Compliance by Design: Industry Response to Energy Efficiency Standards*

Transcription

1 Compliance by Design: Industry Response to Energy Efficiency Standards* By KATE S. WHITEFOOT, MEREDITH FOWLIE, AND STEVEN J. SKERLOS* Policies designed to improve industrial environmental performance are increasing in scope and stringency. These policies can significantly influence engineering design decisions as firms re-optimize their products and processes to meet compliance requirements at minimum cost. This paper demonstrates the importance of accounting for these design responses in the analysis of policy impacts. As a case in point, we model automotive firms medium-run compliance choices under the reformed Corporate Average Fuel Economy (CAFE). Physicsbased simulations are used to characterize the potential for improving fuel efficiency through design changes. These engineering simulation results are coupled with a partial-equilibrium, static oligopoly model in which firms choose prices and key vehicle design attributes. We simulate firms pricing and mediumrun design response to the CAFE regulation. Results indicate that firms rely primarily on changes to vehicle designs to meet the reformed standards, with a smaller contribution coming from pricing strategies designed to shift demand towards more fuel-efficient vehicles. The analysis draws attention to several factors that could offset fuel efficiency improvements achieved under the reformed CAFE standards, such as an increase in the market share of larger vehicles and the strategic response of firms choosing to violate the standard. * Whitefoot: National Academy of Engineering, 500 Fifth St, NW, Washington, DC, ( kwhitefoot@nae.edu). Fowlie: University of California, Berkeley and NBER, 327 Giannini Hall, UC Berkeley, Berkeley CA, ( fowlie@berkeley.edu). Skerlos: University of Michigan, 1301 Beal St, Ann Arbor, MI ( skerlos@umich.edu). The authors would like to thank Severin Borenstein, Lawrence Goulder, Mark Jacobsen, Ryan Kellogg, Chris Knittel, James Sallee and seminar participants at the University of Michigan and the University of California, Berkeley for their helpful suggestions. We also thank Ashley Langer for valuable help with the demand estimation, Bart Frischknecht and Deokkyun Yoon for assistance with vehicle simulations, and Kevin Bolon for help managing the data sets. Financial support for this work is provided by the National Science Foundation and the University of Michigan Rackham Graduate School and is gratefully acknowledged. All findings and conclusions are those of the authors and do not necessarily represent the National Academy of Engineering. 1

2 I. Firm response to energy-efficiency standards Design-based standards are increasingly being used to improve the end-use efficiency of durable goods. Examples include standards for household appliances, lighting products, light-duty and heavy-duty vehicles. How firms respond to these standards can have significant implications for how efficiently mandated energy-intensity reductions are achieved and who bears the costs. In general, firms can comply with energy-efficiency standards by shifting production towards their more efficient products or modifying the attributes of their products via product design changes. This paper examines both price and design responses, placing particular emphasis on how design responses manifest in an imperfectly competitive market context. In much of the economics literature that investigates the response of a differentiated-products industry to a regulatory intervention or change in market structure, firms ability to change product or process design is underemphasized or ignored (e.g., Goldberg 1998; Jacobsen 2012; Nevo, 2000). Recent work on the automotive industry indicates that vehicle design changes have played a significant role in determining fleet fuel-efficiency trends, including gains achieved under the Corporate Average Fuel Economy (CAFE) standards (Knittel 2009; Klier and Linn 2012). This paper presents an approach to explicitly incorporate this design response into the modeling and analysis of energyefficiency standards. The design decisions that pertain to products typically targeted by energyefficiency standards, such as automobiles and household appliances, are more technologically complex than those traditionally explored in the literature on endogenous product attribute choice (e.g., Seim 2006; Sweeting 2007; Fan 2008). 2

3 We develop an empirically tractable approach to model these more complex design decisions. To accomplish this, we draw heavily from the engineering design literature that investigates the design of various technologically complex products (e.g., Asteausu, Astiazaran, and Besga 1992; Assanis et al. 1999; Gholap and Khan 2007). We demonstrate how detailed engineering simulation models of product design can be constructively integrated into economic models of strategic industry interactions. Our general approach is straightforward. Using an engineering simulation model that is used by the automotive industry to support the powertrain development process, we implement thousands of vehicle design simulations over a range of feasible vehicle design configurations. These simulated data are then used to estimate a flexible approximation to the medium-run vehicle design process that can be integrated into an economic model of industry interactions. Our modeling of automotive design tradeoffs is germane to recent work by Klier and Linn (2012) and Knittel (2009). Whereas these authors use bundles of attributes observed in the marketplace to econometrically estimate tradeoffs between fuel economy, weight, and engine power, we use the outputs of physicsbased engineering simulations. 1 For our purposes, an engineering-based approach confers two advantages. First, many combinations of product attributes are not observed in the marketplace, but are technologically feasible and potentially optimal under counterfactual policies. Engineering simulation models can, in principle, identify technologically possible combinations of attributes that have yet to manifest in existing product designs. Second, correlations between unobserved attributes (e.g., electronic accessories) and attributes of interest (e.g., 1 This is certainly not the first paper to make use of engineering estimates in a detailed economic analysis. For example, engineering estimates of costs have been used to benchmark electricity sector performance (see, for example, Wolfram (1999) and Borenstein (2002). Engineering models have also been used to simulate the response of electricity producers and automotive firms, respectively, to policies limiting NO x emissions (Fowlie et al., 2012). 3

