The Evolution of Market Power in the U.S. Automobile Industry (1980-2018) PDF

THE EVOLUTION OF MARKET POWER IN THE U.S. AUTOMOBILE INDUSTRY* PAUL L. E. GRIECO CHARLES MURRY Downloaded from https://academic...

THE EVOLUTION OF MARKET POWER IN THE U.S. AUTOMOBILE INDUSTRY* PAUL L. E. GRIECO CHARLES MURRY Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 ALI YURUKOGLU We construct measures of industry performance and welfare in the U.S. auto- mobile market from 1980 to 2018. We estimate a demand model using product- level data on market shares, prices, and attributes, and consumer-level data on demographics, purchases, and stated second choices. We estimate marginal costs assuming Nash-Bertrand pricing. We relate trends in consumer welfare and markups to trends in market structure and the composition of products. Although real prices rose, we find that markups decreased substantially, and the fraction of total surplus accruing to consumers increased. Consumer welfare increased over time due to improved product quality and improved production technology. JEL codes: L11, L62, D43. I. INTRODUCTION From 1980 to 2018, the U.S. automobile industry experi- enced numerous technological and regulatory changes and its market structure changed dramatically. The goal of this article is to examine whether these changes led to discernible changes in industry performance. This work complements a recent aca- demic and policy literature analyzing long-term trends in mar- ket power and sales concentration from a macroeconomic per- spective (Autor et al. 2020; De Loecker, Eeckhout, and Unger 2020) with an industry-specific approach. Several papers and commentators point to a competition problem where price-cost margins and industry concentration have increased during this period (Economist 2016; Covarrubias, Gutiérrez, and Philippon 2020). Our estimates indicate a significant decline in markups over the past four decades, in contrast to estimates computed us- ing methods and data from the recent macroeconomics literature. * We thank Naibin Chen, Andrew Hanna, Davide Luparello, and Arnab Palit for excellent research assistance. We thank Aviv Nevo and Matthew Weinberg for comments. Portions of our analysis use data derived from a confidential, propri- etary syndicated product owned by MRI-Simmons. © The Author(s) 2023. Published by Oxford University Press on behalf of the Presi- dent and Fellows of Harvard College. All rights reserved. For Permissions, please email: [email protected] The Quarterly Journal of Economics (2024), 1201–1253. https://doi.org/10.1093/qje/qjad047. Advance Access publication on September 18, 2023. 1201 1202 THE QUARTERLY JOURNAL OF ECONOMICS Furthermore, our approach—also in contrast to the recent literature—admits a measure of consumer surplus over time. We find that consumer welfare in the U.S. automobile market has in- creased significantly over this period, primarily due to improve- Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 ments in product quality and production technology. To estimate trends in industry performance in the U.S. new car industry, we specify a heterogeneous agent demand system and assume Nash-Bertrand pricing by multiproduct automobile manufacturers. The key inputs into the demand estimates are ag- gregate data on prices, market shares, and vehicle characteristics over time, microdata on the relationship between demographics and car characteristics over time, microdata on consumers’ stated second choices, and the use of the real exchange rate between the United States and product origin countries as an instrumental variable for endogenous prices. With the demand system in hand, we infer product-level markups from the first-order condition of each firm’s profit maximization problem. We find that median markups as defined by the Lerner index (L = p−mc p ) fell from 0.325 in 1980 to 0.185 by 2018 (Figure VI, Panel A). However, as we detail below, although markups are a useful proxy for market efficiency when products are fixed over time, they are a conceptually unattractive measure over long pe- riods of time when products change. We use our model to consider trends in consumer and producer surplus directly. To quantify changes in welfare over time, we use a decomposition from Pakes, Berry, and Levinsohn (1993) to develop a measure of consumer surplus that is robust to changes in the attractiveness of the out- side good. This approach leverages continuing products to capture changes in unobserved automobile quality over time. However, it is not influenced by aggregate fluctuations in demand for au- tomobiles, for example, business cycle effects such as monetary policy or changes in alternative transportation options. We find that the fraction of efficient surplus (the sum of producer surplus, consumer surplus, and deadweight loss) going to consumers went from 0.62 in 1980 to 0.82 by 2018 and that average consumer sur- plus per household increased by roughly $8,000 over our sample period. The increase in consumer surplus is predominantly due to the increasing quality of cars and improved production technol- ogy. We confirm the patterns in Knittel (2011) that horsepower, size, and fuel efficiency have improved significantly over this time  MARKET POWER IN U.S. AUTOMOBILES 1203 period. We use the estimated valuations of these car attributes to put a dollar amount on this improvement. Furthermore, we use market shares of continuing products to estimate the com- bined valuation of improvements in other characteristics such as Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 electronics, safety, or comfort features that are not readily avail- able in common data sets (e.g., audio and entertainment systems, antilock breaks, rearview cameras, driver assistance systems). Improvements on these dimensions are quantitatively large. In addition, we estimate improved production technology from vari- ation in marginal cost over time controlling for product attributes. Counterfactuals that eliminate the observed increase in import competition or the increase in the number of vehicle models have small to moderate effects on consumer surplus. Counterfactuals that eliminate the increase in automobile quality and the tech- nological improvements in production have the greatest effect on consumer surplus. A number of caveats are warranted for this analysis. First, our main results assume static Nash-Bertrand pricing each year and rule out changes in conduct, for example, via the ability to tacitly collude. However, for robustness, we present a number of alternative assumptions on conduct, all of which indicate de- clining markups when conduct is fixed over time. Second, we do not model the complementary dealer, parts, or financing markets where the behavior of margins or product market efficiency over time may be different than for the automobile manufacturers. By studying long-run trends in market power and market efficiency using the workhorse toolbox of supply and demand es- timation, we provide an alternative perspective on the analysis of the recent literature on the rise in aggregate markups based on production-side modeling and accounting data on revenues and costs, for example De Loecker, Eeckhout, and Unger (2020) and various subsequent studies. This approach infers markups at the firm level under the assumption that firms optimally choose the quantity of variable inputs in production to minimize costs. The assumptions of these two approaches are nonnested; we pro- vide a comparison of our markup estimates in the U.S. automo- bile industry with those constructed by De Loecker, Eeckhout, and Unger (2020) in Section V.F. Our perspective is rooted in the methods developed in industrial organization that grew out of the critique of the structure-conduct-performance literature, for ex- ample, Demsetz (1973), and for a historical perspective see Berry, Gaynor, and Scott Morton (2019). Our approach also allows for 1204 THE QUARTERLY JOURNAL OF ECONOMICS an understanding of the mechanisms that contribute to trends in market power and consumer surplus. In particular, we highlight the importance of characterizing consumer welfare, which is only possible by estimating demand curves. Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 This work thus complements research that raises mea- surement issues and proposes alternatives in the production paradigm, such as Traina (2018), Raval (2023), Demirer (2022), Bond et al. (2021), Doraszelski and Jaumandreu (2021), and Foster, Haltiwanger, and Tuttle (2022). Although we focus on a single industry, our results suggest that more work should be done to carefully measure market power and welfare in important industries to provide an alternative measurement from the pro- duction approach and identify the mechanisms that drive trends in market power and efficiency. There are now other recent examples of researchers using demand and supply to characterize trends in markups in specific industries.1 Brand (2021) and Döpper et al. (2023) analyze mul- tiple grocery categories for a selection of retail outlets over the period 2006 to 2017 and 2019, respectively. Miller et al. (2022) analyze the cement industry over 1976 to 2016. Ganapati (2021) studies the wholesaling sector over 1997 to 2007. Across a variety of industries, these papers point out that technological changes over decades affecting product qualities and costs are large and important to control for when inferring market power. This arti- cle corroborates this finding in the auto industry by documenting the large changes in product quality over time as well as signifi- cant cost-reducing technological improvements. Relative to these papers, this article uses household-level data on purchases, de- mographics, and second choices to estimate a demand specifica- tion with rich heterogeneity and employs standard instrumen- tal variable identification strategies. This study also compares its markup estimates with production function–based estimates as reported in De Loecker, Eeckhout, and Unger (2020) and analyzes the determinants of the change in consumer surplus over time. Bet (2021) compares markup estimates from a demand approach with those from a production approach for domestic airlines and finds that under Nash-Bertrand pricing, markups from the de- mand approach are flat for large carriers, while under the produc- tion approach, markups for large carriers are increasing over the 1. In an earlier contribution, Berry and Jia (2010) analyzed changes in de- mand and market power in the U.S. airline industry between 1999 and 2006.  