A Laboratory Investigation of Compliance Behavior under Tradable Emissions Rights: Implications for Targeted Enforcement

Transcription

1 Unversty of Massachusetts Amherst Department of Resource Economcs Workng Paper No A Laboratory Investgaton of Complance Behavor under Tradable Emssons Rghts: Implcatons for Targeted Enforcement James J. Murphy 1 and John Stranlund 2 Abstract: Ths paper uses laboratory experments to test the theoretcal observatons that both the volatons of compettve rsk-neutral frms and the margnal effectveness of ncreased enforcement across frms are ndependent of dfferences n ther abatement costs and ther ntal allocatons of permts. Ths concluson has mportant mplcatons for enforcng emssons tradng programs because t suggests that regulators have no justfcaton for targetng ther enforcement effort based on frm-level characterstcs. Consstent wth the theory, we fnd that subjects volatons were ndependent of parametrc dfferences n ther abatement costs. However, those subjects that were predcted to buy permts tended to have hgher volaton levels than those who were predcted to sell permts. Despte ths, we fnd no statstcally sgnfcant evdence that the margnal effectveness of enforcement depends on any frmspecfc characterstc. We also examne the determnants of complance behavor under fxed emssons standards. As expected, we fnd sgnfcant dfferences between complance behavor under fxed standards and emssons tradng programs. Keywords: enforcement, complance, emssons tradng, permt markets, standards, commandand-control JEL Classfcaton: C91, L51, Q58 1 James J. Murphy, Department of Resource Economcs Unversty of Massachusetts, 219A Stockbrdge Hall Amherst, MA E: murphy@resecon.umass.edu P: F: John K. Stranlund, Department of Resource Economcs Unversty of Massachusetts, 214 Stockbrdge Hall Amherst, MA E: stranlund@resecon.umass.edu P: F:

2 February 2005 A Laboratory Investgaton of Complance Behavor under Tradable Emssons Rghts: Implcatons for Targeted Enforcement JAMES J. MURPHY * Department of Resource Economcs & Center for Publc Polcy and Admnstraton Unversty of Massachusetts-Amherst JOHN K. STRANLUND Department of Resource Economcs Unversty of Massachusetts-Amherst Abstract: Ths paper uses laboratory experments to test the theoretcal observatons that both the volatons of compettve rsk-neutral frms and the margnal effectveness of ncreased enforcement across frms are ndependent of dfferences n ther abatement costs and ther ntal allocatons of permts. Ths concluson has mportant mplcatons for enforcng emssons tradng programs because t suggests that regulators have no justfcaton for targetng ther enforcement effort based on frm-level characterstcs. Consstent wth the theory, we fnd that subjects volatons were ndependent of parametrc dfferences n ther abatement costs. However, those subjects that were predcted to buy permts tended to have hgher volaton levels than those who were predcted to sell permts. Despte ths, we fnd no statstcally sgnfcant evdence that the margnal effectveness of enforcement depends on any frm-specfc characterstc. We also examne the determnants of complance behavor under fxed emssons standards. As expected, we fnd sgnfcant dfferences between complance behavor under fxed standards and emssons tradng programs. Keywords: enforcement, complance, emssons tradng, permt markets, standards, commandand-control JEL Codes: C91, L51, Q58 Acknowledgements Prmary fundng for ths research was provded by the U.S. EPA Scence to Acheve Results (STAR) Program grant #R Addtonal support was provded by the Cooperatve State Research Extenson, Educaton Servce, U. S. Department of Agrculture, Massachusetts Agrcultural Experment Staton, and the Department of Resource Economcs under Project No. MAS00871, and by the Center for Publc Polcy and Admnstraton, Unversty of Massachusetts-Amherst. Mara Alejandra Velez, Carre Pugls and Mara Clauda Lopez provded outstandng research assstance. Glenn Caffery programmed the software for ths project. Wendy Varner provded valuable admnstratve support. The authors take full responsblty for any errors or omssons. * Correspondence to James J. Murphy, Department of Resource Economcs, 80 Campus Center Way, 219A Stockbrdge Hall, Unversty of Massachusetts-Amherst, Amherst, MA 01003, USA. Phone: (413) , Fax: (413) , E-mal: murphy@resecon.umass.edu.

3 A Laboratory Investgaton of Complance Behavor under Tradable Emssons Rghts: Implcatons for Targeted Enforcement 1. Introducton Although often overlooked, one of the most mportant desgn elements of any regulatory polcy s how complance wth the polcy wll be enforced. Wthn the context of desgnng emssons tradng programs, several authors have provded theoretcal analyses of complance ncentves, the consequences of noncomplance, and strateges for enforcng emssons tradng programs (e.g., Keeler 1991, Malk 1990, 1992, and 2002, vanegteren and Weber 1996, Stranlund and Dhanda 1999, Stranlund and Chavez 2000). Taken as a whole, ths lterature suggests that frms ncentves toward noncomplance n market-based regulaton, as well as the desgn of enforcement strateges to counteract these ncentves, are qute dfferent from complance and enforcement of other polcy nstruments, partcularly command-and-control regulatons. Whle there s a substantal body of economc theory about complance and enforcement n emssons tradng programs, there s lmted emprcal evdence regardng the determnants of complance decsons n these programs. In ths paper, we report the results of a seres of experments desgned to examne some key factors that may affect complance decsons. Our prmary nterest s whether dfferences n the ndvdual characterstcs of frms generate dfferent complance choces. Ths s an mportant ssue because, on the face of t, one may suspect that frms wth dfferent producton processes, abatement technologes, or ntal allocatons of permts may have dfferent complance ncentves. If ths s true, then regulators wll be motvated to choose a targeted enforcement strategy, n partcular targeted montorng effort, whch s condtoned on frm-level characterstcs. Regulators mght use nformaton about frm-level determnants of complance behavor n dfferent ways dependng on what ther enforcement objectves are. If some frms tend toward 1

