Predicting cartels in Dutch industries : a pro-active approach

(1)

Predicting cartels in Dutch industries

A pro-active approach

M.Sc. Economics Thesis Myrthe Michelle Baars University of Amsterdam

15 July, 2018

Abstract

Competition authorities should develop pro-active methods in determining in which markets collusion takes place and implement new and different methods in order to ensure effectiveness of enforcement. This thesis discusses two economic cartel detection instruments: the Competition Index (CI) and a Probit regression. These pro-active instruments aim to detect industries that are prone to anticompetitive behaviour. Nine indicators are used as a basis for this screening approach. The analysis is applied on the Dutch economy, which is divided into 222 industries using the Standaard Bedrijfs Indeling (SBI) 2008 version 2018 as defined by Statistics Netherlands (CBS). The second part of the paper presents the results of both methods for the Dutch economy. Finally, the outcomes and predictive power of the two methods will be compared with each other. We conclude that the CI and the Probit regression give different results and there is no positive and significant relationship between the CI and the presence of a cartel.

Keywords: collusion, detection, cartels, pro-active method, screening, probit JEL Classification: L41, K21, K42, C70

Student name: Myrthe Michelle Baars Phone number: +31624375241

Supervisor UvA: Dr. A. M. Onderstal Supervisor ACM: Dr. J. van Sinderen Studentnumber: 10648267

(2)

2

Statement of originality

This document is written by Myrthe Baars who declares to take full responsibility 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 in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

1. Introduction

A cartel is a group of companies which join together making agreements on prices, limiting production or sharing markets or customers. Instead of competing with each other, cartel members depend on the multilateral agreements which decreases their incentive to provide new or better products and services at competitive prices. Subsequently, the consumers or other business end up paying more for less quality, resulting in lower overall welfare in the economy. For this reason, an important objective of competition authorities is to discover and punish collusive behaviour (Motta, 2004). Because cartels are illegal, they are generally highly secretive and evidence of their existence is hard to find. Up until today, the leniency program is the most effective cartel detection measure as almost 60% of the cartel infringements are discovered through leniency. This program provides (partial) amnesty for a cartel member who reports a cartel and therefore gives incentives to whistle-blowing. Many competition authorities rely on the leniency programs because of their high detection rate and the evidence they provide on the existence and functioning of the cartel (Grout and Sonderegger, 2005). However, this reactive policy relies on complaint of consumers and producers which makes the competition authorities very dependent on these signals to detect collusion.

Instead of only reacting to signals from consumers and producers, it would be wise to develop a method that is independent of such signals. Therefore, competition authorities should develop a pro-active method in determining in which markets collusion takes place and implement new and different methods in order to ensure effectiveness of enforcement. In the future it might be the case that the complaints from the consumers and producers or the use of the leniency programs declines1_{, making the detection of collusion for the authorities harder. Another reason for a new}

method is that when firms become more and more familiar with competition enforcement practices they are more likely to anticipate competition authorities strategies. Furthermore, one can state that the cartels that have been discovered using the signals or the leniency programs could be the ‘failed’ cartels which are badly organized; the well-designed cartels go undetected (Stephan, 2009). In short, leniency is less likely to be successful in identifying profitable and effective cartels. This thesis will construct two pro-active methods which may be used complementary to competition authorities’ existing detection instruments.

An empirical screen is an example of a pro-active cartel detection method. The ability to flag unlawful behaviour through economic and statistical analyses is commonly known as screening. If potential collusive firms know that their market data is being monitored, they know there is an enhanced probability that they will get caught, which may discourage the formation of the cartel.

1 _{Henk Don, Member of the Board of ACM, stated in an interview on 14 November 2012 that in some jurisdictions in the}

Netherlands there is already a concern that the number of leniency applications are getting fewer. Systematic research is required to assess the true effectiveness of leniency programs.

(6)

6 Combining reactive with proactive cartel detection will enhance existing leniency programs and makes anti-cartel policy more effective around the world (Abrantes-Metz, 2013).

The Competition Index, developed by the Netherlands Competition Authority (ACM), is a pro-active approach known for its cartel detection and deterrence objective. The Competition Index functions as a data screening device in order to detect industries that are prone to anticompetitive behaviour (Petit, 2012). Currently, the Index is used as a reference source: a combination of sources is used to detect cartels and the Competition Index is one of the sources used. The Competition Index consists of nine indicators that all relate to competition as can be seen in Table 1.

Table 1 - Indicators used in the Competition Index (Petit, 2012)

INDICATORS

1 Number of trade associations

2 Product prices in the Netherlands versus European Union-averages 3 Herfindahl-Hirschman index (HHI)

4 Number of firms 5 Import rate 6 Market growth 7 Churn rate 8 Survival rate

9 R&D as a percentage of sales

The indicators in Table 1 are considered as indications for the likelihood of anticompetitive behaviour. Despite the fact that past evidence of cartel detection provides some indication of where cartels may exist, it would be helpful to more fully understand the indicators that lead to cartel formation in the first place. This information can then be used to identify the location of the cartels that have not yet been identified. Many factors will depend on specifics of the individuals involved and unique features of the case, but there may be common economic indicators that can help inform the process of deciding where to allocate effort in the cartel detection process. Petit (2012) focused her investigation on data that was available then and used data from the year 2008. However, in April this year a new database from the Statistics Netherlands (CBS) was released with more recent data and more detailed variables for each industry classification.

The above introduction brings me to the following research question: “Which economic indicators predict the likelihood of anticompetitive behaviour and how can we implement this into a pro-active method which identifies Dutch industries valuable of closer inspection?”

(7)

7 This thesis uses two approaches. The first approach constructs the Competition Index, composed of nine indicators that relate to competition, to compose a ranking of industries in terms of likelihood of anticompetitive behaviour. The second approach uses existing evidence of discovered cartel cases and economic data to identify factors that are relevant to the identification of cartels and then uses the economic model to provide predication of the probability of cartels within each of a large number of industry classifications. Next, the outcome of the two approaches, industries with a high likelihood of anticompetitive behaviour, will be compared with each other to verify whether the two approaches point out the same industries valuable of further investigation. Moreover, we will also make a comparison between the predictive powers of the two approaches.

This thesis is organized as follows. The next chapter provides a detailed review of existing literature on the stages of cartel detection and the proactive cartel detection methods. Chapter 3 applies the Competition Index to the Dutch economy. Chapter 4 uses data from discovered cartel cases to measure the probability of a cartel in the Dutch industries. Chapter 5 provides the results of the two methods and lists the top 20 industries prone to collusive behaviour. Moreover, Chapter 5 compares the results and the predictive powers of the two approaches. And last, Chapter 6 will summarize the conclusions of the analysis and gives the limitations of the thesis together with the prospects for further research.

