GAMES

AND

ECONOMIC

BEHAVIOR

2,29-46

(1990)

Endogenous

Timing in Duopoly Games: Stackelberg Cournot Equilibria*

or

JONATHAN H. HAMILTON AND STEVEN M. SLUTSKY

Depurtment
of

Economics,

University of Floridu. Guinesville,

Florida 3261 I

Received December 13, 1988

Simultaneous versus sequential play is studied in an extended game. In a preplay stage, players decide whether to select actions in the basic game at the first opportunity or to wait until observing their rivals first period actions. In one extended game, players first decide when to select actions without committing to actions in the basic game. The equilibrium has a simultaneous play subgame unless payoffs in a sequential play subgame Pareto dominate those payoffs. In another extended game, deciding to select at the first turn requires committing to an action. Both sequential play outcomes are the equilibria only in undominated strategies. Journal of Economic Literature Classification Numbers: 026, 611. 0 1990 Academic Press, Inc.

I.

INTRODUCTION

Economists have studied a wide variety of models of oligopoly and duopoly. One important feature distinguishing different models is whether firms move simultaneously or whether they move sequentially with earlier movers acting as leaders and later movers acting as followers. With quantity competition, this distinguishes Cournot from Stackelberg outcomes, and with price competition, this distinguishes Bertrand competition from price leadership, including kinked demand curve models. Much of traditional duopoly analysis, such as Fellner (1965), treated this feature as exogenously given and compared outcomes of simultaneous and sequen* The authors thank the National Science Foundation and the Division of Sponsored Research (University of Florida) for financial support. We also thank Mark Gradstein, James Friedman, Eric Maskin, an anonymous referee, and audiences at BELLCORE, Florida, and the Public Choice Society meetings for helpful comments. Any remaining errors are our sole responsibility. 29 0899-8256190 $3.00
Copyright 0 1990 by Academic Press. Inc. All righls of reproduction in any form reserved.

30

HAMILTON

AND

SLUTSKY

tial move games in a variety of circumstances, including whether the strategic variables were prices or quantities and whether the firms outputs were substitutes or complements in demand. Recently, there has been recognition that whether duopolists play a simultaneous or a sequential move game should not be exogenous but should result from the firms decisions. Mainly, this has involved studying whether each player prefers to move first and be a leader or to move second and be a follower, For example, Gal-Or (1985) and Dowrick (1986) study this question for standard duopoly models when both firms set quantities or set prices. Gal-Or (1985) shows that, with identical firms, the slopes of the reaction functions determine whether the leader or follower does better. When reaction functions slope down, as in quantity competition with substitute goods, a leader earns higher profits than a follower. When reaction functions slope up, as in price setting games with differentiated substitute products, the follower has higher profits. She does not compare either of these profit levels to those in follower-follower outcomes. Dowrick (1986) compares Nash (follower-follower) games, Stackelberg leadership, and Stackelberg warfare. His analysis, also based on properties of reaction functions, compares payoffs to leaders and followers. Problems with his specification of Stackelberg warfare are discussed below. Boyer and Moreaux (1987) consider first mover advantages when each firm chooses both price and quantity. No pure strategy equilibrium exists when firms move simultaneously. They do not compare the sequential move outcomes with the mixed strategy outcome that arises in simultaneous play. Reinganum (1985) considers second mover advantages in research and development competition. While comparing profits of leaders and followers is of independent interest, it does not really answer the question of endogenously determining who moves first. Since a firm cannot unilaterally go from being a leader to making its rival a leader, it is not sufficient for a firm to wish to move second to turn its rival into a leader. Clearly, if both firms prefer to move first or to move second, neither can achieve this alone. In choosing to move at the first opportunity, a player eliminates only the possibility of being a follower. If he chooses to move at the last opportunity, he excludes the possibility of leading. Even if both firms prefer the outcome in which firm A goes first and firm B second to that in which B goes first and A second, it does not follow that this solution will be reached in a noncooperative setting. These payoffs must also be compared to the simultaneous move equilibrium. We embed simultaneous and sequential play of a two-player game into two different extended games where players first select the timing of choosing actions in the basic game. Comparisons of payoffs with simultaneous and sequential play in the basic game determine what order of play

