Detecting cartels using economic analysis
Master’s Thesis Economics
Abstract
In this paper I test different methods that have been suggested in the literature to detect collusion in a market, using only data on prices and costs. These methods are applied to the Dutch gasoline market. Using the results of research by Heijnen et al. (forthcoming), I divide the dataset in a part that is suspected of collusion and a part that is not suspected. In this way,
I can confirm if the results are driven by collusion. Using these tests, I do not find any convincing evidence for collusion. This paper concludes that using only price and cost data to
detect collusion is a difficult approach. The results of such a test are ambiguous to interpret, and although these methods are in theory applicable in some cases, in most cases additional
data is required.
Cor Schoonbeek
Student number: 1478427 December 2009
Supervisor: dr. M.A. Haan
1 1. Introduction
Antitrust authorities are always on the lookout for cartels and other collusive behavior. While it is clear that cartels exist and that they have an anticompetitive effect, locating them can be quite a difficult affair. Often, cartels get detected through, for instance, buyer complaints or through the use of a leniency program that provides an incentive to insiders to come forward with information about the cartel. Another method, suggested in recent literature (Harrington, 2005), is for the antitrust authority to take a more active stance, and use economic data to locate suspicious firms. Since the formation of a cartel changes the behavior of the participants, the analysis of price data could be used to detect possible cartels. Detecting cartels in this manner does not result in hard evidence. As Motta (2004) explains, convicting a firm using only economic data is not a feasible approach, since a certain suspect price pattern could also be caused by a firm that behaves competitively. However, economic analysis can still be used to screen for suspect firms; if this analysis gives reasonable grounds for suspicion, an investigation can be started to provide hard evidence, for instance by surprise inspection of these identified firms. The results of this investigation then determine if there really is a cartel, or if the firm is simply acting in its own interest without communicating with other firms.
2
The price data that I use is a dataset containing daily observations on almost all of the Dutch gasoline stations, which is created with the use fleet card data. This data is also used by Heijnen et al. (forthcoming), and I use their research as a starting point. These researchers apply the concept of variance screens, as explained by Abrantes-Metz et al. (2006), to examine if there are areas in the Netherlands where the stations use a suspect pricing behavior. In other words, these authors search for possible locations where the gasoline stations could have formed a cartel. They do this by locating firms with a lower than average price variance. The firms which have the 10% lowest price variance are marked as being suspect. In this paper, I look at methods that can be used besides the method of Heijnen et al. to detect collusion in a market. I use the results of Heijnen et al. to create a sample of firms that is suspect, because these firms have a low price variance. To control for possible false positives, I also test the methods on gasoline stations that are not suspected. My dataset is thus divided in two parts, a suspected and a non-suspected part.
The paper proceeds as follows: section 2 explains my approach in more detail and gives an overview of the previous literature in the field. In that section, I try to locate theories and models which can be useful in determining if there is collusion in the dataset. Section 3 is a continuation of this section, where I describe the models in more detail. The fourth section describes the data. Section 5 shows and discusses the results. Section 6 concludes.
2. Literature
Cartels are bad because they limit the competition in a market. The type of cartel that I look at in this paper is an agreement between competitors to coordinate (and raise) prices, so as to increase profits. Increasing prices, however, also creates the possibility of cheating by a member of the cartel (Harrington, 2005). That is, the higher price would give an incentive to one of the cartel members to slightly reduce its price below the level of the cartel, thereby capturing the entire market for that period. This results in higher profits for that member, since he does not have to share the gains with the other cartel members. To reduce this possibility of deviation, the cartel initiates a punishment phase after someone has deviated, lowering the price to the competitive level, or even lower. This punishment phase lasts a certain number of periods, after which the cartel price can be reinstated again.
3
profits during this phase; and the discount rate of the firm, which is used to discount the loss of profits in the punishment phase to the present. The cartel chooses the profits and the length of the punishment in such a way that the firms will have an incentive to sustain the collusion. This occurs when the one-period gain from deviating does not cover the loss of income from the punishment phase.
In mathematical form, this incentive compatibility constraint becomes the following. Consider a cartel member that receives a profit of πc when he chooses to collude, and a (higher) profit of πd when he deviates. However, deviation will result in a punishment phase for T periods, in which the firm will receive profits πp that are the same as those at the non-collusive level, or perhaps even lower (the cartel can decide to punish more severely by pricing at, or even below, marginal cost) . In this notation, the constraint becomes (following Harrington (2005))
(
c p)
d c T π π π π δ τ τ − ≥ −∑
=1 , (1)where δ is the firm’s discount rate. In words, it states that the loss that occurs because of the punishment phase should be larger than the gain by deviating, so that the firms will continue to collude.
Although it is tempting to label all firms with higher than average prices as part of a cartel, this approach is not feasible, since the price could be higher for that firm because of a large number of reasons; for instance, the firm could have less local competitors or be situated in an area with a larger demand. Therefore, other tests are required to determine the existence of a cartel.
4
coordinate a price change than it is for a station manager that acts competitively. The latter can make pricing decisions on his own, while the former has to discuss what to do with the price. If a cartel participant decides to change his price on his own, this could be seen as a breakdown of collusion. This is one reason why the price of a cartel has a low variance. Harrington (2005) adds another reason that explains why a cartel has a lower price variance: the firms in a cartel want to avoid detection, and therefore they will not raise their prices abruptly in response to positive cost shocks. If the public sees a price series that constantly increases, they would find this series suspect and could report it to the antitrust authority, which in turn could result in an investigation into the cartel. Therefore, these firms will respond to cost shocks, but will do this in a less abrupt fashion than the competitive firms. There is a caveat to this theory: while a low price variance could be an indication of collusive behavior, there could also be other reasons for the low variance in those regions. Furthermore, colluding firms cannot always be distinguished by a low price variance. Harrington (2005) points out that under certain circumstances, price variance could even be higher than in the case of no collusion. Thus, while a low price variation could be an indication of a collusive agreement, further testing is required to confirm (or deny) these suspicions.
5
Figure 1. Black squares indicate the stations that have the lowest 10% variation coefficients, grey dots are the other stations. Source: Heijnen et al. (forthcoming).
