Dynamic Efficiencies of the 1997 Boeing-McDonnell Douglas Merger

Transcription

1 Dynamic Efficiencies of the 1997 Boeing-McDonnell Douglas Merger Yonghong An Wei Zhao September 26, 2017 Abstract This paper evaluates the welfare effects of the 1997 merger between Boeing and McDonnell Douglas in the medium-sized, wide-body aircraft industry. To quantify merger efficiencies, we first develop an empirical model of multi-product firms, allowing for learning-by-doing in a dynamic game. This model allows us to disentangle merger efficiencies generated from accelerated learning-by-doing from market power effect. We find that merger efficiencies outweighed market power effect for the Boeing-McDonnell Douglas merger. Taking account of all static and dynamic effects, we find that net consumer surplus increased by as much as $5.14 billion. By contrast, a static model ignoring learning-by-doing predicts a consumer loss of approximately $7.61 billion. These results indicate that failing to account for dynamic effects can lead to biased results and to erroneous policy decisions regarding the welfare effects of proposed mergers. Keywords: Learning-by-doing, multi-product firms, dynamic merger efficiencies. JEL classification: K21; L13; L41 Zhao is extremely indebted to his advisors Joseph E. Harrington, Jr., Przemyslaw Jeziorski, and Richard Spady for their continual guidance, support, and encouragement. We are grateful to C. Lanier Benkard for making the labor input data of Lockheed L-1011 available. We deeply appreciate Edmund S. Greenslet, publisher of The Airline Monitor, for numerous inspiring discussions about the aircraft industry and the merger. Also, we would like to thank C. Lanier Benkard, Jiawei Chen, Yingyao Hu, Elena Krasnokutskaya, Yiyang Ellen Li, Yao Luo, Xun Tang, Michael A. Williams, Jonathan Wright, Ruli Xiao, and participants of seminars at Johns Hopkins University for helpful comments, discussion, and suggestions. All remaining errors are ours. Department of Economics, Texas A&M University, College Station, TX 77843; y.an@tamu.edu. Competition Economics LLC, 2000 Powell Street, Suite 510, Emeryville, CA 94608; wzhao@cecon.com. 1

2 1 Introduction A primary benefit of mergers to the economy is their potential to generate significant efficiencies... which may result in lower prices, improved quality, enhanced service, or new products. (2010 U.S. Horizontal Merger Guidelines) One fundamental question in antitrust practice is to evaluate the short-run impact of mergers on consumer welfare. Comparing pre-merger prices with forecasts of post-merger prices is the dominant approach to this evaluation, e.g., see Nevo (2000). These forecasts are often obtained in a static way by holding firms costs and the qualities of their products fixed. 1 However, mergers may generate economies of scale in production due to learningby-doing, i.e., efficiencies. Such merger efficiencies are realized over time dynamically and are endogenous to firms decisions in the post-merger environment. The most significant impacts of such mergers on consumers welfare are how the post-merger market structure affects product cost over time, rather than how they affect post-merger prices. Despite their importance, few empirical studies address dynamic efficiencies in evaluating mergers. The static approach mentioned above is not applicable without knowing the changes of costs due to a merger. We aim to fill the gap in this paper: we evaluate the welfare effects of the 1997 merger between Boeing and McDonnell Douglas in the aircraft industry, where dynamic efficiencies may be present due to learning-by-doing. Our analysis focuses on the market of medium-sized wide-body (hereafter medium-sized ) aircraft. In the pre-merger market, Boeing and McDonnell Douglas were the first and third largest producers of civilian jets, respectively. Their merger was approved in August 1997 and Boeing acquired McDonnell Douglas. After the merger, only Boeing and Airbus remain as major competitors in the global market for large commercial aircraft. The primary impact of the merger on market structure was the elimination of McDonnell Douglas, whose only wide-body product in production was the MD-11, a medium-sized aircraft. Shortly after the merger, the Boeing company shut down production of the MD-11. We begin our analysis by investigating the change in aircraft prices after the merger. In five years after the merger, the medium-sized aircraft produced by both Airbus and Boeing experienced significant price drops. Meanwhile, the prices of other wide-body aircraft declined only slightly and narrow body aircraft prices rose. Our regressions show that the annual decrease in prices for medium-sized aircraft was 2.36 millions dollars larger in the post-merger period than in the pre-merger period. Focusing only on the post-merger period, the prices of medium-sized aircraft dropped 4.46 million dollars more each year than prices of all other aircraft did. These results suggest that the merger might have generated some efficiencies that offset the market power effect in the post-merger market. However, the reduced-form analysis does not allow us to disentangle dynamic efficiencies from market power effect. To further explore the merger s dynamic efficiencies, we set up a dynamic oligopoly game to describe the market of medium-sized aircraft. We incorporate in the model 1 Although there are some variants to this approach, e.g., it may be recognized that some products would be removed or some immediate cost reductions would be realized, the evaluation still takes the form of a static analysis. 2

