Environmental regulation and productivity: The case of electricity generation under the CAAA-1990

    Environmental regulation and productivity: the case of electricity generation under the CAAA-1990 Pedro Ignacio Hancevic PII: DOI: Reference:

S0140-9883(16)30266-3 doi:10.1016/j.eneco.2016.09.022 ENEECO 3453

To appear in:

Energy Economics

Received date: Revised date: Accepted date:

18 January 2016 14 September 2016 21 September 2016

Please cite this article as: Hancevic, Pedro Ignacio, Environmental regulation and productivity: the case of electricity generation under the CAAA-1990, Energy Economics (2016), doi:10.1016/j.eneco.2016.09.022

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

SC

RI

PT

Environmental regulation and productivity: the case of electricity generation under the CAAA-1990 By Pedro Ignacio Hancevic∗

NU

Centro de Investigaci´on y Docencia Econ´omicas

D

MA

September, 2016

Abstract

AC CE P

TE

This paper measures the impact of the 1990 Clean Air Act Amendments on coal-fired boilers’ productivity and output. The Act led to generating units adopting a number of different pollution abating behaviors, one of which was an input change to lower SO2 emitting coal. A key feature of the production technology is that each boiler is designed to burn a particular variety of coal, with significant deviations from the targeted coal characteristics resulting in productivity losses. Using data for the 1985-1999 period, I present empirical evidence of the policy impact. The main findings are that productivity declined between 1% and 2.5%, on average, and output losses ranged from 1% to 6% for affected boilers, varying across regions and over time. JEL codes: D24 L94 L51 Q51 Keywords: productivity, production function estimation, environmental regulations, sulfur dioxide, electricity generation, coal-fired boilers.

∗ E-mail:

[email protected]. Phone: +52 (449) 994-5150 ext. 5194. I would like to thank Amit Gandhi, Kenneth Hendricks, Alan Sorensen, Jean-Fracois Houde, and seminar participants at University of Wisconsin-Madison, University of California at Davis, Inter American Developmnet Bank, Universidad Javeriana, Colmex, CIDE, and the ICAE 2015 held in Milan, for their helpful comments and suggestions. All remaining errors are my own.

1

ACCEPTED MANUSCRIPT

1

Introduction

AC CE P

TE

D

MA

NU

SC

RI

PT

This paper evaluates the impact of the Title IV of the 1990 Clean Air Act Amendments (commonly referred to as Acid Rain Program, or simply ARP) on the performance of coal-fired electricity generating units during the first phase of the program, i.e. from 1995 to 1999. Historically, coal has been the main source of electricity generation in U.S. And in particular, it roughly accounted for 50% of the industry during the period of analysis. Additionally, coalfired generating units produced most SO2 emissions in an industry that explained more than 70% of total emissions in the country. Given the importance of coal in both energy and pollution production, understanding how the Act affected the behavior of coal-fired electricity generating units (EGU) in general and measuring the associated efficiency losses in particular, are central to the planning and analysis of current and future energy policies. Why might environmental regulations affect productivity? Under the ARP, EGUs had three main compliance alternatives: burn low-sulfur coal to guarantee an emission level below the permits allocated, purchase pollution permits to cover the excess of emissions, or retrofit a flue gas desulfurization system (a.k.a. scrubber) to reduce emissions up to 99%. A salient feature of the production technology is each EGU is designed to burn a particular type of coal. Coal is characterized by a number of attributes, including heat, sulfur and ash contents, moisture, grindability, among others. When a EGU departs significantly from its design coal specifications it suffers an efficiency loss. With regards to the SO2 regulations, if a EGU decides to switch from high- to low-sulfur coal, and if the characteristics of the cleaner coal differ substantially from those the unit was designed for, its productivity will likely decrease. Installing a scrubber system entails a different type of efficiency loss. While a scrubber allows a EGU to burn its ideal coal type, it consumes additional electricity which reduces the net output generated. Naturally, if the EGU simply buys pollution permits to cover the emissions that exceed the initial allocation and no other further action is taken, then its productivity will remain unaffected.1 In this study I provide empirical evidence for the impact of the first phase of ARP on productivity and output. I carefully estimate a production function augmented to include two additional features that affect the total factor productivity (TFP). One is a variable that quantifies the relative deviations from each EGU’s unconstrained (ideal or business as usual) coal type. The other is a dummy for scrubber usage to account for the electricity losses explained 1 During

the period of analysis fuel switching was the most popular compliance strategy, with less than 10% of affected coal-fired units adopting a scrubber.

2

ACCEPTED MANUSCRIPT

1.1

SC

RI

PT

by these devices.2 Also, since the sample period was marked by market restructuring in several states, I incorporate controls to capture these potential confounding effects. I find that more than 90% of the targeted units were affected by the environmental regulation, with output losses ranging from 1% to more than 6% and productivity declining between 1% and 2.5%, on average. The magnitude of the impact depends on the region where power plants are located. Market restructuring and deregulation policies had a small and positive effect on TFP.3

Related literature

AC CE P

TE

D

MA

NU

The ARP has extensively been studied from several different angles (see for example Carlson et al. [11] or Schmalensee and Stavins [33]). This paper is closely related to studies that evaluate the impact of environmental regulations on the efficiency of power utilities. Some of these papers, like Sueyoshi et al. [34], Bernstein et al. [7], or Yaisawarng and Klein [35], use data envelopment analysis or best-practice frontier techniques to evaluate the performance of electric power plants under different environmental regulations. A seminal paper by Gollop and Roberts [19] measures the effect of SO2 emission restrictions on the rate of productivity growth in the U.S. electricity generation industry over the period 1973-79. They propose a firm-specific measure of regulatory intensity, which combines the severity of the emission standard, the extent of enforcement, and the unconstrained emission rate relevant to the firm. In this paper, I construct a boiler-level measure of regulatory intensity akin to Gollop and Roberts’ original approach, but incorporate the variable directly into the production function, and therefore, I am able to directly measure the effect on productivity. I also add to the literature of production function estimation by exploiting a rich data set that allows me to solve the ‘transmission bias’ problem (i.e. the fact that firm’s productivity is transmitted to firm’s optimal choice of inputs giving rise to an endogeneity problem) by using exogenous input prices as instruments. A significant part of the production function estimation literature is concerned with identification and endogeneity issues, potential multicollinearity of production factors, and with the selection bias that emerges when there are many episodes of firms exiting the market. These topics where first addressed by Marschak and Andrews [29] and, more recently, by Ackerberg et al. [2], Gandhi et al. [18], Olley, and Pakes [31], Levinsohn, 2A

natural concern about potential endogeneity problems with these two variables arises, since unobserved productivity could in principle be related to scrubber adoption and fuel switching decisions. Appendix B addresses this matter. 3 Environmental goals of ARP were successfully achieved during this period, a result that is not surprising since cap-and-trade programs restrict aggregated emissions to be, on average, no greater than the total number of allowances. This paper do not analyze the environmental goals of the ARP.

3

ACCEPTED MANUSCRIPT

4 If

AC CE P

TE

D

MA

NU

SC

RI

PT

and Petrin [28], Blundell and Bond [8], among others. Focusing specifically on the U.S. electric power industry, the empirical attempts to estimate a production function are relatively scarce when compared to other industries. Instead, most of the existing papers have generally estimated cost functions, with an early exception being Nerlove [30]. An advantage of the cost function estimation approach is it typically avoids the endogeneity problem: the error term (the unobserved productivity shock) is usually not present in the cost function.4 Most authors justify using this approach by arguing that power plant managers minimize costs in a context of cost-of-service regulation and take the quantity to be produced as given. However, this argument is less valid under the new regulation regimes of the 1990’s, since several states have deregulated or restructured wholesale electricity markets and those that are still regulated have a certain degree of flexibility in the amount of electricity each unit is called to produce.5 Further, cost function estimation requires additional restrictions on the data,6 such as homogeneity conditions, which are not needed when estimating a production function.7 The production function model of this paper incorporates a hedonic approach which is applicable to environments where characteristics and quality of inputs matter and where a discrete classification of different varieties of inputs is hard to handle in practice and unrealistic. By distinguishing among different groups of units that were differentially affected by the SO2 regulations at different points in time and exploiting the rich heterogeneity of units’ characteristics, I am able to illuminate how environmental policies affect EGUs and to quantify the impact of the first phase of the ARP in particular. The remainder of this paper is organized as follows. Section 2 describes in more detail the market and the environmental regulation under analysis. Section 3 presents the empirical prooutput level is indeed adjusting and firms are constrained to produce a fixed amount, the endogeneity with output at the boiler level would be a worse problem. 5 Cost minimizing assumptions are not superior than regular profit maximizing assumptions for electric utilities. If we think that a plant manager in a regulated market only tries to reduce costs given some fixed electricity price, it is also likely that she will not make a big effort to achieve that goal either. By definition, effort is costly to her and given the usual regulatory design she will be able to include any extra cost in the next rate case, where many of those cost are presumably inefficient. Additionally, it could be also true that some timing pattern exists regarding the amount of effort the manager makes along each rate case. Effort could vary depending upon the proximity of a rate review, making the cost minimizing assumption difficult to handle in practice. A recent paper by Abito [1] explicitly models managers’ behavior in a context of firms’ cost optimization under environmental regulations. 6 An exception is Atkinson and Halvorsen [4], which develops a generalized cost function approach that does not need implausible assumptions about cost minimization behavior. 7 Some papers that follow the dual approach are Bushnell and Wolfram [9] which investigates the changes in operating efficiency at plants that have been divested from utility to non-utility ownership, and Fabrizio, Kira, and Rose [15] which studies efficiency of firms that were deregulated.

