literature review research paper

American Economic Association is collaborating with JSTOR to digitize, preserve and extend access to The American Economic Review.

http://www.jstor.org

American Economic Association

Preemption Games: Theory and Experiment Author(s): Steven T. Anderson, Daniel Friedman and Ryan Oprea Source: The American Economic Review, Vol. 100, No. 4 (SEPTEMBER 2010), pp. 1778-1803 Published by: American Economic Association Stable URL: http://www.jstor.org/stable/27871274 Accessed: 04-02-2016 18:36 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/ info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

American Economie Review 100 (September 2010): 1778-1803

http-J/www. aeaweb. or g/article s.php?doi=10.1257/aer. 100.4.1778

Preemption Games: Theory and Experiment

By Steven T. Anderson, Daniel Friedman, and Ryan Oprea*

El Mut?n, perhaps the world's largest remaining iron ore deposit, was opened to private inves tors in the 1980s but, due to the high cost of developing the remote Bolivian site, there were no takers for two decades. In late 2005, spurred by rising commodity prices, the Brazilian company EBX finally seized the opportunity, preempting rivals based in China and India.1 Numerous similar examples can be found in the annals of mining and oil companies (Raymond F. Mikesell et al. 1971).

The strategic problem is as old as humanity: the value of a grove of figs fluctuates as the fruit

ripens (or is attacked by worms) and the first band of hunter-gatherers to seize it eats better than

neighboring bands. High tech firms today face similar issues when they introduce a new product. Delaying introduction might allow the product niche to expand, but new substitutes or changes in standards might shrink the niche. A rival could preempt the niche or it could vanish entirely due to a disruptive new technology (Clayton M. Christensen 1997). Similar considerations apply to retailers deciding when to open a "big box" outlet in a market too small to handle more than one

store, and also apply to academic researchers investigating a hot new topic. In this paper we study such situations both theoretically and empirically. We formalize them

as preemption games, using standard simplifications to put the strategic issues into sharp focus.2 In our games, the opportunity is available to a known number n + 1 of investors; it has a publicly observed value Vthat evolves according to geometric Brownian motion with known parameters; each investor has a privately known avoidable cost of investing; and the first mover preempts and obtains the entire value V.

Our model builds on an active literature reviewed in Marcel Boyer, Eric Gravel, and Pierre Lasserre (2004) that studies preemption contests for investments with option values. Unlike most of this literature (e.g., Steven R. Grenadier 2002; Helen Weeds 2002; Romain Bouis, Kuno J.M.

Huisman, and Peter M. Kort 2009), firms in our model are uncertain of their rivals' costs. A trade-off for this added piece of realism is that unlike some of this literature but like much clas sic auction theory, preemption is complete; the winner takes all of the returns from investment. A separate literature considers preemption in very different environments related to R&D (e.g.,

* Anderson: National Minerals Information Center, US Geological Survey, Reston, VA 20192 (e-mail: Sanderson?

usgs.gov); Friedman: Economics Department, University of California-Santa Cruz, 1156 High Street, Santa Cruz, CA 95064 (e-mail: [email protected]); Oprea: Economics Department, University of California-Santa Cruz, 1156 High Street, Santa Cruz, CA 95064 (e-mail: [email protected]). We are grateful to Jerome Bragdon and Todd Feldman for research assistance, to Nitai Farmer, Adam Freidin and James Pettit for programming support, and to the National Science Foundation (grant IIS-0527770) and the University of California for funding. We received helpful comments from Robert Jerrard, Vijay Krishna, Robert B. Wilson, James C. Cox, Donald Wittman, participants at the 2008 UCSB

Experimental and Behavioral Workshop, from seminar audiences at Cal Poly SLO, CUNY-Baruch, George Mason,

Houston/Rice, and Texas A&M Universities, and from three anonymous referees of this Review. All mistakes are our own.

1 Arai, Adriana, and Andrew J. Barden. 2006. "Bolivia Rules Out Brazil's EBX for $1.1 Bin Project (Update2)."

Bloomberg, April 24.http://www.bloomberg.com/apps/news?pid=10000086&refer=latin_america&sid=a2LjvfEFCDW 2 Reality, of course, is always more complex. In the El Mut?n example, Bolivia's newly elected government shut

down EBX's operations in 2006, citing environmental and other concerns, and in late 2007 signed a 40-year concession contract with the Indian firm Jindal. Other parts of the story are more consistent with our model. The number of serious rivals was always reasonably clear. The true costs of the Brazilian and Indian firms (and their Chinese rival, Shandong)

1778

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

VOL. 100 NO. 4 ANDERSON ETAL.: PREEMPTION GAMES 1779

Jennifer F. Reinganum 1981; Drew Fudenberg and Jean Tirole 1985; Heidrun C. Hoppe and Ulrich Lehmann-Grube 2005) and to market entry (e.g., Dan Levin and James Peck 2003).

Bart M. Lambrecht and William R.M. Perraudin (2003) is our direct predecessor. Like us,

they draw on real options theory to investigate preemption of a stochastic investment opportunity by competitors whose costs are private information. Our theoretical results extend theirs by cov

ering more than two competitors, and by relaxing a restrictive technical assumption. Our model, unlike theirs, is explicitly rooted in auction theory as well as in real options theory. Indeed, spe cial cases of our model include Dutch auctions as well as deferral options.

Our theoretical contribution appears in Section I. We characterize the symmetric Bayesian Nash equilibrium of the preemption game with an arbitrary number of players. Players' BNE

strategies take the form of a threshold value at which the opportunity is seized immediately. The

mapping from realized cost to equilibrium threshold is characterized in two different ways: by an ordinary differential equation and also by a recursion equation. We show that the auction and real option special cases lead to useful bounds on the BNE strategies.

Section II describes a laboratory experiment informed by the theory, using software cre ated expressly for the purpose. It presents the main treatments?Competition (triopoly) versus

Monopoly, and High versus Low Brownian parameters?and obtains four testable hypotheses. Section III explains other aspects of the laboratory implementation.

Section IV presents the results. The first three hypotheses fare quite well: the triopoly market structure leads to much lower markups (threshold value less cost) than the monopoly structure; the Brownian parameters have a major impact in the predicted direction in Monopoly but (again as predicted in BNE) have negligible impact in Competition; and the lowest cost investor indeed is far more likely to preempt than her rivals. The evidence is mixed on the last hypothesis: at the low cost end of the scale, investors' markups indeed tend to decline in cost, but the predicted relationship breaks down at higher costs. However, these departures from prediction turn out to have very little impact on subjects' actual earnings.

Following a concluding discussion, an Appendix collects mathematical derivations and the main proofs. Other Appendices available online provide additional mathematical details, discuss the lesser-known econometric techniques, report supplementary data analysis, and reproduce instructions to subjects. Our theoretical contribution incorporates and extends the PhD disserta tion of Anderson (2003). A companion paper, Oprea, Friedman, and Anderson (2009) describes a related laboratory experiment concerning the monopoly (n

= 0) case only.