3 learning-by-doing, a common dynamic forces relevant to industry performance. 2 Multiproduct firms compete in an infinite-horizon dynamic game. In each period, a firm chooses how much to produce for each product. The production decision affects both current and future profit streams through its impacts on the firm s experience level. Experience is a state variable; a firm s past output not only accumulates (learning), but also depreciates over time (forgetting). We model learning-by-doing so that the unit production cost is a decreasing function in experience. Moreover, a firm s production is allowed to have spillover effects in terms of experience accumulation from that firm s other products and its competitors products. The magnitude of these spillover effects could depend on ownership and product characteristics. Firms are assumed to behave according to a Markov Perfect Equilibrium in which they decide on production in each period given the state variables of firms experiences and the stochastic realization of market size and product characteristics. In the model above, the Boeing-McDonnell Douglas merger may hurt consumers because reduced competition creates an incentive for Airbus and Boeing to restrict production and raise prices; this is the traditional market power effect. However, the merger also may generate dynamic efficiencies in several ways. First, there may be an immediate benefit of lowering the marginal cost for Boeing products because of the one-time experience transfer from McDonnell Douglas after the merger. Second, Boeing chooses to shut down the MD-11 shortly after the merger, and the effect is twofold. On the one hand, fewer products mean less variety, which makes consumers worse off. On the other hand, there would be more demand for other medium-sized aircraft, e.g., Boeing 777. More production leads to faster experience accumulation, lower unit cost, and lower future prices, which makes consumers better off. Third, future experience might be shared more effectively between different products within the same firm (within-firm spillover) than between different firms (across-firm spillover), again bringing the costs down after the merger. To quantify the welfare effects of the Boeing-McDonnell Douglas merger, we employ a multi-step procedure to solve the dynamic game and simulate the paths for prices, quantities, profits, and consumer surpluses. The first step is to estimate the demand and cost parameters of the model. The static demand system is captured by a nested logit model and the parameters are estimated using a two-stage least squares (2SLS) approach. Based on these demand estimates, a static merger analysis following the methodology in Nevo (2000) indicates that consumer surplus will drop 15.24% after the merger. The primary detrimental effect of the merger comes from shutting down MD-11. If MD-11 is continued in the post-merger market, then consumer surplus only declines by 2.2%, even if there is no cost reduction due to the merger. Next, we estimate the total variable cost function, the learning curve, and the state transition process of experience with a generalized method of moments (GMM) approach using the data on Lockheed L Traditional industries benefiting from learning-by-doing include aircraft, shipbuilding, semiconductors, fuel cell vehicles, oil drilling, photovoltaics, machine tools, metal products, nuclear power plants, and chemical processing, etc. Recent works in estimating learning-by-doing include Benkard (2000) for aircraft, Thompson (2001) and Thompson (2007) for shipbuilding, and Gowrisankaran et al. (2006) for surgical procedures. See Besanko et al. (2010) for a complete list of learning-by-doing estimation works. 3

4 Our estimates indicate substantial cross-firm spillover: building four aircraft of a different type is as helpful in experience accumulation as assembling one of the same type. The cross-firm spillover is estimated to be negligible. Finally, we solve the dynamic game based on the estimates of demand and cost parameters, and evaluate the efficiencies of the merger. We consider three different scenarios: (i) merger occurs and the MD-11 is shortly shut down, which actually happened; (ii) merger occurs with continued operation of the MD-11; and (iii) no merger. In simulating postmerger firm choices, we also incorporate possible one-time cost synergy as an experience stock transfer from MD-11 to Boeing 777. The time series for equilibrium prices, quantities, profits, consumer surplus, and total surplus are computed for all three scenarios. By comparing scenarios (i) to (iii), we quantify the merger s effect, which is the net effect of the market power and accelerated learning-by-doing. To fully disentangle the efficiencies of learning-by-doing from market power effects, we solve an additional static model that does not allow for learning-by-doing in the three scenarios above. In this static model, a comparison between scenario (i) and (iii) gives the net effect of market power. The difference of the net effects in the dynamic and static models is the impact of learning-by-doing. Our main finding is that the Boeing-McDonnell Douglas merger did bring dynamic efficiencies, which mainly come from faster learning-by-doing after the merger. The dynamic efficiencies generated by the merger are large enough to offset the static market power effect, so the merger is beneficial to consumers. Specifically, without any one-time transfer of experience, the merger would increase consumer surplus by $0.11 billion; the increase is $5.14 billion if experience is transferred completely. By contrast, a static equilibrium model without learning-by-doing predicts a $7.61 billion loss of consumer surplus. All of these qualitative results survive a battery of robustness checks. The main contribution of this paper is to quantify the dynamic efficiencies in the Boeing-McDonell Douglas merger and to demonstrate that a static merger analysis may lead to biased results and erroneous policy decisions. Gowrisankaran (1999) was one of the first papers that theoretically examined the dynamic effects of a merger with numerical analyses within the Ericson-Pakes framework (Ericson and Pakes, 1995), but the impact on consumer welfare was not his central concern. Chen (2009) also examined similar issues theoretically. Stahl (2009) focused on the consolidation process in the broadcast television industry but did not evaluate merger-generated efficiencies. More recently, Benkard et al. (2010) evaluated the medium- and long-run dynamic effects of airline mergers and explored the effect of mergers on market structure, not consumer welfare. Jeziorski (2014a,b) studied merger impacts in the U.S. radio industry, taking account of the markets being two-sided. Nocke and Whinston (2010) provided a new theoretical framework to model dynamic merger decisions where firms choice variables other than merger decisions were assumed to be static. Our paper distinguishes itself from the aforementioned literature by introducing a model that focuses on the endogenous dynamics of cost. Our method and findings shed light on merger evaluation in aircraft and other industries where dynamic efficiencies may exist, especially when the efficiencies are generated from learning-bydoing. Our dynamic oligopoly model and empirical analysis also contribute to a broader lit- 4