4 fuel economy) can make it difficult to identify design tradeoffs econometrically. 2 Physics-based engineering simulations allow us to capture engineering tradeoffs independent of unobserved product attributes. Notably, our methods are more broadly applicable to other policy interventions and other technologically complex products. 3 With our parsimonious representation of engineering design tradeoffs in hand, we are in a position to bridge the gap between engineering and economic analysis of an industry s response to energy-efficiency standards. The former captures the structure of engineering design in detail, but often ignores consumer demand responses and strategic pricing and design decisions of competitors. In contrast, the economics literature has sought to capture the strategic nature of firms interactions in the product market, but has often oversimplified or ignored the design dimension. We nest our model of the vehicle design process within an economic model of a differentiated product oligopoly. This yields an analytical framework that can be used to examine how firms strategically alter both prices and product attributes in response to a designtargeting policy intervention. In the second part of the paper, we use our framework to analyze the 2014 reformed CAFE standards. We find that the sales-weighted average fuel economy increases by 2.1 mpg under the standards, which represents a 9 percent increase from the pre-reform policy baseline in Our analysis also highlights some important interactions between product design choices and the strategic nature of firms interactions. For example, we find that efficiency improvements among firms that comply with the standard are offset somewhat by an increase in fuel 2 Klier and Linn (2008) exploit an engine dataset to estimate tradeoffs between endogenous attributes using variation in observed attributes of vehicle models with the same engine program. One potential drawback of this approach is that unobserved attributes, such as electronic accessories that can impact fuel economy, are often correlated with observed attributes such as horsepower and weight. 3 For example, Gholap and Khan (2007) develop a model that characterizes the tradeoffs between energy efficiency and production costs using detailed design simulations of a refrigerator. This type of model could be used in concert with econometric estimation of consumer demand for refrigerators to investigate the design and pricing response of refrigerator manufacturers to energy-efficiency standards. 4

5 consumed by vehicles sold by firms choosing not to comply with the standard. A second offsetting effect manifests because some consumers value acceleration performance more than fuel economy and are willing to pay considerably more for it. This creates an incentive for compliant firms to produce some very low fuel-economy vehicles even as they increase the fuel economy of other vehicles to comply with the standards. These effects could be reduced, respectively, by increasing the fine that noncompliant firms are required to pay and by adding minimum standards in addition to the average standards. To explicitly illustrate the value added by our engineering design model, we implement a second simulation exercise in which firms ability to tradeoff acceleration performance and fuel efficiency is shut down. This exercise is important because these tradeoffs are routinely ignored in analysis of CAFE (e.g., Austin and Dinan 2005, NHTSA 2008, NHTSA 2012). We find that ignoring these tradeoffs leads to a significant overestimation of the costs of the policy regulation. The estimated costs in terms of losses in producer profits and consumer surplus per emission reductions are twice as large when we do not account for tradeoffs between fuel economy and acceleration. II. An overview of the vehicle design process Before delving into the details of the engineering model and the associated estimation, this section provides an overview of the automotive development process. We then introduce our analytical framework in general terms. A. The Big Picture: Vehicle Design and Redesign Generally speaking, the automotive design process is a structured sequence of interrelated decisions, many of which constrain choices made at later stages (Braess 2005; Sörensen 2006; Weber 2009). The process typically begins with 5

6 concept development, followed by a system-level design that defines the geometric layout of the vehicle (including target vehicle dimensions), followed by a detailed design of all subsystems. 4 Figure 1 provides a stylized representation of this design process. The figure is somewhat misleading insofar as it suggests that the process proceeds in sequential, clearly defined stages. In fact, iteration loops and overlapping tasks complicate the process. What is important for our purposes is that there are certain design decisions that must be finalized before the detailed engineering design of vehicle subsystems can be finalized. This distinction between long-run and medium-run design decisions is an organizing principle for our analysis. We will focus exclusively on medium-run design choices after Stage A in Figure 1. We take as given the design parameters that are determined in the earlier stages of the design process. These include the vehicle segment (e.g., mid-size sedan), the powertrain architecture (e.g., conventional gasoline, hybrid, or diesel), and key internal and external dimensions, (Braess 2005; Sörenson 2006; Weber 2009). Conditional on these parameters, we focus on two attributes that can be manipulated in the medium run: fuel efficiency (measured as gallons of fuel consumed per 100 miles) and acceleration (measured as the time in seconds to accelerate from 0-60 mph). Although there are other vehicle attributes that can be modified in the medium run (such as vehicle body styling), fuel consumption and acceleration are the most important in terms of determining compliance with the CAFE standards. 5 4 For a newly designed vehicle model, the development process begins with targets for specific vehicle attributes, such as the vehicle segment (e.g., compact), powertrain architecture (e.g., hybrid), variations (e.g., four-door sedan), major dimensions, transmission types (e.g., automatic, torque classes) and engine versions (Braess 2005; Weber 2009). For a redesigned model, the development process begins with the determination of any changes to major properties of the vehicle and specifications for subsystems, such as how many drivetrain configurations or engine options will be available. 5 Vehicle weight is another attribute that significantly affects the fuel efficiency of a vehicle. However, additional lightweighting (i.e., lightweighting that would not have occurred under the older CAFE standards) is not included in our simulations. As we explain below, using NHTSA s (2009) estimates of the cost effectiveness of additional lightweighting, our results would not change substantially if additional lightweighting were explicitly modeled. To the extent that lightweighting options are more cost effective than NHTSA s analysis implies, our results represent upper bounds of profit losses and consumer surplus losses. 6

