhausen, 2008). To the extent that these b ehavioral elements a/ect rms environmental compliance decisions, managers may discount- or ignorethe implicit pro duction subsidy. Finally, if there is any uncertainty regarding how the cap-and-trade regulation will b e implemented or mo di ed in the future, managers may discount the implicit pro duction subsidy to re ect this regulatory risk. In sum, whereas it is standard to assume that the explicit compliance cost and implicit pro duction subsidy are weighted equally in rms pro duction decisions, this may not b e the case in practice. The obligation to hold p ermits to o/set uncontrolled emissions may a/ect short-run compliance decisions di/erently than the implicit subsidy conferred by p ermit allo cation up dating. If rms discount- or ignore- the implicit subsidy conferred by contingent allo cation up dating, p ermit allo cation incentives will not have the intended e/ects on market outcomes. 3 Empirical application: The NOx Budget Program The data in this study come from a ma jor U.S. emissions trading program: The NOx Budget Program. In 1998, the U.S. Environmental Protection Agency (EPA) determined that 23 eastern states were contributing signi cantly to ozone non-attainment problems. These states were issued "NOx budgets" and required to design and implement regulations that would reduce seasonal NOx emissions to budget levels. Although states had exibility in cho osing their compliance strategies, they were invited to meet their compliance obligations by joining an EPA-administered cap-and-trade program. All states accepted the invitation. 1 4States in the NOx Budget Program (NBP) were required to accept program design features outlined in a mo del rule that was issued by the EPA. These features include p ermit trading proto cols and emissions rep orting standards. For instance, throughout the program, a NOx p ermit authorizes the holder to emit one ton of NOx during "ozone season" (i.e. May to Septemb er). 1 4Compliance is on ly re qu ired in th e sp ring and su mmer wh en ave rage temp eratu res rise an d NOx emiss ion s contribu te to smog f ormation . 13 1 5At the end of each season, all regulated source must hold su¢ cient p ermits to o/set ozone season emissions.There are no spatial trading restrictions in the NBP; p ermits are freely traded among all participating sources in all participating states. 1 6The mo del rule also required standardization of intertemp oral trading restrictions across participating states. Emitters cannot b orrow against future allo cations. Emissions in year t must b e o/set using p ermits of vintage t or earlier. Permits can b e banked, although the use of banked p ermits is sub ject to a "progressive ow control" (PFC) constraint designed to discourage the excessive use of banked p ermits in a particular year. 3.1 Permit allo cation in the NBP 1 7In the pro cess of designing the NBP mo del rule, the US EPA commissioned an ex ante analysis of p ermit allo cation design alternatives. A detailed numerical simulation mo del of the electricity sector was used to evaluate various allo cation regimes, including grandfathering, output-based allo cation up dating, and fuel input-based allo cation up dating (US EPA, 1999). Simulation results indicated that the p ermit allo cation design choice would appreciably a/ect market outcomes. Consumer electricity prices were pro jected to b e 3.4 p ercent lower in an allo cation up dating regime as compared to grandfathering. This amounts to a transfer from pro ducers to consumers of $1.25 billion. The study also indicated that emissions leakage to neighb oring, unregulated states would b e reduced under allo cation up dating; simulated electricity pro duction in regulated jurisdictions was 10 p ercent higher under up dating as compared to grandfathering. With fewer emissions reductions coming from demand-side conservation/substitution and/or substitution of imp orts for domestic pro duction, more of the mandated emissions reductions had to come from regulated 1 5If a facility s emiss ion s excee d its p erm it allo cation , th e facility mu st p urch ase add itional NOx p e rmits in the p ermit marke t. Complian ce h as b ee n ne arly p erfec t ove r the duration of the p rogram; the f ew c as es of noncomp lianc e h ave b e en attrib uted to accounting errors. 1 6By law, if th e nu mb e r of p ermits in th e re gion- wid e bank prior to ozone s eason exce eds 10 p erce nt of the total (i.e . p rogram -wide) cap for that s eason, a non -lin ear disc ou nt factor is ap plied. Th e P FC ratio is comp uted as 10 p e rce nt of th e season al cap divid ed by the size of the ban k. This ratio d e n es the f raction of b anked p e rmits th at can b e us ed to o/set a ton of p ermits that se ason. Th e remaining p ermits can b e use d to o/s et only a h alf ton.The disc ount factor is ap plie d at the f ac ility level. For example, if a single rm hold s 100 p erm its and th e ratio in year t is d e n ed to b e 0.5, that rm can us e 50 b an ked p ermits to o/set e missions in year t on a on e f or on e b asis. 1 7Bec ause ove r 90 p e rc ent of emis sion s re gulate d un der th is p rogram c om e from e le ctricity p ro du cers, EPA analys is fo cu sed exclus ively on the electric ity s ector. This an alysis ac counted f or b oth short and long-run re sp on ses to p ermit allo cation ince ntives. 14 pro ducers. Supply-side abatement costs were pro jected to b e 18 p ercent higher under up dating versus grandfathering, largely due to an increased share of electricity supply provided by relatively clean- and relatively more costly- natural gas units. 1 8Whereas many imp ortant program design features were de ned at the federal level, states were ultimately given broad exibility with regards to p ermit allo cation design. The EPA recommended allo cation up dating based on heat inputs, but states were free to deviate from this recommendation.Several states chose to pursue alternative approaches; p ermit allo cation rules vary signi cantly in terms of overall regime choice (i.e. grandfathering, or contingent up dating), the frequency of allo cations, and the basis for distributing the allowances. 4 Data 1 9I use data from the rst four years of the NOx Budget Program (i.e. 2003-2006) and fo cus exclusively on electricity pro ducers serving restructured wholesale electricity markets in the eastern United States (i.e. the New York, New England , and Mid-Atlantic or "PJM" markets).Data sources and details are provided in App endix 4. State-level permit al location regimes Table 1 rep orts state-level NOx budgets (which were pre-determined and do not change over the study p erio d) and information regarding state-sp eci c p ermit allo cation design choices. Whereas smaller states chose grandfathering (due in part to the management resources required to administer a more complex p ermit allo cation up dating pro cess), a ma jority of states chose some form of contingent allo cation up dating based on either output or fuel inputs. 1 8Fu el-bas ed u p dating was ch osen over outpu t-base d u p dating primarily b ecau se, h is torically, emiss ions regulation s h ad b e en de ne d in te rms of m as s e mis sions p er un it of he at in put. 1 9Elec tricity gen erating fac ilitie s (E GUs) comp rise 87 p erce nt of the e mis sions s ou rc es and over 90 p ercent of the NOx em is sion s regulate d un der the NBP (EPA, 2007). E GUs re gulate d un der the NBP op e rate in a variety if elec tricity marke ts . W he re as u nits in the samp le s upp ly res tru ctured wholesale electric ity markets, oth er f ac ilitie s in the p rogram are rate-regulated p ro du cers se rving vertically inte grated, econ omic ally regu lated electric ity marke ts , while oth er un its are own ed and op erated by p ub lic entities and op erate on a n on-p ro t basis. Pro duc tion at rate -re gulate d plants an d pu blic entities are more c entrally co ord in ate d and in ue nce d by an array of ec on omic, regulatory, an d in stitu tion al fac tors. I cho ose to fo cus on restructu re d electric ity markets b e caus e EGUs in th ese marke ts are more like ly to h ave sh ort-run ob jec tives con sistent with pro t maximization. 15 Unit-level operations and attributes 2 0Table 2 summarizes some imp ortant unit-level op erating characteristics by p ermit allo cation regime. Two op erating attributes that are particularly relevant to this analysis are the NOx emissions rate (i.e. p ounds of NOx emitted p er MWh electricity pro duced) and heat rate (i.e. btus of fuel burned p er kWh of electricity pro duced).It is fairly standard in the empirical literature to treat these unit-sp eci c p erformance parameters as immutable features of the pro duction technology. However, emissions rates and heat rates can b e a/ected by op erating decisions made by the plant manager, including the choice of fuel characteristics, utilization rates, and combustion tuning. Purely exogenous factors (such as ambient temp erature) can also play a role. 2 1To construct unit-sp eci c summary measures of these op erating characteristics, separate seasonal regression equations are estimated for each unit. This estimation exercise, describ ed in more detail in App endix 5, obtains unit-sp eci c, season-sp eci c p oint estimates of emission rates and heat rates under average op erating conditions.Capacity-weighted summaries of these estimates are presented in Table 2. The supp ort of the distributions overlap considerably across allo cation regimes, which facilitates an empirical comparison of short-run pro duction decisions made by similar units facing di/erent p ermit allo cation incentives. Emissions permit prices NOx p ermits are actively traded in a liquid p ermit market.2 22 3Brokers facilitate the ma jority of arms length transactions, administer escrow accounts, and provide market analysis. A variety of transaction structures exist in the market, including forward settlements, calls, puts, comp osite structures such as straddles, and vintage swaps. Table 3 rep orts average sp ot NOx p ermit prices by vintage (in nominal dollars p er ton). NOx p ermit prices are falling over the sample p erio d, largely due to abatement costs that proved to b e lower than anticipated, and lower than exp ected temp eratures in the early years of the program.This table also helps to illustrate the e/ect of 2 0A u nit s h eat rate me as ures the e¢ c ie nc y with wh ich the un it trans forms fue l into electric ity. The lower th e he at rate, the more fu el e¢ c ie nt th e gen erator. 2 1Sp eci c ations that allow rates to vary acros s years are also e stimate d. 2 2In 2007, th e volume of "ec onomically signi cant" immed iate settlement trade s (i.e. trade s b e twee n versu s within rms ) re ache d 247,000 tons (EPA 2008). 2 3Evolution M arkets LLC p rovide s inf ormative month ly an alyses of th e NOx B udge t Program p e rmit marke t. 16 2 4the progressive ow control (PFC) constraint on p ermit prices. As early as 2003, p ermit market participants correctly anticipated that the PFC constraint would start to bind in 2005. This explains the large vintage 2004/2005 spread in 2003 and 2004.. Compliance costs and production incentives 2 5To estimate the cost of purchasing p ermits to o/set the emissions asso ciated with generating a MWh of electricity, each unit s NOx emissions rate is multiplied by the NOx p ermit price. Table 4 summarizes these unit-level cost estimates. On average, explicit compliance costs amount to a 7 p ercent increase in total variable (i.e. fuel, op erating and maintenance) costs.However, among units with particularly high emissions rates, this increase can exceed 40 p ercent. Estimating the implicit pro duction subsidies conferred by contingent up dating is more complicated. The size of the pro duction subsidy varies with state p ermit allo cation rules, statesp eci c NOx budgets, annual pro duction levels, and unit-level heat rates (in input-based up dating regimes). Individual states allo cate their resp ective NOx "budgets" (listed in Table 3) using formulas of varying complexity. For each unit, an estimate of the numb er of future p ermits earned p er unit of current pro duction is constructed using the corresp onding state budget Es, average ozone season pro duction (or heat input) aggregated across NBP sources in the state, and the sp eci c details of states p ermit allo cation up dating proto cols. For example, if NOx p ermits in state s are allo cated based on the average heat input in the preceding L years, the e/ect of an incremental increase in current pro duction at rm i in year t on future p ermit entitlements is assumed to b e hi measures the fuel inputs required to generate a unit of output at unit i , and HstEHss t ;where himeasures the total quantity of fuel inputs used by NBP sources in state s over the course of the ozone season in year t. An imp ortant-and plausible- assumption is that rms take the size of the p er-MWh 2 4In years when th e PFC cons traint bind s, ban ke d p ermits trade at a con siderable disc ount. The PFC ratio was 0.25 an d 0.27 in 2005 an d 2006, resp ectively. In b oth ye ars, p ermits we re use d to o/se t e missions at a d is counted rate (4,168 and 1,950 p e rmits in 2005 and 2006, re sp e ctive ly). In M arch of 2005, th e EPA re le as ed its new Clean Air Interstate Rule (CAIR), inten ded to su bsu me th e NBP in 2009. CAIR e liminated progres sive ow control. 