The car market in the Czech Republic : an empirical analysis and merger simulation

(1)

University of Amsterdam

Faculty of Economics and Business

The car market in the Czech Republic: An

empirical analysis and merger simulation

MASTER’S THESIS

submitted in partial fulfilment for the degree of Master of

Science in Economics in specialization Markets and Regulation

Author: Jan M´alek (11375205)

Supervisor: Prof. Dr. Jo Seldeslachts Academic Year: 2016/2017

(2)

Statement of Originality

This document is written by Student Jan M´alek who declares to take full

responsi-bility for the contents of this document.

I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used to create it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

Acknowledgments

I would like to express my sincerest gratitude to Mr. Ondˇrej Bˇehal and the company

Economia, a.s. who were so kind and allowed me to use their database of price lists. Without their help, it would not be possible to write this thesis.

Furthermore, I would like to thank to Prof. Jo Seldeslachts for his comments and supervision of my thesis.

Last but not least, I owe much to my parents who have been continuously sup-porting me in my study efforts by all means.

(4)

Abstract

This thesis focuses on the empirical analysis of the Czech market for new cars and involves competitive assessment of the effects of the takeover of Opel by PSA Group announced in February 2017, using a merger simulation approach. It employs a unique dataset created specifically for the needs of the assessment and covering the market from 2012 to 2017. The results support the hypothesis that the merger causes a price increase but the magnitude is modest - less than 2% in all segments but mul-tipurpose vehicles (MPV), where the increase amounts to 4%. Consumers lose 7.5m Euros and producers benefit with more than 200m Euros. Required efficiencies for price neutrality of the merger are small - below 1.5% (4.1% for MPV). In conclusion, it is unlikely that the merger would significantly increase the market power of the firms and that an impediment to effective competition would arise. The merger can be cleared in the Czech market.

JEL Classification L13, L41, C53, D22

Keywords Automobiles, Nested logit, Bertrand

compe-tition, Horizontal Merger, Merger simulation, Competition Economics

Author’s e-mail janmalek.jm@gmail.com

(5)

List of Tables

2.1 Most popular models in 2017Q1 . . . 5

2.2 List of produced models . . . 7

2.3 Brands by market share . . . 8

2.4 Companies by market share . . . 8

4.1 Sample coverage . . . 16

4.2 Variables overview and summary statistics . . . 17

4.3 Cross-correlation table . . . 18

5.1 NL, instumented. Price variable princ and market size 3% of house-holds. Dependent variable ln(sj) − ln(s0). . . 27

5.2 Effects of brands on the mean valuation. . . 31

5.3 Network overview . . . 31

5.4 Elasticities, all years . . . 32

5.5 Supply side results for 2016 . . . 33

5.6 Multiproduct firm SSNIP TEST for a relevant market . . . 35

5.7 Market shares per company in each segment as of 2016. [%]. Based on the sample. . . 36

5.8 HHI in the merger-relevant segments in 2016 . . . 36

5.9 Effects of the merger . . . 38

5.10 Merger effects: 10% efficiencies . . . 41

5.11 HHI in the Skoda demerger in 2016 . . . 42

5.12 Effects of the demerger of Skoda . . . 44

A.1 Statistics by year . . . III

A.2 Statistics by segment . . . III

A.3 Market shares of all sold brands by year . . . IV

A.4 Market shares per company and segment on total sales of 2016. [%] . V

A.5 NL, non-instrumented. Price variable princ and market size 3% of

households. Dependent variable ln(sj) − ln(s0). . . VI

A.6 Sensitivity analysis to the market size. Dependent variable ln(sj) −

ln(s0) . . . VII

A.7 First stage regressions. . . VIII

(7)

List of Tables vii

A.9 Supply side results by year . . . X

A.10 Markups of the companies . . . XI

(8)

Chapter 1 Introduction

In February 2017, General Motors (GM) announced the intention to sell its remaining operations in Europe – the company Opel/Vauxhall– for approximately 2.2 billion Euros to the French PSA Group. By acquiring Opel, PSA is going to become Eu-rope’s second-ranked carmaker by sales with a 16 percent market share beneath only Volkswagen’s 24 percent.

Such a merger naturally raises many questions about the impacts on consumers and the resulting level of competition, especially as it takes place in such a

concen-trated and economies-of-scale prone industry as the car market1_.

Due to the magnitude of the merger, the European Commission is supposed to

examine the effects and approve or prohibit it at the EU level2_{. However, detailed}

analysis of the merger at a national level only is also of significant policy interest. This thesis focuses on a competitive assessment of the mentioned merger in the

context of the Czech market3. The motivation is driven by the fact that with 10

million inhabitants and 1.34 million of cars produced, the Czech Republic is the second largest per capita producer of cars in the world (OICA, 2015). The automotive industry accounts for 25% of the country’s manufacturing sector, 20% of exports,

7.5% of GDP and employs roughly 150 thousand people ( ˇCesk´a spoˇritelna, a.s., 2015).

In addition, with 260 thousand cars registered in 2016, the country is the twelfth largest car market in Europe. The Czech market also has a highly asymmetric and unique structure where the Volkswagen group holds more than 46% market share. This is a consequence of the ownership of nationally perceived brand Skoda which enjoys particular acclaim among customers.

1_{Car manufacturing is a capital intensive industry with significant fixed costs. Thus, a more} efficient production can be theoretically reached when the scale of operation of a particular company increases.

2_{According to the European Commission (2004a), the undertakings are obliged to notify the} Competition Authority about the merger if their combined turnover exceeds 5bn Euros worldwide and 250m Euros EU-wide for each company. Both GM Europe (Opel) and PSA exceed this threshold by far.

3_{According to the Czech law, the merging parties have to notify the Czech Competition} Author-ity, UOHS, based on the amount of net revenue. The law speaks about the threshold of 1.5bn CZK for all merging parties which was earned in the Czech market. If the merger would be national specific only, both companies individually surpass this threshold, as table A.9 shows, and would need to notify the UOHS.

(9)

1. Introduction 2

The central research question of this thesis is the following: “What is the effect of the PSA purchase of Opel on car prices and firm markups in the Czech market for new cars? ”. Ex-ante, the hypothesis is that the merger will lead to a small increase in prices and that it will allow PSA to charge a higher margin as the merged entity becomes the fourth biggest seller in the market. However, no structural change is expected as the market share of both PSA and Opel are around 5% each in the aggregate market. The shares are higher in the lower-range and multipurpose segments though.

Merger assessment is a well-established discipline in competition economics and represents what the majority of the practical workload of competition authorities consists of. The (anti)competitive effects of a merger are often evaluated by the em-ployment of a simulation approach, which is also the method this thesis uses. The simulation approach requires extensive information about market data, in particular regarding prices, quantities, elasticities and other characteristics that differ on a case by case basis. In the context of the Czech Republic, previous studies that would provide such information are not available. Thus, the only feasible way to quantify the effects of the merger using the simulation method is to first collect appropriate data, conduct an analysis of demand and supply side, extract the necessary informa-tion and finally, gauge the effects per se. The ambiinforma-tion of this thesis is thus twofold: it attempts to partly fill the existing gap in the empirical IO literature regarding the Czech car market and to contribute to the policy debates by quantifying the competitive effects of the ongoing merger.

The remainder of this thesis is structured as follows. Chapter 2 provides an introduction to the Czech car market. Consumer preferences and their developments are introduced, a general overview of the market and the ownership structure is given, and the trends in the registrations and production are shown. Chapter 3 presents the theoretical framework applicable to this thesis and gives an overview of the relevant literature on discrete-choice modelling and merger simulation. Chapter 4 consists of two main sections. The first one describes the process of creating the dataset and lists all relevant sources. The second part focuses then on the models of both the demand side and the supply side and shows necessary derivations. In Chapter 5, the results of the demand side and the supply side estimation are provided and discussed. Furthermore, this chapter defines the relevant market for the mergers and evaluates the impact of the Opel-PSA merger on consumers and the level of competition. Furthermore, it briefly presents a small policy experiment of demerger of Skoda from Volkswagen. Finally, Chapter 6 summarizes the findings.

