Вы находитесь на странице: 1из 68

The Relative Ecacy of Price Announcements and Express Communication for Collusion:

Experimental Findings

Joseph E. Harrington, Jr. Dept of Business Economics & Public Policy The Wharton School University of Pennsylvania Philadelphia, PA USA harrij@wharton.upenn.edu

Roberto Hernan-Gonzalez Dept of Economic Theory and History Universidad de Granada Granada, Spain roberto.hernangonzalez@gmail.com

14 August 2013


Praveen Kujal Middlesex University London, U.K. pkujal@gmail.com

We appreciate the comments of Cedric Argenton, Hans-Theo Normann, and participants at "Cartels: A Conference in Honor of Robert F. Lanzillotti" (U. of Florida) and the 3rd MaCCI Summer Institute in Competition Policy (Edesheim, Germany). This research was partly conducted while the rst author was visiting the Universidad Carlos III de Madrid as a Cátedras de Excelen- cia, and he would like to thank Banco Santander for funding and the faculty for providing a most collegial and stimulating environment. For nancial support, the rst author thanks the National Science Foundation (SES-1148129) and the second and third authors thank the Spanish Govern- ment (DGICYT-2012/00103/00) and the International Foundation for Research in Experimental Economics (IFREE). The experiments were run when the second and third authors were visiting the Economic Science Institute at Chapman University.


Abstract: Collusion is when rms coordinate on suppressing competition. Coordina- tion typically requires that rms communicate in some manner. This study performs an experimental analysis to determine what modes of communications are able to produce and sustain collusion and how the ecacy of communication depends on the number of rms and rm heterogeneity. We consider two dierent communication treatments: non-binding price announcements and unrestricted written communica- tion. Our main ndings are that price announcements allow subjects to coordinate on a high price but only under duopoly and when rms are symmetric. While price announcements do result in higher prices when subjects are asymmetric, there is little evidence that they are coordinating their behavior. When subjects are allowed to en- gage in unrestricted communication, coordination on high prices occurs whether they are symmetric or asymmetric. By analyzing chat, we are also able to learn how they are coordinating which helps shed light on why collusion with price announcements is relatively ineective when rms are dierent.


1 Introduction

For rms to successfully collude, they must coordinate their behavior, and coordi- nation requires some form of communication. In practice, this communication can involve tacking on a few digits to a multi-million dollar bid (FCC spectrum auction) or announcing future intended prices (airlines) or unilaterally announcing a pricing strategy (truck rental) or sitting in a hotel room and talking about prices and sales quotas (lysine). While the last mode of communication is presumably the most eec- tive, it is also the most clearly unlawful. Firms interested in jointly raising prices then face a tension in that communication which is more likely to result in coordination may also be more likely to result in prosecution. Hence, they may choose to more indirectly communicate when it is sucient to produce at least some collusion. This trade-oraises two questions that we will examine here. First, what are the various forms of communication that can produce coordinated collusive outcomes? In particular, how indirect can communication be and still be reasonably eective? This question is central to antitrust and competition law and, in spite of a legion of legal cases that speak to what practices are and are not lawful, there remains a large gray area where legality is unclear. Second, how does the answer to the rst question depend on the specics of the market? These questions are notoriously dicult to examine theoretically because the equi- librium framework cannot speak to the issue of how rms coordinate in moving from one equilibrium to another which is exactly what is at issue here: What forms of communication will result in rms coordinating a move from a static equilibrium with competitive prices to a dynamic equilibrium with supracompetitive prices? Ex- perimental methods oer a comparative advantage in that subjects engage in exactly the dynamic process of coordination that we are trying to understand. While the subjects are college students and not managers - and thus extrapolating from exper- iments to market behavior is always a precarious leap - experimental methods have more promise than other methods for shedding light on the eectiveness of various communication practices in producing collusion. The specic form of those two questions are addressed here are as follows. In practice, two commonly observed methods of communication for coordinating rm behavior are advance price announcements (as arose in the ATPCO airlines cases) and unrestricted communication using natural language (as practiced by all hard core cartels; for example, lysine, vitamins, and ne arts auction houses). To assess the relative ecacy of dierent modes of communication, the research plan is to compare outcomes when sellers can make price announcements with when they cannot, and to compare unrestricted communication (through online chat) with price announce- ments. When are price announcements eective at producing collusion? When is unrestricted communication particularly eective in producing collusion relative to price announcements? Answers to these questions will shed light on when we can ex- pect rms to engage in the most egregious form of collusion - involving unrestricted


communication - and when they will instead choose less express methods. In con- sidering the relative ecacy of these dierent forms of communication, the primary variation in market structure is the extent of rm heterogeneity though the number of sellers is also examined. While unrestricted communication is surely expected to be more eective than price announcements, less clear is how the incremental value of unrestricted communication depends on rm heterogeneity. Our main ndings are that rms are able to coordinate on a high price with price announcements but only for duopoly and when rms are symmetric. While price announcements do result in higher prices for an asymmetric duopoly, there is only weak evidence that they are coordinating their behavior. When rms engage in unrestricted communication, coordination on high prices occurs whether rms are symmetric or asymmetric and regardless of the number of rms. We also investigate how they are coordinating with unrestricted communication which helps explain why collusion with price announcements is relatively ineective when rms are dierent. Section 2 provides a review of experimental work pertinent to the current study; specically, oligopoly experiments for which sellers can communicate and those for which sellers are heterogeneous. For the reader who just wants the executive summary of that literature and how it relates to the current study, s/he can jump to Section 2.3. Section 3 describes the theoretical model underlying the experiment and derives equilibrium predictions, while the experimental design is provided in Section 4. The results from the experiments are described and discussed in Section 5.

2 Literature Review

Pertinent to this study, we will review oligopoly experiments that speak to either of two questions. First, in an oligopoly setting, what is the impact of communica- tion among rms about price or quantity intentions on the frequency and extent of supracompetitive outcomes? Second, what is the impact of rm heterogeneity on the frequency and extent of supracompetitive outcomes? Our review will begin with experiments pertinent to addressing the rst question - which is quite voluminous - and then discuss experiments that address the second question - which is relatively sparse. There are no experiments that address the interaction of communication and rm heterogeneity on outcomes, which is the primary focus of the current study. 1 Re- cent reviews of the literature on communication of intentions in an oligopoly include Cason (2008), Normann (2008), Haan, Scho onbeek, and Winkel (2009), and Potters


1 With the exception of Harstad, Martin and Norman (1998), all studies that allow for communi- cations among subjects have assumed identical rms. However, that study does not vary the extent of asymmetries and thus cannot address the questions of the current study.



Communication of Intentions in Oligopoly Experiments

2.1.1 Experimental Design

Experiments that allow for communication of intentions dier in many features. These dierences pertain to the communication protocol (that is, exactly how subjects are allowed to communicate), market features, and model features. Market features are dened to be those that correspond to sources of variation in real world markets and thus one could say that two experiment that dier in terms of market features actually pertain to dierent real world markets. In contrast, model features concern articial elements of a model that, in practice, do not pertain to sources of variation in real world markets. In terms of market features, experiments have assumed subjects engage in a price game (that is, prices are simultaneously selected and rms produce to meet demand), a quantity game (and price is set to equate supply and demand), or a price game with a maximal number of units (that is, a rm produces to meet its demand up to the maximal quantity it selected). When price is selected, the market institution is almost always posted price but there are a few experiments that have specied a double oral auction. 2 While most experiments assume two sellers (or subjects), some allow for more (as many as eight) and may examine the impact of the number of sell- ers. Products may be either homogeneous or dierentiated, and market demand may be perfectly inelastic or responsive to price (using a linear specication). Subjects interact in a single market with the exception of one study that has subjects compete in three markets in order to create multi-market interaction (Cason and Davis, 1995). Most experiments assume complete information (all prot functions are known to all sellers) but a few allow for private information; for example, a seller’s cost is private information and a buyer’s valuation is private information. 3 Finally, a number of experiments have investigated the eect of competition policy for cartels by intro- ducing probabilistic penalties if subjects choose to communicate, as well as leniency programs (in which case a subject can avoid a penalty by turning in the other subjects involved in the communication). Turning to model features, subjects may be given either a prot function or a prot table. A prot function means a mathematical formula for calculating a subject’s prot once inserting the actions (price or quantity) of all subjects; while a prot table (for a two subject experiment) is a matrix in which rows are a subject’s own action, columns are the other subject’s action, and an entry in a cell is a subject’s prot. Experiments may also vary in terms of the size of the choice set. They range from as few as two actions or prices (Andersson and Wengström, 2010) to as many as 101 prices (Fonseca and Normann, 2012). Demand can be either automated or live.

2 A double oral auction has sellers post prices at which they are willing to sell, buyers post prices at which they are willing to buy, any seller can accept an outstanding buyer o er, and any buyer can accept an outstanding seller oer. 3 In that case, communication can extend beyond intentions though it is still cheap talk.