MARKET POWER IN U.S. AUTOMOBILES 1205 period 2013 to 2019. Relative to these other papers, our work es- timates the role of technological progress in improving consumer surplus by decomposing over time changes in demand shocks into improvements in unobservable quality and changes in the value Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 of the outside option. This decomposition is important for inter- preting the economics of how changing prices and markups trans- late into consumer welfare when products and technology change over time. Our research is also closely related to Hashmi and Biese- broeck (2016), who model dynamic competition and innovation in the world automobile market using a logit model over the pe- riod 1982 to 2006. Relative to their work, this article focuses on analyzing the evolution of consumer surplus and markups rather than modeling dynamic competition in quality. Furthermore, in addition to analyzing a longer time period, this article uses mi- crodata and second-choice data to estimate demand following Bordley (1993) and Berry, Levinsohn, and Pakes (2004), uses a different instrumental variable to account for price endogeneity, and decomposes time effects in demand separately into changes in unobservable quality and changes in the value of the outside option. II. DATA We compiled a data set covering 1980 through 2018 consist- ing of automobile characteristics and market shares, individual consumer choices and demographic information, and consumer survey responses regarding alternate “second choice” products. This section describes the data sources and presents basic de- scriptive information. II.A. Automobile Market Data Our primary source of data is information on sales, manu- facturer suggested retail prices (MSRP), and characteristics of new cars and light trucks sold in the United States over 1980– 2018 that we obtain from Ward’s Automotive. Ward’s keeps dig- ital records of this information from 1988 through the present. To get information from before 1988, we hand collected data from Ward’s Automotive Yearbooks. The information in the yearbooks is nonstandard across years and required multiple layers of dig- itization and rechecking. We supplemented the Ward’s data with 1206 THE QUARTERLY JOURNAL OF ECONOMICS additional information, including vehicle country of production, company ownership information, missing and nonstandard prod- uct characteristics (e.g., electric vehicle eMPG and driving range, missing MPG, and missing prices), brand country affiliation (e.g., Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 Volkswagen from Germany, Chrysler from the United States), and model redesign years. Prices in all years are deflated to 2015 US$ using the core consumer price index. To construct market shares, we define the market size as the number of households in the United States divided by 2.5, which reflects the fact that the av- erage household owns nearly two cars and the average tenure of car ownership during this time period is roughly five years. 1. Product Aggregation. Vehicles sold in the United States are highly differentiated products. Each brand (or “make”) pro- duces many models and each model can have multiple variants (more commonly called “trims”). Although we have specifications and pricing of individual trims, our sales data is at the make- model level. Similar to other studies of this market, we make use of the sales data by aggregating the trim information to the make- model level; see Berry, Levinsohn, and Pakes (1995 (BLP), 2004), Goldberg (1995), and Petrin (2002). We aggregate price and prod- uct characteristics by taking the median across trims. Table I displays summary statistics for our sample of vehicles at the make-model-year level. An example of an observation is a 1987 Honda Accord. There are 6,130 cars, 2,243 SUVs, 680 trucks, and 641 vans in our sample.2 The average car has 52,089 sales in a year, and the average truck has 140,207 sales. Trucks and vans are more likely to be from U.S. brands and less likely to be assembled outside of the United States than cars and SUVs. Two percent of our sample has an electric motor (including hybrid gas- powered and electric only). We present a description of trends in vehicle characteristics in Section III. 2. We use Ward’s vehicle style designations to create our own vehicle designa- tions. We aggregate CUV (crossover utility vehicles) and SUV to our SUV designa- tion. Truck and van are native Ward’s designations. We designate all other styles (sedan, coupé, wagon, hatchback, convertible) as car. Some models are produced in multiple variants. For example the Chrysler LeBaron has been available as a sedan, coupé, and station wagon in various years. However, no model is produced as both a car and an SUV, or any other combination of our designations, in our sample.  TABLE I SUMMARY STATISTICS Mean Std. dev. Min Max Mean Std. dev. Min Max Cars, N = 6,130 SUVs, N = 2,243 Sales 52,088.60 72,750.83 10.00 473,108.00 Sales 51,629.61 66,932.79 10.00 753,064.00 Price 35.85 18.76 11.14 99.99 Price 40.41 14.94 12.75 96.94 MPG 22.67 6.82 10.00 50.00 MPG 18.01 4.98 10.00 50.00 Horsepower 178.21 83.41 48.00 645.00 Horsepower 232.33 74.92 63.00 510.00 Height 55.76 4.21 43.50 107.50 Height 69.01 4.38 53.00 90.00 Footprint 12,870.08 1,710.41 6,514.54 21,821.86 Width 13,790.90 1,785.69 8,127.00 18,136.00 Curb weight 3,181.94 639.51 1,488.00 6,765.00 Curb weight 4,246.05 854.30 2,028.00 7,230.00 U.S. brand 0.40 0.49 0.00 1.00 U.S. brand 0.40 0.49 0.00 1.00 Import 0.59 0.49 0.00 1.00 Import 0.59 0.49 0.00 1.00 Electric 0.02 0.14 0.00 1.00 Electric 0.01 0.12 0.00 1.00 Trucks, N = 680 Vans, N = 641 Sales 140,207.22 184,123.33 12.00 891,482.00 Sales 59,103.39 86,940.25 10.00 891,482.00 Price 27.81 9.82 12.02 69.43 Price 36.05 17.13 11.14 99.99 MPG 17.83 4.36 10.00 50.00 MPG 20.94 6.58 10.00 50.00 Horsepower 189.17 90.31 44.00 403.00 Horsepower 192.18 83.88 44.00 645.00 Height 68.39 6.33 51.80 81.00 Height 60.95 8.41 43.50 107.50 Footprint 15,086.14 2,478.91 8,437.30 20,000.00 Footprint 13,392.63 1,968.92 6,514.54 21,821.86 MARKET POWER IN U.S. AUTOMOBILES Curb weight 4,043.42 1,114.94 1,113.00 7,178.00 Curb weight 3,561.21 897.77 1,113.00 8,550.00 U.S. brand 0.65 0.48 0.00 1.00 U.S. brand 0.44 0.50 0.00 1.00 Import 0.35 0.48 0.00 1.00 Import 0.55 0.50 0.00 1.00 Electric 0.00 0.00 0.00 0.00 Electric 0.02 0.13 0.00 1.00 Notes. An observation is a make-model-year, aggregated by taking the median across trims in a given year. Statistics are not sales weighted. Prices are in 2015 000’s US$. Physical dimensions are in inches, and curb weight is in pounds. 1207 Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 1208 THE QUARTERLY JOURNAL OF ECONOMICS II.B. Price Instrument To identify the price sensitivity of consumers, we rely on an instrumental variable that shifts price while being plausi- bly uncorrelated with unobserved demand shocks. We employ a Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 cost shifter related to local production costs where a vehicle is produced. For each automobile in each year, we use the price level of expenditure in the country where the car was manufac- tured, obtained from the Penn World Tables, version 9.1, variable pl_con, lagged by one year to reflect planning horizons. Follow- ing Feenstra, Inklaar, and Timmer (2015), we refer to this as the real exchange rate (RXR). RXR is equal to the purchasing power parity (PPP) exchange rate relative to the United States divided by the nominal exchange rate relative to the United States. RXR varies with two sources that are useful for identifying price sen- sitivities. First, if wages in the country of manufacture rise, the cost of making the car will rise, which will raise the RXR via the PPP rising. Therefore, the RXR captures one source of input cost variation through local labor costs. Another source of variation is through the nominal exchange rate. If the nominal exchange rate rises, so that the local currency depreciates relative to the dollar, a firm with market power will have an incentive to lower retail prices in the United States, thereby providing another av- enue of positive covariation between the RXR and retail prices in the United States. Exchange rates were used as instrumental variables for car prices in Goldberg and Verboven (2001), which is focused on the European car market, and in Berry, Levinsohn, and Pakes (1999), along with wages. In Figure I, we display the lagged RXR for the most popular production countries, where the size of the plot marker is proportional to the number of prod- ucts sold from each country and the black dashed line represents the U.S. price level. Although our measure of RXR is relative to the United States, U.S. RXR is also changing over time due to inflation. We demonstrate the behavior of the RXR instrumental vari- able in a simple setup in Table II. We estimate a logit model of demand, as in Berry (1994), first via OLS and then using 2SLS with RXR as an instrumental variable for price. We include make fixed effects and year fixed effects. Within make there is variation in real exchange rates within and across time. Within time varia- tion is due to the fact that different models of the same make are assembled in different countries. For example, BMW assembles  MARKET POWER IN U.S. AUTOMOBILES 1209 Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 FIGURE I Real Exchange Rates Lagged real exchange rates are from Penn World Tables 9.1. The size of dots corresponds to the relative number of sales by production country, except for the United States. vehicles for the U.S. market in Germany and the United States, General Motors has produced U.S.