7 B. Manipulating fuel efficiency in the medium run Automotive manufacturers use engineering simulation models to inform powertrain design and development decisions. These models are too computationally demanding to be incorporated directly into our economic modeling of firms design decisions. In this section, we explain how we distill engineering relationships between design parameters and vehicle attributes into a system of estimable equations that can be nested within a standard economic model of the automotive industry. We begin by making a conceptual distinction between continuous and discrete design decisions as they pertain to medium-run fuel efficiency improvements. An automaker can make discrete improvements in fuel efficiency by incorporating extra technology features into vehicle design. Examples include high-efficiency alternators, low resistance tires, and low-friction materials in the engine (NHTSA 2008). An automaker can also make more continuous adjustments to fuel efficiency and acceleration by tuning parameters in the powertrain (e.g., engine displacement, final drive ratio, and programming of the electronic control unit). 6 Both discrete and continuous fuel-efficiency improvements are represented graphically in Figure 2. The solid lines represent iso-technology curves. Movements along the baseline technology curve capture the engineering tradeoffs between fuel efficiency and acceleration for a vehicle with no extra technology features. The addition of technology features effectively shifts the baseline curve down (in the direction of fuel efficiency improvements). Transforming this simple picture into a model that can be implemented empirically requires additional notation. Let j index a vehicle model and engine 6 For example, consider a given vehicle design such as the Honda Accord. If Honda wants to increase the fuel efficiency of the Accord, it could decrease the displacement size of the engine, or it could simply change the programming in the powertrain electronic control unit to favor fuel efficiency over acceleration performance. These changes can be approximated as continuous adjustment of fuel efficiency. Each of these adjustments to improve fuel efficiency will cause some loss in acceleration performance. 7

8 option (e.g., the Ford Escape with a 3.5 L engine). Let x j denote the powertrain design parameters that are manipulable in the medium-run. Let t j index the suite of discrete technology features that affect fuel consumption and/or acceleration. Let represent the fixed design parameters decided earlier in the design process. As noted above, fuel consumption (fuelcons j ) and acceleration (acc j ) can be modified in the medium run via powertrain tuning and the choice of technology features. 7 We define the following two relationships: (1), ; (2), ; As described in detail in section III, for any choice of acceleration performance and technology features, there is only one choice of x j that minimizes the production cost associated with a given level of fuel efficiency. We will show that equation (2) can be substituted into equation (1) to obtain: (3), ; This simplification reduces the set of design choice variables and thus the computational complexity. 8 We use an engineering vehicle simulation model, called AVL Cruise, that is used by major automotive manufacturers to inform powertrain design (Mayer, 2008). This model is used to simulate fuel consumption performance under a wide range of vehicle segment-specific powertrain configurations. The data generated 7 Although there are some other vehicle attributes that can be modified in the medium run (such as the styling of the vehicle body), we expect that fuel consumption and acceleration performance are the most important in terms of being affected by the CAFE standards in the medium run considered in this analysis. 8 We could instead have reformulated the optimization with as a decision variable instead of. This convention is arbitrary and does not affect the formulation. 8

9 by thousands of simulations are used to estimate the flexible function. This estimation exercise is described in detail in Section 3. In addition to modeling the impact of medium-run design decisions on fuel efficiency and acceleration performance, we also need to account for the cost implications of these design decisions. The broken lines in Figure 2 represent isocost curves. Note that it is possible to improve fuel efficiency without increasing production costs through the use of technology features, but doing so will reduce acceleration performance. In Section III, we estimate the relationship between vehicle design parameters and vehicle production costs. C. The firm-level optimization problem To model firms product pricing and design decisions in response to energyefficiency standards, we nest the engineering relationship summarized by equation (3) and the associated cost function within a standard differentiated product oligopoly model. We assume that multiproduct firms choose prices, acceleration, and technology features to maximize profits over the set of vehicles they produce according to the following formulation. 9 (4) max,,, ; 0,, ;, ;, ;, 9 Modeling the market as a two-stage game where firms make design decisions before pricing decisions is beyond the scope of this paper. Morrow et al. (forthcoming) compared the results of a single-stage and two-stage game in the automotive market ignoring the CAFE standards where a single firm makes similar design decisions analyzed in this paper and all firms choose prices. The authors found that profit-optimal design decisions were the same in both the single-stage and two-stage games. 9