Most common is an automated specication in that there is a deterministic demand function that maps from prices for rms into the number of units for each rm. In

a few experiments, buyers are live in that they are also subjects. There are either 3 or 4 buyers who are endowed with a valuation and receive a monetary payment that

is proportional to the surplus received on a purchase; that is, the valuation less the price paid. 4 Finally, a key model feature to any experiment designed to explore collusion is how the horizon is modelled. Theory predicts that collusion will not emerge in a nitely repeated game in which the stage game has a unique Nash equilibrium. Most experiments have assumed an indenite horizon whereby the experiment runs for some number of periods for sure (say, 20) after which, in each period, a random device determines whether the experiment goes another period (for example, a role of a die whereby the experiments end when 1 is rolled). That random process is enacted in each period until termination occurs. However, some experiments assume

a long but nite number of periods. Though theory does not predict any collusion

in those experiments, there is ample experimental evidence that cooperation does emerge. With that model feature, it is common to exclude the last few periods in order to prevent conclusions being contaminated by endgame eects. An alternative to having an explicitly repeated setting is to simulate it through

a two-stage design, as is done in Andersson and Wengström (2010) and Cooper and

Kühn (2011). In the rst stage of Cooper and Kühn (2011), subjects play a price game where there are three prices (Low, Medium, High). The unique Nash equilibrium is both rms price Low and joint prot is higher if both were to price Medium and is higher yet if both were to price High. After playing that game (just once), subjects observe the stage 1 choices and then place a coordination game in stage 2. This coordination game has three actions (L, M, H) and all are symmetric Nash equilibria. These equilibria are Pareto-ranked and the payofrom (L,L) is (for some discount factor) equal to the present value of playing (Low, Low) forever from the stage 1 game. Similarly, the payofrom (M,M) equals the present value of playing (Medium, Medium) forever and the payofrom (H,H) equals the payofrom playing (High, High) forever. Thus, if players go to (L,L) in stage 2, it can be interpreted as the grim punishment (static Nash equilibrium forever), while going to (M,M) or (H,H) is like collusion ad innitum. 5 It is an open question whether results for this two-stage

4 While this is listed as a model feature, one could argue that it is a market feature. Perhaps retail markets with many small buyers can be thought as having automated buyers, and some intermediate goods markets with a few buyers is better described by live buyers. However, the more substantive dierence is whether buyers can make counter-oers or must take price as given. Experiments with double oral auctions (all of which involve live buyers) are better models of intermediate goods markets with large buyers, while those with posted price (sellers make take it or leave it oers) are a better representation of markets with many small buyers. 5 Analogously, Andersson and Wengström (2010) have two actions (L, H) in the stage 2 game and both are symmetric Nash equilibria which are Pareto-ranked with the same relationship between (L,L) and (Low,Low) forever and (H,H) and (High,High) forever.


design are similar to those for an indenite horizon. We conclude this summary of experimental design by describing the many dierent communication protocols that have been used. We have partitioned them into four categories: Simple Price Announcement, Iterative Price Announcement, Strategy Announcement, and Chat. To be clear, in all of these cases the announcements are non-binding. A Simple Price Announcement (SPA) protocol involves one or more subjects announcing a price and possibly subjects responding to an announcement by arming or rejecting it. The canonical case of a SPA protocol is when all subjects announce a price each period prior to then selecting a price (for example, this is one of the protocols in Cooper and Kühn, 2011). The timing is a bit dierent in Cason (1995) and Cason and Davis (1995) where, in one treatment (referred to as Discrete), each subject chooses a price (and maximum quantity) for the current period and announces a price (and maximum quantity) for the ensuing period. In their Continuous treatment, subjects can announce as many prices as they like during a one (or two) minute period where the last price announced is the actual price charged. 6 Harstad et al (1998) also allow for multiple price announcements but do not place an upper bound on the length of the announcement phase; it goes until there has not been a message for some specied length of time (initially, two minutes and then 30 seconds in later periods). In addition to a treatment in which prices are announced, they have a more structured communication treatment in which price changes are announced (relative to the preceding period’s price) and where subjects can respond with "match" or "no match" to any such announcement (again, non- binding). A relatively structured protocol is also used in Holt and Davis (1990) where one selected subject makes an announcement of the form: "$_ is an appropriate price for the market for this period." and the other two subjects simultaneously and independently reply with either "agree," "disagree and price should be lower," or "disagree and price should be higher." Prices are then selected. An Iterative Price Announcement (IPA) protocol has the property that the com- munication phase has multi-stages and the messages that were sent in an earlier stage restrict the messages that can be sent in the current stage. In Bigoni et al (2012), two subjects are asked to choose a "minimum acceptable price" from the price set {01  12} which is then shared with the other subject. 7 In round 2 of the commu- nication phase, subjects again choose a "minimum acceptable price" but where the selected price message must be at least as great as the lowest price message in the previous round. This process applies to all rounds in the communication phase. Thus, the message space weakly contracts during the communication phase which lasts for 30 seconds, after which price is chosen (unconstrained by what was conveyed dur-

6 This is the one protocol in which messages are not purely non-binding in that the nal message is the actual price. However, if it is clear what is the last moment at which a price can be selected then that "last moment" is the "transaction price" phase and all preceding "moments" comprise the "price announcement" phase. 7 Also see Bigoni et al (2013).


ing the communication phase). Hinloopen and Soetevent (2008) engage in a similar process but, in addition, have a "maximum acceptable price" and consider the inter- section of subjects’ sets rather than the union as in Bigoni et al (2012). Specically, three subjects announce a lower and upper bound to price (which could be the same

so that it is a single price). If there is a non-empty intersection of these three sets then there is a second round in which again subjects announce lower and upper bounds but now from that intersection. This process continues until there is a single price or the intersection is empty or a minute transpires, whichever occurs rst. In both of these studies, the purpose for restricting communication in this manner is to make the environment more amenable to producing collusion. As their primary objective

is to investigate the eect of nes and leniency programs, if the experiment produced

very little collusion then it would be dicult to assess the eect of these programs

with a reasonable number of runs. However, one wonders to what extent the proper- ties of collusive outcomes are inuenced by an articially constrained communication process. The Strategy Announcement (SA) protocol is where subjects announce not a price but a strategy for the game or, more generally, some set of contingency plans. This protocol is only used in Andersson and Wengström (2010) and Cooper and Kühn (2011) whose simple two-stage game structure with two and three actions, respec- tively, makes for a simple strategy space. Andersson and Wengström (2010) assume

a subject announces prior to stage 1 what action it will take in the stage 2 game

depending on the action chosen by the other subject in the stage 1 game. In Cooper and Kühn (2011), strategy announcements prior to the stage 1 game are contingent on both subjects’ actions in stage 1; specically, subjects are prompted with: "if we choose the preceding [actions I think should be chosen] in period 1, I think we should choose the following in period 2" and "if we DO NOT choose the preceding [actions I think should be chosen] in period 1, I think we should choose the following in period 2." (They refer to a period what we are calling a stage.) Another distinction be- tween these two studies is that strategy announcements are simultaneous in Cooper

and Kühn (2011), while Andersson and Wengström (2010) impose a sequential struc- ture in which one subject is randomly selected to send an initial message which the other subject reads before sending a message (that sequential structure is enacted only when both subjects choose to send messages). They rationalize this sequential structure on the grounds that it helps to avoid miscoordination. Finally, the Chat protocol allows for either oral or written communication using natural language with only minimal restrictions such as no profane language, not oering side payments, and not making threats. Friedman (1967), which is the rst oligopoly experiment to allow for unrestricted communication, had a limited number of rounds of communication by virtue of the need to hand carry messages: 8

A period consisted of each subject sending two messages to the other,

8 Friedman (1967), p. 385.


then each choosing his price. One subject was designated to send the rst message. The second would read and reply. A second round of messages would ensue after which the rst subject would choose a price, then the second would do so. The messages were carried back and forth by a project assistant. After collecting both prices, he would inform each of the price charged by the other; then the next period would begin. If someone wanted to send no message, he was instructed to write an X.

Some later studies - such as Issac and Plott (1981) and Issac, Ramey, and Williams (1984) - moved to face-to-face communication. In those experiments, the subjects gathered at a table between periods to engage in a time-constrained conversation (four minutes in Issac et al, 1984). Subsequent studies have made the standard chat protocol to be written and anonymity-preserving as in Friedman (1967) though now it is conducted through computers in the form of instant messaging with the requirement that subjects do not provide identifying information. Typically, there is an upper bound on the chat phase of 1-3 minutes which generally proves ample as revealed by a content analysis of chat. A unique feature of Anderson and Wengström (2007) and Cooper and Kühn (2011) - again associated with their two-stage design - is that they consider two Chat protocols. One protocol only allows communication prior to the stage 1 game; thus, it provides an opportunity to coordinate. The second

protocol also allows for communication prior to the stage 2 game. Since the outcome of the stage 2 game is intended to correspond to continued cooperation or a punishment (in the event of non-compliance with the collusive outcome), communication prior to

it allows for the possibility of renegotiation which can then undermine the credibility

of any punishment discussed during the pre-play communication phase of the stage 1 game. An additional dimension to any communication protocol is whether communica- tion is free or costly. Most experiments assume it is free but a few involve a direct or indirect cost. Using a SPA protocol, Anderson and Wengström (2007) assumes one subject decides whether to propose a price and, if so, what price to propose; any proposal incurs a cost to that subject. In the event there is a price proposal, the

other subject can, at a cost, either accept or reject the proposal, or it can not reply which incurs no cost. Apesteguia, Dufwenberg, and Selten (2007), Hinloopen and Soetevent (2008), Dijkstra, Haan, and Scho onbeek (2011), and Bigoni et al (2012) all assume an indirect cost. In those experiments, a communication phase is preceded by

a pre-communication stage in which all subjects decide whether or not to engage in

the communication phase. If and only if all subjects choose to do so is there a com-

munication phase, which involves an IPA protocol in Hinloopen and Soetevent (2008) and Bigoni et al (2012) and a Chat protocol in Apesteguia, Dufwenberg, and Selten (2007) and Dijkstra, Haan, and Schoonbeek (2011). Furthermore, if all subjects have chosen to communicate then they all become liable for probabilistic monetary penal- ties. Thus, the cost is not per message - as in Anderson and Wengström (2007) - but rather is an expected xed cost associated with the decision to communicate. These


studies are interested in exploring the eect of competition policy - in particular, cor- porate leniency programs - and thus have created a model encompassing the illegality of communication with its associated penalties. 9

2.1.2 Results

Before describing results for when subjects can communicate, let us begin by stating a robust nding from the voluminous literature on oligopoly experiments without communication. These studies consistently nd that prices are often above static Nash equilibrium levels but only if there are two sellers. The evidence is overwhelming that duopolies are able to tacitly collude. A recent study (Rojas, 2012) even showed that tacit collusion can occur when demand is changing over time, and subjects prove quite sophisticated in adjusting their behavior to the current state of demand. 10 Summarizing the literature on collusion without communication, supracompetitive outcomes are common when there are two subjects, very rare with three subjects, and non-existent with four or more subjects; see, for example, Huck, Normann, and Oechssler (2004) and the review paper of Engel (2007). 11 In summarizing oligopoly experiments with communication that use a SPA proto- col, Cason (2008) accurately states: "Although price signaling often increases trans- action prices, this increase is very often temporary. [Steady-state] behavior may be unaected by non-binding price signaling in many environments." Allowing subjects to announce prices does raise the average price but the eect largely occurs early on as prices eventually decline. By the end of the experiment, price has often returned to the no communication benchmark level though may also remain above it (which could be an artifact of ending the experiment before this declining price path has con- verged). Subsequently, there have been only two papers using a SPA protocol; one of which conrms this nding and the other does not (but, as will be described, makes