4

ACCEPTED MANUSCRIPT

PT

duction function model while Section 4 shows and discusses the estimation results. Section 5 evaluates the impact of the policy in question. Finally, Section 6 concludes.

Historical background and market overview Sulfur dioxide regulations

SC

2.1

RI

2

AC CE P

TE

D

MA

NU

In response to growing concerns over SO2 emissions in the 1970’s, Congress passed two important amendments to the Clean Air Act. The first amendments, enacted in 1970, established The New Source Performance Standards (NSPS). Starting in 1971, the NSPS required a subset of large, new EGUs to keep emission rates below 1.2 pounds of SO2 per million Btu (lbs/mmBtu). There was no restriction on the compliance method and EGUs were able to either install scrubbers or burn low-sulfur coal. However, with the amendments passed in 1977, the NSPS were modified in such a way that all new EGUs constructed after 1978 were (virtually) forced to install scrubbers. Consequently, many utilities reacted to the NSPS by extending the lifespan of older EGUs and postponed the construction of new ones. The high levels of pollutants in the air during the eighties were mainly caused by the older units excluded from the NSPS. As a result, new amendments to the Clean Air Act in 1990 targeted those older units. The ARP of 1990 established a nationwide market for allowances where affected EGUs had to offset emissions with pollution permits at the end of each year. It departed from previous regulations in that the method of pollution abatement was flexible and the amount of abatement for each EGU was not predetermined.8 As a result, some of them opted for blending different fuels incorporating a higher proportion of low-sulfur coal in their mix, some others chose to retrofit a scrubber or reduced the output level (trimming strategy)9 , while others simply bought allowances from other polluting sources.10 8 The

NSPS were essentially overridden under the ARP, although existing sources still have to comply with command-and-control state regulations. 9 Utilities choosing to drastically reduce output from a unit reached by Phase I of the ARP were required to incorporate a “substitute unit”, i.e. a different unit initially not subject to Phase I. 10 See for example Ellerman et al. [14] for an analysis of the behavior and performance of the market for emissions permits, quantification of the emission reduction, compliance costs, and cost savings associated with the trading program during the first three years of the program. Also, see Schennach [32] for a complete analysis of the economics of pollution permit banking. Frey [17] in turns studies the scrubber adoption choice in a context where cap-and-trade programs (Acid Rain Program) overlaps complex and stringent command-and-control regulations (CAAA-1977). Keohane [24] and [26] compare pollution control techniques choices by regulated firms under emissions standards, emissions taxes, and cap-and-trade schemes.

5

ACCEPTED MANUSCRIPT

Market restructuring

MA

2.2

NU

SC

RI

PT

The ARP was implemented in two phases. Starting in 1995, Phase I originally required the reduction of emissions to a group of 263 large and dirty boilers. Each affected unit was allocated a number of permits sufficient to achieve an emission rate equal to 2.5 lbs of SO2 per million Btu according to the unit’s average heat input registered during the 1985-1987 period. One permit entitled the unit to emit one ton of SO2 and could be used in the allocated year, banked for future use, or traded with other firms. The second phase of the program began in 2000 and included all units larger than 25 MW of nameplate capacity. Under Phase II, each unit was allocated permits sufficient to emit at a rate of 1.2 lbs of SO2 /mmBtu, once again using the base years 1985-1987 for the allocation computation.

AC CE P

TE

D

Until early 1990s the U.S. electricity market was predominantly dominated by vertically integrated monopolies over generation, transmission and distribution of electricity. State regulators determined electricity prices based on accounting costs of service. The main concern at that time was the high electricity prices observed in some states. As a result, several market restructuring and deregulation episodes occurred.11 Fabrizio, Rose, and Wolfram [15] tests whether thermal power plants became more efficient after market restructuring occurred. They found some evidence that labor and materials were used more efficiently after deregulation but could not find evidence for fuel input. A possible reason could be that fuel consumption decisions are usually made at the boiler level rather than at the plant level. Boilers differ in technology and are designed to burn a particular type of fuel. Since most facilities have more than one boiler, several ‘fuel efficiencies’ might coexist in a single plant producing blurry estimates for fuel input demand. Market restructuring episodes and environmental regulations might be related. Since the goal of this work is to identify the impact of the ARP on units’ performance, I necessarily have to control for market restructuring effects.12 11 Several

papers discussed the incentives of IOU under different regulatory schemes. See for example Joskow [23], Hendricks [22], Laffont and Tirole [27], and Cicala [12]. 12 Fowlie [16] studies the NO Budget Program under the Title IV of the CAAA-1990. It found some evidence x that deregulated IOU plants operating in restructured electricity markets were more reluctant to adopt capital intensive environmental compliance options than regulated IOU plants or publicly owned plants.

6

ACCEPTED MANUSCRIPT

2.3

Data

Coal-fired generating units

SC

2.3.1

RI

PT

This paper utilizes data from different sources: the Energy Information Administration (Form 767), the Federal Energy Regulatory Commission (Form 423 and Form 1), the Rural Utilities Service, and the Environmental Protection Agency.

D

MA

NU

For expositional purposes, the relevant set of EGUs can be divided into two groups. The ‘treated group’ comprises the affected unscrubbed EGUs that were regulated under Phase I of the ARP (a.k.a. ‘Table A’ units) and those boilers that opted in as substitute or complementary units during the 1995-1999 period. The ‘control group’ includes Phase II EGUs (i.e. units built before 1971 that did not enter during the first phase); NSPS units (i.e. units built after 1971 with capacity larger than 73 MW); and all those EGUs that already had a scrubber system at the time of the implementation of the ARP.

AC CE P

Variable

TE

Table 1: Summary statistics for treated and control groups in 1990

Nameplate rating (MW) Capacity factor Age Heatrate Share of coal Coal heat content (btu/lbs) Coal sulfur content (%) Coal ash content (%) Labor(1) (# workers) Investor owned utility (%) Market restructuring (%)

Treated (n=297) mean sd 299.5 0.540 29.6 10.67 0.992 11761.1 2.1 9.8 74.7 84.2 37.3

(254.6) (0.190) (7.9) (1.74) (0.028) (960.3) (0.8) (2.1) (51.1)

Control (n=500) mean sd 352.4 0.507 26.5 11.12 0.954 10968.7 1.1 9.2 89.9 76.2 30.6

(268.4) (0.203) (11.7) (4.67) (0.082) (1939.0) (0.8) (3.4) (69.6)

Source: Form 767, EIA. (1) The original variable is at the plant level. I impute labor to each boiler according to its share in the plant’s annual net generation. Table 1 shows some descriptive statistics for coal-fired EGUs in 1990.13 Control units are, 13 The

variable share of coal in table 1 refers to the ratio between heat obtained from coal and the total heat

7

ACCEPTED MANUSCRIPT

Mine locations and coal characteristics

NU

2.3.2

SC

RI

PT

on average, larger, younger, and burn cleaner coal varieties than treated units. Since there is a trade-off between heat and sulfur contents, control units consume coal varieties with lower heat content. Nevertheless, the average heat rates (i.e., the ratio between the amount of energy consumed and the net electricity generated) are very similar in both groups. The quantity of labor used is relatively similar in the two groups of EGUs. Finally, there are no major differences in terms of the percentage of boilers operated by investor-owned utilities (as opposed to publicly- or cooperatively-owned utilities), and the percentage of EGUs located in a state that deregulated or restructured its electricity generation market.

MA

There are four main coal producing regions: Western Region, Powder River Basin (PRB), Interior Region, and Appalachian Region. Additionally, most coal plants are located in the eastern half of the country (see Figure 1).

AC CE P

TE

D

Figure 1: Coal production regions and power plant locations

Source: Department of Energy, Energy Information Agency (EIA)

Coal quality and sulfur content vary both across and within regions. These variations are key elements explaining the heterogeneity in SO2 abatement costs. Using data from the FERC’s Form 423, Table 2 presents some descriptive summary statistics about monthly coal shipment data at the plant level for the years 1985-1999. Central Appalachia -parts of West Virginia, obtained from coal, petroleum, and natural gas. In my sample, just a few EGUs blend coal and natural gas or petroleum (i.e., co-firing), and the proportion of fuel used other than coal is less than 15%, in all cases.