(8)

8

2. Literature review

The following chapter reviews relevant literature on the detection of collusion. First, we will address the economics of collusion and the sustainability of a collusive agreement. Second, the stages of cartel detection will be discussed. Third, the current cartel detection methods will be presented. Fourth, we will discuss some characteristics and indicators that may point to competition or collusion. And last, we will discuss the empirical findings of the current cartel screening methods.

2.1 The Economics of Collusion

From an economic point of view, collusion describes a situation where the prices of the firms are higher than some competitive benchmark. The definition of collusion according to the Dutch competition law is the following: “all agreements between undertakings, decisions by associations of undertakings and concerted practices which may affect trade between Member States and which have as their object or effect the prevention, restriction or distortion of competition within the common market”2_.

Restricting or distortion of competition can generally be accomplished in three ways: increasing the price, limiting output and the allocation of market shares (Ivaldi et al., 2003). When firms make an explicit agreement, they are able to set a higher price than under competition resulting in a lower quantity of products supplied (demand declined because of the higher price) and larger profits for the cartel. Instead of setting the price to a fixed level, the firms can also agree on limiting the output. When firms agree to limit the overall output this increases the price (the good becomes more scarce in the market) and eventually increases the profits for the cartel. Another way to increase cartel profits is to allocate parts of the market to the cartel members, for example by dividing the market geographically. The cartel members then become monopolists in its market and can set the price at the level it prefers. To conclude, the primary objective of a cartel is to increase profits, but this does not necessarily end up in price changes.

Collusive agreements like the one described above can be divided into two categories: explicit collusion and tacit collusion. An explicit collusive agreement occurs when the firms involved used direct communication when reaching the agreement. When firms are able to restrict competition without the use of direct communication, we call this tacit collusion. Firms do not necessarily have to talk to each other for a collusive outcome to be sustainable. Tacit collusion can arise if detection of a deviation is fast and if market punishments of deviations are likely and credible. However, if firms are unable to communicate with each other, they can make mistakes and select a price which is not jointly optimal for the firms. It might be very costly for firms to use market signals to

(9)

9 coordinate on a different price. Instead, under explicit collusion firms can talk to each other and coordinate on their jointly preferred equilibrium without using costly market signals (Motta, 2003).

For collusion to occur two criteria have to be met. First, there must be a possibility to detect deviations from a collusive action. Second, there must be a credible punishment which follows from a deviation (Motta, 2003). Punishment must be sufficiently likely and costly to outweigh the short-term benefits from “cheating” on the collusive path. These short-short-term benefits, as well as the magnitude and likelihood of punishment, depend in turn on the characteristics of the industry. For the punishment to be effective, it must imply a significant profit loss for the deviating firm compared with the profit that it would have obtained from sticking to the collusive agreement. A simple form of punishment consists of the disruption of the collusive agreement and the return of “normal” competition and profits (Ivaldi et al., 2003). Firms will anticipate the situation in which a collusive price will be maintained as long as none of them deviates, but if one of the firms tries to obtain short-term profits by undercutting the price, the collusive agreement ends.

Economic analysis allows a better understanding of the nature of punishment mechanisms and their common features. We will use a base case situation where firms sell a homogenous product with the same variable cost to illustrate the effects and the factors that will affect collusion. Consider a simple symmetric duopoly situation where two firms, i and j can choose to either compete or collude. Table 2 shows the situation in a typical two firm matrix game where the first value in the brackets represents the payoff of firm i and the second the payoff of firm j.

Table 2 - General 2x2 prisoner’s dilemma game in normal form

Collude Compete

Collude (b, b) (d, a)

Compete (a, d) (c, c)

The game is a prisoners dilemma when the profits are ranked in the following way: a > b > c > d and 2b > a + d > 2c. In a simple one period or one shot game setting, collusion does not constitute with a Nash equilibrium and therefore rational firms will never collude under perfect information. Two firms will only collude if the expected profits of collusion are higher than the expected profits of deviation. By assumption, the incentive constraint is violated in the one shot game: the deviation profit in the one shot case, a, is larger than b representing the profits under collusion. Hence, firms will fall back to the unilateral output or price levels and the competitive equilibrium emerges. When the game is played repeatedly, i.e. periods are indexed by 𝑡 ∈ {0,1,2, … , 𝑇}, collusion still won’t occur. So if the game has exactly one Nash equilibrium, then in each Nash equilibrium of the repeated game the players select their equilibrium strategies (von Neumann & Morgenstern, 1944).

(10)

10 One possibility is to adjust the one shot game into a supergame by infinitely repeating the stage game as depicted in Table 2 (formally 𝑇 → ∞). Consider an industry where n firms play an infinite horizon game. Call 𝜋_./_{and 𝑉}

./ respectively the current profits and the present discounted value of profits that firm i receives if it chooses a certain collusive agreement, given that the other firms also collude. If firm i deviates when all other firms stick to the collusive agreement it receives 𝜋_.1_, and 𝑉_.2 is the present discounted value of firm i’s profits in the punishment phase, that is in all periods that follow the deviation period. Denote with 𝛿 ∈ (0,1) the discount factor, which is identical for all firms in the industry. If firms are risk free and have access to a credit market with interest rate r, the discount factor is equal to 𝛿 =₇₈₉7 . Collusion can arise only if each firm will prefer to play the collusive action rather than to deviate from it (and be punished for ever). Therefore, it must be that case that the Incentive Constraint (IC’s) holds and that for each firm the following condition holds:

𝜋./+ 𝛿𝑉./ ≥ 𝜋.1+ 𝛿𝑉. 2

𝑖 = 1, … , 𝑛.

The lower the deviation profit the firm makes relative to the collusive profit and the lower the profit in the punishment phase, the more likely that collusion will be sustained. The intuition behind this is that the harder the punishment, the stronger the deterrence from cheating on the collusive agreement. The above incentive constraint can be rewritten as:

𝜋.1− 𝜋./ ≤ 𝛿A𝑉./− 𝑉.

2_{B 𝑖 = 1, … , 𝑛,}

which states that the gains from deviating obtained today must be lower than the losses from deviation from the collusive agreement incurred from tomorrow onwards. This condition must be satisfied for all the firms in the industry otherwise one or more deviations will occur and collusion cannot be sustained. Another way to rewrite the same incentive constraint is:

𝛿 ≥𝜋.

1_{− 𝜋}

./ 𝑉./− 𝑉.

2≡ 𝛿.∗ 𝑖 = 1, … , 𝑛.

The condition for the sustainability of collusion is most often expressed in the above form. Collusion can only arise if the discount factor is large enough, that is, larger than the ‘critical discount factor’, 𝛿.∗. If a firms’ discount factor is higher than the threshold, any collusive price can be sustained. If instead the firms’ discount factor is lower than the critical threshold, no collusion is sustainable. This is very intuitive: if the discount factor is very low, firms do not give importance to what will happen in the future, and they will prefer to cheat so as to reap all the benefit today. Hence, then collusion will not arise.