I. Theoretical Results

This section analyzes two situations. In the first, called monopoly, a single investor / has sole access to an investment opportunity. In the second, called competition, two or more investors with private information concerning their own costs have access to the same opportunity, and the first to seize it renders it unavailable to the others.

were, in no small part, private information due to confidential subsidies arranged by their own governments as well as

confidential understandings with the Bolivian government. The firms faced the hazard that the Bolivian government

might renationalize El Mut?n, or declare it a protected national park, before any of them could seize the opportunity

(Barden and Arai 2006; Patrick J. McDonnell 2008. "Bolivia opening up huge iron ore deposit to mining." Los Angeles Times, May 28. http://articles.latimes.com/2008/may/28/business/fi-iron28 (accessed October 31, 2008).).

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

1780 the american economic review september 2010

A. Monopoly

An investor / with discount rate3 p > 0 can launch a project whenever she chooses by sinking a given cost Ct > 0. The present value V of the project evolves via geometric Brownian motion with drift parameter a < p and volatility parameter a > 0:

(1) dV= aVdt + aVdz,

where z is the standard Wiener process. That is, the value follows a continuous time random walk in which the appreciation rate has mean a and standard deviation o per unit time. At times t > 0

prior to launching the project, the investor observes V(t). If she invests at time t, then she obtains

payoff [ V(t) ?

Ci]e~pt. The project is irreversible and generates no other payoffs. Thus, the task is to choose the investment time so as to maximize the expected payoff.

The solution goes back to Claude Henry (1974) and has been widely known since Robert McDonald and Daniel Siegel (1986); see Chapter 5 of Avinash K. Dixit and Robert S. Pindyck (1994) for a detailed exposition. The optimal policy takes the form: wait until V(t) hits the threshold

(2) VM(Q) = (1 + w)Ct,

then launch immediately. Note that the threshold is proportional to cost, and that the wait option premium w > 0 is an algebraic function of the volatility, drift and discount parameters <r, a and p.

Specifically,

(3) w =

j^j, where ? = \

- +

B. Competition

Now consider the case that each investor has n > 1 rivals. All investors / = 1,2,...,ai -f- 1 have access to the same investment opportunity, whose value V again evolves according to geo metric Brownian motion (1). Each investor / again knows her own cost Ch but doesn't know the other investors' costs Cj9j ^ i. She regards them as independent draws from a cumulative dis tribution function, H(C), with a positive continuous density function h(C) on support [CL, q/]. The first investor to launch, say at time tt > 0, obtains payoff [V(f/)

- C^e~p%i and the other

investors obtain zero payoff. All this is common knowledge. The resulting preemption game is denoted r(?9n,ti).

The preemption game has a unique symmetric Bayesian Nash Equilibrium (BNE). It is char acterized by an increasing function V*(C/) that maps the investor's cost into a threshold value, above which she immediately invests. We now sketch the derivation and offer some intuition; the

Appendix spells out the derivations and the main proofs.

a + 2p J2. > 1.

3 Recall that the discount rate reflects pure time preference, the expiration hazard, and possibly risk aversion. As

explained in the next section, the laboratory experiment focuses on the expiration hazard, sometimes informally referred to as "preemption by Nature."

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

VOL. 100 NO. 4 ANDERSON ETAL.: PREEMPTION GAMES 1781

Using notation m, V0 and C defined below, the expected discounted payoff E[ V{t) can be written out as the following objective function:

(4) F(m\Chn) = \V*(m) -

?] V\m)

1 - H{m) 1 - H(C)

The choice variable in (4) is m G [CL, Cv], interpreted as the cost-type that the investor chooses as her potential "masquerade."4

The first factor in the objective function (4) is simply the profit (or "markup") [V*(m) ?

C,] obtained at the time of successful investment. The second factor, [V0/V*(m)]?, accounts for the time cost of delaying investment and the expiration hazard, given that V0 is the current value of the investment project. The Appendix shows that the monopolist's value function consists of only these first two factors. It also notes that with competitors (n > 0), the restriction p > a can be relaxed and consequently (4) is valid for ? > 0, while of course (2-3) is valid only for ? > 1.

The third and final factor, [(1 -

H(m))/(l -

H(C))]n, is the probability that the n rivals all have higher costs (and therefore will not preempt), conditioned on the fact that none of them has

already invested. That conditioning is reflected in the denominator. Let V > VQbc the "highest peak" so far achieved by the random walk. Then C is the corresponding cost, i.e., V

= V*(C).

Since the preemption game is over as soon as the first investor moves, it turns out that the BNE threshold strategy is independent of V, C and V0 within the relevant range.

The key to obtaining the BNE is the best response (or "truthtelling") property that investor i maximizes (4) at m

= Ct. The associated first-order condition can be expressed as the following

ordinary differential equation (ODE):

(5) V"(Q = nh(C.) [V(Q

- C,]V(Q

For reasons explained in the next subsection, we also impose the boundary condition

(6)

This boundary value problem has a unique solution V* that characterizes the symmetric BNE threshold for our preemption game. The Appendix shows that it can also be expressed as a con

ditional expectation and that it satisfies

(7) V*(C) = c + Jc [v\y)\ 1 - H(y)

Li-ff(c)j dy.

THEOREM 1: Let the cumulative distribution function H have a continuous density h with full support [CL, Cy], where 0 < CL < Cv < oc. Let it be common knowledge among all investors i = 1,..., n + 1 that Vs investment cost C, is an independent random variable with distribution H

and is observed only by investor i. Then

1. for any ? > 0, boundary value problem (5-6) has a unique solution V* : [CL, Cy] ?> R,

4 Of course, the threshold value itself is the natural choice variable but, as in auction theory, it turns out that m is

more convenient since V*(-) is invertible.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

1782 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2010

2. a function V* satisfies the recursion equation (7) if and only if it solves the boundary problem (5-6), and

3. the premption game T[?,n,H] has a symmetric Bayesian-Nash equilibrium in which each investor i's threshold is V* evaluated at realized cost Cv

To obtain V* numerically, one can use the Euler method of integrating the ODE (5) backward from the upper boundary value (6). Alternatively, one can take an initial approximation (such as

the auction solution V defined in the next subsection), substitute it for V* in the last expression in

(7) to obtain a better approximation, and iterate. The BNE threshold function V* is a fixed point of this mapping.

Lambrecht and Perraudin (2003) use slightly different methods to derive an ODE that is con sistent with equation (5) for the special case that (a) there are only two investors (duopoly), and (b) H has an increasing modified hazard rate Ci[h(Ci)]/[l

? H(Ci)]. Their methods rule out

asymmetric BNE. Our method considers only symmetric BNE but has several compensating advantages. It covers (n -f 1) > 2 investors, allows arbitrary hazard rates, connects explicitly to auction theory, and leads to a streamlined and unified analysis of useful special cases.