9 Apesteguia et al (2007) is singular in that it does not provide repetition of interaction (or simulate it as in, say, Cooper and Kühn, 2011) and thus rely on collusion emerging in a one-period structure. Specically, they assume sellers simultaneously decide whether to form a cartel and, if and only if all sellers choose to form a cartel, then they have unrestricted chat. Sellers then choose price. If they did form a cartel, sellers then simultaneously decide whether to report the presence of a cartel. 10 It is important to note that this occurs in a simple environment with only three actions in the stage game. 11 Singular in the literature on oligopoly experiments, Friedman et al (2012) nds collusive out- comes regularly emerge with three subjects. They consider a repeated quantity game in which subjects know the game is symmetric but do not know payofunctions. A unique feature of the experimental design is that periods are only 4 seconds in length and paired interaction occurs for 400 periods; in comparison, 40 periods is a long horizon in other experiments. Consistent with previous experiments with low information, they nd that quantities lie between static Nash equi- librium (Cournot) and Walrasian (perfect competition) levels for the rst 25 periods. However, when interactions get into the range of 50 periods, quantities start falling and eventually are close to monopoly levels for duopolies and lie between monopoly and Cournot levels for triopolies. Thus, quite contrary to preceding work, collusive outcomes arise and persist with three sellers.


a substantive change in the structure). Recalling the two-stage design of Cooper and Kühn (2011), they nd this pattern either when subjects announce stage 1 prices or announce strategies (prices for stage 1 and contingent actions for stage 2). With the rst protocol, around 37% of subjects choose the High price in the rst round (and recall that there are only three prices) - compared to 15% when there is no communication - but that steadily falls to where it is under 5% after 8 rounds of the experiment which is not far from under 2% for the no communication treatment. 12 The High price is chosen more frequently when subjects announce strategies but there is still the same deterioration of collusion though around 12% of subjects are eventually choosing High, which is above that for when subjects are not allowed to communicate. In sum, Cooper and Kühn (2011) nds, for a quite dierent experi- mental design, that experience tends to undermine collusion when subjects engage in very restricted communication in the form of price announcements. An exception to that nding is Anderson and Wengström (2007). Recall that they assume one subject is randomly selected to propose a non-binding price (or say nothing) and the other subject, in the event there is a proposed price, either accepts or rejects or does not respond. These messages are non-binding and are costly. For the no cost treatment, the usual pattern is observed with prices starting high and then declining. 13 When messages are costly (and there is a low cost and high cost treatment), prices tend to start at around the same level as the zero cost treatment but, by the end of the experiment, are not lower. When messages have a low cost, prices initially decline but then recover. With the high cost treatment, prices initially rise, then decline, and recover. It is dicult to interpret the pattern but it is clear that making messages costly can allow price announcements to sustain prices higher for a longer time compared to when messages are free. This result seems consistent with the theory of forward induction in that those subjects who are more condent about supracompetitive prices being coordinated upon will incur the cost of sending messages. Thus, the messages themselves are more informative. The general nding for experiments with a SPA protocol are roughly conrmed by the two studies using an IPA protocol. Conditional on all subjects choosing to communicate (so that there is a communication phase), both Hinloopen and Soetevent (2008) and Bigoni et al (2012) nd that prices start high and tend to decline over time. A comparison with the prices for the no communication treatment is a bit precarious because the communication treatment also comes with probabilistic penalties and thus costly communication (though of a very dierent variety than Anderson and

12 It is important to point out that what "additional rounds" means in Cooper and Kühn (2011) is dierent from other experiments with a repeated game structure. In other experiments, it means more periods in which the same two subjects choose price. In Cooper and Kühn (2011), each round involves the two-stage game which is designed to simulate an indenite horizon game. Thus, multiple rounds means multiple inde nite horizon games where, with each round, subjects are randomly matched. The two features are related in that both involve more experience. 13 As there was no treatment without communication, we do not know whether price converged to the level for when communication is not allowed.


Wengström, 2007). Nevertheless, prices do seem to converge to the levels for the no communication treatment in Hinloopen and Soetevent (2008) and in Bigoni et al (2012) except for one competition policy treatment. 14 Finally, we come to results for the Chat protocol. Quite contrary to the other pro- tocols, unrestricted communication between the subjects is shown to be extremely eective at producing successful collusion. Supracompetitive prices are frequent, are at a high level (though often below the monopoly price), and persist over time. This nding is true for almost all of the studies including Friedman (1967), Issac et al (1984), Davis and Holt (1998), Cooper and Kühn (2011), Dijkstra et al (2011), and Fonseca and Normann (2012). While having more sellers tends to reduce prices - just as with the no communication treatment - prices are still well above "no communi- cation" levels even when there are as many as eight sellers (Fonseca and Normann,


are as many as eight sellers (Fonseca and Normann, 2012). Figure 1 from Fonseca and Normann

Figure 1 from Fonseca and Normann (2012)

To observe the striking eect of unrestricted communication, consider Figure 1 from Fonseca and Normann (2012) which reports average price across experiments with communication (Talk) and without communication (No Talk) and when the number of sellers ranges from 2 to 4 to 6 to 8. Goods are identical, the monopoly price is 100, and the static Nash equilibrium price is 0. Without communication, prices are very close to the competitive price when there are 4, 6, and 8 subjects but, when there are two sellers, prices are half way between the competitive and monopoly price. This is consistent with the body of work on oligopoly experiments without communication which nds that tacit collusion is common but limited to

14 The latter statement is based on private communication with Maria Bigoni (28 December 2012). When a rm that reports a cartel is given a reward (in addition to avoiding the penalty) then prices remain above levels for the no communication treatment.


the case of two sellers. When subjects can engage in unrestricted chat, prices are vastly higher. Even when there are eight sellers, price has risen more than half of the dierence between the monopoly price and the price without communication. 15 Davis and Wilson (2002) use a common metric to measure the impact of commu- nication which is the Monopoly Eectiveness Index (MEI). MEI is dened as (realized prot - competitive prot)/(monopoly prot - competitive prot), where competitive prot is that associated with the static Nash equilibrium. MEI measures the fraction of monopoly prot that is captured. By the end of the sessions in their study, the MEI was .7 higher in the communication treatment than when communication was not allowed. This large eect of communication is also reected in the commonality of prices. Without communication, identical prices from all four sellers are observed in only 3 out of 40 periods (where data is used from the last 10 periods for the four sessions run). With communication, prices are identical in 9 out of 10 periods in 3 out of 4 sessions.

are identical in 9 out of 10 periods in 3 out of 4 sessions. Figure 1

Figure 1 from Cooper and Kühn (2011)

To compare Chat with price and strategy announcements (as well as when there is no communication), Figure 1 from Cooper and Kühn (2011) is reproduced. Recall that they have a two-stage design that simulates a repeated game setting. In each round, subjects are randomly matched to play the two-stage game. Results are re- ported for rounds 10-20 where all subjects were engaged in the no communication

15 Fonseca and Norman (2012) interestingly nd that the incremental value to communicating is highest with 4 sellers. The incremental value is measured by the dierence in price (and prot) between the communication treatment and the no communication treatment. Though the price with communication is highest with two subjects, it is also relatively high without communication. However, when there are 4, 6, or 8 subjects, the no communication price is very close to the static Nash equilibrium price. Thus, the incremental gain from chat is higher with 4 subjects than with 2 subjects. It proves to be higher with 4 than with 6 or 8 because the supracompetitive price with communication declines reasonably fast with more subjects.


(N) treatment through round 10. Starting with round 11, new treatments were in- troduced: price announcements (P1), strategy announcements (P1C), chat prior to stage 1 (Pchat), and chat prior to both stages 1 and 2 (Rchat). For each round, the percentage of subjects who selected the High price in stage 1 (or period 1) is reported. Price announcements initially involve higher prices (relative to the no communication treatment) but then it steadily declines. When subjects can chat, prices start high and generally remain high. There are two studies - Issac and Plott (1981) and a particular treatment in Davis and Holt (1998) - that do not nd that unrestricted chat produces strong and stable collusive outcomes, but both of them have unique features that a priori one would suspect would make collusion more dicult. Both of these studies assumed live buyers; for example, Davis and Holt (1998) had three sellers and three buyers. In their treatment of three minutes of open chat (after which sellers would post prices), Davis and Holt (1998) found sellers to price close to the monopoly level. Now modify this design so that, after prices are posted, a seller can oer a discount to a buyer and this discount is never observed by the other sellers. For 5 out of 6 groups, any initial collusion collapses and prices end up approaching the competitive level. While posted (or list) prices are close to the monopoly level, transaction prices are much lower. That collusion is more dicult with those hidden discounts is not surprising as the environment is now one of imperfect monitoring in that a seller’s past prices are not publicly observed by all sellers. 16 Issac and Plott (1981) assume a double oral auction so both sellers and buyers make oers. Successful seller collusion was not observed in that setting. Issac et al (1984) used the same experimental design and compared prices with the posted price format (only sellers make oers which buyers accept or reject) with those for the double oral auction and nd that prices are higher under the posted price format and that chat is eective at producing high and sustained prices. It would then seem that the active role by buyers in the double oral auction format made collusion by sellers dicult in spite of the availability of unrestricted communication. On a related issue, Andersson and Wengström (2010) and Cooper and Kühn (2011) explore whether the opportunity to communicate before a possible punishment phase makes collusion more dicult. This is a test of the theory that renegotiation makes punishments less credible which then undermines the collusion that is erected upon a threatened punishment. Interestingly, their ndings are in conict. Andersson and Wengström (2010) compares outcomes for two communication protocols: 1) pre-play communication prior to stage 1 in which each subject announces the action s/he will take in stage 2 depending on what action the other subject chose in stage 1; and 2) the preceding protocol augmented with pre-play communication prior to stage 2 in

16 They also nd that collusion is more eective when the treatment is altered to allow sellers’ past quantities sold to be public information. Now, cheating through private discounts can be partially monitored by observing a rm’s quantity. As a result, 3 out of 6 groups had near-monopoly prices but prices were still lower than in the original communication treatment without discounts.


which each subject announces an action for stage 2. They nd that the frequency with which the high price is chosen in stage 1 is greater when subjects cannot communicate prior to stage 2. While this result is consistent with the theory that the possibility of renegotiating a punishment undermines collusion, it is statistically weak. In contrast, Cooper and Kühn (2011) nd that communication prior to stage 2 does not make collusion more dicult and, in fact, it seems to enhance collusion (compare Pchat and Rchat in Figure 1 from Cooper and Kühn, 2011). As predicted by theory, the option to communicate prior to performing a punishment does result in fewer punishments but it does not harm collusion. It is notable that, contrary to the use of a SA protocol as in Andersson and Wengström (2010), Cooper and Kühn (2011) allow for unrestricted chat. Andersson and Wengström (2010) conjecture these dierent communication protocols is the source of the dierence in results because "even though renegotiation is protable in terms of actions," a player might be deterred by "the perceived cost of a verbal punishment." (p. 16). This does raise an important design issue. While unrestricted chat is more descriptively realistic of the type of communication that occurs in cartels, it may also insert forces (such as chastisement and guilt) that could be irrelevant for market settings.