(11)

11 Consider a standard duopoly case. Suppose that two firms produce the same homogenous good with the same marginal cost 𝑐. Price competition then would lead these firms to price at marginal cost (𝑝 = 𝑐), and there will be no supra-competitive profits for the two firms. Now, if these firms compete repeatedly they may be able to sustain a higher (collusive) price 𝑝/_{> 𝑐, both earning a higher} collusive profit of 𝜋/_{= (𝑝}/_{− 𝑐)𝐷(𝑝}/_{). Deviating from this price would trigger a price war and would} lead the firms to return back to the competitive price 𝑝 = 𝑐 (Friedman, 1971). By sticking to this tacit agreement each firm would earn profits in T periods with a multiplicative factor 𝛿J_{. (Ivaldi et} al., 2003). With profits of:

𝜋/ 2 + 𝛿 ∗ 𝜋/ 2 + 𝛿K∗ 𝜋/ 2 + ⋯ = 𝜋/ 2 (1 + 𝛿 + 𝛿K+ ⋯ ).

Collusion arises at equilibrium if no firm has an incentive to deviate from the behaviour indicated by the trigger strategies. If one firm deviates from the agreement by slightly undercutting the other, it captures the entire market and thus the entire collusive profit 𝜋/_{, but the resulting price war} will eliminate any future profit. Since the two firms are identical, we just need to consider the incentive constraint of one firm:

𝜋/

2 (1 + 𝛿 + 𝛿K+ ⋯ ) ≥ 𝜋M+ 𝛿 ∗ 0.

That is, if

𝛿 ≥ 𝛿∗_≡1 2.

In this example, firms are able to sustain a collusive agreement when the weight they put on future profits, measured by their discount factor, is above a certain threshold. The critical threshold for the discount factor, 𝛿∗_{, is equal to 1/2. Concluding, the critical thresholds 𝛿}∗_{tells us how ‘easy’ it is} to sustain collusion. The lower the critical threshold, the easier it is to sustain collusion. The determination of this critical threshold thus provides a natural way for assessing the scope for collusion. If we measure the impact of the industry characteristics on the likelihood of collusion, we look at how the industry characteristics would affect the critical threshold. A factor that facilitates collusion will decrease the critical threshold, while an industry characteristic that makes collusion more difficult will increase it (Ivaldi et al., 2003). In section 2.5 the main relevant industry characteristics will be reviewed in detail.

2.2 Stages of cartel detection

According to Harrington (2005) the detection of cartels should involve three stages: screening, verification and prosecution. Operating effectively in all three stages is crucial in disrupting

(12)

12 existing cartels and deterring new cartels from forming. The purpose of screening is to identify markets where collusive behaviour is suspected and to point out markets worthy of closer inspection. Verification involves trying to exclude competition as an explanation for observed behaviour and to provide economic evidence in support of collusive behaviour. This evidence can involve estimating a competitive benchmark and comparing the behaviour of suspected colluders or the estimating of both collusive and competitive models to see which better fits the data. The verification method uses more complex indicators to make sure whether collusion is really present in the specific market or not. Since the method is by definition more data- and time-intensive, it can only be applied to a limited number of industries. Therefore it is not practical to engage in such an exercise, except if there are already some suspicions e.g. some evidence that collusion may be present in an industry (Harrington, 2005). Although verification rules out any ‘false positives’ that have been selected in the screening stage, the results of verification do not provide enough legal proof. Additional proof such as contracts, correspondence among cartel members or invoices from hotels, where the cartel meeting took place is needed. Such evidence is typically collected when a potential case is referred to the European Commission or competition authority (Friederiszick & Maier-Rigaud, 2007). Surprise inspections are the most effective and sometimes the only means of obtaining the necessary evidence to sanction the cartel. The proposed screening process would use a mixed instruments methodology, combining both reactive and pro-active instruments, as indicated by Friederiszick & Maier-Rigaud (2007), Schinkel (2008) and Grout and Sonderegger (2005). Friederiszick and Maier-Rigaud (2007) furthermore state that there is no reason why more indirect evidence of cartel activity cannot create a sufficient degree of suspicion to justify an inspection by the competition authority. Additionally, European law does not contain any explicit rule on the exact level of suspicion the European Commission or competition authority needs to have for a decision ordering unannounced investigations to be lawful. The last stage, prosecution, involves the gathering of proof of the illegal behaviour that is sufficient to persuade the courts that there has been a violation of the law.

2.3 Methods of cartel detection

Cartel screens are classified in the literature into two different categories: structural and behavioural screens (Harrington, 2008). The purpose of structural screening is to identify markets where anticompetitive behaviour is more likely to appear. This method looks at the structure of the industry at hand ‘scoring’ the likelihood of collusion based on factors such as high market concentration, homogeneous products, excess capacity, stable demand, and other commonly used collusive markers. This method is useful to create an initial list of ‘suspect’ industries and to complement reactive detection measures. Structural indicators cannot be directly influenced by the market participants and the screening is relatively simple and easy to implement. Usually the data used is public and simple to collect. On the other hand, behavioural screening looks at markets’ and participants’ behaviour. This is translated into observable data to investigate if the observed

(13)

13 behaviour is more or less likely to have been produced under an explicit agreement. Behavioural screens use data on prices, bids, quotes, spreads, market shares, volumes and other data to identify patterns. This method requires data outside of the time of suspected collusion and the approach can test for a breakpoint in the data like for example an exit, merger or the formation of a trade association (Abrantes-Metz, 2013).

Besides the classification into behavioural or structural screens, there is another categorization of approaches in the cartel detection based on economic criteria, namely the top-down and bottom-up approaches. The top-down approach screens several industries in order to identify industries prone to collusive behaviour. While the top-down approach provides helpful insight, such as to pinpoint industries where competition authorities can focus their enforcement priorities, they are often affected by severe shortcomings. First, the level of aggregation is generally too high to identify specific antitrust markets and, in addition, industry classifications do not match antitrust relevant markets. Top down methods thereby do not allow identifying a specific market for cartel inspection, but can only indicate broad indicators. Second, an empirical analysis across various industries required well-defined screening measures. In principle, these screens can be re-engineered if the industry experts or economic consultants know how the general approach of these screens works. They can take the screens into account when designing a cartel to make sure the cartel is shielded against detection methods. Third, the relationship between economic indicators and the probability of anticompetitive behaviour is often not linear. And fourth, by relying on discovered cartel data in the derivation of indicators for detection of cartel activity in other industries, results are possibly affected by a measurement bias: the cartels detected in the past are unlikely to provide a representative sample for all active cartels (Friederiszick & Maier-Rigaud, 2007).