As can be seen from this figure, the Randstad area contains the largest amount of these suspected stations. One explanation for this could be that the higher density of stations (the Randstad area is highly populated) makes communication and observation of each other easier, thus simplifying collusion. On the other hand, a larger number of stations also increases the possibility of having maverick firms, that compete heavily for market share. Therefore, it is not necessarily the case that a large number of stations will make collusion easier or more difficult. Other suspected areas can be found over the whole of Noord-Brabant and parts of Limburg, and the city of Apeldoorn also seems highly suspect. In contrast, the northern provinces (Groningen, Friesland, and Drenthe) do not contain many suspected stations. There are also a couple of other cities (Enschede, Middelburg) that stand out because they do not contain any black squares.
6
false positives. To reduce this risk, I also use the different tests on some cities that are not suspect. Table 1 lists the locations chosen for this research.
Suspected cities Non-suspected cities Den Haag Groningen
Rotterdam Leeuwarden
Leiden Enschede
Zoetermeer Bergen op Zoom Apeldoorn
Table 1. List of the cities contained in the dataset.
In the remaining part of this section, I will describe a couple of the different approaches that can be found in the literature to test for collusion. The point to keep in mind is that I look for tests that do not require many different data sources. Data on, for instance, demand is difficult to obtain, and the screening process should not be too difficult or costly to implement. The methods that I describe below rely for the most part on price data (from the dataset described above), and freely available (marginal) cost data.
In section 5, I use the different methods of this section to test if I can find collusion in my dataset. I use these test on the suspected and the non-suspected part of the data. If I obtain results pointing to collusion in the suspected regions, and results pointing to competition in the non-suspected regions for the same test, this will be an indicator that there is indeed collusion in those suspected regions. A nice feature of this approach is that I can directly see, by comparing the different regions, if a positive result for a test is in reality caused by collusion in those regions, or by some underlying feature of the gasoline market in the Netherlands that causes a positive outcome without there being collusion. If the results for a tests indicate collusion in both parts of the dataset, then it can be assumed that the test is wrong, or that there is widespread collusion in the Netherlands that is not picked up by the Abrantes-Metz et al. screen. The latter, however, seems unlikely.
7 2.1 Asymmetric pricing and Edgeworth cycles
The retail gasoline market has been the subject of a large amount of research. This research is in part fueled by the observation of seemingly asymmetric pricing. This is the idea that gasoline prices respond faster to a cost increase than to a cost decrease, which at first sight seems to be the case in the gasoline market: it seems to car drivers that the prices rise in jumps but only drop with a cent at a time. Furthermore, the gasoline market is a favored area for research because it has easily observable prices (since they are posted on large signs outside of the stations, and more and more regularly on the internet) and because fuel is a fairly homogeneous product.
Bacon (1991) was one of the first to focus on the asymmetric response of the fuel price to costs. That is, he tested the hypothesis that the gasoline price increases quickly following a cost increase, but decreases slowly following a cost decrease. Using two-weekly data, he found evidence for an asymmetric speed of adjustment effect. His research spurred a large strand of other papers trying to find asymmetric price responses.
This research is interesting for my purpose because of the various explanations for why this asymmetry occurs. Lewis (2004) compares three different explanations, and creates a method to determine which one of these is correct. Two of these explanations are search models (one of which created by himself, the other by Bénabou and Gertner (1993)), while the other explanation is that this pricing pattern is the result of collusion. The reasoning behind this latter explanation is that when the marginal costs increase above the current price, a firm will choose to raise its price, since continuing to charge low prices will be costly. However, when costs decrease, firms use the past period price as a focal point to tacitly collude, keeping prices higher than they would be in a competitive setting. This results in a quick increase in price when costs increase, but a slow decrease in price when costs decrease. Note that it is assumed here that the collusion is not perfect: in that case, the price would not slowly drop, but firms would simply set a price together and keep that price. Although Lewis concluded that the asymmetry for the area he researched was the result of search behavior, his paper does offer a useful method to test if an asymmetric result is caused by collusion.
8
The literature does not reach a solid conclusion, however. Where one paper finds overwhelming support for asymmetry, another paper will reject the hypothesis of an asymmetric adjustment to costs.
An interesting paper in this regard is that of Eckert and West (2004a), who add several more explanations for the finding of an asymmetric price adjustment. This paper considers the possibility that the price fluctuates without a cost fluctuation. One suggestion they offer is that of Edgeworth cycles. These cycles were first modeled by Edgeworth (1925), and were brought back to attention by Maskin and Tirole (1988). The model posits that firms are engaged in a battle for market share. Their products are homogeneous, and thus the price determines the market share of a firm. When prices are equal, all firms will sell the same amount, but when a firm charges a lower price, he will capture the entire (or a large part of the) market. Therefore, firms constantly undercut each other until price arrives at marginal cost. Then, one firm will start charging a high price again, and the other firms will follow, restarting the cycle. This creates a sawtooth pattern, as shown in graph 1.
Time
P
ri
c
e
Graph 1. Example of an Edgeworth cycle.
9
result is interesting, since it would mean that the conclusion of collusion, after finding an asymmetric price adjustment, could be wrong. If pricing follows an Edgeworth cycle instead, this could point to a competitive environment: prices fluctuate because firms are undercutting each other. Therefore, one should first see if there is an Edgeworth cycle in the data before testing for asymmetry. Noel (2007b) uses a formal approach to detecting these cycles, but this is a rather involved test. Instead, I graph the data and then look for these cycles. As shown in Noel (2007b), such a pricing pattern should be clearly visible.
The above discussion suggests a first test for collusion. Provided that I do not find an Edgeworth cycle in the data, I can test for asymmetry, and, following Lewis (2004), test if this possible asymmetry is caused by collusion. The finding of an Edgeworth cycle in the data, however, will lead to an ambiguous conclusion. On the one hand, Noel (2007a) and others surmise that Edgeworth cycles, because of the furious undercutting, indicate a competitive market. This seems to be the most logical conclusion. However, Wang (2008) shows with proof from an Australian cartel case that a price fixing scheme was used by gasoline stations that resembled an Edgeworth cycle. With this cartel, firms called each other to coordinate a price rise, but then allowed the price to slowly decline until the next rise. Although this does not seem to be the most effective form of cartel, there was illegal price fixing involved in this case.