Theorem 1 assumes that H has a continuous positive density on the entire support interval, and that the upper endpoint is finite and the lower endpoint is positive. It follows from Lusin's theorem (e.g., Walter Rudin 1966, 53-54) that such functions are a dense subset of all distribu tion functions on [0, oo), so these assumptions are not especially onerous. Still, it would be of interest to explicitly consider distributions with discrete support (e.g., a finite set of cost "types") or asymmetric cases in which some investors are known to have different cost distributions than other investors. Perhaps "ironing" techniques in the spirit of Roger B. Myerson (1981) would be useful for such extensions.

Consider again the special case n ? 0, monopoly. Since n is a factor in the numerator in

(5), it might seem at first that the expression is identically zero. However, the last paragraph of the Monopoly subsection of the Appendix (ending with equation (21)) notes that in this case the denominator V*(C,-)

? ?[V*(Ci)

? C,]

= 0 and that this last expression characterizes the

monopoly solution. The Appendix shows that, at any realized cost Ch the monopoly solution VM(Ci) is an upper

bound on the competitive solution V*(C,-). The intuition is simple but revealing. In monopoly, an

investor increases her threshold up to the point that the greater profit margin just balances the

greater threat of "preemption by Nature," i.e., the expiration hazard. With competition, the inves tor must also consider preemption by other investors, so she finds the balance at a lower threshold.

The special case ? = 0 is also instructive. Here the middle factor of the objective function (4)

disappears, yielding

Apart from the denominator, which disappears in the first order condition, this is a familiar

objective function from auction theory. Thus the ODE (5) collapses to the ODE for bid functions in a reverse Dutch (or reverse first price) auction:

C. Special Cases, Intuition, and Bounds

(8) F(m\C?n) = [V? -

C,] 1 - H{m) T

A-H(C).

(9) v"(Q = [V(Q - q

nh{Ci) [l-H(Q]

'

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

VOL. 100 NO. 4 ANDERSON ET AL.: PREEMPTION GAMES 1783

The boundary condition (6), of course, remains unchanged. Hence, with ? = 0, we obtain an

isomorphism between threshold values in preemption games and bids in auctions. The "auction" solution V(Q) to boundary value problem (9, 6) is another upper bound on the

solution to the more general boundary value problem (5, 6). The intuition again is that the inves tor chooses the threshold to balance greed (a larger profit margin) and fear (of preemption) at the

margin. Completely eliminating a preemption threat (in this case from Nature, by setting ? = 0,

e.g., by setting a > 0 =

p) sets an upper bound on the balance point. The BNE threshold V*(C^) also has a natural lower bound, given by the classic Marshallian

(i.e., zero net present value) investment rule V?(C/) =

Ct. Recall that the boundary condition (6) imposes this rule, but only at the highest possible cost Cv. The intuition for the boundary condi tion is compelling. An investor with the highest possible cost faces Bertrand competition: every rival will find it profitable to undercut any positive markup she might seek.

These bounds and special cases are summarized in the following

THEOREM 2: Under the hypotheses of Theorem 1, the BNE threshold V*(Q) is an increas

ing, continuously differentiable function bounded below by the Marshallian threshold function

V?(Ci) =

Ci and bounded above by the monopoly threshold VM(Q). It is also bounded above by the solution V(Ct) to (9, 6) and is tangent to V(C,) at the upper endpoint Cv.

Finally, consider special cases of the cost distribution. If H has a constant density (i.e., is uni

form) or a continuous decreasing density, then it can be shown (see the Online Appendix) that the BNE threshold function V* is concave.

Our experiment employs the uniform distribution H on [CL, Cv\. In this case, (5) reduces to

[V\Q -

Q]V(Q XQ

- ?[V\Q

- C

In the special case ? ? 0, we then get the well-known ODE from William Vickrey (1961)

(11) vxc^lViQ-Clf^-r,

with analytic solution

nQ + Cv

('<?) no -1 vh-ma - ci) l^c

(12) V**(Q) n + 1

COROLLARY 1: Let H be the uniform distribution on [CL, Cv], let n be an integer > 1, and let

V**(Ci) be given by Equation (12). For ? > 0, the Bayesian-Nash equilibrium threshold function

V*(C,), the solution to (10, 6) on [CL, Cv], is concave, bounded above by V**, tangent to V** at

CUf and bounded below by the Marshallian threshold function V7?(Q) =

Ct.

Thus when the cost distribution is uniform, an upper bound (and often a good approximation) of the BNE threshold V* is the function

(13) V^Q^miniV^V^C,)}.

The monopolist threshold binds in (13) at lower cost realizations for high values of ? and low values of n.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

1784 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2010

Panel A Panel B

Figure 1. BNE Threshold Values for Triopolies with Costs Drawn Independently from the Uniform Distribution on [50,80]

Note: "High" refers to the case ? = 2.25 with option premium w

= 0.8, and "Low" refers to the case ? = 3.00 with

option premium w = 0.5.

It is well known from auction theory that the bid function V(C) converges to the Marshallian threshold function V?(C)

= C as the number of rivals n ?> oo. Theorem 2 implies that the same is true of the BNE threshold function V*(C). Since the bound V converges to the lower bound V? in the large number limit, we conclude that the BNE markups V*(C)

? C converge to zero as the number of rivals increases.

D. Numerical Example

Figure 1 shows the numerical solution V* to the boundary value problem (10, 6) for triopoly (n = 2), given the uniform cost distribution on [CL, Cv}

= [50,80], for ?

= 2.25 and ? = 3.00.

The first ? value is referred to as "High" because it generates higher monopoly thresholds (via option premium w

= 0.8) than the second ? value, which has w

= 0.5 and is referred to as "Low." Panel A includes the monopoly thresholds VM. Panel B shows the BNE threshold func tions at finer resolution, together with the Marshallian lower bound V?(C/)

= C, and the Vickrey

upper bound V**(Ct). One can see that the two BNE threshold functions lie close together and

that, as claimed in Theorem 2 and Corollary 1, both are tangent to the Vickrey bound at the upper endpoint.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

VOL. 100 NO. 4

Table 1

ANDERSON ETAL.: PREEMPTION GAMES

I?Binomial Parameters and Number of Subjects Studied in Each Treatment

1785

Treatment 77 p q Replications

Low 0.03 0.524 0.007 36 subjects

High 0.03 0.513 0.003 36 subjects

E. Discrete Approximations

Brownian motion is an idealization. Our experiment uses a close binomial approximation of the continuous time process. Specifically, it has a fixed time interval At

= 0.003 minutes (i.e., 200 milliseconds) for each discrete step of the value path, and three binomial parameters:

(i) the step size 77 > 0 of the proportional change in value, i.e., the current value V becomes either (1 + 77)Vor (1

? 77)Vat the next step;

(ii) the uptick probability p G (0,1), i.e., the probability that the next step is to (1 + 77) Vrather than to (1

? 77) V; and

(iii) the expiration probability q G (0,1), i.e., the probability that the current step is the last, and the opportunity disappears.

As noted in the Online Appendix, the discrete binomial parameters (p,q,rj,At) map into the Brownian parameters (a,cr,p). The Brownian parameters in turn have a sufficient statistic, ?, which determines the monopoly (n

= 0) threshold and the BNE (n > 0) threshold. Thus, our

choices of n and the binomial parameters in the laboratory allow us to manipulate the equilib rium thresholds and thus to test the theory's predictive power.