5 erature on learning-by-doing encompassing a wide array of industries. The most closely related papers are the pioneering research of Benkard (2000, 2004). Benkard (2000) introduced the concept of forgetting to explain the rise in cost for the Lockheed L Benkard (2004) allowed for a learning curve in a dynamic oligopoly model with four single-product firms, estimating welfare under several counterfactual scenarios with a social planner and a monopoly. Our paper follows this methodological path but differs from these papers by applying the analysis to evaluate effects of an actual merger and by explicitly modeling the spillover effect of learning curves. Specifically, in addition to the within-model spillover assumed in Benkard (2004), we allow for the existence of a cross-product or cross-firm spillover effect. In our model, merger efficiencies are likely to arise through accumulation of experience due to combining output and the spillover effect. Our empirical results of learning curves confirm the existence of cross-product spillover. Although the spillover effect of the learning curve has not been widely investigated for the aircraft industry, 3 it has been modeled and estimated for other industries, including semiconductors (Irwin and Klenow, 1994), shipbuilding (Thornton and Thompson, 2001), fuel cell vehicles (Schwoon, 2008), steel (Ohashi, 2005), and health care (Chandra and Staiger, 2007). However, those papers are not targeted at evaluating mergers in the context of a dynamic game. And, none of them simultaneously estimated within-firm spillover and across-firm spillover, which could be significant factors in calculating merger efficiencies, as our estimation results indicate. Another difference between our paper and Benkard (2004): we solve an oligopoly dynamic model with multi-product firms while in Benkard (2004) each firm only produces a single product. Multi-product firms in a dynamic model impose significant challenges to empirical analysis, and we hope that our strategies will be useful for other similar applications. Finally, we focus on the medium-sized wide-body aircraft industry, while Benkard (2004) analyzes the entire wide-body aircraft industry. The remainder of the paper is organized as follows. Section 2 introduces the industry background and the data, and then provides a descriptive analysis of the merger s effect on prices. Section 3 presents the dynamic oligopoly model. Section 4 estimates the demand system and analyzes the static merger effects, and Section 5 estimates total variable costs and learning curves. Section 6 evaluates the dynamic merger effects, and Section 7 concludes. 2 The Industry and Data 2.1 The medium-sized aircraft market A wide-body aircraft is a large jet airliner with two passenger aisles. Following the introduction of the first wide-body aircraft, Boeing 747 in 1969, only four firms (Airbus, Boeing, Lockheed and McDonnell Douglas) were active in the industry. Lockheed exited the market in Nine wide-body aircraft (A300, A310, A330, A340 and A380 of 3 The international trade literature studies knowledge spillover in the sense of technology transfer across countries and across industries. See Grossman and Helpman (1995) for a review of that literature and Niosi and Zhegu (2010) for a review of the aircraft industry specifically. 5

6 Airbus, B747, B767 and B777 of Boeing, and MD-11 of McDonnell Douglas) were in production during the period. Wide-body aircraft are further split into small (around 250 seats), medium (around 300 seats), and large (around 450 seats) categories by the number of seats and the maximum flying range, as illustrated in panel (a) of Figure 1. The horizontal line in the figure marks the nautical distance between Beijing and New York, which is used as a benchmark separating transatlantic and transpacific routes. The figure shows that medium and large aircraft have longer range and are more suitable for transpacific routes than small aircraft. The primary impact on market structure of the 1997 Boeing-McDonnell Douglas merger was the elimination of McDonnell Douglas. Its only wide-body aircraft in production was the MD-11, a medium-sized aircraft. Prior to the merger, there were three other models in the medium-sized group: A330, A340 and B777. In 1997, the quantity share of medium-sized aircraft was 31.7% among wide-body aircraft. In the medium-sized group, the quantity shares of Airbus, Boeing and McDonnell Douglas were 40.7%, 22.2% and 37.1%, respectively. Boeing inherited the MD-11 and three narrow-body aircraft models (MD-80, MD-90, and MD-95) from McDonnell Douglas after the merger. Several months later, the new Boeing company decided to phase-out the MD-11, which was a competitor of the B777, and later introduced several submodels of the B777. Airbus also introduced submodels of the A330 and A340 after the merger. From , there were twelve submodels of the four models in the medium-sized category. 4 The six Airbus submodels were A , -300, A , -300, -500, and Their combined quantity share in the medium-sized group was 50% in That same year, Boeing s five submodels B , -200ER, -200LR, -300, and -300ER accounted for the remaining 50% quantity share. In Table 1 we summarize some important characteristics of these submodels, based on the data that we describe later in this section. The table shows that the majority of the submodels (seven out of twelve) were introduced after the merger. MD-11 is the first product in the medium-sized group, while B777 is the last to enter the market. Number of engines is an important characteristic because it is a key indicator of fuel efficiency. Twin-engine aircraft generally are more efficient than aircraft with more engines. We maintain the assumption that the medium-sized aircraft compete in an independent market, and restrict our analysis to this market. To examine the overlap between medium-sized aircraft with the other two wide-body groups, we define a ratio R w for any route as the number of medium-sized flights over the total number of wide-body flights. If R w is close to 0 or 1 for a route, then the medium-sized aircraft hardly compete with other wide-body aircraft; R w close to 1/2 indicates medium-sized aircraft are actively competing with aircraft in the other two wide body groups. A histogram of the resulting 4 Other than the aircraft shown in panel (a) of Figure 1, Boeing introduced B787 in 2011 and Airbus answered with A350 that entered the market in The B787 was originally introduced to replace the now defunct B and B and it was originally called Boeing Y2. It may eventually supplant the B (the shorter range version). We treat B787 and A350 as new generation upgrades of B777 and A330 respectively in our extended model with production upgrade (see the online appendix). In this sense, there is more than one aircraft model number (e.g. A330, A350) matching the same product in the model due to generation upgrade. We will still call these products B777 and A330 for simplicity whenever there is no ambiguity. 6