8

ACCEPTED MANUSCRIPT

NU

SC

RI

PT

Kentucky, Virginia and eastern Tennessee- was the primary coal-producing region during the period of analysis. This region produces both medium and low-sulfur coals. After railroad deregulation in the 1980s, the major source of low-sulfur coal has shifted to the PRB -Montana and Wyoming-. Northern Appalachia -parts of Ohio and Pennsylvania- and the Interior Region -mainly Illinois and Indiana-, are prime sources of high-sulfur coal. Other states like Alabama, Texas, and Oklahoma, and also states in the West North Central Region have supplied medium and high-sulfur coals during the sample period, however, their production has steadily declined since early 1990s.

MA

Table 2: Coal characteristics by region (1985-1999) Heat (btu/lbs) mean sd

Ash (% in wt) mean sd

Sulfur (% in wt) mean sd

12310 12351 11234 9036 12227 6362 10960

(778) (634) (536) (933) (372) (690) (651)

11.2% 10.5% 9.1% 7.0% 12.2% 16.5% 12.3%

2.4% 1.5% 2.4% 0.4% 1.2% 1.0% 3.9%

Coal production region

Underground

Spot market

Emission rate (lbs/mmbtu) mean sd p05 p95

Northern Appalachia Central Appalachia Interior Powder River Basin - West AL TX and OK West North Central

58.4% 58.2% 49.0% 9.1% 60.3% 0.0% 0.0%

29.9% 24.3% 25.3% 16.2% 11.3% 0.0% 21.4%

3.77 2.42 4.12 0.79 1.82 2.48 6.82

AC CE P

TE

Northern Appalachia Central Appalachia Interior Powder River Basin - West AL TX and OK West North Central

D

Coal production region

(4.2%) (2.9%) (1.8%) (4.2%) (1.5%) (4.1%) (3.3%)

(1.67) (1.77) (1.50) (0.31) (0.93) (1.09) (1.19)

1.85 0.98 1.62 0.40 0.80 1.23 4.76

(1.0%) (1.1%) (0.9%) (0.2%) (0.6%) (0.4%) (0.6%)

6.82 6.09 6.33 1.49 3.38 5.10 8.32

Source: Form 423, FERC.

The trade off between sulfur content and heat content is remarkable for the PRB coal. In contrast, it is possible to find medium- and low-sulfur coals with high Btu content in the Central Appalachia Region. While PRB coal comes mainly from surface mining, Northern and Central 9

ACCEPTED MANUSCRIPT

NU

SC

RI

PT

Appalachia and the Interior coals are almost equally extracted from underground and surface. Typically underground mining is more costly than surface mining because of the larger amount of physical capital and labor needed. Finally, spot purchases represented roughly 20% during this period, so medium and long-term contracts were the most usual type of transaction. Figure 2 illustrates the evolution of delivered coal costs in nominal cents per million Btu. The costs of both coal types (i.e. low- and high-sulfur) declined during the period of analysis.14 Low-sulfur coal was initially more expensive. Then, after a period of relatively similar costs, low-sulfur coal gradually became more expensive at the time of Phase I started. Finally, there was a faster drop in the cost of low-sulfur coal (see Busse and Keohane [10]).15

Figure 2: U.S. Delivered Coal Costs: Low-Sulfur to High-Sulfur Ratio

MA

(nominal prices in cents per million btu)

(a) Delivered Coal Costs

1.2

TE

D

180

(b) Ratio

120

Delivered Coal Cost Ratio

140

1.1

AC CE P

Delivered Coal Cost

160

.9

.8

100

.7 1985

1990

1995

2000

Low−sulfur

1985

1990

1995

2000

High−sulfur

14 Transportation 15 The

1

costs declined after rail rate reductions, facilitating long distance shipments. share of low-sulfur coal in total coal consumption increased from 33.6% to 53.4% between 1985 and

1999.

10

ACCEPTED MANUSCRIPT

2.3.3

Deviations from unconstrained coal type

AC CE P

TE

D

MA

NU

SC

RI

PT

EGUs are designed to burn a specific variety of coal. The crucial point is that many factors -for instance, market conditions, transportation costs, coal availability- that were originally considered when choosing the coal variety at the time the boiler is constructed may change along with other relevant institutional facts. Environmental regulations constitute a clear example of how a coal variety initially optimal from the technology/technical perspective could eventually be substituted for a different variety once the policy is implemented.16 There is a capital cost of converting a boiler to burn lower-sulfur coal. This cost depends on the differences in quality between the high- and low-sulfur coals used. For example, a switch made with coals purchased from Central Appalachian is not very costly to the unit since both high- and low-sulfur coals extracted from this region are quite similar in terms of quality. On the other hand, switching from Central Appalachian high-sulfur coal to PRB low-sulfur coal -a variety very different from eastern coal- would require a much larger capital investment to retrofit the boiler. In order to capture the main mechanism through which SO2 regulations impact unit’s total factor productivity, I create a variable that measures the distance from the actual coal used relative to the unconstrained ideal coal type for each unit. I define R -hereafter referred to as relative deviation- in the following manner: E ∗ − E jt j R jt = ∗ Ej

(1)

where E jt is the observed emission rate and E ∗j is the free emission rate. Emission rates in turn are calculated in pounds of SO2 per million Btu, which is an implicit characteristic of each variety of coal. EIA’s Form 767 provide data on heat and sulfur contents at the unit level on a monthly basis. I am able then to compute monthly emission rates for each EGU. The emission rate observed in month m is simply the ratio between the sulfur content and the heat content multiplied by a conversion factor (a scalar).17 Since the format of other data used in this analysis is on an annual basis, I compute annual average emission rates, E jt , for each boiler. The unconstrained emission rate, E ∗j , is computed using the median emission rate from the 16 In

extreme cases, it is also possible to observe a EGU reconverting its technology to burn a different fuel (e.g. natural gas) or even exiting the market. See for example Dardati [13] for an empirical paper that evaluates exit/entry decisions in the power generation industry comparing different environmental schemes in US and Europe. 17 To calculate emission rates I use the conversion formulas proposed by the EPA.

11

ACCEPTED MANUSCRIPT

RI

PT

pre-regulation period 1985-1989.18 It is precisely the stability of E jt observed during this time frame what validates the approximation of E ∗j using these years as a benchmark (see Figure 7 in Appendix C which shows the distribution of the relative absolute deviation at the boiler level, |E jt −Med(E jt )| , for t = 1985, ..., 1989). Med(E jt )

AC CE P

0

TE

D

MA

Relative deviation (variable R) .05 .1 .15 .2 .25 .3 .35

.4

NU

.45

SC

Figure 3: Evolution of the relative deviation measure (Mean values over treated and control units)

1985

1987

1989

1991

Treated units

18 There

1993

1995

1997

1999

Control units

are at least two alternatives for the computation of the unconstrained emission rates which in my opinion are less desirable. One method would be to use the emission rate for the least expensive fuel type purchased by the plant or a set of plants owned by the same utility in a certain area. This alternative violates the observation that there are technological constraints and also that the technically ideal type of fuel is not necessary the less expensive one. A second alternative is to use some summary measure (e.g. the average) of emission rates of high-sulfur coals, i.e. only including coals with an emission rate above 1.2 lbs SO2 /mmBtu. Once again, this alternative does not contemplate the technological constraints faced by the boiler. Some units might have been designed to burn low-sulfur coals.

12

ACCEPTED MANUSCRIPT

AC CE P

TE

D

MA

NU

SC

RI

PT

Figure 3 presents the evolution of measure R over time for the control and treated groups described previously. In the case of treated EGUs, a pronounced rise in R corresponds to the implementation of the first phase of the ARP. Fuel switching arose as the most popular compliance strategy for this group. In the case of control units, R remained relatively flat during the entire sample period, although a slight upward trend was also present. Until the early 1990s, R fluctuated between 2% and 4.5%. By the end of the decade, R fluctuated between 5.5% to 8.5%. One can then conclude that control units did not behave as the typical control group used in experiments or program evaluation exercises. The reason for that is, however, simple. All EGUs, not only Phase I, must comply with the second phase of the ARP from 2000 onward. In order to cover emissions exceeding the allocated allowances (if any), some control units opted for fuel switching. Notice that this behavior started ahead of time for both groups: some treated EGUs started to switch fuels in the early 1990s (i.e. before first phase effectively began), while some control EGUs did it so in the mid- to late 1990s (i.e. before the second phase began). A conjecture behind this is that the fuel switching strategy implies the realization of capital investments, and this is especially true when the departure from the “ideal” coal type is significant. Boilers’ conversion works take some time to be completed and new production processes need some time to be fully learned.19 Additionally, generating units could need to be off line as any sort of retrofit investment might have to be phased in due to reliability issues or availability of qualified labor for the required retrofits. Notice that during the period 1985-1989 the value of R is not exactly zero. The variables that characterize coal (in particular btu and sulfur contents) have a continuous range. Using the pre-regulation period as a benchmark and by a simple rule of thumb, any value of the relative deviation below 4% should not be attributed to a response to the ARP and should be consider as an ordinary fluctuation around the unconstrained emission rate instead.20 Why not consider modeling discrete number of varieties? As stated before, coal characteristics are continuous and there are overlaps within and between coal producing regions. Moreover, from an empirical point of view most EGUs purchased coal from the same region before and after the implementation of the ARP. Even though the average sulfur content of coal purchased decreased with the program, most units kept burning coal of similar quality to avoid the realiza19 Although

not totally discarded, the idea that R increased for control units due to the fact that compliance coal became less expensive than high-sulfur coal is not completely appealing. As shown in figure 2, the relative cost changed from 1985 to 1999 only 17%. That number is certainly not enough to justify costly investments that are needed to retrofit the boiler-generator pair. 20 Plant level data from FERC’s Form 423 reveal that two different shipments coming from exactly the same coalmine in a given month present certain differences in terms of coal characteristics. If heat and sulfur contents differ, the corresponding emission rate also differ by construction.