2.2 Firm Heterogeneity

Within the repeated game oligopoly setting without communication, experiments exploring the eect of rm heterogeneity have focused on dierences in costs. Cost

heterogeneity may be modeled in terms of dierences in constant marginal cost, the steepness of a rising marginal cost curve, or dierences in capacities (with a common marginal cost below capacity). Though the number of studies is limited and the results are mixed, the tentative conclusion to be drawn is that the presence of cost heterogeneity (or greater heterogeneity) results in lower prices. Mason, Phillips, and Nowell (1992) considers a quantity duopoly game with lin- ear demand and heterogeneous constant marginal costs. There are three treatments:

both rms have low cost (LL), one rm has low cost and the other high cost (HL), and both have high cost (HH). They nd that, relative to the joint prot maximizing supply, industry supply is higher when rms are asymmetric (HL) than when they are symmetric (LL and HH). Thus, price is lower - relative to the joint prot maximizing price - when rms are asymmetric. Furthermore, the observed quantities under asym- metry are close to the static Nash equilibrium quantities so there is little apparent collusion. These ndings are conrmed in Mason and Phillips (1997) when there is private information so each subject knows only his own mapping from quantity pairs into prot. This lack of information reduces collusion when rms are symmetric but

- given quantities are still close to Nash equilibrium levels when rms are asymmetric

- prices remain higher compared to when rms have dierent costs. Consistent with this nding is Fonseca and Normann (2008) who examine a price game with homogeneous goods. Firms have identical marginal cost but potentially


dierent capacities. Industry capacity is kept constant across the symmetric and asymmetric treatments. They nd that an asymmetric capacity distribution results in lower prices. Dugar and Mitra (2009) also have a price game with homogeneous goods but rms have dierent constant marginal costs and no capacity constraints. All treatments involve rms with dierent marginal costs but they allow the degree of asymmetry to vary. In particular, the cost of the low cost rm is kept the same and the cost of the high cost rm is allowed to take one of three values. 17 The average price and average lowest price are found to be statistically indistinguishable across cost treatments and, in the case of a low or medium gap in costs, are above the static Nash equilibrium prices (so there is some collusion). In this case, more heterogeneity did not seem to have an eect. The one study to nd that rm asymmetry might actually result in higher prices is Argenton and Müller (2012). They consider a duopoly price game with homogeneous goods, linear demand, and a cost function of the form  2 where is allowed to vary across rms. A key assumption is that a rm is required to produce to meet all demand. 18 As a result, there are multiple static Nash equilibria. In particular, there is a range of prices above marginal cost that are equilibria because a rm that undercuts the price experiences a near-doubling of its demand and thus is forced to produce and sell many units at a price below its marginal cost. Prices above the highest static Nash equilibrium price are commonly observed whether rms’ cost functions are identical or dierent. There is no evidence that symmetry produces higher prices and, in fact, there is some evidence that prices are slightly higher in the asymmetric treatments. In contrasting this nding with the other studies, it is important to keep in mind the very dierent structure as reected in the existence of multiple static Nash equilibria. In sum, there are not that many studies that have explored the eect of rm het- erogeneity on achieving collusive outcomes. Two studies clearly nd that asymmetric cost functions make collusion less common (that is, prices are lower), one study nds no eect of increased asymmetry, and one study (with a very dierent modelling of the market) nds weak evidence that rm asymmetry makes collusion more common.

2.3 Summary of Findings and Open Questions

The following results are distilled from the literature:

17 Given that the static Nash equilibrium price is the price just below the high cost rm’s cost (for the parameter values used in the experiment), it would have been a better design to hold xed the cost of the high cost rm and vary the cost of the low cost rm for then the static Nash equilibrium price would be the same across treatments. As demand is perfectly inelastic with a maximum willingness to pay of 50, the monopoly price is independent of all rms’ costs. 18 While that is an optimal response when a rm’s marginal cost is constant (assuming price is at least as high as marginal cost), it is not when marginal cost is increasing. As they note, there are some market settings for which regulation requires rms to meet all demand.



Without communication, prices above static Nash equilibrium levels commonly occur when there are two sellers but very rarely occur with more than two sellers.

2. Compared to prices when sellers do not communicate, allowing sellers to an- nounce prices results in initially higher prices but then prices decline to levels mildly above or close to levels when communication is prohibited.

3. Making communication costly tends to raise price.

4. Compared to prices when sellers do not communicate, unrestricted chat pro- duces signicantly higher prices that persist over time.

5. Compared to when rms are symmetric, asymmetric costs (tentatively) result in lower prices.

Pertinent to the current study, the literature has not addressed the following ques- tions:

What is the eect of rm heterogeneity on the ecacy of communication? 19

What is the eect of rm heterogeneity on the ecacy of unrestricted commu- nication compared to price announcements? 20

Do price announcements allow rms (whether symmetric or asymmetric) to eectively collude when there are more than two rms? 21

3 Theory

In this section, the theoretical model underlying the experimental design is described along with the predictions coming out of the equilibrium analysis.

19 The only experiment with communication and asymmetric rms is Harstad, Martin and Norman (1998) but they do not explore the interaction between heterogeneity and communication. 20 This question is of a similar nature to one posed in Fonseca and Normann (2012) in that it is looking at the interaction of an element of market structure with the relative ecacy of com- munication. They ask: What is the e ect of the number of sellers on the e cacy of unrestricted communication compared to no communication? 21 The only study with price announcements and more than two rms is Hinloopen and Soetevent (2008) which has three rms and nds the rise in prices due to communication is temporary. But they also assume probabilistic p enalties and have a very structured communication protocol.


3.1 Model

The setting is a market with 2 rms oering homogeneous products. Production of one unit requires one unit of capacity. A rm has an unlimited number of units of capacity at cost 0 and a limited number at units at a lower cost . These cost levels are the same across rms but the amount of low cost capacity is allowed to vary. Firm ’s cost function is then of the form:

() = ½

+ 0 ( )



[0  ] 

where 0 and 0

() is market demand when the lowest price in the market is . Consistent with the experiments that we conducted, price comes from a nite set {01   }

for some  Also, assume 0 and 0   It is assumed that market demand is decreasing in the market price when demand is positive. Dene to be the monopoly price based on marginal cost of 0 and assume it is uniquely dened:

= arg max (0 ) ()

0 .

It is assumed that the number of low cost capacity units is suciently small so that all units are used up at the joint prot maximum; specically, () P =1 22 This condition implies the joint prot maximizing price is Though rms may dier in terms of their low cost capacity, they have the same ordering over a common price and, in particular, agree that is the most preferred common price. Firms simultaneously choose price and how much they are willing to supply at that price. Thus, a pure strategy is a pair (  ) where is a rm’s price and is its maximum quantity, and a rm produces to meet all demand up to . Firm ’s residual demand is given by


(1    ; 1    )

min 0max


| Γ ( )|


( ) X  ( ) X X

( )

( )

Γ( ){}


(0 ) =

{{1  } :  0 } is the set of rms pricing strictly below 0

Γ (0 ) = {{1  } : = 0 }

is the set of rms pricing equal to 0


By (1), when rm ’s price is the lowest in the market then its demand is


⎧ ⎨ µ

|Γ ( )| ( )  ( ) X


Γ( ){}

22 This assumption is satised by the parameters used in the experiment.


so that it equally shares market demand among the lowest priced rms or, when it is larger, it receives the residual demand after the other lowest price rms supplied the market according to their maximal quantities. When rm ’s price is not the lowest in the market then one rst subtracts othe supply of the lower priced rms and, if that is positive, then rm ’s demand is the larger of the remaining demand equally divided among the rms setting the same price as rm and the remaining demand after subtracting othe supply of the other rms setting the same price as rm  23 A key assumption is that a rm’s share of market demand when all rms charge

a price of 0 is at least as large as its low cost capacity: (0 )  max {1    } 024 The key implication is that, in equilibrium, a rm’s residual demand when it prices at 0 is at least as large as its low cost capacity. In equilibrium, if  0 then it is not optimal for rm to supply more than units since for each unit it sells beyond , it receives a price less than the cost of producing that unit. Thus, if = 0 and rm is a lower-priced rival to rm (  0 ) then rm supplies 25 Thus, subtracting out the supply of rms pricing below 0 , the remaining demand is

(0 ) X


(0 )

If = 0 then rm ’s demand is at least (2) divided by the number of rms pricing

at 0 , which is |Γ (0 )|. Hence, rm has demand at least as large as its low cost

capacity when


|Γ (0 )|


(0 ) X

(0 )

which can be rearranged to

⎠ ⎞

(0 ) X + | Γ (0 )|

(0 )

If (0 )

max {1    } then this condition holds for all (0 ) and for all Γ (0 )

and (0 ). 26

3.2 Static Equilibrium

In the static game, a pure strategy is of the form (  ) × < + and a mixed strategy is a probability distribution over × < + Let () denote the maximal

23 In the experimental design, a rm’s expected residual demand equals (1) in that the demand of rms that charge the same price is sequentially allocated among those rms. 24 This assumption is satised by the parameters used in the experiment. 25 Note that all rms pricing below 0 are setting their maximal quantity equal to their low cost capacity which, given ( 0 ) P =1 , means that all of them have residual demand exceeding their low cost capacity. 26 It can be shown that there is not a weaker assumption than ( 0 )  max {1    } that ensures a rm’s demand at a price of 0 is at least as great as its low cost capacity.


quantity associated with rm choosing price (whether as part of a pure or mixed strategy). Note that when  0  () is part of an optimal strategy as long as it as least as large as residual demand (for all possible realizations of the other rms’ strategies). However, setting the maximal quantity at least as large as market demand () weakly dominates setting it below market demand. When = 0 , a maximal quantity at least as great as weakly dominates setting it below . We will focus on Nash equilibria in which () () when 0 . 27

Theorem 1 Consider a Nash equilibrium in which () () 0 in the support of rm ’s strategy, . Then each rm assigns probability one to prices in {0  0 + 1}.