7 R w is presented in panel (b) of Figure 1, illustrating that the ratio concentrates at 0 and 1 and our assumption of independent market is valid: only 38.5% of the routes are in the interval [0.2, 0.8] and 26% in [0.3, 0.7]. 2.2 Data sources The data on quantity, price, market size, and characteristics for aircraft from 1991 to 2009 mainly come from Airline Monitor, Avmark, and official websites of Boeing and Airbus. The quantities of aircraft are constructed using the annual fleet and deliveries data from the Airline Monitor. The market size is approximated by the total number of used and new wide-body aircraft with data from the Airline Monitor. Annual average aircraft value data for each type is provided by Avmark and are used as plane prices. The characteristics of planes, including number of seats, maximum range, number of engines, fuselage, empty operating weight, first flight year, and fuel efficiency, come from the official websites of Boeing and Airbus, as well as various online sources. Another characteristics of aircraft, the product generation, is determined using the fuel efficiency data from the Airline Monitor and operating cost difference claims reported in Boeing and Airbus newsletters. Jet Airliner Production List provides the first flight date of every wide-body aircraft produced, which we take as the date of production. Production rates and experience are constructed using quantity data and date of production. The cost information of aircraft mainly comes from the production data of Lockheed L-1011, the third wide-body airliner to enter commercial operations, after the Boeing 747 and the McDonnell Douglas DC-10. Specifically, the direct man hours incurred by Lockheed in the production of each L-1011 aircraft are used in Benkard (2000) for labor input. 5 Some cost and demand shifters are also required for our analysis. These shifters include the present and lagged terms of U.S. manufacturing wage rates from the Bureau of Labor Statistics, aluminum prices from IMF s International Financial Statistics Online Database, present and lagged terms of world and regional GDP from IMF s International Financial Statistics Online Database and oil price data from the Energy Information Administration. 2.3 A reduced-form analysis As we discussed in introduction, the dynamic efficiencies due to learning-by-doing may offset the market power effects such that post-merger prices decline. Therefore, declined prices after the merger might suggest the existence of efficiencies generated by the merger. In this section, we investigate whether prices of medium-sized aircraft declined due to the merger, by a descriptive analysis and then difference-in-difference regressions. First, we plot in Figure 2 prices of both narrow and wide-body aircraft, with deliveries before and/or shortly after the merger. The vertical line marks the Boeing-McDonnell Douglas merger and the prices cover the period from 1991, the year that medium-sized aircraft (MD-11) were first delivered, to 2002, the fifth year after the merger. The two 5 We are grateful to C. Lanier Benkard for making this data available. 7

8 subplots on the top row illustrate significant price drops for medium-sized aircraft produced by Airbus (left) and Boeing (right). In five years after the merger, the average prices of these two groups decrease by 23.4% and 18.4%, respectively. During the same time period, the average price of other wide-body aircraft declines 3.9% as illustrated in the right subplot at the bottom, while narrow body s average price rises 5.6%. One explanation of the discrepancy in price drop after the merger is that the merger brings efficiencies that lower the costs of medium-sized aircraft. Although the merger could cause prices to rise due to the increased market power of Airbus and Boeing, the substantially declining prices indicate that efficiencies may have dominated the effects of market power for medium-sized aircraft. An alternative explanation is that medium-sized aircraft are at their earlier stages of production compared to others; thus the effects of learning-by-doing on marginal costs are larger, and this leads to more rapid price decreases. To distinguish these two alternative explanations, we use difference-in-difference regressions by regressing the first difference of price on dummy variables post-merger, medium, their interaction, and other covariates that may affect changes of prices. p j,t p j,t p j,t 1 = β 1 I{post-merger} t + β 2 I{medium} j + β 3 I{post-merger} t I{medium} j + β 4 H jt + β 5 d jt + β 6 D jt + β 7 d 2 jt + β 8 Djt 2 + β 9 GDP t + β 10 Oil t + β 11 I{Airbus} j + β 12 I{Boeing} jt + φ t + ɛ jt, (1) where I{post-merger} t = 1 for the years from 1998 to The variable H jt is the number of years product j has been in the market at year t since its first delivery. d jt and D jt are number of aircraft j delivered in year t and the total delivered number up to year t, respectively. The three variables H jt, d jt and D jt capture the effects of learningby-doing as well as forgetting. Including H jt mitigates the concern that medium-sized aircraft learn faster because they are at earlier stages of production. We use the growth rate of world GDP and price change of oil to control the effects of the demand side on the price changes. A linear time trend φ t accounts for the downward trend in price changes. In specification (1), β 1 + β 3 captures differences in price changes before and after the merger for medium-sized aircraft; β 2 + β 3 compares differences in the post-merger price changes between medium-sized and other aircraft. Our sample for the regression includes a panel of 32 submodels of aircraft, spanning the period 1958 to 2011: 14 submodels are narrow body, and of the remaining 18 widebody submodels, 12 are medium-sized. We employ a pooled OLS estimator and present the results in Table 2. The estimated coefficients in the complete specification (the third column) indicate that the annual decrease of price for medium-sized aircraft is 2.36 million dollars ( million dollars) more in the post-merger period than the pre-merger period. Focusing only on the post-merger period, the prices of medium-sized aircraft drop 4.46 million dollars ( million dollars) more every year than prices of all other aircraft. Formal tests demonstrate that these effects are significant at the 1% significance level (p-values are less than 0.01). Therefore, (1) across time, medium-sized aircraft experience faster price decreases in the post-merger period than before it; and (2) across 8