10 The variables,, and c j are respectively the quantity demanded, price, and marginal cost associated with vehicle model. The constraint, ; represents engineering restrictions on powertrain design variables and technology features. Given these engineering restrictions and the fixed design parameters, fuel consumption for a vehicle is determined jointly by the firm s choice of technology features t j and acceleration. The reformed CAFE policy is represented as a constraint for each vehicle class (i.e. passenger cars and light trucks). We define to be the sales-weighted average fuel economy of all vehicles in class that the firm produces, which must equal or exceed the firm s standard for that vehicle class,. This policy is described in more detail in Section IV. Conditional on the long-run design parameters, vehicle demand is determined by vehicle price, fuel consumption, and acceleration. Note that the design parameters x j and technology features t j are not explicitly represented in these demand equations. Technology features and powertrain parameters impact demand through their influence on fuel consumption and acceleration. We do not expect that they would have significant direct, intrinsic value to the consumer. 10 Demand estimation is discussed in detail in Section V. Finally, the engineering and economic tradeoffs associated with the medium-run design decisions are summarized by the functions and which together define a surface that we will subsequently refer to as the production possibility frontier (PPF). Points along the PPF represent the set of values,, that are attainable in the medium run conditional on. III. An estimable model of medium-run design decisions 10 This intrinsic consumer valuation may be more significant in cases of more advanced technology such as electrification of vehicles. Because the medium run considered in this paper is after powertrain architecture is fixed, and therefore advanced technology features such as electrification are exogenous, we do not expect consumer valuation of technology features in of themselves to be significant. 10

11 This section provides a more detailed explanation of how we implement the approach empirically. Our modeling framework is intended to capture the major levers that firms can use to respond to the CAFE standards. As we discuss below, our model is not completely comprehensive. We omit some design parameters and technology features that may be manipulable in the medium-run. This caveat notwithstanding, the model considerably extends the scope of the strategic response that has been analyzed in the literature and confers several advantages described in Section C below. A. Estimating baseline design tradeoffs We begin by specifying the design parameters explicitly considered in the analysis. In our model, two key parameters characterize the design of a vehicle s powertrain: engine displacement size and the final drive gear ratio in the transmission. These two parameters are manipulable in the medium run. Longerrun design parameters vehicle segment and curbweight (i.e. the weight of the vehicle without any passengers or cargo) are held fixed. 11 Our simulation of vehicle design tradeoffs begins with the construction of segment-specific bundles of design variables indexed b=1 B. Each bundle is comprised of a set of values corresponding to the manipulable parameters in (i.e., engine displacement size and final drive ratio) and the fixed design 11 The ability to reduce weight by applying lighter weight materials is not included as an explicit decision variable in our model. Firms have already applied light-weighting when it is cost-effective under the unreformed CAFE standards. This is implicitly captured in our representation of vehicle design and performance. However, the medium-run design model does not consider the ability of manufacturers to further light-weight their vehicles. NHTSA s (2009) analysis indicates that additional light-weighting is less cost-effective, in terms of fuel-consumption reductions per additional cost, than engine friction reduction or high-efficiency alternators. Our results indicate that the policy change does not significantly affect adoption of these technology features: the vast majority of vehicles (approximately 98% of vehicle sales) have no change in these technology features between the reformed and unreformed CAFÉ simulations. This implies that additional light-weighting is not cost-effective for these vehicles (assuming NHTSA s estimated costs). To the extent that lightweighting strategies may be more cost effective than NHTSA s analysis estimates, our results represent upper bounds of profit losses and consumer surplus losses. This would only reinforce our finding that policy costs are significantly overestimated in many analyses that do not include design responses. 11

12 parameters in (curbweight). Parameter values are varied across segmentspecific bundles at small intervals over the range of possible values. We use the engineering simulation model to calculate the fuel consumption per 100 miles (fuelcons) and 0-60 mph acceleration time (acc) of a particular vehicle design, conditional on the assumed bundle of design parameters. We generate almost 30,000 sets of simulation results, each representing the fuel consumption and acceleration of a simulated vehicle in a vehicle segment, s, with a corresponding bundle of design parameter inputs, b. Additional details of the vehicle simulations are discussed in Appendix A. We could in principle use the results of the engineering simulations to estimate the baseline relationships summarized by equations (1) and (2). However, we can lower computational complexity by exploiting properties of the simulation results to reduce the number of decision variables. Specifically, for any plausible combination of engine displacement and curbweight, there is a unique final drive ratio, x f, that minimizes acceleration time. Because adjustments to the final drive ratio have a negligible effect on production costs, assuming that consumers value faster acceleration, a profit-maximizing firm will choose the final drive ratio that minimizes acceleration time given any desired level of fuel efficiency. 12 Conditional on these assumptions, we find there is a one-to-one mapping between acceleration and engine displacement size. We can thus represent the design tradeoffs captured by equations 1 and 2 as one equation (equation 3). Appendix B discusses this derivation in more detail. Several specifications for equation (3) were estimated using the engineering simulation data for the baseline case (i.e. no additional technology features). The following specification performed the best under the Akaike Information Criterion: 12 Parts can often be swapped out to adjust the gear ratio with approximately equivalent production costs. Many aftermarket vendors (e.g., will sell tail-housings with different final-drive ratios for the same price. 12