Contrary to the top-down approaches, the bottom-up approaches focuses on a particular sector or market. Bottom-up approaches do not rely on cross-sector data so they can adopt a more flexible set of criteria. This means that three of the four problems addressed above are solved by the bottom-up method. As said, because of the more flexible set of criteria, the industry experts of economic consultants are not able to discover a general screening-approach. This makes the potential cartels unable to shield against the detection methods as discussed above (Friederiszick & Maier-Rigaud, 2007). Furthermore, with a bottom-up approach the issue of non-linearity of individual indicators can be focused on by a more case-based approach. And last, this method is less affected by the selection bias as theoretical considerations can more easily be taken into account. Clearly, the bottom-up approach also has its shortcomings, those being the limited availability of public market data and the resource requirements for the implementation of such an approach (Friederiszick & Maier-Rigaud, 2007).

Abrantes-Metz (2013) stresses that when designing and implementing screens, there are two golden rules to keep in mind. These rules are obvious when stated, but may sometimes be forgotten.

(14)

14 First, “one size does not fit all”, and second, “if you put garbage in, you get garbage out.” Screens can be helpful and potentially very powerful, but these are econometric tools, with all the usual caveats (data limitations, no hard evidence, resource-intensive, etc.), and they may potentially be misused. The first golden rule states that a screen needs to be adjusted to the situation at hand. In turn, the second golden rule states that, as is always the case in empirical work, a screen is only as good as the choices of what is put into it. Even a cartel screen which is properly designed and implemented and based on reliable theory can still produce incorrect conclusions, just as is the case with any other statistical test. It may indicate that collusion may have existed where one did not (‘false positive’), or it may fail to flag collusion which did exit (‘false negative’) (Abrantes-Metz, 2013).

One example of the power of screens was performed by the Italian competition authority, as documented in Esposito and Ferrero (2006). The authors test the power of the variance screens for prices to detect previously known illegal conspiracies. They ask the question of whether a price variance screen could have identified collusion in two well-known Italian cartel cases involving gasoline and diesel on the one hand and in baby milk on the other. Furthermore, the authors ask whether the screen could have correctly identified who was involved and during which time period. The answer to both questions is “yes”: the screen would have correctly identified these two cartels even before the Italian competition authority did (Abrantes-Metz, 2013).

2.4 Indicators of Collusion

We will review 15 industry characteristics and indicators that the literature (Grout & Sonderegger (2005), Rhoades (1993), Motta (2004), Petit (2012), Ivaldi et al. (2003), NERA (2004), Symeonidis (2003), Brock and Scheinkman (1985)) has singled out as being relevant when assessing the sustainability of collusion within a market. It should be noted that these industry characteristics are neither necessary nor sufficient for collusion to be sustained, but simply affect the likelihood that collusion might be sustained.

1. In particular the literature suggests that a small number of firms in the market facilitates collusion for three reasons. Firstly, a high number of firms increases the probability that firms with different costs of production operate within the market, this decreases the likelihood of collusion. Secondly, a large number of participants make deviations from the collusive agreement harder to monitor. Thirdly, when the number of firms increases, each firm receives a lower share of the pie. This makes deviations more tempting and punishments less costly (Grout & Sonderegger, 2005).

2. However, even if the industry has a large number of firms, collusion can still arise if only a few firms are very large and a lot of firms are very small. This concentration is measured by the Herfindahl-Hirschman Index (HHI). The HHI index ranges between 1/n and 1,

(15)

15 reaching its lowest value when all firms in the industry are of equal size, and reaching unity in the case of monopoly (Rhoades, 1993).

3. The more symmetrical firms are, in the sense of cost structure, capacities, knowledge, etcetera, the more their incentives are aligned and thus makes it easier to form a cartel (Motta, 2004).

4. If entry barriers are low, high prices will attract new competitors in the future. These new competitors will decrease the gains from collusion, making punishment less costly to bear. Therefore, in markets where entry barriers are low, collusion is harder to sustain. Industries where entry is more likely should therefore ceteris paribus be associated with lower probability of collusion (Motta, 2004).

5. Frequent interactions facilitate collusion by shortening the time of reaction to deviations from the collusive agreement. Petit (2012) measures the frequency of interaction by the existence of trade organisations. It has been documented that trade associations are used as a cover for cartel meetings and, more to the point, trade associations have been created for that express purpose (Harrington, 2005). Petit (2012) assumes that even the existence of one single trade association makes a significant difference, compared to the absence of a trade association.

6. Collusion is easier when firms observe each other’s prices and quantities. Therefore market transparency has a positive effect on the likelihood of collusive behaviour (Ivaldi et al., 2003).

7. As stressed above, collusion is easier to sustain when short-term gains from a deviation are small compared with the cost of future punishment. For that reason, collusion is easier to sustain in growing markets, where the profits of today are small compared with tomorrow’s profits. Contrarily, in declining markets (where tomorrows profits will be small anyway) collusion is harder to sustain, because there is no possibility to convince firms from sticking to the agreement (Ivaldi et al., 2003).

8. As a consequence of the impact of growth and decline of the market, is that collusion is less sustainable in markets that are subject to demand fluctuations (Ivaldi et al., 2003). 9. A large inventory or excess capacity is necessary for cartels to be able to punish cartel

members when they deviate. The punishment requires the cartel member to lower their price, which will raise demand and therefore output has to be expanded temporarily. If the cheater would not have the extra capacity, the gain from deviation is small because it is not able to produce the extra demand generated by the deviation. Therefore the exact effect of extra capacity is ambiguous, since it makes cartels less stable. The absence of excess capacity would mean that no member could deviate, but that it cannot be punished either (Brock and Schienkman, 1985).

10. When firms are present on several markets they can more easily sustain collusion. First, multi-market contact increases the frequency of interaction between the firms which facilitates collusion as discussed in 5. and second of all, it may allow softening asymmetries

(16)

16 that arise in individual markets. Example given, one firm may have a competitive advantage in one market and its rival can have its own competitive advantage in the other market. Market-level analysis may then suggest that collusion is hard to sustain but multi-market contact restores in such a case an overall symmetry that might facilitate collusion (Ivaldi et al., 2003).

11. When firms offer differentiated products (vertical differentiation or horizontal differentiation), collusion is more difficult (Ivaldi et al., 2003).

12. Import also has an influence on concentration: when the import penetration is low, this signals that there is little international competitive pressure. The lower the import penetration, the higher the concentration and thus a higher chance of anticompetitive behaviour.

13. Innovation makes collusion on prices less easy to sustain (NERA, 2004). Industries that do not invest a lot in research and development (R&D) are considered to be more suspicious than industries that show high R&D figures. That is, firms that are member of a cartel do not have to spend a lot on innovation in the market and gain profit. Contrary, a firm that is active on a competitive market would have to innovate a lot in order to keep up with the rest (Symeonidis, 2003).

14. Low market share variance and 15. highly symmetric market shares are indicators that measure the dynamics of collusive behaviour. Over time, with other variables changing, one would expect a change in market shares as well. Stable market shares therefore indicate collusive behaviour (Grout & Sonderegger, 2005).