2 5Unit-sp e ci c e stim ate s of variable f ue l op eratin g c osts are ob tain ed by multip lying the u nit-le vel heat rate (se e ab ove) by the corres p ond in g fu el p rice. E stimate s of un it- le ve l variable, non -fu el op eratin g and mainten ance costs (not inc lu ding e nvironme ntal comp lianc e cos ts ) are obtain ed from Platts . 17 2 6subsidy as given.Column 3 of Table 4 summarizes these estimated subsidies, in terms of future p ermits allo cated, p er MWh of electricity generated. In present value dollar terms, the estimated implicit subsidy conferred under this input-based up dating regime is: = Ll =1X i(l ) Ll hitHEt s ; (12) 28 sit where l 27 26 st 27 i(l )l 28 is the exp ected p ermit price in l years and (l ) is the discount rate applied to b ene ts accruing l years in the future.In the nal column of Table 4, net compliance costs (i.e. explicit compliance costs less the implicit subsidy p er MWh) are summarized.These incentives vary signi cantly across facilities. Notably, for several units in allo cation up dating regimes with relatively low emissions rates, the estimated implicit subsidy exceeds the estimated explicit compliance cost such that the net e/ect of the NBP on variable op erating costs is negative. 5 Empirical framework The unique design of the NBP provides several p otentially useful sources of variation. First, the delegation of p ermit design to state-level agencies has yielded signi cant interstate variation in p ermit allo cation rules and related incentives. From a research design standp oint, p ermit allo cation design features would ideally have b een randomized across electricity pro ducers. Although states choice of p ermit allo cation regime was not random, interstate variation in p ermit allo cation design is arguably exogenous, in an econometric sense, to rms short-run pro duction decisions. State level p ermit allo cation design decisions were determined by a variety of factors, including the institutional capacities of the implementing agency and the preferences of p olitically p owerful A rm with a d omin ant marke t p os ition wou ld want to accou nt for the f act th at inc re asing its fue l c onsu mption wou ld inc re ase Hand thus d ecrease th e size of the s ubs id y it rec eived for all of its p ro du ction . Thus we would exp ec t the p erceived sub sidy wou ld b e dec re asing in m arket sh are . To c onstruct an es timate of the se implicit s ub sidies, th e fu tu re s pric e of p e rmits issu ed l years in th e fu ture is us ed to es timate : In c ases wh ere p ermits did n ot trad e far en ou gh into the f uture, the market price f or the p e rmit vintage farthes t in the fu tu re was ap plied to all su bse qu ent vintage s.. Rather th an as sume an arb itrary discou nt rate, th ese c alc ulations p res ent un disc ou nte d estimate s of th e imp licit pro duc tion s ubs id y. Plant man age rs p re sum ab ly disc ount the valu e of fu ture p ermit allo cations, so thes e es timates sh ould b e inte rpreted as ge nerous . 18 constituents. Conditional on pre-determined industry and pro duction technology characteristics, the factors that shap ed a state s choice of allo cation regime should not directly impact unit-level short-run pro duction decisions. Second, the seasonal nature of the program s compliance requirements generates useful intertemp oral variation. There is considerable overlap in the distribution of hourly load levels, and other observable market conditions across ozone season and o/-season (see Table 5). This makes it p ossible to observe unit-level pro duction decisions in hours that di/er in terms of ozone compliance requirements, but are otherwise similar. Third, a subset of NOx emitting pro ducers supplying the New England electricity market are exempt from the NBP for meteorological reasons. Most counties in Maine, Vermont, and New Hampshire were already in compliance with the ozone standard. Prevailing wind and weather patterns ensure that emissions from p oint sources in these states do not contribute signi cantly to U.S. non-attainment problems. The distributions of op erating characteristics that determine shortrun pro duction decisions in the exempt and NBP regulated sub-p opulations overlap considerably (see Table 2), making these exempt units a p otentially useful control group. 2 9Finally, a ma jority of states chose to adopt the EPA s recommended p ermit allo cation approach: heat input based up dating. Under this regime, the pro duction subsidy varies signi cantly with fuel e¢ ciency. Input-based up dating thus generates interfacility, intra-market variation in pro duction incentives that is indep endent of variation in explicit compliance costs p er unit of pro duction. 5.1 Mo deling the data generating pro cess The mo del develop ed in section 2 serves as a go o d starting p oint for a mo del of the data generating pro cess. However, a preliminary lo ok at observed hourly electricity pro duction decisions suggests that some mo di cations are needed. First, physical capacity constraints routinely bind at the unit-level in the short-run. Interior solutions are no longer assumed. Second, the vast ma jority of variation in hourly heat rates o ccurs across (versus within) units (see App endix 5). I therefore 2 9The corre lation co e¢ c ie nt b etwe en u nit-level f uel e ¢ cienc y measu re s and NOx e mis sions rates.is 0.69. 19 assume constant unit-level marginal costs. Once these two mo di cations are made, the mo del closely resembles mo dels used to simulate wholesale electricity market outcomes in comp etitive b enchmark analysis (see, for example, Borenstein, Bushnell, and Wolak, 2002 and Wolfram, 1999) and environmental p olicy simulations (examples include US EPA, 1999 and Burtraw, Palmer and Kahn, 2005). This mo del predicts that pro t maximizing electricity pro ducers follow an on-o/ strategy, pro ducing at full capacity whenever price exceeds a reservation price set equal to the unit s marginal op erating cost. In theory, the intro duction of the NBP increases this reservation price by an amount equal to the unit-sp eci c net compliance cost p er MWh. 3 0Are observed pro duction decisions consistent with this prediction? Figure 2 is generated using a small subset of the data collected from a representative unit over a short (three day) p erio d in the ozone o/-season.3 1The left panel plots capacity utilization and hourly wholesale electricity prices over these 96 hours. The horizontal line represents the estimated marginal o/-season op erating cost of $49/MWh sp eci c to this unit and time p erio d (i.e. the reservation price). The right panel plots capacity factor as a function of the wholesale electricity price less variable op erating costs. The thick black step function plots the relationship predicted by the mo del. The thin S-shap ed curve is a lo cal p olynomial smo oth of the observed data. Note that observed hourly pro duction decisions at this unit deviate systematically from the predictions of the simple, static mo del. 3 2Figure 3 conducts a similar exercise using the complete data set. The vertical axis measures capacity factor. The horizontal axis measures price less variable op erating costs (not including NBP compliance costs). The solid line in each panel plots the lo cal mean smo oth of hourly, unitlevel capacity utilization rates on hourly, unit-sp eci c price-cost margins in the ozone o/-season. These functions are generated separately for grandfathering and contingent up dating regimes, 3 0I chos e a p erio d in the ozone o/-s eason in which the wh ole sale e le ctricity p rice was vasc illating aroun d this un it s the oretical res ervation price (i.e . the pre vailing fu el pric e multiplie d by the u nit s f uel e¢ c ie ncy rating plus variable , n on-f ue l op erating c os ts). 3 1The se su pply d ecisions are obse rved in the ozon e o/-seas on. The ave rage ne t comp lianc e cost in curre d by this un it du ring ozone s eason is es timated to b e $3.16. Th e thick re d line illustrate s how the intro d uc tion of the emiss ion s trading p rogram shou ld , in theory, a/ec t the unit s h ou rly pro duc tion dec is ion s. 3 2To ge nerate thes e gu res , I us e an E pane ch nikov we ight fu nction and a rule- of -thumb band wid th es timator. The smo oth is evaluated at 50 p oints. Pric e cos t margin s are calculated by su btracting a un it s f ue l c os ts and non -fu el variab le op erating costs p er MW h f rom th e h ou rly real- time wholesale electric ity p rice. 20 resp ectively. The broken lines plot the same relationships using data from ozone season (i.e. May to Septemb er) when units are required to hold p ermits to o/set their NOx emissions. On average, we should exp ect that the ozone-season supply functions will lie to the right of their o/-season counterparts. Because the average net e/ect of the NBP on variable op erating costs is substantially higher among units op erating in grandfathering regimes (see table 4), we exp ect that the NBP-induced shift in the supply curve will b e larger in the right panel. The data app ear to b e consistent with the rst prediction, less so with the second. These gures suggest that the simple, static mo del p o orly approximates the underlying data generating pro cess. On average, plant managers are willing to op erate in hours when prices fall b elow marginal op erating costs, and are slow to resp ond when the prices rise ab ove cost. Much of this b ehavior can b e attributed to technological and system op erating constraints that are omitted from this static mo del. At the unit level, ramping limits, start up costs, minimum run times, and other intertemp oral op erating constraints can signi cantly a/ect how a plant resp onds to changing market conditions. At the system-level, transmission constraints, system-security requirements, and other op erating proto cols can a/ect which units get called up on to run in a given hour. Intertemp oral op erating constraints The true data generating pro cess involves a complex, multi-p erio d optimization problem. This so-called "unit commitment" problem (i.e. the dynamic scheduling of electricity pro duction over hours in a day.) has b een extensively analyzed in the op erations research and p ower systems literature. The problem is di¢ cult to solve b ecause of its large dimension, non-linearity, and large numb er of constraints (Sheble and Fahd 1990). One formulation of the unit commitment problem, intro duced in App endix 5, helps to convey the complexity of the dynamic optimization faced by plant managers and system op erators each day. Previous work has demonstrated the p erils of ignoring unit commitment idiosyncrasy when simulating short-run wholesale electricity market outcomes in the context of comp etitive b enchmark analyses (Harvey and Hogan, 2002(Harvey & Hogan 2002); Mansur, 2008). Ignoring unit commitment constraints will also b e problematic in this setting, although for di/erent reasons. The extent to which the solution to the unit commitment problem departs from the static mo del 21 of pro t maximization is likely correlated with factors that determine a plant s resp onse to p ermit allo cation incentives. Ignoring unit commitment constraints could bias estimates of rms resp onse. Sp ecifying a fully structural econometric mo del of unit commitment would b e computationally intensive and would require critical assumptions ab out the nature of b oth the system-level and unit-level op erating constraints and proto cols. Given the limitations of the available data, two complementary empirical strategies are pursued. The rst approach is grounded in economic theory only to the extent that theory identi es which variables should a/ect rms supply decisions. The second fo cuses on one sp eci c dimension of the larger unit commitment problem- the decision to b egin op erating conditional on b eing shut down- and derives a reduced form that can b e implemented empirically as a discrete choice problem. 