(10)

Chapter 2 Market for new personal cars: A

de-scription

Demand side

As of 2016, the Czech market for new cars was the 12th largest in the EU with 259 693 registrations. The number of new car registrations per 1000 inhabitants was 21.9 which is below the European average of 27 but considerably above the average of 11.5 for the other post-communist countries (ACEA, 2015). Figure 2.1 displays the development of total registrations and registrations of domestically produced cars over the period of the data sample (2012-2017Q1).

Figure 2.1: Evolution of (domestic) registrations and production

0 5 10 15 20 25 30 01/ 2012 04/ 2012 07/ 2012 10/ 2012 01/ 2013 04/ 2013 07/ 2013 10/ 2013 01/ 2014 04/ 2014 07/ 2014 10/ 2014 01/ 2015 04/ 2015 07/ 2015 10/ 2015 01/ 2016 04/ 2016 07/ 2016 10/ 2016 01/ 2017 Car s (t h o u san d s) Time Registrations (Left) Local sales (Left)

30 50 70 90 110 130 150 C ar s (t h ou san d s) Production (Right)

Source: Author. Data source: SDA-CIA, AutoSAP

A positive trend in number of registrations is complemented by seasonal fluc-tuations. In particular, registrations are lower in the first and third quarter by an average of 10%, compared to second and fourth quarter. The share of domestic cars on total monthly registrations fluctuates between 40% and 50% with a decreasing trend, although the absolute numbers of registrations grow.

Changes in tastes have also occurred. Figure 2.2 shows the market shares of individual segments and their development. The segmentation corresponds to the

(11)

2. Market for new personal cars: A description 4

well established marketing categories of the Car Importers Association (SDA-CIA) which are used as a standard in the Czech market.

Figure 2.2: Share of segments

0 5 10 15 20 25 30 01/ 2012 04/ 2012 07/ 2012 10/ 2012 01/ 2013 04/ 2013 07/ 2013 10/ 2013 01/ 2014 04/ 2014 07/ 2014 10/ 201 4 01/ 2015 04/ 2015 07/ 2015 10/ 2015 01/ 2016 04/ 2016 07/ 2016 10/ 2016 01/ 2017 Sh ar e [ % ] Time

Mini Small Lower middle Middle

Upper middle Sport+Lux MPV SUV and Terrain

Note: MPV - multipurpose vehicles; SUV - sport-utility vehicles. Source: Author.

The figure identifies two basic clusters by shares. The popularity of the segments of mini, upper middle, luxury and sport cars is low as their market shares amount to less than 7%, 3 % and 1 % respectively. The shares are fairly stable over time.

The opposite holds for the remaining segments in the upper cluster where the trends and the ordering of the most popular segments behave more dynamically. Initially, small cars dominated the market with approximately 25% share in 2012. However, the segment started to decline in 2013, probably to the benefit of lower middle cars. Certain recovery of small cars has appeared since the second half of 2015. Lower middle cars evince the most dynamic development. They increase share on new registrations from approximately 15% to a little more than 25% which means the segment was a market leader between 2013 and mid 2015. After then, the position fluctuated. The share of middle cars shows declining trend over time. Although the segment dominated the market in January 2012, it ended as least popular from the upper cluster in 2017 with 15 % share. The sport-utility vehicle (SUV) segment has been growing at a significant pace on a year-to-year basis and became market leader in 2016. This tendency fully complies with the world-wide

(12)

trend of an increasing popularity of SUVs. The trend for multipurpose vehicles (MPV) is rather stable and oscillates around 15%, even though approximately 2% fall associated with the termination of the production of Skoda Roomster occurred in summer 2015 (Auto.cz, 2015; SDA-CIA, 2017). The reason for stability of the development in MPVs is probably related to the specific nature of the segment - the cars are high-roof, spacious and can mostly transport up to 7 people or be easily modified to a small cargo carrier.

On aggregate level, consumers favour petrol cars over diesel cars. However, this does not hold universally on a segment level basis. As a rule, consumers buy petrol-engine cars in lower segments such as mini, small and lower middle much more than diesel cars. In all the other segments, a larger fraction of diesel engines prevail. The exception is the segment of sport cars which are predominantly sold with petrol engines due to their technical characteristics. The proportion of cars with alternative fuels (LNG, electricity) is negligible. They constitute less than 3% of newly registered cars and the trend is stable over time. Detailed figures for fuel characteristics are given in Appendix A.

The table 2.1 presents the most popular models sold on the market. Although it provides the state as of the first quarter of 2017, the ordering almost does not change over time. The most sold cars in the Czech Republic are all from Skoda, which is a unique pattern in the European context where mostly Volkswagen or Renault branded cars occupy first places.

Table 2.1: Most popular models in 2017Q1

Place Brand Registr. Share Place Brand Registr. Share

1 Skoda OCTAVIA 6 810 10.00% 6 Hyundai i30 1 599 2.34%

2 Skoda FABIA 5 088 7.47% 7 Ford FIESTA 1 430 2.10%

3 Skoda RAPID 2 845 4.18% 8 Hyundai ix20 1 384 2.03%

4 Skoda SUPERB 2 400 3.52% 9 Skoda YETI 1 362 2.00%

5 Volkswagen GOLF 2 052 3.01% 10 Dacia DUSTER 1 320 1.93%

Data source: SDA-CIA.

Supply side and ownership structure

Figure 2.3 presents an ownership structure in a compact matrix. All brands with non-negligible shares are presented and grouped by company. The star at particular brand signifies if a brand is incorporated in the dataset (see section 4.1).

From the ownership perspective, there is not a single Czech company in the market. However, lot of cars are physically produced within the borders of the

(13)

Figure 2.3: Ownership structure

Aston Maritn Aston Martin Nissan Motor Co. Infiniti

BMW BMW * Nissan *

Mini PSA Peugeot Citroën Citroën *

Daimler Mercedes-Benz * DS *

Smart * Peugeot *

Ford Motor Ford * Renault Group Dacia *

Fiat Chrysler Auto. Alfa Romeo * Lada

Dodge Renault *

Ferrari Rolls Royce Rolls Royce

Fiat * SsangYong SsangYong

Chrysler Subaru Corp. Subaru *

Jeep * Suzuki Motor Co. Suzuki *

Lancia Tata Motors Jaguar *

Maserati Land Rover *

Geely Volvo * Tesla Inc. Tesla

General Motors Cadillac Toyota Motor Co. Lexus *

Chevrolet * Toyota *

Opel * Volkswagen Group Bentley

Honda Motor Co. Honda * Lamborghini

Hyundai Motor Co. Kia * Porsche

Hyundai * Seat *

Mazda Motor Co. Mazda * Škoda *

Mitsubishi Group Mitsubishi * Volkswagen *

Note: Sorted by companies in alphabetical order. The star signifies that given brand is included in the dataset. Source: Author based on information from the companies’ websites

(14)

country. There are five factories and five companies manufacturing several car models as table 2.2 shows.

Table 2.2: List of produced models

Company Plant Models

Volkswagen Kvasiny Skoda: Superb, Yeti, Kodiaq, Roomster ; Seat: Ateca Mlada Boleslav Skoda: Fabia, Rapid, Octavia

Vrchlabi VW group automatic transmission systems

Hyundai Nosovice Hyundai: ix20, ix35, i30, Tucson

TPCA Kolin Citroen: C1; Peugeot: 107, 108; Toyota: Aygo

The production of the models in italics was terminated; the regular font stands for the models which are currently being produced. TPCA is a short-cut for a joint-venture of Toyota, Peugeot and Citroen. Data source: websites of manufacturers.

All production shown in figure 2.1 comes from mentioned plants. In 2016, the production accounted for a record of 1 351 124 cars. The difference between pro-duction and domestic registrations series indicates that a majority of domestically produced cars is exported. Even if all imports of cars are taken into consideration, the Czech Republic remains a net exporter by a notable margin - on average, the net exports outnumbered the imports almost seven times in 2016.

