Testing for collusion in russian oil and gas auctions

Testing for Collusion in Russian Oil and Gas Auctions�

Yuriy Horokhivskyy

Pennsylvania State University

Lily Samkharadze

Pennsylvania State University

April, 2008.

Abstract

We use Russian oil and gas �elds auction data for the period 2004-2008 to investigate bidders�

noncompetitive behavior. We �rst document and analyze two unusual phenomena speci�c to the Russian

environment: shill bidding and a¢ liated bidding. The empirical evidence con�rms that a¢ liated and

shill companies do not bid independently. We then test whether the winning bids are more likely to

be generated by noncooperative or collusive behavior. The model is based on Baldwin, Marshall and

Richard (1997). We �nd that the collusive model markedly outperforms the noncooperative model.

1 Introduction

Collusion in auctions and procurements is a widespread problem. Collusion is an agreement among the

bidders aimed at limiting competition in order to maximize pro�ts. A silver bullet for detecting collusion

is impossible to �nd since there are many di¤erent types of collusion and it is very hard to prove without

data on economic returns. The main idea is to identify observable implications of collusive and competitive

behavior and be able to formally establish a signi�cant di¤erence between them. The caveat lies in that

outcomes in most cases heavily depend on the particular characteristics of the economic environment,

surrounding institutions, and on the nature of a good at sale.

In 2002 the Russian government started to use auction mechanism for the allocation of licenses to explore

and develop oil and gas �elds. Ascending price oral auctions were used to sell the rights for 659 �elds in�The authors want to thank the Institute for Financial Studies in Moscow for support. We are indebted to Rober C.

Marshall for invaluable guidance in research. We are grateful to Andrei Karavaev, Vladimir Kreyndel, Alexey Kuchaev, and

Andrey Vavilov for numerous discussions of the Russian oil and gas lease auctions. We thank Barry Ickes, Sung Jae Jun,

Isabelle Perrigne, Joris Pinkse, Mark Roberts, James Tybout, and Quang Vuong who provided helpful comments.

1

our dataset, covering 2004-2008 period. Despite the fact that there haven�t been any cases of prosecution

for noncompetitive behavior, the pervasiveness of bid rigging at Russian oil and gas �eld auctions is widely

recognized by the experts and in the media.

What we can observe in the data is that in some auctions winning bids are extremly low, while in

others it is substantially high. For examples, in many auctions the �nal price was higher than the reserve

price by more than one hundred bid increments1 . At the same time for a substantial part of auctions

the bidding went as far as 1-3 bid increments only. This feature of the data drew our attention to the

importance of investigating a possible collusive behavior. So, we set out to answer the following question.

After controlling for the demand conditions are variations in prices better explained by the collusive or

noncooperative auction model. It turns out most auctions with little competition are the ones where only

two bidders participated2 . According to the Russian legislation, if only a single bidder registers for an

auction, then the auction is not held. So, if a �eld on sale is of interest to a single bidder only, then the

bidder has an incentive to arrange for a second bidder to register but not compete and thus ensure that

the �eld is auctioned. A second bidder may be either a �ctional company (pure shill) or some other real

company (occasional shill). Closer inspection of the data indicates that such shill bidding does happened in

Russsian auctions. We traced many instances when a given pair of companies participates repeatedly in 2B

auctions, the winner is always one company, and �nal prices were extremely low. We used the information

about shill bidders extracted from 2B auctions to model the collusive probability for 3B+ auctions3 .

Another important phenomenon at Russian oil and gas auctions is that in many cases di¤erent sub-

sidiaries of a company participate in the auctions together. We argue that such a¢ liated bidders are very

unlikely to have di¤erent valuations and to bid independently. Therefore, we model the a¢ liated companies

as a single bidding entity whenever such bidders participate in an auction together. To the best of our

ability we collected information on a¢ liation among bidders and used this information to transform the

observed number of participating companies into the number of e¤ective bidders for each auction.

We adopt the methodology developed by Baldwin, Marshall and Richard (1997) who provide a mech-

anism for collusion and testing for it at English auctions under the independent private value paradigm.

Within the single-object framework we estimate the noncooperative and collusive structural models. Re-

duced form speci�cation is used for the process governing the formation of bidding rings. Conditional

on this fact, the models in the paper are structural. E¢ cient use of sempi or nonparametric estimation

methods is problematic in our case because of the small �nal sample of 3B+ auctions. Thus, we employ

1Bid increment is set equal to 10% of the reserve price.2We call such auctions 2B auctions, as opposed to the auctions with 3 and more participants, which we call 3B+ auctions.3We test for collusion in 3B+ auctions only since there is little competitive bidding in 2B auctions.

2

parametric maximum likelihood inference techniques to estimate the two competing models.

Our estimation results show that the model speci�cation with the e¤ective number of bidders performs

better than the one with the actual number of bidders. It con�rms our intuition about the a¢ liated compa-

nies not bidding independently. Most importantly, we �nd that the collusive model markedly outperforms

the noncooperative model. Also, using parameter estimates of the collusive model we approximate the

revenue loss from collusion for the auctioneer. It amounts to 8.4% (equivalent to $234 mln.).

The paper is organized as follows. Section 2 reviews related literature. In Section 3 we discuss the

speci�c features of Russian oil and gas industry relevant to analyzing the auction environment. Section 4

summarizes the data, addresses the sample selection problem and describes the auction mechanism. In

Section 5 we extensively discuss "shill" bidding and "a¢ liated" bidding phenomena, which are important for

understanding the nature of noncompetitive behavior in the data. The reasoning and empirical modeling

in the paper are conditional on the maintained hypotheses stated in Section 6. The competing models are

presented in Section 7. Section 8 explains selection of covariates. In Section 9 we present the main results

of the paper. Finally, Section 10 concludes.

2 Related Literature

While theoretical analysis of collusion has been given a good deal of attention in the literature, empirical

detection of collusive behavior hasn�t been done by many. Graham and Marshall (1987), and Mailath and

Zemsky (1991) o¤er schemes for division of the collusive gains at a second-price auction. Their mechanisms

are individually rational and incentive compatible. Marshall, Meurer, Richard, and Stromquist (1994) use

numerical technique to illustrate that bidder collusion is easier to maintain and more pro�table at a second-

price auction than at a �rst-price auction. Marshall and Marx (2007) reinforce previous intuition and show

in more detail that a cartel which cannot control the bids of its members can suppress all ring competition

at a second-price auction, but not at a �rst-price auction. In turn, cartels which can control bids of its

members are capable of suppressing ring competition at both types of auction. The authors also analyze

the implications of shill bidding for pro�tability and sustainability of collusion focusing on the collusive

mechanisms which don�t condition on price paid or the identity of the winner. Marshall and Marx (2008)

scrupulously analyze how information structures impact the vulnerability of second-price and ascending

auctions to collusion They show that details of information revelation and registration as well as auction

3

procedures play a crucial (sometimes not intuitive) role in the auction�s susceptibility to collusion. It turns

out that limiting the amount of auction information available to the public inhibits collusion but widens the

potential for abuse by a corrupt auctioneer. In addition, Marshall and Marx (2008) apply their theoretical

�ndings to empirical analysis of shill bidding in Russian oil and gas auctions4 .

In their seminal work Hendricks and Porter (1988) detect informational asymmetry among bidders

in the o¤shore oil tracks lease data, and point out that cartels take advantage of being better informed

than individual bidders. The authors have access to the ex-post production data, analyze �rst-price pure

common value auction environment within a reduced form approach and conclude that companies owning

tracts adjacent to the one on sale don�t behave competitively and coordinate their bids.

Porter and Zona (1993, 1997) take the identity of cartel as given, observe that rankings of cartel bids

don�t coincide with rankings of costs, and propose a test to detect collusion in auction markets based on

the rank distribution of the cartel and non-cartel bids They concentrate on phantom bidding scheme and

show that there must exist fundamental di¤erence between the ordering of competitive and cartel bids

conditional on the observed data. Evidence of collusive behavior is then the fact that the highest non-ring

bid is not statistically distinguishable from that of other non-ring �rms, whereas the determinants of the

highest ring bid di¤er from those of "phony" ring bids.