Another interesting result to consider when using a test for asymmetry is Verlinda (2008), who shows that the asymmetric cycle has faster price rises and slower declines when the firms have more market power. This means that even if both parts of the dataset show an asymmetric response, it is still possible, by determining the extent of the asymmetry, to see which part of the dataset has more market power; if the suspect part has more market power, this could point towards collusion.
2.2 Spatial density
10
easier, thereby improving collusion and raising prices. For instance, in the retail gasoline industry, the closer the gasoline stations are to each other the easier it is to check the prices of the other firms. This means that a negative relation between spatial density and price indicates a competitive environment, while no or a positive relation indicates collusion.
Clemenz and Gugler test this proposition on the gasoline market in Austria. They divide the country in regions, and then test how the average price-cost margin (price minus marginal cost) in a region is related to the spatial density. They calculate the average margin for each region, which is the average of the margins asked by the gasoline stations in that region on all of the days of their sample. This results in one observation per region. The relationship between the spatial density and the margin is then examined for the entire country at once. Of course, the drawback of testing this way is that there are only two possible conclusions: if there is a positive or no relation between the density and the margin, the conclusion will be that there is widespread collusion in the entire Austrian retail gasoline market. When a negative relation is found, the conclusion is that there are no cartels to be found in Austria. By testing in such a large fashion, it is very well possible that smaller cartels are not picked up by this method. Clemenz and Gugler do not seem to consider this possibility, and their conclusion is, unsurprisingly, that there is no collusion in the Austrian retail gasoline market. Although the previous discussion shows that their method of testing does not seem to be entirely correct, their paper does suggest an easy second test for collusion. If I divide the suspect and the non-suspect areas into smaller regions (for this research, I have chosen to use the three digit postal code numbers), I can calculate the effect of spatial density on the margin for both groups. If the effect of spatial density on both groups is different, this can be seen as evidence pointing in the direction of collusion.
2.3 Profit margins and cost fluctuations
11
This discussion surrounding equation (1) assumes a static situation, in which demand and cost do not fluctuate. Borenstein and Shepard (1996), however, use a dynamic model. In this dynamic model, the demand and the input costs can fluctuate. The most important modification to the static situation is that the price-cost margin of the cartel can change based on expected future values of demand and cost. This means that the margin can change without any change to the current demand and cost. To see why this happens, assume that there is a cartel, and demand is expected to rise in future periods. In that case, the stream of future collusive profits will increase. If a firm deviates at this point, it will miss out on the higher future profits. Therefore, the expected increased demand makes collusion easier, and the cartel raises the current price-cost margin. The problem with a test based on this observation is that my dataset does not include observations on demand or sales. However, Borenstein and Shepard suggest a second test, based on (for this industry easily observable) cost data. Assume again that there is a cartel, but now let demand be stable and let the marginal costs be expected to increase in the future. Higher costs will reduce the future collusive profits stream. This makes deviation more likely, since the profits that a firm misses in the punishment phase will be lower. To prevent this, the cartel will decrease their margin.
The different pricing behavior between a collusive and a competitive market can thus be summarized as follows: in a competitive market, the margin that a firm charges is based only on current values of demand and costs. In a market with collusion, however, the margin is not only based on current values, but also on expected future values of these variables. In a collusive setting, changes in the margin can occur before a change in the fundamentals.
Using futures data for gasoline from the terminal, this proposition can be easily tested, offering a third method of determining whether there is collusion in the retail gasoline market in certain areas of the Netherlands. Of course, the above discussion assumes that the incentive compatibility constraint is binding. If it is not, it might be that margins do not change at all when expected future costs change.
12
One thing that should be pointed out is that it is not assumed here that every gasoline station has the same costs, or that the cost only consists of the terminal price. However, the differences in costs between gasoline stations lie mainly in the fixed costs. The marginal costs of each firm consist for the most part of the gasoline price at the terminal, which is roughly the same for all firms. It is this marginal cost which is used in the calculation of the price-cost margin. There are slight differences in input costs based on the operating model of the gasoline station. A station that is owned and operated by an oil company presumably pays less for the gasoline than a station that is owned and operated by an independent, but data on this is not available. The only visible price is the posted terminal price, which is charged to firms that are not owned by the oil company. Insofar the changes in marginal costs for the firms follow the changes in the (observable) terminal price, however, this should not matter. Therefore, I use the same marginal cost for every firm, the price for refined gasoline at the terminal.
2.4 Price stability and volatility
13 3. Models
In this section, I will explain and show the theoretical models I use to estimate the results. These results will be shown in the fifth section. This section is structured in the same order as the previous section.
3.1 Testing for asymmetry
To test for asymmetry, I follow the method used by BCG (1997). In their paper, BCG test if the effect of a change in an upstream price on a downstream price is asymmetric or not. They test different asymmetries: for instance, they test if a change in the oil price has an asymmetric effect on the terminal price, or if a change in the terminal price has an asymmetric effect on the retail price. The latter is the response I am interested in, since the terminal prices are the correct input costs for a gasoline station.
BCG use average retail and terminal prices for a group of cities, measured twice per month. Although I use their model, I use weekly data for a panel of gasoline stations. The basic methodology is the same, but in my case the variables point to the various stations, and not to the average price of all these stations.
Mathematically, the theoretical model of BCG starts from this simple theoretical long-run relationship:
i i
i W
P =
φ
0 +φ
1 +ε
. (2)Here, P stands for the retail price, W is the wholesale or terminal price, the subscript i denotes the firms and ε is an error term. There is no time subscript since this is a general relationship. It simply states that in the long run, the retail price is dependent on the terminal price that the firm has to pay.
Earlier literature, for instance Bacon (1991), continue from this long-run relationship with a partial adjustment model. Such a model takes the form of
(
it it)
itit
it P W P
14
The downside of this model is that the wholesale and the retail price will always move in the same direction. Consider a wholesale price increase in one week and a price decrease in the next week. According to this partial adjustment model, the retail price will increase in the first week and decrease in the second. However, in reality it could well be that the retail price is slow to adjust, and that it would still be increasing at the time that the wholesale price goes down. Because of this restriction, BCG (1997) use a model with multiple lags, in which the adjustment path is less restricted. The advantage of the partial adjustment model is that it returns after a while to the long-run relationship between wholesale and retail prices. To implement a similar mechanism, the model of BCG uses an error correction term. Their model has also been used, or adapted into slightly different forms, by others to test for asymmetry; see, for instance, Noel (2007a), Deltas (2008), Radchenko (2005), and Balke et al.(1998).