Production exhibits an upward-sloping trend but is subject to a considerable seasonality. Every year July, August and December are connected to a decrease in production. On average, production is lower by 26% during those months in comparison to the rest of the year. This is determined by the presence of company-wide holidays on the production sites during these months. The extraordinary drop in July 2016 is associated with the fact that all Skoda production sites were shut down between July 4 and July 22 as a major modernization took place (Skoda Auto, 2016).

As table 2.2 suggests, an important aspect of production is that the companies try to maximize efficiency by utilizing the individual platforms to assemble cars of similar segment and comparable characteristics. An appropriate illustration is the mini-car joint-venture of TPCA or the production of Seat Ateca in Kvasiny plant together with its twin, Skoda Kodiaq.

The following paragraphs build on figure 2.3 and present the market structure in a detailed way. It is important to stress that the number of offered brands and models do not necessarily correlate with a market share of a particular company.

(15)

Market structure

The Czech car market seems concentrated and this pattern does not change much over years. Tables 2.3 and 2.4 give an overview of market shares from different perspectives. Additional figures and tables are available in Appendix A.

Table 2.3: Brands by market share

Brand 2012 2013 2014 2015 2016 2017 Skoda 30.91 30.33 30.21 32.02 31.68 29.68 Volkswagen 8.73 9.07 9.51 10.29 10.24 11.94 Hyundai 8.71 9.86 9.85 8.53 8.08 7.34 Ford 7.31 5.74 6.54 6.68 5.88 6.11 Dacia 2.19 3.43 4.83 4.55 4.77 5.13 Renault 6.01 3.67 3.28 3.50 3.93 3.97 Kia 4.92 3.84 3.48 3.29 3.44 3.20 Opel 2.20 2.70 3.59 3.51 3.26 3.18 Peugeot 3.86 4.40 3.73 3.10 3.12 3.41 Mercedes-Benz 1.86 1.92 2.01 2.09 2.89 3.13 Seat 1.54 2.65 3.20 2.93 2.85 3.06 BMW 2.24 2.27 2.36 2.55 2.48 2.52 Audi 2.19 2.00 2.05 2.17 2.45 1.94 Toyota 2.27 2.53 2.08 1.89 2.20 2.87 Citroen 3.28 3.63 2.83 2.18 2.04 1.89 Nissan 1.82 1.72 1.58 2.17 1.92 2.25 Fiat 1.49 1.60 1.44 1.47 1.31 1.27 Suzuki 1.34 1.39 1.22 1.05 1.29 1.23 Mazda 0.69 1.08 1.21 1.14 1.20 1.08 Mitsubishi 0.64 0.64 0.89 1.05 1.01 0.75 Note: Market shares in percentages, ordered based on val-ues of 2016. The year 2017 includes only first quarter. Only first 20 most sold brands reported. Table A.3 in Ap-pendix A lists all sold brands. Data source: SDA-CIA.

Table 2.4: Companies by market share

Company 2012 2013 2014 2015 2016 2017 Volkswagen Group 43.59 44.23 45.11 47.58 47.38 46.81 Hyundai Motor Co. 13.63 13.70 13.33 11.82 11.53 10.54 Renault Group 8.26 7.18 8.16 8.09 8.74 9.18 Ford Motor 7.31 5.74 6.54 6.68 5.88 6.11 PSA Peug. Cit. 7.15 8.03 6.56 5.31 5.22 5.35 General Motors 3.84 3.77 3.82 3.53 3.27 3.24 Daimler 1.90 1.97 2.04 2.13 2.94 3.15 BMW 2.42 2.48 2.52 2.82 2.70 2.84 Toyota Motor Co. 2.37 2.63 2.15 1.96 2.30 2.97 Nissan Motor Co. 1.85 1.75 1.62 2.19 1.93 2.25 Fiat Chrysler Auto. 1.99 2.08 1.94 2.05 1.88 1.78 Suzuki Motor Co. 1.34 1.39 1.22 1.05 1.29 1.23 Mazda Motor Co. 0.69 1.08 1.21 1.14 1.20 1.08 Mitsubishi Group 0.64 0.64 0.89 1.05 1.01 0.75 Honda Motor Co. 0.97 1.13 0.97 0.78 0.87 0.76 Geely 0.88 0.93 0.84 0.75 0.74 0.82 Tata Motors 0.44 0.54 0.47 0.42 0.49 0.47 Subaru Corp 0.51 0.50 0.42 0.44 0.38 0.40 Other 0.20 0.23 0.19 0.20 0.22 0.23 Tesla Inc. 0.00 0.00 0.00 0.00 0.02 0.04 Rolls Royce 0.00 0.00 0.00 0.01 0.01 0.00 Aston Maritn 0.00 0.00 0.00 0.00 0.00 0.01 SsangYong Motor 0.01 0.01 0.01 0.00 0.00 0.00 Note: Market shares in percentages, ordered based on values of 2016. The year 2017 includes only first quarter. Data source: SDA-CIA.

Brand-wise, Skoda is a stable leader with a market share around 30% which is almost triple of the second Volkswagen, with approximately 12% (2017 values). Hyundai is currently the third most sold brand with 7%, although it has been second in previous years. Ford holds continuously the fourth position (6%). The fastest growing brand is Dacia which has been occupying fifth place (5%) for several years although it used to be thirteenth in 2012. Furthermore, next six brands have around 3% market share each. After that, the shares start to drop slowly. The top five brands have a combined market share of 60% which has been stable over time.

The market becomes even more concentrated and asymmetric from the company-level perspective. The Hirschman-Herfindahhl index for 2016 reaches the value of 2561 which is, in view of most practitioners, consistent with calling the industry

highly concentrated1,2_{. The prominent contributor to the concentration is the}

Volks-1_{Herfindahl-Hirschman Index (HHI) is a measure of the size of firms in relation to the industry} and an indicator of the amount of competition among them. It is computed as a sum of squared market shares. If the shares enter in percentages, the index ranges between 0 and 10000 with the lower bound representing perfect competition and the upper bound standing for a monopoly market.

2_{According to the US merger guidelines, the industry is called highly concentrated if the HHI} exceeds 2500. The European guidelines speak then about the value of 2000.

(16)

wagen group which sells nearly every second car on the market. The second largest company by market share, Hyundai Motor Co., serves a 4x smaller fraction of the market but its share is decreasing. Renaults Group slowly grows to their current 9 % market share with the potential to surpass Hyundai if the trends for both companies continue. Ford Motor maintains the stable share of 6%.

The primary objective of this thesis is to quantify the effects of the Opel takeover. For that reason, a detailed decomposition of the market positions of PSA and GM is provided in figure 2.4.

Figure 2.4: Sales and shares of PSA and GM

0 1 2 3 4 5 6 7 8 9 0 2000 4000 6000 8000 10000 12000 14000

PSA GM PSA GM PSA GM PSA GM PSA GM

M ar ket s h ar e Re gi st rat io n s

Citroën Peugeot DS Cadillac

Chevrolet Opel PSA share GM share

2012 2013 2014 2015 2016

PSA has been the fifth biggest seller since 2015 with a market share about 5%. However, its market share has been steadily decreasing and absolute number of registrations has stagnated, despite the introduction of a new brand, DS, in 2014. On the contrary, GM has been able to slowly increase the number of its absolute registrations, mainly due to the performance of Opel. The growth of Opel even compensated for exit of Chevrolet at the end of 2015 and GM thus managed to stay sixth largest seller with a share around 4 %. When the operations of both companies are combined, the newly-emerged entity surpasses Ford and ends up as the fourth largest company with an aggregate market share of 8.59% (as of 2017) selling 4 brands altogether.

(17)

Chapter 3 Literature overview

In reality, there are many car manufacturers of different sizes which operate in the market and compete for customers by offering various sorts of cars for the lowest pos-sible price. Translated into economic terms, this thesis therefore interprets the car market as an oligopolistic structure with unequally-sized firms, selling differentiated goods and competing on prices in Bertrand fashion. A particularly adept class of models that are used for analysis of such market setting are discrete choice models. Discrete choice models grew to become a standard approach to industrial organi-zation analysis due to their tractability and ability to reduce the dimensionality of product space into the smaller space occupied by product attributes (Bonnet et al., 2016).

