In this paper we use standard merger simulation techniques to compare the effects of three hypothetical mergers within the German automobile market. The simulation is conducted in three steps. Firstly, by applying a two-level Nested Logit model we estimate the demand for and then present stylised facts on the purchasing behaviour of German consumers. Secondly, conduct is modelled as Bertrand-type to simulate the pricing behaviour of manufacturers. In the third step, we adjust the market structure according to the respective merger cases and re-estimate supply and demand to simulate price effects. Our results suggest that welfare changes are only significant if German producers are involved. Furthermore, we find extensive differences across the market segments. Whilst mergers have virtually no consequences for the pricing of subcompact vehicles, we predict a considerably large effect within other car segments, such as luxury cars.
The rise and fall of one of the largest economies worldwide crucially depends on the development of the automobile industry. There are few export oriented nations, other than Germany who are dependent on and therefore at the mercy of one single industry. This dependency makes the German car market extraordinarily appealing for the simulation of merger effects. In general, merger activity in Germany is fairly high, in 2007, for instance, 2242 merger notifications were registered by the domestic antitrust authority, and therefore a scenario involving mergers in this industry is realistic.
In this paper, we will explore the effects of a merger within the automobile sector, as to date authorities do not take full advantage of the powerful techniques developed to anticipate the potential consequences from mergers. We will establish a model, which will simulate the German market and compare the predictions across different merger cases.
The paper is structured as follows: The following section reviews relevant literature and the characteristics of the German automobile market. Subsequently, we describe the methodology used for our merger simulation. Section four summarizes the estimated results of three different hypothetical mergers. The concluding section compares the results of the three cases and highlights possibilities for future research.
II. LITERATURE REVIEW AND OVERVIEW
Merger Simulation Literature
Generally, the underlying idea of using a merger simulation model is to forecast the effects on pricing in a three-step procedure. Firstly, a demand model is estimated, the price setting behaviour of the firm is then modelled, and following this, both models are re-estimated utilising the market structure on completion of the merger.
Sophisticated techniques have developed for the analysis of horizontal mergers on homogeneous good markets. Researchers mostly agree on the model choice and even dynamics may be modelled using existing specifications1.
Unfortunately, when considering the literature on heterogeneous products, such powerful models have not yet been established. It is not so much the non-existence of models, but rather the lack of large horizontal mergers within differentiated product industries, due to the rejection by authorities, which make it difficult to evaluate the power of the proposed models.
Extensive research literature applies different model types to horizontal mergers. For example, Werden and Froeb (1994) for the telecommunications sector, Hausman et al. (1994) and Slade (2009) for the brewing industry. However, these papers cannot shed light on the efficiency debate, since they do not focus on evaluating the models when comparing predicted and actual price changes.
Nevertheless, certain comparisons of simulations and actual merger price effects may be found within the current literature. Peters (2006) took advantage of the existence of rich datasets for the airline industry and simulated six actual mergers. Comparing two different models of demand, i.e. GEV and Nested Logit, she found that short-term price changes were anticipated to some extent, however not predicted completely by the simulations. Two mergers within the ready-to-eat cereal industry were explored in Nevo´s (2000) analysis, where he estimated a random-coefficients demand model. In this instance, simulations fared well when evaluated against observed post-merger prices. Finally, applying the Nested Logit model, Pham and Prentice (2010) simulated a merger on the Australian cigarette market to find that the predictions of their simulation were inaccurate, however they still advocate the use of the Nested Logit within alternative industries, due to the reasonability of their demand estimates.
Within the relevant literature, there is large consensus around the use of Bertrand-type pricing models on the supply side. However, as can be seen from the papers mentioned above, researchers disagree on a specific model for the demand side, and therefore make use of a variety of specifications. Yet the performance of a merger simulation crucially depends on the efficiency of demand estimates, which makes the model choice pivotal within the analysis. Therefore, we will review literature concerning the development of different demand models in the following subsection.