II. Treatments and Hypotheses

Two binary treatment variables, Parameters and Structure, allow us to test the major predic tions of the model. Parameters fixes the time step at At

= 0.003 (in minutes) and the step size at 77

= 0.03. Parameters = Low is shorthand for p = 0.524 and q = 0.007, corresponding to ? ? 3.0 and option premium w ? 0.5, as in the numerical example of the previous section.

Likewise, Parameters = High is shorthand for p = 0.513 and q = 0.003, corresponding to

? ? 2.25 and w ? 0.8. These configurations differ considerably from each other, yet both yield value paths "in the money" often enough, and jagged enough, to maintain subjects' interest. Table 1 summarizes this treatment variable for easy reference.

The second treatment variable is Structure. When Structure = Monopoly, the subjects make investment decisions with no rivals (n

= 0) and therefore no risk of preemption by another

investor. When Structure = Competition, the subjects compete in triopolies (n =

2). In each

Competition period the subjects are randomly reassigned to one of three or four separate markets, each with three investors.

Each period each subject's cost is drawn independently from t/[50,80], the uniform distribu

tion with support [50, 80]. Sessions begin with a ten-period Monopoly block, Monopolyl, con

tinue with 25 periods of Competition, and end with Monopolyll, another ten-period Monopoly block. The data analysis focuses on the Monopolyl (hereafter simply Monopoly) and Competition block. The online Appendix analyzes Monopolyll data and obtains parallel (but somewhat more

diffuse) results.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

1786 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2010

50 55 60 65 70 75 80 Cost

Figure 2. Predicted Markups by Treatment

Figure 2 plots markups, i.e., threshold value less cost. For both High and Low parameters, the dotted lines represent Monopoly markup, V*M(C)

? Ch and the solid lines represent BNE mark

ups in Competition, V7*(Ci) ?

Cv The figure shows that the Competition markups are every where much lower than the corresponding Monopoly markups. Our first hypothesis is that the

markups observed in the experiment will have the same ordering as these theoretical constructs.

HYPOTHESIS 1: (Structure.) Markups chosen by subjects in the Monopoly treatments signifi cantly exceed markups chosen in the Competition treatments.

Another striking aspect of Figure 2 is that the Monopoly line for the High parameter vector

is far above the corresponding Low line, while under Competition the two lines are very close

together. The second hypothesis is that the experimental data will reflect this aspect of the theory.

HYPOTHESIS 2: (Parameters) Markups chosen in the High Monopoly treatment significantly exceed those in Low Monopoly, while chosen markups have the same distribution in the High Competition data as in Low Competition.

Recall that the symmetric BNE strategy V*(Cf) is increasing in cost Cf. A direct implication is that the investment opportunity is always seized by the lowest cost investor. Allowing for some

behavioral noise, we obtain the following efficient sorting hypothesis:

HYPOTHESIS 3: (Efficiency) Under Competition, the most efficient (lowest cost) subject is the one most likely to preempt the others.

A final observation from Figure 2 is that, under our parameters, Competition markups decrease in costs. The reason is that the BNE threshold slope V*'(Cf) < 1; indeed the approximate thresh old function (13) has a slope of 2/3, so the theoretical markup slope in Competition is close to

-1/3.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

VOL. 100 NO. 4 ANDERSON ETAL.: PREEMPTION GAMES 1787

Figure 3. Subject Screen under Competition at the End of a Period

HYPOTHESIS 4: (Monotonicity.) Under Competition, observed markups are decreasing in cost.

A complementary hypothesis, drawn from real options theory and observable in Figure 2, is that Monopoly markups are instead increasing in cost. This is a central hypothesis tested by

Oprea, Friedman, and Anderson (2009) who find strong empirical support.

III. Implementation

Experiments were conducted using customized software called InvestmentTiming. Figure 3 shows the user interface. The lightly shaded band (colored blue on subjects' screens) indicates the cost range, [50,80], which was held constant throughout the session and announced publicly. The horizontal line (colored red on subjects' screens) represents the subject's own cost that period; its status as private information was also announced publicly.

The current value of investment, V(t), was represented by a jagged line (colored green on sub

jects' screens) that evolved from the right, as on a seismograph. During each period the value line was initialized at 50 (the lower bound of the cost distribution) and evolved from there according to the binomial parameter vector, High or Low, chosen for that session. The vertical axis rescaled if the value line ever rose out of the displayed bounds.

Subjects were not allowed to invest when the value line was below their own cost, to prevent negative earnings, nor could they invest after the random ending time. At all other times, subjects could attempt to invest by tapping the space bar at their computer terminal.

In the Monopoly treatment, an investment attempt prior to period end was always successful,

immediately netting a subject V(t) ?

C, points. Subjects in the Competition treatment were not

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

1788 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2010

told whether or when their competitors invested until after the period was over.5 This semi

strategy method gives us access to more data, while still giving subjects the real-time choice

experience that we feel helps them adapt to the stochastic environment.6 After the period ended, subjects were told the time at which each subject in their group

attempted investment, the value at which they attempted investment, the costs of each competi tor, and the resulting profits: V(t)

? Cj to the subject who invested first, and zero to the others. In

the Monopoly treatment, of course, there were no other subjects in the group. All cost draws, value sequences and period endings were made only once for each parameter

set, and were repeated in all sessions for that treatment. In one session under High parameters, a software malfunction during period 30 (toward the end of the Competition block) led to four

missing periods which have been dropped from the dataset.

Subjects were given instructions in two parts, reproduced in the Online Appendix. First, instructions pertaining to monopoly periods were distributed and read aloud. Binomial param eters and exchange rates for the session were written on a white board and were pointed out several times before play began. Following six unpaid practice periods, we conducted the first

Monopoly block. Then the second part of the instructions, pertaining to Competition, was dis tributed and read aloud. The instructions were identical across sessions; only the parameters written on the white board varied.

Experiments were conducted at the University of California, Santa Cruz, using inexperienced undergraduate subjects recruited from a large online database of volunteers. Subjects each were

paid a $5 showup fee. They earned five cents per point in Monopoly periods and (to maintain a comparable payout rate in triopoly) 15 cents per point in Competition periods. Low sessions

averaged one hour and 30 minutes, while High sessions averaged close to two hours. On average,

subjects in the Low treatment earned $15.21, and subjects in the High treatment earned $12.57.

IV. Results

Subjects frequently failed to invest prior to the random end of the period. Efforts to invest at

particularly high values are therefore censored in many periods, while efforts to invest at lower values are more likely to be observed. Consequently the sample of observed investment decisions is downward biased. We employ two different techniques to correct the bias. The first is to construct product limit

(PL) estimates (E. L. Kaplan and Paul Meier 1958) of the empirical cumulative distribution func tion of markups. As explained in the Online Appendix the PL procedure uses all data (including all instances of censoring) to construct maximum likelihood nonparametric estimates. The esti mates are graphed for each treatment in Figure 4.