13

ACCEPTED MANUSCRIPT

Model: the production function

MA

3

NU

SC

RI

PT

tion of capital investments.21 Another concern regarding the validity of the relative deviation as a measure that explains departures from the ”ideal” coal type arises when considering the evolution of the relative prices of high and low sulfur coals that were available to each firm. In other words, if big changes in relative prices were observed, the notion of “technically ideal” coal type should be replaced by a notion of “economically ideal” coal type. As it was shown in Figure 2, relative price changes naturally existed since the main implication of the environmental policy was the fuel switching phenomenon. However, the effect of a higher demand for low sulfur coal, in particular the varieties extracted from the western region, was cushioned by significant decreases in coal transportation rates.

AC CE P

TE

D

The relevant economic agent in the model is the pair boiler-generator, henceforth referred to as unit, EGU, or simply boiler. I only consider boilers and generators associated with each other in a one-to-one relationship. This simplification allows me to measure inputs and output for each unit in a straightforward and accurate way. The EGU’s output, Y , is measured by net annual megawatt-hours. The inputs entering as arguments in the production function are capital, labor, and total heat -K, L, and H, respectively.22 Capital input is measured by the generator’s nameplate capacity in MW.23 Capital decisions are made in advance of output decisions, essentially at the time of the unit construction or retirement. Less frequently, some capacity adjustments occur once (twice at most) during the unit’s lifespan. Labor is reported at the plant level in my sample, and the allocation among units in a given plant is not perfectly delimited. Because environmental regulations are enforced at the boiler level, I am interested in measuring output at the unit level rather than at the plant level. Therefore, I impute labor to each boiler according to its share in the plant’s annual net generation. Labor input is of course used for different tasks in a power plant; some tasks are shared by all 21 Table

7 in Appendix C presents a transition matrix for coal purchases using the seven coal producing regions defined in this paper. To compute the matrix, I considered four sub-periods: 1985-1989; 1990-1994; 19951999; and 2000-2005. The matrix shows certain stability for units buying from Appalachian Region (Central or Northern), and from PRB-West Region. However, many units originally buying from the Interior Region switched to PRB-West after the implementation of the environmental program. 22 Other potential inputs like materials are not available in my sample. 23 Capacity could be measure by the maximum steam flow capacity of the boiler. It turns out that the correlation between maximum steam flow capacity and nameplate capacity is above .99, so the distinction between them is irrelevant.

14

ACCEPTED MANUSCRIPT

MA

NU

SC

RI

PT

units, but most of them are unit-specific tasks. For instance, tasks related to the unloading of the railcar (or barge), coal handling, coal burning, ash removal, maintenance of the unit and of auxiliary equipment, and pollution abatement equipment operation are all specific to the unit.24 Although labor input is partially decided in advance of output decisions, it still could be correlated to productivity shocks generating an endogeneity problem. Fuel decisions are made in real time in response to dispatching and operation changes. According to the production function literature, I consider fuel as a flexible, non-dynamic input, whereas capital is an inflexible, dynamic input. Labor input lies in between these two extreme cases. The variable H jt corresponds to the total amount of heat used by boiler j to generate electricity in year t and is expressed in million of British thermal units (mmBtu). Suppose there are m fuel types or varieties, then boiler j’s total heat used is H jt = θ1 X jt1 + ... + θm X jtm , where X i is the quantity of fuel used and θi is the heat content per unit of fuel i, for i = 1, ..., m. The generic production function is as follows:

D

T

(2)

TE

Y jt = F(K jt , L jt , H jt ) exp{γ0 + γR R jt + γ f f jt +C jt γc + ε jt }

AC CE P

The variable R is the measure of relative deviation from the unconstrained coal variety introduced before. The dummy variable f is equal to one when the unit has an active scrubber system and equal to zero otherwise. C jt is a vector with year dummies and controls (e.g. institutional variables accounting for different regulatory regimes, ownership types, environmental regulations, and any other relevant characteristics of coal-fired generating units that potentially affect the TFP). My main interest is in the TFP term (i.e. the exponential part in expression 2). I choose a Cobb-Douglas form for the function F(·).25 For estimation purposes, the concrete 24 Maintenance

and repair expenditures may be inversely related to the net amount of electricity generated since most of the times the unit needs to be out of service for major works. However, we do not have access to data on these variables at the boiler level. A very rough estimation at the plant level implies that materials and maintenance expenditures can be retrieved as the residual of non-fuel expenses once labor costs are subtracted. This quick estimation indicates that operating costs other than labor and fuel represent (on average) 2.5% of the plant’s total revenues from electricity sales. 25 The Cobb-Douglas form makes the interpretation of input coefficients (elasticities) straightforward. However, it imposes the constraint that elasticity of substitution are all equal to one. Other functional forms provide more flexibility at the expense of additional assumptions and extra regularity conditions that need to be properly checked. I also estimated a translog production function (results available upon request) and no major changes were obtained in the coefficients of interest (i.e. those associated with variables R and f ). Separability restrictions (global, linear and nonlinear) were also checked, and the corresponding test results indicated Cobb-Douglass functional form is a suitable choice (see Berndt and Christensen [5] and [6] for a formal treatment of these issues).

15

ACCEPTED MANUSCRIPT

form for equation 2 is T

(3)

PT

y jt = γ0 + αh h jt + αk k jt + αl l jt + γR R jt + γ f f jt +C jt γc + ε jt

3.1

AC CE P

TE

D

MA

NU

SC

RI

where y = ln(Y ), l = ln(L), k = ln(K), and h = ln(H). Notice that productivity enters in a Hicks-neutral fashion. This representation of the production function reflects the assumptions made about the underlying economic model of production. For example, one may think of log-productivity as if it were composed of two components: an unobserved shock (η) which realizations occur after output decisions are made, and a second component (ω) that is anticipated and observed by the firm (but not by the researcher) at the beginning of each period and before output decisions are made. Additionally, one may assume further that the anticipated log-productivity term is fixed over time for each unit, so a panel fixed effect model can be used, i.e. ω jt = ω jt−1 for all t. An alternative representation for ω assumes anticipated productivity evolves according to some dynamic process, e.g. a first order Markov process in which case ω jt = φ(ω jt−1 ) where φ(.) is some arbitrary function to be determined. Another possibility is to contemplate some serial correlation at the unit level for the unexpected innovation term (η) -assuming for example an ith order autoregressive process- and include unit fixed effects for the anticipated part, i.e. ω j constant over time. I will explain below the specific assumptions about the error term and discuss the estimation strategy.

Production function estimation

Variables R and f are key pieces of the model. They help explain the impact of the ARP on productivity and are, by construction, choice variables. A natural question that emerges is whether these variables are endogenous. Concretely, are more productive boilers more prone to adopt scrubbers ( f = 1) or to switch fuels (R > 0)? This concern is theoretically addressed in Appendix B. The maintained assumption throughout the paper is that both variables are econometrically exogenous. From the description of production inputs made before, K is unlikely to be correlated with log-productivity shocks. L is certainly more flexible than capital, it is correlated with current output and (partially) responds to ε. Similarly, fuel input decisions are contemporaneous to productivity shocks. If it is true that output decisions are made after the unit observes its efficiency, then the manager may increase the quantity of fuel and labor used in response to positive productivity shocks. Although some portion of labor input is decided in advance of 16

ACCEPTED MANUSCRIPT

The construction of appropriate instruments for fuel input

AC CE P

3.1.1

TE

D

MA

NU

SC

RI

PT

productivity shock realizations (usually based on previously made seasonal forecasts), I cannot     dismiss the possibility that E ε jt |l jt 6= 0 and that E ε jt |h jt 6= 0. Some of the commonly proposed estimation methods to tackle the simultaneity problem are: instrumental variables; panel data models (including first differences, individual dummy variables, fixed effects, and random effects models); correlated random effects model (which combines the attributes of fixed effects and random effects models); different approaches based on GMM26 ; control function techniques27 ; and methods that exploit the first order conditions of a profit maximizing firm.28 I adopt a GMM approach and use two-period lagged values of labor input as instruments for current values of this variable. The formal assumption is   that E ε jt |l jt−2 = 0.29 Additionally, exogenous coal prices at the plant level are used as instruments for fuel input. Coal prices are likely to be highly correlated with fuel input demand but uncorrelated, for instance, with how efficiently a boiler burns fuel to heat the water that produces saturated steam to generate electricity. Another concern is the selection bias that may arise if the decision to retire a boiler is driven by the realization of unobserved productivity shocks. In our sample, the selection bias does not seem to be particularly problematic. Electric utilities operated in a more stable environment than firms in other industries where the selection problem is more acute.30,31