5.2 A descriptive mo del of short-run supply decisions In theory, price-taking pro ducers cho ose output levels based on historic, current, and forecast electricity market conditions, op erating costs, and intertemp oral op erating constraints. My ob jective with the rst empirical exercise is to assess extent to which statistical patterns in the data are consistent with the standard theory. I de ne the following short-run supply function: t e( ei s+ si) + "it; (13) cfit = i + X0 iti + O Zt where i indexes the electricity generating unit, t indexes the time p erio d, and cf is the capacity factor (i.e. the p ercentage of maximum capacity at which the unit is op erating). The deterministic comp onent is comprised of two parts. The rst is a exible function that predicts a rm s op erating capacity, given Xiti+ X0 itwhen the cost of emissions is zero. Let f cfitirepresent this conditional predicted value. The matrix X includes observable variables that a/ect short-run supply decisions but are unrelated to the NBP (including fuel prices, day ahead and real time electricity prices, temp erature, etc.). The second deterministic comp onent in [13] allows predicted pro duction levels to deviate system22 atically from f cfit as the NOx price deviates from zero. The binary variable D _O Zit i and si s e it 33 34 Let Hi 33 35 34 35 23 eq uals one during ozone season and zero otherwise. The NOx p ermit price is . Unit-level emissions rates and implicit subsidies (in terms of future p ermits earned p er unit of electricity generated) are represented by e, resp ectively. The co e¢ cients allow for asymmetric weighting of the explicit compliance costs and implicit pro duction subsidies. We should exp ect to b e negative and to b e p ositive. The error term "is intended to capture the e/ects of unobserved determinants of the op erating decision (including unanticipated deviations from least-cost system dispatch and unit level pro ductivity sho cks). This descriptive mo del can b e used to assess whether the statistical relationships in the observed data are qualitatively consistent with the theory. However, b ecause the assumed linear conditional mean sp eci cation has not b een formally derived from the underlying unit commitment problem, any attempts to draw causal inferences will require additional assumptions. 5.3 A reduced form mo del of the participation decision The second approach is more directly tied to the underlying theory. Consider the unit commitment problem faced by a single electricity generating unit (see App endix 5).The advantage of the second approach is that, conditional on the assumptions of the mo del, there is a basis for investigating causal relationships b etween p ermit allo cation design features and rms short-run pro duction decisions. The disadvantage is that, by conditioning on b eing inactive, only a fraction of the available data are used to estimate the mo del. denote the relevant time horizon (measured in hours) for the unit commitment problem solved by unit i . Because most electricity generating units are incapable of resp onding quickly to changing market conditions, pro duction levels in one hour will constrain pro duction p ossibilities H >0 hours into the future. This es timation fram ework is s im ilar to the ap proach us ed by Man sur (2008) to m o d el the s hort- run pro d uction de cisions of e le ctricity gene rating u nits in the PJM wholesale market. Here I as sume th at hou rly s up ply de cisions are controlled by p lant managers in sofar as th ey c an sub mit b id s to the in dep end ent system op e rator to achieve their de sired pro duc tion le vels. Among the most n imble u nits, H = 0 and the pro d uction d ecision red uc es to th e static b e nchmark mo de l. Let qmeasure the output level at unit i at the b eginning of hour t. The control variable dititmeasures the change in output at unit i in hour t. The transition equation that determines the evolution of the state variable qitover time is thus qit+ dit= qit+1: Let D de ne the decision space. The set of p ossible pro duction level changes available to unit i in hour t, D, will dep end on time invariant parameters of the op erating constraints iitand the state variable qit(qit; i: The family of unit-sp eci c constraint sets is fD)g. Pro ts earned from the sales of electricity generated at unit i in hour t are:it it = (Pit ci t tei + s tsi)(qit + dit) Uiyit Fi; dit 2 f Dit(qit; i)g (14) i ; ); t ; X F 0 included to allow rms to weigh explicit compliance costs and where the parameters are implicit subsidy asymmetrically in their pro duction decisions. The X0 itmatrix includes stat d variables observed byi b oth plant manager and the econometrician, including fuel pric i the ermit prices, day ahead and real time electricity prices. ( t t Note that the contemp oraneous p ermit price tsis used to proxy for rms exp ectations q regarding future p ermit prices in [14]. This implies that managers value future p ermit allo cations at contemp oraneous p ermit prices and employ a discount rate of zero. This simplifying assumption solves a multicollinearity problem; current and future p ermit prices are highly correlated. However, rms presumably use a non-zero discount rate when valu b ene ts accruing in future p erio ds. Moreover, Table 2 illustrates how futures prices can deviate from sp ot prices in this p ermit market. Consequently, captures b oth a discount factor and the average exp ected di/erence b etween current and future p ermit prices. i in [14] equals 1 if the unit turns on in hour t; otherwise y The binary variable yit and Fi it; dit it= 0. Startup costs and xed costs are represented by U, resp ectively. The plant manager s ob jective is to maximize multi-p erio d pro ts (q0; X0i0; d) =HiX) sub ject to the transition function T (q) and the constraint set Dit(qit; it=0F (qit; X0 it; dit): Within this framework, theory makes clear predictions ab out when an inactive unit should 24 st ar t pr o d u ci n g. A pr o t m a xi m izi n g m a n a g er wi ll c h o o s e to in c ur th e c o st s of in iti at in g o p er at io n s if th e pr o ts fr o m d oi n g s o e xc e e d th e pr o ts a ss o ci at e d wi th re m ai ni n g o ut of th e m ar k et . F o c u si n g e xc lu si v el y o n th e p ar tic ip at io n m ar gi n (a n d th u s c o n di ti o ni n g o n qit = 0) gr e at ly si m pl i e s th e d er iv at io n of a n e sti m a bl e re d u c e d fo r m m o d el . L et j to in d e x th e p ar tic ip at io n c h oi c e sit u at io n a n d t in d e x th e h o ur s re le v a nt to th e p ar tic ip at io n d e ci si o n: t = 0: :: H: I d e n e c h oi c e s p e ci c v al u e fu n cti o n s v (y ; X 00 ji3 6) to c a pt ur e th e e x p e ct e d pr o ts a ss o ci at e d wi th p ar tic ip at io n c h oi c e: v (1; X00 j) = E0 H t=1 F [ ; (1) ; X00 j) + (1) ;FX00 j) + X (qt d t 1 j] (15) ( 0 ; d 0 The X00 atri x incl ude s all stat t v (0; X00 j) = E0 [F (0; 0; X00 j; ) + t = 1 X H F ( q t ; d t m (0) ; X00 j) + j: ij 0 e vari abl es rele van t to the part icip atio n cho ice ma de in hou r 0. The se incl ude fuel pric es and p erm it pric es (wh ich do not var y ove r the tim e hori zon H ), the real tim e ele ctric ity pric e in hou r0 and the day ahe ad fore cas t ele ctric ity pric es in hou rs 1:: H. The opti mal pro duc tion cho ice in hou rt con diti ona l on the initi al part icip atio n cho ice y is d(y ). A dec isio n sp eci c sho ck yca ptur es the e/e cts of uno bse rve d fact ors a/e ctin g exp ect ed retu rns to eith er star ting to pro duc e or rem aini ng ina ctiv e. The se fact ors cou ld incl ude pla nt e¢ cie ncy sho cks, uns che dul ed out age s, or op erat or erro rs. I ass um e the se sho cks are add itive and ind ep end entl y dist ribu ted N(0 , ). T o c o m pl et e th e m ot iv at io n of th e e c o n o m et ri c m o d el ,I d e n e a la te nt v ar ia bl e yt o m e a s ur e th e di /e re n c e in th e s e c o n di ti o n al v al u e fu n cti o n s: y = v (1; ij X00 ij The xe d e/e ct i cap ture s unit -sp eci c star t-up cos ts, xed op erat ing cos ts, and oth er uno bse rve d, tim e-in vari ant fact ors in u enc ing the part icip atio n dec isio n. The set of ele ctric ity pric e co e¢ cie nts P t, t= 0 :::H ; cap ture the p erio d by p erio 00 ij) = i v (0; X) + Ht=0 P ij X tP t c ijcij e je i + s jsi + ij (16 ) d di/e ren ces in opti mal out put lev els (co ndit ion al on the part icip atio n dec isio n ma de in p erio d 0) for a part icul ar unit and cho ice situ atio n: P t= (qt( 1) + d t( 1)) (qt( 0) + d t( 0)). Let ijm eas ure the e/e ct of the part icip atio n 3 6 T h e i s u b s c r i p t s h a v e 2 5 b e e n d r o p p e d f o r e a s e o f e x p o s i t i o n . : decision on pro duction over the time horizon Hi c; e; and ij = Ht=1X(qij(1) h +d ij h(1) ij, qij h t i ij ij h(0) d , and s i ij i s ij t(y 0 j )g . The random state variable and 1 ij j 00 ij i j H t=0 P ij tPt c ijcij e je + s jsi! ; (17) serves as an y j ij Pr(yij (0)). The reduced form parameters co e¢ cients are de ned as ; resp ectively. Note that the values vary b oth across units and across the Jchoice situations faced by a given unit i . This is b ecause fuel prices, p ermit prices, and electricity prices will vary across choice situations, and this will have implications for the tra jectories of optimal pro duction choices f dis assumed to b e normally distributed (arising from the di/erence b etween ): The observed binary choice variable y> 0 : y= 1fy ij 1g: The probability that an inactive unit i facing choice situation j will b egin to op erate is given by: i = 1jX ; ) = i +X where is the cumulative distribution function of the standard normal distribution. 6 Estimation 3 7Explanatory variables and sto chastic comp onents o ccur at nested micro (i.e. unit-hour) and macro (i.e. electricity generating unit) levels. Each individual unit is observed making pro duction decisions over several thousand hours, so there are su¢ cient data to carry out a unit-by-unit analysis of relationships b etween short-run supply choices and changing electricity market conditions. However, indep endent variation in p ermit allo cation-related incentives exists only in very limited quantities at the micro-level.The empirical analysis must therefore exploit macro-level variation in unit-level op erating characteristics and state-level p ermit allo cation design choices. Estimation strategies are designed to make e¢ cient use of these hierarchical data without incurring large computational costs. 3 7The re is some with in -un it variation in emiss ion s rates and he at rates (and thus p er-M Wh comp lianc e costs) across ye ars. This variation is extreme ly limited as c ompare d to the signi c ant cross- sec tional variation in e missions rate s and implicit pro d uction s ubs idies. 26 6.1 Estimating the descriptive mo del of unit-level pro duction decisions The iparameters in equation [13] are likely to vary signi cantly across pro duction technologies and op erating proto cols. There are su¢ cient data to estimate these parameters separately for each unit. This could b e done in one step using fully p o oled OLS, but this is cumb ersome b ecause it would involve thousands of interactions b etween the variables in Xitand unit-sp eci c dummy variables. Estimating the mo del in two steps is computationally easier and has exp ositional advantages. 