Panel A of Figure 4 suggests that investment behavior is considerably different under

Competition than under Monopoly. For example, the median markups (where the graph crosses the horizontal line at 0.5) are about 5 or 6 for both Competition treatments versus about 12 for Low Monopoly and about 21 for High Monopoly. The observed ordering seems consistent with the first two hypotheses.

Panel B of Figure 4 compares observed distributions (solid lines) with the predicted distribu tions (dashed lines), obtained as PL estimates on artificial data generated by applying theoretical functions V* and to the realized cost draws. In the Competition panel, the predictions for

High and Low parameters are very close together. The CDFs for observed markups for High

5 As the Appendix Remark Al shows, this design choice does not alter the BNE. 6 Implementing the full strategy method would require us to constrain the strategy space, e.g., to a choice of thresh

old V(C), excluding, a priori, nonstationary and other sorts of strategies that subjects might otherwise use.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

VOL. 100 NO. 4 ANDERSON ET AL.: PREEMPTION GAMES 1789

Panel A

0 10 20 30 40 50 60 701 > / 4o

Markups

Panel B Competition Monopoly

Markups Markups

Figure 4. Product Limit Cumulative Distribution Functions

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

1790 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2010

and Low parameters are also close to each other, and only slightly more diffuse than the predic tions. In the Monopoly panel, the predictions for High parameters are about 15-20 points higher

(i.e., to the right of) those for Low parameters at each percentile. Above the thirtieth percentile, the CDFs for observed markups have the same ordering and about the same spacing above the

eightieth percentile, but for the most part they fall well below (i.e., to the left of) the theoretical

predictions. Our second approach to the censoring problem is to examine a data subsample in which, for

exogenous reasons, censoring is rare. Theory guides the subsampling: in BNE for Competition there is no censoring in periods in which the maximum, v, of the value line is above 80. Likewise in the Monopoly prediction there is no censoring when v exceeds 120 in the Low treatment or 144 in the High treatment. We refer to the sets of such periods as "feasible" samples, since in theory investment was feasible for all investors. Note that v is uncorrelated with cost and, because it is unknown to subjects ex ante, it should also be uncorrelated with the chosen markup.

The feasible samples turn out, in fact, to be nearly uncensored. In Competition periods, we observe nearly 97 percent of 660 decisions in the subsample (versus 85 percent of 1,039 deci sions overall), while in Monopoly periods we observe over 98 percent of 72 decisions (versus 61 percent of 519 overall). Therefore standard econometric tools can be used confidently on the feasible samples.

A. Treatment Level Hypothesis Tests

The PL estimates in Figure 4 seem consistent with the first two hypotheses. To formalize the

inferences, we estimate the following nested random effects regression on the feasible sample:7

(14) Vjit ?

Cjit = 7 + ipHighj + nMonopolyt + ?Highj Monopolyt + i/j + ut + ejit.

The left-hand side is the observed markup, Highj is an indicator variable taking a value of 1 in High parameter set sessions and 0 otherwise, and Monopolyt is an indicator variable for

Monopoly periods. The variable Vj is a random effect on sessions and ut is a random effect on individual subjects, both assumed to be normally distributed with a mean of zero. Together these account for within subject and within session correlations. Finally, ejit is a normally distributed,

mean zero disturbance term.

Results are displayed in Table 2. Estimates of treatment effects are constructed from the coef ficient estimates as follows: Competition Low

= 7; Competition High =7 + 7/;; Monopoly Low = 7 + n\ and Monopoly High

= 'y + i/j + K + S. Thus Hypothesis 1, that Monopoly markups

exceed Competition markups, translates to 7 + n > 7 (or k > 0) for Parameters = Low, and

translates to 7 + ip -f n + S > 7 + i> (or n + ? > 0) for Parameters =

High. Table 2 confirms the Low case: k is significantly greater than zero at the one percent level. A Wald test confirms the High case: k + ? is greater than zero (p

? 0.000).

FINDING 1: Consistent with Hypothesis 7, markups are significantly lower under Competition than under Monopoly given either Low or High Parameters.

The second hypothesis is that investment values are sensitive to binomial param eters under Monopoly but not under Competition. Equation (14) translates the hypothesis as

7 Here we pool data from all treatments and make cross treatment comparisons. For consistency, we uniformly apply

the most stringent criterion, v > 144. Using the less stringent criteria where applicable (e.g., v > 80 in Competition blocks) does not appreciably change the estimates and does not alter the results of the hypothesis tests reported below.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

VOL. 100 NO. 4 ANDERSON ETAL. : PREEMPTION GAMES

Table 2?Estimates (and standard errors) from Nested Random Effects Model (14)

1791

Variable Coefficient Estimate

Intercept 7 5.929

(1.542)***

High ^ -0.866

(1.839) Monopoly K 12.979

(2.128)***

High x monopoly 5 9.184

(2.784)***

Note: Variables High and Monopoly are dummies for the Parameter and Structure treatment variables.

*** Significant at the 1 percent level.

** Significant at the 5 percent level.

* Significant at the 10 percent level.

y + ^ + K + ?>~/+K(pr'ip + ?>Q), but 7 H- ̂ is no greater than 7 (or ip is insignificantly different from zero). A Wald test indicates that i? + S is indeed significantly larger than zero

(p =

0.019). Table 2 reports that -0 is insignificantly different from zero. Thus:

FINDING 2: Consistent with Hypothesis 2, investment values in Monopoly periods are signifi cantly larger under High parameters than Low parameters, while there is no significant differ ence in Competition periods.

B. Preemption, Markups and Cost in Competition

The third hypothesis is that the lowest cost investor (the efficient one) will usually preempt her rivals. Figure 5 shows the fraction of times that the preemptor has the lowest, middle, and highest cost. In both treatments the lowest cost investor wins roughly 80 percent of the time, while the

highest cost investor wins about four percent of the time in the Low treatment and hardly ever in the High treatment. Equally important, the exceptional cases are mostly when costs are very close. On average, the difference between the middle and low cost draws is (Cv

? CL)/(n + 2)

= (80

- 50)/(3 + 1)

= 7.5 points, but in periods when a second lowest cost subject preempts the lowest cost subject, the median difference is only 2.5. Likewise, when a highest cost subject preempts the lowest cost subject, the median difference is only 4, as compared to an a priori aver

age of (2) (7.5) = 15 points. To summarize,

FINDING 3: Consistent with Hypothesis 3, the lowest cost investor usually preempts the other

investors, and the highest cost investor rarely preempts the other investors.