Table 3 - Indicators of collusion

(1)

INDICATORS Effect on collusion

Structural Indicators:

1 Small number of firms

+

2 High Herfindahl-Hirschman Index (HHI) +

3 Symmetry of firms +

4 High entry barriers +

5 Frequency of interaction +

6 Market transparency +

7 Market growth +

8 Demand fluctuations -

9 Inventory or excess capacity +/-

(17)

17

11 Product differentiation -

Behavioural Indicators:

12 Low import penetration +

13 Low R&D expenditures +

14 Low market share variance +

15 Highly symmetric market shares +

2.5 Empirical findings from cartel regressions

In this section, the current empirical findings and articles on screening will be summarized. We will use these studies to provide insights into the usefulness of the various methods. The studies that are discussed below are presented chronologically.

Symeonidis (2003) was the first to use a regression for finding industries that are prone to anticompetitive behaviour. He uses a comprehensive dataset of collusion across British manufacturing industries in the 1950’s to examine the impact of several industry characteristics on the incidence of collusion. Agreements between competitors that were formally registered in compliance with UK restrictive Trade Practice Act of 1956 are used as an indication of the likelihood of anticompetitive behaviour. These agreements between firms were considered lawful at the time. Collusive agreements between firms were common in British industry in the mid-1950’s. Almost half of the manufacturing sector was subject to agreements that significantly restricted competition. The econometric results, based on the comparison of cartelised and non-cartelised industries, suggest that collusion is more likely the higher the degree of capital intensity and less likely in advertising-intensive industries. In contrast with the literature on theoretical indicators for collusion, he finds no clear link between concentration and the likelihood of collusion once he controls for capital intensity. Finally, there is weak evidence that collusion may be less likely in R&D-intensive industries than in low-R&D industries. He came to this result using data on price agreements and industry indicators using a Probit regression model with the basic specification.

The later study by Grout and Sonderegger (2005) is similar to the regression by Symeonodis, however there is one big difference between the two papers. Grout and Sonderegger (2005) use data on failed and/or discovered cartels which induces a strong selection bias (Harrington, 2008). While Symeonodis (2003) has data on all the implied and explicit agreements made, as explained above. However, after cartels were found illegal, the data collection on both discovered and undiscovered cartels was not available anymore. Nevertheless, the results of Grout and Sonderegger are useful in our analysis. Cartel cases from the European Commission from 1990 to 2005 and the US Department of Justice cases from 1994 to 2005 are used in the regression. The regression predicts

(18)

18 the probability of cartel formation with an econometric model using the presence of a cartel as the binary dependent variable. They try to explain the probability of cartelization using various industry specific variables such as industry turnover, volatility measures, measures of entry barriers, cost measures and concentration measures, through regression analysis with the methods (ordered) Logit and (ordered) Probit. Grout and Sonderegger find that variability in growth is shown to have a negative impact on cartels. Furthermore they find that, contrary to the literature, growth in turnover also has a positive impact. In contrast, traditional entry barriers, the level of stocks per firm and the level of R&D per firm have little effect on cartel formation. An interesting finding of the regression is that the results indicate that elements of employment also seem to matter a great deal. That is, employee costs are significant. Even though the current literature does not explain this relation, one can think of factors why higher employee costs are related to collusion. For example, industries with higher employment cost per employee have higher paid staff who may be acquainted of better information which increases the likelihood of a cartel being discovered and the relevant evidence being uncovered (Grout and Sonderegger, 2005).

Petit (2012) also conducted a collusion probability analysis. She uses a structural approach to construct a Competition Index where nine economic indicators for collusion are used as basis for the screening approach. These economic indicators can be found in Table 1. She applied the Competition Index to the Dutch economy, which is divided into 502 industries at the 4 digit level. The Index acts as a supplementary instrument to complaints, signals and whistle-blowers, which may be used to find possible problematic cases regarding anticompetitive behaviour. Petit uses membership functions to transfer the raw data into numbers between zero and one, where zero indicates a likelihood of anticompetitive behaviour being absent and an outcome of one indicates the likelihood of anticompetitive behaviour being present. Using the Competition Index she finds that the industries: ‘manufacturing of malt’, ‘manufacturing of non-distilled fermented beverages’ and the ‘manufacturing of lime’ are the industries that are most prone to anticompetitive behaviour.

Antonielli and Mariniello (2014) use a dataset of manufacturing sectors from five European countries (France, Germany, Italy, Spain and the United Kingdom) between 2000 and 2011 and identify a number of key sector-level features that have a positive impact on the likelihood of collusion. They use Antritrust Risk Indicators to rank sector’s predisposition to collusion. The indicators that they use are proxies of the market concentration, the likelihood of entry, the stability of demand and supply and the market symmetry. To measure the level of concentration within an industry they use the average price-cost margin for the period 2000 – 2011, the industry concentration ratio for 2010 and the HHI for 2010. For the likelihood of entry they build two indicators: the firms’ size and import penetration. Firms’ size is computed as the average size of companies within the sector for the period 2000 – 2011, for import penetration they divided the yearly average of sector imports by the sector production. For the market stability they computed

(19)

19 two indicators: variance in market size and variance in import penetration. High variance levels are supposed to indicate lower market predictability and lower likelihood of collusion. To test for the symmetry within the sector Antionielli and Mariniello construct an asymmetry indicator based on Gini’s coefficient. They employ it on the distribution of the production shares of the top four companies in the sector. If the indicator is close to zero, this means that the four observed companies have identical production shares and the market is perfectly symmetric. When the indicator approaches 100 this means that there is a huge gap between the market shares held by the biggest company and the one held by the smaller ones. After raking the sectors on the basis of the above indicators, using Eurostat NACE 2-digit classification, the sectors that appear more exposed to collusion risk are ‘tobacco’, ‘pharmaceuticals’, ‘beverages’ and ‘chemicals’.

FOD Economie (2015) used a top-down screening method to test the likelihood of collusion in the Belgian market. They used statistical indicators (number of firms, HHI, import penetration, capital intensity, price-cost margin) and dynamic indicators (weighted market share of exit and entry firms, survival rate, market share volatility, churn rate) to screen the market. The horizontal screening investigates which sectors are less competitive and thus are worthy of further scrutiny. FOD Economie divides Belgian markets into two groups: industrial industries and service industries which are classified using Eurostat NACE 5-digit division. Two methods are used; an analysis per indicator and an aggregated method in which the different indicators are merged and integrated into a single variable. The first method, the analysis per indicator, gives a conclusion per indicator and gives the industries with a high likelihood for anticompetitive behaviour. This is done for the nine indicators. Some industries are ranked high in at least five indicators, for example the ‘manufacturing of sanitary ware’, ‘manufacturing of margarine and other food fats’, ‘production of lead, zinc and tin’, and the ‘production of gas’. The second method, the aggregated method, uses the different individual indicators and integrates them into one combined indicator. This combined indicator shows the functioning of the market in a global way and allows for a ranking of the different industries of the Belgium economy. Before the different indicators can be assessed in one method, the indicators have to be normalized to be able to compare the different indicators. The following rescaling method is used for this normalization:

𝑌O.=

𝑌.− 𝑌P.O 𝑌PQR− 𝑌P.O,

where 𝑌O. is the normalized indicator 𝑌 for industry 𝑖 is, 𝑌. the initial value of the indicator for industry 𝑖, 𝑌P.O the observed minimum value of the indicator 𝑌 and 𝑌PQR the observed maximum value of the indicator 𝑌. The normalization method scales the individual indicators into numbers between zero and one, making them easier to interpret. A value of zero indicates a low likelihood of anticompetitive behaviour while a value of one indicates a high likelihood. Next, the standardized indicators are aggregated into one integrated indicator which ranks industries with a higher

(20)

20 likelihood of anticompetitive behaviour. The results from the aggregated index correspond to the results from the analysis per indicator, pointing out similar industries as noted above.