6.1.1 First stage of estimation The primary purp ose of this rst stage is to obtain unit sp eci c estimates of the ico e¢ cients. Unit-sp eci c equations are estimated using micro-level data: cfit = 1 i + X0 iti + i t D _O + "it: (18) Zt The dep endent variable measures unit-level, hourly capacity factor (on a scale of 0-100). The Xit matrix includes 28 covariates: marginal op erating costs, wholesale price (contemp oraneous and two lags), forward price (contemp oraneous and two leads), daily average price (contemp oraneous and lagged), and year xed e/ects. With the exception of the xed e/ects, these variables enter as third-order p olynomials. 3 8To address p otential endogeneity concerns, ISO-sp eci c hourly demand forecasts are used to instrument for electricity prices. Lo cal marginal electricity prices may b e endogenous if unobserved supply sho cks a/ect b oth market clearing prices and unit-level supply decisions. Day ahead demand forecasts should b e indep endent of unobserved supply sho cks but highly correlated with realized demand conditions and thus electricity prices in a given hour. Serial correlation is another concern given the time series nature of these data. Newey-West auto correlation consistent standard errors are estimated assuming a twelve hour lag. 3 8Note that th is error structu re do es n ot accou nt for contemp oraneou s correlation across the un it-le ve l equations. 27 6. 1. 2 S e c o n d st a g e of e sti m at io n In th e s e c o n d st a g e, th e fo ll o wi n g e q u at io n is e sti m at e d u si n g a fe a si bl e g e n er al iz e d le a st s q u ar e s (F G L S) e sti m at or : The estimated b 0 i i 0= e+ i ei s+ si i + i: (19) o e¢ cie nts are regr ess ed on a con sta nt, a me asu re of the unit -sp eci c NO x emi ssio ns rate e; and the unit -sp eci c imp licit pro duc tion sub sidy s. Emi ssio ns rate s at 3 9 40 c unit s exe mpt fro m the NO x Bud get Pro gra m are set to zer o and the vari abl e is de me ane d suc h that the con sta nt ter m cap ture s the ave rag e rela tion shi pb etw een an incr em ent al cha nge in the NO xp erm it pric e and op erat ing cap aciti es am ong unit s with ave rag e NO x emi ssio ns rate s.In mor e e xibl e sp eci cati ons , the co e¢ cie nts are allo we d to var y with obs erv abl e unit cha ract eris tics. T h e s e c o n d st a g e re si d u al s c o nt ai n tw o c o m p o n e nt s: a s a m pl in g er ro r c o m p o n e nt (i. e. th e di /e re n c e b et w e e n th e tr u e v al u e of th e e sti m at e d d e p e n d e nt v ar ia bl e a n d th e tr u e v al u e) a n d th e id io sy n cr at ic v ar ia ti o n th at w o ul d h a v e o bt ai n e d re g ar dl e ss of w h et h er th e d e p e n d e nt v ar ia bl e w a s e sti m at e d or o b s er v e d di re ctl y. If th e s a m pl in g v ar ia n c e di /e rs a cr o ss u ni ts , th is c o m p o n e nt wi ll b e h et er o sk e d a sti c. T o a d dr e ss th is is s u e, a F G L S e sti m at or is u s e d to in c or p or at e in fo r m at io n a b o ut th e v ar ia n c e st ru ct ur e re s ul te d in th e r st st a g e e sti m at io n ( H a n u s h e k, 1 9 7 4) .T h e w ei g ht in g m at ri x u s e d in th e s e c o n d st a g e is gi v e n b y: 2 N k (21) X ) 1b + tr This is a dis ad vantage of es timating rst s tage re gres i sion e qu ations sep arately. GX I treat un it- le ve l emiss ion s rates as xe d. No attemp t is mad e to ac count for the sto ch astic p rop e rties of the se unit-sp eci c op eratin g param eters (se e App end ix 4). Future work will e xp lore alternative app roach es to incorp orating this variation . For e xamp le , Joskow and S ch male nse e (1985) demon strate a cons is te nt (ad ju sted least-s qu ares ) te ch niqu e f or uesing e stimated plant op erating charac teristics as ind ep en de nt variab le s in c ross sec tion regre ssion an alysis. 4 0A more common app roach to adj usting s tan dard e rror e stimate s wh en th e d ep e nde nt variab le is e stim ate d involves weighting sec ond s tage ob se rvation s u sing the invers e of th e es timated s tan dard error of the de p end ent variable (S axonh ouse , 1976). However, this ass umes th at th e total residu al is h eteroskedas tic (versu s j ust the comp on ent that is explain ed by s amplin g error).If th e varian ce du e to s ampling e rror ac counts f or a relatively small fraction of th e total res id ual varian ce in the se cond s tage, this rewe ightin g can ge nerate m is le ad ing s tan dard error es timates . 28 b = b 2 = P i b G + 2I ; (20) 39 2 i b Pi where bG is the variance covariance matrix from the rst stage, e2 iare the rst stage standard errors of the estimated co e¢ cient that serves as the dep endent variable in the second stage, iare the residuals from an OLS estimation of the second stage, k is the numb er of regressors in the second stage and X is the matrix of regressors. This estimator weights more precise rst stage estimates more heavily, but only to the degree that sampling error is an imp ortant comp onent of the overall second stage residual. Standard errors are also clustered at the facility level to account for the fact that the idiosyncratic comp onent of the error term may b e correlated among b oilers lo cated at the same facility. 6.1.3 Results Table 6A summarizes results from estimating [18] using 2SLS. For ease of exp osition, this table rep orts summary statistics for only a subset of the co e¢ cient estimates: the unit-sp eci c constants, fuel costs ($/MWh), hourly wholesale electricity market price ($/MWh), and the NOx price interacted with the ozone season indicator ($/ton). Corresp onding standard errors are also rep orted. 4 1These rst stage results lo ok sensible. The marginal cost co e¢ cient is negative on average, the co e¢ cient on (instrumented) electricity price is p ositive.The co e¢ cient on the NOx p ermit price interaction is also negative on average. Table 7 rep orts results from the second stage of the estimation. The dep endent variable is the unit-sp eci c NOx price co e¢ cient obtained in the rst stage. The most restrictive sp eci cation (1) includes only the demeaned NOx emissions rate and a constant. The intercept is not statistically signi cant at the 10-p ercent level, suggesting that there is no statistical relationship b etween observed pro duction decisions and ozone season NOx prices among NBP units with average NOx emissions rates. This is surprising; we should exp ect these units would pro duce less (more) when increases (decreases) in NOx p ermit prices increase (decrease) marginal op erating costs. The NOX emissions rate co e¢ cient is statistically signi cant and negative at the 1-p ercent 4 1The wh ole sale electric ity pric e co e¢ cie nt is p articularly di¢ cu lt to inte rp ret b ec ause th is variab le is h igh ly corre lated with the lagged prices , day ahe ad lead pric es, an d h igh er ord er te rms that are alos includ ed in the regre ssion. M arginal costs als o ente r as a third order p olynomial. 29 elevel, suggesting that the relationship b etween NOx p ermit prices and hourly output is signi cantly more (less) negative among units with higher (lower) emissions rates. Adding the implicit subsidy variable (sp eci cation 2) marginally improves the t of the mo del. The co e¢ cient on the implicit subsidy is p ositive (as exp ected) but very imprecisely estimated. The co e¢ cient b ecomes more negative (as exp ected), but not signi cantly so. 4 2Column (3) investigates the extent to which these co e¢ cient estimates vary systematically with observable unit characteristics (namely NOx emissions rates and op erating capacity).Because b oth the NOx emissions rate variable and the pro duction capacity variables are demeaned, the constant term now estimates the conditional mean value of iamong units with average NOx emissions rates and pro duction capacities. Of the interaction terms, only the quadratic emissions rate term is signi cant, suggesting a negative and convex relationship. Patterns of statistical signi cance, and most p oint estimates, are not signi cantly changed when p otentially endogenous wholesale electricity prices are used in the rst stage versus instruments (see Column 4). The last four columns are intended to address concerns ab out p otentially spurious correlation. Note that [13] mo dels the e/ect of a change in NOx prices on hourly capacity factor somewhat crudely as a vertical shift inf cf it4 3. This shift should only o ccur in hours when a unit is close to the margin. Otherwise, an incremental change in the p ermit price should have no e/ect on short-run pro duction decisions (see Figure 2). For many units in the sample, price-cost margins are far b elow or ab ove zero in ma jority of hours. In these cases, we should exp ect to nd a very weak or non-existent relationship b etween p ermit price variation and unit-level capacity factor (averaged across all hours). Inclusion of these non-marginal observations biases estimates toward zero. If unit-level price-cost margins are correlated with emissions rates or implicit subsidies, this would confound the interpretation of the estimation results. I take two complementary approaches to assessing this issue. First, unit-sp eci c averages of squared price-cost margins are added to the mo del (columns 5 and 6). Estimated co e¢ cients 4 2Se veral alte rn ative s p ec i cation s we re tried but n one imp roved the t of th e mo de l. For example, NOx emiss ion s rate s and implicit sub sidies were also interac te d with heat rate s an d S O2 e mis sions rates. 4 3The corre lation c o e ¢ cient b etween un it-leve l NOx rate s an d a unit-sp eci c meas ure of squared hourly pric e cos t margin s is 0.55. Th e corre lation co e¢ c ie nt b e tween the implicit sub sidy measu re an d the pric e cost margin meas ure is 0.38. 30 are not signi cantly a/ected. Second, the mo del is re-estimated using a data set that omits observations asso ciated with price cost margins that are large in absolute value (i.e. greater than $50/MWh or less than -$50). These estimates are rep orted in columns (7) and (8). As exp ected, all co e¢ cient estimates are larger in absolute value. Taken together, these results suggest that the negative NOx emissions rate co e¢ cient and the p ositive subsidy co e¢ cient are not a result of spurious cross-sectional variation in price-cost margins. 4 4To put these results in context, consider a fairly small change in the NOx price of $100/ton (the average NOx price over this p erio d is $2837/ton; the standard deviation is $1437). For a unit with an average NOx emissions rate in a grandfathering regime, this NOx price change would increase ozone-season op erating costs by approximately $0.22/MWh.For a unit of average op erating capacity op erating in a grandfathering regime, this incremental NOx price change is asso ciated with a decrease in op erating capacity of 7% . If this unit is in an up dating regime and receives 2.9 lbs worth of p ermits p er MWh generated (the average subsidy observed in the data), this decrease falls to 3%. Using the more limited data set that omits hourly observations that are far from the margin, these estimated decreases are on the order of 10% and 5%, resp ectively. sIn sum, the statistical patterns found in these data are generally consistent with standard theory. The estimated NOx emissions rate co e¢ cient is negative and consistently statistically signi cant across all sp eci cations. This is consistent with the hyp othesis that plant managers will account for environmental compliance costs in their short-run pro duction decisions. The coe¢ cient on the implicit pro duction subsidy is p ositive, somewhat smaller in absolute value, and less precisely estimated. The statistical signi cance of the eco e¢ cient is not robust to all sp eci cations. These results o/er relatively weak supp ort for the hyp othesis that the implicit subsidy conferred by allo cation up dating will mitigate the negative e/ects of environmental compliance costs on short-run pro duction decisions. Finally, whereas the value of svaries systematically with unit capacity and emissions rate, attempts to explain variation in the co e¢ cient are largely unsuccessful. 