A major critique of using input prices as instruments to estimate production functions is that input prices usually do not have enough exogenous variation across firms (or boilers in this paper). If all EGUs face the same coal price there is little hope for identifying the parameters of interest using IV or GMM techniques. However, it turns out that delivered coal costs are highly dependent on the location of the unit, or more precisely, on the distance from power plants to coalmines. This is especially true for low-sulfur coal extracted in the PRB and for those units located in the West Region. The variation in heat and sulfur content among coals 26 See

for example Arellano and Bond [3], and Blundell and Bond [8] for example Olley and Pakes [31], Levinsohn and Petrin [28]; Ackerberg, Caves and Frazer [2] 28 See for example Gandhi, Navarro and Rivers [18], and Zhang [36]. 29 Some autocorrelation of order one is observed for the error term. Hence, the use of two-period lags for labor input minimizes the possibility of having an endogeneity problem. 30 The problem of selection bias is explored in depth in the seminal paper by Olley and Pakes [31] which studies the telecommunication equipment market in United States. 31 Both problems mentioned before, i.e. transmission bias and selection bias, have been largely studied in the literature. The former, however, has received more attention because its solution is somehow easier to handle in practice. 27 See

17

ACCEPTED MANUSCRIPT

Identification strategy

D

3.1.2

MA

NU

SC

RI

PT

from different regions is also a key source of heterogeneity in coal prices. In general, a unit does not buy different varieties of coal from the same region and different varieties of coal from different regions all at once.32 Hence, in order to have a set of exogenous coal prices for each boiler I must estimate and impute coal prices econometrically. Following a strategy similar to Keohane [24], I estimate price regression equations for coal from each region presented in table 2. In the estimated regressions I only use spot transactions to reflect the true available prices each unit face at each point in time (Appendix A presents a sample price regression for the Central Appalachian Region). Based on the corresponding regression coefficients I generate predicted coal prices at the boiler level for coals from appropriate districts using the corresponding average sulfur, ash, and btu contents. Based on the characteristics of the unit, each unit is left with the least expensive low-sulfur coal and least expensive high-sulfur coal available. Those coal prices, the two-period lag of labor and the interactions of those variables are the excluded instruments used to estimate equation 3.

AC CE P

TE

There is a lot of temporal and spatial heterogeneity among the EGUs that I use to identify the coefficients of interest in the production function. In my sample, there are hundreds of units each with its own legal and institutional environment. Pollution reduction mandates and electricity market restructuring, however, were not randomly assigned to the units. Regarding SO2 regulations, both the assignment of emission standards (CAAA-1970) and the allocation of allowances (CAAA-1990) critically depended on boilers’ characteristics and their geographical location -e.g. the concentration of SO2 emissions from air contaminant sources. With regards of market restructuring episodes, earlier works suggest that they are correlated with higher electricity prices in the cross section.33 To address these concerns, I exploit a rich panel data on output, inputs and characteristics describing most EGUs during an extensive period of time. My panel includes observations before and after the implementation of both market restructuring and ARP. I incorporate two control variables similar to those used in Fabrizio et al. [15].34 The first is a binary variable equaling one when the generating unit belongs to an IOU and is located in a State that restructured its electricity market (=1 since the year of the first formal hearing). 32 Other

variables that account for coal price variation are: ash and moisture contents, the type of purchase (spot or contract), shipment type (rail, barge, truck), among others. 33 See for example Bushnell and Wolfram [9] for a detailed analysis. 34 The construction of the two market restructuring dummies was based on information provided by Status of Electricity Restructuring by State, EIA.

18

ACCEPTED MANUSCRIPT

Estimation results

SC

4

RI

PT

The second control is a dummy that takes on a value of one if the unit is publicly owned after 1992, and zero otherwise. Hence, the ‘treated group’ in this case is composed of restructured IOU units, while the ‘control group’ consists of non-restructured IOU units and publicly (or cooperatively) owned units.35

AC CE P

TE

D

MA

NU

In this section, I present the results from estimating the production function in equation 3. I implement a two-step GMM approach using the set of excluded instruments described in the previous section and compute robust standard errors clustered at the boiler level. This estimation strategy solves the endogeneity problems mentioned before and is more flexible than other alternatives that assume some specific autoregressive process for the unobserved productivity term. I also avoid making stronger assumptions such as cost minimization or profit maximization behavior, or strict monotonicity conditions for auxiliary functions (e.g. investment function, materials demand function, etc.) which are typically needed to ‘invert out’ the unobserved productivity shocks. Table 3 presents the estimation results for one OLS and four GMM specifications. Column GMM-1 estimates the model using dummies that assume the value one for Phase I units and Phase II units once the CAAA-1990 was passed and after it was effectively implemented. Column GMM-2 incorporates the relative deviation measure, R. Column GMM-3 has the two market restructuring dummies described in the previous section. Column GMM-4 includes both the relative deviation measure and the market restructuring dummies, and column OLS-4 presents the OLS estimates using a similar specification. All the models include a scrubber dummy. Some comments about the estimation results. First, scrubbing and fuel switching negatively affect boiler productivity and output. Both coefficients γ f and γR are negative and statistically significant, indicating environmental policies of the sort analyzed in this paper do matter for TFP. I analyze this result in more detail in section 5. Second, there is some evidence of a positive market restructuring effect. However, its magnitude is relatively small (see models GMM-3 and GMM-4). This result is in line with Fabrizio et al. [15] who find small differences in fuel input demands between restructured and non-restructured firms. Third, model GMM-1, which does not include R but incorporates dummy variables for Phase I and II units, is very 35 The

reason for the inclusion of this second control variable is to capture deviations in the behavior of publicly or cooperatively owned units after the restructuring process occurs.

19

ACCEPTED MANUSCRIPT

AC CE P

TE

D

MA

NU

SC

RI

PT

informative. Conditioning on the covariates, both groups are more productive on average than NSPS units. A possible reason is that most NSPS units were constrained right from the onset -many years before Title IV regulation was passed. Fourth, productivity of Phase I and Phase II units started to decrease in advance of the corresponding program implementation dates. Fifth, electricity generation presents increasing returns to scale (with the corresponding elasticities approximately equal to 1.1). Sixth, the distribution of productivity is not dramatically dispersed because technology is relatively more similar among electricity generating units than other more sophisticated industries producing highly differentiated products. Concretely, the ratio between the 99th and the 1st percentiles of the total factor productivity distribution is approximately 1.8, whereas the corresponding ratio between the 90th and the 10th percentiles is approximately 1.15. By comparing the OLS estimates with the GMM estimates, it is apparent that the transmission bias discussed before is severe and that the GMM approach solves (or at least mitigates) the problem. Output-elasticities with respect to capital and labor increase when GMM is implemented, at the expense of output elasticity with respect to fuel which drastically decreases. Hence, output-elasticity of capital is between .21 and .25 in the GMM models while it is slightly negative and not significant in the OLS regression; output-elasticity of labor is .02 in OLS, while it is approximately .09 in the GMM specifications; and output-elasticity of fuel is between .77 and .81 in the GMM models, whereas it is 1.06 in the OLS regression. Finally, the value of the coefficient of determination in the OLS regression is very close to 1, implying that there is practically nothing left in the error term once the part explained by the model is accounted for. This suggests that technology in the electricity generating industry is rather simple in the following sense: one can predict output with the only inclusion of labor, capital, and fuel as inputs. For the remainder of this paper I use the estimates of model GMM-2 which incorporates the variable R but excludes the market restructuring dummies.36

36 The

choice of specification is, to some extent, arbitrary and responds to the following reasons. First, in the specificacion GMM-4, the variable market restructuring is not statistically significant, whereas the variable publicly-owned utility is significant at the 1%. In any case, these two binary variables are only capturing part of the intercept, which is not particularly relevant for the subsequent simulations of the paper. Second, the coefficients used in the next sections are γR and γ f , which are very similar in both specifications, GMM-2 and GMM-4.

20

ACCEPTED MANUSCRIPT

GMM-1

GMM-2

GMM-3

GMM-4

RI

Parameter

PT

Table 3: Production function parameter estimates. OLS and GMM specifications.