Demand Estimation Literature
Whilst demand estimation for homogenous goods is fairly common, models for heterogeneous goods have not yet reached comparable efficiency, despite receiving growing attention from researchers. Issues with estimation arise due to potential market power created by the product differentiation. Certain characteristics may lead consumers to stick with products despite alternatives being available at a lower price. There are many ways to model demand in these circumstances, but an overview of the variety of models is beyond the scope of this paper, and therefore, the focus will be on discrete choice approaches, since they are the most prominent.
Each type of differentiated product demand model faces two key problems. The first stems from the number of parameters required to be estimated. Since demand for each product is a function of all prices, the number of parameters grows exponentially with products on the market. Secondly, idiosyncratic consumer preferences are pivotal for differentiated product demand, but are not observed, and therefore cannot be included in the model. To overcome these problems, models impose strong assumptions, which often reduce the reliability of policy conclusions. Within the discrete choice literature this is managed by restricting the parameter space.
Initially models of discrete choice were used in the field of psychology, and it was McFadden (1974) who introduced them into the field of economics. Whilst the basic discrete choice model, i.e. Logit or Conditional model, is pretty straightforward to estimate its assumptions may lead to severe inefficiency in many applications. Particularly, the independence of irrelevant alternatives assumption (IIA), i.e. relative probabilities of existing alternatives unaffected by adding new options, has been viewed critically ever since.
The Nested Logit model, which was developed by Ben-Akiva (1973) and McFadden (1978) maintains the tractability of the simple Logit, but adds more flexibility. In this model, the IIA is relaxed for products belonging to identical nests, which have to be specified by the econometrician prior to the estimation. It is more flexible, because preference correlation may be modelled within the nests. However, preferences for products within different nests are still subject to the IIA.
Further developments within the discrete choice literature brought the most flexible model in terms of preference correlation, unfortunately however, this comes at the cost of high computational demands. Berry, Levinsohn and Pakes (BLP, 1995) developed the Random Coefficients Model, in which substitution among products is a function of characteristics and independent of nesting specification. However, several recent papers do criticise the BLP model, claiming that it cannot efficiently estimate demand due to problems in specification2.
Work on the Automobile Market
Within the context of automobile demand, discrete choice models are a good fit. It is reasonable to assume that the buying decision within the car market is of a discrete nature, in that consumers will choose to purchase one car or nothing at all. Ironically, BLP first estimated their model on the automobile market, which arguably is suitable for the less complicated Nested Logit. Deciding on which model to use for the demand on the German automobile market, it is helpful to review previous research, focussing on the demand estimation for cars.
Conditional Logit models have not reached the level of acceptance as the other two models and are therefore rarely used. It is unrealistic to assume that all cars on the market are strictly differentiated and elasticities are independent of alternatives. Guerzoni and Soellner (2009) and Moral and Jaumandreu (2007) provide two examples of the application of the conditional Logit model to the car market, however their results should be treated with caution due to the mentioned restrictiveness of the model assumptions.
Within the car market, the candidates for the demand side model are reduced to Nested Logit and BLP. Certain analyses explicitly focus on the question which of the two is more efficient for vehicle demand. Grigolon and Verboven (2011) examined a dataset for the European automobile market, and found quite different results concerning estimated substitution patterns. However, they concluded that both models yield sufficient estimates to draw correct implications from policy counterfactuals, such as merger simulations. Wojcik (2000) advocates the Nested Logit model in her comparison on the US car market, however, her results are often criticised due to problems in specification.
Research concentrating solely on either of the models, found advantages as well as shortcomings in both approaches. Papers such as BLP (1995) on the one hand and Verboven (1996) or Fershtman and Gandal (1998) using Nested Logit on the other hand, demonstrate that estimations yield reasonable results for both models, but they also encountered certain problems.
Following a detailed survey of the literature, one may conclude that neither of the two alternatives prevailed. For this reason, we take a indifferent stance with the models, from an efficiency point of view. However, the great virtue of the Nested Logit lies within its computational simplicity, which supported our decision in favour of this particular model. Tractability remains a striking argument within this context, therefore eliminating the BLP model from our analysis even if it could possibly offer slightly more efficient results.