13 (5) The parameter wt measures vehicle curbweight, which varies across the bundles of attributes that define the simulations. The term represents the error associated with approximating the calculations performed in the vehicle simulations with this parametric function. To allow for correlation in the residuals across simulations that assume the same curbweight, we cluster by wt. Equation (5) predicts observed vehicle performance reasonably well (R 2 =0.76). Appendix E discusses comparisons between estimated and observed performance attributes in more detail. B. Fuel-saving technology features Having estimated the baseline trade-offs, we now incorporate the technology features that can be added to improve fuel efficiency. The National Highway Transportation Safety Administration (NHTSA) provides estimates of how the addition of specific technology features affects fuel economy in each vehicle segment (NHTSA, 2008). These estimates are based on values reported by automotive manufacturers, suppliers, and consultants. Appendix A explains how we integrate these data from NHTSA (2008) with the engineering simulations in order to determine how the addition of one or more technology features affects the position of the iso-technology curve relative to the baseline described above. Our engineering simulations incorporate most but not all of the medium-run technology features considered by NHTSA (see Appendix A). 13 Some features are omitted because of simulation and data constraints. 14 If one or more of the design 13 The medium-run considered in our analysis assumes that the vehicles engine and transmission architecture are fixed so while NHTSA (2008) considers applying higher speed transmissions, and diesel and hybrid powertrains to a conventional vehicle, these are not included in our model. 14 The ability to apply variable valve timing, variable valve lift, direct injection, or turbocharging to conventional vehicles are not included as explicit decision variables in our model. Firms have already applied these technology features when it is cost-effective under the unreformed CAFE standards. This is implicitly captured in our representation of vehicle design and performance prior to the reformed CAFE standards. However, the medium-run design model does not consider the ability of manufacturers to further apply these technology features to their vehicles. 13

14 features we omit are cost effective in some instances, our estimated cost of complying with CAFE will be biased upwards. If not, our analysis is unaffected. It is computationally infeasible to incorporate all of the discrete simulated technology choices explicitly into our policy simulations. For the purpose of tractability, we introduce two simplifications. First, we consider only those combinations of technology features that are cost effective meaning that there is no lower cost combination that could achieve the same or better level of acceleration performance and fuel efficiency. This reduces the set of technology feature combinations under consideration from to 20-76, depending on vehicle segment. A second simplification approximates the set of cost-effective technology feature combinations with a single variable, tech. The tech variable takes on an integer value between zero (the baseline case) and the maximum number of costeffective combinations of technology features for each vehicle class. 15 Technology feature combinations are ordered by decreasing fuel consumption for the same acceleration time, which is also increasing in cost. Therefore, a higher value of tech corresponds to a lower fuel consumption and higher cost vehicle conditional on acceleration time. We show in Appendix C that the particular specification we use to estimate the relationships of the continuous tech variable to fuel consumption and cost preserves important properties of the discrete technology combinations. We augment equation (5) to accommodate the addition of one or more technology features as follows: (6) 15 Note that a particular value for tech maps to a specific combination of technology features (e.g., low resistance tires and a high efficiency alternator) and does not represent the number of technology features. 14

15 Estimated values of the parameters in equation (6) are reported in Table 3. The model fits the vehicle data in each class well (R 2 >0.89) with the exception of the two-seater class (R 2 =0.44). However, the two-seater class comprises less than 1% of vehicle sales so the poorer fit of this class should not significantly affect the policy simulation results. All parameter estimates have expected signs. The positive sign on the interaction between the technology variable and acceleration indicates that implementing more costly combinations of technology features reduces fuel consumption with decreasing returns, as illustrated in Appendix C. The positive sign on the weight parameter and negative sign on the weight-acceleration interaction imply that the iso-technology curves in Figure 2 shift up and rotate clockwise with vehicle weight. This indicates that heavier vehicles will have worse fuel consumption given the same acceleration time, as expected, but that this effect increases for vehicles with faster acceleration. The different estimates on the constant and acceleration coefficients across segments imply that achieving relatively fast acceleration results in a greater loss of fuel consumption for larger segments (e.g., trucks and SUVs) than for smaller segments (mid-size, compact, and sports cars). Comparisons between model predictions and observed market data are summarized in Appendix E. C. Comparison of engineering versus econometric approaches The engineering-based approach we take to model vehicle design tradeoffs stands in contrast to the econometrically estimated models of similar tradeoffs found elsewhere in the economics literature (Gramlich 2008; Klier and Linn 2012). There are two key advantages to our approach. First, using physics-based simulations to identify the engineering tradeoffs between vehicle attributes allows us to model tradeoffs and attribute combinations 15