Bajari and Ye (2003) develop tests for collusion based on searching for patterns in the bidding data

that are not consistent with their model of competitive bidding. They identify two conditions, which

should be satis�ed under competitive bidding, conditional independence and exchangeability; then check

the conditions to identify bidders which may be the members of cartel. As the last step, Bajari and Ye

contrast the model of competition with the collusive model in which potential cartels were identi�ed by the

�rst two checks.

Unlike Porter and Zona (1993,1997), and Bajari and Ye (2003) we don�t have complete history of bids,

rich bidder speci�c data or ex-post production information. What we do know is the identity of the winner,

identities of all the participants, and the winning bid. In our analyses we use the approach developed in

Baldwin, Marshall and Richard (1997). They study forest timber sales at ascending price oral auctions,

adopt an independent private value framework for their environment and draw attention to the fact that

practically no sales are truly single object. By estimating several structural models the authors test for

collusion while allowing for supply e¤ects, and convincingly �nd that the collusive model with no supply

e¤ects performs best.

4Marshall and Marx (2008) work with the same dataset we do in the current paper.

4

3 Russian Oil and Gas Industry5

Russia�s oil and gas industry is a critical part of the world energy market. Russia is the second largest

oil producing country and is the world�s clear leader of the gas sector in terms of reserves, production and

export. One of the notable features of Russian oil and gas industry is its extreme polarization, i.e. a small

number of large companies account for most of the oil and gas production in the country, whereas the

share of numerous small companies is practically negligible. This phenomenon is an essential feature of our

dataset and helps to understand the extent of heterogeneity among the bidders. Looking at the oil industry

separately, about 76% of the country�s oil is produced by the �ve largest oil companies (Lukoil, Rosneft,

TNK-BP, Surgutneftegaz and Gazpromneft). Furthermore, 10 vertically integrated oil companies together

with their a¢ liates account for up to 90% of production. Smaller companies often tend to operate �elds

which are not very attractive to the majors. The situation is even more severe in the gas sector, which is

absolutely dominated by the state owned company OJSC Gazprom with 86% share of total gas production.

The other main gas producers are OJSC Novatek and the four largest oil companies.

Another important thing to note is that production and transportation costs are very high in oil and

gas industries. Costs signi�cantly a¤ect the companies�willingness to pay for a given oil/gas �eld, i.e. their

valuations. Therefore, it is important to understand the nature of the costs incurred by the oil and gas

companies. As a result of the wave of mergers and takeovers, which took place in 1999-2004 in Russia, the

major oil and gas companies today are highly vertically integrated. They carry out exploration, production,

transportation and re�nement of oil and gas, as well as marketing of �nal products through their a¢ liates.

Whereas most large vertically integrated companies own transportation and re�ning facilities (e.g. all the

existing gas pipelines on the territory of Russian Federation are exclusively owned by Gazprom), small

and medium-size oil and gas producing companies have limited access to the pipelines, re�ning capacities

and storage terminals. It is often the case that major corporations are reluctant to make their capacities

available, charge unreasonably high tari¤s for pumping through their territory, thus boosting production

and transportation costs for smaller producers. To sum up, the Russian oil/gas companies incur high costs

of drilling and transporting the resources. These costs are private and vary substantially across companies.

5 Information about the speci�cs of the Russian oil and gas industry was obtained from Wehbe and Maggs (2007) and the

conversations with energy experts from the Institure for Financial Studies in Moscow.

5

4 Data

4.1 Data Sources

Our data was collected in 2007-2008 by the Center for the Study of Auctions, Procurements and Competition

Policy (CAPCP) at the Penn State University in cooperation with the Institute for Financial Studies in

Moscow and is currently available on the CAPCP webpage.6 The data sources include online and printed

government bulletins, as well as other o¢ cial reports in mass media (below we describe in detail both the

data sources and the auction procedure).

4.2 Auction Procedure

In Russia each oil and gas �eld auction is announced through mass media by the federal or corresponding

regional Subsoil Resources Management Agency. The announcement describes characteristics of the �eld

(location, resources, structures), the rules of the auction (participation requirement, date, application dead-

line, reserve price, deposit, increment), and the terms of use (duration of use, exploration conditions, fees

and payments.) The companies apply to participate in an auction, and the agency decides which of them to

admit to the auction based on their application information. The admitted companies willing to participate

pay the deposit in the amount of reserve price. There is a registration procedure on the day of the auction,

which starts one hour before the auction. Registration is held in such a way that every bidder can learn the

identities of all the participants. The auction mechanism is an ascending price oral auction with reentry.

The auctioneer calls out each asking price starting with the reserve price plus one increment (which is set

equal to 10% of the reserve price). The bidders lift assigned paddles to indicate their willingness to pay

that amount. The auctioneer takes larger and larger bids. The winner is the bidder who placed the highest

bid. After the auction, the winner pays the di¤erence between his last bid and the deposit and gets the

license; all other participants get their deposits back. After the auction, the names of the participants and

the winner, the reserve and �nal prices of a given auction are published in government bulletins, which are

the source of our data.

4.3 Data Selection

Our dataset contains information on 659 auctions held during 2004-2008 period for licenses to explore and

develop oil and gas �elds in Russia. These auctions are characterized by the following information: region,6http://econ.la.psu.edu/CAPCP/RussianData/index.html

6

date, reserve price, �nal price, the identities of all participants and the winner, amount of estimated oil and

gas reserves, distances to the closest oil and gas pipelines, oil and gas �elds, road, railroad and settlement.

We also added the average January temperature in the region where a �eld is located.

It is worth noting that our data covers about 90% of all oil and gas auctions held in Russia during the

2004-2008 period (Trutnev (2008)), so it�s a very representative sample. However, the major drawback of

this dataset is that for many observations government bulletins record the results of an auction inaccurately,

in the sense that some covariates are missing. In a number of cases we were able to �ll in the missing variables

using additional sources (mass media, the companies�web-sites), but often we couldn�t �nd all the necessary

information. Thus we had to eliminate a substantial number of observations to construct the �nal dataset

for estimation, which clearly calls into question randomness of our �nal sample. Below we address sample

selection issue in a greater detail. Table 1, in particular, presents a detailed roadmap of the reasons for

elimination and the number of observations discarded at each step.

To explore the nature of missing data, we �rst try to identify the main covariates "responsible" for the

missing values. For 118 auctions out of 659 we don�t have the �nal price. Interestingly, when the �nal

price is missing, almost all other variables (number of bidders, amount of estimated oil and gas reserves,

etc.) are also missing. It appears that what we know about these 118 auctions is largely the name of the

�eld and the date when the auction was held. Putting it di¤erently, we do not observe ex-post information

about the course and the results of these auctions because they were not properly recorded in government

bulletins (which are the source of our data). For the lack of the �eld and auction speci�c information we

cannot analytically address a possible sample selection problem at this point.

Moving on, 541 auctions left can be divided into 3 disjoint subsets: 206 auctions with 2 participants

(2B), 262 auctions with at least 3 participants (3B+) and 73 auctions with the number of participants

unknown. In Section 5 we scrupulously discuss the di¤erences between 2B and 3B+ auctions and establish

the fact that there is little competitive bidding in 2B auctions. It doesn�t make much sense to analyze them

using standard auction theory, and therefore we do not use 2B auctions for the �nal estimation.

In terms of the estimated volume of oil and gas on sale 3B+ �elds are twice bigger than 2B �elds,

which, in turn, are slightly bigger than the �elds sold at the 73 auctions with unknown number of bidders.

Judging from other characteristics of these 73 auctions we can say that they are a mix of 2B and 3B+

auctions with the majority accounted for by 2B auctions. Since we retain only 3B+ auctions for the �nal

estimation, the group of 73 is discarded.