-sold vehicles in Canada, Mex- ico, and South Korea (among other countries), and many of the Japanese and South Korean brands produce some of their mod- els in the United States, Canada, and Mexico. Lacetera and Syd- nor (2015) provide evidence that vehicle manufacturers maintain quality standards when producing vehicles in different countries. The first column in Table II shows the first-stage relevance of the instrumental variable. The sign is positive as predicted by the theory with a first-stage F-statistic of 14.09. We cluster the standard errors at the make level. The first stage implies a pass- through of RXR to prices of 0.117, which is consistent with esti- mates in the literature (Goldberg and Campa 2010; Burstein and Gopinath 2014). The difference in the price coefficient in the last two columns demonstrates that using the IV moves the coefficient estimate on price in the negative direction, which is expected be- cause the OLS coefficient should be biased in the positive direc- tion if prices positively correlate with unobserved demand shocks conditional on observable characteristics. Comparing the mean own-price elasticities between the OLS and IV estimates confirms the importance of controlling for price endogeneity.  1210 TABLE II LOGIT DEMAND First stage Reduced form OLS IV Price −0.334 (0.042) −1.696 (0.598) RXR 0.411 (0.110) −0.697 (0.232) Height −0.199 (0.048) −0.064 (0.066) −0.120 (0.069) −0.401 (0.161) Footprint −0.117 (0.066) 0.348 (0.081) 0.318 (0.082) 0.149 (0.149) Horsepower 0.768 (0.116) −0.097 (0.070) 0.149 (0.067) 1.206 (0.472) MPG 0.113 (0.036) −0.062 (0.057) −0.018 (0.062) 0.130 (0.116) Curb weight 0.803 (0.111) −0.493 (0.142) −0.233 (0.140) 0.868 (0.541) Num. of trims −0.115 (0.020) 1.097 (0.045) 1.060 (0.044) 0.902 (0.091) Release year −0.081 (0.040) −0.173 (0.054) −0.195 (0.056) −0.311 (0.091) Yrs. since design 0.000 (0.012) −0.145 (0.017) −0.145 (0.017) −0.144 (0.024) Sport 0.480 (0.090) −0.679 (0.105) −0.523 (0.102) 0.134 (0.323) Electric 0.765 (0.176) −1.031 (0.255) −0.791 (0.245) 0.267 (0.560) Truck −0.416 (0.154) −0.485 (0.099) −0.631 (0.107) −1.190 (0.359) SUV −0.111 (0.117) 0.561 (0.100) 0.515 (0.105) 0.372 (0.214) Van −0.268 (0.161) 0.037 (0.126) −0.060 (0.143) −0.417 (0.330) Mean own price elas. — — −1.204 −6.107 Implied pass-through 0.117 (0.032) THE QUARTERLY JOURNAL OF ECONOMICS First stage F-statistic 14.086 Notes. Unit of observations: year-make-model, from 1980 to 2018. Number of observations: 9,694. All specifications include year and make fixed effects. Standard errors clustered by make are in parentheses. All continuous car characteristics are in logs and price is in 2015 $10,000. Variables are logged and standardized. Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025  MARKET POWER IN U.S. AUTOMOBILES 1211 II.C. Consumer Choices and Demographics We collect individual-level data on car purchases and demo- graphics from two data sources: the Consumer Expenditure Sur- vey (CEX) and MRI’s Survey of the American Consumer (MRI). Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 These data sets provide observations on a sample of new car purchasers for each year, including the demographics of the pur- chaser and the car model purchased. CEX covers 1980–2005 with an average of 1,014 observations a year. MRI covers 1992–2018 with an average of 2,005 observations a year. We construct micro- moments from these data to use as targets for the heterogeneous agent demand model, following Goldberg (1995), Petrin (2002), and Berry, Levinsohn, and Pakes (2004). There are some general patterns from these data that motivate specification choices for the demand model. For example, that the average purchaser of a van has a larger family size suggests families value size more than nonfamilies. That the average price of a car purchased by a high-income versus low-income buyer suggests higher-income buyers are either less sensitive to price or place higher value on characteristics that come in higher-priced cars. That rural house- holds are more likely to purchase a truck suggests stronger pref- erence for features of trucks by rural households. To approximate the distribution of household demographics, we sample from the CPS, which contains the demographic infor- mation from 1980–2018 that we use from the CEX and MRI sam- ples. Average household income (in 2015 dollars) increases from $55,382 to $81,375 from 1980 to 2018. Average household age in- creases from 46 to 51; average household size falls from 1.60 to 1.25; the percent of rural households decreases from 27.9 to 13.4. We account for these trends by explicitly including evolving con- sumer heterogeneity in income, family size, and rural status as part of our model. II.D. Second Choices We obtain data on consumers’ reported second choices from MaritzCX, an automobile industry research and marketing firm. MaritzCX surveys recent car purchasers based on new vehi- cle registrations. The survey includes a question about cars that the respondents considered but did not purchase. We use the first listed car as the purchaser’s second choice. These data have previously been used, such as in Leard et al. (2023) and Leard (2022), and are similar to the survey data used in 1212 THE QUARTERLY JOURNAL OF ECONOMICS TABLE III SECOND CHOICES, SELECTED EXAMPLES Model and year Modal second choice Next second choice (Modal + next)/n 1991 (N = 29,436) Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 Ford F Series Dodge Ram Pickup Chevrolet C/K Pickup 0.35 Honda Accord Toyota Camry Nissan Maxima 0.19 Dodge Caravan Ford Aerostar Plymouth Voyager 0.15 Mercedes-Benz BMW 5 Series Lexus LS 0.17 E Class Toyota 4Runner Ford Explorer Nissan Pathfinder 0.34 Nissan 300ZX Alfa Romeo 164 Chevrolet Corvette 0.20 1999 (N = 20,413) Chevrolet Silverado Ford F Series Dodge Ram Pickup 0.76 Toyota Camry Honda Accord Nissan Maxima 0.38 Plymouth Voyager Ford Windstar Dodge Caravan 0.42 Audi A6 BMW 5 Series Volvo 80 0.28 Chevrolet Tahoe Ford Expedition Dodge Durango 0.36 BMW Z3 Porsche Boxster Mazda MX-5 Miata 0.42 2005 (N = 42,977) Toyota Tacoma Nissan Frontier Ford F Series 0.35 Ford Focus Toyota Corolla Honda Civic 0.22 Honda Odyssey Toyota Sienna Chrysler Town & Country 0.71 Lincoln Town Car Cadillac DeVille Chrysler 300 Series 0.44 Honda CR-V Toyota RAV4 Ford Escape 0.38 Porsche Cayenne BMW X5 Land Rover Range Rover 0.43 2015 (N = 53,391) Ford F Series Chevrolet Silverado Dodge Ram Pickup 0.64 Toyota Prius Honda Accord Hybrid Honda CR-V 0.11 Toyota Sienna Honda Odyssey Chrysler Town & Country 0.64 Volvo 60 BMW 3 Series Audi A4 0.16 Nissan Frontier Toyota Tacoma Chevrolet Colorado 0.69 Chevrolet Camaro Ford Mustang Dodge Challenger 0.46 Toyota Prius PHV Chevrolet Volt Nissan Leaf 0.32 Notes. The data are from MaritzCX surveys in 1991, 1999, 2005, and 2015. Vehicles selected are high- selling vehicles that represent a range of styles and attributes. The last column displays diversion to the two most popular second choices, conditional on diversion to any vehicle. Berry, Levinsohn, and Pakes (2004).3 After we merge with our sales data, we use second-choice data from 1991, 1999, 2005, and 2015, representing 29,396, 20,413, 42,533, and 53,328 purchases, respectively. In Table III we display information about second choices for many popular cars of different styles and features to give a sense 3. The MaritzCX survey asks respondents about vehicles that the respondents considered but did not purchase. One of the questions is whether the respondent considered any other cars or trucks when shopping for their vehicle. Respondents answer this question either yes or no. For those who answer yes, the survey asks respondents to provide vehicle make-model and characteristics for the model most seriously considered.  MARKET POWER IN U.S. AUTOMOBILES 1213 for how strong substitution in vehicle style appears in the data. For each year, we display the modal second choice, the next most common second choice, and the share who report these two cars as second choices over the total responses for that car. For example, Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 in 1991, the Dodge Ram Pickup is the modal second choice among the respondents who purchased a Ford F Series. The Chevrolet CK Pickup is the second most popular second choice, and together, these two second choices make up 69% of reported second choices for the Ford F Series. From this sample of vehicles, second choices tend to be similar types of vehicles (i.e., trucks, cars, SUVs, vans). Also, there is substantial variation in the share that the two most frequent choices represent. For example, in 1991, the F Series and Dodge Ram represent 76% of reported second choices for the Chevrolet Silverado in 1999, but the Civic and Corolla only rep- resent 22% of second choices for the Ford Focus in 2005. The gen- erally strong substitution toward similar vehicles is crucial for identifying unobserved heterogeneity in the demand model we present in Section IV. III. EMPIRICAL DESCRIPTION OF THE NEW CAR INDUSTRY, 1980–2018 This section describes trends in the U.S. automobile industry from 1980 to 2018 related to market power and market efficiency. We first discuss changes in prices and market structure. Then we discuss trends in product