γf γph1 γph1-post89 γph1-post94 γph2 γph2-post89 γph2-post94 γR γiou-drg γmnc-post92 γ0

-0.0358*** 0.0508*** -0.0239* -0.0069 0.0138 -0.0043 0.0222

-0.4329

6,816 0.763

NU

1.0585*** (0.0062) .02066* (0.0098) -0.0184 (0.0096)

(0.0042) -0.0568*** (0.0044)

-0.0579*** (0.0070)

-0.0476*** (0.0049) 0.0076 (0.0090) 0.0278*** (1.2567) -0.5989

-0.0386*** 0.0122** 0.0088 -3.2726***

PT ED

MA

(0.0064) -0.0550*** (0.0043) -0.0532*** (0.0079) (0.0120) (0.0075) (0.0222) (0.0101) (0.0127) -0.0499*** (0.0088) 0.0127*** 0.0284*** (1.3257) -0.7651 (1.1109) -0.8016

CE

Observations R-squared Hansen’s J

SC

0.7743*** (0.1271) 0.8129*** (0.1089) 0.8134*** (0.1235) 0.7955*** (0.1324) 0.0966** (0.0360) 0.0898** (0.0308) 0.0868** (0.0347) 0.0926** (0.0373) 0.2540* (0.1157) 0.2104* (0.1019) 0.2140* (0.1159) 0.2287* (0.1240)

AC

21

αh αl αk

OLS-4

(0.382)

6,813

0.406

6,816 (0.524)

0.563

(0.0098) (0.0055) (0.0094) (1.3491)

6,813 (0.448)

0.312

(a) Robust standard errors clustered by electricity generation unit are shown in parenthesis. (b) Significance levels: *** p<.01; ** p<.05; * p<.10 (c) All models include year dummies (not shown in the table) to control for year fixed effects. (d) Hansen’s test statistic χ2(1) and the corresponding p-value associated (shown in parenthesis).

(0.0117) (0.0058) (0.0057) (0.0674)

6,813 0.9961 (0.576)

ACCEPTED MANUSCRIPT

5

Impact of SO2 regulation on productivity and output Productivity changes for distinct compliance alternatives

PT

5.1

MA

NU

SC

RI

The average marginal effect on TFP of a change in R is approximately -2.2%.37 Since f ∈ {0, 1}, the corresponding effect on TFP is, by design, a discrete change. The relevant exercise is to compute the change in productivity due to scrubber adoption.38 It is important to highlight that the purpose here is not to evaluate the long-run compliance choice, where (present and expected) capital, operating and maintaining costs, allowance prices, and electricity prices, influence the decision.39 Instead, the goal is to measure and compare the potential impact that different compliance options would have had on coal-fired boilers’ productivity.40

Table 4: Productivity changes. Affected units during Phase I of ARP.

D

Factual

TE

1- Scrubbing 2- Fuel switching

AC CE P

Counterfactual

mean

sd

pctl 10 pctl 90

-2.50% 0.14% -2.68% -2.33% -0.97% 0.63% -1.86% 0.80% mean

sd

pctl 10 pctl 90

a- Full scrubbing -2.47% 0.17% -2.68% -1.98% b- Full fuel switching b.1) 2.5 lbs of SO2/mmbtu -1.09% 0.50% -1.78% 0.30% b.2) 1.2 lbs of SO2/mmbtu -1.53% 0.34% -1.94% -1.12%

Table 4 presents estimates of changes in TFP due to different scrubbing and fuel switching scenarios. The first two rows of the table show estimates of the observed choices made by the units during the first phase of the Title IV of the CAAA-1990. Productivity losses due to 37 The

computation of the average marginal effect of a change in R is as follows: γR · exp{γ0 + γR E [R jt ] +

T

E [C jt ] γc } 38 The expression used to compute the effect of f on TFP is: (exp(γ ) − 1) · exp{γ + γ E [R ] + E [C ]T γ }. f 0 R jt jt c T

Similarly, the effect of R on TFP is computed as follows: (exp (γR · R) − 1) · exp{γ0 + E [C jt ] γc } 39 See [20] for a formal analysis and estimation of compliance costs in a dynamic discrete choice setting. 40 Productivity loss is an inherent part of the regulation compliance cost. Surprisingly, most studies estimating the ARP cost barely mention the potential productivity loss as a significant part of the compliance cost. Some of the most cited empirical works are: [11], [14], and [25]. None of them makes productivity loss explicit in the estimation models. An early exception is [19], but that paper analyzes previous amendments to the CAAA in 1970 and 1977, and the results are not directly comparable to the ones presented in this paper.

22

ACCEPTED MANUSCRIPT

AC CE P

TE

D

MA

NU

SC

RI

PT

scrubbing are clearly larger than those related to fuel switching. The key factor in my model explaining this result is that while the productivity changes caused by fuel switching depend on the magnitude of R, scrubbing implies a discrete change. Recall that scrubbing drastically reduces sulfur dioxide emissions (up to 95%) and therefore over-compliance in both phases of the program is guaranteed. The second part of the table presents some counterfactual situations. I compute the productivity decline that would have happened if all units had adopted scrubbers (row a), and if all units had switched to low sulfur coal so as to comply with the environmental regulation (b.1 and b.2). As stated in the previous paragraph, different fuel switching scenarios require the specification of different (implicit) emission rates, or in other words, different values of R. Therefore, the distribution of R plays a significant role. In the simulations shown in table 4, I mimic the allowances allocation rule used in Phase I and Phase II of the ARP. Concretely, I assume each boiler receives allowances as if its emission rate standard were 2.5 and 1.2 lbs SO2 /mmbtu, respectively.41 The main assumption is that each boiler switches to some variety of low sulfur coal so that it achieves the required implicit emission rate. Notice that the three counterfactual situations implicitly assume there is no allowance trading. As in the factual case, scrubbing has a larger negative impact on TFP. As expected, the tighter standard implicitly required in Phase II increases the productivity losses due to fuel switching. However, even in this more stringent compliance scenario, TFP losses under a full scrubbing regime are considerably larger.42 It is worth emphasizing that the numbers presented in the counterfactual segment of table 4 should be considered with caution. They are a rough approximation of the counterfactual changes in average boiler’s productivity. A deeper technical analysis made on a case by case basis would be needed in order to evaluate the real retrofitting possibilities each EGU has. In that sense, the estimated productivity losses due to fuel switching presented in this section are probably underestimated.

5.2

Effect of ARP on output

Using the estimates of the production function parameters, I evaluate the effect of the ARP on output. A simple but very informative exercise is to “turn off” variable R for all units and set γ f = 0 for those Phase I units that installed scrubbers after 1990, and then compute the 41 Allowances

were allocated to existing units using the years 1985-1987 as baseline. fair comparison between fuel switching and scrubbing would be to assume that the change in R under fuel switching is such that the (lower) scrubbing emission level is achieved. That exercise is unrealistic to say the very least. Emission levels achieved under scrubbing are not feasible under any fuel switching alternative. 42 A

23

ACCEPTED MANUSCRIPT

Y jt∗ −Y jt Y jt∗

= 1 − e(γR R jt +γ f f jt )

PT

counter-factual output level that would have been produced without the implementation of the environmental policy in question. The impact on output is computed as follows

AC CE P

TE

D

MA

NU

SC

RI

where Y ∗ and Y are the counterfactual and the factual output levels, respectively. Before 1990, most units have R equal to zero and some units have values very close to zero. The reason behind this was explained in Section 2.3.3: for each unit the unconstrained emission rate is approximated by the median emission rate during the 1985-1989 period. Thus, small departures from the unit’s unconstrained emission rate might exist even during those pre-regulation years. Figure 4 depicts the evolution of the average policy impact on output for the period 1990-1999 including only the units affected by the first phase of the ARP. It emphasizes the input biasing character of the environmental regulation where the majority of affected units opted for a fuel switching strategy. Scrubber adoption was less significant but still important in terms of output loss.43 Trading of pollution permits acted as a complementary strategy. Phase I units were, all things equal, the dirtiest emission sources. Considering output losses below 0.3% as normal output fluctuations, it is evident that some boilers reacted in advance of the effective implementation of the ARP. Although one can distinguish an important jump in the average output loss in 1995, the trend is positive since early 1990s, suggesting certain policy anticipation by the firms: This result was also found in Ellerman et al. [14], and is not at odds with the way electricity generating units make their decisions. Fuel quality adjustments require capital investments that take some time to be fully operative. There might also be some processes of learning by doing, contract renegotiation delays, or even frictions when relationships with new suppliers (i.e. coalmines) are established. As a result, electric utilities might retrofit their boiler-generator pairs sometime before the effective implementation of the environmental program, just to make sure they will be able to comply with the regulation mandates. The differentiated impact of the environmental policy across geographical regions is presented in Figure 5.44 The Great Lakes and the West North Central were, on average, the most 43 First,

recall from Section 5.1 that the average productivity loss associated with scrubbing was considerably higher than that associated with fuel switching. Second, among the affected units, only 10% effectively adopted a FGD system during the first phase of the program, but they accounted for more than 20% of the total electricity generated. These two facts help explain why the size of the output loss caused by scrubbing is more than one half of the output loss due to fuel switching (see Figure 4) 44 Regions are defined as follows: Northeast: PA, NY, NJ, CT, RI, MA, VT, NH, ME. West: AZ, CA, CO, ID, MT, NV, NM, OR, UT, WA, WY. Great Lakes: IL, IN, OH, MI, WI. West North Central: IA, KS, MN, MO,

24

ACCEPTED MANUSCRIPT

Figure 4: Decomposition of the SO2 regulation impact: R and f effects

1991

1992

AC CE P

1990

TE

0.00

D

0.50

MA

NU

Output loss in % 1.00 1.50 2.00

SC

RI

2.50

PT

3.00

(Mean values over Phase I units. Weighting variable: Y jt )

1993

1994

scrubber adoption

1995

1996

1997

1998

1999

fuel switching

Note: the threshold for normal output fluctuations is < 0.3% (green dash line).

affected regions. The policy impact on output in the South and Northeast followed a similar pattern but the magnitude of the impacts were relatively smaller. Observing the location of coalmines (see Figure 1) and comparing the characteristics of the coal produced in each region (see Table 2), it is clear that units located in the Northeast have a closer access to low- and medium-sulfur coal from Central Appalachian mines whereas units located in the West North Central Region are far away from clean coal sources (either Appalachian mines or Powder River Basin mines). Notice that the West Region is not included in Figure 5, and the reason is that there is no Phase I unit in this geographical area.45 NE, ND, SD. South: AL, AR, DE, FL, GA, KY, MD, MS, NC, OK, SC, TN, TX, VA, WV, LA. 45 Boilers in the West Region had an easy access to low-sulfur coal from PRB, and many of them were built

25

ACCEPTED MANUSCRIPT

PT

Lastly, Figure 6 shows how the policy impact on output was distributed across the affected units in the year 1999. Table 5 breaks down the distribution of output loss into quintiles using net electricity generation as weights.