4 hgher volatons than others, then a regulator that s prmarly motvated to detect those wth hgher volatons wll drect a greater share of hs enforcement effort toward these frms. On the other hand, suppose that a budget-constraned regulator seeks to dstrbute hs enforcement effort to mnmze aggregate volatons. 1 In ths case, the regulator s not necessarly concerned about whether some frms have hgher volatons than others. Hs man concern s to drect hs enforcement effort to where t s most productve. That s, he wll montor more closely those frms that are more responsve to ncreased enforcement than others, regardless of volaton level. Conceptually, Stranlund and Dhanda (1999) show that the ndvdual complance choces of rsk neutral compettve frms n emssons tradng programs are ndependent of dfferences n any frm-level characterstc. Consequently, regulators have no reason to condton ther enforcement effort on frm-level characterstcs. Ther reasonng s straghtforward. Snce complance n emssons tradng programs means that a frm holds enough permts to cover ts emssons, a rsk neutral compettve frm s margnal beneft of noncomplance s what t has to spend for permts to make sure t s complant; that s, the prevalng permt prce. Thus, a frm s complance decson s made by comparng ths permt prce wth the margnal expected penalty for emssons n excess of permts. Snce ths margnal beneft-cost comparson does not depend on anythng unque to a partcular frm, the complance decson s ndependent of any frm characterstc. Ths result contrasts sharply wth the effects of frm-level characterstcs on complance wth command-and-control standards. A rsk neutral frm s decson about whether to comply wth a fxed emssons standard should be determned by the relatonshp between ts margnal 1 In the case of unformly mxed pollutants, ths s equvalent to seekng to reduce the envronmental damage from noncomplance. If polluton s not unformly mxed, then regulators may wsh to target enforcement effort at frms that have greater mpacts on envronmental qualty. We do not address ths possblty n ths paper. 2

5 reducton n abatement costs from exceedng the standard and the margnal expected penalty t faces for the resultng volaton. Therefore, frms wth hgher margnal abatement costs or who face strcter standards wll have a greater ncentve to be noncomplant. In ths way, frm-level characterstcs are mportant determnants of complance wth fxed standards (Garve and Keeler 1994). A recent paper by Gray and Shadbegan (2004) fnds strong support for ths concluson n ther analyss of complance behavor by pulp and paper manufacturers. Our expermental results are largely consstent wth the general hypothess that ndvdual complance decsons n emssons tradng programs are ndependent of dfferences n frms characterstcs, but not completely so. We fnd that ndvdual volatons are ndependent of parametrc dfferences n ther abatement costs. However, ndvdual volatons are not ndependent of the ntal allocaton of permts. What appears to matter most about the ntal allocaton of permts s how t determnes who wll sell permts and who wll buy them. We fnd that those subjects who are predcted to buy permts tend to have hgher volaton levels than those who are predcted to sell permts. We also test whether the margnal productvty of ncreased enforcement n reducng ndvdual volatons vares accordng to dfferences n abatement costs and ntal allocatons of permts, and fnd that the margnal productvty of enforcement s statstcally ndependent of frm-specfc characterstcs. We also ran a seres of experments n whch subjects faced fxed emssons standards to compare complance behavor under emssons tradng and emssons standards. These experments were dentcal to our market experments, except that subjects could not trade ther ntal allocaton of permts. The results of these experments provde an emprcal study of how dfferent complance behavor s under the two types of polces. Consstent wth theoretcal expectatons (e.g., Garve and Keeler 1994), subjects wth hgher abatement costs had sgnfcantly hgher volatons and were much more responsve to ncreased enforcement. 3

6 Clearly our results have mportant mplcatons for desgnng enforcement strateges for compettve emssons tradng programs. There s lttle emprcal justfcaton for targetng frms n these programs based on dfferences n ther characterstcs. Whle our results suggest that an enforcer whose objectve s to detect and penalze frms that tend toward hgher volatons may wsh to montor permt buyers more closely, rrespectve of abatement costs, there appears to be no justfcaton for targetng frms to maxmze the productvty of enforcement resources. Under fxed emssons standards, however, there s substantal justfcaton for targeted enforcement strateges that are condtoned on nformaton about frms abatement costs Although expermental technques have been used to evaluate other polcy ntatves, ncludng some aspects of emssons tradng programs (e.g., Cason 1995, Cason and Plott 1996, Ishkda et al. 1998, Isaac and Holt 1999), these technques have not yet been wdely appled to ssues of regulatory enforcement, much less to complance behavor n emssons tradng programs. 2 However, laboratory experments are partcularly useful for an nqury nto the determnants of complance n these programs, because the opportuntes for emprcal analyses of complance behavor n exstng tradng programs appear to be very lmted at present. Some programs have acheved such hgh rates of complance that there s smply s not enough varaton n complance decsons to conduct meanngful econometrc analyses. For example, the enforcement strateges of the Sulfur Doxde Allowance Tradng program and the RECLAIM program of southern Calforna were clearly desgned to acheve very hgh rates of complance, and have been largely successful n dong so (Stranlund, Chavez, and Feld 2002). Other emssons permt markets have not yet fully developed. Ths s the case for the emssons tradng program for total suspended partculates n Santago, Chle (Montero, Sanchez, and Katz 2002). In fact, because so few trades have taken place n ths program, Palacos and Chávez (2004) were 2 See Alm and McKee (1998) for a revew of the expermental lterature that focuses on tax complance. 4