characteristics. III.A. Prices and Market Structure Inflation-adjusted average prices in the automobile indus- try rose from 1980 to 2018. At the same time, concentration decreased. Figure II displays these patterns. In Panel A, we docu- ment that the average MSRP rose from around $17,000 in 1980 to around $34,000 in 2018 (in 2015 US$, deflated by the core CPI). The bulk of the change in average price occurred before 2000, although the upper 25% of prices continued to rise after 2000. At the same time, the Herfindahl-Hirschman index (HHI) mea- sured at the parent company level fell from over 2,500 to around 1,200, see Panel B. The C4 index saw a similar decrease over the same time period, from around 0.80 to 0.58. In Panel C, we docu- ment the main source of decreasing concentration. While the total number of firms in this industry fell slightly from 1980 to 2018, 1214 THE QUARTERLY JOURNAL OF ECONOMICS Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 (A) (B) (C) (D) FIGURE II Prices and Market Structure, 1980–2018 Panel A displays share-weighted average price along with the interquartile range. Panel B: HHI (bold line and left scale) and C4 (dashed line and right scale) are defined at the parent company level, for example, Honda is the parent com- pany of the Honda and Acura brands. In Panel C, the number of products corre- sponds to a model available in a given year in our sample. The style definitions referred to in Panel D are described in the text. Data are from Ward’s Automotive Yearbooks, and the sample selection is described in the text. there were about twice as many products in 2018 as there were in 1980. In 1980, the “Big Three” U.S. manufacturers accounted for a large portion of sales, whereas by 2018, sales were more evenly dispersed among domestic and international firms, consis- tent with patterns in other manufacturing industries (Amiti and Heise 2021). III.B. Physical Characteristics of Vehicles That prices rose while concentration fell might seem counter- intuitive at first pass; however, prices are also a function of phys- ical characteristics, quality, and production technology. There are two main trends regarding the physical characteristics of cars. The first is the rise of the SUV, which was a nearly nonexis- tent vehicle class in 1980 and by the end of our sample repre- sented roughly half of all sales. Second, cars and trucks have be- come larger and more powerful without sacrificing fuel efficiency (Knittel 2011).  MARKET POWER IN U.S. AUTOMOBILES 1215 Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 (A) (B) (C) (D) FIGURE III Physical Vehicle Characteristics, 1980–2018 Panels A–C display average characteristics for available models in our sample. Panel D is the percent of each feature installed on total “cars” sold (i.e., not trucks, SUVs, or vans). Factory-installed features were compiled from Ward’s Automotive Yearbooks from various years. For example, in 1980 61% of “cars” sold had air conditioning. The number of products available to consumers increased from 1980 to 2018. A major contribution to this change is the rise of SUV production, particularly smaller SUVs that are designed to compete with sedans. Our SUV category aggregates SUVs (typ- ically larger vehicles built on pickup truck frames, like the Toy- ota 4Runner) together with CUVs (smaller than SUVs and built on sedan frames, like the Honda CRV). In Figure II, Panel D, we display the number of products by vehicle style over time. In the early 1980s fewer than 25 SUVs were available to con- sumers (typically large truck-like vehicles) and after 2000 there were nearly 100 SUVs available in the market. Figure III displays selected product attributes over time. Av- erage horsepower and footprint (length times width) increased substantially from 1980 to 2018. Average horsepower more than doubled for cars and roughly tripled for trucks from 1980 to 2018; see Figure III, Panel A. Cars became larger, SUVs and vans be- came smaller during the 1980s and then grew, and the average truck size grew substantially from 1980 to 2018. At the same 1216 THE QUARTERLY JOURNAL OF ECONOMICS time that horsepower and size increased, average fuel economy remained roughly constant, which largely reflects federal regula- tory standards for fleet fuel economy, first enacted in the Energy Policy and Conservation Act of 1975. Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 In addition, attributes not related to size and power changed substantially from 1980 to 2018. In Figure III, Panel D, we show the percent of cars (i.e., not trucks, SUVs, or vans) sold with the following features, for 1980, 1990, 2000, 2010, and 2014: air con- ditioning, power windows, antilock brakes, cassette player stereo system, side airbags, memory seats, and rear camera.4 The per- centage of cars with many of these features increased from 1980 to 2018, however, both technology and trends in preferences af- fected the rate of adoption differently for different features. For example, air conditioning reached near universal adoption by 2000, but rear cameras are a recent addition. Safety features, like side airbags, were quickly adopted through the 1990s as federal safety regulations tightened. The cassette player, once a luxury feature, faded from cars as CDs became popular, disappearing by 2010. In our demand model, many of these features will be sub- sumed into a quality residual that summarizes all characteristics not captured by readily available data like horsepower and vehi- cle size. IV. MODEL Our framework is a differentiated product demand and oligopoly pricing model following Berry, Levinsohn, and Pakes (1995), which is standard in the industrial organization litera- ture. IV.A. Consumers Consumer i makes a discrete choice among the Jt options in the set Jt of car models available in year t and an outside “no- purchase” option (indexed 0), choosing the option that delivers the maximum conditional indirect utility.5 Utility is a consumer-specific linear index of a vector of vehi- cle attributes (xjt ), price (pjt ), an unobserved vehicle-specific term 4. These data were collected from Ward’s Automotive Yearbooks of the corre- sponding years. 5. Our model focuses on consumers’ selection of a manufacturer’s product. In particular, we abstract away from financing, leasing, and dealership choice.  MARKET POWER IN U.S. AUTOMOBILES 1217 (ξ jt ), and an idiosyncratic consumer-vehicle-specific term ( ijt ). (1) ui jt = βit x jt + αit p jt + ξ jt + i jt. The index i denotes an individual in a given year. We specify Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 and estimate parametric distributions of taste parameters β i and α i across individuals that depend on time-varying demographics and allow for unobservable heterogeneity. In our preferred spec- ification, the parameters governing these distributions are fixed over time, but we also report estimates including time-varying components to parameters of the distribution of α i and β i. We as- sume that  ijt are independent draws from the standard Gumbel distribution. Utility of the no-purchase option is ui0t = γ t +  i0t , where γ t reflects factors that change the utility of the no-purchase option from year to year, including business cycle fluctuations, urbaniza- tion, and durability of used automobiles. The average unobserved quality of new automobiles is also changing over time. We denote the mean utility of the choice set in year t relative to the base year as τ t so that ξ jt = τt + ξ˜jt and assume that E[ξ˜jt |z jt ] = 0, where zjt is a vector of instruments including xjt , year dummies, and an instrument for price (i.e., RXR). It is well known that discrete-choice models only identify util- ity relative to the outside good. Without further restrictions, we would be unable to separately identify yearly average unobserved quality, τ t , and the value of the outside option, γ t. To address this issue, we follow Pakes, Berry, and Levinsohn (1993) and add the restriction that (2) ∀ j ∈ Ct : E[ξ jt − ξ jt−1 ] = E[(τt − τt−1 ) + (ξ˜jt − ξ˜jt−1 )] = 0, where Ct is the set of continuing vehicles offered in both year t and t − 1 that have not been redesigned by the manufacturer. Consider a model j ∈ Ct as a product nameplate and design gen- eration appearing both in t − 1 and t.6 This restriction captures the fact that models in a model generation have substantively the same design from year to year, although it allows for idiosyn- cratic changes in features, marketing, or consumer taste. That is, while ξ jt can change from year to year, innovations in ξ jt are mean zero across years in a model generation. This restriction 6. Vehicle models are periodically redesigned. Within a design generation and across years, models share the same styling and the same (or very similar) at- tributes. A typical design generation is between five and seven years. 