It is straightforward to show that there are only two pure-strategy equilibria; one in which all rms price at 0 and one in which all rms price at 0 + 1.

Theorem 2 Consider a pure-strategy Nash equilibrium (1    ; 1    ) in which ( )  Then either (1    )=(0    0 ) or (1    )=(0 + 1   0 + 1).

There are also mixed-strategy Nash equilibria in which rms randomize over {0  0 + 1} To characterize them, dene

(0 )


(0 +1)

´ + (0 )


The next theorem provides sucient conditions for all rms to randomize.

Theorem 3 If ( ) 2 Y then there is a Nash equilibrium for which (  ) =


(0 + 1 00 ) where 00 (0 + 1) with probability






and (  )=(0  0 ), where 0 (0 ) with probability







27 The following theorems have not been shown to hold when ( 0 ) [  ( 0 )) That case has some subtle challenges which have prevented us from delivering a characterization theorem.


The probability that rm chooses 0 + 1 is decreasing in and, given that

is increasing in , then the probability is decreasing in its low cost capacity. Note

that if 1 = ··· = then 1 = ··· = (= ) and the condition ( ) 2 Y


simplies to 2  1 which is always true given that (01). By continuity, the condition also holds when rms’ low cost capacities are not too dierent. In addition, if = 2 then the condition becomes 1   which again is always true. When 2 Y holds for some but not all then, it is conjectured, there is


a Nash equilibrium in which some rms choose 0 + 1 for sure and the other rms

randomize over {0  0 + 1}. In summary, static Nash equilibrium (for which maximal quantities for prices at or above 0 are at least equal to market demand) predicts each rm prices at 0 or 0 + 1 and, therefore, the market price is 0 or 0 + 1. For the non-degenerate mixed

strategy equilibrium, a rm’s expected price is decreasing in its low cost capacity but the eect is small since, for all rms, expected price lies in [0  0 + 1].

3.3 Dynamic Equilibrium

With static equilibria, rms are predicted to price at either 0 or 0 + 1, and that prediction does not depend on a rms’ level of low cost capacity. Variation in the amount of low cost capacity does aect payos-a rm with one more unit of low cost capacity can expect to earn an additional 0 - but does not inuence the prices that rms set or the quantities they sell (at least for pure strategy equilibria) because all rms expect to produce at a common marginal cost of 0 . In this sub-section, we explore how, when rms interact repeatedly, variation in their low cost capacities can

inuence behavior. Assume there is an indenite horizon in which the probability of continued interac- tion in the next period is (01) In each period, a rm chooses price and a maximal quantity; let (  ) denote the action pair for rm in period  A history in period is composed of all past prices and maximal quantities selected and quantities sold by all rms, and a rm’s information set is that history. The focus will be on pure strategies

where a pure strategy is of the form © =1 and : ¡ × < + × < + ¢ 1 × < + .

Each rm acts to maximizes its expected prot stream and we will examine subgame perfect equilibria. To consider equilibrium outcomes in this setting, consider the rms acting col- lectively - as a cartel - to result in payos that exceed static equilibrium payos. Imagine the cartel choosing an outcome path that maximizes some cartel welfare function subject to the path being implemented by a subgame perfect equilibrium. A cartel welfare function can be thought of as a reduced form for rm bargaining. The question is: For various specications of the cartel welfare function (that is, prefer- ences) and the subset of equilibrium outcome paths from which the cartel can choose



(that is, the choice set), how do the resulting outcome paths depend on a rms’ low cost capacities? In particular, do rms equally share market demand? Or do rms with more low cost capacity have more or less market share? Let us initiate this exploration by considering a specic scenario. Suppose the cartel welfare function is the objective from the Nash bargaining solution and the choice set is the set of all stationary outcome paths implementable using the grim punishment. 28 Let us further limit our attention to rms choosing a common price but possibly setting maximal quantities less than 1 of market demand (though at least as great as their low cost capacities). 29 Thus, the outcome path may have rms with unequal market shares. If the bargaining threat point is the static Nash equilibrium when rms charge a price of 0 the cartel’s problem is:


( 1  )

Y [(0 ) + (0 ) (0 ) ]


subject to (  1    ) satisfying:

(0 ) + (0 )


(1 0 ) (1) + (0 ) + ( 1 0 )





X = ()


As specied in (3), the Nash bargaining solution is the outcome that maximizes the product of the dierence between a rm’s prot and its threat point which means maximizing the product of rms’ incremental gains from colluding. (4) ensures that the outcome can be generated by a subgame perfect equilibrium in which deviation from the outcome path results in a static Nash equilibrium price of 0 forever. The presumption is that rms appropriately set maximal quantities to produce the vector

of quantities: (1    )=(1    ). 30

This problem can be simplied to:


( 1  )

Y (0 )


28 This specication was used in Harrington (1991) for the duopoly case when 1 = 0 and 2 = (that is, constant marginal cost that di ers between rms). Also see Thal (2011) where optimal punishments are considered. 29 Restricting the analysis to a common price is for purposes of simplication. The result that is to be derived holds as well if the collusive outcomes considered are allowed to have prices vary across rms. 30 (4) considers deviations in which a rm undercuts the collusive price. A second type of deviation is for a rm, which is producing below its demand, to set a higher maximal quantity and thereby produce more while maintaining the collusive price. It can be shown that this deviation is inferior to undercutting the collusive price and producing to meet market demand.


subject to:

(0 )

1 (1 0 ) (1) ; and



= () .

Given that the objective and constraints are independent of rms’ low cost capacities, the solution must then be independent of those low cost capacities and, therefore, it is symmetric. Hence, when the process by which rms select a collusive outcome is described by Nash bargaining and the feasible equilibrium outcomes are those sustainable by the grim punishment, rm heterogeneity does not matter and the resulting outcome is symmetric. While rms’ traits then need not aect collusive behavior, the preceding analy- sis has considered only one possible model of collusion. There are two general ways in which asymmetry in rms’ traits could possibly translate into asymmetry in the collusive outcome. First, rms may care about relative prots and not just absolute prots. In particular, a rm may not agree to an outcome that has it earn a signif- icantly lower level of prot than other rms. Given that a rm’s prot is increasing in the amount of its low cost capacity, this would result in an inverse relationship between a rm’s collusive market share and its low cost capacity share. Thus, rms may need to coordinate in unequally sharing market demand in order for rms en- dowed with fewer units of low cost capacity to produce more and thereby reduce the dierence in rms’ prots. A second way in which asymmetric outcomes could emerge is through the equilib- rium constraints. If an equilibrium has, for all histories, all rms producing at least as much as its low cost capacity then the equilibrium conditions are independent of the amount of low cost capacity; the additional prot that a rm receives from having more units of low cost capacity is realized whether it abides by the collusive outcome or deviates from it. However, consider a strategy prole in which the punishment is that the deviator produces zero for some number of periods and, after doing so, there is a return to the collusive outcome. Now equilibrium conditions depend on a rm’s low cost capacity because a rm with more low cost capacity foregoes more prot when it produces zero. By aecting the set of equilibrium outcomes from which the cartel selects, rms’ traits may then result in an asymmetric outcome. To elaborate on this point, consider the following strategy prole designed to support a collusive outcome in which all rms set a common price and rm ’s share of market demand () is If a rm deviates from the outcome path then the punishment has the deviator choose (  )=(0 + 1 (0 + 1)) and the non- deviators choose (  )=(0  (0 )) for one period - so the deviator sells zero and the non-deviators share market demand at a price of 0 - and then there is a return to the collusive outcome. This punishment applies whether a rm deviates from the original collusive path or the punishment path. There are three equilibrium conditions. First, the equilibrium condition on the


collusive path is:

(1 + ) [(0 ) ()+(0 ) ] (1 0 ) (1) + (0 )

(1 0 ) (1) (0 )

(1 + ) (0 ) ()


A rm that undercuts the collusive price realizes an increase in prot from the increase

in its demand but then, as part of the punishment, realizes one period of zero prot.