Hypothesis 4 predicts that under Competition, the target markup M =

V(C) ? C is decreasing

in the cost of investment C; indeed, the BNE functions decline almost linearly. In testing this

hypothesis, we focus on the feasible subsample (periods in which v > 80). Figure 6 shows the

average markups by cost range for each treatment. The first impression is that there is no relation, and that markups are approximately constant. To check this impression, we estimate the follow

ing nested random effects regression

(15) va ~

Ca = a + bCit + Vj + it,- + eit

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

1792 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2010

Lowest r Second Highest

Cost rank

Figure 5. Probability of Preempting Both Rivals as a Function of Cost Rank

Note: The fraction of cases in which the highest cost competitor successfully invests in the

High treatment is 0.006, too small to be seen in the figure.

where w, and Uj are random effects on subjects and session respectively (both assumed normal with mean zero) and e is an error term distributed N(0,o%). The column labeled (1) in Table 3

reports the cost coefficient estimates, which are negative but significantly smaller (in absolute

value) than the predicted coefficients shown in the previous column. Indeed, one can see from the standard errors that these coefficient estimates are not significantly different from zero.

Moreover, the intercept estimates are less than half of the BNE intercepts. An alternative hypothesis is that subjects use cognitively simpler constant markup strategies.

The Online Appendix characterizes the Nash Equilibrium (NE) in threshold strategies con strained to be constant across costs. In the NE of the constrained game, the slope coefficient is of course zero and the intercept turns out to be 7.15 (resp. 7.81) in the Low (resp. High) treatment. In column 1, under both treatments, the slope is insignificantly different from zero, and 95 percent confidence intervals for the intercepts cover the NE values. Thus, column 1 is consistent with the constrained Nash equilibrium.

Before rejecting Hypothesis 4, however, one should recall that the incentives for playing BNE

strategies are not constant across costs. Figure 6 includes a dashed line charting the expected earnings from BNE play according to equation (4). Earnings drop rapidly, so deviations from the BNE strategy are cheap at higher costs. Could weak incentives at higher costs confound our statistical inference? A closer look at Figure 6 confirms that the BNE prediction does much better in the lower cost part of the range, where expected earnings are far larger. Indeed, in the Low treatment, the open circles lie parallel to the prediction line over the range E(ttbne) > 1 (or C < 60), and in the High treatment this relationship continues over the wider range E(ttbne) > 0.1 (or C < 70).

To follow up, we rerun (15) on the subsample in which E(irBNE) > 0.1, and report the results in column 2 of Table 3. As predicted in Hypothesis 4, the slope coefficient is indeed much more

negative than in column 1 and is quite significant in the High treatment. In the more stringent subsample E(irBNE) > 1.0 reported in column 3, the estimated slope coefficient is almost exactly the BNE value in the Low treatment and is marginally more negative than the BNE value in the

High treatment.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

VOL. 100 NO. 4 ANDERSON ETAL.: PREEMPTION GAMES

Panel A. Low Panel B. High

1793

10 4

Q_ 13 co

2

BNE markup - BNE expected earnings

10 4

cl 3

Cost Cost

Figure 6.

Notes: Average observed markups by cost interval are shown as open circles for the feasible

subsample, v > 80, of Competition periods. The BNE markups appear as solid lines, and the

expected earnings from BNE play (given by equation (4)) appear as dashed lines.

To extend this line of thought, suppose that behavioral noise is inversely related to expected earnings. Specifically, assume e

~ N(0, o-2v/E(ttbne)), i.e., the error term e is heteroskedastic with a

variance inversely proportional to expected earnings in BNE, and re-run (15) accordingly. Column 4 in Table 3 reports the results. In both treatments, there is a negative relationship between markups and costs that is economically and statistically significant. In the High treatment, both intercept and slope are indistinguishable from BNE levels by Wald tests. In the Low treatment the empirical intercept is lower than the BNE intercept and the slope term is somewhat less negative.8

FINDING 4: Except at higher costs (where expected payoff is very small), the observed markups are consistent with BNE in the High Competition treatment, and they also decline in cost in the

Low treatment Overall, the payoff-weighted evidence supports Hypothesis 4.

Finding 4 begs the question of how much money subjects leave on the table due to behavioral

noise. Would subjects significantly improve earnings by reducing the noise? To put it another

8 Our estimates also imply the predicted positive relationship between value and cost. Our markup regressions are of

the form Vit -

Cit ? a + ?Cit. Adding Cit to both sides we see that Vit

= a + (? + 1)C?. Adding 1 to any of our cost

coefficient estimates from Table 3 yields a positive number confirming the model's prediction.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

1794 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2010

Table 3?Coefficient Estimates (and standard errors) from Nested Random Effects Regressions (15) for the Feasible Sample, v > 80

BNE Overall ?(tt) > 0.1 ?(tt) > 1

(1) (2) (3) Weighted

(4) Panel A. Low

Estimates

Intercept

Cost

Wald tests

Intercept = BNE

Cost = BNE

Observations

20.65

-0.253

8.148***

(1.878) -0.0377

(0.0282)

0.0000

0.0000

386

11.096***

(3.142) -0.0872*

(0.051)

0.003

0.001

267

21.4388**

(8.281) -0.2742*

(0.149)

0.924

0.887

117

14.392***

(2.876) ?0.144***

(0.050)

0.030

0.029

386

Panel B. High

Estimates

Intercept

Cost

Wald tests

Intercept = BNE

Cost = BNE

Observations

22.14

-0.273

7.9283***

(2.540) -0.0561

(0.0391)

0.0000

0.0000

251

15.8773***

(3.239) ?0.192***

(0.053)

0.055

0.129

165

25.2618***

(8.143) -0.359**

(0.144)

0.703

0.552

92

18.1915***

(3.397) ?0.229***

(0.0583)

0.246

0.451

251

Notes: Wald Tests (e.g., "Intercept =

BNE") report p-values from Wald test of the equality of estimated coefficients to the BNE coefficients. Columns 1-4 report results for the full feasible sample, for subsamples in which BNE play pro vides expected returns exceeding 0.1 and 1, and for a weighted regression in which error amplitude is assumed to be

inversely proportional to expected return in BNE.

***Significant at the 1 percent level.

**Significant at the 5 percent level. * Significant at the 10 percent level.

way, is BNE a better response to the empirical distribution of investment choices than subjects' actual strategies?

To answer such questions, we calculate the gains to BNE: what each subject would have earned had she unilaterally played the BNE throughout the experiment, less what she actually earned.

Figure 7 shows that the median gain is quite small: about 1.24 points (or $0.18) per session in Low and 1.52 points (or about $0.22) in High. A Wilcoxon test cannot distinguish these medians from zero (p

= 0.4943 in Low and p ? 0.1896 in High). Overall, then, behavioral noise seems

rather inexpensive.

FINDING 5: Against the empirical distribution of play, the differences between actual

earnings and the earnings obtained from BNE strategies are economically and statistically insignificant.

The finding pertains to individual deviations to BNE. Would earnings significantly increase if all subjects coordinated on BNE strategies? In the Low treatment, such coordination would lead

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

VOL. 100 NO. 4 ANDERSON ET AL.: PREEMPTION GAMES 1795

Panel A. Low

12n

104

13 ct

d

Panel B. High

12t

104

5* 8H c CD

I 6i 44

24

0-S -30 -20 -10 0 10 20 30

BNE gains -30 -20 -10 0 10 20 30

BNE gains

Figure 7. Histograms of Gains from Unilaterally Deviating to BNE Strategies Calculated by Subject

the median subject's earnings to increase by 2.36 points, which is marginally different from zero

(p ?