ENDOGENOUS

TIMING

IN DUOPOLY

GAMES

31

occurs. For one game, Pareto dominance of sequential play payoffs over those of simultaneous play makes sequential play the equilibrium of the extended game. For the other game, the two sequential play outcomes are always the only undominated pure strategy equilibrium outcomes. Independently, McAndrews (1988, 1989) has undertaken a similar analysis of extended games closely related to ours. While our results and his overlap, there are some differences. He distinguishes between leadership and simultaneous organizations. We also consider whether simultaneous play will be early or late. We specify our results in terms of reaction functions and the Pareto preferred set at the simultaneous equilibrium. His results which complement ours are based on the derivatives of the payoff functions. Section II presents the model and the formal definitions of the games. Section III analyzes the extended game with observable delay and Section IV analyzes the game with action commitment. Section V presents our conclusions. II.
THE MODEL AND THE RULES OF THE GAMES

Our purpose is to consider formally the question of when we should expect duopolists to play simultaneously and when we should expect them to play sequentially. In the underlying basic games we consider, each players choice set is a continuous interval on the real line,. Possible interpretations for duopoly are that each firm chooses its price, quantity, or product type. Assume that the profit function for each player is strictly quasiconcave in its own action. Thus, reaction functions are continuous and a Nash equilibrium in pure strategies exists for simultaneous play. Furthermore, sequential play by the duopolists also has a pure strategy equilibrium. We assume that each of these three games has a unique equilibrium and that these differ from each other. The basic duopoly game is extended by having firms choose both an action in the basic game and a time to choose that action. If the firms choose actions at different times, then the player choosing a later time observes the action chosen by the firm playing first, giving rise to a sequential play subgame. If the firms choose actions in the same period, a simultaneous play subgame occurs. The choice of times is really a preplay stage and does not take place in real time. Thus, profits from simultaneous play are the same whether the firms
An immediate consequence of this assumption is that each firm strictly prefers its Stackelberg leader payoff to that in simultaneous play, as stated below in Lemma I. Without this assumption, it is possible that simultaneous play and one of the sequential play games yield the same outcome. This adds to the set of possible equilibria, although these new equilibria use dominated strategies.

32

HAMILTON

AND SLUTSKY

choose as quickly as they can or whether they delay as long as possible. Note that the firms can select the period in which they choose their actions in the basic game, but not whether they are leaders or followers since this involves decisions of the other firm as well.* We consider two different extended games. In one, the firms announce at which time they will choose an action and are committed to this choice, but they need not specify the action they will take if they choose in the first period. After the announcement, firms then select their actions knowing when the other will make its choice. We thus call this the extended game with observable delay. That the firms are committed to move at the time they announce is not restrictive since neither firm wants to renege in any circumstance. If simultaneous play results from the announcement stage, then firms have rejected being followers if play is early or leaders if play is late. If early simultaneous play is announced and one firm delays until after the other has moved, then the outcome does not change. Similarly, if a sequential move outcome results from the announcement stage, either firm could have achieved a simultaneous outcome and rejected it. Because a later mover observes the action of an earlier mover, the leader cannot switch his action without changing the followers choice.3 We consider only subgame perfect equilibria in the extended games. There is no discounting cost from waiting until the second period to choose actions and to receive the payoffs. Formally, define l?r = (N, S, P1) to be the extended game with observable. delay. The set of players is N = {A, B}. Let (Y and p be compact, convex intervals of 68 which are As and Bs possible actions in the basic game. We will relate these intervals to typical oligopoly choice variables after presenting the basic results. The payoffs to the two players depend only on the actions taken in the basic game. The payoff functions are a: CYx p + [WI and b: (Y x p -+ 58. T = {F, S} is the set of possible times at which to choose an action. Let i be the period in which A has chosen to select an element of (I! and j be the period in which B has chosen to select an element of p. The set of strategies for player A is SA = {F, S} x QA, where GA is the set of functions that map {(F, F), (F, S), (S, F) x /3, (S, S)} into CY.Bs strategy set is Sk = {F, S} X cPB, where aB is the set of functions that map {(F, F), (F, S) x (Y, (S, F), (S, S)} into p. Define SA = (i, 4A) E SA, where (bA E @A, and sB = (j, $B) E SB, where C#J~ @s. The payoff functions are E
2 Gal-Or (1985) suggests a game of this nature but does not analyze it. Hamilton and Slutsky (1988) analyze a similar game where the final stage is a bimatrix game. A major focus of that paper is outcomes when the basic game has a unique equilibrium which is in mixed strategies. 3 If the unique equilibrium of simultaneous play is in mixed strategies, then this may not be true, as discussed below.