Now that we have determined the long-run relationship, the next step is to determine how a wholesale price change interacts with the retail price. It is assumed that this adjustment process takes several periods. A full adjustment of the retail price to the wholesale price takes N periods. Then, the change in the retail price ∆Pt =Pt −Pt−1, can be modeled as:
∑
= − ∆ = ∆ N n n it n it W P 0β
. (4)Here, β is a parameter that determines how P responds to W. This equation acknowledges that the current change in P is not only dependent on the current change in W, but also on the changes of W in the previous periods.
Of course, this formulation creates a symmetric response, since there is no mechanism to differ between positive and negative price changes. Since the interest is in asymmetric responses, BCG alter this function slightly:
15
This is a basic model to test if the retail price responds asymmetrically to an increase or decrease in the wholesale price. If both β’s are the same, then the response is symmetric, but if they are different, the response is asymmetric.
BCG modify this function to take account of a number of econometric issues. The most important change they make is adding an error correction term, as explained above. The error correction term accounts for the tendency to return to the long-run relationship between wholesale and retail prices. This term uses the basic relation described in equation (2), and modifies the model to be:
(
)
(
)
∑
= − − − − − + − + + − − + ∆ + ∆ = ∆ N n it it it in it n n it n it W W P W P 0 1 1 0 1 1φ
φ
ε
θ
β
β
(6)By comparing this equation with equation (2), it can be seen that the error correction term is the lagged residual of that equation. With some more changes, the equation that is to be estimated becomes
(
)
(
)
(
)
it P j t j j it it N n N n n it n n it n n it n n it n it MONTH W P P P W W Pε
η
φ
φ
θ
γ
γ
β
β
+ − − − + ∆ + ∆ + ∆ + ∆ = ∆∑
∑
∑
= − − = = − − − + − + − − − + − + 2 , 1 1 0 1 1 0 1 (7)MONTH is a list of dummies indicating the month of the year to control for seasonal effects. The added lagged and signed changes in the retail price are there to make sure the error term is random. Slightly rearranging this equation, we arrive at a somewhat clearer form to estimate the parameters:
(
)
(
)
(
)
it P j t j j it it N n n n n it n n it n n it n n it n it MONTH W P P P W W Pε
η
θ
φ
θ
φ
θ
θ
γ
γ
β
β
+ − − − + ∆ + ∆ + ∆ + ∆ = ∆∑
∑
∑
= − − = = − − − + − + − − − + − + 2 , 1 1 1 1 0 1 1 1 0 1 (8)16
the appendix. All the parameters required for this function can be taken from the regression results, except for
φ
1. To calculate this value, I simply divide the coefficient on Wit-1 by the coefficient on Pit-1.If this regression yields the result that there is an asymmetric price adjustment in the dataset, then the question arises if this asymmetry occurs because of collusion or because of some other reason. As stated in section 2.1, Lewis (2004) devised a method to determine the cause of the asymmetry. He looks at three different models that could cause asymmetric price adjustment; the first two of these are price search models, and the last one is a collusion model. The pricing behavior of the firms is different between these models. The collusion model is set apart in the way that firms reduce prices: in both price search models, firms reduce prices gradually, while a price reduction in the collusion model can be abrupt if it is caused by a breakdown in collusion. In this case, the price reduction should only be observed in a subset of stations, since that shows that collusion for that submarket has broken down. Lewis (2004) suggests a simple test to determine the pricing pattern. If there is a breakdown of collusion in a city between a couple of the firms, these firms will start charging competitive prices, which are lower than the collusive price and which follow changes in costs more closely. Meanwhile, the firms that are still colluding are charging higher, more rigid prices. By creating a correlation matrix with the average, minimum and maximum price, as well as the costs, it can be seen if the lowest price in a market adjusts in a greater degree to cost than the highest price in the market. If that is the case, then this pattern corresponds to a theory of collusion.
Of course, the assumptions made here are that the collusion is tacit, and that it cannot be maintained for long. Price reductions are mainly caused by breakdowns in collusion. Although such patterns are possible, not all collusive agreements behave this way, making both of these assumptions quite strong.
3.2 Testing for spatial density
17
competitors, prices will go down. Apart from spatial density (S) and transport costs (T), prices are also dependent on marginal costs (c) and on market concentration (C). Market concentration indicates the distribution of the market shares of the firms. When there are fewer firms that serve a larger part of the market, market concentration increases. The effect of market concentration on the price is somewhat unclear. On the one hand, a higher market concentration can cause prices to increase, since it is easier for a firm in a highly concentrated market to tacitly collude. However, a large market concentration can also occur when a firm is very efficient and has low prices, thereby capturing a large market share. In this case, the firm acts competitively.
Summarizing, Clemenz and Gugler arrive at the following function for price:
(
S,T,c,C,...)
P
P= (9)
It is expected that the price increases with higher marginal costs c, and with higher transport costs T. Higher travelling costs for the consumer results in an increased local monopoly of the firm, since the increased costs make alternative firms more expensive to the consumer. The effect of S, the spatial density, is the more interesting part. In the case of competition, a larger spatial density (i.e.: more firms closer together) should lead to a decrease in price, since the local monopoly of a firm becomes smaller. When firms are colluding, however, an increase in spatial density will have no effect on price, since the firms simply fix their price. It might even be that there is a positive effect of spatial density on price. This would occur if the shortened distances between firms make coordination easier, thereby improving collusion and increasing prices.
Clemenz and Gugler do not use daily data, although they do have access to this type of data. Instead, they choose to compute average prices for every region, taken over the entire time span of their dataset. Although the reasoning behind this decision is not made explicitly clear in their paper, it seems logical that they use prices at a regional level because it is much easier to create spatial density statistics that way. Furthermore, they take the average of the prices because the focus is on equilibrium prices. I will follow this logic in my test.