1218 THE QUARTERLY JOURNAL OF ECONOMICS separately identifies average quality of the choice set, τ t , from the average consumer valuation of the outside good, γ t. Identification follows from a two-step argument. First, following the usual logic of discrete-choice models, τ t − γ t is identified. Second, given that Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 ξ˜jt can be constructed from identified objects, the moment condi- tion over continuing products equation (2) identifies τ t (subject to the normalization that τ 0 = 0). As this argument for identifica- tion is constructive, we will follow it closely when estimating the model. Separating average unobserved quality and the value of the outside option is important because we expect that unobserved product attributes change over time, as in Figure III, Panel D. It is important for us to incorporate this concept into consumer welfare. Second, the time effects capture aggregate economic con- ditions that influence the total sales of vehicles, but that are ar- guably not relevant for assessing the functioning of competition in the industry. We model consumer heterogeneity by interacting household demographics and unobserved preferences with car attributes. Our baseline specification is:  (3) αit = ᾱ + αh Dhit h  (4) βik = β̄k + βkh Dhit + σk νik , h where subscript k denotes the kth car characteristic (including a constant) and h indexes dimensions of consumer demograph- ics (e.g., income). Allowing for observed heterogeneity allows sub- stitution patterns to differ by demographics. The distribution of Dit is taken from the CPS. In practice, we do not interact ev- ery demographic with every car characteristic. See Table IV for a complete listing of demographic-characteristic interactions and unobserved heterogeneity that we include in the model. Allowing for unobserved heterogeneity allows for more flexible substitution patterns. Unobserved taste for automobile characteristics, ν ik , are assumed to be independent draws from the standard normal distribution. Our baseline specification holds the parameters underlying the distributions of β i and α i fixed over time. That said, the dis- tributions themselves can change over time because of changing demographics. For example, increasing income inequality will  TABLE IV COEFFICIENT ESTIMATES Demographic interactions β σ Income Inc. sq. Age Rural Fam. size 2 FS 3–4 FS 5+ Price −3.112 — 0.094 −0.462 2.065 — — — — (1.124) (0.010) (0.133) (0.122) Van −7.614 5.538 — — — — 1.737 3.681 5.840 (0.598) (0.133) (0.165) (0.176) (0.176) SUV −0.079 3.617 — — — — — — — (0.339) (0.087) Truck −7.463 6.309 — — — 3.007 — — — (0.898) (0.310) (0.340) Footprint 0.534 1.873 — — — — 0.481 0.459 0.636 (0.261) (0.118) (0.053) (0.054) (0.054) Horsepower 1.018 1.246 — — — — — — — (0.954) (0.361) Miles/Gal. −0.965 1.645 — — — — — — — (0.211) (0.151) Luxury — 2.624 — — — — — — — (0.047) Sport −3.046 2.617 — — — — — — — (0.549) (0.075) EV −5.549 3.798 — — — — — — — (1.406) (0.511) MARKET POWER IN U.S. AUTOMOBILES Euro. brand — 1.921 — — — — — — — (0.054) U.S. brand — 2.141 — — — — — — — (0.048) Constant — — 0.362 — — — — — — (0.034) Notes. Brand and year dummies included. Standard errors are constructed by bootstrapping the microdata and are clustered at the brand level. All continuous car characteristics are in logs and standardized, and price is in 2015 $10,000. Footprint is vehicle length times height in square inches. Income is normalized to have zero mean and unit variance. 1219 Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 1220 THE QUARTERLY JOURNAL OF ECONOMICS lead to increasing dispersion in the α it distribution over time. We estimate additional specifications where we allow the parameters to vary over time and for price to enter indirect utility in logs rather than levels. Allowing preferences to vary over time pro- Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 vides greater flexibility in the estimation of markups, since firms will react to these changes when setting price. However, it will also imply changes in surplus due only to changes in the param- eters of the utility function. We discuss the details of these al- ternative specifications and report results in Online Appendix B. Our estimates of markups are similar across specifications, so we perform the bulk of our analysis using the baseline specifi- cation that maintains stable-over-time parameters with clearer consumer welfare implications. For a given year, market shares in the model are given by integrating over the distribution of consumers who vary in their demographics, unobserved tastes for characteristics, and idiosyn- cratic error terms,  exp(βit x jt + αit p jt + ξ jt ) (5) s jt =  dF (i ). i exp (γt ) + l∈Jt exp (βit xlt + αit plt + ξlt ) Shares conditional on consumer demographics can be com- puted by replacing the population distribution with the appropri- ate conditional distribution F(i|Dit ∈ ·). Moreover, second-choice shares conditional on a given first-choice vehicle can be computed similarly by integrating consumers’ choice probabilities, when the first-choice vehicle is removed, over the distribution of consumers, weighted by their probability of making that first choice. IV.B. Firms On the supply side, we assume automobile manufacturers, indexed by m, play a static, full information, simultaneous-move pricing game each year. Manufacturers choose the price for all vehicles for all of their brands, Jtm , with the objective of maximiz- ing firm profit. Observed prices form a Nash equilibrium to the pricing game. We assume a constant marginal cost, cjt , associated with producing a vehicle in a given year. The pricing first-order condition for vehicle j is:  ∂s jt (6) s jt + (p jt − c jt ) = 0. ∂ pkt k∈Jt m  MARKET POWER IN U.S. AUTOMOBILES 1221 These first-order conditions will be used in conjunction with the estimated demand system to solve for marginal costs for each product. Marginal costs will then be used to compute markups and for counterfactual analysis. For a subset of counterfactual Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 analysis, we parameterize marginal costs to depend on vehicle co- variates including elements of xjt and cost shifters excluded from demand, which we describe in detail in Online Appendix C.2. Our assumption of Nash-Bertrand pricing rules out cartels or other changes in conduct over the time period.7 If firms became more or less collusive, then the implied marginal costs inferred by assuming a static Nash equilibrium in prices would be mislead- ing. We consider alternative conduct assumptions for robustness and analyze alternative models of conduct in counterfactual anal- ysis. However, we do not attempt to measure changes in conduct as in Bresnahan (1982), Lau (1982), or Duarte et al. (2023). V. ESTIMATION AND RESULTS We estimate the model using generalized method of moments (GMM), closely following the procedures outlined by Petrin (2002) and Berry, Levinsohn, and Pakes (2004). Our estimation proce- dure is implemented in three steps. We briefly outline each step here and provide a full description in Online Appendix A. In the first step, we jointly estimate consumer heterogeneity and the mean consumer valuations. We compute the conditional demographic and second-choice moments from the model and con- struct a GMM estimator matching these to their analogues in the consumer-level choice data. We employ micro-moments from two sources: (i) demographic information linked to car purchases from MRI and CEX and (ii) second-choice information from the Mar- itzCX survey. An example of a moment for the first source is the difference between the observed and predicted average price of vehicle purchases for each quintile of the income distribution. For 7. We also rule out the effect that voluntary export restraints (VERs) in the 1980s and corporate average fuel economy (CAFE) standards have on optimal pricing. See Goldberg (1995) and Berry, Levinsohn, and Pakes (1999) for supply- side models of VERs and Goldberg (1998) and Gillingham (2013) for models of CAFE standards. In both cases, the marginal costs that we recover reflect the shadow costs of adhering to these restrictions. 1222 THE QUARTERLY JOURNAL OF ECONOMICS the second source, we match the correlations in car characteris- tics between the purchased and second-choice cars.8 In the second step, we estimate ᾱ and β̄ and year fixed ef- fects by regressing the estimated consumer mean valuations on Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 product characteristics, prices, make dummies, and year dum- mies. Our assumption that xjt and the RXR are uncorrelated with product-level demand shocks provides the classic moment condi- tions for 2SLS. The year fixed effects absorb the structural param- eters for annual variation in mean car quality, τ t , and preference for outside good, γ t. In the third step we use the empirical analogue of the continuing-product condition equation (2) to separately estimate τ t and γ t from the estimated year effects. We compute standard errors using a bootstrap procedure. We resample the microdata, including the sampled households in the CEX and MRI surveys as well as the MaritzCX survey, and rees- timate the model following the same three-step procedure. We account for the sampling variation in ξ jt in the second step of the estimation procedure. In the 500 bootstrap draws of the mi- crodata, we use a nested parametric bootstrap, clustering at the make level, of the second-step estimation. V.A. Parameter Estimates Table IV presents the coefficient estimates for mean coeffi- cients (column β), random coefficients (column σ ), and the demo- graphic interactions (remaining columns). The demographic esti- mates are intuitive and match clear patterns in the microdata. Higher-income and older consumers are less price sensitive for the relevant range of incomes. Larger family size households have stronger preferences for vans and vehicle footprint. Rural house- holds have a stronger preference for trucks. In general, we esti- mate large and economically meaningful coefficients representing unobserved heterogeneity, which rationalizes very strong substi- tution patterns observed in the second-choice data. The largest random coefficients appear on vehicle style, suggesting that con- sumers substitute most strongly within vehicle style. The random coefficient associated with trucks is double the magnitude of the interaction of truck with a rural consumer dummy variable, sug- gesting that unobservable taste heterogeneity is quantitatively 8. See Online Appendix Table A.4 for a complete list of micro-moments.  