(21)

21

3. The Competition Index applied to the Dutch

industries

To develop a proactive method, we will construct two different approaches. First, we will apply the Competition Index to the Dutch industries using recent data on the industry indicators as explained in this chapter. Second, we will use a Probit regression model to estimate the probability that a cartel exists in an industry, using data on discovered cartels as will be explained in Chapter 4. We will then compare the results and the predictive power of the two methods and check if they give comparable results. This will allow us to determine in which industries the likelihood of anticompetitive behaviour is highest.

The objective of the Competition Index is to make a selection of industries where the likelihood of anticompetitive behaviour is higher than in other industries. In other words, the goal is to screen the Dutch economy for industries that are prone to anticompetitive behaviour. Our subset of industries certainly contains cartels, which can be proven from our dataset of discovered cartel cases in Dutch industries. However, this subset will also include industries where the conditions for anticompetitive behaviour, according to our Competition Index, are satisfied but there is no violation of competition law. Yet, it might well be the case that existing cartels are considered unlikely by the output of our Competition Index. The Competition Index therefore involves ‘false negatives’, as already explained in Section 2.3. On the other hand, ‘false positives’ are not a problem. The Competition Index functions as a screening method and focuses on circumstantial evidence. Therefore it is essential to undertake further comprehensive research in order to make a final judgement about the actual competitiveness of the industries and possible violations of competition law.

The remainder of this chapter is structured as follows. Section 3.1 gives a description of the indicators that are used for the construction of the Competition Index. The raw numbers of those indicators are standardized into comparable numbers between zero and one, as explained in Section 3.2. A weighted average of those number results in a ranking list of all industries in the Dutch economy as can be seen in Section 3.3.

3.1 Description of the indicators

This section discusses the mechanism behind each of the nine indicators that are used in the construction of the Competition Index. We selected the following indicators because the literature points them out as important collusive markers and because of data availability. The indicators can be divided into three groups: ‘market concentration’, ‘likelihood of entry’ and ‘dynamics/market stability’. Unfortunately, we were unable to gather data for a number of indicators that are discussed in Section 2.4. Specifically, the indicators: frequency of interaction (being the number of

(22)

22 trade organizations in the industries), symmetry of firms, market transparency, excess capacity, multi-market contact and product differentiation are not taken into account in the construction of the Competition Index.

Market concentration

The first group of indicators consists of indicators that measure the level of concentration in an industry. According to the theoretical and empirical literature, concentration is an important factor in the formation of collusion. A higher degree of market concentration is associated with higher likelihood of collusion. With a smaller group of players it is easier to reach and coordinate an anticompetitive agreement. Deviation from an anticompetitive agreement is less profitable if concentration is high: the remaining slice the market player would grab by undercutting its competitor is smaller when markets are more concentrated. This means that cartels are generally more stable when markets are more concentrated. To measure the market concentration in an industry we use three indicators: number of firms in 2014, the HHI for 2014, the average price-cost margin for the period 2009-2014 and the Boone indicator for 2014.

1. Number of firms: The indicator ‘number of firms’ concerns the number of Dutch firms active in the domestic market of each specific industry. Industries with a relatively low number of firms have a higher likelihood of collusion.

2. Herfindahl-Hirschman Index (HHI): The HHI measures market concentration of an industry based on market shares. We assume that the higher the HHI, the less competitive a market is, or potentially can be and thus a higher likelihood of collusion. The index is calculated by the sum of squared market shares of the firms 𝑖 within industry 𝑗:

𝐻𝐻𝐼V= W 𝑠.,KY7ZK O

.[7 .

3. Average price-cost margin (PCM): Although Harrington (2006) stresses that there are many industries that may have high price-cost margins, only a few appear to be cartelized; this Competition Index uses the price-cost margin as an indicator for anticompetitive behaviour. This is because the cost margin is low when competition is intense so higher price-cost margins may show anticompetitive behaviour.

𝑃𝐶𝑀V= 1

6∗ W 𝑃𝐶𝑀V. KY7Z

`[KYYa

Generally, heavier competition reduces the PCM of all firms. But since more efficient firms may be able to skim of parts of the profits stemming from their efficiency lead, this may lead to a higher PCM. As such, the estimations of the PCM will typically underestimate the

(23)

23 PCM and the level of competition itself. The cost margin is calculated as the price-cost margin divided by the price (the Lerner-index).

4. Boone’s indicator: Boone’s Indicator (Boone, 2008) is based on the relationship (elasticity) between performance, in terms of profits, and efficiency, measured as variable costs. To obtain the elasticity, the log of profits (measured by return on assets) is regressed on the log of variable costs for different firms at one moment in time. The elasticity is the estimated coefficient computed from the first derivation of a trans-log cost function. The intuition behind the indicator is, first, that more efficient firms (that is, firms with lower marginal costs) gain higher profits and, second, that this effect is stronger the heavier the competition in that market is. An increase in the Boone indicator implies a deterioration of the competitive conduct of financial intermediaries because the effect of reallocation is stronger (Boone, 2008). Following Schaeck and Cihák (2010) the Boone indicator is estimated the following:

ln 𝜋V,KY7Z= 𝛼 + 𝛽 lnA𝑐V,KY7ZB,

where 𝜋_V` measures profits of the industry at year 2014, 𝛽 is referred to as the Boone indicator, and 𝑐V,KY7Z denotes the marginal cost for the year 2014. Since we cannot observe marginal costs directly, we use average variable costs as a proxy.

Likelihood of entry

The second group of indicators measures the likelihood of entry in an industry. The easier entry into an industry the more difficult to sustain collusive prices. If entry barriers are low, high prices will attract new competitors in the future. These new competitors will decrease the gains from collusion, making punishment less costly to bear. Therefore, in markets where entry barriers are low, collusion is harder to sustain. Industries were entry is more likely should therefore ceteris

paribus be associated with lower probability of collusion (Motta, 2004). The likelihood of entry in

an industry is measured through three indicators: the import intensity in 2014, the churn rate in 2014 and the relative R&D expenses in 2014.