18 k k k k k P c S C M =ln( − ) =
β
+β
ln +β
ln +ε
ln 0 1 2 (10)The subscript k denotes the area (in my case these are the three digit postal code areas). M is the margin that a firm charges, which consists of price minus marginal cost. The explanatory variables are the spatial density S, calculated as the number of gasoline stations per square kilometer in area k, and the measure of market concentration C, for which I use the Herfindahl-Hirschman Index. The market shares used in the calculation of this index are not created by using the sales of each station. Instead, it is assumed that each station in the region has the same output; therefore, I use the number of stations owned by the same company to
calculate the market share of that company. This number is calculated as
2 1 ,
∑
= = k N n k k b k N N HHI .Here, k stands for the three digit postal code area, and b stands for specific brands, so that Nb,k is the amount of firms owned by b in area k, and Nk is the total amount of firms in area k. Clemenz and Gugler also proxy for differing transport costs by using a variable indicating the share of Alps and woods in district k. Such a proxy is not feasible in my dataset, since I only use data for cities. I assume, however, that because there are only cities, transport costs are more or less the same for each city. This approach is strengthened by the fact that this variable was insignificant in Clemenz and Guglers results.
Because I want to test if both parts of my sample behave differently, I slightly modify the equation (10) to be:
(
k)
k(
k)
k k k k k P c V V S V C M =ln( − ) =β
+β
+β
+β
ln +β
+β
ln +ε
ln 0 1 2 3 4 5 (11)Here, Vk indicates the suspected regions of the sample. Using this equation, I can directly test if the suspected part behaves differently. If β1, β3, and β5 are significantly different from 0, then there is evidence that the suspected areas indeed respond to the concentration and density measures in a different way. This could then be taken as evidence of collusion.
3.3 Testing for changing profit margins
19
while I use the price of each station. Secondly, Borenstein and Shepard use only monthly data, while I use weekly data. The higher detail of my data will hopefully yield clearer results. As already explained in section 2.3, the best approximation for the marginal cost is the (observable) terminal price for refined gasoline. Denoting again the retail price as P and the terminal, or wholesale, price as W, this means that the margin is calculated as Mit =Pit −Wt , where M is the margin, and i and t denote, respectively, the station and the time period.
Theoretically, the margin in a competitive world is dependent on demand and on cost, and in a collusive setting it is also dependent on expected values of these variables, as explained in section 2.3. However, due to a lack of demand data, I cannot calculate this relationship fully. Instead, I follow Borenstein and Shepard (1996) in modeling the relation between the current and future costs and the margin, but leave out the demand part of the equation. I experimented with using monthly dummies to compensate for demand fluctuations. Since fluctuations in demand are mainly caused by seasonality, this procedure should remove a large part of the missing information. The results were not significantly changed, however.
The margin can be estimated as
it it it it N n n it n N n n it n N n n it n N n n it n it W P EXPW P P W W M
ε
α
α
α
γ
γ
β
β
+ + + + ∆ + ∆ + ∆ + ∆ = − − + = − − − = + − + = − − − = + − +∑
∑
∑
∑
1 2 1 1 1 0 1 1 0 0 (12)As before, the plus superscript denotes an increase and the minus superscript a decrease. N is the amount of lags used. EXPWit+1 is the expectation of W in the next period, which I calculate with the use of the price of futures on the oil price. ∆ indicates the change since the last period.
20
on current margins from expected future terminal prices. Finally, the lagged retail and wholesale prices (the sixth and seventh terms) are added as an error-correction term.
Borenstein and Shepard create an estimate of the expected future wholesale price by using the following equation: it j j j it it it it it it it it it it it MONTH C b W b C b C b C b C b W b W b W b W b W
ε
δ
+ + + + ∆ + ∆ + ∆ + ∆ + ∆ + ∆ + ∆ + ∆ =∑
= − − − − − − + − + − − − − − + − + − 12 1 1 1 2 1 2 1 2 1 2 1 10 9 8 7 6 5 4 3 2 1 (13) Here, C is the crude oil price and MONTH are monthly dummies. Borenstein and Shepard use this equation to include all information from past periods to correctly predict the next period’s price. The predicted expected future terminal price is then created by using the fitted values of this regression. However, I propose another method to create this estimate, which is easier to implement. If you assume that the terminal price is related to the crude oil price byit
it C
W =
β
0 +β
1 , (14)and you take expectations of this:
1 1 0 1 + + = + it it EC EW
β
β
, (15)then all that remains is to create a prediction of next period’s crude oil price. The most logical prediction of this price is the futures price of crude oil. In other words, I assume ECit+1 = Fit, where F is the futures price of gasoline. This price is observable, and should, theoretically, already contain all the past information. This includes more information than the equation that Borenstein and Shepard use; their equation only uses past terminal and crude prices, while all available news should be incorporated in the futures price. Thus, using the futures price results in an easier and more accurate calculation of the predicted future wholesale price than the method of Borenstein and Shepard.
21 4. Data
The price data comes from a novel source: it consists of fleet card data, collected from the website of Athlon Car Lease. Athlon Car Lease is a large leasing company in the Netherlands with a fleet of around 129.000 cars. The lessees use a fleet card to pay for their gasoline, and this bill is forwarded to Athlon Car Lease, providing the company with daily data for a large number of gasoline stations. The company started to publish this daily data online since August 10, 2005 (Soetevent et al., 2008). Soetevent et al. have downloaded this data daily from Athlon’s website, resulting in a dataset that covers a large number of stations in the Netherlands almost daily. These authors have graciously provided me with this data. The dataset contains observations for 3585 gasoline stations in the Netherlands. By comparison, the NMa (2006) reported that, according to the Bovag (the Dutch industry association for the automotive sector), the Netherlands had a total of 4319 retail gasoline outlets in 2005. They also cite the VNPI (the Dutch Union for the Petroleum Industry), who state that there were 3625 gasoline stations in 2004. The discrepancy is explained by the fact that independent operators are not a member of the VNPI, and are apparently not counted. In any case, the dataset contains a large number of all the gasoline stations. Assuming the Bovag is correct, the dataset contains 83% of the total amount of gasoline stations. Presumably, the missing stations are small stations that are not (regularly) frequented by lessees of Athlon Car Lease, and thus are probably no candidates for collusion.
Because it frequently occurs that stations have missing data for a day, I convert the daily data to weekly data, taking the average of the daily prices for each week. The price data is for the Euro 95 fuel type, spanning the period of October 2005 to August 2007. Some of the figures in the text are drawn using daily data, but the regressions on time series all use weekly data. Even on a weekly basis there are still some missing observations, but since this seems to occur randomly, I do not expect this to have a large impact.