Since the early work of McFadden (1973), an extensive literature has emerged providing flexible discrete choice models well suited to estimation and inference. Powerful and later popular contributions were made by Berry (1994); Berry et al. (1995) and Petrin (2002) whose models encompass both sides of the market. On the demand side, the models work with unobservables and heterogeneity in consumer preferences. Consumers are treated as utility maximizers who choose a particular product based on its characteristics. This allows rich substitution patterns which can be retrieved from the estimated demand structure. On the supply side then, firms’ behaviour is mostly modelled as a one-stage Bertrand game in an oligopolistic environment, based on early insights of Rosse (1970) and Bresnahan (1981; 1982). By combining both sides, one can infer marginal costs (without actually observing them) and examine a range of market counterfactuals - which is an essential feature in the context of a merger simulation (Berry & Haile, 2014).

The framework in Berry et al. (1995) explicitly presents the random coefficient logit model, the nested logit model and simple logit model. All those model types are based on aggregate sales data. In such a setting, one observes market shares, market characteristics, product prices and characteristics, and product or market level cost shifters, but not individual choices or firms’ costs (Berry & Haile, 2014).

There exists a vast body of literature that employed discrete choice models for analyses of various industries - automotive (Berry et al., 1995; Verboven, 1996;

(18)

Gold-3. Literature overview 11

berg & Verboven, 2001; Petrin, 2002; Berry et al., 2004; Ivaldi & Verboven, 2005b; Brenkers & Verboven, 2005; 2006; Reynaert & Verboven, 2014; Verboven & Grigolon, 2014; Xavier et al., 2014; Guajardo et al., 2015; Ohashi & Toyama, 2017; Konishi & Zhao, 2017); beer production (Rojas, 2005; Rojas & Peterson, 2008; Beelitz et al.,

2009; Kusuda, 2011; Toro-Gonz´alez et al., 2014), cereals (Nevo, 2000; 2001; Shum,

2004) and housing market (Yates & Mackay, 2006). Furthermore, because of the power of the model framework to provide insights in demand and supply, the models were applied to much broader range of topics than just traditional IO matters. Ex-amples include various policy questions - market power, mergers, welfare gains for new goods/technologies, network effects, media bias, trade policy and many others. Detailed overview can be found in Train (2003), Berry & Haile (2014) and Berry & Haile (2015).

There are several developments in the literature. For example, Berry & Haile (2009; 2014) consider identification in the case micro-level data are accessible. A completely new thinking about product differentiation was initiated by the work of Pinkse et al. (2002) and Pinkse & Slade (2004) who introduced the elements of spatial economics to the literature and allowed the transition from discrete-choice to continuous-choice modelling. Within this framework, known also as “distance metric” approach, differentiation is perceived as a distance in the broadest sense of the word. Differentiation can take place in the geographic, attribute, demographic or temporal space. Additional detailed discussion of current trends in the literature can be found in Bonnet et al. (2016).

Merger simulation

Merger simulation is one of the profound examples of a structural modelling approach in IO. Structural models try to build an economic supply-demand based model of the respective market and model the overall equilibrium. Merger simulation is usually executed as a sequence of three steps. Firstly, the demand system is estimated and cross price and own price elasticities are empirically deducted. The second step fo-cuses on the oligopoly pricing modelling, mostly by assuming Bertrand competition, in order to retrieve the implied marginal costs of individual companies. The last step entails a comparison of pre-merger and post-merger equilibria, after the changes in the market structure and/or possible cost efficiencies are taken into consideration.

Before the merger simulation was introduced, the assessment of merger effects was based primarily on the definition of the relevant market and comparison of market shares using straightforward measures such as HHI. This approach was deemed as insufficient already a long time ago (Werden & Froeb, 1996) as it did not allow for a proper quantitative analysis of post-merger effects and for the evaluation of strategic

(19)

3. Literature overview 12

responses. Particularly problematic is this approach in the context of industries with differentiated products. Markups can be high in a relatively unconcentrated market if the sold products are not close substitutes (Pinkse et al., 2002).

Merger simulation was introduced by Hausman et al. (1994) and triggered a rich discussion shortly afterwards (Werden & Froeb, 1994; Shapiro, 1995; Werden, 1996; Werden & Froeb, 1996; Hausman & Leonard, 1996). Many subsequent studies emerged which used the framework to analyse various aspects of mergers. Several examples of papers include Crooke et al. (1999); Nevo (2000); Epstein &

Rubin-feld (2002); Slade (2006); Verboven & Grigolon (2014); Bj¨ornerstedt & Verboven

(2016) and many others. Real application of the simulation method in practice then comprises numerous cases of the European Commission (e.g. Volvo/Scania, Lagarde/Natexis/VUP, Nuon/Reliant, Oracle/PeopleSoft). More detailed overview can be found in Budzinski & Ruhmer (2010) or Russo et al. (2010).

Advantages of a merger simulation are manifold. Firstly, it is appealing for

its tractability and clear structure. As Crooke et al. (1999) notes, merger simula-tion is based on three key assumpsimula-tions: the chosen form of competitive interacsimula-tion, the shape of marginal cost curves and the demand system in the market. These assumptions allow use of well-understood techniques for inspection of the market. In addition, Epstein & Rubinfeld (2002) emphasize that merger simulation can of-fer assessments of competitive effects and remedies that are beyond the reach of other methods of inquiry - e.g. merger-specific efficiencies, product repositioning and adequacy of proposed divestitures. Budzinski & Ruhmer (2010) list further rea-sons, among others that the main strength of a merger simulation is that it forces economists to be clear about made assumptions which contributes to the trans-parency of the assessment. Lastly, a major advantage is that the framework of the simulation can be freely adjusted on case by case basis rather that being applied as a universal non-adjustable tool. This feature is also stressed in Walker (2005).

The assumptions, transparency and complexity of the simulation models are, however, at heart of the criticisms. Practitioners have criticized that the Bertrand model often assumes more competition than is present in real markets; that the treatment of production costs is too simplistic; or that a particular demand system is unrealistic (Cheung, 2016). Moreover, there is an ongoing debate about the struc-tural modelling approaches in IO. Some authors propose to abandon the strucstruc-tural modelling approach and rather suggest to use “treatment effect” designs which would avoid the necessary specification assumptions (Ashenfelter et al., 2009). However, as Nevo & Whinston (2010) note, such approach heavily depends on availability of information on past, comparable and structurally similar mergers. Nonetheless, this is rarely the case in the merger assessment practice. Furthermore, mergers are

(20)

3. Literature overview 13

not exogeneous effects as the treatment approach would require (Nevo & Whinston, 2010).

In conclusion, merger simulation is still a well accepted tool for evaluating the merger effects as it provides much deeper insight than measures based on market share paradigm (Nevo & Whinston, 2010). However, a general consensus prevails that a more extensive ex-post evaluation of the performance of these models is needed in the long term perspective.

(21)

Chapter 4 Methodology

4.1 Dataset

Because of the absence of an appropriate dataset that would satisfy the requirements of this analysis, a dataset covering the Czech market for personal cars and small

multipurpose vehicles was created1_{. Both listed car types are included in the M1}

category of the European vehicle categories2_.

Except several key variables, such as prices and quantities, it is not ex-ante apparent what data in particular about the car market and car characteristics should be collected. For that reason, the methodology of Brenkers & Verboven (2006); Verboven & Grigolon (2014) and others is used, who were repeatedly analysing the car market (see Chapter 3 for a detailed overview), and constructed structurally

similar datasets for the Western European markets3.

Variables

The dataset contains four groups of variables: quantities, prices, car characteristics and (socio)economic indicators.

Quantities are defined as new car registrations at the lowest possible level of data – per car type (i.e. Skoda Octavia, Opel Astra, BWM X3, ...). The depth of the data does not allow to distinguish between various types of engines or levels of equipment for the whole period in question. Thus for consistency reasons, basic model of particular vehicle is always taken as representative. The data follow the

1_{“Small multipurpose vehicles” stands only for minivans.} _{Vans are not taken into account} in the construction of the dataset. Examples of vans are: Ford Transit, Mercedes-Benz Sprinter, Volkswagen Transporter. Minivans include for instance Dacia Dokker, Fiat Doblo, Peugeot Partner, Citroen Berlingo.