7 forced to assume that sources n ths program faced fxed standards to conduct an emprcal analyss of complance behavor n ths program. We are aware of only two papers that use laboratory experments to examne complance behavor n emsson permt markets. Cason and Gangadharan s (2004) experments nvolve emssons tradng when emssons are stochastc, permts can be banked, and subjects performance s audted based on ther past complance hstory. Ths desgn allowed them to dentfy nterestng nteractons between random emssons shocks, permt bankng, and complance. Our approach s much smpler: emssons are determnstc, bankng permts from one perod to the next s not allowed, and audts are completely random wth a known and constant probablty. Ths approach allows us to generate fundamental results about the potental for targeted enforcement n emssons tradng programs that Cason and Gangadharan do not consder. Murphy and Stranlund (2004) s closely related to our current work n terms of desgn, but has a dfferent focus. That work consders the effects of changng enforcement strateges on the permt market, and n turn the ndrect effects of enforcement on complance behavor. Although dentfyng the drect and ndrect market effects of enforcement s also mportant to us n ths paper, Murphy and Stranlund (2004) do not examne how the margnal productvty of enforcement may vary across frms, and hence, they do not draw conclusons about whether targeted enforcement s approprate n emsson tradng programs. In addton, ther work focuses solely on mperfect complance. We nclude treatments n whch subjects are predcted to be perfectly complant so that we can examne frm-level determnants of complance over the full range of enforcement strateges, from weak enforcement wth sgnfcant noncomplance to strong enforcement that s suffcent to nduce perfect complance. Fnally, nether Cason and Gangadharan (2004) nor Murphy and Stranlund (2004) conduct experments wth fxed 5

8 emssons standards to draw conclusons about how complance decsons dffer under marketbased and command-and-control regulaton. The rest of ths paper s organzed as follows. In the next secton we present a postve theory of complance n a compettve emssons tradng program to develop the hypotheses to be tested. In secton 3 we provde detals of the experments we desgned to conduct these tests. The results of the experments are presented and dscussed n secton 4. Secton 5 concludes. 2. Theory and Hypotheses Consder a fxed number of rsk neutral frms n a perfectly compettve emssons tradng program. 3 The abatement costs of frm are summarzed by cq (, θ ), whch s strctly decreasng and convex n the frm s emssons q. Heterogenety of the frms s captured by the parameter θ, and we assume that total and margnal abatement costs are ncreasng n ths parameter. A total of Q emssons permts are dstrbuted to the frms, free of charge. Frm receves l 0 permts ntally, and holds l permts after tradng n a complance perod s complete. Compettve behavor n the permt market establshes a constant prce per permt p. If frm s noncomplant, then ts emssons exceed the number of permts t holds and the magntude of ts volaton s v = q l > 0. If the frm s complant, q l 0 and v = 0. To check for complance, each frm s audted wth a known probablty π. A frm that s found to be n volaton s assessed a penalty, f(v, φ). There s no penalty for a zero volaton, but for postve 3 None of the results that we present n ths secton are new, so we do not provde rgorous dervatons. Interested readers should consult Stranlund and Dhanda (1999) for a complete analyss. It s mportant to note that the theoretcal model we use n ths paper can easly be appled to other tradable property rghts programs. For example, recent papers by Hatcher (2005) and Chavez and Salgado (2004) are drect applcatons of the lterature on complance and enforcement of emssons tradng to ndvdual transferable fshng quotas. Thus the results of ths paper apply more broadly than just emssons tradng. 6

9 volatons, the penalty s postve, strctly ncreasng and strctly convex. An ncrease n the parameter φ ncreases both total and margnal penaltes. Assumng that each frm chooses postve emssons and permts and never over-comples, then s problem s to mnmze ts expected costs, cq (, θ ) + pl ( l0) + π f( q l, φ), subject to q l 0. It s more convenent for our purposes to focus on the choces of volatons and emssons rather than on permt demands and emssons. Therefore, we rewrte the problem as mn cq (, θ ) + pq ( v l0) + π f( v, φ), subject to v ( v, q ) 0. The followng frst-order condtons are both necessary and suffcent to determne optmal choces of volaton and emssons: L = p+ π f ( v, φ) 0, v 0, L v = 0; [1] v v v L = c ( q, θ ) + p= 0. [2] q q These condtons mplctly defne a frm s optmal choces of volaton and emssons as v = v ( θ, π, φ, p) and q = q ( θ, π, φ, p). Note that these choces do not depend at all on a frm s ntal allocaton of permts, smply because l 0 does not enter ether frst-order condton. Ths mples that ndvdual volaton levels and the margnal productvty of enforcement across frms do not vary wth changes n the ntal allocaton of permts. 4 For the purposes of ths study, the most mportant result from ths smple model s that the decson to comply and the choce of volaton level are ndependent of any frm-specfc characterstcs (whch are reflected n the parametersθ and l 0 ). From [1] t s easy to establsh 4 Ths s an extenson of Montgomery s (1972) result that emssons choces are ndependent of the ntal allocaton of permts to envronments n whch frms can be noncomplant. It s well known that ths result does not hold n the presence of market power (Hahn, 1984) or transacton costs (Stavns, 1995). Smlarly, complance choces wll not be ndependent of the ntal allocaton of permts n the presence of market power (van Egteren and Weber, 1996; Malk, 2002; Chavez and Stranlund, 2003), or transacton costs (Chavez and Stranlund, 2004). 7