5. Import intensity: calculated by the total import of firms in industry 𝑗 divided by total turnover for the year 2014 in industry 𝑗:

𝐼𝑚𝑝𝑜𝑟𝑡 𝑖𝑛𝑡𝑒𝑛𝑠𝑖𝑡𝑦V=

𝑖𝑚𝑝𝑜𝑟𝑡V,KY7Z ∑O.[7𝑡𝑢𝑟𝑛𝑜𝑣𝑒𝑟V,KY7Z

.

A high import intensity suggests that the industry tends to have relatively lower barriers to entry to foreign competitors. Because exporters are exposed to for example different costs shocks, it is reasonable to assume that reaching a collusive agreement with exporters is comparatively more difficult. It would therefore be more difficult for domestic producers to

(24)

24 explain price changes by exporters and detect potential deviation from the collusive agreement that may not be justified by a change in production costs.

6. Churn rate: The churn rate measures the number of firms entering and exiting compared to the number of firms present in industry 𝑗:

𝐶ℎ𝑢𝑟𝑛 𝑟𝑎𝑡𝑒V=

𝐸𝑛𝑡𝑟𝑦V,KY7Z+ 𝐸𝑥𝑖𝑡V,KY7Z 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝐹𝑖𝑟𝑚𝑠V,KY7Z .

It indicates the presence or absence of entry and exit barriers. In the case of a high churn rate, entry and exit barriers can be considered to be low and consequently, the lower the probability of anticompetitive behaviour (Petit, 2012).

7. R&D intensity: Industries that do not invest a lot in research and development (R&D) are considered to be more suspicious than industries that show high R&D expenses. The logic behind this is that a firm that is active on a competitive market would have to innovate a lot in order to keep up with the rest. Firms with high R&D expenses are assumed to be competitive to obtain a “first movers advantages”. The avoidance of R&D expenses might point to possible collusive behaviour. The indicator is calculated by dividing the total R&D expenses of industry 𝑗 in 2014 by the total turnover of industry 𝑗 in 2014:

𝑅&𝐷 𝑖𝑛𝑡𝑒𝑛𝑠𝑖𝑡𝑦V=

𝑇𝑜𝑡𝑎𝑙 𝑅&𝐷 𝑒𝑥𝑝𝑒𝑛𝑠𝑒𝑠V,KY7Z ∑O.[7𝑡𝑢𝑟𝑛𝑜𝑣𝑒𝑟V,KY7Z

.

Dynamics/market stability

The third group of indicators measures the dynamics of the industry. Collusion is mostly found in mature industries in which the business demography is stable. In a stable market it is easier to make long-term agreements compared to developing markets. When industries are subject to frequent and unpredictable demand or supply shocks, detection of deviation is more difficult, hence collusion is less stable. Furthermore, it is often the case that in stable markets the same few firms have been active on the market for a long time. This makes it easier to form a cartel given that the firms get to know each other’s characteristics and behaviour which lowers the information asymmetry. The amount of dynamics in an industry is measured through two indicators: market growth and the ratio of bankruptcy.

8. Market growth: The market growth in industry 𝑗 is determined by calculating the growth of the total turnover in industry 𝑗 over five years (2009-2014):

𝑀𝑎𝑟𝑘𝑒𝑡 𝑔𝑟𝑜𝑤𝑡ℎV=

∑O.[7𝑡𝑢𝑟𝑛𝑜𝑣𝑒𝑟V,KY7Z− ∑O.[7𝑡𝑢𝑟𝑛𝑜𝑣𝑒𝑟V,KYYa ∑O.[7𝑡𝑢𝑟𝑛𝑜𝑣𝑒𝑟V,KYYa

.

(25)

25 So, high market share volatility may be a sign of collusive behaviour. The reason behind this is that with other variables changing, you would expect a change in market shares as well. Stable market shares therefore indicate collusion (Harrington, 2006).

9. Ratio bankruptcy: Although there is no clear empirical evidence on the relation of bankruptcy in an industry with the likelihood of collusion, one can think of arguments for this relationship. One example is that when the ratio of bankruptcy is high in the industry, firms tend to combine their powers to prevent to be declared bankrupt. Therefore, if the ratio of bankruptcy is relatively high in an industry, the likelihood of collusive behaviour is higher. On the other hand, one might argue that a very low value of bankruptcy might be the effect of collusion in the market. Therefore very low ratios of bankruptcy might also signal collusive behaviour. We believe that there is always a small ratio of bankruptcy in the market. The indicator is calculated by dividing the firms that went bankrupt in industry 𝑗 in 2014 by the total firms in industry 𝑗 in 2014:

𝑅𝑎𝑡𝑖𝑜 𝑏𝑎𝑛𝑘𝑟𝑢𝑝𝑐𝑡𝑐𝑦V=

𝑇𝑜𝑡𝑎𝑙 𝑓𝑖𝑟𝑚𝑠 𝑡ℎ𝑎𝑡 𝑤𝑒𝑛𝑡 𝑏𝑎𝑛𝑘𝑟𝑢𝑝𝑡V,KY7Z 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝐹𝑖𝑟𝑚𝑠 V,KY7Z .

3.2 Aggregation of the indicators

Before we can use the above described collusive indicators to determine the likelihood of anticompetitive behaviour in a particular industry, we need to find a way of aggregating the individual indicators into one index. Specifically, how do we take all the separate values of the indicators, for example the number of firms or the R&D-intensity, into one number that tells us how likely collusion is in an industry. In order to build an index, we need two ingredients: a way to process and scale the value of each indicator and a method to aggregate these separate values. In this thesis we follow the aggregation method used by Vermeulen (2007) and Petit (2012). This method uses a so-called fuzzy decision making technique, or membership function for the aggregation of the data.

Suppose we use linear functions to scale the indicators into the final index, the outcome of the regression is the following:

𝑦| = 0.10 − 0.0001 ∗ 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑓𝑖𝑟𝑚𝑠 + 𝛽}K∗ 𝑥K+ ⋯ + 𝛽}O∗ 𝑥O.

According to this estimation result, if the number of firms increase, the probability, 𝑦, of a cartel in that industry decreases. The variable ‘number of firms’ is multiplied by a fixed coefficient and the effect on the probability is linear. To illustrate, in an industry with 1000 firms, the effect of the number of firms on the probability of the presence of a cartel is 10 times smaller compared to an industry with 100 firms. In reality, however, the probability of a cartel does not change when the

(26)

26 number of firms exceed a certain value. The likelihood of anticompetitive behaviour is the same in an industry with 100 firms as it is with 1000 firms. Therefore it would be optimal to use the indicators in a non-linear method. One of such methods is the use of membership functions.

We apply membership functions (or fuzzy sets) to facilitate comparison and easy interpretation of the different industry indicators. This method, extensively described by Kaymak & Sousa (2003), has two important features. First, it provides the possibility to use non-linear functions to scale the indictors. The fuzzy approach allows us to use any type of function to rescale the values of indicators as already explained above. Second, it allows us to combine the nine indicators in a more complex way than by just taking the weighted average.