16 that are technologically possible, but have not yet been implemented in existing vehicles. This is important because these new combinations of design attributes may become optimal under the policy regimes we are interested in analyzing. In fact, fuel economy standards historically forced the frontier of vehicle design (Klier and Linn, 2012). Manufacturers have stated that they will rely on further advancing this frontier in order to meet the reformed CAFE standards. Second, our engineering approach allows us to isolate the tradeoffs between fuel consumption, acceleration, and production costs without conflating changes in unobserved attributes that typically affect both fuel consumption and consumer demand. Econometric approaches are more limited in their ability to account for correlation between endogenous attributes and unobserved attributes. 16 It is instructive to compare the tradeoffs we estimate using data generated by physics-based vehicle simulations with estimates constructed using an econometric approach. Note that equation (6) can be estimated using either simulated data or performance attributes observed in the marketplace. We estimate this equation first using simulated data, and then using observed vehicle data from MY2006. Figure 3 illustrates the estimated tradeoffs for the SUV segment using both of these approaches. Appendix F discusses the econometric estimation results in more detail. As Figure 3 illustrates, the estimated tradeoff between fuel consumption and acceleration is relatively less steep when we use the econometric approach. In other words, the econometric approach implies that designing a vehicle s powertrain to achieve lower fuel consumption (without adding any additional 16 A concrete example helps to clarify this point. The 2010 Chrysler 300 Touring with a 2.7 L engine option has a combined fuel economy of 21.6 mpg and a 0-60 acceleration of 10.5 s, whereas the 3.5 L engine option has 19.7 mpg and 8.5 s. However, engine options are correlated with unobservable attributes; in addition to a larger engine, the 3.5 L Touring also contains a suite of electronic accessories including anti-lock brakes, electronic traction control, light-sensing headlamps, and an upgraded stereo system. The addition of these extra accessories could consume additional energy, further reducing fuel economy. Moreover, these accessories typically increase demand, violating the certeris paribus assumption in counterfactual policy simulations. 16

17 technology features) leads to a relatively greater loss of acceleration performance. A possible explanation is that vehicles sold in MY2006 that have higher acceleration performance are generally equipped with other luxury trimmings, such as sound-reducing materials and adaptive suspension, that cumulatively can significantly increase fuel consumption. A failure to account for these unobserved attributes can result in a biased (steeper) estimated tradeoff between fuel consumption and acceleration. As noted above, an advantage of the engineering approach is that, provided the underlying simulation model is correctly specified, we can better account for correlations between endogenous attributes and unobserved attributes. IV. Vehicle production costs Conceptually, the production costs of vehicle model j can be decomposed into three parts: (7) The first component represents the portion of production costs that are dependent on medium-run powertrain design decisions. The second component captures the production costs associated with each technology feature. The third component represents the portion of costs that are independent of the medium-run design decisions that are the focus of this paper. A. Cost implications of medium-run design decisions We draw on Michalek et al. (2004) to capture medium-run powertrain costs. The authors use cost data from manufacturing, wholesale, and rebuilt engines with varying displacement sizes to estimate a model that relates engine production costs to engine displacement x d and fixed design characteristics : 17

18 (8), Similar to the derivation of equation (5) discussed in section III, we map engine displacement to acceleration and other vehicle attributes in order to reduce the number of decision variables and thus lower computational complexity. Maintaining the conditions that firms maximize profits, consumers value lower fuel consumption and faster acceleration, and adjustments to the final drive ratio have a negligible impact on production costs, there is a one-to-one mapping between engine displacement and segment-specific combinations of acceleration performance and fixed design parameters in. We can thus write: (9),, Substituting equations (9) into (8), we can construct estimates of engine costs for each bundle of attributes b within each vehicle segment s. To summarize the relationship between engine costs and medium-run vehicle design decisions, we regress our cost estimates on the determinants of engine displacement. (10) Many parametric forms of this relationship were tested and the specification in equation (10) performed the best using the Akaike Information Criterion. B. Fuel-efficiency improving technology costs Production costs associated with the addition of specific technology features are taken from NHTSA s (2008) analysis, shown in Table 1, which were used in costbenefit analyses of the reformed CAFE regulations. NHTSA generated these vehicle-segment specific estimates using reported values from automotive manufacturers, suppliers, and consultants. 18