4 4This . 31 in crease is very small relative to average variable op erating c osts wh ich exc eed $80/MWh 6.2 Estimating the reduced form mo del of the participation decision To estimate the reduced form mo del of the participation decision, all unit-hours in which a unit b egins an hour with a capacity factor ab ove zero are dropp ed from the data set. This amounts to omitting more than 80 p ercent of observations at coal- red units, approximately 25 p ercent of observations at natural gas red units, and 14 p ercent of hourly observations at oil- red units. Electricity generating units in the study sample vary signi cantly in terms of the pro duction technologies they use and the wholesale electricity markets they serve. This variation presumably b egets variation in residual variances and the ijPparameters. The large quantity of data makes it p ossible, in principle, to consistently estimate the reduced form , ; and cco e¢ cients in [17]separately for each unit. Estimation in two stages is much easier to implement, although this relative simplicity comes at a cost of b eing unable to account for contemp oraneous correlation across the unit-level equations. In the rst stage, unit-sp eci c probit equations are estimated: j yij = 1f i + Ht=0 P Pt c icij 1i D _O + ijg : X it Zj In the second stage, unit-sp eci c estimates of 1 iare regressed on a constant, unit-level emissions rates (interacted with an NBP indicator variable) and estimated implicit subsidies. An FGLS estimator is used to incorp orate information ab out the variance structure obtained in the rst stage. Interpretation of these second stage estimates is complicated by two p otential identi cation problems. The rst is inherent in all discrete choice mo dels: parameters are identi ed only up to a scale factor (Maddala 1983). Consequently, comparisons of co e¢ cient estimates across units confound the magnitude of the regression co e¢ cients with residual variation. The second identi cation problem p ertains to the structure of the reduced form estimating equation. Permit price co e¢ cients are confounded with the ijparameters:4 5If resp onsiveness to p ermit allo cation in4 5The ratio of the marginal cost an d NOx p e rmit price c o e ¢ cient is id enti ed in prin ciple; s cale parameters an d parameters canc el ou t. One ap proach wou ld involve u sing th is ratio as th e dep end ent variab le in th e sec ond s tage in orde r to overc ome th ese identi cation p roblems . Howe ver, th is is di¢ cu lt to imp le ment in p ractice . Bec ause the distrib ution of b oth c o e ¢ cients inclu de z ero, s ome e stim ate s of th is ratio can b e in n itely large. 32 centives varies across units with di/erent emissions rates and/or di/erent implicit subsidies in ways that are correlated with the unobserved residual variances from the rst stage and/or the ijparameters, co e¢ cient estimates may b e biased. In this section, I present some alternative sp eci cations and auxiliary tests which are generally supp ortive of the identifying assumptions. 4 6Table 6B summarizes the estimation results from the rst stage. These estimates are generated using a sp eci cation that assumes a time horizon of six hours.Hourly demand forecasts are used as instruments for real time and day ahead electricity prices. As exp ected, estimated marginal cost co e¢ cients are negative on average. Estimated wholesale electricity price co e¢ cients are of the same order of magnitude and are p ositive on average. The average NOx price co e¢ cient estimate is also negative and highly variable across units. 6.2.1 Results Estimation results from the second stage are presented in Table 8. The dep endent variable is the estimated 1 ico e¢ cients from the rst stage. The most restrictive sp eci cation includes only a constant and demeaned NOx emissions rate. Here again, the constant term is not statistically signi cant. This implies that an incremental change in the NOx price has no statistically signi cant e/ect on the latent value y as de ned in [16] among units with average NOx emissions rates. The NOx emissions rate co e¢ cient is negative and statistically signi cant, as exp ected. This implies that the e/ect of a change in the p ermit price on the latent y value is signi cantly more (less) negative among units with relatively high (low) emissions rates. A less restrictive sp eci cation (2) allows the NOx price co e¢ cient to vary with the implicit subsidy intro duced by contingent allo cation up dating regimes. The emissions rate co e¢ cient increases slightly in absolute value and remains highly statistically signi cant. The co e¢ cient on the subsidy variable is p ositive as exp ected, but imprecisely estimated and not statistically signi cant at a 10 p ercent con dence level. 4 6The choice of a six h our time horiz on was inf ormed by an in form al analys is of ramping con straints . Un it-s p ec i c ram p rates were c ru dely es timated by lo okin g at hou r-to-h our changes in pro d uction, e xp re ssing th ese change s as a p ercentage of installe d capac ity, an d d e n in g th e ramp in g c onstraint to b e th e 95th p erc entile hou rly chan ge. The mean ramp ing constraint is 50%, m eaning that th e average p lant is c ap able of increasin g outp ut by half its installed c apacity over the spac e of one h our. The vast ma jority of estimate d ramp rates excee d 16 p e rc ent. 33 Several alternative sp eci cations were estimated so as to allow the NOx rate and subsidy co e¢ cients to vary with observable unit-level characteristics. One of the sp eci cations that b est t the data allows these co e¢ cients to vary with installed capacity (column 3). Results indicate that pro duction decisions of larger emitting pro ducers are more resp onsive to a change in emissions p ermit prices (and thus compliance costs) as compared to smaller pro ducers. With the addition of these interaction terms, the subsidy e/ect b ecomes statistically signi cant at a 5 p ercent con dence level. This suggests that larger pro ducers are also more likely to resp ond to the implicit subsidy. Columns (4), (5), and (6) explore the robustness of these results to varying assumptions ab out the endogeneity of rst stage covariates and the assumed time horizon. Column (4) rep orts results obtained when electricity prices are assumed to b e exogenous in the rst stage. Columns (5) and (6) rep ort results under varying assumptions ab out the time horizon. The NOx rate co e¢ cient is negative and highly statistically signi cant across all sp eci cations. The statistical signi cance of the implicit subsidy and the interaction terms are somewhat less robust to these sp eci cation changes. In order to use these results as a basis for causal inference regarding the impacts of unit-level emissions rates and subsidies on the relationship b etween p ermit prices and short-run participation decisions, I intro duce some additional assumptions. I assume that the residual variance and the ij4 7parameters are distributed indep endently of p ermit allo cation incentives. If this is not the case, strong correlations b etween NOx price co e¢ cients estimated in the rst stage and emissions rates or implicit subsidies could b e spurious.To investigate this p ossibility, additional sp eci cations are estimated. Although the ijparameters and residual variances are not observable, variables that are plausibly strongly correlated with these factors are observed in the data. For example, fast ramp rates should b e asso ciated with smaller parameters. Larger op erating capacities should b e asso ciated with larger ijijparameters. Columns (7) and (8) constitute indirect tests of these indep endence assumptions. Ramp rates parameters, are added to and unit capacity, two variables plausibly correlated with the ij the 4 7For e xamp le , if the ij 34 parameters are system atically smalle r among units with h igh er e mis sions rates , th is wou ld resu lt in a n egative NOx emis sion s rate c o e ¢ cient in th e s econ d s tage that has noth in g to do with a cau sal relations hip b e twee n h igh emiss ion s rate s an d the nature of pro duc ers resp on se to chan gin g NOx p ermit pric es. mo del in (7). Although b oth have the exp ected sign, only the ramp rate co e¢ cient is statistically signi cant. The other co e¢ cient estimates are not substantially a/ected. Residual variances could b e exp ected to vary across regional electricity markets with di/erent dispatch proto cols and pro cedures. Indicator variables for the New York and New England market are added to the mo del in (8). The New York co e¢ cient is p ositively and weakly signi cant. This could b e a result of smaller residual variance among units supplying New York, or it could b e capturing the average e/ect of the state s input-based allo cation up dating regime. The other co e¢ cient estimates are not signi cantly a/ected. Taken together, these results provide weak supp ort for the aforementioned indep endence assumptions. Conditional on my additional identifying assumptions, the estimation results suggest that electricity suppliers do resp ond to b oth explicit compliance costs and implicit pro duction subsidies asso ciated with the intro duction of this cap-and-trade program. The NOx emissions rate co e¢ cient is negative and strongly statistically signi cant across all sp eci cations. The p oint estimate of the e/ect of the implicit subsidy is consistently p ositive and smaller in absolute value. Taken at face value, these p oint estimates are roughly consistent with a discount rate of 10-15 p ercent (p ermit allo cations are up dated with a three to four year lag on average). However, based on these fairly noisy estimates, the hyp othesis that plant managers ascrib e equal weight to the implicit subsidy and the explicit compliance costs in their short-run pro duction decisions cannot b e rejected. 7 Conclusions Policymakers, industry representatives, and other stakeholders are increasingly interested in understanding how the design of p ermit allo cation proto cols a/ects p ermit and pro duct market outcomes. A growing theoretical literature o/ers insights into how rms should resp ond to di/erent p ermit allo cation rules, including "up dating" regimes that simultaneously p enalize emissions while rewarding pro duction. Empirical evidence has thus far proved elusive. This pap er analyzes the pro duction decisions of hundreds of electricity generators in a large emissions trading program that is particularly well suited to a study of the short-run impacts of p ermit allo cation design. A simple partial equilibrium mo del demonstrates the theoretical, short-run implications of 35 contingent up dating vis a vis more traditional p ermit allo cation designs (i.e. auctioning and grandfathering) and serves to motivate the empirical work. Two complementary empirical strategies are used to assess whether the detailed hourly data collected from electricity generators in the NOx Budget Program are consistent with standard theory. The rst investigates statistical relationships in the data. I nd that an increase in the emissions p ermit price is asso ciated with larger (more negative) changes in the short-run pro duction decisions of units with higher emissions rates. I also o/er somewhat weaker evidence that the relationship b etween NOx price changes and short-run supply decisions is less negative among units that implicitly receive a pro duction subsidy via p ermit allo cation up dating. These results are generally consistent with our theoretical understanding of how p ermit allo cation incentives a/ect market outcomes. The second strategy imp oses more structure so as to provide a basis for causal inference. I fo cus exclusively on the participation margin of the unit commitment pro cess that presumably generates the observed short-run supply decisions. I derive and estimate a reduced form mo del of electricity pro ducers decisions to resume op erations conditional on b eing inactive. Estimation results provide further, robust evidence in supp ort of the hyp othesis that pro ducers account