10 that a frm wll be complant f and only f p π f v (0, φ) ; that s, a frm s complant f and only f the permt prce s not greater than margnal expected penalty of a slght volaton. Note that there s nothng n ths decson rule that s unque to an ndvdual frm. Therefore, a frm s decson about whether to comply s ndependent of any of ts characterstcs. Ths ndependence result extends to the volaton level of a noncomplant frm as well. To see ths let [1] hold wth equalty and substtute the frm s choce of volaton, v = v ( θ, π, φ, p), to obtan p π f ( v ( θ, π, φ, p)) 0. [3] v Dfferentate ths dentty wth respect to θ to obtan π fvvv θ = 0, whch mples v θ = 0. Ths result ndcates that frms volatons are ndependent of parametrc dfferences n ther abatement costs. Conceptually, therefore, there s no reason for regulators to beleve that some frms wll be more lkely to be noncomplant or tend toward hgher volatons even though they may have very dfferent abatement cost functons or ntal permt allocatons. Hence, a regulator that s motvated to target hs enforcement resources to detect ncdences of noncomplance or hgher levels of volaton cannot do so productvely on the bass of frm-level characterstcs. On the other hand, a regulator that s motvated to target hs enforcement effort where t wll be most productve n reducng aggregate volatons wll not necessarly care about ndvdual volaton levels, but rather wll be focused on the margnal productvty of hs enforcement effort across frms. Ths matter s complcated somewhat by a unque feature of enforcng marketbased polces: enforcement wll have a drect effect on ndvdual volatons, as well as an ndrect effect through ts mpact on the equlbrum of the permt market. From [3] obtan v = f / π f < 0 and v = f / f < 0, whch confrm the unsurprsng π v vv φ results that, holdng the permt prce constant, frms respond to ether ncreased montorng or ncreased penaltes by decreasng ther volatons. In prncple, frms could reduce ther vφ vv 8

11 volatons by purchasng more permts or reducng ther emssons. However, [2] ndcates that frms always choose ther emssons to equate ther margnal abatement costs to the permt prce. Ths rule has nothng to do wth the enforcement strategy frms face. Therefore, agan holdng the permt prce constant, frms reduce ther volatons n response to ncreased enforcement by purchasng more permts, not by decreasng ther emssons. If every frm ncreases ts demand for permts wth ncreased enforcement, then the equlbrum permt prce must ncrease. To defne the equlbrum permt prce, frst defne the vector θ = { θ } N, where N s the set of regulated frms. Gven that a total of Q permts are ssued to the frms, and the enforcement authorty has commtted tself to montorng each frm wth probablty π and mposng penaltes wth parameter φ, the equlbrum permt prce s p = p( θ, πφ,, Q). Ths s the mplct soluton to the permt market clearng condton, l ( θ, π, φ, p ) = Q, where l ( θ, π, φ, p) s permt demand by frm. It s easy to demonstrate N that the equlbrum permt prce s ndeed ncreasng n the enforcement parameters; that s, p π > 0 and p φ > 0. Moreover, an ncrease n the permt prce motvates frms toward hgher volatons. To demonstrate ths, obtan v = 1/ π f > 0 from [3]. The ntuton here s that the p permt prce represents the margnal cost of complance; thus, a hgher permt prce ncreases the ncentve toward noncomplance. The equlbrum volaton of a frm s defned as v θ π φ Q = v ( θ π φ p θ π φ Q ) vv (,,, ),,, (,,, ). The margnal productvty of ncreased enforcement on the volatons of frm s therefore v = v + v p and π π p π v = v + v p. The drect effect of ncreased enforcement s to reduce the φ φ p φ volaton of an ndvdual frm ( v < 0 and v < 0 ), but the ndrect effect motvates hgher π φ volatons ( v p > 0 and v p > 0 ). It s easy to show that the drect effect domnates the p π p φ 9

12 ndrect effect so that ndvdual volatons fall wth ncreased enforcement. Murphy and Stranlund (2004) confrm ths hypothess n the laboratory. Now let us turn to possble dfferences n the margnal productvty of enforcement across frms. Suppose that v π > v for two frms, and j. That s, at the margn, the regulator s j π montorng effort s more productve n reducng the volaton of frm than frm j. If the regulator s objectve s to reduce aggregate volatons and the margnal cost of montorng the two frms s the same, the regulator would be motvated to montor frm somewhat more closely than frm j. Conceptually, however, snce a frm s choce of volaton s ndependent of ts type and ts ntal allocaton or permts, the margnal productvty of the regulator s enforcement s ndependent of these parameters as well. 5 j Ths mples that dfferences n vπ and vπ cannot be attrbuted to dfferences n s and j s abatement costs or ther ntal allocatons of permts. Matters are very dfferent for frms that face fxed emssons standards. If frm faces an emssons standard s, ts complance problem s to choose emssons to mnmze cq (, θ ) + π f( q s, φ), subject to q s. Equvalently, t chooses ts volatons to mnmze cv ( + s, θ ) + π f( v, φ). The frst-order condton s c ( v + s, θ ) + π f ( v, φ) 0; f ths s strctly postve, q then v = 0. Clearly, the frm s noncomplant f and only f c ( s, θ ) > π f (0, φ); that s, f and only f the frm s margnal abatement costs evaluated at the standard s greater than the margnal expected penalty of a slght volaton. It s straghtforward to show that a noncomplant frm s volaton s ncreasng n ts margnal abatement costs [ v θ < 0 ] and decreasng n the emssons standard [ v < 0 ]. Thus, a regulator who s motvated to detect and penalze noncomplant frms s and those that tend toward hgher volatons should target frms wth hgher margnal abatement v q v 5 That s, v θ = 0 mples vπθ = vφθ = v p θ = 0 and consequently, v πθ = 0 10