ENDOGENOUS

TIMING

IN

DUOPOLY

GAMES

33

and
p,(s) = b[$A(i, H@A(F, j), S), didi, +I#, j)l S, +A(F, if (i, j) E W, FL 0, 91

b[+A(S, F, $qs(s, F)), b(s,

S))l F)l

if (i,j) if (i,j)

= (F, S)

= (S, F).

Figure 1 displays the extensive form of the game. At each of the nodes, order of play begins. We consider only subgame perfect Nash equilibria, so there is only one possible outcome in each subgame. We can thus consider the extensive form of Fig. 2 instead of the complete version of Fig. 1. The normal form of the extended game is:
dl - d4, a subgame which is the basic game with a particular
Player B Player A First Second First a,> bc a ,b, Second 01, bf a,, b,

FIG.

1. The extensive form of the extended game with observable delay.

34

HAMILTON

AND

SLUTSKY

a,

al

at

ac

bc
FIG. 2. The reduced

bf
extensive form

bl
for observable

bc
delay.

The subscripts c, 1, and f denote Cournot, Stackelberg leader, and Stackelberg follower payoffs in the basic game, respectively. We shall refer to the Coumot outcome as simultaneous play and the Stackelberg outcomes as sequential play, where the timing concerns the order of choice of actions (ai, pj) in the basic game. We will also denote the Cournot, Stackelberg leader, and Stackelberg follower actions in the basic game as aye,aI, and ozr, respectively. In our second extended game, the firms cannot simply announce when they wish to choose actions in the basic game. Instead, a firm can play early only by selecting an action to which it is then committed. We thus refer to this as the extended game of action commitment. In contrast to the previous game, a firm that plays early does not know whether its rival is also playing or will delay. It is possible in this game that a firm wishes to alter its action after learning its rivals decision. However, given complete information, if a unique equilibrium exists in the extended game, each firm can compute this and will not wish to change its decision. The formal definition of this second game is the following. I2 = (N, S*, P*) is the game of action commitment; the players N are the same as above. Let W be the action of waiting until the second period to choose an action from (Y(or fi). As set of strategies is Si = {ai, W X P\I(pi)}y where cziE (Yand VA is the set of functions that map j3 U W into (Y. Similarly, Bs set of strategies is Sk = {pi, W X Ia(ai)}, where /3i E p and *\~r,is the set of functions that map cy U W into p. Define sA E Si and sB E Si to be the strategies chosen by the players and s = (sA,Ss).Letaii=(YiifSA=LYiand

ENDOGENOUSTIMING

IN DUOPOLY

GAMES

35

FIG. 3. The extensive form of the extended game with action commitment.

l)a(Pi) if SA = W X PA, where $A() E qA(.); SiItlilaIfy, pi = pi if Ss = pi and $B(ai) if SE = W X $B(ai), where I/Q(.) E IB(*). Payoffs are PA(s) = a[aii, pi] and P&) = b[aii, pi]. Figure 3 presents the extensive form of this game. A final property of both extended games is symmetry in terms of possible times to choose the action in the basic game, even if the basic duopoly game has asymmetries in payoffs or strategic variables. Neither has the advantage of being able to commit to an action or to commit to the time to choose that action before its rival can, as might occur in a game between an incumbent and a potential entrant. We assume that strategies in the extended game are chosen simultaneously, but we briefly consider extending the game that determines the timing of play. III.

THE

EXTENDEDGAMEWITHOBSERVABLE

DELAY

We can now formally analyze the equilibria in the first extended game. Comparisons among payoffs in the simultaneous play game and the two sequential play games are essential for this analysis. By assumption, we have ruled out the possibility of mixed strategy equilibria or multiple equilibria in the basic duopoly game.
LEMMA

I.

multaneous

Each players leadership payoff exceeds his payoff play. That is, al > a, and bl > 6,.

in si-

36

HAMILTON

AND

SLUTSKY

Proof. Let us consider As payoff in the two cases. The Nash equilibrium of simultaneous play is the intersection of the two reaction functions. If A is a leader, it chooses its most preferred point on Bs reaction function. Thus, the Nash equilibrium of simultaneous play is feasible for the leader. Under the assumption that the simultaneous play equilibrium differs from that when A is the leader, it must be true that A does strictly better as a leader. Q.E.D. Theorems II, III, and IV establish the timing of choices depending the comparisons of payoffs with the three different orders of play. on

THEOREM 11. If each firm prefers its simultaneous move equilibrium payoff to its payoff as a Stackelberg follower, then the equilibrium of the extended game with observable delay is that both firms move at thejrst opportunity yielding the same payoffs as those in the simultaneous move game.