2_{The definition of the M1 category is the following: Vehicles designed and constructed for the} carriage of passengers and comprising no more than eight seats in addition to the driver’s seat, and having a maximum mass (”technically permissible maximum laden mass”) not exceeding 3.5 tons (European Commission, 2007).

3_{The older dataset covering Belgium, France, Germany, Italy and the U.K between 1970 and} 1999 is publicly available on the following website: https://sites.google.com/site/frankverbo/data-and-software. The newer dataset covering altogether nine Western European countries between 1998 and 2011 is not public.

(22)

4. Methodology 15

distinction between petrol/diesel engines though and adopt the fuel-type that has been sold more at particular vehicle throughout the whole period of the sample. Electric cars are excluded from the dataset because of their technological dissimi-larity and negligible market share. The sources for car registrations are publicly available databases of the Czech Ministry of Transportation and repositories of the

Car Importers Association4.

Prices correspond to gross list prices, including VAT, suggested by the producers as the final retail prices. Unconditional discounts to all customers offered by the producers in official catalogues are taken into account. In contrast, special discounts, for instance a price reduction in exchange for an old car, are not considered as not all customers qualify for them universally.

The availability of prices was a major threat to the construction of the dataset. Car makers typically publish only current price lists on their websites and remove their older versions. Therefore, it was not possible to rely only on publicly available sources to obtain historical price series. To resolve this issue, I contacted the company Economia a.s., an owner of car-comparison platform www.vybermiauto.cz, which agreed to provide price data for the purpose of this thesis. I collected the data at the headquarters of the company during February 2017. The remaining price data for March 2017 were then amassed from the manufacturers’ websites.

Vehicle characteristics include the measures of car size (curb weight, length, height, width), engine performance (number of cylinders, displacement, horsepower, acceleration and maximum speed) and efficiency indicators (fuel consumption and emission production). In addition, dummy variables based on a perceived origin of the car are created on the country as well as on a continent basis. The source of the characteristics are various internet websites (www.auto-data.net, www.automobile-catalog.com, www.parkers.co.uk) as well as the websites and price lists of individual car producers.

The last category covers the socio-economic variables such as GDP, population and number of households that are necessary to compute the ratio of price/income per capita and the size of potential market respectively. The series are obtained from the Czech Statistical Office.

Availability and coverage

The availability of price data determined timespan and coverage of the dataset. Although only 5 years of data were accessible (2012-2017Q1), their quality was high. This allowed to decrease the time unit to the level of quarters, which is an important

4_Ministry’s _repository: http://www.mdcr.cz/Statistiky/Silnicni-doprava/Centralni-registr-vozidel; Car Importers Association’s repository: https://www.sda-cia.cz/repository?lang=EN

(23)

4. Methodology 16

distinction from other car-market studies, e.g. Brenkers & Verboven (2006) and Verboven & Grigolon (2014), who were working with yearly data. Intuitively, the deaggregation into quarters is plausible as the car producers usually change prices several times a year (often on a quarterly basis), new car models or facelifts are introduced during the year and also the car registrations exhibit a seasonal pattern as seen in Chapter 2. Moreover, quarterly deaggregation increases the number of observations four times. As a result, quarterly treatment of data should capture more variability and produce more reliable results than yearly treatment would.

The cross-sectional dimension of the dataset encompasses 303 car types sold in the Czech republic during 2012-2017. The brands and vehicles not covered represent only a small portion of total registrations. The only structural exception with non-negligible market share (0.14%) are the Porsche cars for which the prices were not available. The size of the sample accounts for 95.79% of the whole market. Table 4.1 displays the sample coverage by year.

Table 4.1: Sample coverage

Year 2012 2013 2014 2015 2016 2017Q1 Total

Total registrations 174 009 164 736 192 314 230 857 259 693 68 059 1 089 668

Dataset 165 686 157 659 184 909 221 039 249 933 64 954 1 044 180

Coverage ( %) 95.22 95.70 96.15 95.75 96.24 95.44 95.79

Summary statistics

Table 4.2 gives the reader an idea of the shape of the data and the overall state of pertinent variables before any regression analysis is performed. Meaning and full description, including units of measure, are given in the table’s footnote.

The dataset consists of 4892 observations and is not balanced which is determined by the fact that the product portfolio of each producer develops over time as new models are introduced and the production of old ones is terminated. Generally, more new products were initiated than phased-out as the number of models in the dataset has been steadily increasing (from 214 in 2012Q1 to 236 in 2017Q1).

Table A.1 in Appendix A shows the most important variables grouped by year. Several trends are evident from the data. Whilst the mean and standard deviation of the ratio of price to per capita income remained relatively stable over time, average registrations grew fast. The car market has been subject to a significant technological advancement. On average, cars became more efficient - the horsepower, maximal speed and width increased while fuel consumption and the range of fuel consumption declined.

(24)

4. Methodology 17

Table 4.2: Variables overview and summary statistics

Variable Mean Std. Dev. Min. Max.

price 0.73 0.58 0.14 4.31 princ 1.84 1.46 0.37 10.58 regist 213.45 552.64 0 8269 cyl 4.23 0.91 3 10 disp 1822.39 665.79 898 5204 hp 144.71 71.42 60 540 weight 1451.91 343.83 830 2570 length 4427.57 376.17 3415 5137 width 1806.69 84.11 1535 2090 height 1552.62 136.68 1183 1981 fuel 5.76 1.44 3.5 14.9 accel 10.79 2.71 3.5 19.6 speed 193.37 27.94 140 319 luggage 440.28 172.94 100 1400 GDP 4150938.1 155004.84 3981303 4443919.68 pop 10535.13 23.94 10506.14 10585.19 households 4375122 0 4375122 4375122 N 4892

Variables: price - price of a car in millions of CZK; princ - price/ yearly GDP per capita; regist - number of registered cars (represents sales); cyl - number of cylinders; displ - displacement in cm3; hp -horsepower in PS; weight - weight in kg;length - length in mm; width - width in mm; height - height in mm; fuel - fuel consumption in l/100km; accel - acceleration in sec/[0-100km/h]; speed - maximum speed in km/h; luggage luggage space in l (VDA norm); GDP -yearly GDP in million of 2010 CZK; pop - population in thousands; households - number of economic households as of 2011

(25)

4. Methodology 18

One of the key variables, not shown in the previous table, is the segmentation variable which categorizes the cars into mutually exclusive groups based on their characteristics. As will be explained later in section 4.2, this variable serves as a nesting parameter in the nested logit model. To resemble the market as close as possible, the cars are categorized according to the commonly accepted classification of the Car Importers Association (SDA-CIA). The basic summary statistics of most important variables by segments are depicted in the table A.2 in the appendix.

The majority of variables varies across segments - which is a desirable feature. Illustrative graphical example is shown in Figure A.5. On average, prices, perfor-mance characteristics and engine power grow with the exclusivity of the segments. On the contrary, the number of registrations decreases. The trunk space grows with segments as well, with the exception of sport cars where the low capacity of luggage space is defined by their construction.

Table 4.3 indicates the pairwise relationships between the characteristics and the most obvious links in the data. The price (princ) and quantity sold are weakly neg-atively related but price (princ) and all engine-specific characteristics are strongly positively correlated. Acceleration and height relate to price negatively - on aver-age, the cars in higher segments (and especially the sports segment) are lower and accelerate faster.