13 costs and/or lower standards. Contrast ths to compettve emssons tradng for whch regulators cannot use dfferences n margnal abatement costs or the ntal allocaton of permts to target noncomplant frms or those wth hgher volatons. In further contrast to permt markets, the margnal productvty of ncreased enforcement wth standards s lkely to depend on frms abatement costs and emssons standards. Although, n general, t s not possble to dentfy the frms for whch a regulator s enforcement effort s more effectve, f frms abatement costs and the penalty functon are quadratc, as we assume n our expermental desgn, then t s straghtforward to show that v πθ < 0 and v π < 0. These ndcate that the margnal productvty of ncreased montorng n reducng volatons s greater for frms wth hgher margnal abatement costs or that face lower standards. The same holds for the margnal productvty of ncreased penaltes. Thus, to use hs enforcement resources most effectvely, a regulator s justfed n targetng frms wth hgher margnal abatement costs or lower standards. Our results about volaton choces, the margnal productvty of enforcement, and the potental for targeted enforcement are summarzed n the followng hypotheses: s Hypothess 1: Under a compettve emssons tradng program, the volaton levels of rsk neutral frms are ndependent of dfferences n ther abatement costs and ther ntal allocatons of permts. Hypothess 2:Under a compettve emssons tradng program, the margnal productvty of ncreased enforcement (ether ncreased montorng or ncreased penaltes) on the volaton levels of rsk neutral frms s ndependent of dfferences n the frms abatement costs and ther ntal allocatons of permts. Hypothess 3: Under fxed emssons standards, the volaton levels of rsk neutral frms are hgher for frms wth hgher margnal abatement costs and/or lower emssons standards. 11

14 Hypothess 4: Under fxed emssons standards, the margnal productvty of ncreased enforcement (ether ncreased montorng or ncreased penaltes) on the volaton levels of rsk neutral frms s greater for frms wth hgher margnal abatement costs and/or lower emssons standards. 3. Expermental Desgn and Procedures 3.1 Experment desgn The experments were desgned to test our hypotheses about the relatonshps between ndvdual frm characterstcs and volaton choces, but the subjects were placed n a more neutral envronment. We framed the experments as a producton decson n whch permts conveyed a lcense to produce, rather than an emssons decson, to avod ntroducng potental bases due to ndvdual atttudes about the envronment or emssons tradng. Durng each perod of the experment, subjects smultaneously chose to produce unts of a fcttous good and traded n a market for permts to produce the good. Partcpants could produce as many unts of the good as they wshed (up to a capacty constrant) regardless of the number of permts that they owned. However, at the end of the perod, each ndvdual was audted wth a known, exogenous probablty. If an ndvdual was audted and found to be non-complant (.e., total producton exceeded permt holdngs), then a penalty was appled. Subjects receved a beneft from ther choce of producton, q, whch was generated from a lnear margnal beneft functon b ( q) = 18 β q. Each experment had eght subjects dvded evenly nto two types; subjects were randomly assgned a type. Subjects wth a hgh margnal beneft functon ( β = 1) also had a greater producton capacty (17 unts). Subjects wth a low H margnal beneft functon ( β L = 2 ) could produce up to 8 unts. Producton was constraned to be a whole number. Hgh margnal benefts from producton n our experments are equvalent to 12

15 hgh margnal abatement costs n emssons tradng. Therefore, throughout the remander of the paper we denote subjects wth hgh margnal producton benefts as HghMAC subjects, and subjects wth low margnal producton benefts as LowMAC subjects. The treatment varables n ths paper are the probablty of an audt, the margnal penalty functon, the ntal permt allocaton, and the total supply of permts. Table 1 summarzes the expermental desgn. Each of the twelve cells was repeated three tmes. Although our man focus s on complance behavor n permt markets, we also ran several experments n whch subjects faced fxed standards to compare complance behavor under emssons tradng and emssons standards. To be complant n the market experments, subjects were requred to possess suffcent permts, l, to cover ther producton choces. Lmtng the total number of permts mposed an aggregate producton standard. We consder two aggregate standards. In the low aggregate standard experments ( Q L = 28 ), denoted LowStandard, each of the four HghMAC subjects was allocated three permts, and the four LowMAC subjects were each gven four permts. We call ths the (nearly) unform ntal allocaton. In the hgh aggregate standard experments ( Q H = 56 ), denoted HghStandard, there were two dfferent ntal allocatons of permts. Wth a unform ntal allocaton, each of the eght subjects n an experment started wth seven permts. Wth a nonunform ntal allocaton, the HghMAC subjects began wth 13 permts, and the LowMAC subjects had a sngle permt. Reallocatng the ntal allocaton of permts n ths way changes the predcton about whch subjects would sell permts and whch would buy permts. In the compettve equlbrum, the LowMAC subjects would be the net sellers of permts wth the unform ntal allocaton, and net buyers of permts when the ntal allocaton s nonunform. 13

16 To check for complance, each subject s record was examned wth a known probablty π. If a subject was audted and found to be non-complant, that s q > l, then she was assessed a penalty that was generated from a lnearly ncreasng margnal penalty functon, f ( q l) = F + φ( q l). By changng the parameters of the margnal expected penalty functon, π f ( q l) = π[ F + φ( q l)], we developed four enforcement strateges, whch we label Hgh, Medum(π H ), Medum(π L ), and Low. 6 The Hgh enforcement strategy nvolved a hgh audt probablty and a hgh margnal penalty functon (π H = 0.70, F = 17.5, φ = 1.43). All Hgh margnal expected penalty treatments were parameterzed to nduce perfect complance by rsk neutral agents. Wth the hgh aggregate standard, however, the ncentves for perfect complance were much stronger. For the other three levels of enforcement, subjects were expected to choose to be noncomplant. The treatments Medum(π H ) and Medum(π L ) nvolved the same margnal expected penaltes, but Medum(π H ) had a hgher montorng probablty and a relatvely low margnal penalty functon (π H = 0.70, F = 6, φ = 1.43), whereas Medum(π L ) had a lower montorng probablty and a hgher margnal penalty functon (π L = 0.35, F = 12, φ = 2.90). Our ntenton here was to examne whether the subjects reacted dfferently to montorng and penaltes. The Low margnal expected penalty treatments had the weakest enforcement strategy wth the low montorng probablty and a low margnal penalty functon (π L = 0.35, F = 2, φ = 2.90). Enforcement parameter values were chosen, n part, so that the expected margnal penalty functons are parallel to each other each has a slope of about one. 6 The enforcement parameters were the same for each subject type wth the excepton that, snce LowMAC subjects could only produce up to eght unts, only the frst eght steps of the penalty schedule was dsplayed for them. 14