Proof. From Lemma I, A is strictly better off outcome differs from that in (S, S). The same is true of the theorem guarantees that A prefers his (F, F) (S, F) and B prefers his (F, F) payoff to that from dominant strategy for both of them.

in (F, S) since that for B. The condition payoff to that from (F, S). Thus, F is a Q.E.D.

Another case with players having identical preferences over the order of play also exists. If each prefers being a follower rather than playing simultaneously, then both sequential play games and a mixed strategy over playing immediately or delaying form the multiple equilibria of the extended game.
THEOREM III. If both firms prefer their Stackelberg follower payoffs to the simultaneous play payoffs, then both sequential play subgames are Nash equilibria of the extended game with observable delay. There is also a mixed strategy equilibrium in which firms randomize over F and S.4

Proof. A prefers the (F, S) outcome to the (S, S) outcome and B prefers the (S, F) outcome to that with (S, S). Consider (F, S); neither player gains by changing his time of move. Similarly for (S, F). Hence, both are equilibria. There must also exist a mixed strategy equilibrium Q.E.D. since neither player has a dominant strategy. When the players have identical preferences over the order of play, as in Theorems II and III, the extended game does not have a unique equilibrium which replaces the simultaneous play outcome. In one case, neither sequential play outcome Pareto dominates the simultaneous play outcome
4 The outcomes of the mixed strategies are probabilities of playing either sequential game or of playing the simultaneous move game. Stackelberg warfare where both play the leader action does not occur.

ENDOGENOUS

TIMING

IN

DUOPOLY

GAMES

37

which is therefore the extended game equilibrium. In the other, both sequential play outcomes Pareto dominate the simultaneous play outcome and multiple equilibria exist in the extended game. To get a unique endogenous sequential play equilibrium, firms must have different preferences over leading and following. Firm A playing first and firm B delaying can be a unique equilibrium of the extended game only if A prefers leading to simultaneous play to following and B prefers following to simultaneous play.
THEOREM IV. Sequential play will be the unique subgame perfect Nash equilibrium of the extended game with observable delay ifand only if one player choosing first and the other delaying has an outcome which Pareto dominates the simultaneous play game and the reverse order of play yields an outcome which does not.

If only one sequential play outcome Pareto dominates simultaneous play, one player has a dominant strategy of delaying. The other player then wants to choose first. If either player prefers his simultaneous play payoff to the sequential play payoff, that player can change the time of his play to get the simultaneous play outcome. If both sequential play outcomes are preferred by the players to the simultaneous play outcome, Q.E.D. multiple equilibria result as in Theorem III.
Proof.

This is precisely parallel to the result in Hamilton and Slutsky (1988) when the second stage is a 2 x 2 matrix game instead of one with continuous action spaces as here. In that case, even when the matrix game has a unique equilibrium in mixed strategies, if it is Pareto dominated by some pure strategy payoffs, then the players adopt sequential play in the extended game. Simultaneous play results from the extended game only in the circumstances of Theorem II and then only with both players playing at the first opportunity. For both players to choose at the last period is never an equilibrium in the extended game, due to our assumption that the unique equilibrium in the basic simultaneous play game uses only pure strategies. In this case, a leader can always achieve at least his simultaneous play payoff and thus if one player delayed, the other player desires to play first. This is not true when the simultaneous play subgame has a mixed strategy equilibrium. A leader cannot duplicate the simultaneous play payoff since the follower would observe and react to the realization of the leaders strategy. If the leader in both sequential games has a payoff lower than the expected simultaneous play payoff, then the extended game equilibrium is that both wait and play at the second turn. We can now examine whether duopolists will play simultaneously or sequentially in terms of the basic conditions of duopoly rivalry. A large variety of duopoly games have been studied, with quantities, prices, and

38

HAMILTON

AND

SLUTSKY

qualities or locations as the choice variables for the firms. Let cr and p be each firms choice variable in the basic duopoly game. Whether (Yor p is price or quantity, and whether the goods are substitutes or complements, determines the slopes of reaction functions and how the set Pareto dominating the simultaneous play outcome compares with the reaction functions. The results in Theorems II-IV relate to the results in Gal-Or (1985) and Dowrick (1986) based upon slopes of the reaction functions in duopoly games.
THEOREM V. In any duopoly game, whether symmetric or asymmetric, in which each$rm chooses a single variable and a unique equilibrium exists on the interior of the action spaces:

(A) if both reaction functions have slopes of the same sign, then either(i) neither reaction function intersects the Pareto superior set relative to the simultaneous move equilibrium, so the unique equilibrium of this extended game is the simultaneous move equilibrium or (ii) both reaction functions enter the Pareto superior set, so this extended game has multiple equilibria. (B) if the reaction functions have slopes opposite in sign, then one and only one reaction function enters the set of outcomes Pareto superior to the simultaneous move outcome, so the unique outcome in this extended game is that the player whose reaction function enters the Pareto set moves second and the other player moves first. Proof. Relative to the simultaneous move equilibrium at the intersection of the reaction functions, the Pareto superior set lies in the intersection of the isoprofit curves through that point. Since these are tangent to a horizontal and a vertical line, the Pareto superior set must lie entirely within one of the four quadrants. If it is in the first or third quadrant, then any upward sloping reaction function enters this set, and a downward sloping reaction function does not, while if the Pareto superior set lies in the second or fourth quadrant, the reverse is true. Regardless of the slopes, the Pareto superior set can be in any of the four quadrants. If it is in the first or third quadrant and both reaction functions slope down, then neither intersects the Pareto superior set, while if both slope up, then both enter the Pareto superior set. The reverse holds for the second or fourth quadrant. If reaction functions have opposite slopes, one enters the Pareto superior set and one does not, regardless of which quadrant contains the Pareto superior set. Given this, the nature of the extended game Q.E.D. equilibria follow from Theorems II to IV. These results can be illustrated graphically. For any player with a single choice variable, Fellner (1965) presents four types of quasiconcave indifference maps and the associated best reply functions. These four com-

ENDOGENOUS

TIMING

IN

DUOPOLY

GAMES

39

a FIG. 4. The possible reaction functions and payoff isoquants.

pletely characterize the possibilities if the simultaneous move game is to have a unique pure strategy equilibrium. In a duopoly, there are fewer distinct possibilities than the 16 arising from all possible combinations of these. Relabeling players or transforming variables (replacing (Ywith d = K - a) can yield a qualitatively identical graph in many cases. In Fig. 4, the four possible maps for a single player are shown in the four panels. For our purposes, there are qualitatively only three different combinations, which are presented in Fig. 5. In Fig. 5a, the set of Pareto superior outcomes has no intersection with the reaction functions, as in Theorem V(Ai). The standard quantity setting duopoly game with substitute products corresponds to this. In that context, we should expect to observe the simultaneous play game, unless one firm has an unchallenged ability to move first. For example, an existing firm might block an entrant or reduce the entrants profits if it could commit before the entrant. Without such an asymmetry in commitment ability, endogeneity in the order of play through communication will not be enough to get a leader-follower outcome. In Fig. 5b, a portion of both reaction functions lies in the Pareto

40

HAMILTON

AND SLUTSKY

(a)

a FIG. 5. Reaction functions and Pareto sets.

superior set, as in Theorem V(Aii). A symmetric Bertrand equilibrium with differentiated products (see Cheng (1985) or Singh and Vives (1984)) corresponds to this. The follower has the advantage of being able to undercut the leaders price and thus does better than the leader. The leader in turn must do at least as well in the sequential game as in the simultaneous play game since as a leader, he could play his simultaneous play equilibrium strategy. Hence, a follower in the sequential game does better than a simultaneous player. In Fig. 5c, only one reaction function intersects the Pareto superior set, as in Theorem V(B). The standard oligopoly games where both players choose prices or both choose quantities never fit this third case. One game which does fit is the asymmetric game of Singh and Vives (1984, p. 550) in which one player chooses price and the other quantity. For substitute differentiated products, Fig. 5c describes the case where q1 and p2 are the strategic variables. Having the quantity setter play first yields a Pareto superior outcome relative to simultaneous moves.

ENDOGENOUS

TIMING

IN DUOPOLY

GAMES

41

If a weak form of symmetry holds, we obtain stronger results. Define a game to be qualitatively symmetric if both firms reaction functions have slopes of the same sign and if the direction of change in profits for a firm due to an increase in its rivals action is the same for both firms. A game with identical players is qualitatively symmetric, but qualitative symmetry is much more general. In the standard Cournot duopoly model, qualitative symmetry holds even if the firms have different costs. Increases in output by one firm always lower the others profits along the others reaction function.
THEOREM VI. Consider a qualitatively symmetric duopoly game with a unique equilibrium in the interior. If reaction functions slope down, then neither intersects the set Pareto superior to the simultaneous move outcome. If reaction functions slope up, then both intersect this Pareto superior set.