9 aircraft, the prices of medium-sized aircraft drop more when we only focus on the postmerger period. The estimates also indicate that product history H jt does not affect the price changes significantly. This implies that the price drops observed in Figure 2 is not simply because the medium-sized aircraft are at their earlier stages of production. By focusing on all the wide-body aircraft rather than just the medium-sized ones in the regression (replace I{medium} j by I{wide} j ), we obtain similar results but the scale of the annual price decrease is smaller, as shown in the last three columns in Table 2. To check the robustness of our regression results, we redefine the dummy variable I{post-merger} t by varying the post-merger period. The regression results do not change qualitatively as long as the post-merger period is less than seven years. In summary, we conclude from the estimates in Table 2 that prices of medium-sized aircraft declined significantly after the merger, and such declines cannot be simply explained as medium-sized aircraft were at their earlier stages of production, but may also be results of merger efficiencies. The merger efficiencies might be large enough to offset the effects of increased market power on prices of medium-sized aircraft. The sources of these merger efficiencies could be (1) accelerated accumulation of experience through producing more medium-sized aircraft after MD-11 was phased out, i.e., accelerated learning-bydoing (e.g., see Benkard, 2000) and/or (2) one-time experience transfer for Boeing after the merger, e.g., Boeing consolidates facilities to improve overall management after the merger, and the company said the plan would bring lower costs and keep Boeing competitive (Aviation Week & Space Technology, March 30, 1998). The impact of increased market power depends on consumers preference. It is essential to disentangle dynamic efficiencies from market power effects and to further quantify both for the purpose of merger evaluation. However, the data on prices and quantities only do not allow such quantification without modeling consumers demand and firms strategies. Thus in the remainder of this paper we first propose a structural model for the market of medium-sized aircraft, and then estimate the model and evaluate the welfare effects of the merger. 3 The Model 3.1 Setup This section presents a dynamic framework for the market of medium-sized aircraft as the basis of our dynamic merger analysis. Our framework closely follows Benkard (2004) in model setup, transition of state variables and estimating strategies, and adds two important features (1) firms produce multiple products, and (2) experience are transferable between different products (models or submodels), both within and across firms. The industry is composed of I 2 multi-product firms competing in discrete time over an infinite horizon, t = 0, 1, 2,,. Firm i owns a subset of the available J( I) products, which is denoted as J i with a cardinality J i. The quantity and price of product j at time period t are denoted as q jt and p jt, respectively. The market structure of the industry at t is characterized by the state vector ω t = (ω 1t, ω 2t,, ω Jt, M t ). The state variable of a product j, ω jt = (E jt, ξ jt ) consists of two components. E jt is the 9

10 (production) experience level of product j, which evolves endogenously as a function of the past experience and the current production of the industry, and ξ jt summarizes the unobserved characteristics of product j. M t is the overall market size, a common state variable shared by all the products. In addition, product j s observed characteristics are exogenously given and denoted as a vector Z jt. All firms decide the quantities of their products. In every period, the game can be divided into two stages: First, the nature draws shocks on demand (M t and ξ jt ) and all draws are immediately observed by all firms. Second, firms compete in a simultaneous quantity competition game. Then the experience level for each product is realized based on the quantity choices and is revealed to all firms. By an abuse of notation, we define firm i s decision in quantity choices at t as a vector Q it (q i1t, q i2t,, q iji t). The industry quantity vector at t is then denoted as Q t = (Q 1t, Q 2t,, Q It ) (Q it, Q it ). The flow profit of a firm from product j is π j (Q, ω) = p j (Q; Z j, ξ j, M)q j C j (q j ; E j ), (2) where we suppress the index t whenever there is no ambiguity. p j (Q; Z j, ξ j, M) is the inverse demand function. The cost C j (q j, E j ) is the sum of a fixed cost and total variable cost T V C(q j, E j ), which relies both on the quantity and the experience level. Let ρ denote the discount factor of firms, then the joint optimal quantity policies for firm i are solved from the following maximization problem: { } max Qi π j (Q i, Q i, ω) + ρ V i (ω )F(ω ω, Q)dω, (3) j J i where V i (ω) denote firm i s value function and it is the solution to the Bellman equation: { } V i (ω) = max π j (Q i, Q i, ω) + ρ V i (ω )F(ω ω, Q)dω. (4) Q i j J i F(ω ω, Q) is the transition process of ω, and ω is the state variable in the next period. This model incorporates both learning-by-doing and firms forward-looking behavior. The quantity choices affect dynamic market cost structure through the mechanism of learning-by-doing. Such a model is also applicable to many industries in which learningby-doing is important, e.g., manufacturing industries such as aircraft, shipbuilding, semiconductors, and chemical processing. In the dynamic game described above, firms maximize their expected discounted value of profits by choosing the joint optimal policy (a sequence of quantities at each time period), conditional on their expectations of the evolution of competitors. We use the Markov perfect Nash equilibrium (MPE) for our dynamic model. We refer interested readers to Benkard (2004) for details of issues related to the existence of MPE and multiplicity of equilibria. We assume away entry and exit decisions on both firm and product levels in our model. Nevertheless, we allow a firm to switch any of its product to a potential entrant good by setting the quantity of that product to be zero in any period, and to reverse the process by setting a positive quantity in any future period. We elaborate on several reasons why 10