for b oth the explicit emissions tax and the implicit pro duction subsidy conferred by this emissions trading program. I fail reject the hyp othesis that negative emissions incentives and p ositive production incentives conferred by allo cation up dating are weighed equally in short-run pro duction decisions. This analysis is not without its limitations. Precise estimation of the welfare implications of the p ermit allo cation design decisions I analyze is b eyond the scop e of this pap er. A burgeoning literature endeavors to predict welfare impacts of p ermit allo cation design choices in a range of counterfactual p olicy contexts using detailed, theory-based simulation mo dels. Although the results of these studies cannot b e con rmed or refuted based on evidence presented here, the empirical ndings lend supp ort to some pivotal assumptions underpinning these simulation exercises. Finally, an imp ortant motivation for this study is to inform future p ermit market design decisions, particularly those that will determine how hundreds of billions of dollars worth of carb on 36 credits will b e allo cated under prop osed Federal climate change legislation. The pap er provides empirical evidence that rms are resp onding to the relatively mo dest p ermit allo cation incentives conferred by the NOx Budget Program. We might exp ect this resp onse to carry over to a larger greenhouse gas emissions trading program where the stakes will likely b e much higher. However, the regional emissions trading program analyzed in this pap er and prop osed Federal program designed to limit greenhouse gas emissions di/er along a numb er of imp ortant design dimensions. Additional research is needed b efore we can de nitively anticipate how rms will resp ond to p ermit allo cation incentives in the next generation of cap-and-trade programs. References Bernard, A. L., Fischer, C. & Fox, A. K. (2007), Is there a rationale for output-based rebating of environmental levies? , Resource and Energy Economics 29(2), 83 101. Bohringer, C. & Lange, A. (2005), Economic implications of alternative allo cation schemes for emission allowances , The Scandinavian Journal of Economics 107(3), 563 581. Borenstein, S., Bushnell, J. B. & Wolak, F. A. (2002), Measuring market ine¢ ciencies in california s restructured wholesale electricity market , American Economic Review 92(5), 1376 1405. Bunn, D. & Fezzi, C. (2007), Interaction of Europ ean Carb on Trading and Energy Prices , SSRN eLibrary . Burtraw, D., Palmer, K., Bharvirkar, R. & Paul, A. (2002), The e/ect on asset values of the allo cation of carb on dioxide emission allowances , Electricity Journal pp. 51 62. Burtraw, D., Palmer, K. & Kahn, D. (2005a), Allo cation of CO2 emissions allowances in the regional greenhouse gas cap-and-trade program, Discussion Pap ers dp-05-25, Resources For the Future. Burtraw, D., Palmer, K. & Kahn, D. (2005b), CO2 allowance allo cation in the regional greenhouse gas initiative and the e/ect on electricity investors, Discussion Pap ers dp-05-55, Resources For the Future. 37 Bushnell, J. & Chen, Y. (2009), Regulation, allo cation, and leakage in cap-and-trade markets for co2, Technical rep ort, EI@Haas WP-183, Washington, DC. Commission, C. P. U. & the California Energy Commission (2008), Joint california public utilities commission and california energy commission sta/ pap er on options forallo cation of ghg allowances in the electricity sector, Technical rep ort, R.06-04-009 and D.07-OI IP-01. Cramton, P. & Kerr, S. (2002), Tradeable carb on p ermit auctions: How and why to auction not grandfather , Energy Policy 30(4), 333 345. Dinan, T. & Rogers, D. (2002), Distributional e/ects of carb on allowance trading: How government decisions determine winners and losers , National Tax Journal 2, 199 221. Duxbury, D. & Summers, B. (2004), Financial risk p erception: Are individuals variance averse or loss averse? , Economics Letters 84(1), 21 28. Environmental Protection Agency, U. (1999), Economic analysis of alternate metho ds of allo cating nox emission allowances, Technical rep ort, ICF Consulting. Fell, H. (2009), Eu-ets and nordic electricity: A cvar analysis , Energy Journal . Fiegenbaum, A. (1990), Prosp ect theory and the risk-return asso ciation : An empirical examination in 85 industries , Journal of Economic Behavior & Organization 14(2), 187 203. Fischer, C. (2003), Output-based allo cation of environmental p olicy revenues and imp erfect comp etition, Discussion Pap ers dp-02-60, Resources For the Future. Fischer, C. & Fox, A. K. (2007), Output-based allo cation of emissions p ermits for mitigating tax and trade interactions , Land Economics 83(4), 575 599. Gersbach, H. & Requate, T. (2004), Emission taxes and optimal refunding schemes , Journal of Public Economics 88(3-4), 713 725. Goulder, L. H., Parry, I. W. H., Williams I I I, R. C. & Burtraw, D. (1999), The cost-e/ectiveness of alternative instruments for environmental protection in a second-b est setting , Journal of Public Economics 72(3), 329 360. Hanushek, E. A. (1974), E¢ cient estimators for regressing regression co e¢ cients , The American Statistician 28(2), 66 67. 38 Harvey, S. & Hogan, W. (2002), Market p ower and market simulations, Mimeo, Center for Business and Government, John F. Kennedy Scho ol of Government, Harvard University. Hirshleifer, D. (2001), Investor psychology and asset pricing , Journal of Finance 56(4), 1533 1597. Jensen, J. & Rasmussen, T. N. (2000), Allo cation of co2 emissions p ermits: A general equilibrium analysis of p olicy instruments , Journal of Environmental Economics and Management 40(2), 111 136. Maddala, G. (1983), Limited-Dependent and Qualitative Variables in Econometrics, Cambridge University Press, New York. Montgomery, D. W. (1972), Markets in licenses and e¢ cient p ollution control programs , Journal of Economic Theory (3), 395 418. National Commission on Energy Policy (2007), Allo cating allowances in a greenhouse gas trading system, Sta/ pap er, NCEP. Neuho/, K., Martinez, K. K. & Sato, M. (2006), Allo cation, incentives and distortions: the impact of eu ets emissions allowance allo cations to the electricity sector , Climate Policy 6(1), 73 91. Quirion, P. & Demailly, D. (2006), CO2 abatement, comp etitiveness and leakage in the Europ ean cement industry under the EU ETS: grandfathering versus output-based allo cation , Climate Policy 6, 93 113. Sarin, R. K. & Web er, M. (1993), E/ects of ambiguity in market exp eriments , Manage. Sci. 39(5), 602 615. Sijm, J., Neuho/, K. & Chen, Y. (2006), Co2 cost pass through and windfall pro ts in the p ower sector, Cambridge Working Pap ers in Economics 0639, Faculty of Economics, University of Cambridge. Sterner, T. & Muller, A. (2008), Output and abatement e/ects of allo cation readjustment in p ermit trade , Climatic Change 86(1-2), 33 49. U.S. Environmental Protection Agency (2009), The United States Environmental Protection Agency s Preliminary Analysis of the Waxman-Markey Discussion Draft in the 111th 39 Congress, The American Clean Energy and Security Act of 2009, Technical rep ort, http://www.epa.gov/climatechange/economics/economicanalyses.html. Wolfram, C. D. (1999), Measuring duop oly p ower in the british electricity sp ot market , American Economic Review 89(4), 805 826. Zachmann, G. & von Hirschhausen, C. (2008), First evidence of asymmetric cost pass-through of eu emissions allowances: Examining wholesale electricity prices in germany , Economics Letters 99(3), 465 469. 40 Fig ure 1 : Emi ssions constrained social welfa r e maxi mizing outc ome Notes : The downward sloping solid li ne represents the emissions constraint. The soci ally optimal allocation of production o ccurs at the point where the emissions constraint is just tan gen t to a leve l set of the economic surplus function. This point is intersecte d by the vector from the origin. The broken line , with a slope of -1, conne cts all poin ts that correspo nd to an aggregate output qua ntity equal to that associated with th e optimum outco me. Points l y ing on the emissions constraint above (belo w ) th e optimal point are associated with more (less) consu mption and more (less) supply side abatement tha n is consisten t with w elfare maximi zation . 41 Figure 2 : Hourly Producti on 0 .2 .4 .6 .8 1 0 20 40 60 80 0 12 24 36 48 60 72 84 96hour Decisi ons wholesale electricity price 0 .2 .4 .6 .8 1 -20 -10 0 10 20 30price-cost margin at a Repr ese nta tive Unit capacity factor capacity factor Notes : Hourl y production decisions at a single uni t (measured as capacity factor) and th e corresponding hourly wholesale electrici t y price over a four day peri od are plotte d in the left panel. The horizontal li ne represents the theoret ical reservation price duri ng these off - season hours (i.e. th e unit’s constan t marginal ope r ating cost). This cost of $ 49/MWh is e s timated using th e fuel input pri ces that prevailed in this four day period, the unit- spe cific heat rate, and other variable (no n - fuel) operations costs. The thin black li ne in the right pan el plots these same data in capa city fa ctor, price- co st margin (i. e . the hourly wholesale electri city pri c e less the marginal operating costs incurred) space. This is a mean smooth of c apacity factor on price - cost margins. The thi ck bla ck line represents the on off producti on protocol implie d by th e benchmark model of a profit maxi mi zing, pri ce taking produce r . Comparing these two functions he lps to illustrate how observed production decisions deviate systematica lly from the predictions of the simple, static model. Finally, the thick red liner epresent th e expected impact of the NBP on thes e short- run supply f unctio ns. In theory, the uni t’s reservation pri ce should increase by an amount equa l to the net compliance co sts per MWh . 42 Capacity season factor Ozon Grandf ather i ng Regimes Contingent Upd at i ng Regimes Figure 3 : Electricity Supply Decisi ons in Grandf ath e ring and Contingent All ocation Updati ng Regimes Notes : Th ese figures plo t a local mea n smooth of hourly, unit - level capaci ty factors on hourly, unit - specific pri ce - cost margins , where costs include fuel costs and other variable operating costs but excl ude any costs or implici t subsi dies associated with the NOx Bu dget Program. The shaded regions represent 95 percent confidence intervals. The se graphs are ge n e r a te d usi ng hourly from almost 600 electrici ty generati ng units over the study time period (2003- 2 006), excludi ng the top and bottom fiv e per centiles of observations. I use an Epane chni kov weight function and rule- of- thu m b b a ndwidth estimator. The right panel s ummarizes the produ ctio n de cisions at 73 units operating in states where NOx permits are grandfathere d. The lef t panel summar i zes producti on decisions at 515 units operating in states where permit alloca tions are peri odically u pda ted based on lagged inpu t or output choices. -150 -100 -50 0 50 100Price-cost margin ($/MWh) 43 ($/MWh) Off season 0 .2 .4 .6 .8 0 .2 .4 .6 .8 -150 -100 -50 0 50 100Price-cost margin Off season95% CI Ozone season 95% CI Ozone season Table 1 : Permit allocati on regim e chosen by New York, New Eng l and , and Mi d- Atl antic states State Annu al state NOx bu Chosen perm Electricity dget for Electricity ge nerating uni it allocation regime market ts (tons NOx) CT 4,253 Outp ut - based up datin g MA 12,861 Outp ut - based up datin N g ME N/A N/A NH N/A N/A RI 936 Grandfatheri ng VT N/A N/A e New York NY 30,405 Inpu t - based updati ng w PJM DC 233 Grandfatheri ng DE 4,463 Grandfatheri ng MD 14,520 Grandfatheri ng NJ 8,200 Outp ut - based up datin g E PA 47,244 Inpu t - based updati ng VA 17,091 Inpu t - based updati ng n Notes: Annual, state- level NOx b udge ts do not chan ge over the s tudy . Amongg states that chose contin gen t u pdatin l g, implementation details vary considerably. For example, whereas so a me states update annually, others u pdate in thre e or four year blocks. n d 44 period Table 2 : Operating summ ary statistics by per m it all ocati on regi me : CENSUS # Units Summer Off - seas Hea Allocati on Capacit on capacit t rate* (b y y tu/kW h) regime (MW) factor 11,384 ( 2,444) 353 155 (206) Outp ut - based updati ng Grandfatheri ng Exe m pt 140 134 (176) 69 200 (182) 18 181 (234) 0 0 5 $ 0 3 $ 2 2 7 4 3 2 2 0 0 $ 1 8 3 12,080 ( 3,447) 3.07 ( 3.60) 1.57 (4.67) 19% (26%) 10% (18%) 21% (26%) 23% (29%) $ 5 2 0 0 6 8 4 3 6 8 2 2 0 0 4 $ 2 1 8 0 5 $ 2 $ 2 7 7 1 $ 2 9 5 1 2 2 0 0 2.01 ( 2.04) 2.82 ( 4.54) Notes: Standard deviations in pare nth e ses. Su mmar y statistics are gen erated using da ta fr om 610 fossil fuel - fired ele ctricity gener a ting units su pplying the New York, New En gland, or PJM mark ets durin g the study period (2003- 2006). Self ge nerati ng and co - ge nerating units are exclude d from the sample. * Emissions rate and hea t rate summary statistics are weighted by installed ca pacity. Table 3 : NOx Allowance Prices 2003- 2 006 (Nominal $/ton) Permi t vintage Transac tion year 2003 2004 2005* 2006* 2 - NOx rate* (lbs NOx/MWh) 12,857 ( 3,090) 11,5 92 (5,388) Inpu t - based updati ng Ozone season $ 1 9 0 6 $ 3 1 6 3 0 $ 1 2 0 4 $ 1 5 0 7 $ 3 0 1 7 $ 2 6 6 2 $ 1 7 5 0 2 0 0 8 $ 2 7 0 5 $ 2 2 9 9 $ 1 5 7 0 2 0 0 9 $ 2 3 1 4 $ 2 2 3 2 $ 1 5 1 8 Notes: This table reports average an nual permit prices by NOx permi t vintage. Conte mporaneous permit pri ces appear in bold. Th e asterisk denotes years in which the progressive flow control (P FC ) constraint wa s bindin g. The PFC ratio was 0.25 and 0.27 in 2005 and 2006, respec tively; ban ked permits are traded at a discoun t in these years. 