Second, the equilibrium condition on the punishment path when rm is the rm that deviated is:

[(0 ) ()+(0 ) ] (0 )

(1 ) (0 )

(0 ) ()


The deviator is tempted to price at 0 not 0 + 1in order to have demand that allows

it to sell output from its low cost capacity and thereby earn (0 ) . It is deterred

from doing so by the prospect of earning zero prot next period rather than collusive prot. Third, the equilibrium condition on the punishment path when rm is not the rm that deviated (and = 2) is:

(0 ) + [(0 ) ()+(0 ) ] µ ( 0 + 1)


(0 + 1) 2(0 )

2(0 ) ()

+ (0 )


When = 2a non-deviator can increase its current prot by raising price from 0 to 0 + 1 and sharing market demand with the deviator while still selling its low cost capacity (now, at a slightly higher price). When 3, a non-deviator loses all demand when it raises price to 0 + 1 because there are other non-deviators still pricing at 0 ; hence, (7) is absent when 3 because compliance by a non-deviator

is automatic. 31

In examining how a rm’s low cost capacity inuences these equilibrium condi- tions, rst note that (5) and (7) are less stringent when a rm has more low cost capacity because the punishment from deviating from the outcome path is harsher for that rm since it foregoes more prot by producing zero for one period. Specif- ically, if  then the punishment lowers prot more for rm than for rm

by an amount equal to (0 ) ( ). By analogous logic, equilibrium condition (6) is more stringent when a rm has more low cost capacity because the path has

it produce zero in the current period; thus, with more low cost capacity, it is more

31 The latter statement does require that the 2 rms pricing at 0 have sucient capacity to meet market demand.


tempted to lower price to 0 so that it has demand which allows it to sell the output of its low cost capacity. Evaluate these equilibrium conditions at the parameter values that will be used in

the experiment - () = 150   0 = 54

= 10  = 8 - and set the collusive price

at the joint prot maximum: = = 102The equilibrium conditions are now:

(102 1 54) (150 101) 8 (54 10)

(1 + 8) (102 54) (150 102)

55531 0084877

(1 8) (54 10)

102) 0047743

8 (102 54) (150

(150 54 1) 16 (54 10)

16 (102 54) (150 102)

02577 019097

As long as 41the binding constraint is the rst one which ensures a rm wants to set the collusive price. Evaluating it at the asymmetric capacities to be used in the experiment - (1  2 ) = (186) - this is an equilibrium if and only if 1 403 and 2 504With equal market shares, the low cost rm is amply satised - it only needs a market share of at least 40.3% to nd it optimal to set the collusive price - but the high cost rm does not have enough market share for it to nd it optimal to abide by the collusive outcome. Thus, if rms use a punishment in which the deviator sells zero for one period then the joint prot maximizing price is sustainable if and only if the rm with fewer units of low cost capacity is given more than 50% of market share. While any amount between 50.4 and 59.7% will work, if, as discussed above, relative prots are a consideration in the selection of an outcome then the market share may need to exceed the minimum 50.4% that will ensure stability. The conclusions of our analysis of dynamic equilibrium are imprecise but perhaps informative nevertheless. First, there is a wide class of scenarios whereby the col- lusive outcome is symmetric even when rms have dierent cost functions (at least as heterogeneity is modelled here). If rms focus on equilibria in which they always produce at least as much as their low cost capacity (such as with symmetric equilib- ria constructed on the grim punishment) and the selection of an outcome does not depend on relative prots (such as with Nash bargaining) then the prediction is that the collusive outcome will involve equal market shares. Second, scenarios have been identied whereby the collusive outcome has the rm with fewer units of low cost ca- pacity assigned a higher market share, but scenarios have not been identied whereby it is the rm with more units of low cost capacity that is favored. If the selection of an outcome considers relative prots then the higher cost rm may receive a higher market share in order to reduce the dierence in prots. If the punishment used in equilibrium has the deviator produce zero for some length of time (and non-deviators meet demand at a lower price), it is the higher cost rm’s equilibrium condition that is most stringent which means it may then require providing more market share to the higher cost rm in order to induce it to set the collusive price. There could,


of course, exist punishments whereby it is instead the lower cost rm’s equilibrium condition that is more stringent, but thus far they have not been found. In sum, the theory is inconclusive as to whether the rm asymmetries modeled here should result in symmetric or asymmetric outcomes but, if it does require asymmetric outcomes, it seems likely to imply that the higher cost rm will receive a higher market share.

4 Experimental Design

4.1 Environment

The experiment consists of a multi-period posted oer market with xed matching. If the market has sellers then participants are matched and the match is kept xed throughout the session. Participants are told that the experiment will last for at least 40 periods after which there is an 80% chance in each period of the experiment continuing to the subsequent period. The shortest experiment ran for 40 periods while the longest lasted for 53 periods. Sellers oer identical products and face market demand () = 150 , and are informed that the buyers are simulated. 32 Each seller’s cost function is a step-function with the low cost step equalling 10 and

  high

the high cost step equalling 54. Seller is assigned low cost units and cost units so the cost function is

() = ½


10+ 54 ¡ ¢ if



© 01    © + 1   +


In all treatments, industry capacity is xed at 24 units of low cost capacity and 180 units of high cost capacity. Thus, market demand and the industry cost curve are as depicted in Figure 4.1.

Figure 4.1: Industry Cost and Demand

depicted in Figure 4.1. Figure 4.1: Industry Cost and Demand 3 2 There are then 150

32 There are then 150 computerized buyers with one buyer with a valuation of 150, one with a valuation of 149, and so forth.


In each period, subjects simultaneously choose a price and a maximal quantity (to be sold). A subject’s total number of units sold equals the minimum of its demand and the maximal quantity it selected. Subjects are told that low cost units will be sold rst, and any excess demand will not be carried over to the next period. Sellers only incur costs for the units sold. Subjects have 60 seconds to select a price and a maximal quantity, and there is only one price-maximal quantity oer posted by a subject in each period. If a subject choose not to post an oer then s/he earns zero prots for that period. Once subjects post their price-maximal quantity oers, the market clears using computerized buyers. Buyers rst purchase from the low price seller until demand or the low price seller’s maximal quantity is reached. If there is any residual demand, the process is repeated for the next lowest price seller. This process continues until all demand is met at the prevailing prices or maximal quantities are achieved. Buyers only purchase units if the price is below their valuation for those units. In case of a tie, the system alternates between sellers buying a single unit from each seller (with identical prices) until all available units are exhausted. Subjects are informed about the tie-breaking rule and that the buyers are computerized. At the end of each period, each subject learns the price-maximal quantity oers of all subjects as well as all subjects’ results in terms of units sold and prot earned. They can also review the entire history at any point in time. The environment that subjects face is common knowledge; in particular, they all know market demand, the number of sellers, and each seller’s cost function. Subjects are provided with a prot calculator where they can input price-maximal quantity oers for all sellers and learn the resulting prots. They are told: "The prot calculator allows you to estimate your (and others’) prots. To do so you can input your price and quantity and make guesses for the other sellers." The calculator allows them to try various combinations of oers and learn the eect on prots. By the theory in Section 4, if maximal quantities are not binding then static Nash equilibrium predicts that rms assign positive probability only to prices of 54 and/or 55. We will take 55 to be the "competitive price." An advantage of having the step- wise marginal cost function is that, contrary to when marginal cost is constant, the static Nash equilibrium price is not weakly dominated. For example, if there are two symmetric rms and both price at 55 (and set the maximal quantity at least as high as 95) then each earns expected prot of

µ 150 55


55 10 × 12 54 × µ 150 55 12 = 5755


while prot is zero by pricing above 55 (as residual demand is zero) and prot is only 528 by pricing at 54 (and is even lower by pricing below 54).

4.2 Treatments

There are three sources of treatments - number of rms, seller cost heterogeneity, and information. The number of rms varied between 2, 3, and 4. In the symmetric


treatment, all sellers have the same number of low cost and high cost units. The asymmetric treatment - which was run only for the case of a duopoly - assumes that both rms have total capacity of 102 units with rm 1 having 18 units of low cost capacity and rm 2 having 6 units of low cost capacity. The market structure treatment allows the number of sellers to vary between two, three, and four. The various market structure and cost treatments are shown in Table 4.1.

Table 4.1: Market Structure and Cost Treatments




= 2

= 3

= 4

= 2

¡   ¢ = (1290)

¡   ¢ = (860)

¡   ¢ = (645)

¡   ¢ = (1884) ¡   ¢ = (696)





There are three informational treatments.

No Communication: All information is common and sellers have access to the entire history; specically, sellers can observe all past price-maximal quantity oers, transaction prices, quantities sold, and prots. Sellers cannot communi- cate in any form with their rivals. Sellers simultaneously choose price-maximal quantity oers and have a maximum of 60 seconds to make a decision. If oers are made earlier, the system goes directly to determining the market outcome (that is, allocating demand according to the price-maximal quantity oers se- lected and calculating prot) and informing sellers of the outcome. Sellers also have the option of not posting an oer by clicking on the “Do not send an oer” button.

Price Announcement: Sellers are informed that each period of the experi- ment consists of two stages. In the rst stage (Price Announcement), sellers simultaneously choose (or not) to make a single non-binding price announce- ment regarding the price they will select in the market competition stage. Thus, communication between sellers is exclusively numeric and no additional infor- mation can be transmitted. If any sellers choose to announce a price, the an- nouncements are simultaneously released to the other sellers. All sellers know that all price announcements are non-binding, and that they can choose not to make an announcement. The rst stage lasts for 60 seconds, however, it moves to the second stage if all announcements are made before the time limit. As in the No Communication treatment, the second stage has them simultaneously make price-maximal quantity oers. All information is common and sellers have access to the entire history, including all subjects’ announcements.

Chat: Sellers are informed that each period of the experiment consists of two stages. In the rst stage, they can participate in an online chat room where they can communicate with the other seller(s) for 60 seconds. The communication protocol is explicitly explained to the participants: “You are free to discuss


any aspects of the experiment, with the following exceptions: you may not reveal your name, discuss side payments outside the laboratory, or engage in inappropriate language (including such shorthand as ‘WTF’). If you do, you will be excused and you will not be paid”. After the rst stage chat session, they simultaneously make price-maximal quantity oers. All information is common and sellers have access to the entire history.

The No Communication treatment describes the usual environment in which rms can only coordinate by signaling through their actual transaction prices. The Price Announcement treatment captures a feature of some markets in which rms make non-binding announcements about future prices. For example, advance price an- nouncements have been deployed and argued to have produced coordinated supra- competitive prices in steel (Scherer, 1980), airlines (Borenstein, 2004), and diesel and petrol fuel in Taiwan (Fair Trade Commission, 2004). In our experiments, price an- nouncements can only aect seller behavior because buyers are simulated and, even if buyers were live, they would be irrelevant to buyer behavior. It is then best to think of the Price Announcement treatment as relevant to markets in which these announcements are not received by buyers (for example, they occur through a trade association) or where such information is of little value to buyers. We intentionally did not allow rms to also announce maximal quantities because such quantity an- nouncements are very uncommon though have occurred in the automobile industry (Doyle and Snyder, 1999). The Price Announcement treatment is designed to give rms an instrument by which to coordinate that is short of express communication. The issue is whether price announcements are suciently informative to induce co- ordinated behavior. Finally, the Chat treatment models explicit collusion in that rms can engage in unrestricted communications in order to coordinate on a collusive outcome and engage in an exchange of assurances. The primary objectives of this study is to assess how eective are price announcements (relative to no communica- tion) for producing supracompetitive outcomes, how eective is chat (relative to price announcements) for producing supracompetitive outcomes, and how the incremental value of communication depends on rm heterogeneity. There are three sources of variation in treatments: market structure (that is, number of sellers), rm heterogeneity (that is, symmetric vs. asymmetric cost func- tions), and communication protocol. Table 4.2 summarizes the dierent combination of treatments used in the experiments along with the notation we will use when referring to the treatment. In [ ] is the number of experiments run with that treat- ment. Given the large number of possible combinations, the market structure-rm heterogeneity treatments were chosen to make the best use of our budget by avoiding treatments that were unlikely to provide new information. For example, if rms for a treatment yielded competitive results then we did not run the treatment with


more than rms as it is likely to produce competitive results.