From now on, we focus on the 262 3B+ auctions. For 90 of them we either a) couldn�t locate a �eld

on the map (Palshin (2007)) to measure distances to the geographical objects of interest (reason �poorly

7

mapped �elds, our human error), or b) are missing some of the identities of the bidders, which we will need

in important sense later (here the absence of identities is most likely due to the coding problems). In Table

2 below we compare the 90 auctions with the rest of the 3B+ auctions along the di¤erent dimensions of the

descriptive statistics. The 90 auctions with missing covariates are 25% smaller in terms of the reserve price

and the amount of oil and gas on sale. Smaller �elds were more di¢ cult for us to locate on the map; also,

they may have not being mapped. We conclude that missing covariates don�t create a sizeable selection

problem and arrive at 172 usable observations for estimation.

Table 1. Data Selection Steps

# of Auctions remaining Reason for Removal # of Auctions Removed

659 Initial Dataset 0

541 Missing Final Price 118

335 Auctions with 2 participants 206

262 Missing # of Participants 73

215 Missing Some of the Identities 47

172 Exact Geographic Location not Found 43

Table 2. Sample Selection

Auctions with Missing Covariates Similar to those without Missing Covariates

90 3B+ Auctions with

missing covariates

172 3B+ Auctions without

missing covariates

Average # of Bidders 4.3 4.2

Volume of Reserves, mln. barrels 32 43

Reserve Price, ths. $ 1,936 2,502

F/R 12.2 13.7

4.4 Summary Statistics

Table 4 summarizes the �nal sample of 172 auctions with the number of participants ranging from 3 to 10.

Summary statistics show that the auctioned �elds are highly heterogeneous and there is a lot of variation in

the data. For instance, the �nal price paid ranges from $19,292 to $252,000,000.7 The amount of estimated7All monetary values in the paper are measured in constant 2004 US dollars.

8

oil and gas reserves available on the �eld varies from 0.18 mln. barrels to 508 mln. barrels of equivalent

fuel, with an average of 43 (please refer to Appendix B for the exact de�nitions and the explanation of units

of measurement). Tracts also di¤er signi�cantly in terms of their remoteness from the neighboring �elds,

transportation arteries and other objects of infrastructure.

5 Evidence on Noncompetitive Bidding

Upon closer inspection of patterns in the data we �nd two very interesting facts of unusual participation

and bidding behavior. Throughout we refer to these phenomena as "shill" bidding and "a¢ liated" bidding.

These are two di¤erent types of noncompetitive behavior, which drew our attention to the relevance of

studying collusion in Russian auctions.

Shill Bidding.

Now we turn back to the 2B auctions and provide the rationale for treating them as "suspicious". Also,

we formally de�ne the shill bidding phenomenon. The de�nition of shill bidders here is our own and is

driven by the data.

De�nition 1 Shill bidders are a pair of bidders which participated in a 2B auction with the �nal to reserve

price ratio less than 1.4.

Figures 3 and 4 compare histograms of �nal/reserve price ratios (F/R) for 2B and 3B+ auctions. The

auctions are arranged in the order of decreasing F/R along the horizontal axes. Compare almost symmetric

graph of 3B+ auctions with the heavily skewed picture in case of 2B auctions. The 1.4 cuto¤ comes from

inspecting the Figure 3. The upper contour of the histogram goes smoothly with little rise from right to

left all the way until it hits the bump at F/R=1.4, and then goes up dramatically. In other words there are

disproportionately many 2B auction with F/R< 1:4.

According to the de�nition we were able to �nd shill pairs in 148 auctions with a total number of distinct

shill pairs equal 102.

F/R can be thought of as a rough measure of competitiveness, conditional on the characteristics of a

�eld. The reserve price at Russian oil and gas auctions is set unaggresively as evidenced by the fact that

the average number of increments by which the �nal price di¤ers from the reserve one is 72. Calculations

of the reserve prices are consistent across auctions and don�t involve any strategic considerations.

For 2B auctions F/R is 1.96, whereas for 3B it is equal 12.3. This fact alone clearly does not unambigu-

ously imply noncompetitiveness in 2B auctions. One possible explanation for such depressed winning bids

9

could be simply the fact that these �elds are of low quality and only certain regional bidders are interested

in acquiring them. For example, average volume of the reserves at 2B auctions is half that at 3B auctions.

However, in addition to very low �nal to reserve price ratio 2B auctions have several peculiar features which

are the main reasons for treating them as "suspicious".

Law Requirement : According to the Russian legislation an auction can be conducted only if at least

two bidders are present. That is, if a single bidder registers for the auction, the auction is not held. So, if

a given oil/gas tract is of interest to a bidder, who suspects that no other company may be interested in

the tract, then this bidder will have an incentive to arrange for another company to participate in order to

ensure that the �eld is auctioned. To see that this is happening, we discuss several patterns persistent at

2B auctions.

Repeated participation: By going over the 2B auctions more carefully, we noticed that in several cases a

certain pair of bidders participated in several auctions, and the same member of the pair always won with a

price just 1-3 increments above the reserve8 . This phenomenon can be explained as follows: since creating

a shill bidder entails some �xed cost, a company creates an arti�cial bidder and then uses it repeatedly.

Another nature of shilling may be that a company enters a quid pro quo lasting relationship with some

other real company.

Fictional Bidders: Perhaps the most striking feature of shill bidders is that many of them never win

an auction but participate occasionally. For example, 58 out of 102 "shill losers" have never won a single

auction during the 2004-2008 period. In addition, 29 out of 58 have participated in one auction only,

another 12 participated multiple times, but in 2B auctions only. It is hard to imagine that there exists

a real independent oil/gas producing company, which has never won a single oil/gas tract; moreover,

attempted to purchase several times. All this suggests that a number of bidders are �ctional. (Fictional

companies are widely use in 3B+ auctions as well. In fact, there are four companies �F, I , C, and T

�which didn�t win a single auction, never were in 2B auctions, but participated in 8, 8, 15, and 16 3B+

auctions respectively.)

In view of the above empirical regularities, we make the following conjecture: shill bidding is used as

a device of reducing competition and increasing the probability of winning, or simply for ensuring that

the auction is held. We should note here that after examining 2B auctions closely we believe that shill

8For example, company "Bashneft" participated in 2B auctions 14 times and won all of them with F/R� 1:3: In 6 auctions

"Bashneft" competed against "Bashmineral", and in the other 8 auctions it bid against "Bashneftgeo�zika". "Bashmineral"

and "Bashneftgeo�zika" haven�t won a single auction in the dataset, participated only in the auctions were "Bashneft" was

present; also participated along with "Bashneft" in a number of 3B+ auctions.

10

bidding can take di¤erent forms. Certain shill pairs seem more organized in the sense that they participate

in multiple auctions, while there are occasional shill pairs, which we observe in one auction only. We think

that this di¤erence can be attributed to di¤erent underlying motives for using a shill bidder. As an example,

under scenario 1 company A creates an arti�cial company B or rewards a real company C to ensure that

a particular �eld is auctioned. Under scenario 2 company A enters a lasting collusive relationship with a

real company B. One way or the other, we are convinced that there is little competitive bidding in most 2B

auctions. Hence, there is no point in analyzing those auctions using standard auction models (cooperative

or collusive), which of course presume that there is some competition among bidders. This is exactly the

reason for treating 2B auctions as "suspicious" and separating them from the �nal dataset of 3B+ auctions.

Nevertheless, we are going to elicit a very important piece of information from 2B auctions. To be more

precise, we argue that since shill pairs act noncompetitively in 2B auctions, they are highly likely to bid

jointly and noncompetitively in 3B+ auctions as well. We are going to take the presence of a shill pair in

a given 3B+ auction into account in a way that is discussed at the end of current section.

A¢ liated Bidding.

By careful inspecting the bidders�identities in di¤erent auctions we noticed another interesting phenom-

enon. In many cases the participants in an auction are o¢ cially a¢ liated companies. Surprisingly, such

practice isn�t prohibited by the legislation in Russia. The de�nition of an a¢ liated pair is stated as follows

De�nition 2 If a company A owns (or possesses more than 50% of shares of) companies B and C, then

A-B, A-C as well as B-C are said to be a¢ liated.