22
this spot price is a good indication of marginal cost for integrated refiners/sellers and for companies that only own gasoline stations. The spot price is expressed in dollar cents per gallon. I convert this data to euros per liter using the daily reference euro-dollar exchange rate as published by the European Central Bank. As with the retail price data, I create a weekly series by taking the average of the daily prices per week.
Graph 2 shows the average price on each day for the suspected and the non-suspected gasoline sites. Also shown is the terminal price. Although not that much information can be gained from this figure, it is already apparent that the average price of the suspected stations lies a few cents higher than the average price of the non-suspected stations. This does not have to mean anything, however: The list of suspected cities contains larger cities than the list of non-suspected cities, and thus the higher average prices can also arise from a higher demand for gasoline in those cities.
23
For the test on spatial density I use data on the size in square kilometers of the postal code areas of the Netherlands, which was provided to me by dr. Stelder of the Rijksuniversiteit Groningen.
The test that follows Borenstein and Shepard (1996) uses cost predictions for the future. As explained in section 3.3, I opt to use data on futures prices to create these cost predictions. This futures data is obtained from the website of the Energy Information Administration, who report it in dollars per barrel. As before, I convert these values to euros per liter. The data used is the ‘Light, Sweet Crude Oil, Cushing, Oklahoma’ weekly series.
The final dataset has a total of 259 gasoline stations, and an average of about 81 observations per station.
5. Results
5.1 Preliminary testing
Before moving to the more involved tests, I perform the analysis of Eckert and West (2004b) on my data, as explained in section 2.4. I follow their approach of calculating the mode price (modemt, where m is the market [either suspect or not suspect] and t is the day), and the difference of each station’s price and the mode price (diffimt, where i indicates the station). The distribution of this difference is shown in table 2.
Suspect stations Non-suspect stations diffimt < 0 33.50 % 37.51 %
diffimt = 0 17.63 % 32.62 % diffimt > 0 48.87 % 35.85 %
Table 2. Distribution of the difference between the daily price and the daily mode of the price, separated between suspected and non-suspected stations.
24
the suspected stations price above the mode price in almost half of the cases, while this is only ~36% for the non-suspected stations.
Theoretically, when firms are (tacitly) colluding, we would expect prices to converge to one another and to remain stable. This would result in a large percentage of the firms setting the same price. This is also the result that Eckert and West (2004b) find. Vancouver, the city that they suspect of colluding, has a much larger percentage of firms that set the same price than Ottawa. In table 2, this kind of effect is not visible, and the results are the other way around: the stations in the suspected part of the dataset set the mode price considerably less often than the ones in the non-suspected part.
Apart from looking at the frequency in which firms set the same price, the frequency at which this price itself changes can also be an indicator for collusion. If the mode price is stable over time, this could indicate price fixing by firms. On the other hand, a fluctuating mode price indicates a more competitive market. We expect, therefore, that the mode price is more volatile for the non-suspected firms than for the suspected firms. Graph 3a and b are graphs of the daily mode price for a subset of one year of the data (1 October 2005 - 30 September 2006).
Graph 3a Graph 3b
Graph 3a and b. Graphs of the daily mode price. Panel (a) indicates the non-suspected stations, panel (b) the suspected ones.
25
41.4%. Contrary to what we expected, the mode price seems to be more stable for the suspect part of the sample.
However, these results can also be caused by the discrepancy in the number of observations: the suspected part contains 65,342 observations, while the non-suspected part has only 18,489 observations. This could explain the large number of suspect firms in table 2 not setting the mode price. Another reason for this result, and the result of graph 3a and b, could be the fact that different cities are added together.
Therefore, I have also looked at the difference between the mode and retail price on a per city basis. Apart from Den Haag and Rotterdam, which have a large number of observations, the amount of observations of the other cities is more or less the same. Thus, the results cannot be driven anymore by a difference in observations. The results on a city basis again pointed in the same direction as the previous results. The non-suspected cities had a much more stable price than the suspected cities. However, closer inspection reveals that this result is not very reliable. Most gasoline stations only have an observation for a few days in a week. This results in a mode price that fluctuates more heavily than the real prices do. To see why this is, consider a theoretical city which only has two firms. These firms do not change their price, and the price of each firm is different. Each day, one of those firms’ prices is observed; this price automatically is the mode price. Although both firms’ prices are stable, the mode price itself fluctuates. Examining the pricing pattern of specific cities, I find the same kind of result. A related problem is that often there are two, or even more, mode prices. With a low amount of observations per day, multiple mode prices can easily surface.
26
stations of the same brand there sometimes is a difference in the relative price they set. For instance, the city of Bergen op Zoom has a cheap and an expensive Esso station. These observations suggest that the gasoline price in Bergen op Zoom is in fact much more stable than the mode price shows.
Graph 4. Graph of the daily mode price for the city of Bergen op Zoom. If a day contained multiple mode prices, the minimum mode was chosen.
Because of gaps in the daily data, the stability of the price cannot be determined in a reliable way, creating an inconclusive result for using the analysis of Eckert and West (2004b). It is interesting to see that table 2 results in a different conclusion for which cities are suspect and which are non-suspect than Heijnen et al. (forthcoming). Since it is quite possible that this outcome is caused by the difference in the number of observations, I continue to use the results of Heijnen et al. to make a distinction between suspect and non-suspect cities. Since I use the same tests on both parts of the sample, this should not make a difference in any case.
27
This first test yields inconclusive results. Therefore, I now move on to the more involved tests. Before running the regressions I check the data for econometric problems. Two of the three tests use panel time series data. This means that non-stationarity can be a problem. The common (augmented) Dickey-Fuller unit root test does not work well with panel data. Maddala and Wu (1999) compare different unit root tests that are feasible for use with panel data. They conclude that a test suggested by Fisher (1932), which combines the p-values of individual unit root tests, is a better test than the Levin-Lin and Im-Pesaran-Shin tests. Furthermore, the Fisher test, in contrast to the Im-Pesaran-Shin test, does not require a balanced panel.