Following the method of Petit (2012) and Vermeulen (2007), the raw numbers of those indicators per industry are standardized into comparable numbers between zero and one based on literature and our own empirical research. Where zero indicates a likelihood of anticompetitive behaviour being absent, and one indicates the likelihood of anticompetitive behaviour being present. The higher the value of the indicator, the higher the likelihood of collusion. A membership function f outlines the value of the indicator x into the range between zero and one: 𝑓7: 𝑥7→ [0,1]. This can be done in a non-linear fashion, but also by using a linear function. Membership functions allow for subjective relationships and thus the values are debatable. Yet, we apply this concept for the Competition Index because it is the most suitable alternative. The validity of the Membership functions is discussed in Section 5.2.2.

Below the membership function of the nine indicators are given. The slopes of the membership function is determined based on empirical insights and literature. The reasoning of the shape of these slopes is explained in Section 3.1 and the formulas of the slopes can be found in Appendix 8.1. In the following, a brief description of the individual slopes is given. Graph 1 shows that a HHI-score exceeding 1000 (for scaling purposes this number is visualised as 0.1 in the graph) becomes a problematic value for the likelihood of collusion. Graph 2 illustrates that if an industry has less than 400 firms, the likelihood of collusion is high. Graph 3 shows that when the market is stable the likelihood is high. We assume that there is small growth present in the market (1%). Graph 4 illustrates that there is a, more or less, negative linear relationship between the churn rate and the likelihood of collusion if the churn rate is below 1. Graph 5 shows an upward sloping function for the average price-cost margin where a positive margin (>0) becomes problematic. Graph 6 depicts a linear line between the R&D intensity and the likelihood of anticompetitive behaviour in the case of R&D-intensity rates of below 6. R&D-intensity values exceeding 6 are not considered to indicate anticompetitive behaviour. Graph 7 visualizes that import penetration rates lower than 1 increase the likelihood of anticompetitive behaviour. Graph 8 shows that values below 0.1 of Boone’s Indicator decreases the likelihood of anticompetitive behaviour. Finally, Graph 9 depicts that relatively high and relatively low ratios of bankruptcy will increase the likelihood of

(27)

27 anticompetitive behaviour. We assume that there is always a small fraction of bankruptcy (0.025), which is the stable situation.

Figure 1 - Membership functions

One of the comments often heard on this approach to use literature and our own empirical research is that the weights and membership functions are rather arbitrarily chosen. While this is true to a certain level (after all, we indirectly use the results of empirical studies to determine the slope and the weights), the question remains whether the use of a regression analysis is less arbitrary. This limitation will be discussed further in Section 5.2.

3.3 Data sources and descriptives

In order to calculate the Competition Index, we need to gather data for each industry and indicator of the Index. The data we used to calculate the Competition Index for each industry is gathered from Statistics Netherlands (CBS) and is processed into useful data for the Index. For the purpose of our analysis, we have partitioned the Netherlands’ economy into groups of comparable industries. The industry classification we have used is the Standaard Bedrijfs Indeling (SBI) 2008 version 2018 as defined by Statistics Netherlands (CBS). Many countries use their own classification systems for the companies operating in their geographical area. The SBI classification is comparable to industry classifications such as the European Industry Classification (NACE), and

.1 .2 .3 .4 .5 .6 .7 .8 .9 1 C om pe tit io n In de x 0 .1 .2 .3 .4 HHI Graph 1 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 C om pe tit io n In de x 0 1000 2000 3000 4000 Number of Firms Graph 2 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 C om pe tit io n In de x -.4 -.2 0 .2 .4 Market Growth Graph 3 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 C om pe tit io n In de x 0 .5 1 1.5 Churn rate Graph 4 .2 .3 .4 .5 .6 .7 .8 .9 1 C om pe tit io n In de x -1 -.5 0 .5 1

Average Price Cost Margin Graph 5 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 C om pe tit io n In de x 0.00 2.00 4.00 6.00 8.00 R&D Graph 6 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 C om pe tit io n In de x 0 .5 1 1.5 2 Import Penetration Graph 7 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 C om pe tit io n In de x -4 -2 0 2 4 Boone's Indicator Graph 8 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 C om pe tit io n In de x 0 .01 .02 .03 .04 .05 Ratio Bankruptcy Graph 9

(28)

28 divides the economy in sectors based on their activity. The classification starts with 21 general sections marked with a letter. These sections are separated into 2-digit industries which are subsequently separated into 3, 4, 5 and maximum 6-digit industries. Examples of such activities are the ‘growing of vegetables, roots and tubers’ (0113), ‘processing of potatoes, vegetables and fruit’ (103), ‘wholesale of vegetables and fruits’ (46311) and ‘shops selling potatoes, fruit and vegetables’ (4721). Each sector is divided into different underlying sectors, so that at the most disaggregated level we would ideally approximate the relevant market.

In order to effectively determine the location of possible collusive behaviour, we need to carry out the analysis at the most disaggregated level as possible. Due to data restrictions and availability, our analysis is executed at the 3-digit level. However, for some industries the data was available at the 4-digit level. For that reason we used the values at the 4-digit level if that was a possibility. A limitation of this industry classification is that this does not necessarily capture the relevant market, in competition policy terms. Clearly, this limits the effectiveness of our index. The SBI 2008 captures 261 industries at the 3-digit level3_{. All the data are for the year 2014, except for the}

indicators ‘average price-cost margin’ and ‘market growth 2009-2014’ which are averages or growth rates for the years 2009 until 2014.

Table 4 - Data descriptives

The data descriptives of the nine indicators of the Competition Index can be found in Table 4. It can be seen that the sample of the indicators fluctuates per indicator. For 145 industries the only available indicator is the ‘number of firms’. Therefore we excluded these 145 industries. Appendix 8.2 provides an overview of the excluded industries. The appendix shows the excluded industries ordered by the number of firms in the industries. Our analysis will be constructed for the remaining 222 industries in the sample. Furthermore, Table 4 shows a high mean of the indicator ‘number of firms’. One might argue that the industries are not aggregated enough based on the high average

3_{As can be seen in Table 4 there is a sample of 367 industries for the indicator ‘number of firms’. This has to do with the}

4-digit industries that are also in this analysis.

(1) (2) (3) (4) (5)

INDICATORS N Mean Deviation Standard Min Max

Number of firms 367 4871.05 12327.92 0.00 95756

HHI 220 0.09 0.11 0.00 0.74

Average price-cost margin 222 -0.53 7.49 -88.42 61.42

Import intensity 162 0.23 0.18 0.00 1.91 Churn rate 181 0.13 0.05 0.00 0.28 R&D intensity 222 2.25 7.09 0.01 75.29 Market growth 2009-2014 213 0.22 0.64 -0.56 8.32 Boone’s Indicator 222 -1.81 4.22 -29.97 7.59 Ratio Bankruptcy 222 0.01 0.01 0.00 0.04

Predicting cartels in Dutch industries : a pro-active approach