Among our bidders we have six companies, which are all o¢ cial a¢ liates of Gazprom, and we often

observe them participating together in the same auction. There are many examples like this in the data.

One possible reason for a¢ liated companies to participate in the same auction is to ensure that the auction

is held. There may be other explanations not known to us. It should be noted that our dataset does not

contain information on a¢ liation. We conducted an intensive project to collect information about a¢ liated

pairs among the bidders. A¢ liation information was obtained from (i) the website of a parent company (if

possible), (ii) government bulletins or (iii) multiple consistent references in the media. We managed to

compile a list of 510 distinct a¢ liated pairs of our bidders. There are 113 unique bidders which are a¢ liated

with some other company in the dataset. For a given company which has several a¢ liates we consider all

possible pairwise combinations on the dimensions parent-a¢ liate and a¢ liate-a¢ liate. Our conjecture is

that such a¢ liated pairs don�t bid independently and competitively. Unfortunately, we observe only winning

bids which leaves our conjecture unveri�able, but intuitive.

11

To complete the story of a¢ liated and shill pairs, we illustrate why information about such pairs is useful

for our empirical analysis of collusion, and how we are going to utilize this information. Table 6 and Figure

5 show how the auctions where a shill or an a¢ liated pair was found compare to the rest of 3B+ auctions

in terms of their competitiveness. Note that, on average, the former group has 30% lower F/R than the

latter. The information we are going to extract from a¢ liated bidding is the following. As it is unrealistic

to expect that a¢ liated companies bid independently, an a¢ liated pair is treated as a single bidding entity.

This gives rise to the di¤erence between what we call the "e¤ective" and the "actual" number of bidders.

For instance, consider an auction with bidders a, b and c, where a and b are a¢ liated. Then, according

to our de�nition the actual number of bidders is three, while the e¤ective number of bidders is two. In

addition, we construct a collusive regime indicator dummy variable, called bidder proximity dummy, which

takes on value 1 if in a given auction at least one shill pair was found, 0 otherwise. We conjecture that

given that there is at least one shill pair among the bidders in a given auction, the probability that bidders

will involve in some type of collusive behavior in this auction is higher compared to the auctions in which

no such pair was found. So, we assign higher collusive regime to the auctions with bidder proximity dummy

equal 1. Construction of a better collusion speci�c variable is extremely data-demanding, and is not possible

without further enriching the dataset.

6 Maintained Hypotheses

We brie�y state and then discuss maintained hypotheses. All subsequent modeling and reasoning is condi-

tional on these hypotheses.

1. Number of bidders who are willing to pay at least the reserve price (who were admitted to the auction

and actually participated) is what we observe and call ni:

2. The bidders are ex-ante symmetric in terms of the distribution of their valuations.

3. Bidders have independent and private valuations (IPV).

4. Variations in oil and gas �elds supply (in the range that they occur in the data) don�t a¤ect bidders�

valuations.

Discussion

1. Number of bidders. Auction records contain the list of bidders who were admitted, who registered on

the day of the auction and actually sat in the room during the auction, regardless of their bidding activity.

12

2. Ex-ante symmetry. Valuation is de�ned as a dollar amount per barrel of equivalent fuel. Since symmetry

means that bidders draw their valuations from the same distribution and valuations are measured per unit,

the mere fact that bidders are of di¤erent sizes, of course, does not imply asymmetry. Imagine a world,

where there are constant returns to scale, and �rms di¤er in size only. Then all the companies are going to

have identical distribution of valuations irrespective of the fact that some produce ten times what the others

do. While we don�t have reasons to think that there are asymmetries in production technology depending

on the size of a �rm, it�s evident that Russian companies di¤er signi�cantly along many other dimensions.

In particular, there are 509(!) di¤erent bidders in 659 auctions; 257 bidders have never won a single auction;

of them 181 have participated in one auction only. It actually suggests that more than half of our bidders

are either �ctional or very negligible players. Tables 7, 8, and 9 summarize the market structure and

illustrate market concentration. To sum up, the group of 10 major companies together with their a¢ liates

have won 39% of all auctions and brought in about 70% of auctioneer�s revenue. It implies that the biggest

most attractive tracts are usually won by the big 10 companies (which is in a perfect line with the general

description of Russian oil and gas industry from Section 3). In view of this heterogeneity it is not very

realistic to expect that the bidders draw their valuations from the same distribution. Nevertheless, as a

starting point we maintain this assumption. One might want to di¤erentiate several groups of bidders (e.g.

the "big 10", "intermediates" and "�ctionals") and treat bidders as symmetric within a group. Such more

accurate formulation would clearly be able to explain the data better but is much harder to do. This is an

interesting direction for future research.

3. Independent private values vs. common values. Certainly, there is a common value component in natural

reserves sales but the importance and the extent of it should be studied carefully for each particular

environment. A number of authors, such as Bajari and Hortacsu (2003), Gilley and Karels (1981) and

Paarsch (1991) suggested examining variations in bid levels as the number of bidders varies as a reduced

form test for common values. Such simple regression approach, which is based on the winner�s curse

phenomenon, "works well for second price and English auctions" (Pinkse, Tan (2002)). We can�t use the

above approach since we have only the winning bid information in the data and, more importantly, don�t

have exogenous variation in the number of bidders. We resort to several informal but important arguments

to justify IPV assumption for Russian oil and gas �elds auction market:

� In Russia the government conducts seismological and geological exploration of a �eld before auc-

tioning it. The amount and structure of estimated reserves are announced publicly in o¢ cial sources before

the auction. All potential bidders have equal access to this information and have the same view about its

13

precision and reliability. There is no asymmetry.

� Bidders are not allowed to conduct their own geo/seismic surveys or do the drilling on the tract

to be auctioned (one seismic survey costs more than $10 mln.). On top of that, government has access to

the production and extraction information from adjacent tracts and uses it for estimation of the reserves

on a tract on sale. The law mandates that companies report production data as well as various technical

information from operating wells in order for the government a) to keep proper track of the rate of

depletion of the resources and b) to enable better assessment of the nearby geological formations. Even

though the owner of the adjacent �eld certainly enjoys some level of informational advantage over the other

auction participants, the government�s e¤orts to inform all bidders maximally and equally greatly dilutes

this advantage.

� Costs of drilling, storage and transportation substantially a¤ect bidders�pro�ts and, hence, their

valuations. Moreover, as we mentioned before, private costs di¤er signi�cantly across �rms. They often

have di¤erent access to the pipelines and di¤erent technologies employed on di¤erent �elds. Costs also

heavily depend on the bidder-speci�c drilling, prospecting and development strategies, �nancial constraints,

opportunity costs and, last but not the least for Russia, on the established connections with local and federal

o¢ cials (giving access to preferential treatments on o¢ cial and uno¢ cial levels).

We take the above arguments as suggestive rather than conclusive evidence that there is a signi�cant

IPV component in Russian auction data. Li, et al. (1999,2000) similarly argued and showed that the

variability of bids in OCS auctions is mostly explained by the variability of �rms�private information.

4. Variations in supply. Baldwin, Marshall and Richard (1997) consider increases in supply as a possible

cause of low winning bids at forest timber sales in the Paci�c Northwest, and construct an additional multi-

unit supply speci�cation of their model. We don�t observe spikes in the supply of the oil and gas reserves in

our data; instead, supply was increasing steadily over the course of 2004-2008 period. More importantly, we

think that for Russian oil and gas �elds environment increase in supply is not associated with lower prices,

at least the increase within the range we observe in the data. We tried to include the supply of reserves in

the region of an auction in a given year as an additional covariate9 in the estimations, but it turned out to

be insigni�cant in all speci�cations.

9We are talking here about the covariates a¤ecting the mean of the distribution of bidders�valuations. See Section 7 for

more details.