11 there is no need to directly model entry and exit. First, it requires huge initial capital and a complete set of frontier technologies to start a new business. The state of the art technologies employed in aircraft design and manufacturing also work as entry barriers. Second, the learning curve feature acts as an entry barrier because it implies that an entrant cannot make any profit until after a long period. As for exit, no evidence exists that the only two remaining firms, Airbus and Boeing, will exit the market, particularly considering their important political strategic status. Although entry and exit on the product level is not directly modeled as a firm choice variable, we allow their existence in a restricted format. A product can be switched between a potential entrant and an active good through quantity choices, as discussed above. Note that in the model, when quantity of a product is effectively zero, it has no effect on choices of other products or consumer surplus. 6 Hence, we feel comfortable in assuming away exit and entry and instead focusing on experience, which we believe to be much more important in the dynamics of the industry. 3.2 Transition of state variables We maintain the assumption that the three state variables experience E, unobserved characteristics ξ j and market size M evolve independently and thus discuss the transition process separately for them State transition of market size and unobserved characteristics As argued in Benkard (2004), the steady growth of market size M t is of second-order importance compared to business cycle fluctuations. So we first de-trend the market size variable M t, which is approximated by the total number of used and new wide-body aircraft, to reflect 1994 values, and then discretize it to three states. Such an approach results in a stationary state variable M t ; the finite values it takes greatly reduce the complexity of the problem. We further assume that the discretized M t follows a firstorder Markov process. The state variable ξ jt is the unobserved characteristics of product j at time t. Its fluctuation represents changes in consumers taste driven by exogenous shocks from various sources, e.g., major accidents or technological problems specific to a product or a firm; operating-cost-related macroeconomic shocks that lead to preference of twin-engine aircraft; or the temporary spur in international travel driven by the business cycle that makes relatively larger planes more attractive, etc. Ideally, the exogenous fluctuation ξ jt should be modeled as a state variable for each product. Unfortunately, adding one more state variable for each product leads to the well-known curse of dimensionality problem, and estimation would be intractable. Under the assumption that the unobserved factor 6 In the online appendix, we extend the dynamic model to allow for generation upgrade of aircraft. Then the model is consistent with the introduction of new aircraft as future generations of the current types. This is because products in the model are captured by characteristics, and their advances are captured by generation upgrade. So introduction of a new generation model replacing the old one is viewed as a quality improvement. 11

12 ξ jt is marginal in explaining variations of market share 7, we impose restrictions on joint transitions of all ξ jt and model its transition by two alternative approaches: (1) ξ jt is a constant that is subject to some mean-zero i.i.d. shocks ξ jt = ξ j + ξ jt where ξ j is mean of the series ξ jt with respect to t; (2) ξ jt = ξ j + ξ jt is maintained but we introduce a (discrete) preference rank state variable to approximate ξ jt. We present our main results based on the first approach, and the results using the second approach as a robustness check Experience To model the experience transition, we first assume that the experience is accumulated, i.e., E j,t+1 depends on the last period s experience E jt. Furthermore, we explicitly allow the spillover of experience across products and/or firms. The literature on the aircraft industry has documented such spillovers. For example, after the 1997 Boeing-McDonnell Douglas merger, the Joint Strike Fighter (JSF) program of Boeing has taken advantage of McDonnell s Phantom Work and experience with the AV-8B Harrier and F-18E/F programs (Aviation Week & Space Technology, February 16, 1998, page 71). The spillover from McDonnell Douglas to Boeing is realized through significant technology, manufacturing, systems and lesson learned experience from McDonnell s engineers that could be provided to JSF designers (Aviation Week & Space Technology, December 23, 1996, page 13). Moreover, technology diffusion also enables possible experience transfer across products of different firms, e.g., Neven et al. (1995) discuss the possibility of learning between Airbus and Boeing. The transfer of experience between different products (models or submodels), both within and across firms, implies that the experience E j,t+1 depends on the entire industry quantity vector Q t. By taking into account the accumulation and spillover of experience, the law of motion for experience can be expressed as Pr(E 1,t+1,, E J,t+1 E 1t,, E Jt, Q t ) = J Pr(E j,t+1 E jt, Q t ). j=1 Note that using E jt as a state variable in dynamic game helps us to avoid tracking the entire product history. To introduce a parametric form of the process Pr(E j,t+1 E jt, Q t ), we need to incorporate the following factors: (1) E j,t+1 is increasing in both E j,t and any q k,t, k = 1, 2,, J, that is, experience accumulates over time through both direct learning from production (q jt ) and spillover from production of other goods (q kt, k j); (2) producing a similar product to j contributes more to E j,t+1 ; and (3) experience may depreciate due to forgetting, which can be incorporated in the transition of E jt as E j,t+1 / E j,t < 1. We incorporate all the factors above and model the transition of experience as follow. 8 E j,t+1 = δe jt + J k=1 θ jkq kt, E j1 = 1, j, (5) 7 This assumption is tested empirically in Section The spillover effects measured by the parameters here are net effects in the sense that increases in quantities of other aircraft also may spur competition for experienced workers in the labor market. Thus, the parameters represent net effects of experience spillover and labor market competition. 12