19 We treat the costs of adjusting powertrain tuning variables and costs of technology features as additively separable. Engines are manufactured separately from other subsystems of the vehicle before assembly. Most of the technology features we consider do not require changes in engine design or affect the assembly of the engine with other vehicle subsystems. This is consistent with our assumption that costs are additively separable. There are only two exceptions engine friction reduction and cylinder deactivation that do affect the engine subsystem. Even in these cases, it is reasonable to approximate technology costs as additively separable from the baseline production cost of the engine. For example, engine friction can be reduced by using higher-performing lubricants, the costs of which are independent of the medium-run powertrain decisions considered. 17 Combining these two sets of cost estimates, equation (11) summarizes the production cost implications of the medium-run design decisions we consider: (11) Estimation results are summarized in Table 4. These estimates fit all vehicle classes reasonably well (R 2 >0.83). Production costs increase with the level of technology implementation by design. As expected, production costs decrease as acceleration times get slower. The positive sign on the weight term and negative sign on the weight-acceleration interaction term indicate that incrementally improving acceleration is more costly in heavier vehicles, and this effect is magnified for vehicles with relatively faster acceleration times. All parameter coefficient estimates are statistically significant at the 90% level. 17 The case of cylinder deactivation poses a larger challenge for treating technology costs as additively separable from engine costs. Given large changes in engine displacement achieved by switching the engine architecture (e.g., replacing a V-8 engine with a V-6) would slightly reduce the costs of cylinder deactivation due to a smaller number of cylinders. However, even with this cost reduction, cylinder deactivation is the highest-cost technology feature considered and therefore would not significantly affect counterfactual results. 19

20 C. Baseline production costs The final vehicle production cost component, ω, does not depend on mediumrun design decisions. This cost component is derived from the first-order conditions of firms profit maximization. This derivation, discussed in detail in Appendix D, follows Jacobsen (2012). V. Vehicle demand In this section, we introduce an empirical model of vehicle demand. Our specification builds on the seminal work by Berry et al. (1995), and subsequent work by Berry et al. (2004), Train and Winston (2007), Langer (2011), and others. A. Data sources We use two sources of data to estimate the model of vehicle demand: a household-level survey conducted by Maritz Research, and vehicle characteristic data available from Chrome Systems Inc. The Maritz Research U.S. New Vehicle Customer Study is a monthly survey of households that purchased or leased new vehicles. This survey provides information on socio-demographic data, household characteristics, and the vehicle identification number (VIN) for the purchased vehicle. The survey also asks respondents to list up to three other vehicles considered during the purchase decision. We use data from September August 2006 (what we call model-year or MY2006). Approximately one-third of respondents listed at least one considered vehicle. Because the survey oversamples households that purchase vehicles with low market shares, we take a choice-based sample from this data such that the shares of vehicles purchased by the sampled households matches the observed MY2006 market shares. We supplement the survey data with information on vehicle characteristics using Chrome System Inc. s New Vehicle Database and VINMatch tool. Vehicle 20

21 alternatives are identified using the reported VIN, distinguishing vehicles by their make, model, and engine option, with a few modifications. We eliminate vehicles priced over $100,000, which represent a small portion of market sales, and remove seven vehicle alternatives that were not chosen or considered by any survey respondent. We further reduce the data set by consolidating pickup truck and full-size van models with gross vehicle weight ratings over 8,000 lb to only two engine options each. The remaining set of vehicle models and engine options represent 97% of vehicle sales. Summary vehicle data are described in Table 2. B. Empirical model Following Berry et al. (1994), we model vehicle demand at the household level and then aggregate up to the market level. New automobiles are described as bundles of attributes. Consumers are assumed to choose the vehicle that maximizes the utility derived from these attributes. As noted above, the design parameters x i and technology features t i affect utility only indirectly through their effect on vehicle attributes such as fuel efficiency and acceleration. The utility obtained when no new vehicle is purchased is defined as u i0. Conceptually, the utility that consumer n derives from vehicle model j can be decomposed into four components, defined as: (12) The first component, δ j, represents the portion of utility for a specific vehicle model that is common across all consumers: (13) 21

22 The are observable vehicle attributes, such as price and fuel efficiency, and the attribute-specific coefficients are common across all consumers. The captures the utility of vehicle attributes valued by the consumer, but not observed by the econometrician. Examples include vehicle handling, acoustic performance, and electronic accessories. One might expect the to be correlated with the vehicle attributes of primary interest: vehicle price, fuel efficiency, and acceleration. This gives rise to a well-documented endogeneity problem. Following Berry (1994), we exploit the linearity of equation (13) in. We note that the endogeneity of observable attributes such as price and fuel efficiency can be dealt with using two-stage least squares provided we have valid instruments. The second component in equation (12) captures the component of utility that varies systematically with observable consumer characteristics z n. Interactions between consumer characteristics and vehicle attributes play an important role in determining substitution patterns. Our specification includes interactions between price and income, children in the household and SUV and minivan segments, and allows preferences for pickup trucks to vary across rural and urban areas. The third component captures the effects of interactions between vehicle attributes a j and household characteristics we cannot observe. This allows for some random variation in consumer preferences for specific vehicle attributes. The random coefficients are assumed to be distributed normally in the population according to the distribution f(. Finally, the disturbance term ϵ nj captures idiosyncratic individual preferences which are assumed to be independent of the product attributes and of each other the effects of unobserved determinants of utility that vary randomly across consumers. We assume that these idiosyncratic errors have an i.i.d. Type I extreme value distribution. This assumption yields the familiar logit functional 22