17 3.2 Experment procedures Partcpants were recruted from the student populaton at the Unversty of Massachusetts, Amherst. Subjects were pad $7 for agreeng to partcpate and showng up on tme, and were then gven an opportunty to earn addtonal money n the experment. These addtonal earnngs ranged between $5.68 and $17.49, wth a mean of $13.67 (σ = 1.46). Earnngs were pad n cash at the end of each experment. Each experment lasted about 2 hours. The experments were run n a computer lab usng software desgned specfcally for ths research. To famlarze subjects wth the experments, we ran a seres of tranng experments. In the frst stage of the traners, students read onlne nstructons that ncluded nteractve questons to ensure that they understood the nstructons before proceedng. After everyone had completed the nstructons and all questons were answered, the tranng experment began. These practce rounds contaned all the same features as the real data experments wth the excepton that we used a dfferent set of parameters. The data from the traners were dscarded. For the real data sessons, we recruted partcpants from the pool of traned subjects. Subjects were allowed to partcpate n multple sessons. A total of 176 subjects partcpated n 36 eght-person experments. Pror to the start of the real data experments, subjects were gven a summary of the experment nstructons. 7 The expermenter read these nstructons aloud and answered any questons. Each subject was gven a calculator, a pencl and paper. Each experment conssted of 12 dentcal rounds. At the start of each perod, the eght subjects were each gven an ntal allocaton of permts and $10 n expermental cash. A unque feature of our experments s that the producton decsons and permt market tradng were unbundled nto two separate, but smultaneous, actvtes. We dd ths to allow for 7 Avalable upon request. 15

18 the possblty that the producton levels and permt holdngs could dffer, thereby ntroducng a complance decson. Durng the perod and concurrent wth the producton decson, subjects had the ablty to alter ther permt holdngs by tradng n a contnuous double aucton. In the aucton, ndvduals could submt bds to buy or asks to sell a sngle permt. The hghest bd and lowest ask prce were dsplayed on the screen. A trade occurred whenever a buyer accepted the current ask or a seller accepted the current bd. After each trade, the current bd and ask were cleared and the market opened for a new set of bds and asks. The tradng prce hstory was dsplayed on the screen. Each perod lasted a total of fve mnutes. The permt market was open for the entre perod, but producton had to be completed n the frst four mnutes. The one-mnute reconclaton perod gave subjects a fnal opportunty to adjust ther permt holdngs. After each perod ended, random audts were conducted and penaltes were assessed. All nformaton relatng to audt outcomes was prvate. As noted earler, we also conducted a seres of experments wth a fxed standard for the subset of treatments n Table 1 that have a unform ntal permt allocaton. These standards experments were the same as ther market counterparts, except that the permts were not tradable. Therefore, the ntal permt allocaton became the fxed standard. We use these treatments to test Hypotheses 3 and 4. Ths desgn yelds observatons on ndvdual decsons and market outcomes; of partcular nterest n ths paper are volatons and permt prces. Gvng subjects dfferent producton beneft schedules and ntal allocatons of permts allow us to test for the mpacts of these dfferences on ndvdual volatons under both emssons tradng and fxed emssons standards (Hypotheses 1 and 3). Moreover, varyng the enforcement parameters allows us to test whether 16

19 the margnal productvty of enforcement s senstve to dfferences n producton benefts and ntal permt allocatons or fxed standards (Hypotheses 2 and 4). 4. Results We begn our analyss of the results wth a dscusson of general patterns n the permt prce and ndvdual volatons. We use some of these patterns to motvate our econometrc model specfcatons when estmatng these varables. The econometrc analyss s used to test our specfc hypotheses. Because our theoretcal development suggests that an ndvdual s volaton decson s condtoned on the permt prce, we frst estmate ths prce and then use the estmated prce as an nstrumental varable when modelng volaton choces. Because ndvduals make decsons n mult-round experments, we control for repeated measures usng lnear random effects models. We omt the data from the frst perod to mnmze the effects of learnng and prce dscovery. Ths omsson does not have a qualtatve effect on any of our conclusons. In addton to the data reported n ths paper, we also ran a separate seres of perfect complance experments for the low and hgh standard wth unform permt allocatons. These experments were dentcal to those descrbed n ths paper, except that emssons were assumed to exactly equal the fnal permt balance; that s, noncomplance was not allowed and audts were unnecessary. As expected, observed prces and quanttes n these experment quckly converged to the compettve equlbrum. Therefore, we are confdent that any devatons from the compettve predcton n the dscusson below reflect treatment effects related to the complance decson and are not artfacts of the subject pool or expermental desgn. 8 8 The results from our perfect complance experments are avalable upon request. 17