Table 4.3: Cross-correlation table

Variables price princ regist disp cyl hp height width fuel speed accel luggage

price 1.00 princ 1.00 1.00 regist -0.19 -0.19 1.00 disp 0.89 0.89 -0.20 1.00 cyl 0.84 0.85 -0.14 0.87 1.00 hp 0.93 0.93 -0.21 0.91 0.82 1.00 height -0.06 -0.06 -0.05 0.06 -0.00 -0.11 1.00 width 0.64 0.64 -0.13 0.66 0.56 0.64 0.22 1.00 fuel 0.49 0.50 -0.16 0.60 0.58 0.60 0.15 0.31 1.00 speed 0.79 0.79 -0.15 0.75 0.66 0.88 -0.37 0.58 0.36 1.00 accel -0.77 -0.77 0.18 -0.75 -0.63 -0.86 0.27 -0.59 -0.37 -0.92 1.00 luggage 0.17 0.17 0.04 0.28 0.20 0.13 0.57 0.50 0.04 0.07 -0.07 1.00

4.2 Model

4.2.1 Demand

As discussed in chapter 3, this thesis employs a model from the class of discrete choice models to estimate demand and retrieve elasticities. In particular, the nested

(26)

4. Methodology 19

logit (NL) model is utilized as it proved useful in modelling of demand in the car market in numerous previous studies (list given in Chapter 3). Nested logit enables to realistically model the decision making processes consumers undergo by utilizing one or two levels of nesting categories across which the preferences of consumers are correlated (Verboven & Grigolon, 2014). Firstly, consumers decide which segment fits their needs the best; secondly they may decide about the origin (sub-segment) of the car and lastly they choose a particular car maximizing their utility based on its characteristics. The suitability of one or two-level nesting depends on the particular dataset and on the meaningfulness of the obtained results.

There are two elementary demand specifications that appear in the NL appli-cations - unit demand or constant expenditure. Unit demand specification has a rich history of applications as it dates back to Berry et al. (1995) who also apply it in the context of automobile markets. Unit demand specification implies that con-sumers buy a single product or they do not buy at all (and hence prefer the outside option yielding zero utility). On the other hand, constant expenditure differs from unit specification in the sense that consumers spend a fraction of income to pur-chase a particular good. The potential market is determined by the available budget

rather than the number of consumers. Bj¨ornerstedt & Verboven (2016) apply such

a framework in the context of the Swedish market for analgesics.

Besides this, a major advantage of nested logit is its computational simplicity following from the linearisation ot the regression equation. Furthermore, the model is much less restrictive in substitution patterns than the non-nested version of logit models; it allows products in the same segment (“nest”) to be closer substitutes than products in other groups. The group segmentation restricts the patterns to some extent though - cross price elasticities within the same group are still symmet-ric and substitution outside a group is symmetsymmet-ric to all other groups (Verboven & Grigolon, 2014). Another important feature of the nested logit is that it deals with

the independence of irrelevant alternatives assumption5_{. The assumption continues}

to hold in the same nest but is not satisfied between goods in different nests - which is an important distinction from simple non-nested logit models (Ivaldi & Verboven, 2005a).

The random coefficient nested logit (RCNL) model would allow for much more flexibility than nested logit because it also takes into account the correlations be-tween the continuous product characteristics. However, the cost is an upsurge in computational difficulty as such a model needs to be numerically solved. As Knittel & Metaxoglou (2014) indicate, this can be a major challenge as random coefficient

5_{This property implies that the ratio of the choice probabilities for n products remains the same} irrespective of some n+1th product being added to the choice set.

(27)

4. Methodology 20

models are highly non-linear and the optimization algorithms often do not converge. Furthermore, Verboven & Grigolon (2014) discuss in a detailed way the differences in performance of various discrete-choice models and conclude that there is a statistical evidence to reject the NL against the RCNL model. However, the results of NL are closest to the RCNL from all of the tested logit models. More importantly in the context of this thesis, Verboven & Grigolon (2014) conclude that the implications of nested logit for the market definition and merger simulation are comparable to RCNL and that nested logit should thus be sufficient to obtain reliable policy results. Specification

The demand estimation is based on the nested logit model as described in Berry

(1994); Verboven (1996); Brenkers & Verboven (2005; 2006) and Bj¨ornerstedt et al.

(2014) with some small modifications relevant for application in this thesis.

There is one national market with consumers i, i = 1, 2, ..., N . As the unit demand specification indicates, each consumer i can choose one of the j products, j ∈ {1, 2, . . . J + 1}, in time t = 1, 2, ..., T , where the J + 1th product represents the outside option. The products are classified into G + 1 exhaustive and mutually exclusive groups, where the goods j belong to g = 1, 2, ..., G + 1 and the G + 1th group is reserved solely for the outside good. The size of the national market M is given by the number of potential customers which entails all buying customers and

customers who prefer to buy an outside good yielding zero utility6. The addition

of the outside good is necessary if one wants to study the aggregate demand with this model framework. As Berry (1994) explains, the outside good ensures that the demand is not fully determined by differences in prices only.

Consumer i derives indirect utility from buying the good j in time t of the form

uijt= δjt + ¯ijt, (4.1)

where the first term represents mean utility that is common to all consumers and the second term stands for the mean-zero consumer-specific utility term.

The common part of the utility is based on the following formula

δjt = x0jtβ − αpjt + ξjt. (4.2)

The transposed vector x0_jt is a K-dimensional vector of observed product

character-6_{The definition of the size of a potential market is rather vague in the literature. It can be} estimated or assumed. In most applications, the number of households is taken and divided by some factor. The result then constitutes the potential market of size M, with M being naturally bigger than the number of actually sold goods. Despite the fact that the market size is determined heuristically, the model results are robust to the definition of the potential market in most studies.

(28)

4. Methodology 21

istics, including e.g. the origin or brand dummies, pjt is the price and ξjt is the term

for the valuation characteristics that are unobserved to the econometrician. The parameters β and α then determine the magnitude of the effect of the respective variables.

As shown in Brenkers & Verboven (2005) and Verboven & Grigolon (2014), the consumer specific part of the utility can be rewritten in the following way

¯

ijt = ζigt+ (1 − σ)ijt. (4.3)

The ζigt and ijtare i.i.d random variables. The meaning is then that ζigt is consumer

i’s utility, common to all products belonging to group g, whereas the term ijt is

consumer i’s utility, specific to product j. The nesting parameter σ, 0 < σ < 1 measures the correlation of preferences for cars for the same group g. As σ approaches 1, preferences for products of the same group become perfectly correlated and the goods in particular segment are perceived as perfect substitutes. In contrast, as σ approaches zero, the goods become independent and customers switch between them

irrespective of segments7_.

Furthermore, following Berry (1994), the equation (4.3) can be rewritten to the following form

uijt = x0jtβ + ξjt− αpjt+

X

g

[djgtζigt](1 − σ)ijt, (4.4)

where djgtis equal to one if j ∈ g and zero otherwise. In this form, the indirect utility

function contains directly the consumer heterogeneity for the continuous

character-istics in x and consumer heterogeneity for the group dummy variables djgt.

In every period t, consumer i chooses product j so that his utility is maximized. On aggregate level, the probability of choosing car j coincides with the market share

sj (Brenkers & Verboven, 2005). The probability of choosing product j, sj, is equal

to the product of the probability of choosing group g, ¯sg, and the probability of

choosing product j given group g, ¯sj|g:

sjt(δ, σ) = ¯sgt(δ, σ) · ¯sj|gt(δ, σ) = D_gt1−σ P gD 1−σ gt exp(δjt 1−σ) Dgt = exp( δjt 1−σ) Dσ gt[ P gD 1−σ gt ] , (4.5) with Dgt= X g exp δ_jt 1 − σ (4.6)

(29)

4. Methodology 22

and the outside good’s utility normalized to zero, δ0t = 0. Moreover, this specification

implies that s0t(δ, σ) = P 1

gD 1−σ gt

.

Following Berry (1994) again and taking the natural logarithms of sjt and s0t and

summing them, the following is obtained:

ln(sjt) − ln(s0t) =

δjt

1 − σ − σln(Dgt). (4.7)

The unknown value for σln(Dgt) can be retrieved from the expression for group share

sgt and by recognizing that it can be rewritten using the term for s0t:

ln(sgt) = ln(Dgt1−σs0t) = (1 − σ)ln(Dgt) + ln(s0t). (4.8)

Finally, by combining (4.7) and (4.8) the expression for mean utility is retrieved

δjt = ln(sjt) − σln(sj|gt) − ln(s0t) (4.9)

which, together with (4.2), leads directly to an equation that is subject to estimation

ln(sjt/s0t) = β0+ x0jtβ − αpjt+ σln(sj|g,t) +

N −1

X

i=1

γibrandi+ ξt+ ∆ξjt, (4.10)

where ξjt = ξt+ ∆ξjt. The time fixed effects can be understood as a tool to control

macro-economic fluctuations that affect the decision to purchase a new car. The brand dummies are taken out of the vector x for clarity reasons.