0.0912). In the High treatment, by contrast, the median subject would earn 4.24 points less were they all to coordinate on the BNE (p

= 0.0006). Thus, BNE play would not substantially

improve average earnings in either treatment.

V. Discussion

Our exploration of preemption games uncovered several new regularities. On the theoretical

side, we were able to extend previous work to obtain precise predictions of behavior in compe tition. In Bayesian Nash Equilibrium (BNE), each investor waits until the value of the invest

ment opportunity hits a specific threshold that depends on that investor's private cost. Using both a recursion equation and an ordinary differential equation, we characterized the symmetric BNE threshold function for an arbitrary number of competitors and over relevant parameter ranges. Special cases of the parameters give threshold functions identical to known expressions in real options theory and in auction theory. These expressions in turn provide useful bounds and

approximations. For example, in triopoly with uniformly distributed costs, the BNE markup of threshold over cost decreases by about $1 for each $3 increase in cost.

The laboratory experiment confirmed most of the theoretical predictions. Observed invest ment thresholds indeed were much lower in triopoly competition than in monopoly. Changes in the parameters driving the stochastic value process had a strong effect (in the predicted direc

tion) in monopoly but (again as predicted) no detectable effect in competition. We also confirmed the sort of efficiency predicted in BNE: lower cost investors are far more likely to preempt than their higher cost rivals.

Other laboratory findings provide more qualified support for the BNE theory. The advantage to following the BNE strategy is much greater at lower costs, and behavior there cannot be dis

tinguished from BNE. At higher costs, however, expected earnings drop quickly and the BNE

prediction appears to break down. Thus the data seem consistent with BNE when it matters most, and the data deviate more when deviations make little difference. The upshot is that our subjects on average earn as much as in BNE.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

1796 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2010

Future laboratory work could investigate behavior for different market structures. We focused on triopoly (n

? 2) and suspect that higher values of n will push behavior to noisy approxima

tions of the Marshallian threshold. We skipped the duopoly case (n ?

1) to gain more separation, but future work could investigate it. Perhaps, as in many other laboratory duopolies,9 there will be frequent attempts to collude. Likewise, future work could investigate different parameter con

figurations. We chose certain parameter variations a priori, and the results generally conformed to theory. Conceivably the behavioral impact of other parameter variations could differ from the theoretical predictions in interesting ways.

Our experimental design did not support a serious investigation of learning behavior. Behavior

generally conformed to equilibrium theory even early on, but the long term trends are unclear. Future research could investigate learning in the preemption game in sessions with 50-100 peri ods. Such sessions could also manipulate feedback on other players' costs and markup choices in order to identify patterns of social learning. An expanded strategy method might also offer new insights into players' reasoning. After receiving her cost draw, each player could be offered a menu including the option to seize the opportunity manually (as in our experiment) or to pro gram a threshold agent (by filling in the threshold value) or to program any other sort of para metric agent that interests the researcher.

Our study examined winner-take-all preemption for several reasons: it brings the central stra

tegic issues into sharp focus, it connects nicely to real options theory and to auction theory, and in many important cases it is a reasonable simplification of reality. (To this list we add that win ner-take-all preemption permits clean laboratory implementation via the semi-strategy method; see Remark Al in the Online Appendix.) However, many other important cases involve partial preemption. For example, in larger communities a second or third big box store may be viable, and in some new research areas a second (or even a seventh) paper may be valuable. A clear next

step is to extend our work in this direction. To conclude, our work bridges the gap between two active fields that previously had little con

tact: real options and auctions. Perhaps our results will lead to more two-way traffic, empirical as well as theoretical.

Monopoly.?Let the gross value of investment V be governed by the stochastic differential

equation

where z is the standard Wiener process. A key insight from real options theory10 is that when future values are discounted at rate p > 0, the expected present value of waiting for appreciation R > 1 is R~?, where

Appendix

(16) dV= aVdt + aVdz,

(17)

9 For recent examples, see Martin Dufwenberg and Uri Gneezy (2000) and Steffen Huck, Hans-Theo Normann and

J?rg Oechssler (2004). 10 Samuel Karlin and Howard M. Taylor (1975, 357-364) use the Laplace transform and the Martingale theorem to

obtain the result. Appendix A of our working paper, Anderson, Friedman, and Oprea (2008), includes a streamlined

two-page derivation of the Monopoly optimum.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

VOL. 100 NO. 4 ANDERSON ETAL.: PREEMPTION GAMES 1797

In particular, if the current gross value is VQ and the investor will seize the opportunity when V hits threshold Vx > V0, then the discount factor is

(18) Ys.

When there are no rivals (n =

0), an investor with avoidable fixed cost C > 0 seeks to maximize the expected discounted profit E[(V

? C)e~pt). In view of (18), the problem reduces to finding a

threshold Vx to maximize [Vx ?

C}[V0/Vx]?. The associated first order condition is

(19)

with solution

(20)

0 = 1 v,

- c

V, = Vu = -( ? ?-l

?_

c.

If ? > 1, then (20) gives the optimal threshold we seek. If ?? [0,1], e.g., if a > p in (17), then the expected discounted profit increases in V] over the entire domain [C, oo), so there is no finite

optimal threshold. Note that (19-20) can also usefully be written

(2i) vyc) -

?[vM(c) -

c] = o.

The Preemption Game.?Let T[?,n,H] be the preemption game in which each player ("inves tor") i

= l,...,n+ 1 chooses a measurable threshold function Vtr. [CL,Q/] ??

[0,oc), draws real ized cost Ci independently from cumulative distribution H on [CL,Cu], and seizes the investment

opportunity at the first time tt that Vin (16) hits V^C,). In view of (18), player f s (expected) payoff is

?Pt[t,<tj Vy^i], (22) [VAQ -

c,] vxc.)

given initial value V0 in (16). We seek a symmetric Bayesian Nash equilibrium (SBNE) of

T[?9n,H]9 that is, a single threshold function V* such that if all other investors j ^ i choose

Vj =

V*(Cj) then investor / maximizes payoff (22) at any cost draw Cf G [CL,Q/] by choosing threshold V*(C,).

To characterize the SBNE, suppose that all rivals use the same increasing differentiable thresh old function V(Cy). Let V0

= V(t0) be the current gross investment value, let V

? maxse^0 t^V{s)

be the highest value yet observed, and let C

= V~l(V) be highest cost draw that would already have

led a rival to invest. Investor i can assume that all C; > C, since otherwise the game is already over and his choice is moot.

The unconditional probability that any one rival j has cost C, higher than C is 1 ?

H(C), and the probability conditional on C} > C is (1

? H(C))/(l

? H(C)). If investor / chooses

threshold Vx =

V(m), the probability that any particular rival will not preempt therefore is

(1 -

H(m))/(l -

H(C)l and

(23) Pt[ti<tj V;Vi] = 1

- H(m)

11 - H(C)

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

1798 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2010

It is convenient to combine this preemption probability with the discount factor and write

(24) 1 -

G(m) =

lV(m)i 1 - H(m)

L 1 - H(C)

Assume that the cost distribution H has density h with full support on the interval [CL, Cjj], where 0 < CL < Cu < oo. The definition (24) of G then ensures that G has a positive and continuous derivative g. Note also that 1

- G(Q/)

= 0 since 1 - H(CV) = 0.