14

7 Alternative Models.

Note that theoretical models as well as notation in this section are not intended to appear ours and mostly

parallel those of Baldwin, Marshall, and Richard (1997) whose methodology we adopt.

7.1 Collusion mechanism and distributional assumptions.

Consider single-object English auction in a form of the "button" model as de�ned by Milgrom and Weber

(1982) where n bidders independently draw their valuations for an object from a common distribution and

each bidder�s signal is his private information. Under noncooperative scenario Vickrey�s (1961) logic gives

bidders incentive to bid their true valuations, and the price paid at the auction is equal to the second-

highest valuation. Now suppose that a subset of size l of bidders forms cartel, the rest of the players

acting noncooperatively. The ring can suppress all within-ring competition at English auction using the

mechanisms of Mailath and Zemsky (1991) or Marshall and Marx (2007). They are individually rational,

incentive compatible, and induce cartel members to report their valuations truthfully. The bidder with the

highest report is instructed to participate in the auction and bid up to his valuation, all the other ring

members are instructed not to bid. The winner of the auction will pay the second-highest valuation of

the bidders who bid actively. Cartel gains from collusion only when it can depress the price compared to

the noncooperative outcome, which happens only if at least two highest valuations belong to the cartel

members. We call the number of highest valuations contained in the coalition at auction i the e¤ective

coalition size Ki. Price paid is then equal to the (Ki + 1)th highest overall valuation.

We have no structural intuition for the data generating process of the e¤ective coalition size Ki and

will use reduced-form speci�cation of the actual process. Thus, our theoretical models can be viewed as

structural, conditional on K:We �rst make distributional assumptions about unobserved to econometrician

distribution of private values, then endogenize parameters of the discrete random process governing the

formation of bidding coalitions.

Random variables are denoted by capital letters and their realizations at auction i by lowercase letters

with subscript i: Volume of natural reserves on a �eld, the real reserve and �nal prices are represented

by voli, ri and Pi Key assumption is that distribution of bidders� valuations per unit of reserves, vi,

is lognormal with mean �0zi and variance �2. Mean of the distribution thus depends on the vector of

covariates zi selection of which is described in Section 8 below. We know that price paid for a �eld

at auction takes form of a certain order highest statistic of n independent draws from assumed lognormal

distribution, truncated from below by ri. Transforming winning bids and reserve prices to obtain convenient

15

objects coming from a standard normal distribution we get

Ui =ln( Pivoli

)� �0zi�

;

si =ln( ri

voli)� �0zi�

:

Therefore, we are interested in the density function of the �th-highest order statistic of n independent

draws from a standardized normal distribution now truncated from below by s, which is

f�(ujr; n) = n�n� 1�� 1

��(u) � [1� �(u)]��1 � [�(u)� �(s)]n�� � [1� �(s)]�n

where �(�) and �(�) are distribution and density functions of a standard normal distribution. Also, let

Vi1 � Vi2 � � � � � Vini denote the order statistics of Ui

7.2 The Noncooperative Model

The theory of bidding at ascending price oral auctions predicts that Ui = Vi2 when bidders behave nonco-

operatively. Ui is then the only endogenous object and relevant likelihood can be written as

L(�; �;Data) =Y

i2Data

voli�pi

� f2(uijri; ni)

7.3 The Collusive Model

E¤ective coalition size Ki generates price equal to (Ki + 1)th order statistic of ni independent draws from

the distribution of private values. We assume that there is only one coalition, all non-ring bidders play

noncooperatively. Bidders join a coalition Ci with auction speci�c probability pi: Each bidder�s decision is

independent of those of other bidders at an auction. This is clearly a shortcoming of such modeling but our

16

dataset lacks bidder-speci�c information rich enough to make the decisions to join the ring a¢ liated. We

don�t allow for ex ante asymmetries between ring and non-ring members. Additionally, pi�s are independent

across auctions, conditional on observables we parametrize pi with.

For arbitrary auction i in Table 10 below we show all possible e¤ective coalition sizes, probabilities of

such events and respective implied prices (qci� denotes the probability that in auction i �th highest order

statistic is paid).

Table 10. Collusive Prices

Event Pr(Event) � qci� Price

:[(Vi1 2 Ci) \ (Vi2 2 Ci)] 1� p2i Vi2(� = 2; ki = 0 _ ki = 1)

��1Tj=1

(Vij 2 Ci) \ (Vi� =2 Ci) p��1i (1� pi) Vi�(� = 3! ni; ki = 2! ni � 1)

��1Tj=1

(Vij 2 Ci) pnii ri(� = ni + 1; ki = ni)

Now standardized random variable Ui is a mixture of di¤erent order statistics and has the following

density

hc(ui) =

niX�=2

qci� � f�(uijri; ni); ui > si

The probability pi is parametrized as

pi =e

0wi

1 + e 0wi

where wi is a vector of covariates. Then, likelihood for the collusive model is

L(�; ; �;Data) =Y

i2Data

voli�pi

� hc(ui):

17

8 Selection of Covariates

8.1 Distribution of valuations

We expect the following covariates to a¤ect the mean of ln(vi). Exact de�nitions and units of measure-

ment can be found in Appendix B.

Volume of natural reserves o¤ered for sale on a �eld. Importance of the size of the auctioned �eld for

bidders is obvious, but there are several reasons why it is important. On the one hand, there exist increasing

returns to scale in developing a deposit and extracting resources (as we learned from the experts). On the

other hand, a number of factors decrease willingness to pay per unit of the reserves as the �eld size increases.

Namely, big oil and gas drilling projects require attracting large amounts of �nancial resources (often debt

with higher interest rates); are more likely to be subject to later political revisions (forceful renegotiation,

or even dubious prosecution followed by alienation of property rights as exampled by YUKOS); have greater

exposure to the risks of inaccurately measured or falsely reported amount of the reserves.

Appendix A gives detailed account of di¤erent categories of the reserves and how they are treated here.

An important thing to know is that di¤erent categories of oil and gas reserves are all converted to the

equivalent fuel units. Also, the categories of the reserves can be provisionally divided into two qualitatively

di¤erent groups depending on the methodology and reliability of their measurement. We expect that these

two groups may in�uence the valuations di¤erently.

� Volume_proven is part of the volume represented by the categories of reserves which are well

explored and precisely estimated;

� Volume_probable is part of the volume represented by the categories of reserves which are classi�ed

as promising (or forecast) and are measured less reliably.

RailroadDistance. Intuitively, the farther a �eld is from railroad the costlier it�s to transport the necessary

equipment to the site and to move extracted resources to the railroad if it�s chosen as a preferred mean of

transportation.

Warm is a dummy for mean January temperature in the region of a �eld being higher then 0�F . Costs of

drilling, building additional roads and secondary pipelines are much higher in the regions with extremely

cold climate (even forti�ed equipment fails during the extreme temperature drops).

OilpipeDistance is the distance to the nearest oil pipeline, length of the �rst link in the transportation

chain of the extracted natural reserves. Companies have to use railroad or custom-build secondary pipeline

18

to transport extracted resources to a pipeline in place.

Gaspipe Distance is the distance to the nearest gas pipeline, a¤ecting transportation costs similarly to

the Oilpipe Distance.

Oil�eld Distance is the distance to the nearest oil �eld under exploration, which we view as a proxy

for the availability of infrastructure (in the widest possible sense) needed for oil development projects. For

example, secondary pipeline or additional railroad branch may have already been built from the neighboring

�eld, and you can pay for using them instead of building your own. Since secondary pipelines and railroads

aren�t always charted on the map which we used to measure the distance variables, use of distance to the

nearest �eld as a proxy is justi�ed.

Gas�eldDistance is the distance to the nearest gas �eld under exploration, which is used as a proxy for

the availability of infrastructure needed for gas development projects.

Most of the �elds in our data contain both oil and gas reserves, but some �elds are exclusively oil- or

gas-bearing. To deal with this issue from now on we interact distance variables with the following relevant

dummies:

OilDummy: Coded as one when there is only oil on the tract, or oil and gas but the quantity of

oil is at least four times as large as the quantity of gas as measured in equivalent fuel units; zero otherwise.