As Brenkers & Verboven (2005) note, it is important to stress that for the model to be consistent with utility maximization, α has to be positive and σ has to lie between 0 and 1.

It is also possible to derive the expressions for elasticities from the specified model.

As Bj¨ornerstedt & Verboven (2016) show, the demand elasticities can be computed

and written in aggregate form as follows:

ηjk = ∂qj ∂pk pk qj = −αpk 1 1 − σD 1 jk − σ 1 − σsj|gD 2 jk− sk (4.11) with D_jk1 =    1 if j = k 0 if j 6= k and D_jk2 =    1 if j, k ∈ g 0 if j, k /∈ g

(30)

4. Methodology 23

In the further text, ηjj is used to denote the own price elasticity, ηjk to denote the

cross price elasticity in the same segment and ηjl to denote the cross price elasticity

between different segments. Instruments

An essential assumption in the regression model is that the explanatory variables are uncorrelated with the error term. The variables in vector x are assumed to be predetermined and thus satisfy the condition of being exogeneous (Berry, 1994; Berry et al., 1995). Although many studies have discussed the viability of such an assumption (Gandhi & Houde, 2016), it is common in the literature to work with it. As Verboven (1996) notes, there should not be a concern about this assumption in the short run because firms cannot immediately adjust characteristics of sold cars as a response to a shock aggregated by the error term.

Nevertheless, exogeneity is clearly not satisfied in the case of prices and the

segmentation variable sj|g,t. In the former case, this is caused by the fact that the

price of a car is not related only to the observed characteristics in x, but also to unobservable characteristics such as design, colour, delivery time and advertising

captured by ξjt. Moreover, as the observed characteristics affect the marginal cost

of production, simultaneity bias can be another source of endogeneity (Verboven,

1996). In the latter case then, sj|g,t is endogeneous because the segments are a

choice variables in the model.

To deal with endogeneity, prices and the segmentation variables need to be in-strumented. Conventional instrumental-variable candidates would be supply side variables, i.e. demand shifters, which satisfy the condition of relevance and exogene-ity in standard settings. However, it is hard to observe some of these variables and collect corresponding data. This obstacle was removed by Berry et al. (1995) and Verboven (1996) who show that in an oligopolistic setting such as the car market, only demand-side instruments are sufficient. This connects to the fact that in such an environment, pricing for a product j does not solely depend on its own charac-teristics, but due to oligopolistic interdependence also on the characteristics of the other products sold by the same firm and on characteristics of competing products. Following the work of Verboven (1996) and Brenkers & Verboven (2006), this thesis employs the following instruments: i) number of products of the competing firms in the same segment, ii) number of products of the same firm in the same segment, iii) sums of the characteristics of the competing firms’ product in the same segment and iv) sums of the characteristics of the products of the same firm in the same segment.

(31)

4. Methodology 24

4.2.2 Supply

Similarly to Nevo (2001); Brenkers & Verboven (2006) and Bj¨ornerstedt et al. (2014),

supply side is based on the standard profit-maximizing behaviour of firms that are competing in a Bertrand fashion in a multiproduct environment in every period t:

Πf t(p) =

X

j∈Ff

(pjt− cjt)qjt(p), (4.12)

where Πf(p) is the profit of firm f , the sum represents the set of price-cost

margins of the subset of cars Ff that are owned by firm f and qj(p) stands for the

demand.

The corresponding first-order conditions have the following form:

qj(p) + X k∈Ff (pk− ck) ∂qk(p) ∂pj = 0. (4.13)

In matrix form, this can be compactly written as

q(p) + (Θ∆(p))(p − c) = 0, (4.14)

where q(p), p and c are a Jx1 vectors of demands, prices and marginal costs respec-tively, ∆(p) is JxJ Jacobian of first derivatives, Θ is the JxJ ownership matrix and is elementwise multiplication.

This specification does not assume any form of collusive conduct among pro-ducers. Although it may be a strong assumption, there is no reason to be ex-ante suspicious about the firms’ conducts for two reasons. Firstly, the Czech competi-tion authority, UOHS, has never initiated any type of investigacompeti-tion in the sector in the area of horizontal or vertical agreements. Secondly, it is unlikely that potential members of a sustainable cartel would file a complaint to the Competition Authority in an ongoing public procurement with the accusation of predatory pricing, as is currently the case between two largest companies in the sector - Skoda auto (part of Volkswagen) and Hyundai (iDnes.cz, 2017). Moreover, in the context of the merger simulation, the assumption of absent collusion can be seen as an effort for a con-servative assessment of the merger effects. Under collusion, simulated price effects would be higher because companies do not compete so harshly as they partly or fully take into consideration competitors’ profits.

Given the identification of demand, substitution patterns and the oligopoly model of competition, first-order conditions allow to retrieve marginal costs from the

(32)

4. Methodology 25

Bj¨ornerstedt et al. (2014) in detail. The resulting equation yields:

c = p + (Θ∆(p))−1q(p). (4.15)

Furthermore, by inverting equation (4.14) and isolating the price, the merger effects can be directly predicted by computing the post merger price equilibrium. In the case of the current merger assessment, the equilibrium is dependent on the

changes in the structure of the ownership matrix from Θpre _{to Θ}post_{, where Θ}post

(33)

Chapter 5 Results

5.1 Demand

Table A.5, depicted in the appendix, reports the estimates of basic non-instrumented OLS nested logit. These outputs are provided merely for illustrative purposes as they are generally biased.

Table 5.1 shows six regression equations in which the endogeneous variables are instrumented by the instruments specified in section 4.2. There are three pairs of regressions and each pair corresponds to the specification in table A.5 (i.e. (1) in A.5 has two counterparts (1a) and (1b) etc). The left column in the pair uses two-stage least square estimation whereas the equations in the right column are estimated using the generalized method of moments. All models use standard errors robust to heteroskedasticity.

The signs of all coefficients have expected and consistent direction in all the models but the magnitude varies. The significance of the estimates is satisfactory as the significance level reaches less than 0.1% for most coefficients.

The effects of individual car characteristics on the mean valuation correspond to expectations in most cases. The performance characteristics, namely the cylinder number and the horsepower, have both positive effect on the valuation. Thus, the stronger the car is, the higher the valuation to consumers. Magnitudes seem also realistic. Cylinder number accounts mostly to 3,4 or 6,8 in luxury cars segment, while horsepower reaches up to lower hundreds. One additional cylinder should thus increase the mean valuation for a car more than one additional horsepower. The displacement coefficient is negative and significant in models (2a)-(3b) although it is also directly linked to the power of engine. The reason for this result is specific to the Czech market. In the Czech Republic, every car owner is annually charged an insurance premium for every car in possession. The amount is derived form the displacement of an engine and the relationship is positively related - cars with higher displacement are charged higher insurance rate. The variable displacement thus essentially reflects the long-term cost of ownership and should have negative effect on the mean valuation.

(34)

5. Results 27

Table 5.1: NL, instumented. Price variable princ and market size 3%

of households. Dependent variable ln(sj) − ln(s0).