The results of the test indicate that my weekly retail price series is not stationary, while my weekly terminal price series is. The weekly terminal price series, however, is the same for every individual gasoline station in the panel; it is one time series which I use for every panel. An ordinary augmented Dickey-Fuller test on this time series indicates that it contains a unit root (and thus is not stationary), a result that is at odds with the Maddala-Wu test. Since the ADF tests the original time series I use, I assume that test is correct and recognize the terminal price series as stationary. The predicted next period’s terminal price is non-stationary according to both the Maddala-Wu and the augmented Dickey-Fuller tests. Stationarity tests show that the first differences of these three series are all stationary, indicating that all three series are I(1).
The non-stationarity of these series could result in spurious results. However, since all these series are related to each other, it could very well be the case that they are cointegrated. I test this by using a least squares regression, with the weekly retail price as dependent variable and the weekly terminal price and the expected terminal price as independent variables. A stationarity test on the residuals of this regression shows that the residuals are stationary, indicating that the three variables are cointegrated1. This can also be seen from graph 5, which shows the average retail and terminal prices for each week and the margin between the two. This margin is more stable than the retail and terminal prices themselves. Because both models use error correction mechanisms, the non-stationarity will be accounted for.
1
28
Graph 5. Graph indicating the average retail price, average terminal price and average margin per week.
5.2 Results for the asymmetry test
I do not find anything resembling Edgeworth cycles after close inspection of graphs like graph 2, which is on an averaged basis, or on other graphs of individual gasoline stations. The pattern that emerges in these graphs of the price data seem to be mainly dependent on cost changes. Therefore, I proceed with the estimation as described in section 3.2 and deem the possible finding of an asymmetric adjustment to cost changes as not being caused by an underlying cyclical Edgeworth pattern in the data.
One important step for this test is to determine the correct amount of lags to use. The Akaike information criterion is minimized when using just one lag, indicating that running this model with one lag has the best fit. Furthermore, a Hausman test to choose between fixed and random effects indicates that fixed effects are the consistent model to use.
29
heteroskedasticity is present in my sample. This is accounted for by estimating with the Huber-White estimator.
The results of this regression are in table 3, reported for the suspect and the non-suspect sample.
Coefficients Non-suspect Suspect
∆W0+ 0.7396 0.7475 ∆W-1+ 0.5635 0.5483 ∆W0- 0.4297 0.4664 ∆W-1- 0.7630 0.7511 ∆P-1+ -0.2098 -0.1762 ∆P-1- -0.257 -0.2908 P-1 -0.1509 -0.1591 W-1 0.0997 0.1128
Table 3. Regression results for equation (8). Monthly dummies not reported
P-values for all variables were 0.000, except for a few slightly insignificant dummy variables.
As can be seen from the table, the coefficient values for both samples are quite close. When calculating and graphing the cumulative adjustment function (details in the appendix), this similarity is also apparent.
Graph 6a Graph 6b
Graph 6a and b. Graphs of the cumulative adjustment functions. Panel (a) indicates the non-suspected sample, panel (b) indicates the suspected sample.
30
I use the delta method to construct 95% confidence intervals. Because of the mathematics involved, which get more complex for each week further away from the shock, I only report these for the first couple of weeks. After week 3, the difference between increases and decreases stays more or less the same, and since the 95% confidence intervals become noisier for each added period, any result that holds for the third week will also hold for the following weeks. These confidence intervals are presented in table 4.
Non-suspect Suspect Cumulative adjustment 95% confidence Cumulative adjustment 95% confidence B0+ 0.7396 0.7060 - 0.7732 B0+ 0.7475 0.7289 - 0.7660 B1+ 1.1361 1.0912 - 1.1810 B1+ 1.1580 1.1331 - 1.1828 B2+ 0.9812 0.9394 - 1.0231 B2+ 1.0143 0.9910 - 1.0375 B0- 0.4297 0.3867 - 0.4728 B0- 0.4664 0.4425 - 0.4903 B1- 1.1375 1.0835 - 1.1915 B1- 1.1739 1.1432 - 1.2047 B2- 0.9171 0.8625 - 0.9717 B2- 0.9754 0.9440 - 1.0068
Table 4. Results for the cumulative adjustment function, including 95% confidence intervals obtained using the delta method.
Notation is the same as in the appendix, Bi+ indicates the cumulative adjustment i periods after a positive shock.
Comparing the adjustments for the non-suspect and the suspect sample, it is evident that the 95% confidence intervals overlap for each period. Therefore, one cannot distinguish between the adjustment pattern of the non-suspect and the suspect sample. Since both samples behave in the same way following a change in the terminal price, the asymmetry is apparently not caused by collusion. The asymmetry can be caused by a number of different reasons, for instance the ones that are mentioned in section 2.1.
Summarizing, testing for different responses in the adjustment to a cost shock does not show any differences between the suspect and the non-suspect gasoline stations.
5.3 Results for the spatial density test
31 Coefficients β0 β1 β2 β3 β4 β5 Value 0.00064 (0.964) 0.03295 (0.045) 0.00072 (0.888) 0.00468 (0.501) 0.00841 (0.303) 0.00273 (0.777)
Table 5. Regression results for equation (11). P-values reported in parentheses
Using a Breusch-Pagan / Cook-Weisberg heteroskedasticity test, I cannot reject the null of a constant variance. Therefore, I conclude that there is no heteroskedasticity in this sample. These results show that most of the coefficients are highly insignificant. This means that I cannot establish a relation between the degree of spatial density and the margin. Interestingly, the only significant coefficient is β1, the coefficient that indicates whether the margin is different for the suspected sample compared to the non-suspected sample. The coefficient shows that the margin is higher for the suspected part, a fact that was also shown in graph 2. However, the results of this test do not show if this margin is higher because of collusion or because of some other reason.
One possible reason that this test yields no significant results is that the amount of data points is quite low. By aggregating the data, I end up with a total of 259 regions, divided into 190 suspected and 69 non-suspected areas. This highlights one problem with this test: to accurately test for the presence of collusion, not only data on prices is required, but also on area sizes and the amount of firms in these areas. The size of the areas chosen should not be too large or too small; ideally, one area should only contain firms that influence each other. However, this choice cannot be freely made because the data is limited to zip code areas or larger regions. To get a good number of observations, a large area should be tested, which defeats the goal of finding local cartels. Although this test is intuitively straightforward, it seems to be only appropriate for large samples.
As in the previous section, the results of this second test also do not point in the direction of collusion.
5.4 Results for the profit margins test
32
in the specification. Finally, as before, a likelihood ratio tests indicates the presence of heteroskedasticity, for which I again account by using the Huber-White estimator.