MARKET POWER IN U.S. AUTOMOBILES 1223 Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 FIGURE IV Distribution of Price Sensitivity The plot displays smoothed kernel regression of 10,000 draws from the esti- mated distribution of α i , by year, for the baseline specification with constant ᾱ over time. important. Electric vehicles also have a large estimated random coefficient. Although we fix model parameters over time in our main specification, the distribution of price sensitivity and other tastes does change due to changes in the distribution of consumer demo- graphics over time. For example, Figure IV presents the distribu- tion of consumers’ price sensitivity, α i , in 1985, 2000, and 2015. Over the data period, there was a shift in the distribution toward less price sensitivity, which is a reflection of higher incomes and an older population. This, together with changes in the product set, drives changes in the elasticity of demand over time. To en- sure that our main results on markups are not driven by our as- sumption of fixed parameters, we perform robustness checks by allowing more flexibility in α and other parameters. We report results on estimated markups below for these alternate specifi- cations in the main text, and we discuss the technical details in Online Appendix B. Our estimates of own-price elasticities for the earlier years in our sample are similar to BLP, Goldberg (1995), and Petrin (2002). The average share-weighted own-price elasticity across our entire sample is −5.06. Table V displays elasticities for the aggregate market and for a group of parent companies. Berry, Levinsohn, and Pakes (2004), on the suggestion of analysts at 1224 THE QUARTERLY JOURNAL OF ECONOMICS TABLE V SELECTED ELASTICITIES Year Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 1985 1995 2005 2015 Average own-price elasticity −4.23 −5.30 −5.78 −5.36 Market elasticity −1.07 −1.44 −1.38 −1.29 Ford −3.51 −4.21 −5.29 −4.75 GM −2.64 −3.75 −4.60 −4.72 Toyota −3.40 −5.06 −4.67 −4.40 Volkswagen −4.15 −5.42 −5.54 −5.45 Hyundai — −3.43 −3.93 −4.11 Notes. “Average own-price elasticity” is the percent change in sales for a 1% increase in price, averaged across each available product (share-weighted). “Market elasticity” is the percentage change in the sales of all vehicles for a 1% increase in the price of all vehicles. Manufacturer-specific elasticities represent the percent change in sales for all cars of that manufacturer for a 1% increase in price for all cars of that manufacturer. General Motors, calibrate their model by targeting an aggregate price elasticity of −1 for 1993. Our estimates roughly validate this assumption. Demand elasticities became more elastic from 1980 to 2005 in these categories with most of the change from 1985 to 1995. They are level or decline slightly thereafter. V.B. Decomposition of Time Effects The restriction in equation (2) decomposes the time effects into average improvements in unobservable car quality and rel- ative movements in the utility of the outside good over time— potentially due to business cycle factors or changes in the utility of not purchasing a new car. Figure V displays the results of this decomposition. We find that unobservable vehicle quality is steadily increasing, roughly linearly, by a cumulative total of about $25,000. The value of the outside option also generally increases over the time period with noticeable deviations from trend during the 1990–1991 and 2007– 2009 recessions. Our model points to a substantial improvement in the qual- ity of automobiles over the sample period, equal to approximately the mean price of a new car in the early part of the sample pe- riod. The economic meaning of this increase is that a consumer faced with the choice between two new automobiles of the same observable characteristics (e.g., size, horsepower, fuel economy) but with average unobserved quality (e.g., airbags, sound system,  MARKET POWER IN U.S. AUTOMOBILES 1225 Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 FIGURE V Quality and Aggregate Components of Time Effects Average unobserved quality, τ t , and value of the outside good, γ t , in 2015 dollars. See the text for estimation details. durability) of 1980 versus 2018 would place a significantly higher value on the 2018 vehicle. To quantitatively assess the plausibil- ity of the estimated unobserved quality component, we manually collected data from the Kelly Blue Book website in 2021 for mint- condition used automobiles produced every five years between 1992 and 2017. We then regressed the Kelly Blue Book private- party transaction value against characteristics and dummy vari- ables for the year of production. The year of production dummy variables should capture the average unobserved product differ- ences across years of production. The full specification is pre- sented in Online Appendix C.4. We find that the year of produc- tion dummies rise by $19,638.88 between 1992 and 2017, which is nearly the increase we estimate for the value of unobserved product improvements, suggesting the estimate is not implausi- bly large. A number of narratives also support such large increases. Automobiles have become safer through features such as im- proved airbag technology, body construction, rearview cameras, and blind-spot sensors. According to the National Highway Traf- fic Safety Administration (NHTSA), fatalities not involving alco- hol impairment per vehicle miles traveled (VMT) have decreased 40% between 1982 and 2019 from 1.27 per hundred million VMT 1226 THE QUARTERLY JOURNAL OF ECONOMICS to 0.74 per hundred million VMT.9 Unobserved comfort improve- ments include power steering, durable interior materials, and electronic features such as Bluetooth audio systems and power or heated seats. Many of these features had not even been invented Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 at the start of the sample. Finally, car durability is likely an important aspect for both the increased quality of new cars and the value of the outside good (which includes driving used cars). We would expect increased car durability to increase the value of a car. Between 1980 and 2018, data from the NHTSA implies that the average time a con- sumer keeps a new car has risen from 3.9 to 5.9 years, consistent with increased durability. This is part of the improvement in un- observed quality captured by our quality adjustment, τ t , along with improvements in safety, comfort, and electronics. However, as cars become more durable, households will replace them less often, which has the effect of making the outside option appear more attractive. We expect this effect to be captured in the out- side good part of the time effect, γ t. The outside-option series is broader than durability, however. In addition to improvements in the attributes of used cars, the outside option is influenced by alternative transportation methods, such as public transport or ride-sharing, or changes in the commuting needs of the popula- tion. It will also be affected by business cycle fluctuations or mon- etary policy that may lead consumers to accelerate or postpone new car purchases. V.C. Model Fit We target correlations between the attributes of purchased cars and stated second choices for survey years 1991, 1999, 2005, and 2015. The first column of Table VI presents the average cor- relation across years for each attribute we target. These corre- lations suggest strong substitution patterns among vehicles with similar characteristics. As seen in the second column of Table VI, our estimated model is able to match these moments well. To emphasize the importance of observed and unobserved consumer heterogeneity in our model, we compare our fit to a series of more restrictive models. In the third column, we present the implied correlations from a model with only demographic heterogeneity 9. Although this could also be due to safer driving behavior or safer road con- struction, the rise of distracted driving because of mobile handsets likely pushes in the opposite direction.  MARKET POWER IN U.S. AUTOMOBILES 1227 TABLE VI ATTRIBUTE CORRELATION BETWEEN FIRST AND SECOND CHOICES Alternative specifications Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 Only dem. & Only Data Model footprint RC demographics Logit Van 0.720 0.727 0.048 0.008 −0.008 SUV 0.642 0.640 0.018 −0.007 −0.010 Truck 0.843 0.798 0.246 −0.013 −0.024 Footprint 0.710 0.693 0.665 −0.002 −0.018 Horsepower 0.599 0.588 0.384 0.009 −0.012 MPG 0.647 0.657 0.362 0.003 −0.013 Luxury 0.484 0.493 0.031 0.005 −0.005 Sport 0.277 0.291 0.001 −0.004 −0.004 Electric 0.373 0.192 0.002 −0.001 −0.001 Euro brand 0.336 0.353 0.018 0.000 −0.003 U.S. brand 0.479 0.472 0.121 −0.010 −0.012 Notes. Data from MaritzCX survey, 1991, 1999, 2005, 2015. The numbers are the average across these four years. “Model” column represents the predictions from the model presented in Table IV, and the first column of Online Appendix Tables A.1 and A.2. The “Logit” column contains model predictions from a simple logit demand specification, with no observed or unobserved heterogeneity. The “Only demographics” column contains model predictions from a model with the same demographic interactions as our main specification but without any unobserved heterogeneity. “Logit” and “Only demographics” are estimated without moments on second choices. and a random coefficient on footprint. This model is roughly able to match the second-choice correlation on footprint but under- states the remaining second-choice correlations, even those one would expect to be highly correlated with footprint (e.g., horse- power, miles per gallon, and truck). The fourth column drops the random coefficient on footprint. Surprisingly, this model achieves essentially none of the second-choice correlations reported in the data. This is despite the fact that it matches demographic pat- terns well, as reported in Online Appendix Table A.4. Indeed, it is only a slightly improved fit for second choices over the logit model in the fifth column, which restricts substitution by assuming independence of irrelevant alternatives. We conclude that observable heterogeneity alone is insufficient to generate substitution patterns implied by the second-choice survey data. Online Appendix Table A.5 shows that the model matches the second-choice correlations separately in each year that we have second-choice data. Online Appendix Table A.4 displays the fit of all of the demographic moments we match. 