13 where 1 if j = k, θ θ jk 1 if j and k are different submodels of the same model, θ 2 if j and k are different models of the same firm, if j are k models from different firms. θ 3 The parameters θ 1, θ 2 and θ 3 capture spillover across submodels (e.g., A and A ), across-product in the same firm (e.g., A330 and A340), and across-firm (e.g., MD-11 and B777), respectively. Spillover of experience does not rely on similarity of two submodels in their characteristics. We also consider a more general transition function where transition between two products depends on their distance in two dimensions of characteristics: number of seats and maximum ranges. The details are in the online appendix. The transition function of experience in (5) nests the commonly used specification in the literature E j,t+1 = δe jt + q jt as a special case where θ 1 = θ 2 = θ 3 = 0, i.e., there is no spillover of experience across production. In such a specification, learning is reflected by the positive correlation between E j,t+1 and q jt, and forgetting is captured by 0 < δ < 1. This accumulation function implies that experience accumulates as more aircraft are produced and also depreciates due to organizational forgetting. Our specification in (5) enables us to test the existence of spillover across production once we estimate θ 1, θ 2, and θ Estimation strategies We follow a multiple-step approach to estimate the parameters and solve the dynamic game introduced above. In the first step, we parametrize demand and cost functions of the dynamic oligopoly model, and estimate the corresponding parameters. Under the assumption that the structures generating the estimates are unchanged in the dynamic model, we take the estimates in the first step as primitives to solve the equilibrium of the dynamic game. Benkard (2004), among others, uses this approach to estimate the parameters and solve dynamic games. Such an approach is computationally convenient since the dynamic game only needs to be solved once, and there is no parameter searching in solving the dynamic game. One may take an alternative approach to estimate the parameters directly in the dynamic game. A commonly used method is to build a likelihood function or a set of moment conditions containing the parameters. Then one solves a constrained optimization problem with respect to the likelihood function or moment conditions by treating equilibrium conditions as constraints. When the optimization problem is solved, optimal parameter values are found together with the corresponding equilibrium of the dynamic game. This approach is employed for example in Goettler and Gordon (2011). Nevertheless, in the aircraft industry variables are serially correlated, thus the entire time series of a variable is just one observation of its evolution to construct the likelihood function or moment conditions. This limits both credibility and the number of moment conditions that can be constructed. Hence, this approach is not suitable for the aircraft industry, but it could 13

14 be more attractive for more localized industries containing many geographic markets. 4 Demand Estimation and Static Merger Analysis In this section, we estimate the demand system introduced in Section 3 and evaluate static efficiencies of the 1997 Boeing-McDonnell Douglas merger using the methodology pioneered in Nevo (2000). 4.1 The demand function We assume that consumers do not engage in intertemporal substitution and their choices of demand are solely based on current period product characteristics. The characteristics of product j at a given time period are classified into two categories: Z j = (X j, G j ) is observed, where G j is the relative generation of product j, 9 X j represents all other exogenous characteristics of product j, and ξ j captures characteristics unobserved to econometricians, such as product suitability. To model the demand function, we maintain that the product characteristics (X j, G j, ξ j ) and the market size M in a period are given, and the inverse demand function p j = p(q; X j, G j, ξ j, M) is single valued and taken as given by firms. We model yearly aircraft demand using a one-level Nested Logit discrete choice model, as in Benkard (2004), Hendel and Nevo (2006), and Gowrisankaran and Rysman (2012). Such an approach permits a closed-form solution of the inverse demand function, thus greatly reduces computational burdens compared to a random coefficient model when solving the dynamic game. Consumer a s utility function from aircraft j at time t is v ajt = δ jt + ω ajt, δ jt = ϕg jt + X jt β αp jt + ξ jt, ω ajt = ζ agt + (1 σ)ɛ ajt, (6) where g refers to group (new medium-sized aircraft or outside good). The error terms ζ agt and ɛ ajt are the random tastes at the group and submodel levels respectively, where ɛ ajt is i.i.d. extreme value and ζ agt has the (unique) distribution such that ζ agt + (1 σ)ɛ ajt is extreme value. The nesting parameter σ captures the consumers preference correlation across submodel. Since G jt is a plane s generation level measuring quality, the impact of future generations on demand is modeled as the product of differences in generations and ϕ, which captures gaps in quality between generations. X jt are observed characteristics including seats, maximum range, number of engines and fuel efficiency (the number of liters consumed per seat per 100 kilometers). p jt is the average price for aircraft j in year t and all prices are converted into 1994 U.S. dollars. ξ jt is the unobserved component affecting demand with its different values capturing variations in consumer preference over brand and plane characteristics. Note that although characteristics are fixed for an 9 The detailed discussions on the definition of G j are included in the online appendix. 14