20 4.1 General patterns n market experments Table 2 presents some summary statstcs of permt prces. Note that mean and medan prces tend to be hgher than the compettve equlbrum predcton, but they move as expected: prces are hgher when the aggregate standard s reduced and when the margnal expected penalty s ncreased. Note, however, that wth the HghStandard treatments, for a gven margnal expected penalty, mean and medan prces are hgher for the nonunform ntal allocaton of permts. Ths, of course, s not consstent wth the conventonal hypothess that behavor and market outcomes under emssons tradng programs are ndependent of the ntal allocaton of permts. Table 3 presents some summary statstcs for ndvdual volatons. Mean and medan volatons also move as expected. Gven the aggregate standard, they tend to be lower wth hgher margnal expected penaltes, and gven the enforcement strategy, they tend to be lower wth the HghStandard. Conceptually, ths latter result s due to the lower permt prces resultng from a hgher aggregate standard. We note an nterestng pattern n mean and medan volatons that plays an mportant role n how we analyze and nterpret volaton choces. The compettve equlbrum predcton s that ndvdual volatons should be unform across subjects n each treatment. However, mean and medan volaton levels clearly dffer by subject type, but whether HghMAC subjects tend toward hgher or lower volatons than LowMAC subjects appears to depend on the ntal allocaton of permts. In partcular, consder the eght unform allocaton treatments. The HghMAC subjects are predcted to be the net buyers of permts n these treatments, and they clearly tend toward hgher volatons than the LowMAC subjects. On the other hand, for the four nonunform allocaton treatments, the LowMAC subjects are predcted to be the permt buyers, and they are the ones that tend toward hgher volatons. It appears that the dfferences n volatons by subject type may be determned n part by how the ntal allocaton of permts 18

21 determnes whether subjects wll be net buyers or sellers of permts. Ths speculaton s easly tested and we wll do so shortly. 4.2 Market regresson results and tests of hypotheses Table 4 presents the results of a lnear random effects model of the permt prce. The dependent varable s the prce of each trade n perod t = 2,,12 of group j = 1,,27. The margnal expected penalty, the aggregate standard, and the ntal allocaton are modeled as fxed effects. Assumng rsk neutral subjects, snce the Medum(π H ) and Medum(π L ) enforcement strateges have the same margnal expected penaltes, n theory both should lead to dentcal market outcomes. A test of the null hypothess that the coeffcents on the dummy varables for the Medum(π H ) and Medum(π L ) treatments are equal cannot be rejected at any conventonal level of sgnfcance (Wald χ 2 (1) = 0.92, p = 0.34). Thus, decreasng montorng and ncreasng penaltes, but leavng the margnal expected penalty functon unchanged had no affect on market prces. The prce estmaton results n Table 4 confrm the mpressons we reached by comparng average prces across treatments. The coeffcents for the three margnal expected penaltes are ncreasng, ndcatng that ncreased enforcement puts upward pressure on the equlbrum prce. The NonUnform coeffcent s postve and sgnfcant, as we expected from our perusal of the average prce results, even though theory predcts otherwse. Lastly, the coeffcent on HghStandard s negatve and sgnfcant, ndcatng the unsurprsng result that permt prces fall wth a greater supply of permts. Table 5 presents the results of a two lnear random effects models for ndvdual volatons. Model 1 estmates the effects of the treatment varables on volatons. Model 2 adds the nteracton effects needed to test Hypothess 2, whch requres that we capture all possble combnatons of subject type and treatment cells n Table 2. Usng an nstrumental varable 19

22 approach, PrceHat s the estmated prce from the model n Table 4. NetSeller s a fxed effect that equals one f the subject s predcted to be a net seller (HghMAC subjects for the nonunform allocatons and LowMAC subjects for the unform allocatons). HghMAC s a fxed effect that equals one for HghMAC subjects. Note the negatve and sgnfcant mpact of enforcement, and the postve and sgnfcant mpact of prce. These are consstent wth the theoretcal model we presented n secton 2. Whle comparng mean and medan volatons across treatments, we suggested that dfferences n volatons by subject type may be determned n part by how the ntal allocaton of permts determnes whether subjects wll be net buyers or sellers of permts. Ths s confrmed by the estmaton results. In Model 1, the NetSeller coeffcent s negatve and sgnfcant. Although NetSeller s not sgnfcant n Model 2, a test of the jont hypothess that NetSeller and the four terms nteracted wth NetSeller are all equal to zero s rejected (Wald χ 2 (5) = 27.6, p = 0.00). On the other hand, the coeffcent on HghMAC s small and not statstcally sgnfcant n both Models 1 and 2. 9 Hypothess 1 asserts that ndvdual volatons are ndependent of any subject-level characterstc. Our tests of ths hypothess are mxed. We have strong support for the hypothess that volatons are ndependent of dfferences n subjects margnal abatement costs. However, ndvdual volatons are not ndependent of how the ntal allocaton of permts determnes whch subjects wll sell permts and whch wll buy them. The mplcatons of these results for targeted enforcement are clear. A regulator that s motvated to fnd and penalze those frms wth hgher volatons has some justfcaton for targetng frms that buy permts, but has no reason to target frms on the bass of nformaton regardng ndvdual abatement costs. 9 Wth Model 2, each of the four nteracton terms contanng HghMAC s not statstcally sgnfcant. We also fal to reject the jont hypothess that HghMAC and these four nteracton terms are all equal to zero (Wald χ 2 (5) = 0.37, p = 0.996). 20

23 Although we fnd that permt buyers tend to have greater volaton levels, t does not necessarly follow that a regulator that wshes to drect hs enforcement effort to where t s most productve should target these frms. The estmates of the margnal productvty of enforcement n Table 6 reflect the senstvty of volaton choces by subject type to changes n enforcement pressure for a gven aggregate standard and ntal allocaton. The values n the HghMAC and LowMAC columns are the estmated margnal changes n ndvdual volatons as enforcement s ncreased. Of course, these margnal effects have to nclude both the drect effects of enforcement and the ndrect effects through the change n the permt prce. Lettng change n enforcement, gven an aggregate standard and ntal allocaton of permts, we calculated v / E+ ( v / p)( p/ E), = { HghMAC, LowMAC}, for each ncrease n E denote a enforcement, aggregate standard, and ntal allocaton of permts. v / E s the drect effect of the change n enforcement, whle ( v / p) ( p/ E) s the ndrect prce effect. Values for v / E and v / p were calculated usng Model 2 n Table 5, and p / E was calculated usng the prce equaton n Table 4. The last column of Table 6 contans the p-values for a test of the hypothess that the margnal productvty of enforcement s equal across subject types. 10 Note that the estmates of these margnal effects are consstently hgher for the HghMAC subjects when the ntal allocaton of permt s unform, but these effects are hgher for the LowMAC subjects when the ntal allocaton s not unform. However, none of the dfferences n margnal productvtes between subject types are statstcally sgnfcant. Thus, we have strong support for Hypothess 2 the margnal productvty of enforcement s ndependent of dfferences n subjects abatement costs 10 Although Table 6 compares subjects based on margnal abatement costs, the results would be dentcal f we nstead compared them based on whether they were predcted to be net sellers or buyers. Ths s because all HghMAC subjects are net buyers wth a unform allocaton and net sellers wth a nonunform allocaton (the reverse s true for LowMAC subjects). 21