In section 3.3 I already explained that Borenstein and Shepard use demand data when calculating the determinants of the price margin. Because I do not have sales data, I also experimented with adding twelve monthly dummies to the regression to proxy for the demand. The results were not significantly changed, though. The results are reported in table 6.
Coefficients Non-suspect Suspect
W0+ -0.1616 -0.1654 W-1+ 0.5889 0.5771 W0- -0.4476 -0.4112 W-1- 0.6866 0.6866 P-1+ -0.1956 -0.1750 P-1- -0.1821 -0.2080 EXPW -0.0788 -.0674 P-1 0.8624 0.8569 W-1 -0.7971 -0.7985
Table 6. Regression results for equation (12) P-values for all variables are 0.000
The idea of the test was to see if the expected future terminal price would have an effect on the current price. The effect was predicted to be negative, since a higher future cost would reduce current margins (see section 2.3). If this variable has a significant effect, this could be evidence for collusion. As is shown in table 6, the coefficient is highly significant, both for the non-suspect and the suspect sample. This is a surprising result, since the variance screen did not point out a possible collusive effect in the non-suspect sample.
33
next period’s terminal price and the current terminal price. This coefficient is about 0.96, indicating that the calculated expected terminal price does not add that much to the equation. Nevertheless, I tested equation (12) using this estimate of the future cost. The regressions using the prediction with one lag all show a highly insignificant coefficient for this parameter. The regression using the prediction with two lags shows an insignificant coefficient for the part of the sample that is not suspect, while it shows a significant effect for the suspect part of the sample. Although the latter is the result that we are looking for, since it indicates collusion in the suspect sample, it is a rather weak result; the predicted future cost was estimated using insignificant coefficients. Also, these results are influenced by multicollinearity. Following the estimation of equation (12), I calculated the correlation of the estimated coefficients of the predicted and the current terminal price. In all cases that used the proxy of equation (13) this correlation was above .95. This indicates that the coefficient estimates of these two variables are not very reliable, since their individual effects are hard to disentangle. The correlation is around .75 when using futures as a proxy for the next period’s terminal price.
Summarizing the above results, it is apparent that the constructed predicted terminal price using the method of Borenstein and Shepard yields unreliable results. This proxy is highly correlated with the current terminal price and does not add much to the regression. Choosing for the alternative of using futures data results in a proxy that is less correlated with the current terminal price, although the correlation of .75 is still high. It is expected that multicollinearity has less of an effect on the regression that uses this coefficient. The regression results from using the futures proxy indicates that both parts of the sample seem to respond to expected changes in future costs. This could mean that there is widespread collusion in the gasoline market, but it could also be the result of a different effect. One should be careful to draw strong conclusions from this result, though, since it can in part be driven by the high correlation.
34
suggested prices contain more information than just summarizing changes in the spot price. Therefore, these authors conclude that the suggested prices have a coordinating effect, which can be explained as being tacit collusion. It is quite possible that these oil companies are forward looking in setting a suggested retail price and take expected developments into account. Since these suggested prices are followed by gasoline stations, the gasoline stations, in turn, also let their prices be dependent on future expected costs. The results of table 6 thus confirm the findings of Faber and Janssen.
Although this finding indicates that tacit collusion seems to be present in the Dutch retail gasoline market, it does not help in pointing out locations of specific cartels. Since all firms incorporate future expected costs in their retail price (presumably through the suggested prices), equation (12) cannot be used to determine if there are explicit cartels. Although the suggested price will reduce competition, it is unclear what the strength of this effect is. Even if this effect is small, the firms will still be responding to the expected future costs. The NMa already studied the effect of suggested prices (NMa, 2003), and while it is believed that these suggested prices indeed raise the price level, the NMa also concluded that halting this practice will not result in a more competitive environment. However, because of the price differentials between stations, this tacit collusion does not result in a fully collusive price and firms would be able to improve on this situation by forming a cartel. If these kinds of explicit cartels exist, they are not detectable using the method of Borenstein and Shepard (1996). Their method therefore has limited usability. The market has to conform to a specific structure before this test can be performed. Also, this method requires a proxy for expected future costs; as shown above, constructing a good proxy can be troublesome, and using a wrong proxy does not yield reliable results.
35
actually only performed on the price, and not on the margin. Adding the margin to the left hand side only serves to muddy the results. The theoretical method of Borenstein and Shepard is correct in looking at the margin, but the model itself is not that logical.
6. Summary and conclusions
This paper locates different methods found in the literature that use economic analysis to detect collusion in a market, and applies these methods to the Dutch gasoline market. The results of Heijnen et al. (forthcoming) are used to pinpoint locations that are suspected of containing a cartel. I use the different methods to further test these locations for evidence of collusion. To control for wrong conclusions from the results of the tests, a control group of non-suspected locations is also added. The first three methods that I use do not find any compelling evidence for collusion. The fourth test indicates that there might be widespread tacit collusion present in the Dutch gasoline market. This is presumably caused by the suggested prices set by oil companies. Since all gasoline stations use these suggested prices in calculating their own price, these suggested prices can be used to coordinate tacit collusion. Apart from this nationwide form of tacit collusion (which has already been studied by the NMa), I find no evidence for local cartels. This, of course, does not have to mean that the gasoline market is cartel-free. Additional data, for example data on demand or sales, could help paint a clearer picture.
Detecting collusion using only price data is a difficult approach. This is mainly caused by the fact that collusion can result in multiple different pricing patterns. For instance, Harrington (2005) describes different pricing patterns, or markers, that could appear when a cartel is present, but he also makes it clear that these patterns require certain conditions to be present. Since cartels do not follow clear-cut patterns, it is difficult to draw conclusions from price data alone.
Using economic analysis to detect collusion could be a usable approach in some other markets, but this is very dependent on specific conditions in those markets. The outcomes of these tests can be very ambiguous and difficult to interpret, and they are thus not very reliable. This paper does offer a simple suggestion to determine the cause of an observed price pattern, removing a part of this unreliability: comparing suspected and non-suspected firms with each other will show if the observed pattern is caused by collusion or by something else.
36
by testing at what point the behavior of firms becomes different. This complicates the approach, but it also removes the uncertainty in the selection of firms, resulting in a clearer selection of the suspect sample.