Proof, The Pareto superior set lies in the first quadrant either if both reaction functions slope up and an increase in one firms action increases the other firms profits along the others reaction function or if both reaction functions slope down and the firms profits increase along its reaction function as the rivals action increases. When the direction of the profit changes is reversed, the Pareto superior set lies in the third quadrant. With qualitative symmetry, the Pareto superior set never lies in the second or fourth quadrant. The result follows immediately. Q.E.D. Under qualitative symmetry the outcome in this extended game is a unique equilibrium yielding the simultaneous game payoffs or multiple equilibria including both sequential equilibria. These arise in the same circumstances that followers or leaders do better in Gal-Ors (1985) results for identical players. As a final point, we assumed simultaneous play in the announcement stage itself. What if this game were extended so firms could choose when to make their announcements? Extending the game does not change the outcome.5 When neither sequential outcome Pareto dominates simultaneous play, both players have a dominant strategy to choose an action in the basic game at the first opportunity. They will play whatever strategy is required to achieve this no matter how many times the game is extended. When only one sequential play outcome Pareto dominates simultaneous play, only one player has a dominant strategy to choose first. This players payoff is the same with either simultaneous or sequential announcement. The remaining case has multiple equilibria in the announcement game and, in general, these are not eliminated by extending the game further.
5 See Hamilton and Slutsky (1988) for a more formal analysis of the game which is extended to allow players to choose when they announce when they will select an action.

42

HAMILTON AND SLUTSKY

IV.

THE EXTENDEDGAMEWITH

ACTION COMMITMENT

This second extended game is close in spirit to Dowricks (1986) analysis. In his games, leadership means committing to a particular action, whether or not the rival attempts to lead or follow. Thus, the leadership action must be chosen without observing the timing of the opponents move. Dowrick calculates the leadership actions under the assumption that the rival acts as a follower and waits to observe the leaders action. When both firms try to be leaders, Stackelberg warfare is the outcome. But both firms playing their leader strategies can only occur through error. The equilibrium choice of an action when leading must be consistent with the rivals action. The Stackelberg leader level is only appropriate if the player knows that the rival chooses the follower level or is waiting to select until after the leader picks his action. Without this knowledge, a leader must take account of the possibility that the rival will not wait.
THEOREM VII. In the extended game with action commitment, multiple equilibria exist, including both playing in the jirst period by playing their simultaneous play equilibrium actions, and each waiting and the other playing its Stackelberg leader choice in the jirst period. No other pure strategy equilibria are possible. Proof. First, consider ((Ye, PC), the simultaneous play equilibrium. Since PC is a best reply to oC, neither waiting nor any other choice in the first period can raise Bs payoff, and similarly for A. Thus, this is an equilibrium. Second, consider the situation when A plays ozIin the first period and B waits and then plays its follower strategy Dr. When B waits, A knows it is a leader with an optimal strategy of aI. To (~1,pr is the best reply, which can be played at the first turn or after observation. However, if it were played at the first turn, al would not be the best reply to it. Similarly, B playing PI and A waiting and then playing its follower strategy or is an equilibrium. Third, no other pair can be a pure strategy equilibrium. Assume (oi, fij) f (a,, PC) is an equilibrium. Then pj must be a best reply to ai, and CX~ must be a best reply to pj. This contradicts (a,, PC) being the unique simultaneous move equilibrium. If A waits, the only possible equilibrium action is B playing PI, and similarly if B waits. Thus, these are the only equilibria in pure strategies. Q.E.D.

This set of pure strategy equilibria can be refined by sequentially deleting dominated strategies.6 After two rounds of deletions for each player,
6 van Damme (1983) demonstrates for two-player nonzero sum matrix games that trembling hand perfect equilibria use only undominated strategies. If players might make errors in playing their strategies, opponents consider this in choosing their own strategies. Here,

ENDOGENOUS

TIMING

IN

DUOPOLY

GAMES

43

at As first turn, the set of undominated strategies is to wait or to play any action in the interval ((Ye,a,] (or [cx,, a,) if ai < a,). For B, the undominated set is to wait or to play at the first turn any action in the interval (PC, PiI (or [PI, PC) if PI < PC). 0th er strategies are dominated either by waiting until the second turn or by an action in these intervals. Since playing (Y, and PC in the first round are not in the set of undominated strategies, (a,, &) is eliminated as an equilibrium if undominated strategies are required. Even eliminating only primarily dominated strategies is sufficient to rule out the Cournot equilibrium.
THEOREM VIII. Zf the basic duopoly game has a unique equilibrium in the interior of the action space, then in the extended game with action commitment, the two Stackelberg equilibria are the only pure strategy equilibria in undominated strategies. Playing the Cournot equilibrium strategy at the first turn is dominated by waiting to play after ones rival.