GasDummy: Coded as one when there is only gas on the tract, or gas and oil but the quantity of

gas is at least four times as large as the quantity of oil as measured in equivalent fuel units; zero otherwise.

EnergyDist is the distance to the nearest high voltage power line. It may in�uence the cost of powering

up the equipment on the exploration site.

SettlementDist is the distance to the nearest settlement. This distance may impact the cost of hiring

labor and be a proxy for a general level of infrastructure development.

RoadDist is the distance to the nearest road. Companies usually build the roads to connect the exploration

site to the existing ones.

Obviously, our auction models predict that covariates would have a non-linear impact on ln(vi): Nonethe-

less, we use ordinary linear regression to guide us in the narrowing down to the �nal set of covariates as well

as to make an educated guess about starting values for maximum likelihood numerical estimation procedure.

Linear regression coe¢ cients here can�t represent true and pure causal e¤ects. One reason for it is already

mentioned non-linearity in data generating process. Also, we have to remember that dependent variable

is ln(vi) = ln(Pi=voli). Pi is the price paid at the auction and is not a random draw from the assumed

19

lognormal distribution but a second highest order statistic (in case of noncooperative behavior) or mixture

of highest order statistics (in collusive speci�cation) of a number of such draws. In the regression we control

for the endogenous number of bidders as it may be indicative of the level of competition at the auction and

be in�uenced by the unobserved to the econometrician but observed to the bidders information.

Regression results in Table 11 below only include the covariates that were kept for maximum likelihood

estimation. The depressing e¤ect of the size of the �eld on valuation is clearly prevailing over the increasing

returns to scale e¤ect as evidenced by the negative sign of the coe¢ cient on Volume. Location in the

region with warmer climate has a huge positive impact on valuation, increasing it by a factor of more than

2. As expected, the number of bidders is picking up unobserved heterogeneity and competitiveness e¤ects.

For one thing, more attractive tracts draw in more bidders, for another, more bidders mean greater second-

highest order statistic realized, and more competitive auction. Distance to the railroad shows anticipated

negative but insigni�cant e¤ect. Finally, distances to the pipelines and nearest �elds appear to explain

a lot of variation in valuations, yet one can�t give an intuitive explanation for some of their signs. We

are convinced that it is mostly due to the omitted variables problem. There exist important unobserved

region- and auction-speci�c factors at play which aren�t captured in the data and for which distances are

proxies �preventing us from making unambiguous comments about coe¢ cients on distances. For example,

we usually rationalize that the farther a �eld is from a pipeline the costlier it is to develop (lower bidders�

valuations). This logic is undoubtedly true ceteris paribus. Yet here lies a logical twist. It�s known that

huge gas and oil �elds are easier to �nd, more attractive to the buyers, and often put for sale �rst by a

myopic politician. When a �eld goes under exploration, it usually gets connected to a major pipeline by a

secondary one. In such case, the proximity of a �eld on sale to a pipeline may imply that local terrain has

been well explored and there are lower chances that the �eld in question indeed possesses large estimated

reserves, otherwise it would have been found earlier. Speaking about distances to the nearest �elds we

have to note similar presence of di¤erent potentially contradictory forces confounding the analyses of pure

e¤ects of regressors. Again, remoteness from the relevant infrastructure in place should de�nitely depress

prices ceteris paribus. At the same time, as we already mentioned �eldDistances may also serve as a proxy

for pipeDistances in case of poorly mapped pipelines. If this "proxy" e¤ect dominates the "infrastructure"

e¤ect we may be back to the above analysis of pipeDistances.

20

Table 11. Ordinary Least Squares Estimates (N= 172)

Dependent variable

Covariates ln (v)

Constant -0.06

(-0.14)

Volume -0.002

(-1.48)

RailroadDistance -0.003

(-0.41)

Warm 0.82

(2.81)

OilpipeDistance -0.006

(-1.37)

GaspipeDistance 0.002

(2.67)

Oil�eldDistance 0.004

(4.11)

Gas�eldDistance -0.006

(-3.81)

Number of Bidders 0.24

(3.66)

R2 0.24

White�s t-statistics in parentheses

8.2 Modeling Collusive Probability

Volume is the �rst candidate thought to impact the collusive probability. Large tracts appear to be

particularly attractive for the bidders to collude on.

As for the collusion speci�c covariates, ideally we would want to have some universal measure of "prox-

imity" for each pair of bidders. Universal "proximity" dummy would let us use proximity map on the

dimension winner-participants to construct relevant collusive covariate. Unfortunately, available bidder

speci�c information is too scarce in order to construct such a measure. It�s usually argued that if bid

21

increment is small, then after the last non-ring member exits, the two ring members can be recorded as

�rst- and second-highest bidders at the auction at the price of two bid increments. Such scheme would

considerably confound inference for the regulatory body. To the best of our knowledge there has not been

a single case of prosecution for collusion by the Department of Justice, which makes us think that Russian

oil and gas companies don�t engage in such confusing activities. In this light, the fact that second-highest

bidder�s identity is not recorded in our data isn�t a problem. Lacking a perfect "proximity" measure we

instead use bidder proximity dummy which takes on the value one if there is a shill pair (as de�ned in

Section 5) found in the auction, zero otherwise. Two bidders forming a shill have already revealed their

noncompetitive behavior in a 2B auction, which gives us reasons to think that their presence is indicative

of potential collusion at a 3B+ auction. So, bidder proximity dummy appears as additional covariate in pi.

9 Results

Main results are reported in Table 12 below. Columns 1 and 2 contain maximum likelihood estimates of

the noncooperative model, columns 3 and 4 include the same estimates for the collusive model. Actual

number of bidders N is used in column 1. In column 2 we recognize the fact that some bidders in an

auction happen to be a¢ liated, have the same valuation and thus don�t constitute separate bidder entities.

E¤ective number of bidders Naff is used in column 2. The collusive probability is modeled as a function

of a constant variable only in column 3. Bidder proximity dummy is added to the constant in column 4.

Comparing columns 1 and 2 the �rst thing to notice is that model with Naff performs better than

the model with N as suggested by the increase in the log likelihood value. This result con�rms our

intuition about a¢ liated companies not bidding independently. Now we brie�y interpret the estimated

coe¢ cients on the covariates. Coe¢ cient on warm is positive and signi�cant. Strikingly enough, location

of a �eld in a warm climate zone increases its value by a factor of 7, illustrating the level of importance

of natural conditions in determining the production costs in Russia. Notice that volume_probable is

insigni�cant across speci�cations while volume_proven is consistently signi�cant and negative. To remind,

volume_proven is a more precisely estimated part of the reserves compared to volume_probable part. If

bidders possessed asymmetric information about the estimates of the reserves it would rather concern

volume_probable part of it. Yet variation in the valuations is mostly explained by the volume_proven

which is exactly measured and known to all the bidders. We interpret this result as supportive of the IPV

framework that we adopt in our analyses. Additional million barrels of volume_proven on a �eld at sale

generates 2 percent lower valuations per unit of the reserves.

22

Collusive models in columns 3 and 4 clearly outperform noncooperative speci�cations as evidenced by

the increase in the log likelihood function and the signi�cance of the collusive probabilities. Also, the model

with bidder proximity dummy is preferred to the baseline collusive model. Signi�cance of the collusive

covariates is not of much interest itself, because even if = 0 then p = 1=2, which doesn�t imply the

absence of collusion. Therefore, we care about the signi�cance of p in the �rst place. Since there is just one

dummy variable in collusive probability there are only two di¤erent possible values of p : ph = 0:43 and

pl = 0:27, obtained using point estimates of from column 4. We estimate standard errors of ph and pl

with the help of �rst-order Taylor series expansion of ph and pl with respect to the coe¢ cients. Collusive

probabilities turn out to be highly signi�cant in both collusive speci�cations. Presence of a shill pair in

an auction increases the collusive probability by 60%. T-statistics for bidder proximity dummy is equal to

1.22 which suggests that �nding a better proximity measure would further improve the performance of the

collusive model. We also tried using volume of the �eld and dummy for the presence of a¢ liated pairs as

collusive covariates but they were insigni�cant.