23 form for the vehicle choice-share probabilities, P, conditional on z, v, and the parameters to be estimated, θ, as shown in equation 14. (14) P Pr,,θ exp 1 exp The predicted market share of a vehicle, j, is. Following Train and Winston (2007), the utility formulation is extended to include information about ranked choices when these data are available for a respondent. The ranking is specified as u u u u for all j i, h,,h where i is the chosen vehicle; h 1 is the second ranked vehicle (the vehicle that would have been chosen if vehicle i was not available) and h m is the m ranked vehicle. Therefore, the probability that respondent n purchased vehicle i and ranked vehicle h 1 through h m is defined as: (15)...,,, The first argument represents the probability that the consumer purchased vehicle i, given all available vehicle models and the outside good. The second argument represents the probability that the consumer would have purchased vehicle h 1 if vehicle i and the outside good were not available, and so on. 18 When no ranking data are available for a respondent, the likelihood consists of only the first term in equation (15). C. Identification strategy 18 The outside good is excluded from the denominator of every term but the first because we do not observe whether the respondents would have chosen not to purchase a vehicle if their first choice was not available. 23

24 One distinguishing feature of our demand estimation pertains to our choice of instruments. It has become standard in the literature to use functions of nonprice attributes, including vehicle dimensions, horsepower, and fuel economy, as instruments for price (e.g., Berry et al. 1995; Berry et al., 2004; Train and Winston 2007). The validity of these instruments is defended on the grounds that, if a researcher is exclusively concerned with short-run pricing decisions, attributes such as fuel efficiency can be considered exogenous to the analysis. This approach to instrumentation has been criticized on two accounts. The first has to do with the timing of vehicle design decisions. Automotive manufacturers can adjust fuel economy and acceleration somewhat in the same time frame as setting suggested retail prices. 19 A second concern is that car manufacturers choice of unobservable vehicle trimmings, such as adaptive suspension, may depend on previously determined attributes, such as vehicle dimensions or horsepower, rendering these attributes invalid as instruments. Our empirical strategy addresses the first concern directly. We use only those vehicle attributes that are determined by longer run product-planning schedules to instrument for variation in price, fuel efficiency, and acceleration performance. 20 More precisely, we use the moments of vehicle dimensions of same-manufacturer vehicles (din, dsqin) and different-manufacturer vehicles (dout, dsqout), powertrain architecture (i.e., hybrid, turbocharged, and diesel), and drive type (i.e., all wheel drive or 4-wheel drive). These design features can be considered fixed in the medium run. Our strategy does not directly address the second concern. In order for our instruments to be valid, we must assume that these variables are uncorrelated with unobserved attributes that consumers value. D. Demand-side estimation 19 For example, by adjusting the programming of the electronic control unit in the powertrain. 20 Literature detailing the automotive design process allows us to address the first criticism of instrument choice, described in Section II. A remaining assumption in our approach is that these longer run attributes are not correlated with choices of unobserved attributes in the medium run. 24

25 Table 5 reports our estimated demand model parameters. Several of the parameters that capture the effects of interactions between vehicle attributes and observable consumer attributes are found to be statistically significant, including the ratio of price to income (p/inc), and the interactions between minivans and children, SUVs and children, and pickup trucks and living in a rural location. Only the standard deviation of the fuel consumption coefficient is found to be statistically significantly different from zero. It has come to light that these kinds of random-coefficient discrete choice demand models can yield parameter estimates that are very sensitive to the choice of optimization routine and starting values (Knittel and Metaxoglou 2008). In light of these concerns, we estimate the model using a series of randomly selected initial values to test the robustness of our estimates. These results are reported in Appendix G. All of the initial values we consider deliver the same estimate solutions within 1e Infinite-norms of the gradients for each solution were on the order of 1e-3 to 1e-4, and the Hessians at these solutions were all verified to be positive definite, verifying the algorithm converges to the maximum likelihood. 22 Table 6 summarizes the first-stage estimation results. Endogenous choice variables are price, fuel consumption, and inverse 0-60 mph acceleration time. The first stage F-test statistics are 20.68, 19.62, and 21.38, respectively. Table 7 reports the second-stage IV estimates. The SUV indicator variable is positive and significant suggesting a preference for SUVs over the omitted category (sedans) when the observable attributes are controlled for. The minivan indicator is negative and significant. The parameter for two-seater sports cars is negative and the pickup-truck indicator is slightly positive, but neither is significant. 21 Initial values were randomly selected from a uniform distribution from Initial values outside of this range were also tested but these initial points often produced values of the log-likelihood that were near negative infinity, indicating a very poor fit to the model, which prevents the algorithm from solving the estimation problem. 22 The algorithm minimizes the negative log likelihood (equivalent to maximizing the log likelihood) thus a positive definite hessian satisfies the second order conditions. 25