Table 4.2: Experimental Treatments






= 2

= 3

= 4

= 2

No Communication

SNC2 [12]

SNC3 [8]


ANC2 [13]

Price Announcements

SAN2 [12]

SAN3 [8]

SAN4 [6]

AAN2 [12]


SCH2 [12]


SCH4 [6]

ACH2 [12]

4.3 Procedures

Our subject pool consisted of students from a major American university with a diverse population. Participants were recruited by email from a pool of more than 2,000 students who had signed up to participate in experiments. Emails were sent to a randomly selected subset of the pool of students. Subjects were recruited for a total of two hours. The experiments took place in May 2011. In total, 242 students participated in 73 duopoly, 16 triopoly and 12 quadropoly experiments. The instructions were displayed on subjects’ computer screens and they were told that all screens displayed the same set of instructions. They had exactly 20 minutes to read the instructions (see Appendix B). A 20-minute timer was shown on the laboratory screen. Three minutes before the end of the instructions period, a monitor entered into the room announcing the time remaining and handing out a printed copy of the summary of the instructions. None of the participants asked for extra time to read the instructions. At the end of the 20-minute instruction round, the experimenter closed the instructions le from the server, and subjects typed their names to start the experiment. The interaction between the experimenter and the participants was negligible. Average payos (including the show-up fee) varied from a low of $18.85 (which was for triopoly with the No Communication treatment) to a high of $34.35 (which was for duopoly with the Chat treatment).

5 Results

In this environment, a seller is trying to determine both the optimal price given the other sellers’ anticipated prices, and what prices it should anticipate being set by the other sellers. There appears to be a fair amount of learning of the rst kind as reected in the experimental output and in the messages from the Chat treatment. In the Chat treatment, even when sellers communicated their intent to maximize joint prot, there was a lot of discussion about what price would actually achieve that objective. Presumably, when there is no chat, an individual seller is also trying to "gure out" the best price, given beliefs as to the other sellers’ prices. What this suggests is that the experimental output in the early periods is a conuence of


learning, competition, and cooperation, while the output in later periods is more representative of what we are interested in which is competition and cooperation. Therefore, results will be reported for periods 1-20, 21-40, and 1-40 (recall that the length of the horizon is 40 periods for sure and is then stochastically terminated).

5.1 Baseline: No Communication

For the No Communication (NC) protocol, Table 5.1 reports the average market price for the symmetric duopoly (SNC2), symmetric triopoly (SNC3), and asymmetric duopoly (ANC2). 33 In examining these prices, recall that the competitive (static Nash equilibrium) price is 54-55 and the monopoly price is 102. As reported in Table 5.1, average market price is at least 54 in all treatments so the transaction price is at least as high as the static Nash equilibrium price. With a symmetric duopoly, average market price is 66.7 over periods 1-40 and 69.5 for periods 21-40; while it is 61.5 and 63, respectively, for when rms are asymmetric. With a symmetric triopoly, average market price is 58.5 over periods 1-40 and 56.3 for periods 21-40 which is very close

to the competitive level. Conducting a t-test for the hypothesis that average market price exceeds a price of 56, it is soundly rejected for a triopoly and soundly accepted for both symmetric and asymmetric duopolies. 34 Examining the histograms in Figure 5.1 for market price, the price distribution has a peak around 55 for symmetric triopoly and asymmetric duopoly, while the peak is closer to 60 for symmetric duopoly. As we move from symmetric triopoly to asymmetric duopoly to symmetric duopoly, there

is a shifting of mass to higher prices.

Consistent with previous ndings in the experimental literature, supracompetitive prices occur with two sellers but not with three sellers. We also nd for the case of a duopoly that prices are higher when rms’ cost functions are identical though it is only

barely statistically signicant (which is not surprising given only 12 observations). For periods 1-40, prices are higher under symmetry by 8.5% (= (66.7 - 61.5)/61.5) with

a p-value of .103 (see Table 5.4), and are higher for periods 21-40 by 10.3% (with a p-value of .128). 35

Property 1: For the case of no communication, average market price exceeds the competitive level in duopoly (symmetric and asymmetric) but not in triopoly (symmetric).

33 The market price is the sum of rms’ prices weighted by the rm’s market share. The average market price for a group is the market price averaged across all periods and is the unit of observation for calculating the average, median, and standard deviation in Table 5.1. 34 In conducing this test, one group (that is, a matched set of subjects interacting for 40+ periods) is an observation so there are 12 data points. 35 Keep in mind that the degree of asymmetry is rather mild in that rms only dier in the number of low cost capacity units and, at any relevant symmetric outcome, rms are producing well beyond their low cost capacities in which case they face the same marginal cost.


Property 2: For the case of duopoly and no communication, average market price is higher when rms have identical cost functions than when they have dierent cost functions.

5.2 Signaling: Price Announcements

Now we turn to the central part of the analysis which is assessing the eect of the communication protocol on behavior and how the communication’s eect depends on market structure. It is important to emphasize that our interest lies in determining whether rms collude, and collusion is more than high prices; it is a mutual under- standing among rms to suppress competition. Prices could be high, and yet subjects are not colluding. For example, rms may periodically raise price with the intent of coordinating on some supracompetitive outcome but never succeed in doing so. High average prices are then the product of failed attempts to collude. Or sellers may engage in randomized pricing that periodically results in high prices - thus producing high average prices - but again there is not the coordination that we would associate with collusion. 36 In the ensuing analysis, sellers are said to be colluding when they achieve high and stable prices. This could mean they consistently set identical prices, and equally share demand. Or rms could set dierent prices with the rm with the lower (but still high) price restricting its supply so that the rm with the higher price has residual demand. Or rms could alternate over time with one rm selling to the market and the other rm pricing itself out of the market or not participating. Recognizing the dierent forms that supracompetitive outcomes can take, various measures will be used in our analysis. As an initial step let us focus on collusion that takes the form of rms setting identical supracompetitive prices. To identify the extent to which price announce- ments make such collusion more common, we will report average market price and two measures of coordination: the number of periods for which sellers set the same price (Same ) and the longest number of consecutive periods for which sellers set identical prices (Duration ). If sellers achieve a high average price and high measures of Same and Duration, this is compelling evidence that they are colluding. If sellers achieve a high average price and low measures of coordination than it could either be that rms are not colluding or are colluding in a dierent manner. 37 In going from the No Communication to the Price Announcement treatment, Table 5.2 reports that the average market price under duopoly substantially increases, whether rms are symmetric or asymmetric. Over periods 1-40, price rises from 66.7 to 76.2 under symmetry (though the p-value is .133, see Table 5.4) and from 61.5 to 67.5 under asymmetry (p-value = .004). For periods 21-40, price rises from 69.5

36 While such randomizing pricing is not predicted by the theory, it appears to t the pricing of some groups. 37 Of course, a low average price and high coordination is consistent with competition.


to 79.5 under symmetry (p-value = .273) and from 63.0 to 70.6 under asymmetry (p-value = .017). While the increase in average price is actually larger when rms are identical - compare 9.9 (or 14.2%) for symmetric rms with 7.6 (or 12.1%) for asymmetric rms - the standard deviation is much larger under symmetry which may be why the dierence is statistically signicant only for the asymmetric duopoly case. We will return to this point later. Having the ability to make price announcements proves insucient to produce collusion when there are more than two rms. For symmetric triopolies, average price is 57.7 (periods 21-40) which is close to average price without announcements (56.3) and to the competitive price (55). Similar results were found in six sessions conducted with four symmetric rms. In sum, price announcements matter when there are two sellers - whether symmetric or asymmetric - but not when there are more than two sellers. The general nding in the literature - collusion in the absence of communication is unlikely when there are more than two sellers - is robust to allowing sellers to communicate by announcing prices. While price announcements are producing distinctly higher average prices for duopolies, is this collusion? Examining the coordination measures, the evidence is compelling that symmetric duopolies are colluding, but that is not the case with asymmetric duopolies. As shown in Table 5.3, there is almost a doubling in the num- ber of periods in which rms in a symmetric duopoly set identical prices; it increases from 8.3 to 16.1. It is even more impressive if we focus on periods 21-40 where the frequency of identical prices rises from under 25% of periods (4.6 out of 20 periods) to more than 50% (10.3 out of 20 periods). The Duration measure tells the same story; the average maximal number of consecutive periods for which rms set the same price goes from 2-3 to 8-10 periods. In contrast, price announcements do not produce any change in the coordination measures when rms are asymmetric. Though, given the small number of observations, the dierences for symmetric rms are not statistically signicant by the usual standards (see Tables 5.5 and 5.6), the evidence is suggestive that price announcements are producing more coordination. 38 Of course, the lack of evidence for increased coordination in asymmetric duopolies may just reect the inadequacy of our measures. Same and Duration are designed to detect coordination on identical prices. Perhaps, due to cost dierences, asymmetric duopolies collude with dierent prices and choose maximal quantities so as to allocate market demand, or instead alternate in supplying the entire market. If rms have settled down to such supracompetitive outcomes then this will be reected in high and stable industry prot. Figures 5.2 and 5.3 report the mean and standard deviation of industry prot over

38 In identifying the presence of collusion, this nding highlights the importance of including measures of coordination as well as price levels. For example, Fonseca and Normann (2012) nd reasonably supracompetitive prices with duopoly without communication but an examination of the price series (made available to us by the authors) shows that there is very little coordination. There is something going on but it is not a straightforward story of tacit collusion.


periods 21-40 for asymmetric and symmetric duopolies, respectively, and for both the No Communication and Price Announcement treatments. 39 Collusion is associated with the northwest quadrant where industry prot is high with low volatility. Exam- ining Figure 5.2, price announcements raise average industry prot for asymmetric duopolies but there are no observations of high and stable prot (relative to when rms are not permitted to make price announcements). Instead, price announcements are causing higher and more variable prot. In contrast, price announcements result in higher and less variable prot for symmetric duopolies. More specically, there are four groups in Figure 5.3 in which prot is high and the standard deviation is lower than in any of the 12 groups in the No Communication treatment. In sum, we nd clear evidence that price announcements are signicantly increasing the extent of collusion for symmetric duopoly but little evidence that it is doing so for asymmetric duopoly.