After �nding the evidence of collusion it is interesting to know how big exactly the impact of bid rigging

was on the outcome? Using our parameter estimates from column 4 we ran counterfactual analyses trying

to approximate the loss in revenue for the auctioneer. We "ran" all auctions in the data twice: once with

ph = 0:43; pl = 0:27, followed by setting collusive probabilities to zero. The resulting loss in revenue in the

data is 8.4%, or $234 million. On �gure 1 you can see how the revenue loss would change in the collusive

probability (the dot represents actual loss).

The collusive probability decreases the highest-order statistic paid at the auction and does so by way

of coalition suppressing the bids of its members. At the moment when a ring of size Nactual is formed its

members don�t know how their valuations compare to those of non-ring bidders. In other words they don�t

know Neffective = K, which is the number of highest-order statistics possessed by the ring. E¤ectiveness

of a collusive ring is revealed only during the auction, and it is e¤ectiveness alone which matters both for

the ring and for the auctioneer. On Figure 2 we take a look at how expected ring size stacks up against

expected e¤ective ring size as collusive probability changes (for a hypothetical auction with the number of

bidders equal to 10). Cases when nobody joins a ring or just one bidders joins are treated similarly as the

ring of size 1.

23

Table 12. Maximum Likelihood Estimates (N = 172)

Standard N Standard Naff Collusive Naff

Variables in mean (1) (2) (3) (4)

Constant -6.72 -6.15 -2.94 -3.13

(-2.46) (-2.42) (-2.06) (-2.13)

Volume_proven -0.036 -0.032 -0.021 -0.022

(-3.38) (-3.36) (-2.56) (-2.67)

Volume_probable 0.004 0.002 -0.001 0.001

(0.1) (0.1) (-0.04) (0.04)

RailroadDistance 0.001 0.001 0.001 0.001

(0.54) (0.55) (0.62) (0.56)

Warm 2.04 1.98 1.70 1.84

(2.07) (2.02) (2.47) (2.59)

OilpipeDistance -0.008 -0.0034 -0.003 -0.003

(-1.57) (-1.5) (-0.82) -0.85)

GaspipeDistance 0.007 0.007 0.005 0.005

(2.47) (2.31) (2.42) 2.46)

Oil�eldDistance 0.012 0.01 0.007 0.008

(1.6) (1.57) (1.37) 1.31)

Gas�eldDistance -0.01 -0.01 -0.009 -0.009

(-1.36) (-1.34) (-1.48) (-1.47)

St. dev. 3.46 3.47 2.76 2.78

(6.13) (6.01) (6.89) (6.92)

Variables in probability

Constant -0.6 -0.97

(-2.09) (-1.73)

Bidder proximity dummy 0.7

(1.23)

ph 0.35 0.43

(5.41) (5.16)

pl 0.27

(2.46)

Log likelihood -504.02 -495.65 -484.21 -482.68

t-statistics in parentheses 24

10 Conclusion

We analyzed the data on Russian oil and gas auctions held over the course of 2004-2008 period and

found two empirical regularities, shill bidding and a¢ liated participation, which are important features of

the environment. We collected a¢ liation information to arrive at the e¤ective number of bidders at the

auctions, and used the patterns of noncompetitive behavior from two-bidder auctions to help detect collusive

behavior in our �nal dataset of the auctions with three and more bidders. According to the methodology

by Baldwin, Marshall, and Richard (1997) the noncooperative and collusive structural empirical models

were constructed in order to understand what type of bidders�behavior explains the winning bids better.

The estimates obtained convincingly favor the collusive model.

One way to improve current analyses is to �nd more bidder-speci�c information to devise a better collu-

sive regime indicator. Also, collusion in the data is observationally equivalent to "strong" bidders �"weak"

bidders case, which would invariably lead a researcher to discard the distributional homogeneity assumption

even if it is, in fact, true. We have to acknowledge that both factors may be at play simultaneously. In

view of this paper�s approach it would then be interesting to divide the bidders into several categories and

model them di¤erently. One subset of bidders could be the group of 10-15 biggest companies who won

many auctions. Another subset would comprise companies which won just several �elds in the dataset.

First group can then be modeled as drawing its valuations from the conditional distribution which �rst-

order stochastically dominates that of the second group. The last subset would consist of the companies

which haven�t won a single auction. Another line of future work is to model the �rst group of companies as

playing dynamic game and the second one as solving static optimization problem, like in Jofre-Bonet and

Pesendorfer (2003).

We hope that our results will drive more attention to the necessity of detecting collusion. Bid rigging

hurts either a private auctioneer or, indirectly, a taxpayer in case when a government is selling or buying.

Detecting collusion is as important as deterring it. Reliable methods of empirical detection of noncom-

petitive behavior would both help directly deter collusion and importantly guide the process of optimal

mechanism design.

25

Appendix A

Classi�cation of the Reserves

When Federal Subsoil Resources Management Agency solicits bidders for an auction it announces the

estimated reserves content of the �eld on sale. For oil and gas separately the breakdown is made into

the following categories �A, B, C1, C2, C3, D1;D2. The mapping from the Russian classi�cation into

the international one is shown in Table 1 below. Russian classi�cation is based solely on the amount of

geological reserves, without a view to the extent of extractability or economic e¢ ciency and current level of

technology. Categories of the natural resources di¤er by the type and quality of geological, seismological and

production information on which their estimates are based. Categories thus vary in accuracy, reliability and

credibility of their measurement. There is widely used in the industry system of coe¢ cients for conversion

of categories into equivalent oil or gas units.10

Eq.oil.units = 1� (A + B + C1) + 0.7� C2 + 0.3� C3 + 0.1� (D1 + D2)

Eq.gas.units = 1� (A + B + C1) + 0.7� C2 + 0.3� C3 + 0.1� (D1 + D2)

Eq.fuel.units = Eq.oil.units + 0.9� Eq.gas.units1000

Coe¢ cients are the same for oil and gas. Equivalent oil is measured in weight units (tons), equivalent

gas is measured in volume units (cubic meters).The last formula brings oil and gas under the common

denominator and is based on the energy content.

Table 3. Russian and International Classi�cations of Reserves

Appendix B10Coe¢ cients were obtained from experts at the Institute for Financial Studies in Moscow.

26

Variable De�nitions and Summary Statistics

Ni : The number of bidders who are willing to pay at least the reserve price for �eld i and who registered

at the auction, actual number of participants. Ni and their identities are known to all the bidders at the

beginning of the auction.

Naffi : The e¤ective number of bidders who are willing to pay at least the reserve price for �eld i and who

registered at the auction, Naffi � Ni. We obtain Naff

i by taking into account a¢ liation between bidders

participating in the same auction. Whenever two or more bidders in the auction are a¢ liated (one owns

more than 50% of the other) we merge them into a single bidder entity.

Pi : The winning bid (in constant 2004 dollars). Units: cents.

ri : The reserve price set by the auctioneer (in constant 2004 dollars). Units: cents.

vi : Bidders�valuations per unit of estimated reserves, in cents per barrel of equivalent fuel.

Volume: The quantity of natural reserves (oil and gas) as estimated by the Federal Subsoil Resources

Management Agency in mln. barrels of equivalent fuel. Information taken from the auction announcement

and aggregated across di¤erent reserves as well as di¤erent categories of the reserves.

vol: Volume in barrels. vol = Volume�1; 000; 000

Volume_proven: Part of Volume represented by the categories of reserves which are well explored and

precisely estimated (categories A,B,C1,C2), in million barrels of equivalent fuel.

Volume_probable: Part of Volume represented by the categories of reserves which are classi�ed as

promising (categories C3, D1, D2), in million barrels of equivalent fuel.