24 and ntal allocaton of permts. Ths suggest that a regulator that seeks to drect hs enforcement effort to where t s most productve does not have a statstcally sgnfcant justfcaton for dong so on the bass of frm-level characterstcs. 4.3 Standards We do not expect ths no targetng polcy prescrpton to hold when permts are not tradable. Table 7 presents summary statstcs for our fxed standards experments wth a unform allocaton of both low and hgh aggregate standards; we dd not conduct any nonunform allocaton treatments. 11 Recall that these standards experments were dentcal to the permt market experments, except that subjects were not able to trade ther ntal permt allocaton. As predcted, HghMAC subjects tended to have greater volaton levels. The estmaton results n Table 8 confrm ths: the coeffcent for HghMAC subjects s postve and hghly sgnfcant. Moreover, the negatve and sgnfcant coeffcent for HghStandard ndcates that ndvdual volaton levels tend to decrease as the standard s relaxed. Therefore, we have strong support for Hypothess 3 and the enforcement mplcaton that a regulator who s motvated to fnd and penalze frms wth hgher volatons has justfcaton for targetng frms wth hgher margnal abatement costs and lower emssons standards. We use the nteracton effects of Model 2 n Table 8 to test for dfferences n the margnal productvty of enforcement. Smlar to Table 6 for the case of emssons tradng, Table 9 ncludes the margnal productvty of enforcement under fxed standards and presents the results of tests for the equalty of these values across subject types. Hypothess 4 posts that volatons of subjects wth hgher margnal abatement costs wll be more responsve to changes n enforcement. Ths s true wth our expermental desgn except n two cases. Wth a low standard, 11 Ths mples that we have one low and one hgh ndvdual standard for both HghMAC and LowMAC subjects. 22

25 our predcton s that the change from the low margnal expected penalty to ether of the medum margnal expected penaltes has the same effect on volatons for both subject types (see Table 7); ths s smply an artfact of the expermental desgn, n partcular that emssons choces are dscrete. As expected, when enforcement changes from Low to Medum(π H ) there s not a statstcally sgnfcant dfference n the margnal productvty of ths change across subject types (p = 0.78). However, there s the unantcpated sgnfcant dfference n margnal productvtes when enforcement changes from Low to Medum(π L ) (p = 0.02). More to the man pont of Hypothess 4, however, we fnd that margnal productvty of enforcement s sgnfcantly hgher for the HghMAC subjects for sx of the eght cases n whch we expect to observe ths dfference. Hypothess 4 also posts that the margnal productvty of ncreased enforcement s hgher when ndvdual standards (defned by the ntal permt allocaton) are lower. Observe n Table 9 that for each subject type the effect of each change n enforcement s greater when they face the low standard. However, these dfferences are statstcally sgnfcant n only two of the fve comparsons for the HghMAC subjects, and n only one of the fve comparsons for the Low MAC subjects. Although are results concernng Hypothess 4 are mxed, we do have strong support for the polcy recommendaton that, under fxed standards, regulators who are motvated to drect ther enforcement resources to where they wll be most effectve are justfed n targetng frms wth hgher margnal abatement costs. For the purposes of ths paper, however, the most mportant message s that whle frm-level characterstcs are mportant determnants of the effectveness of enforcement under command-and-control standards, the margnal productvty of enforcement does not depend on these characterstcs under a compettve emssons tradng program. 23

26 5. Conclusons Complance behavor n emssons tradng programs s fundamentally dfferent from complance behavor under emssons standards. Conceptually, the complance choces of rsk neutral frms that operate n a compettve emssons tradng program are ndependent of parametrc dfferences n ther abatement costs and ther ntal allocatons of permts. Therefore, regulators have no justfcaton for targetng ther enforcement efforts ether to fnd and punsh those frms that tend to hgher volatons, or to maxmze the productvty of ther enforcement efforts n reducng aggregate volatons. On the other hand, snce the complance choces of frms that face fxed emssons standards depend on ther abatement costs and the standards they face, regulators wll be motvated to target frms wth hgher margnal abatement costs and those that face lower standards. The results of our laboratory experments generally support the conclusons of theoretcal models of rsk neutral complance behavor n emssons tradng programs, but not completely so. We found that the subjects volatons are ndependent of dfferences n ther abatement costs, but ther volatons are not ndependent of how the ntal allocaton of permts determnes whch of them wll sell permts and whch wll buy permts. Those subjects that were predcted to buy permts tended to have hgher volaton levels than those who were predcted to sell permts. Despte ths, we found that ndvdual dfferences n abatement costs or ntal allocatons of permts do not have a statstcally sgnfcant mpact on the margnal productvty of ncreased enforcement n reducng ndvdual volatons. Concernng targeted enforcement, then, regulators may be motvated to drect ther enforcement efforts at frms that are net buyers of permts t they wsh to fnd and penalze the most sgnfcant volators, but there s no theoretcal reason or statstcal evdence that targetng frms based on frm-specfc 24