Proof. If B plays PCat the first turn or waits to play at his second turn, A is indifferent between playing (Y,at the first or second turn. For all other actions of B, A does better to wait and observe Bs play in order to play

according to his reaction function at the second turn. Thus, LX, is dominated by waiting. If B waits and plays fir, then A does strictly better to play aI than any other action. Thus, the Stackelberg equilibria with sequential play do not use dominated strategies. Q.E.D. The results here where the strategy space is an interval differ from those in Hamilton and Slutsky (1988) where each player has only two possible strategies. In that context, a unique trembling hand perfect equilibrium often exists. Whether the simultaneous move equilibrium is Pareto dominated by some pure strategy outcome is crucial to the results. Here, Pareto dominance is irrelevant and multiple equilibria always exist because of our assumption that the simultaneous move equilibrium and the two sequential equilibria are all distinct. In a 2 x 2 matrix game, a simultaneous move equilibrium in pure strategies must also be one of the sequential move equilibria. At least one player is then indifferent between waiting and moving first when his rival waits. In this paper, in the same circumstances, moving first is strictly better. In 2 x 2 games, the two extended games yield essentially the same results. Here the results in the two extended games are quite different.
we consider the stronger property that strategies remain undominated after iterated deletion of dominated strategies. If only one player has a dominant strategy and his payoff in the simultaneous move game is less than his best payoff under the dominated strategy, then that player can gain even if no pure strategy outcome Pareto dominates the simultaneous move Nash equilibrium. In the extended game of Section III, the player with the dominant strategy gains only when a Pareto improvement is possible.

44

HAMILTON

AND

SLUTSKY

With observable delay, the nature of the basic duopoly game is important for the results. The Pareto dominance of the simultaneous move equilibrium is an important consideration, and this in turn depends on the slopes of reaction functions. Simultaneous play and a unique sequential equilibrium are possible. In the game with action commitment, the properties of the basic duopoly game are irrelevant (as long as the simultaneous and sequential move equilibria all differ). The same multiple equilibria always exist. Dowrick (1986) compares the Cournot and Stackelberg leader-follower outcomes to Stackelberg warfare. His definition of Stackelberg warfare is both firms taking their leadership actions simultaneously. Clearly, this will not be an equilibrium phenomenon. While a coordination problem arises when both players want to be Stackelberg leaders in this extended game, Dowricks type of Stackelberg warfare does not occur in any realizations of a mixed strategy equilibrium. If there exists an equilibrium in which firms randomize over playing or delaying, the rivals probability of playing first affects a firms first period action. In quantity setting games, each firm would reduce its first period quantity below its leadership level because sometimes its rival will also act in the first period. Thus, if a mixed strategy equilibrium exists, firms first period actions differ from the Stackelberg leadership ones and the warfare outcome is not as extreme. With linear demand and equal constant marginal costs, the Cournot equilibrium is the only symmetric equilibrium, and no equilibrium, symmetric or asymmetric, with mixing over the timing of actions exists. In this case, no form of Stackelberg warfare ever occurs.

V.

CONCLUSIONS

Many economists have viewed the application of Nash equilibrium to duopoly with scepticism. The argument is that this equilibrium concept, based on zero conjectural variations, involves naive beliefs by firms about rivals responses and that a more general model would allow nonzero conjectural variations. Johansen (1982) argues persuasively that this criticism is unfounded. A firm can only react to what it observes. It ought not react to mere announcements or to preliminary choices which can be modified. It is impossible for every firm to react to its rivals actions. A reacting firm cannot itself have committed to an action to which its rival is reacting.8 It is possible for one firm to be a leader if it commits irrevocably
8 Another possible justification of a nonzero conjectural variation equilibrium is that it is the reduced form of a repeated game. However, Friedman (1983) argues that it is necessary to analyze explicitly such a repeated game. Arbitrary assumptions about the reduced form can easily be invalid.