RI

Figure 5: Impact of SO2 regulation by region

SC

(Mean values over Phase I units. Weighting variable: Y jt )

4

1996

1998

D

1994

0

2000

TE

1992

Output loss in % 1 2 3

MA

Output loss in % 1 2 3 0 1990

1990

1992

1990

1992

1994

1996

1994

1996

1998

2000

1998

2000

4

South

0

Output loss in % 1 2 3

AC CE P

Output loss in % 1 2 3

4

West North Central

0

Great Lakes

NU

4

Northeast

1998

2000

1990

1992

1994

1996

Note: the threshold for normal output fluctuations is < 0.3% (red dash line).

for PRB coal specifications. A conjecture is that boilers located in the West Region would not have suffered the impact of the ARP as much as units located somewhere else in the country.

26

ACCEPTED MANUSCRIPT

Quintile

2nd

3rd

4th

0.07%

0.39%

1.70%

2.94%

5th 5.09%

RI

Average output loss

1st

PT

Table 5: Distribution of SO2 regulation impact in year 1999

MA D TE AC CE P

0

5

Number of units 10 15 20

25

NU

30

SC

Figure 6: Distribution of estimated output loss: Phase I units, year 1999

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0 Output loss (in %)

Note: the threshold for normal output fluctuations is < 0.3% (red solid line).

6

Concluding comments

Using a sample for the 1985-1999 period, this paper provides empirical evidence for the impact of the first phase of ARP on productivity and output. The model used takes into account relevant technological particularities and market conditions that characterize the industry. Each EGU is designed to burn a particular type of coal. However, there are different coal vari27

ACCEPTED MANUSCRIPT

AC CE P

TE

D

MA

NU

SC

RI

PT

eties and the production technology is such that significant deviations from the targeted design (ideal) coal characteristics harm productivity. The most important contribution of this research on theoretical grounds is the inclusion of the variable R, which represents the observed deviations of coal actually burned from the boiler’s ideal fuel design characteristics. This variable enters the production function directly, affecting the unit’s TFP. The relevance of R is illustrated by the fact that most units affected by the ARP opted (fully or partially) for a fuel switching compliance strategy, which gave rise to important relative deviations from those ideal design characteristics, and ultimately, affected productivity and output. Since the period of analysis was also marked by several episodes of market restructuring, a careful empirical strategy was implemented to account for these confounding influences. The estimation procedure was based on the Generalized Method of Moments developed by Hansen [21] and proved to be effective in dealing with the ‘transmission bias’ problem. Hedonic price regressions were used to compute exogenous coal prices at the plant level, which in turn were used as instruments in the estimations to address the endogeneity problem. With the production function parameter estimates I was able to measure the impact of the first phase of the ARP on output and TFP. Market restructuring processes proved to be slightly significant in determining productivity. Estimated TFP declines due to the ARP were between 1% and 2.5%, on average. Output losses ranged from 1% to more than 6%, varying across regions, over time, and mainly depending on the proximity to low-sulfur coal sources.

References

[1] J. Abito. Welfare gains from optimal pollution regulation. Job Market Paper, Northwestern University., 2012. [2] D. Ackerberg, K. Caves, and G. Frazer. Structural identification of production function. UCLA, Deloitte and Touche, Rotman School of Management, 2006. [3] M. Arellano and S. Bond. Some tests of specification for panel data: Monte carlo evidence and an application to employment equations. The Review of Economic Studies, 58(2):pp. 277–297, 1991. [4] S. Atkinson and R. Halvorsen. Parametric efficiency tests, economies of scale, and input demand in u.s. electric power generation. International Economic Review, 25(3):pp. 647– 662, 1984.

28

ACCEPTED MANUSCRIPT

PT

[5] E. Berndt and L. Christensen. The Internal Structure of Functional Relationships: Separability, Substitution, and Aggregation. The Review of Economic Studies, 40(3):pp. 403– 410, 1973.

RI

[6] E. Berndt and L. Christensen. The translog function and the substitution of equipment, structures, and labor in U.S. manufacturing 1929-68. Journal of Econometrics, 1(1):pp. 81–113, 1973.

SC

[7] M. Bernstein, S. Feldman, and A. Schinnar. Impact of pollution controls on the productivity of coal-fired power plants. Energy Economics, 12(1):pp. 11 – 17, 1990.

NU

[8] R. Blundell and S. Bond. GMM Estimation with persistent panel data: an application to production functions. Econometric Reviews, 19(3), 2000.

MA

[9] J. Bushnell and C. Wolfram. Ownership change, incentives and plant efficiency: The divestiture of us electric generating plants. Center for the Study of Energy Markets, (140), 2005.

D

[10] M. Busse and N. Keohane. Market effects of environmental regulation: coal, railroads, and the 1990 clean air act. RAND Journal of Economics, 38(4):pp. 1159–1179, 2007.

AC CE P

TE

[11] C. Carlson, D. Burtraw, M. Cropper, and K. Palmer. Sulfur dioxide control by electric utilities: What are the gains from trade? Journal of Political Economy, 108(6):pp. 1292– 1326, 2000. [12] S. Cicala. When does regulation distort costs? lessons from fuel procurement in us electricity generation. American Economic Review, 105(1):pp. 411–444, 2015. [13] E. Dardati. Pollution permit systems and firm dynamics: Does the allocation scheme matter? Working paper, 2012. [14] D. Ellerman, P. Joskow, R. Schmalensee, J. Montero, and E. Bailey. Markets for clean air: the U.S. acid rain program. Cambridge, UK: Cambridge University Press, 2000. [15] K. Fabrizio, N. Rose, and C. Wolfram. Do Markets Reduce Costs? Assessing the Impact of Regulatory Restructuring on US Electric Generation Efficiency. American Economic Review, 97:pp. 1250–1277, September 2007. [16] M. Fowlie. Emissions trading, electricity restructing, and investment in pollution abatement. The American Economic Review, 100(3):pp. 837–869, 2010. [17] E. Frey. Technology Diffusion and Environmental Regulation: The Adoption of Scrubbers by Coal-Fired Power Plants. Energy Journal, 34(2), 2013. [18] A. Gandhi, S. Navarro, and D. Rivers. On the identification of production functions: How heterogeneous is productivity? University of Wisconsin-Madison and University of Western Ontario, 2011. 29

ACCEPTED MANUSCRIPT

PT

[19] F. Gollop and M. Roberts. Environmental regulations and productivity growth: The case of fossil-fueled electric power generation. Journal of Political Economy, 91(4):pp. 654– 74, 1983.

RI

[20] P. Hancevic. A Dynamic Approach to Environmental Compliance Decisions in U.S. Electricity Market: The Acid Rain Program. CIDE, Working Paper, 2016.

SC

[21] L. Hansen. Large Sample Properties of Generalized Method of Moments Estimators. Econometrica, 50(4):pp. 1029–1054, 1982.

NU

[22] W. Hendricks. The effect of regulation on collective bargaining in electric utilities. The Bell Journal of Economics, 6(2):pp. 451–465, 1975.

MA

[23] P. Joskow. Inflation and environmental concern: Structural change in the process of public utility price regulation. Journal of Law and Economics, 17(2):pp. 291–327, 1974. [24] N. Keohane. Environmental policy and the choice of abatement technique: Evidence from coal-fired power plants. Yale School of Management, Working paper, 2002.

TE

D

[25] N. Keohane. Cost savings from allowance trading in the 1990 clean air act: Estimates from a choice-based model. In Moving to Markets in Environmental Regulation: Lessons from Twenty Years of Experience, Chapter 8. Oxford University Press, 2006.