German Car Market
Since entrepreneurs such as Gottlieb Daimler or Ferdinand Porsche, Germany has become an international leader for the production of passenger vehicles. Today, the automotive sector is the driving force of German industry. Only the USA and China are ahead of Germany in terms of production. Domestic companies employ around 700.000 people with an annual revenue of above 300 billion Euros.
Almost half of all cars sold in Germany are subcompact or compact vehicles. The smallest market segment comprises luxury cars, which are exclusively manufactured by domestic producers.
Within the German market, there are sixteen producers with significant market shares, five of which offer domestic brands, to account for above 60 percent of total car sales. Due to this strong concentration of sales among few producers we can describe the market as an oligopoly.
The previously mentioned similar performance of Nested Logit and Random Coefficients models is primarily observed within automobile markets. Due to its specifics the decision process within the car market may be fully explained by the Nested Logit model, despite receiving criticism and the fact that its efficiency has also been questioned in applications to alternative markets.
Critiques often claim that substitution patterns may only be modelled sophisticatedly within a nest, but not for the whole market, the model is also error-prone since researchers are required to decide a priori on the less trivial specification of nests. However, due to the division of specific segments containing car models with similar characteristics and more importantly correlated prices, the automobile market is suited to the structure of a Nested Logit model.
Table 1: Nesting Structure
illustration not visible in this excerpt
We specify the structure of the two-level Nested Logit as depicted in table 1. Car segments define the groups at the first level, and countries of origin further separate the segments into subgroups for foreign and domestic vehicles. Particularly for Germany, due to their large domestic car market, it is a crucial decision for consumers, whether to buy a domestic or a foreign car, which is why the second nesting level is defined3. Our specification yields a model in which preferences are strongly correlated across products within the same sub-nest, i.e. cars in the same segment with an identical country of origin. This correlation is diminished if cars within a given segment are manufactured by producers of different origin. Finally, no correlation of taste between cars in different segments is included, which is acceptable, since various automobile datasets highlight that the correlation of cars in different segments is at best very weak. This structure of correlation patterns therefore seems plausible for the car market, and the Random Coefficients model may be avoided.
The different models within the discrete choice literature rely on the same basic assumptions. Consumer i is assumed to confer utility from his decision for product j on all markets n as specified by the following equation:
illustration not visible in this excerpt
This general utility function may be considered to be divided into two parts: one depending on common product attributes and the other varying with consumer’s characteristics. The former contains information of both the observed and unobserved characteristics 4 [illustration not visible in this excerpt], as well as pricing [illustration not visible in this excerpt]. These variables have identical effects on utility across all consumers and are therefore denoted as product mean utility:
illustration not visible in this excerpt
The second part of equation (1) specifies idiosyncrasies within consumer taste. Differences between the discrete choice models arise in the definition of the[illustration not visible in this excerpt] term. In the Nested Logit this taste parameter is separated into consumers´ individual preferences for specific products [illustration not visible in this excerpt] but also for nests, i.e. car segments [illustration not visible in this excerpt], and subnests, i.e. country of origin [illustration not visible in this excerpt]
illustration not visible in this excerpt
Cardell (1997) showed, that all unobservables have an extreme value distribution5.
Both [illustration not visible in this excerpt] parameters play a crucial role within this setting, as they indicate the strength of the correlation of taste across products within the same sub-nest ([illustration not visible in this excerpt] and nest ([illustration not visible in this excerpt] If they are equal to one, products within the same subgroup become perfect substitutes, which then makes product taste irrelevant, and therefore utility only dependent on the consumers´ individual taste for each respective car segment [illustration not visible in this excerpt].
1 see for example Gowrisankaran (1999)
2 see for instance: Knittel and Metaxoglou (2008) and Dubé, Fox and Su (2011)
3 The country of origin is not necessarily the country where the production site is located but rather where the company is headquartered.
4 This regressor is a matrix containing all relevant characteristics which is why the order of regressor and coefficient is reversed in equation (1).
5 Formally proved for the one-level nested logit but the same arguments apply for the two-level version.