15 aircraft, preference over brand or certain characteristics might change across time and affect consumers preference. We view each year as a market. Consumer a chooses product j {0, 1,, J} in market t if v ajt > v akt for all k j, k {0, 1,, J}, where 0 denotes the outside product with δ 0t being normalized to zero. Integrating over the probability of choosing product j for all consumers gives the well-known formula of the market share: ln(s jt /s 0t ) = ϕg jt + X jt β αp jt + σln(s j g,t ) + ξ jt, (7) where for a given market t, s jt, and s j g,t are respectively market shares of product j in the entire wide-body aircraft market and the new medium-sized wide-body aircraft market. s jt = q jt, s 0t = M t J j=1 q jt, s j g,t = M t M t q jt J k=1 q, kt where J is the number of products in the group of medium-sized aircraft. The demand model above rules out the possibility that submodels of the same model, e.g., A and A , are closer substitutes for consumers than submodels of different models, e.g., A and B To relax this restriction and allow greater substitution between submodels of the same model, we also estimate the demand function using a two-level nested logit model (e.g., see Verboven, 1996). The details are in the online appendix. 4.2 Description of Data As in Benkard (2004), we approximate the market size M t by the total number of used and new wide-body aircraft. This approximation is consistent with the assumption that all old and new aircraft are re-sold or rented out each year. 10 If a used aircraft did not change ownership in a year, it is viewed as bought by the firm that owned it. Therefore, market size or total transactions each year equal total number of used and new aircraft. We use demand data from 1991 to 2009 with a total of 12 aircraft submodels, leading to 115 submodel-year observations. The group g in our analysis includes all new medium-sized wide-body aircraft while the outside option contains small or large widebody aircraft, as well as all of the old wide-body aircraft in service. The price p jt and the market share s j g,t need to be instrumented in the demand estimation because they are likely to be correlated with unobserved aircraft characteristics ξ jt. We use hourly wage in manufacturing and its lagged terms, price of aluminum and its lagged terms, characteristics in subgroup and group, counts of products in subgroup and group as our instruments Trade and rent of used aircraft are very common, e.g., almost every MD-11 airliner has changed ownership or is owned by a leasing company. 11 If the instruments are weak, i.e., only weakly correlated with the endogenous variables, then the regression results will be biased. We run a first-stage regression of endogenous variables on the exogenous variables and instruments to test whether our instruments are weak. For both endogenous variables, the 15

16 The characteristics of other planes in group g are taken as instruments under the assumption that the observed characteristics are uncorrelated with the unobserved components. Manufacturing wage and aluminum price are cost shifters for price and are assumed to be orthogonal to ξ jt. All these instruments are widely used in the literature except for the number of other products within the same firm. Here we assume that the number of other products within the same firm is not correlated with unobserved characteristics of a product. It is correlated with the price of a product, because operating cost for an airline (consumer) is generally lower if its fleet consists of a set of planes from the same firm. Thus, a positive externality of products of a firm on other products in the same firm may be present. The correlation parameter σ is identified by covariation between the within-group market share of the plane s j g,t and its total market share s jt. We find from the data that s j g,t and s jt are strongly correlated with a correlation coefficient of Results Demand parameters We report the estimates of demand function in Table 3. The last column of the table presents the standard errors of the explanatory variables in the data, which allow us to compute the standardized coefficients by combining them with the estimates in the first column. The price coefficient is estimated to be 2.09, which is close to the estimate 2.4 for all wide-body aircraft in Benkard (2004). It indicates that price has a statistically significant negative effect on market share. Within-group utility correlation is estimated to be about 0.88, indicating that different aircraft in the medium-size group have a higher substitution effect than they do with others. The high correlation estimate is consistent with what we observe in the data: the market shares of individual medium-sized aircraft vary substantially over time, but the medium group as a percentage of all the wide-body aircraft has much less variation. This implies that the change of individual aircraft s market share is largely at the expense of other medium-sized aircraft, i.e., medium-sized aircraft are highly substitutable. A direct implication is that aircraft would have high price elasticities. Specifically, the estimated aggregate own-price elasticities of demand range from to 6.29 with a mean of The high own-price elasticities have been documented in the existing literature on aircraft market. For example, focusing on all the wide-body aircraft, the own-price elasticities are estimated to be 7.8 in Irwin and Pavcnik (2004) and 10 to 4 in Benkard (2004). Considering that the competition in the medium-sized wide-body market is more intense than in the wide-body market (Irwin and Pavcnik, 2004), our high elasticities are consistent with the existing results. Such results are potentially due to the large stock of used aircraft that is potentially on the market. F-statistics are larger than 10. According to Staiger and Stock (1997), our testing results imply that the maximum bias in IV estimators is less than 10%. We also check the variation in the price of aluminum, which is a crucial cost shifter. The data show that there are large movements in that price: for example, during the period 2002 to 2008 the price increased from $0.6/lb to $1.3/lb. The detailed price changes are readily available online, e.g., 16