1228 THE QUARTERLY JOURNAL OF ECONOMICS Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 (A) (B) FIGURE VI Markups Panel A displays the median, 10th, 25th, 75th, and 90th percentiles of markups over time. Panel B displays share-weighted markups for our baseline specification and alternative specifications described in the text. Refer to Online Appendix B for a detailed description of robustness specifications. V.D. Markup Estimates We infer marginal costs of each vehicle using the first-order conditions in equation (6) at the estimated demand parameters. Online Appendix Table A.8 displays the coefficient estimates from projecting inferred marginal costs on vehicle attributes and cost shifters. Together with the observed vehicle prices and shares, we use the marginal costs to calculate vehicle markups expressed as Lerner indices ( p−mc p ). Figure VI, Panel A displays the distri- bution of markups (median, interquartile range, and 10th–90th percentiles) over time. We estimate that the median markup is falling in our sample, from 0.325 to 0.185. Markups across the distribution also decrease. To ensure that the functional form of our utility function is not the primary driver of the decrease in markups, we estimate alternative specifications of the model and compare the implied share-weighted markups with our baseline in Figure VI, Panel B.10 First, we allow for a linear trend in α, the dot-dashed line, and markups have a similar trend over the sample. Second, we estimate a separate α for each five-year segment of our data, the line with triangle markers. The downward trend in markups per- sists. Third, we use the log of price in the utility function, shown as the light dashed line. Under this specification, the decrease in 10. For those specifications that allow for time heterogeneity in α, we add as- sembly country dummies as additional instruments to increase first-stage power. We discuss this instrument set in Online Appendix B. The baseline results are nearly identical when estimated with this instrument set.  MARKET POWER IN U.S. AUTOMOBILES 1229 Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 (A) (B) FIGURE VII Markups over Time by Vehicle Style and Import Status Share-weighted (by category) mean markups. Vehicle style is defined in the text. “Domestic” are those cars produced in the United States, regardless of brand head- quarters. share-weighted mean markups is less dramatic, 0.32 to 0.27, than in the baseline, 0.42 to 0.22. However, when we have logged price and also allow flexibility in the price parameter with separate pa- rameters for each five-year segment (the dashed line with circles), the trend in markups looks very similar to our baseline specifica- tion. Last, we estimate the entire model for three separate time segments: 1980–1992, 1993–2004, and 2005–2018. The choice of these segments is motivated by the coverage of our survey data. The markups for this specification are the light gray dotted line, and the decrease in markups is slightly more than in the baseline, with the first 13 years having a nearly identical match. Overall, we conclude that our finding of decreasing markups is robust to alternative specifications, so we continue the analysis with our baseline specification. The details of these alternative specifica- tions are in Online Appendix B. In Figure VII, we display share-weighted average markups by vehicle style in Panel A and by import status in Panel B. The decline in markups occurs across all vehicle styles and for both imported and domestically produced vehicles. Starting with Panel A, truck markups were higher than other vehicles at the begin- ning of our sample but fell more steeply throughout the 1990’s. This is likely due to two factors, a steeper increase in the qual- ity and price of trucks and slightly greater competition as the popularity of foreign-manufactured trucks increased. This spe- cific pattern is consistent with the move by Toyota and Nissan to produce trucks in the US to avoid the so-called chicken tax.11 11. The chicken tax is the informal name for the 25% tariff on light trucks imported into the United States. It was originally imposed during the Johnson 1230 THE QUARTERLY JOURNAL OF ECONOMICS Markups for SUVs also experienced a sharp fall during the 1990s, likely due to the massive increase in competition in this segment. The number of SUVs available nearly tripled during this time, and our demand estimates imply strong within-category substi- Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 tution. Turning to Figure VII, Panel B overall, imported vehicles have lower markups than domestically produced vehicles in the early decades of our sample, where our classification is based on the country of production, not the headquarters country of the product. However, domestically produced vehicles experienced a much greater fall in markups over this period, and markups are roughly equal between domestic and imported products in the fi- nal decades of our sample. To assess sampling variability in the estimated markup trend, we use a bootstrap procedure accounting for sampling vari- ability in the demand estimates, demographic data, and the ξ jt residuals. In our baseline results, only a single product out of 9,694 has inelastic demand and all consumer price sensitivities are negative. However, in some of our bootstrap samples, some products have positive elasticities due to some consumers having positive price sensitivities. In these cases, which make up 5.6% of products over all bootstrap draws, the Nash pricing condition cannot be satisfied, and there is no inversion from observed prices to marginal costs.12 This occurs for at least one firm in 14.2% of year and bootstrap combinations. In all bootstrap samples where the inversion is well-defined for all firms in 1980 and 2018, we find that median markups decrease over the sample period. 1. Explaining the Evolution of Markups. What drives the decline in markups? In the model, the exogenous forces that can change markups are changes in the ownership configuration, product entry and exit and associated changes in product charac- teristics, changes in the value of the outside option, and changes in consumer demographics or preferences. In our data and esti- mates, all of these forces are active throughout the time period. administration to retaliate against European countries imposing a tariff on U.S. poultry. 12. One possible route to avoid this issue would be to add restrictions to in- crease the precision of our estimates of price sensitivity. These restrictions could take the form of additional exclusion restrictions or enforcing the supply model as part of the estimation.  MARKET POWER IN U.S. AUTOMOBILES 1231 Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 (A) (B) (C) (D) FIGURE VIII Markups, Prices, and Shares Panel A displays share-weighted mean markups for our baseline model and a model that assumes each product’s price is set independently of all other products. In Panel B, average prices are in 2015 US$. An intermediate observation to understand the estimated change in markups is that the trend is similar if we infer markups assuming single-product firms, as seen in Figure VIII, Panel A. Assuming single-product firms is a good approximation if vehi- cles manufactured by the same parent are not strong substitutes for each other. In the single-product firm case, the Lerner index is equal to the inverse elasticity of the product: p − mc 1 s 1 (7) = = × ds. p elas p dp In the remaining panels of Figure VIII, we plot average prices (panel B), average market shares (C), and average deriva- tives of share with respect to price (D), noting that some intu- ition about the drivers of markups over time can be gleaned de- spite each being both an average and an endogenous function of the underlying preference, technology, and ownership struc- ture primitives. This decomposition also emphasizes that price elasticities are key estimands in determining markups. For our full model, this includes both own-price elasticities as illustrated 1232 THE QUARTERLY JOURNAL OF ECONOMICS in equation (7) and Figure VIII for single-product markups and cross-price elasticities that also enter the markup equation of multiproduct firms. During the period 1980 to 1999, when estimated markups de- Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 creased, average market shares and the average of their deriva- tives with respect to price are stable, while average prices increased. This combination suggests that markups decrease ac- cording to equation (7). The economic reason prices are increasing without shares decreasing and without changes in the derivative of share with respect to price is that vehicle quality is increasing. In the period 2000 to 2019, markups are stable as average mar- ket shares are decreasing, the average of their derivatives with respect to price are increasing, and average prices are roughly stable. In this latter period, although quality is still increasing steadily, the outside option also experiences substantial growth, which can explain the flattened average price trend and offsetting the decline in average shares and increase in the average of their derivative with respect to price. Under the logic of equation (7), this combination leads to flat markups. To study which primitive factors explain the estimated de- cline in