Property 3: When rms can make price announcements then - compared to no communication - rms in a duopoly set higher prices whether they are symmet- ric or asymmetric, but rms coordinate more only when they are symmetric. When there are more than two rms, price announcements do not result in supracompetitive prices.

Let us make two comments regarding our measures of coordination. First, these measures look at the mean and standard deviation for periods 21-40. A duopoly could succeed in colluding late in the horizon and thereby fail to have a high stable prot in this 20 period window. Inspection of the time series for all of the groups reveals only two such cases: one symmetric duopoly (SAN2 group 4) and one asym- metric duopoly (AAN2 group 9). They are both examined below. Second, sellers could be coordinating on an industry outcome with periodicity exceeding one period. For example, rms could cycle between both setting the monopoly price and both setting the competitive price which would produce reasonably high industry prot but with a high standard deviation. Besides the fact that a multi-period cycle would be both more dicult to coordinate upon and sustain and probably less protable, an inspection of prices, quantities, and prots shows no evidence of such a pattern. There is a rather natural explanation for why price announcements are more eec- tive in producing collusion when rms are symmetric. When sellers have identical cost functions, a symmetric supracompetitive outcome is focal, and can be implemented by coordinating on identical prices. However, when sellers have dierent cost func- tions, a symmetric outcome is no longer focal. An asymmetric division of industry prot could be produced in a variety of ways but arguably the most straightforward is for sellers to set identical prices and unequally allocate market demand, which is what has been done with many cartels (see Harrington, 2006). For example, if sellers

39 Note that the monopoly prot is 3360, and when all rms set a price of 54 the industry prot is 1056.


wanted to support the joint prot maximum and have the high cost seller receive 60% of market demand, both sellers could charge the monopoly price of 102, which yields market demand of 48, and have the low cost seller set its maximal quantity equal to 19, which will result in the high cost seller supplying the residual demand of 29. However, this collusive outcome requires coordination of prices and quantities. The diculty in coordinating on equal prices and unequal quantities in the Price An- nouncement treatment is that sellers are only allowed to announce prices. Of course, just because an asymmetric outcome may be the rst-best collusive outcome for an asymmetric duopoly, it does not imply that rms would try to coordinate on it. If it is perceived to be too dicult then they may decide to go for a second-best solution of coordinating on identical prices and equally sharing market demand; some collu- sion is better than competition. However, that is not what we are nding. Under asymmetric duopoly, sellers are not coordinating on a common price and, with one exception to be analyzed below, they are not coordinating on dierent prices either. It is worth pointing out that, in contrast to other experiments that have allowed sellers to make non-binding price announcements, the eect of producing more coordi- nated behavior persists over time. 40 Indeed, the results are stronger for periods 21-40 than for periods 1-20. More informative to this point than the aggregate measures we’ve provided are the price series from the 12 symmetric duopolies in Figure 5.4. In ve out of the 12 groups, sellers eventually had identical prices well above competitive levels. In four out of those ve groups (groups 3, 7, 8, 12), sellers set the same price for the last 20 or so periods with price at or near the monopoly level for three of them (it was around 75 for group 7). (Group 9 is also a candidate for inclusion in that set.) In the other run (group 4), sellers’ prices were common but steadily rising in the last 15 periods as they climbed from the competitive level to the monopoly level. Only in one group did sellers achieve reasonably stable, common, and supracompetitive prices to then experience an unravelling whereby prices retreated back to competitive levels (group 2). Thus, our results show that price announcements can be eective in producing persistent collusive pricing. Let us now return to the issue of the high standard deviation for average price for a symmetric duopoly under the Price Announcement treatment (see Table 5.2). The price paths in Figure 5.4 reveal that, generally, either sellers set high identical prices (groups 3, 4, 8, 9, 12) or set prices near competitive levels with some unsuccessful forays into supracompetitive territory (groups 1, 5, 6, 10, 11). (Group 2 does not fall into either of those two bins, and group 7 would fall into the rst bin except that price is only 75.) This pattern can also be seen in the market price histograms in Figure 5.5. For the symmetric duopoly, the distribution has a mode near the competitive price and one near the monopoly price (both for periods 1-40 and 21-40). In comparison, there is not this stark dichotomy when rms are asymmetric. The distribution in

40 Recall the assessment of the literature in Cason (2008): "Although price signaling often increases transaction prices, this increase is very often temporary. [Steady-state] behavior may be unaected by non-binding price signaling in many environments."


Figure 5.5 for asymmetric duopoly is unimodal for periods 1-40 and far less bimodal for periods 21-40 than under symmetry. It is this dichotomy in outcomes for the symmetric duopoly which is producing

a relatively high standard deviation; either rms have great success in colluding or

very little success. Figure 5.6 nicely depicts this distinction between symmetric and asymmetric duopolies. An observation here is a group’s average market price and the number of periods for which rms set the same price. When rms are symmetric, the observations form two clumps; one with low price and low coordination, and the other with high price and high coordination (with the exception of group 2 which has high price and low coordination). When rms are asymmetric, there is just one large clump without any apparent relationship between average price and the frequency with which rms set the same price. Having found that collusion in symmetric duopolies is more common when sellers can make price announcements, it is useful to examine the pattern of announcements and prices in order to gain some insight into the underlying mechanism responsible for this nding. It was with this objective in mind that we performed a statistical analysis of announcements which, however, proved uninformative. This is not too surprising given that announcements are cheap talk. (Even if announcements do serve to coordinate behavior in some instances, any regularity could easily be lost

if many announcements are meaningless.) Instead, let us engage in some post hoc

analysis of seller behavior in a few groups, which is suggestive but speculative. To frame our thinking, consider two hypotheses regarding how announcements could produce collusion. The rst hypothesis is that a seller believes there is mutual understanding of a desire to collude and the only challenge is to coordinate on a particular price. In that case, one might expect to observe a seller announce a high price and, in anticipation that the announced price coordinates expectations, sellers then price at the announced level. A second hypothesis is that a seller is uncertain that there is mutual understanding regarding collusion and, therefore, acts cautiously by announcing a high price but not pricing at that level until the other seller has

made the same announcement. If both sellers are thinking that way then we ought to observe high prices only when both of them have announced high prices. 41 Illustrative of the rst hypothesis is group 12 from the symmetric duopoly treat- ment (see Figure 5.7). After some initial failed attempts at coordinating through announcements, success occurred in period 24 when rm 1 announced a price of 102 and both rms priced at 102. Thereafter, they priced at that level and, with the exception of one period, rm 1 preceded it with the same announcement. Perhaps illustrative of the second hypothesis is group 9 from the symmetric duopoly treatment. In period 13, rm 2 announced a price of 100 but rm 1 did not make an announcement, and neither set a high price. Again in period 17, rm 2

41 This latter scenario corresponds with the ATPCO case whereby American Airlines would an- nounce a future price increase and enact that price increase if and only if the other airlines also announced a future price increase; otherwise, the proposed price increase would be retracted.


announced 100. Though rms did not raise prices, rm 1 did respond in periods 18 and 19 with an announcement of 95 though again there was no eect on prices. In pe- riod 20, rm 2 priced at 90 even though it made no announcement. Finally, in period 21, both rms announced high prices - rm 1 with 90 and rm 2 with 95 - and rms coordinated on the lower announcement by pricing at 90. From that point onward, announcements and prices steadily rose. What diers from the second hypothesis is that rm 1 raised price in period 20 prior to the simultaneous announcements in period 21. Thus, coordination of expectations could be due to the price increase in period 20 and/or the price announcements in period 21. Another symmetric duopoly, group 4, is interesting. During the rst third of the experiment, announcements were rarely used and, subject to some initially high prices, prices settled down at competitive levels. Then, in period 15, there was an attempt to coordinate on higher prices which ultimately failed. Firm 2 announced and priced at 102 but rm 1 kept price low at 50. In period 16, rm 1 announced 102 but only raised price to 91, while rm 2 continued to price at 102. Firm 2 did not give up, as it announced and priced at 102 in period 17; again rm 1 kept price below that level. Prices then gradually declined. A second attempt to coordinate began in period 25 which proved successful. Firm 1 announced 55 and priced at that level. Though 55 is the competitive price, rm 1’s purpose may have been to signal to rm 2 that its announcement is an accurate predictor of its price. Firm 1 gradually raised its announcement while always pricing at its announcement, as did rm 2. From period 25 to 38, price gradually rose to 102. Every now and then, rm 2 would make a dierent announcement but prices always followed rm 1’s announcement. Clearly, rm 1 emerged as the market leader. Turning to an asymmetric duopoly, group 9 initially had a lack of success in coordinating - in spite of rms using announcements - but eventually one rm took charge and collusion ensued. In period 21, rm 2 (high cost rm) announced 100 and priced at 62 but rm 1 priced at 54 and sold all units. In period 22, rm 2 continued to announce 100 but dropped price to 56. While rm 1 priced at 55, it limited its supply to 18 units which left residual demand for rm 2; rm 1’s prot was 810 and rm 2’s was 416. Starting with period 23, rm 1 began announcing. It announced and priced at 57 and again limited its supply to 18 units, while rm 2 announced 100 but priced at 58. Firm 1’s prot was 864 and rm 2’s prot was 560. From that point onward, rm 1 gradually raised its announcement, always priced at its announcement, and always limited its quantity to 18. While rm 2 was pricing above rm 1’s price, rm 2 always had residual demand due to the limited supply of rm 1; in fact, rm 2 (who had higher cost) made higher prot along this path. The steady-state was reached in period 33 and it was characterized by rm 1 announcing and pricing at 99, rm 2 announcing and pricing at 100, and rms sharing market demand with rm 1 selling 18 units and rm 2 selling 32 units. The steady-state prot was 1602 for rm 1 and 1736 for rm 2. This group clearly colluded and was a