Volume_proven+Volume_probable = Volume

Temperature: Mean January temperature in the region, in Fahrenheit.

Warm: Dummy for mean January temperature in the region of a �eld being higher then 0�F .

RailroadDistance: Distance from the �eld to the nearest railroad, in kilometers.

OilpipeDistance: Distance to the nearest oil pipeline, in kilometers

Gaspipe Distance: Distance to the nearest gas pipeline, in kilometers.

Oil�eld Distance: The smallest distance to the nearest oil �eld under exploration, in kilometers.

Gas�eldDistance The smallest distance to the nearest gas �eld under exploration, in kilometers.

Bidder Proximity Dummy: Equal to one if there is a shill pair (as de�ned in Section 8) found in an

auction, zero otherwise.

OilDummy: Equal to one when there is only oil on the tract, or oil and gas but the quantity of oil is at

least four times as large as the quantity of gas as measured in equivalent fuel units; zero otherwise.

27

GasDummy: Equal to one when there is only gas on the tract, or gas and oil but the quantity of gas is

at least four times as large as the quantity of oil as measured in equivalent fuel units; zero otherwise.

Summary statistics and frequency counts are below.

Table 4. Summary Statistics (N = 172)

Variable Mean Standard Deviation Minimum Maximum

N 4.2 1.6 3 10

Naff 3.8 1.5 2 10

Final Price (in ths. $) 15,553 32077 19 251887

Reserve Price (in ths. $) 2,500 5149 14 42757

Volume (in mln. barrels) 43.7 91.2 0.18 807.2

Temperature (in �F ) 1 11.3 -45 30

RailroadDistance 139 182 0 800

OilpipeDistance 80 140 0 800

GaspipeDistance 130 215 0 800

Oil�eldDistance 65 123 5 685

Gas�eldDistance 73 102 5 800

28

Table 5. Frequency Counts (N = 172)

Variable Frequency

N:

3 81

4 37

5 27

6 12

7 7

8 1

9 4

10 3

Sale Year:

2004 14

2005 36

2006 76

2007 39

2008 7

Warm:

0 61

1 111

Bidder proximity dummy:

0 114

1 58

OilDummy:

0 43

1 129

GasDummy:

0 100

1 72

29

Tables and Figures.

Table 6. 3B+ Auctions with A¢ liate or Shill Pair Found vs. Those Without

Auctions Number Average Volume # of Bidders F/R

With Shill or A¢ liated Pair 110 6.4 4.4 10.3

Without 65 5.7 3.9 15

Table 7. Cumulative Distribution of Auctions Won

Auctions Won # of Companies Auctions Won, Cumulative

0 257 100%

1 155 100%

2 51 71%

3 18 52%

4 9 43%

5 7 36%

6 1 29%

7 3 28%

8 1 24%

9 1 23%

10 1 21%

12 1 19%

14 1 17%

19 1 15%

28 1 11%

32 1 6%

30

Table 8. Market vs. Group of 10 Biggest Companies

Group of 10 All companies Share of Group of 10

# of Auctions Won 146 541 27%

Money Paid, $ mln. 2344 4509 52%

Volume of Reserves Won, mln. ton 9103 17065 53%

F/R 6.1 7.3

Price/barrel 19 27

Average Field Volume, mln. ton 8.3 4.7

Average # of bidders 3.5 3.2

Table 9. Market vs. Group of 10+Their A¢ liates

Group of 10 All companies Share of G10+A¢ liates

# of Auctions Won 209 541 39%

Money Paid, $ mln. 3169 4509 70%

Volume of Reserves Won, mln. ton 11865 17065 70%

F/R 8.3 7.3

Price/barrel 19 27

Average Field Volume, mln. ton 8.1 4.7

Average # of bidders 3.4 3.2

31

References

Athey, S., Haile, P., 2002. Identi�cation of Standard Auction Models. Econometrica, 70, 2107-2140.

Bajari, P., Hortacsu, A., 2003. Winner�s Curse, Reserve Prices, and Endogenous Entry: Empirical Insights

from eBay Auctions. RAND Journal of Economics, 34, 329-355.

Bajari, P., and Ye, L., 2003. Deciding between competition and collusion. Review of Economics and

Statistics 85�4, 971�989.

Bajari, P., Summers, G., 2002. Detecting collusion in procurement auctions. Antitrust Law Journal 70�1,

143�170.

Baldwin, L., Marshall, R., Richard, J-F., 1997 Bidder collusion at forest service timber auctions. Journal

of Political Economy 105�4, 657�699.

Donald, S., Paarsch, H., 1996. Identi�cation, estimation and testing in parametric empirical models of

auctions within the independent private values paradigm. Econometric Theory 12, 517-567.

Gilley, O., Karels, G., 1981. The Competitive E¤ect in Bonus Bidding: New Evidence. Bell Journal of

Economics, 12, 637-648.

Graham, Daniel., Marshall, R., 1987. Collusive Bidder Behavior at Single-Object Second-Price and English

Auctions.Journal of Political Economy, 95, 1217-39.

Jofre-Bonet, M., Pesendorfer, M., 2003. Estimation of a Dynamic Auction Game. Econometrica, 71, 1443-

1489.

Hendricks, K., Boudreau, B., Porter, R., 1987. Information, returns, and bidding behavior in OCS auctions:

1954-1969. The Journal of Industrial Economics XXXV, 517-542.

Hendricks, K., Porter, R., 1988. An empirical study of an auction with asymmetric information. American

Economic Review 78, 865�883.

Hendricks, K., Porter, R., 1989. Collusion in Auctions. Annales d�Economie et de Statistique, nos., 217-30.

Li, T., Perrigne, I., Vuong, Q., 1999. Structural estimation of the a¢ liated private value model with an

application to OCS auctions. Working paper, University of Southern California.

Li, T., Perrigne, I., Vuong, Q., 2000. Conditionally Independent Private Information in OCS Wildcat

Auctions. Journal of Econometrics, 98, 129-161.

Mailath, G., Zemsky, P., Collusion in Second Price Auctions with Heterogeneous Bidders. Games and

Economic Behavior, 3 (1991), 467�486.

Marshall, R., Marx., 2008. The vulnerability of auctions to bidder collusion. Forthcoming in the Quarterly

Journal of Economics

32

Marshall, R., Marx., 2007. Bidder collusion. Journal of Economic Theory 133 (2007), 374�402.

Marshall, R., Meurer, M., Richard,J-F., Stromquist, W., Numerical Analysis of Asymmetric First Price

Auctions. Games and Econ. Behavior 7 (September 1994): 193-220.

Milgrom, P., Weber, R., 1982. A theory of auctions and competitive bidding. Econometrica 50 (5), 1089-

1122.

Paarsch, H. 1991. Empirical Models of Auctions and an Application to British Columbian Timber. Univer-

sity of Western Ontario, Department of Economics, Technical Report 91-19.

Palshin, I., et al, 2007. Atlas. Fuel and Energy Complex of Russia. INCOTEC.

Pinkse, J., Tan, G., 2002. The A¢ liation E¤ect in First-Price Auctions. Mimeo.

Porter, R., Zona, D., 1993. Detection of Bid Rigging in Procurement Auctions. Journal of Political

Economy, 101, 518-38.

Porter, R., Zona, D., 1997. Ohio School Milk Markets: An Analysis of Bidding. Rand Journal of Economics,

30(2), 263-88.

Riley, J., Samuelson, W., 1981. Optimal Auctions. American Economic Review 71, 381-92.

Trutnev, Y., 2008. Interview of the Minister of Natural Reserves. Vedomosti, March.

Vickrey, W., Counterspeculation and Competitive Sealed Tenders. Journal of Finance, 16 (1961), 8�37.

Wehbe, Y., Maggs, P., 2007. Russian Oil and Gas Market Report. UK Trade and Investment.

http://www.econ.psu.edu/CAPCP/RussianData

http://www.ifs.ru/en/

http://www.oilukr.com/serv05.htm

33