AC CE P

[26] N. Keohane. Environmental Policy Design and the Distribution of Pollution Control Techniques in a Regulated Industry. Yale School of Management, Working paper, 2006. [27] J. Laffont and J. Tirole. A theory of incentives in regulation and procurement. Information Economics and Policy, 6(1):pp. 89–93, 1994. [28] J. Levinsohn and A. Petrin. Estimating production functions using inputs to control for unobservables. The Review of Economic Studies, 70(2):pp. 317–341, 2003. [29] J. Marschak and W. Andrews. Random simultaneous equations and the theory of production. Econometrica, 12:pp. 143–205, 1944. [30] M. Nerlove. Returns to scale in electricity supply. Measurement in Economics, Chapter 7. (Christ, C. et al, eds.). Stanford: Stanford University Press), pages pp. 167–98, 1963. [31] G. Olley and A. Pakes. The dynamics of productivity in the telecommunications equipment industry. Econometrica, 64(6):pp. 1263–1297, 1996. [32] S. Schennach. The Economics of Pollution Permit Banking in the Context of Title {IV} of the 1990 Clean Air Act Amendments. Journal of Environmental Economics and Management, 40(3):pp. 189–210, 2000. [33] R. Schmalensee and R. Stavins. The so2 allowance trading system: The ironic history of a grand policy experiment. Journal of Economic Perspectives, 27(1):pp. 103–22, 2013. 30

ACCEPTED MANUSCRIPT

[34] T. Sueyoshi, M. Goto, and T. Ueno. Performance analysis of US coal-fired power plants by measuring three DEA efficiencies. Energy Policy, 38(4):pp. 1675 – 1688, 2010.

RI

PT

[35] S. Yaisawarng and J. Klein. The Effects of Sulfur Dioxide Controls on Productivity Change in the U.S. Electric Power Industry. The Review of Economics and Statistics, 76(3):pp. 447–460, 1994.

AC CE P

TE

D

MA

NU

SC

[36] H. Zhang. Biased Technology and Sources of Growth in Chinese Industries: A FirmLevel Analysis. Pennsylvania State University, 2012.

31

ACCEPTED MANUSCRIPT

Appendix

PT

Estimation of exogenous coal prices.

RI

A

Delivered cost of coal

Std. Err.

0.0130*** -7.9051*** 0.3699 0.0627*** 0.0519** 0.0589*** 0.0478*** 0.0536*** 0.0802*** 0.0744*** 0.0655*** 0.0708*** 0.0775*** 0.0849*** 0.0893*** 0.0854*** 0.0861*** 0.0796***

(0.0022) (2.3268) (0.5287) (0.0179) (0.0192) (0.0169) (0.0159) (0.0105) (0.0121) (0.0115) (0.0117) (0.0118) (0.0095) (0.0083) (0.0076) (0.0058) (0.0059) (0.0082)

D

TE

AC CE P

btu content sulfur content ash content distance x year1985 distance x year1986 distance x year1987 distance x year1988 distance x year1989 distance x year1990 distance x year1991 distance x year1992 distance x year1993 distance x year1994 distance x year1995 distance x year1996 distance x year1997 distance x year1998 distance x year1999

Coef.

MA

Regressor

Observations R2

NU

Dependent variable:

SC

Table 6: Coal price regression for Central Appalachian Region

Regressor

Coef.

Std. Err.

year1985 year1986 year1987 year1988 year1989 year1990 year1991 year1992 year1993 year1994 year1995 year1996 year1997 year1998 year1999

-8.0876 -19.2948 -27.8554 -26.5262 -27.1747 -20.1321 -21.7311 -22.7071 -24.4019 -26.6738 -32.3887 -36.3969 -37.0180 -38.7826 -39.1622

(37.1538) (36.0872) (35.2650) (35.6328) (35.4231) (37.1582) (37.8462) (38.3259) (38.2985) (37.1536) (36.6639) (36.5165) (36.5942) (36.2754) (36.6149)

140,745 0.9741

Source: FERC’s Form 423. Distance is in miles, from plant’s county to coalmine’s county.

B

Compliance choices and the exogeneity of R and f

On the one hand, scrubber adoption and fuel switching decisions are both subject to expensive capital costs. Unexpected productivity fluctuations are unlikely to modify these two long-run compliance choices. On the other hand, most coal purchases are made by contracts which are 32

ACCEPTED MANUSCRIPT

SC

RI

PT

typically executed as follows: a number of shipments with some estimated delivered quantity/quality are agreed. Those quantities, however, are not completely fixed and some adjustments are possible afterwards. Both facts described in the previous paragraph imply that if a boiler receives a positive productivity shock, it will purchase larger quantities of the same variety of coal currently consuming. As an illustration, suppose that a boiler is consuming the variety m of coal. The implicit R is ∗ E − Em R= E∗

TE

D

MA

NU

where E ∗ is the unconstrained emission rate (fixed by definition) and E m is the actual emission rate derived from consuming variety m -i.e. a linear transformation of the ratio between sulfur content and btu content. As a result, a higher demand for the variety m will leave E m , and therefore R, unaffected. To analyze the exogeneity of R and f in more detail, consider the profit function for the unit j in time t once the (long-term) compliance strategy was already decided.46 For ease of the exposition and without loss of generality, let us consider two leading extreme cases: A) Adopt a scrubber and burn ideal coal type so that R jt = 0 and f jt = 1

AC CE P

B) Only purchase low-sulfur coal so that R jt > 0 and f jt = 0 The profit functions under these two strategies are  
ΠAjt = p jt Y jt − q jt H jt − r jt K jt − w jt L jt − zt S jt (1 − φ) − Sˆ jt −CB S jt , K jt ,t  
ΠBjt = p jt Y jt − q jt H jt − r jt K jt − w jt L jt − zt S jt − Sˆ jt −CC K jt ,t

(4) (5)

where S and Sˆ are the gross SO2 emissions before scrubbing, and the number of allocated free emission permits due to grandfathering, repectively. The parameter φ is the scrubber emission reduction rate; CB (·) is the levelized scrubbing cost which includes operating and maintenance costs plus capital cost; CC (·) is the levelized capital cost of fuel switching; p is the electricity generation price in $/MWh; z is the pollution permit price measured in dollars per ton of SO2 ; q is the coal price in $/mmBtu; w is the average wage rate in dollars; and r is the rate of return 46 One

can think of it as a two stage profit maximization problem where units choose the compliance method in the first stage and then the output level in the second stage. In a multi-period setting, compliance decisions are made at the beginning and they are only revised when a new regulation is passed or drastic market changes are observed. Output decisions are more flexible and more frequent (i.e. each period).

33

ACCEPTED MANUSCRIPT

to capital. Thus, for some fixed levels of L jt , K jt and H jt , it is possible to compare the two extreme cases just subtracting one profit equation from the other

RI

PT

    γ0 +ε jt γ f γR R jt = p F(K , L , H )e ∆ΠA−B e − e + q (0) − q (R) H jt jt jt jt jt jt jt jt   h B

i − zt S jt (0) (1 − φ) − S jt (R) − C S jt (0), K jt ,t −CC K jt ,t

(6)

Omitted tables and figures

AC CE P

C

TE

D

MA

NU

SC

The key aspects to consider when choosing the compliance strategy are: i) the value of output losses; ii) the (potential) cost premium for low-sulfur coal; iii) the difference in permits used or saved; and iv) the levelized costs associated to fuel switching and scrubbing. The effect of a log-productivity shock, ε, is a priori ambiguous. Consider a boiler highly constrained that requires a large deviation from its ideal coal type (i.e. a large R) if it were to adopt the fuel switching strategy (option B). Then, higher log-productivity favors scrubber adoption. The opposite happens if following strategy B implies a minor deviation from ideal coal type. In that case, higher log-productivity reduces the incentives for scrubber adoption. Because there are no a priori reasons for supposing a positive (negative) correlation between R and ε or between f and ε, I assume that R and f are econometrically exogenous in equation 3.

Table 7: Coal purchases by producing regions. Transition Matrix. (Periods considered: 1985-89; 1990-94; 1995-99; and 2000-05) Northern Central Appal. Appal.

Northern Appalachian Central Appalachian Interior Region PRB-West Region Alabama Texas and Oklahoma West North Central

0.66 0.11 0.06 0.01 0.01 0.00 0.00

0.28 0.78 0.12 0.00 0.63 0.00 0.00

Interior Region

PRBWest

AL

TX, OK

WNC Region

0.01 0.04 0.55 0.00 0.18 0.00 0.29

0.06 0.06 0.26 0.99 0.04 0.36 0.67

0.00 0.01 0.01 0.00 0.13 0.00 0.00

0.00 0.00 0.00 0.00 0.00 0.64 0.00

0.00 0.00 0.01 0.00 0.00 0.00 0.05

34

SC

RI

PT

ACCEPTED MANUSCRIPT

0

500

AC CE P

Frequency

TE

1000

D

MA

1500

NU

Figure 7: Distribution of the relative absolute deviation of emission rates during the period 1985-1989

0

.05

.1

.15

35

.2

.25

.3

ACCEPTED MANUSCRIPT

Highlights

PT

• Each coal-fired boiler is designed to burn a particular variety of coal • Deviations from the targeted coal characteristics result in productivity losses

SC

RI

• During the phase 1 of the ARP, most boilers switched to lower sulfur coal, others bought allowances, and a few adopted scrubbers • On average, productivity and output dropped between 1

AC CE P

TE

D

MA

NU

• The negative impact on productivity due to scrubbing was larger than that due to fuel switching

36