45 5 7 0 5 $ Table 4 : Es ti mate d NBP compliance c o sts an d prod uction incentives by alloc a tion regim e (1) (2) (3) (4) NOx permi (1) as a perce ntag Future permi Estim ated ne t costs per M e of off ts allocated (tot compliance co Wh season ns) pest Allocati r MWh genera generated variable oper per MWh* on regime ted ating costs Inpu t based updati ng Out p ut bas ed upd ati ng Grandfatheri ng $6.49 ($9.03) $4.89 ( $4.69) $ 7 . 4 6 6.4% ( 6.0%) 6.5% ( 5.9%) 0.001 ( 0.001) 0.002 (0.0 01) $1.40 ( $4.69) $1.45 ($8.15) ( $ 8 . 2 4 ) 8.1% (6.8%) 0 $6.49 ($4.01) Notes: Stand a rd deviations in parenthe ses. Summar y statistics are generate d using data fr om 580 fossil fuel fired ele ctricity gen e r a ting uni ts supplying the New York, New England, or PJM markets duri ng the study period (2003- 2006). Self g e nera ting and co generati ng units are e xcl uded from the sample . Averages across unit - year s are reported; standard deviations are in paren thes es. * To calculate net complia nce costs, future permi ts allocated per unit of output are valued using futures perm it prices. This value is then subtra cted from the expli cit compliance cost per MWh (i .e. the prod uc t of the specific emissions rate and the spot NO x permit pri ce). Exe m pt $0 -- 0 $0 46 uni t - Table 5 : Regi onal whol esa l e elec tricity prices ectricity dem and by seas on Ho urly el ectr icity dem an d and el Wholesale el ectricity pric e ($/MWh) Regio nal (MW) el ectricity mar ket Off - seas on New Engla nd (NEPOOL) New York (NYISO) Ozone season Off - seas on 14,603 ( 15,344 ( 2,513) 3,416) 17,547 ( 2,768) 19,151 (3,936) $60.88 ($ 29.34) Ozone seaso n $57.98 ($ 34.88) $68.00 ( $96.97) $52.03 ( $33.84) Mid- Atl antic (PJM) 29,214 (8,787) 31,694 (9,957) Notes: Standard deviations in pare nth e ses. Th is tabl e summarizes observed pri ces and load levels over the 32,748 hours in the data. Both ad levels are simi larly distribu ted across seasons. 47 prices and lo $70.74 ( $78.17) $55.41 ($43.34) Table 6A : Summary of first stag e es ti mate s: Dependent variable is hourly capacity factor Covariate Average poi Standard Average standard e n t estim rr or ate deviation - 9.24 12.93 4.18 U n i t l e v e l c o n s t a n t 2 5 . 2 4 9 3 . 4 9 p e r a t i n g c o s t ( $ / M W h ) 3.04 17.76 9.39 Local whol esale marginal price ($/MWh) 6 2 . 6 6 M a r g i n a l - 0.003 0.25 0.0008 15,087 NOx price * Ozone indicator ($/to n) Average # obs e rvati ons /unit Note s: The unit of analysis is a unit -hour. The dependent va r i a b l e is the cap a c i ty factor (measu red on a scale of 1 - 100). For ea se of exposition, only th e higher order cost and price variable s, lagged prices, futu res prices, and year dummies are omitted from this table. Table 6B : Summary statis tics for the fi rst stag e e s ti mate s Note s: The unit of analysis is a unit -hour. The dependent va r i a b l e is the bin a ry participation indicator. Each unit - level regression includes: a un it -level fixed effect, marginal operating costs, contemporaneous wh olesale price , hourly forward prices for 1 through 6 hours forward, the NOx permit pric e interacted with the ozone season indicator and year fixed effects. Real time and day d electricity prices are instrumented for using hou r ly demand forecasts. 48 Covariate - 2.98 1.78 1.69 l e v e l Marginal ope r ati n g cost ($/MWh) Local whol esale marginal price ($/MWh) NOx price * Ozone indicator ($/to n) Average # obse rvati ons /unit Average poi n t estim ate c o Standard Average rr or deviation U n i n s t a n - 0.14 0.03 0.02 0.06 0.09 0.05 - 0.0004 0.018 0.0001 13,945 t t - standard e ah e a o Table 7 : Des c riptive model : Second stage es tima ti on resul ts Specification (1) (2) (3) (4) (5) (6) (7) (8) Constant - 0.00 (0.02) - 0.02 (0.02) - 0.05 (0.03) - 0.03* (0.02) 0.03 (0.02) - 0.07** (0.03) - 0.05 0.03 - 0.10*** (0.03) NOx ra te 14.56** 16.27*** 30.02*** 28.46*** 15.68*** 31.35*** 35.83*** 61.87*** (5.65) (5.59) (10.90) (7.24) (5.87) (11.75) (12.87) (17.84) 10.82 (9.84) 23.42** 17.06** 13.41 26.00** 27.14 31.52* Estimated (11.00 (7.96) (9. (11.46 (17. subsidy ) 54) ) (18.72) 45) Pri ce cos t margi n 2 (NOx 0 0 * d rate 1 rate 3 1 727.62) Estimate . 2 5 7 0 0.09* ( (0.05) 0 0 2 7 . 1 7 * 1413.27*** * 2225.61* . 9 ) . 8 0 0 . 8 ( 550.55 (1359.16) First stage V Y N Y . 0 5 5 - ) ( 0.06 (0.06) 0.05 553 N 0 1297.7 1 8 635.06 (1 - 0 . 0 0.05 553 0.03 553 Y N Y N N N 0 . - 5 0 8 0 . ( 2 ) 0.07 (0.09) 0.09 553 0.02 550 R N . 0 (0.001) 938.23 (7718.3 6) 0 8 0.02 (0.05) 0.02 553 0 5 8 ) 0.02 553 ( 0 . 0.0001 (0.0003) - 0.05 (0.03) 0.001 2 - subsidy NOx rat e * ca pacity Estimated subsid y * c apacity 2 0 ) 1 7 9 8 2 5 . 6 467.35 2033.56 (887.56) ) NOx 0 3 N N Y 0.03 550 Y Y Y I Re s t ricte d sample? Note s: The unit of analysis is an electricity gen e rating un it. The depend ent variable is the estimated NO x price effect (during ozone season ) from the unit -le v e l linear regression estimation. In the first stage, forecast demand is used to instrument for loc al marginal electricity prices un less otherwise indicated. Emissions rates (d emean e d) and estimated subsidies are measured in tons of NO x per MWH. Production capacity (also demeaned ) is measured in MW. Th e price-cost margin va r i a b l e is constructed by computing th e average squared pric e -cost margin for each unit (across all hours). FGLS standard errors (in parentheses) are clustere d at the facility le v e l . To estimate specifications (7 ) an d (8 ), all un it -hours in wh i c h the estimated price margin is greater th a n $50 or less than -$5 0 are dropped in the first stage. See text for details. * Statistically significant at the 10 per c ent level. ** Statistically significant at th e 5 perc 0 * 0 * ent level. 49 *** Statistically significant at the 1 perc ent le v e l . Table 8 : Ho w do NOx permit prices on resul ts af fect the participation deci sion? Second stage es tim ati Specification (1) (2) (3) (4) (5) (6) (7) (8) Constant 0.000 (0.000) - 0.003 (0.002) - 0.004*** (0.001) 0.002 (0.001) - 0.002 (0.002) - 0.002 (0.002) 0.002 (0.002) - 0.005* (0.001) NOx rate (tons / 2.21*** 2.44*** 3.04*** 2.05*** 2.87*** 2.97*** 2.73*** 3.09*** MWh) (0.80) (0.82) (0.44) (0.49) (0.54) (0.65) (0.52) (0.45) 1.69 (1.23) 1.92*** 1.45* 1.48* 1.47 1.69** 1.65** (0. (0. (0.99) (0.71 (0.75 Estimated (0.70) 76) 84) ) ) subsidy (tons / MWh) NOx rat e * ca pacity * * 2 ) 0.014*** (0.01) 0 . 0 Estimated subsid y * c apacity 0 1 ( 0 . 0 0 4 0.004) 0.004 (0.004) R a 0 1 C . ) m 0 . 0 0 ( 0 . 0 - 0.002 (0.003) - 0.003 8 0 (0.004) - * 0.010*** ) 0.007* p ( r 0.007* (0.00 4) 0.003 0.008** (0.005) (0.003) (0.003) 0.001 (0.005) a t e - 0 a 0 p 0 N 3 a . * c 0 * i t 0 . 0 . 0 0 0 y 0 0 E ( N . ( 0 . 0 0 . 0 0 2 0 1 ) Y 0 2R 0 0 0.09 . First stage 0 1 0.08 553 6 IV? * ( 0.10 0 0.06 0.12 0.12 553 550 550 550 550 6 3 ) 0.11 0.10 550 N Assume d time horiz on (hours ) 0 2 550 6 2 8 6 6 6 Y Y Y N Y Y Y Notes: The unit of analysis is an electric ity generating unit. The de pendent variable is the coefficient on the inte raction between the NOx permit price and ozone season indicator from irst stage of the estimation. First stage probit equations also include a nt, unit- specific marginal op erating cos t , year indicato rs, and instrume nts for current and future wholesale electric the f consta ity prices (unless otherwise indicated). These first stage equations are estimate d under different assumptions about the relevant time horizon (me a s u re d in hours). Emissions rates (demea ned) and estimated subsidie s are measured in to n s of NOx per MWH. Prod uction capaci ty (als o demeaned) is measured in MW. FGLS standa rd errors (in parentheses) are clus te re d at the facili ty le ve l. See text for de tails * Statistically ent level. 50 *** significant Statistically at the 10 perc ent level. significant at the 1 ** Statistically significant perc ent le v e l . at th e 5 perc A p p e n di x 1 U si n g th e sy st e m of rs t or d er c o n di ti o n s, s ol v e fo r e q ui li br iu m q u a nt iti e s a n d p er m it pr ic e in te r m s of th e m o d el p ar a m et er s: +e c + c ae + e q c = bE ecbe2 c 2+ E dec d q d = bE edbe2 c d d + e2 cced d aece+ be + e2 dcc + be2 d 2 cc +Ee + e2 dcaeced2 + e2 d dcc 2+ d d be 2 dd c 2 c c d c c + ae (2) ( c+ 3) Q dced 2ae =E eccb d bE e 2bec ed2 c 2bedc bE e 2beeccec c) 2 + e+ E edcc + ae2dc+ d ae2 d (1) 2 dcc e c) cc E (bcc2 c+ bcd + +e cc (4) cccd d + e2 ccd d= d + ) b (e e c Substituting these expressions into the economic surplus function and di/ere c ect to the emissions a constraint: c ) ( e c c @ S@ E= y equation [??] , b E ( c ) b (e ed2)+ e2 d + a +e = : (5) ccc (eccd is also B the marginal cost of reducing emissions on the supply side. Thus, in the rst b est case, marginal abatement costs are set equal across all pro ducers and across the supply and demand side of the pro duct market. App endix 2 : Pro t maximization under symmetric up dating yields lower consumer prices and higher supply side abatement costs as compared to the rst b est case. To see this, I solve for the equilibrium quantities and p ermit price in terms of the mo del parameters: 51 =E eccbed2 c S_U PQ 2be + be2 d 2b ec ce d d d 2aeccc 2 2 cc d dc c + + e2 + e2 be2 d ccd dcc ced seced + E eccd+ ae2 d + + ae se2 d c d seced + E 2 c + c 2c c edcc se2 c d e = bE ec secc bE2be se be2 = bE ed ed ae aed c qS _ U P 2be c q S_UP d + bcd 2 S_U P A ss u m in g th e i m pl ici ty pr o d u cti o n s u b sE c ced sE cc+de+ E edcc e + ae2 c+ +e ae2 d (7) e 1+ d be2 d 2+ e2 +e c c > (8) 1ccd bE e + cccd) + s (e 2) + e c + e2 2 2c1 = a (ecc + edd d ) E (bccc ) b (e e c c+ ecc d (9) si d y is st ri ctl y gr e at er th a n z er o, th e p er m it pr ic e u n d er u p d at in g (a n d th u s th e m ar gi n al a b at e m e nt c o st o n th e s u p pl y si d e) is hi g h er u n d er u p d at in g v er s u s th e b e n c h m ar k c a s e. It c a n al s o b e s h o w n th at c o n s u m pt io n le v el s wi ll b e hi g h er u n d er sy m m et ri c u p d at in g: 2c + E edcc 2aeced + ae2 c+ ae2 d s E c + E edcc 2ae 2be2)cceed (a (e c + ae2 c + sE cc (11) c d a e + + e c + e2 d dcc be2 d d cd 2 d 2 c dc c seccd sedcc 2 c be 2bece + be2 + c e c +e d 2d S_U Q PQ = s (e d e =E (10) E e eccbed2 c ) E (bcdcc + + bEc + acccd)) (12) d edcc @ S+ edcc) E (bcd + bEc d + acccd > 0): Thus, QO B A > Q : a ( eccd+ ed 52 d+ cc+ ccc Note that (a (eccd =)) >0 if the emi ssio ns con stra int bin ds in the @E dcc) c bE ( c)2+ e2 cc 2 dc c rs tb est cas e (b eca use ) b( e e+ ed A p p e n di x 3 Pr o t m a xi m iz at io n u n d er sy m m et ri c u p d at in g yi el d s lo w er c o n s u m er pr ic e s a n d hi g h er s u p pl y si d e a b at e m e nt c o st s a s c o m p ar e d to th e r st