(1a) (1b) (2a) (2b) (3a) (3b)

princ -1.94432∗∗∗ _-2.01944∗∗∗ _-1.46112∗∗∗ _-1.41798∗∗∗ _-0.39287 _-0.60915∗ (0.17659) (0.16719) (0.10469) (0.09741) (0.26676) (0.25863) sjg 0.08070 0.23377∗∗∗ 0.64350∗∗∗ 0.65790∗∗∗ 0.40753∗∗∗ 0.45503∗∗∗ (0.04922) (0.04591) (0.03065) (0.02800) (0.04997) (0.04780) cyl 0.94489∗∗∗ 0.88061∗∗∗ 0.54926∗∗∗ 0.54133∗∗∗ 0.13134∗ 0.15913∗∗ (0.05831) (0.05718) (0.03701) (0.03556) (0.06179) (0.06086) disp -0.00012 -0.00007 -0.00032∗∗∗ _-0.00046∗∗∗ _-0.00067∗∗∗ _-0.00065∗∗∗ (0.00012) (0.00011) (0.00008) (0.00007) (0.00011) (0.00011) hp 0.02824∗∗∗ _0.02908∗∗∗ _0.02522∗∗∗ _0.02259∗∗∗ _0.00164 _0.00457 (0.00348) (0.00333) (0.00214) (0.00195) (0.00288) (0.00280) weight -0.00168∗∗∗ -0.00097∗∗∗ -0.00075∗∗∗ -0.00026+ _-0.00053∗ _-0.00047∗ (0.00026) (0.00024) (0.00016) (0.00014) (0.00021) (0.00021) length 0.00207∗∗∗ 0.00239∗∗∗ 0.00147∗∗∗ 0.00169∗∗∗ 0.00035 0.00045 (0.00016) (0.00015) (0.00010) (0.00009) (0.00031) (0.00031) width 0.00301∗∗∗ _0.00251∗∗∗ _0.00270∗∗∗ _0.00265∗∗∗ _0.00276∗∗∗ _0.00231∗∗∗ (0.00062) (0.00060) (0.00038) (0.00036) (0.00069) (0.00069) height 0.00305∗∗∗ 0.00247∗∗∗ 0.00085∗∗∗ 0.00040∗ -0.00137+ _-0.00060 (0.00035) (0.00034) (0.00021) (0.00020) (0.00074) (0.00072) fuel -0.39912∗∗∗ -0.39579∗∗∗ -0.15161∗∗∗ -0.13599∗∗∗ -0.03762 -0.07027∗∗ (0.03232) (0.03081) (0.02151) (0.01920) (0.02483) (0.02429) accel 0.19988∗∗∗ 0.13566∗∗∗ 0.11537∗∗∗ 0.05645∗∗∗ -0.09683∗∗∗ -0.08206∗∗∗ (0.02933) (0.02780) (0.01816) (0.01567) (0.02005) (0.01975) speed -0.02054∗∗∗ _-0.03410∗∗∗ _-0.01850∗∗∗ _-0.02534∗∗∗ _0.00307 _0.00252 (0.00439) (0.00411) (0.00260) (0.00228) (0.00329) (0.00324) luggage -0.00126∗∗∗ -0.00207∗∗∗ -0.00118∗∗∗ -0.00143∗∗∗ -0.00035 -0.00011 (0.00026) (0.00024) (0.00013) (0.00012) (0.00029) (0.00027) Constant -20.49569∗∗∗ -17.02217∗∗∗ -13.38899∗∗∗ -11.72430∗∗∗ (1.48045) (1.39997) (0.92914) (0.82091)

Exclusive no no yes yes no no

Car FE no no no no yes yes

Time FE yes yes yes yes yes yes

Brand FE yes yes yes yes yes yes

Errors Robust Robust Robust Robust Robust Robust

Method iv gmm2s iv gmm2s iv gmm2s

Endogen. test 349.59328 349.59328 309.41009 309.41009 424.57042 424.57042

Weak ID test 18.73594 18.73594 18.28201 18.28201 2.62930 2.62930

RMSE 1.31447 1.23345 0.73468 0.73264 0.41798 0.39542

(35)

5. Results 28

The characteristics related to the size of a car have all positive effects on the valuation which seems plausible as larger cars are generally more comfortable and safer. In contract, weight contributes to the utility negatively.

Another important determinant of the mean utility is the inverse measure of fuel efficiency - the fuel consumption. The coefficient is negative, significant and of a large magnitude. The fact that consumers do not derive utility from less efficient cars is fully expected as fuel efficiency is directly linked to the costs of operation and

also to the environmental issues that are of notable relevance for many customers1_.

Finally, price has a negative effect on consumer mean valuation which is a result congruent with economic theory and expectations. The magnitude of the effect differs though. From a highly significant value of about -2 in specification (1), values close to zero occur in specification (3). The nesting parameter σ, determining the degree of substitutability of products within particular segment, has a low standard error and lies between zero and one in all the cases. Consistent findings of the magnitude of the parameter for the European car market are around 0.55-0.65 (Verboven, 1996; Brenkers & Verboven, 2006; Verboven & Grigolon, 2014). The nesting parameter σ seems thus to be estimated precisely in the case of models (2a) and (2b).

The variation in the magnitude of the most relevant coefficients on price and seg-mentation is directly connected to a critical difference between models (1a,b), (3a,b) and (2a,b). The models (2a,b) comprise a variable exclusive which is a dummy equal to one if car j belongs to the luxury or sport segment and zero otherwise. The results indicate that after being included in the model, the variable exclusive is statistically highly significant and helps to increase both the value and the significance of σ. Hence, by controlling for the fact that a particular car belongs to one of the two highest segments, cars in all the groups become more substitutable. This result is driven by the fact that exclusive cars posses many features that non-exclusive cars do not have and that should be controlled for specifically. One of the prominent exam-ples is the average price of both the sport cars and the luxury cars which exceeds the average in the other segments by a large amount (as can be seen in table A.2). The omission to explicitly control for this particular aspect of data would result in a bias in the aggregate results. Another treatment that yields similar results to (2a) and (2b) would be to control for high segment cars by including the square of prices to allow for non-linear price effects. This approach would, however, invalidate further analysis as the expressions for the elasticities would not be applicable.

Regarding the statistical inference, there are several common features across the

1_{The attempt to control the environmental concerns directly was done by incorporation of a} variable for emissions per 100 kilometres. However, this variable was almost collinear with the fuel consumption and was not significant in any of the regressions. As a consequence, it was dropped from the estimation.

(36)

5. Results 29

instrumented models. First of all, Kleibergen-Papp LM statistic suggests that the hypothesis of underidentitifacation of models (1a)-(2b) can be rejected. The rejection is less strong in the case of (3a) and (3b) as the regressions incorporate fixed affects for all the car models, which substantially increases the number of estimated parameters. Furthermore, the endogeneity tests based on Sargan-Hansen statistic provide strong evidence to claim that a severe endogeneity is present and instrumental treat-ment is necessary. All instrutreat-ments as specified in Section 4.2 were employed first. However, the test for redundancy of the instruments suggested that the instruments based on sums of horsepower, emissions and maximum speed of products produced by the same firm were redundant. Therefore, they were excluded from the list of instru-ments in the reported regressions. The first stage regressions for both endogeneous variables provide evidence that instruments are strongly relevant as the respective F-statistics allow to comfortably reject the hypothesis of the absent explanatory power of the instruments. The results of the first stage estimation are given in table A.7 in the appendix. The weak identification test based on Kleibergen-Paap Wald RK F statistic suggests different conclusions across models. Based on Stock and Yogo critical values, the maximum IV bias of models (1a)-(2b) is about 4-6%. The panel models perform significantly poorer as the bias there exceeds 30%. Utilization of either of models (3) seems thus infeasible.

When deciding which model to employ for further analysis, several factors are taken into consideration. Firstly, both coefficients of the primary interest, σ and α, should be significant and their magnitude should correspond to the findings in the

relevant literature. Moreover, as R2 cannot be used as a reliable indicator in the

IV environment, root mean squared error should serve as a proxy for the “goodness of fit” of the model. Both criteria are best satisfied for models (2a) and (2b). The results of GMM estimation barely differ from IV and therefore there is no reason to proceed with more complicated GMM structure. For simplicity, the model (2a) is used as a basis for computation of substitution patterns and further simulation.

Section 4.2 discussed the issues connected to the definition of a potential market. The results reported in the previous tables used the size of the market equal to 3% of

economic households in every quarter2_{. Table A.6 in Appendix A gives an overview}

of several robustness checks that work with different definition of the market size. The estimates are robust and barely respond to the changes in size. Even though the potential size is only assumed, these results indicate that such a treatment should not be a matter of concern.

Although only models with one-level nesting were presented in the section 4.2 and in the above results, it is noteworthy that two-level nesting is commonly used

The car market in the Czech Republic : an empirical analysis and merger simulation

University of Amsterdam

Faculty of Economics and Business