markups, we turn to counterfactual simulations. To con- sider the impact of concentration, the first counterfactual we perform adjusts the ownership matrix in each year to remove the impact of the growth of competition from foreign brands since 1980. To consider the effect of product proliferation, our second counterfactual holds the number of products fixed over time at the level of 1980. These counterfactuals are described in full in Section VI.B as Mechanisms 1 and 2. These changes to primi- tives do not eliminate the decrease in markups we observed in our baseline results. We display the results in Online Appendix Fig- ure XVIII. Next we simulate a counterfactual where the observ- able characteristics of vehicles in each year are scaled down to match the distribution of characteristics from 1980. Specifically, if a vehicle is in a certain percentile of a characteristic in a given year, we assign the same percentile from the 1980 distribution of that characteristic. As a result of this change, which shifts the dis- tribution of products toward lighter, lower-horsepower vehicles, the increase in marginal costs over time estimated by our model is effectively eliminated, as shown in Figure IX, Panel A. The rea- son marginal costs are flat despite an estimated downward tech- nological trend is that there is an offsetting upward time trend in the RXR, see Online Appendix Figure A.8. Other primitives, like  MARKET POWER IN U.S. AUTOMOBILES 1233 Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 (A) (B) FIGURE IX Counterfactual Markups, 1980 Distribution of Characteristics For each vehicle in each year, we assign the same percentile from the 1980 distri- bution of each characteristic, recompute marginal costs, which are plotted in Panel A, and recompute the pricing equilibrium and share-weighted mean markups which are plotted in Panel B. the number of products and the market structure, are allowed to evolve as they do in the data. This counterfactual, which ef- fectively eliminates the growth in observed product quality, does significantly reduce the fall in markups, as shown in Figure IX, Panel B. The main takeaway of this exercise is that a major driver of the decline in markups is that increasing observable quality of vehicles results in increasing marginal costs that are less than fully passed through to consumer prices. The importance of vehicle quality in driving markup trends highlights the fact that markups are not conceptually attractive proxies for welfare when the product set is changing.13 This fact motivates our focus on the model’s measures of welfare and sur- plus over time to assess industry performance in Section VI. V.E. Robustness to Conduct Assumption In this section, we compare markup estimates under alter- native assumptions of conduct. To summarize the results, while there is a disparity in the level of markups, these alternatives all point toward declining markups over the sample period, as 13. For a simple example of when markups can be misleading, consider a monopolist facing logit demand with u = δ − αp + ε, whose market share is s = exp(δ−α p) 1+exp(δ−α p). The pricing first-order condition is p = c + α(11−s ) = c + α1 (1 + exp(δ − α p)). Suppose the product improves in quality without changing its marginal cost. Totally differentiating the first-order condition with respect to δ, we find ddδp = αs > 0. Since marginal cost is constant, this implies that markups rise. However, since d(δ−α p) dδ = 1 − s > 0 consumer surplus also increases. 1234 THE QUARTERLY JOURNAL OF ECONOMICS Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 FIGURE X Markups: Alternative Conduct Assumptions Estimated share-weighted mean markups for Nash-Bertrand pricing by parent companies (Baseline), the Big Three U.S. automobile manufacturers colluding for every year in our sample (Big 3 Collusion), and joint price-setting by every parent company in our sample (Full Collusion). in the base case of Nash-Bertrand pricing. In the first case, we assume the Big Three U.S. auto manufacturers (GM, Ford, and Chrysler) collude on prices for our entire sample.14 Markups are much higher than our baseline case in the 1980s, but then be- come closer to our baseline case throughout time. This is consis- tent with the decline in the dominance of the Big Three firms over time. Notably, markups at the end of the sample under the assumption that the Big Three collude are lower than the Nash- Bertrand markups at the start of the sample. Therefore, under the assumption that the Big Three were competing in 1980 and organized a pricing cartel in response to import competition after 1980, we would still find a decline in markups between 1980 and 2018. In the second case, we consider markups that are implied if all of the firms colluded on prices. In this case, markups are much higher. However, there is still a decrease in markups over the time period. Figure X establishes that markups decline over time under a variety of constant conduct assumptions. However, it is possi- 14. For Chrysler, we follow the ownership from Chrysler to Daimler to Cere- bus private equity firm, then to Fiat, and assume that the owner of Chrysler col- ludes with all of the ultimate owner’s brands. For example, then the Fiat brand is part of the “cartel” after 2012.  MARKET POWER IN U.S. AUTOMOBILES 1235 TABLE VII AVERAGE MARKUPS WITH DIFFERENT CARTEL ASSUMPTIONS Mean markup HHI Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 1980 baseline 0.42 2,661 2018 baseline 0.22 1,132 2018 hypothetical cartel membership GM + Ford + Toyota 0.26 2,546 Top 3 + Fiat 0.35 3,724 Top 4 + Honda 0.39 4,819 Top 5 + Nissan 0.46 6,000 Note. Computed share-weighted mean markups and HHI with simulated collusion in 2018 for various manufacturer cartels. Fiat is the parent company of Chrysler in 2018. ble that a cartel could have formed during our sample period. We now ask how large such a cartel would need to be to have held markups constant over the period. To quantify this, we con- sider different size cartels in 2018 to measure how many cartel members it would take for a cartel in 2018 to achieve the base- line noncollusive level of markups found in 1980. Specifically, we form cartels with the largest (by sales) manufacturers, adding one manufacturer at a time. The results are in Table VII. One change in conduct from Nash-Bertrand that would produce estimated in- creases in markups would involve a cartel of the six largest par- ent companies (Top Five + Nissan) forming during our sample. Overall, it seems that a price-fixing cartel on the scale needed to keep markups at their 1980 level would be unlikely to escape the notice of antitrust authorities. Indeed, given the greater con- centration among U.S. automakers, it seems more likely that the 1970s and earlier periods would be subject to coordinated pricing decisions.15 V.F. Comparison to the Production-Based Approach De Loecker, Eeckhout, and Unger (2020) (DLEU) use finan- cial data from Compustat to estimate markups. This approach 15. Bresnahan (1987) investigates a potential breakdown in collusion among U.S. automakers in 1955. More recently, in 2013 the Department of Justice secured convictions of nine automobile parts suppliers fixing prices of sales to U.S. auto manufacturers plants (see https://www.justice.gov/opa/pr/ nine-automobile-parts-manufacturers-and-two-executives-agree-plead-guilty- fixing-prices), suggesting that the DOJ would be attuned to coordination in the auto industry itself. The effect of this collusive ring raising manufacturers’ costs of inputs would be captured in our estimates of marginal cost. 1236 THE QUARTERLY JOURNAL OF ECONOMICS Downloaded from https://academic.oup.com/qje/article/139/2/1201/7276495 by Poli MI user on 22 January 2025 (A) (B) FIGURE XI Comparison to De Loecker, Eeckhout, and Unger (2020) Panel A displays share-weighted mean price over marginal cost in our estimates, the estimate for share-weighted mean price over marginal cost in the U.S. auto- mobile industry from De Loecker, Eeckhout, and Unger (2020), and the average estimate across 1971–1990 from Berry, Levinsohn, and Pakes (1995). Panel B dis- plays our estimate of total variable profits, quantity sold multiplied by margins, summed across all products. uses a model of firm production and data on input expenditures and output revenue to estimate price over marginal cost ratios.16 In their baseline results, they estimate an increase in the sales- weighted average price to marginal cost ratio (across all sectors) from 1.21 to 1.61 from 1980 to 2016. In addition to aggregate re- sults, DLEU report estimates for specific industries, including the U.S. auto industry. Figure XI, Panel A displays the time series of average price to marginal cost ratio from their work together with our own measure. We also include an estimate from Berry, Levinsohn, and Pakes (1995), which reports an average price to marginal cost ratio from 1971–1990. Both the level and trends in the price to marginal cost ratio differ from the estimates we derive, though both series are relatively flat from 1995 onward. In the right panel, we plot our estimates for total variable prof- its, which is the sum of price minus marginal cost multiplied by quantity sold over models in a year. Quantity thus enters directly into the right panel but does not enter directly in our estimates in the left panel. Our estimates for total variable profits share some patterns with the DLEU estimates for markups, including an in- crease in the 1980s, a dip and recovery in the 1990s, and a dip and recovery around the Great Recession. 16. For purposes of comparison, this section reports markups as the price to marginal cost ratio pc rather than the Lerner index, p−c p.  MARKET POWER IN U.S. AUTOMOBILES 1237 The two markup estimates rely on different underlying data and nonnested sets