b e st c a s e. T o s e e th is, I s ol v e fo r th e e q ui li br iu m q u a nt iti e s a n d p er m it pr ic e in te r m s of th e m o d el p ar a m et er s: 2 + e2 + e2 dcc d ccd ce d d +E ecc = bE ec bE edbe2 c qA_ U P c s+ + e2 d ae2 d dsc aece + E ed 2be aecdce d + qA_ U P d = bE ed 2 2bece csd d b (ed b e be2 + d e2 ccd ed + c d bE ecbe2 c c (14 eced ) dsc a e 2+ d e2 dcc e e + e2 cs (15 ) +e c +e c d cdb d c (e+ edcc) E cd+ bcc 2c+d cccd) + becsc (bcd ec2)dsc+ be + e2 c + edccsd ccd (17 ) c c2 c cc = a (ecc (16) be be e2)dsd+ es+ ec A_ U P A_ U PQ = c + sd ; = E eccd + E edcc 2aebec2 + ae2 c+ ae2 d 2c ced+ be note that e2 dsc = ecedsd and e2 csd= ecedsc d+ e2 csd 2be d + e2 dsc ecedsc ecedsd (13 ) cedsd ec ed sc 2 dsc 2 c s d : Thus, if the pro duction subsidies are set such thatI the subsidy p er unit of emissions are f equal, the rst-b est pro duction quantities and pro duct price is achieved. Although the pro duct price in this case will b e equal to P;note that the p ermit price will exceed (to comp ensate for the relatively high subsidy paid to the more p olluting rm). If the subsidy p er unit of p ollution is 2+ d e+ e hig . Alternatively, if the subsidy p er unit of p ollution is higher for the more p olluting rm, the her reverse will b e true. for cle App endix 4 :Data set construction an er rm State-level permit al location regimes Permit allo cation design parameters were collected from s state-level implementing agencies. All states have do cumented the sp eci c algorithms and (as equations used to determine facility-sp eci c p ermit allo cations in detail. Agencies were iscontacted directly when implementation details were unclear or could not b e found in the the public record. ca se Unit-level operations and attributes wh en pro 1Hourly, unit-level emissions, heat input, and output data were obtained from the US EPA duContinuous Emission Monitoring System (CEMS).Continuous emission monitoring and rep ctiorting on 1Coal un its rep ort hourly gross an d n et u nit load (in MWe ). Combined -cycle un its are require d to re p ort su th e bsi 53 die s are sy m me tric ), e> e; e> ece dsc, an d co ns um pti on ex ce ed s rst b est lev els 2requirements for EGUs regulated under EPA s Acid Rain Program (ARP) and/or the NOx Budget Program require monitoring of hourly sulfur dioxide mass emissions, carb on dioxide mass emissions, NOx emission rates, heat input, and other op erating conditions. 3I exclude nuclear, hydro, and renewable electricity generators b ecause these zero emitting pro ducers are categorically excluded from the NOx Budget Program. Small pro ducers (i.e. less than 10 or 15 MW) are also excluded from the NBP. Finally, I exclude any units for which wholesale electricity generation is not the primary pro duction activity (such as self generating units). This leaves 580 electricity generating units in the cleaned data set and 15,455,067 hourly observations. Emissions permit prices NOx p ermits are actively traded in a liquid p ermit market.4Brokers facilitate the ma jority of arms length transactions, administer escrow accounts, and provide market analysis. A variety of transaction structures exist in the market, including forward settlements, calls, puts, comp osite structures such as straddles, and vintage swaps. Daily p ermit price data were purchased from Evolution Markets LLC. Electricity market data Hourly data summarizing electricity market conditions (including realized and forecast load, real time prices, and day ahead prices) were collected from the New England, New York, and PJM Indep endent System Op erator websites. Hourly weather station data were obtained from the National Oceanic and Atmospheric Administration. Additional op erating characteristics (including pro duction capacities and fuel characteristics) were obtained from Platts Basecase database. Fuel price data su m of all loads as so ciate d with all c yc le s (steam and electric ) on a cons is te nt bas is . This s ometimes involves converting steam load to an equivale nt e le ctrical load or vice -versa. 2Only those un its that p articipate in b oth th e NB P an d th e th e Acid Rain P rogram are require d to re p ort h ourly op eration s year round . Un its regulate d und er the NBP on ly nee d n ot re p ort ou ts id e of ozon e s eason . Hourly data are also n ot rep orte d whe n plants are take n o/-line f or sche duled outtage s or mainten ance . In total, th e analysis samp le includ es 15.5 million hourly ob servation s. 3The d ata set in clud es April 2003 th rou gh Dec emb er 2006. Th e maximu m numb er of ob servations on a single un it is 32,748 . Th e average numb er of obs ervations p e r un it is much lower (26,646) b e caus e several un its rep ort only du ring oz on e s eason . 4In 2007, th e volume of "ec onomically signi cant" immed iate settlement trade s (i.e. trade s b e twee n versu s within rms ) re ache d 247,000 tons (EPA 2008). 54 5Daily sp ot prices of New York Harb or No. 2 and No. 6 heating oils were obtained from the Energy Information Administration. Daily natural gas sp ot prices were obtained from Platts. 6The OTC NYMEX "Lo ok-Alike" contract, the most actively traded coal pro duct, is used as a measure of coal prices. App endix 5 : Unit-level op erating characteristics The data set used in estimation contains 15,455,067 hourly observations for 580 units. Table A1 decomp oses the observed variation in hourly heat rates (a measure of fuel e¢ ciency) and NOx emissions rates into within and b etween unit comp onents. To generate this table, all observations in which a unit is running at or b elow 5 p ercent capacity factor are dropp ed. When electricity generators are ramping up or down and are op erating at low capacity factors, plants run inef ciently and NOx emissions rates can b e much higher than they are when the plant is running e¢ ciently. Because emissions rates and heat rates observed at these low capacity factors are not go o d indicators of units average op erating characteristics, they are dropp ed. Table A1 : Decomp osing the variance in unit-level op erating characteristics 5For Ne w En gland I us e the n atu ral gas sp ot p rices for Algonquin City Gates an d Ten nes see , zon e 6. For New York gas prices I u se Dominion, n orth p oint. For P JM I use the Transc o Zone 6 non- Ne w York pric es. 6The NYME X lo ok-alike pric e is based on trading us in g th e sp e ci cations in the NYME X f utures contract. Origin ally, NYME X was pred omin antly traded "phys ic al" with c omp anies engaging in bilate ral agre ements. Now a ma jority of NYMEX trad es are trans acted OTC then "given up" to the e xchange for c le aring, wh ich allows marke t partic ip ants to m an age cred it risk over a larger p o ol of c ounte rp arties. 55 NOx emissions rate Heat rate (lbs NOx/MWh) (btu/kWh) Mean 2.79 11,206 Between 5.06 4,952 standard deviation % variation 93% 84% Within 1.42 2,955 standard deviation % variation 7% 16% Observations 5,043,876 5,043,876 Units 580 580 Although the vast ma jority of the observed variation is b etween units, there is some within-unit variation in b oth op erating attributes. This within-unit variation can b e a result of pro duction choices made by the plant manager (i.e. choice of fuel characteristics, the capacity at which the plant is running, combustion tuning, etc.) or it can b e a/ected by purely exogenous factors (such as ambient temp erature). Unit-sp eci c, season-sp eci c (i.e. ozone season or o/-season) measures of these two imp ortant op erating parameters are obtained by estimating regression equations separately for each unit and season. The following general sp eci cation is estimated: = is + X0 its is + "its; can b e interpreted as the i unitzits 56 where zi is the op erating characteristic of interest (i.e. heat rate or emissions rate) and X is a matrix of variables that a/ect a unit s op erational p erformance (ambient temp erature and capacity factor). Because the variables in X are demeaned, the intercept a sp eci c. season-sp eci c value of z under average op erating conditions for unit i: These equations are estimated using all available data and data from hours when units capacity factors exceeded 5%. Unit sp eci c p oint estimates are summarized in table 2. App endix 6 : The Unit Commitment Problem The unit commitment problem (UCP) confronting electricity system op erators has b een extensively studied in the op erations research literature. The problem obtains decisions for a sequence of time p erio ds (typically 24 hours of a day). What happ ens in one hour a/ects what can happ en in subsequent hours; optimal unit commitment in one hour is not indep endent of solutions in other hours. The following UCP sp eci cation was adapted from Hobbs et al. (2001) 8i; t min Xiqit= Dts:t: Xitir8t it+ X qitci it+ X yitUi Xqitit= S D zitt8tM I NizitFi qit + rit zitM AXi 8i; t rit zit M AX S 8i; t Pi q it 1 + M xI N Cii 8i; t qit it 1 8i; t + M xD E C q qit zit 1 + yit 8i; t zit = xit zit zit 1 ak i(qit8i; t X) M xF l 8 k ; i ; owit t where the decision variables are: 57 qitthe MW pro duced by unit i in hour t ritthe MW of spinning reserves itA binary variable indicating whethe provided by unit i in hour t zitA binary variable indicating whether unit binary variable indicating whether u i is dispatched in hour t: y T h e p a r a m e t e r s o f t h e p r o b l e m a r e : D tE le ct ri ci ty d e m a n d in p er io d t: S D D e m a n d fo r s pi n ni n g re s er v e s in p er io d t. Fi tt Fi x e d c o st s of o p er at in g u ni ti in p er io d t ( m e a s ur e d in $/ h o ur ). cit V ar ia bl e o p er at in g c o st s at u ni ti in p er io d t ( m e a s ur e d in $/ M W h) . Ui tT h e c o st s of st ar ti n g u p u ni ti ( m e a s ur e d in $) . M A X T h e m a xi m u m pr o d u ct io n c a p a ci ty of u ni ti . M I Ni iT h e m in i m u m pr o d u ct io n c a p a ci ty of u ni ti (a s s u m e d to b e 0 fo r al li ) M xI N Ci M a xi m u m ra m pu p ra te fo r in cr e a si n g pr o d u ct io n at u ni ti M x D E C M a xi m u m ra m pd o w n ra te fo r d e cr e a si n g pr o d u ct io n at u ni ti . ak iiL in e ar iz e d c o e ¢ ci e nt re la ti n g b u s i in je ct io n to li n e k o w . M x Fl o wi t M a xi m u m M W o w o n li n e k in p er io d t. In th e re st ru ct ur e d el e ct ri ci ty m ar k et s c o n si d er e d in th is p a p er , m ar k et p ar ti ci p a nt s s u b m it bi d s 7t o a n in d e p e n d e nt s y st e m o p er at or (I S O ). U ni t-l e v el pr o d u ct io n a ct iv iti e s ar e c o or di n at e d vi a a t w os et tl e m e nt m ar k et s y st e m . A d a y a h e a d fo r w ar d m ar k et s c h e d ul e s re s o ur c e s a n d d et er m in e s h o ur ly pr ic e s fo r th e fo ll o w in g d a y; a b al a n ci n g m ar k et e n s ur e s th at s u p pl y m e et s u ct u at in g d e m a n d in re al ti m e. O n c e g e n er at or s h a v e s u b m itt e d th ei r s u p pl y bi d s, in d e p e n d e nt s y st e m o p er at or s id e nt if y u ni t c o m m it m e nt s c h e d ul e s to m in i m iz e th e c o st of m e et in g el e ct ri ci ty d e m a n d s u b je ct to th o u s a n d s of u ni t-l e v el a n d s y st e m -l e v el o p er at in g c o n st ra in ts .If r m s a ct a s pr ic et a k er s, s u b m itt in g bi d s th at re e ct tr u e c o st s a n d o p er at in g c o n st ra in ts is a li k el y to b e a pr o t m a xi m iz in g st ra te g y ( G ro s s a n d Fi nl a y, 1 9 9 6) . R e f e r e n c e s B . H o b b s , M . R ot h k o pf , R . O N ei ll, a n d H . C h a o. 2 0 0 1. T h e N e xt G e n er at io n of El e ct ri c P o w er U ni t C o m m it m e nt M o d el s. Kl u w er . 7T h e d i/e re nt re gi o n al e le ct ric ity m ar ke ts re q ui re g e n er at or s to su b mi t bi d s in di/ e re nt f or m s. F or ex a m pl e, th e N e w Y or k IS O s e p ar at e s e n e rg y b id s int o st ar tu p co st cu rv es a n d a n e n er gy cu rv e w hi ch ca n b e a th re e p ar t st e p fu nc tio n or a 6 s e g m e nt pi ec e wi se lin e ar c ur ve . 5 8 G.Gross and D.J. Finlay, 1996, Optimal bidding strategies in comp etitive electricity markets, Pro ceedings of Power Systems Computation Conference (PSCC 96), Dresden, August 19-23, 1996, pp.815-823). 59