ENDOGENOUS

TIMING

IN

DUOPOLY

GAMES

45

to an action first and the action is observed by the rival. Thus only two valid models exist for analyzing single play duopoly, a simultaneous play game and a sequential play game. Of course, many different variations of these can be constructed, depending upon the strategy spaces available to the players. Given the two available models, a crucial question then becomes which to use when modeling any particular situation. One line of analysis to answer this is to determine exogenous factors which would tend to make one firm a leader. For example, incumbents may be able to move before new entrants. Entrants can observe what incumbents have done before they move. Of course, if the existing firms had sufficient postentry flexibility, then entrants will not react to current choices but to their perception of postentry behavior. A second approach, as we have done here, is to view the order of play as endogenously determined and seek to determine factors influencing simultaneous versus sequential play. We developed two games which extend the basic duopoly model by allowing firms to choose the time of play in addition to choosing a specific action. In the game with action commitment, both sequential play subgames are outcomes in undominated strategies; the simultaneous play subgame uses dominated strategies. In the game with observable delay, only when both sequential play outcomes Pareto dominate the simultaneous play outcome are both sequential play subgames outcomes of equilibria in the extended game. In other cases, either the simultaneous play subgame is an outcome of the unique equilibrium in the extended game (neither sequential game payoff vector Pareto dominates that of the simultaneous play subgame) or one of the sequential play subgames is the unique outcome of the extended game (when only one sequential outcome Pareto dominates). When multiple perfect equilibria exist, no clear prediction is possible without further assumptions. However, when delay is observable, then the players may, by their equilibrium choices, play in a particular order that achieves a Pareto preferred outcome. Whether a Pareto dominating outcome exists depends on characteristics of the reaction functions and on the firms isoprofit contours in the action space. When firms supply homogeneous products, simultaneous play will occur in quantity setting games. If duopolists set prices, multiple equilibria (including both sequential play games) exist in the extended game. If one firm chooses price and the other chooses quantity, such as in a game due to Singh and Vives (1984), the unique equilibrium of this extended game has the quantity setter choosing an action first and the price setter second. In the game with action commitment, regardless of Pareto dominance or slopes of reaction functions, both sequential games are outcomes in undominated strategies of the extended game. Our model does not assert that either sequential or simultaneous play is

46

HAMILTON

AND SLUTSKY

preferable as a modeling assumption. It does unveil factors which are important in determining this. The choice of models should depend upon whether the firms have some flexibility in the time of moves, whether they can engage in preplay communication about time of play, whether they can observe delay by rivals, and finally, whether one, both, or neither of the sequential outcomes Pareto dominates the simultaneous move outcome.

REFERENCES BOYER, M., AND MOREAUX, M. (1987). Being a Leader or a Follower: Reflections on the Distribution of Roles in Duopoly, Int. J. Ind. Org. 5, 175-192. CHENG, L. (1985). Comparing Bertrand and Cournot Equilibria: A Geometric Approach, Rand J. &on. 16, 146-151. DOWRICK, S. (1986). van Stackelberg and Cournot Duopoly: Choosing Roles, Rand J. Econ. 17, 251-260. FELLNER, W. (1965). Competition among rhe Few. New York: Kelly. FRIEDMAN, J. (1983). Oligopoly Theory. New York: Cambridge Univ. Press. GAL-OR, E. (1985). First Mover and Second Mover Advantages, Inr. Econ. Reu. X,649652. HAMILTON, J., AND SLUTSKY, S. (1988). Endogenizing the Order of Moves in Matrix Games, University of Florida Working Paper 88-3. JOHANSEN, L. (1982). On the Status of the Nash Type of Noncooperative Equilibrium in Economic Theory, &and. J. Econ. 84, 421-441. MCANDREWS, J. J. (1988). Strategic Role Choice: An Economic Theory of the Formation of First Mover Organizations, Ph.D. thesis, University of Iowa. MCANDREWS, J. J. (1989). Strategic Role Complementarity, mimeo, Federal Reserve Bank of Philadelphia. REINGANUM, J. (1985). A Two-Stage Model of Research and Development with Endogenous Second-Mover Advantages, Int. J. Ind. Org. 3, 275-292. SINGH, N., AND VIVES, X. (1984). Price and Quantity Competition in a Differentiated Duopoly, Rand J. Econ. 15, 546-554. VAN DAMME, E. (1983). Rejnements ef the Nash Equilibrium Concept. Berlin: SpringerVerlag.