Given investment cost Ch investor f s problem reduces to finding a threshold Vx =

V(m) that solves

(25) max [V(m) -

Ct] [1 -

G(m) me[C,Cu}

The FOC is

(26) V'(m)[\ -

G(m)] -

[V(m) -

Q]g(m) = 0.

In SBNE, investor i will find it advantageous to use the same threshold function V as the other investors. Accordingly, insert the "truthtelling" condition m

? Ct into (26) and simplify notation

by setting Ci = C, to obtain

(27) VXC)[l-G(C)]-[V(C)-C]g(C)=0.

Rearranging slightly, write

(28) -Cg(C) = V'(C)[l -

G(C)} -

V(C)g(C) = -fc [V(C)(1 -

G(C))}

and integrate both sides of (28) from C =

C, to Cw using 1 ?

G(CV) = 0, to obtain

(29) ?Cu

Je ^(y) ^ = 0 - V(C)[1

- G(C)].

Hence the SBNE threshold function V = V* must satisfy

(30) V\C) 1

1 - G(C) r ^-?G[y|y > C].

Since // and hence G are smooth strictly increasing functions, equation (30) ensures that V* is also smooth and increasing in C.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

VOL. 100 NO. 4 ANDERSON ETAL.: PREEMPTION GAMES 1799

Of course, the function G is itself defined in terms of V*.11 To obtain an explicit recursion

formula, first integrate by parts to get yg(y) dy =

Cv- CG(C) -

f?U G(y) dy. Then (30) reads

(31)

1 - G(C) C[l

- G(C)} + 3c \V(y) \-H{C)J

= C + \Cu\v\c) 1 - H(y) 1 - H(C)

dy.

The last expression shows that V*(C) is equal to C plus a positive markup, which shrinks to 0 as C approaches its upper endpoint C

= Q/.

Note that V0 and V =

V*(C) drop out of the formula (31) for V*, as well as from the bound

ary value problem (5-6). Therefore the opportunity to observe rivals' actions does not influence

thresholds; V*(C) can be set as soon as the investor draws cost C, and subsequent observations are irrelevant. The logic parallels the strategic isomorphism of Dutch and first price auctions.

To obtain the boundary value problem satisfied by V*, first insert (24) into (25) to obtain the

objective function (4) in the text. Use the product rule to take the derivative of equation (4) with

respect to m and cancel like terms (or, alternatively, take the derivative of InF) and evaluate at the "truthtelling" point m

= C{, to obtain the FOC:

(32) V'(Ci) nhiCt) [ViCd-Q] V\Q) [l-tf(Q)]

= 0.

Solve (32) for V*' to obtain equation (5). As noted in the text, the boundary condition (6) comes from the economics of the situation. At the highest possible cost realization, the existence of rivals known to have equal (or lower) cost induces Bertrand competition and drives the markup to zero.

The following lemma is proved in online Appendix A, using straightforward estimates.

LEMMA Al: Assume that ? > 1, that all rivals use a threshold function with inverse 7 such that 7' > 0, and that the hypotheses of Theorem 1 hold. Let the threshold value y maximize the

competitor's payoff given cost realization C G [CL,Q/]. Then y < V*M(C).

11 Here we depart from the auction literature, e.g., Vijay Krishna (2002, 14-19).

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

1800 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2010

PROOF OF THEOREM 1: The key step in part 1 is to show that the RHS of the ODE (5) is Lipschitz continuous in V*. A

potential problem is that the denominator factor 1 -

H(C) \ 0 as C / Cv. But given the bound

ary condition (6), the numerator factor V*(C) ?

C\0. Indeed,

(33) V*'{Cu) = Jim [V*(C)

- C)V*{C) nh{C)

c/cu\v*(C) -

?[V\C) -C])\\- H{C)i

nhjC^VjCu) V(C) - C

V\CV) -

?{V*(Cu) -

Cv] c/cu \-H{C)

nh(Cv) V*\CV) - 1

\-KCu)] 1-/3 Cu

= -niV'XCv)

- 1],

using L'H?pital's rule. Hence V*'(Cu) =

n/(n + 1), so we have Lipschitz continuity in an e neigh borhood of the upper boundary point.

A second potential problem is that the other denominator factor V*(C) ?

?[V*(C) ?

C] might be zero (or negative). For ? < 1 the factor obviously is positive. For ? > 1 we have

(34) V(C) -

?[V*{C) -C}=y-?[y-C]

= ?C-(?- l)y >?C-(?- \)VM(C)

VM{C) -

?[VM(C) -

C]

0,

since V*(C) = y < VM(C) by Lemma Al, and the last expression is 0 by (21).

Since the denominator is continuous on the closed interval [CL, Cu ?

e], it achieves a positive minimum value and thus is bounded away from zero. It follows that the RHS of the ODE (5) is positive, bounded, and Lipschitz continuous in V*. Consequently, the classic Picard-Lindelof theorem (e.g., see Chapter 8 of Morris W. Hirsch and Stephen Smale 1974) guarantees that a

solution V* to the boundary problem (5-6) exists and is unique. For the second part of the Theorem, suppose that V* satisfies the recursion equation

(31). Note that the integrand f(C,y) in the last term of (31) is the product of two increasing

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

VOL. 100 NO. 4 ANDERSON ET AL.: PREEMPTION GAMES 1801

functions of C, so df(C,y)/dC > 0. Indeed, using the product rule, one obtains df(C,y)/dC = ((?V'(C)/V(C)) + nh(C)/{\

- H{C))) f(C,y). Differentiating both sides of (31), we get

(35) V*'(C) = 1 - f(Q C) + \J dy

for all C G [CL, Q/). The last equality follows from subtracting C from both sides of (31). Divide

(35) through by the markup V*(C) ? C > 0 to obtain (32). As noted earlier, this can be rewrit

ten as the ODE (5). In the limit as C / Cv, the integral term vanishes in (31) and we obtain the

boundary condition (6). Conversely, suppose that V* solves the boundary value problem (5-6). Then the derivation

(27-31) ensures that it also satisfies the recursion formula (31). To establish part 3 of the Theorem, let W(C9 m) be the expected payoff to investor / when she

draws cost C and employs threshold V{ =

V*(m), assuming each other investor j sets threshold

V*(Cj). We complete the proof by showing that setting m = C, i.e., truthtelling, always maxi

mizes W(C,m). Assume that m > C; otherwise the investor already would have ended the game. Consequently

V*(m) > V > V0 and, by (25) and (30),

(36) W(C,m) = [V*(m) ~

C] [1 -

G(m)]

where the last expression uses integration by parts: J^u yg(y) dy ?

Cv ?

mG(m) ?

JCm" G(y) dy.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

1802 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2010

But (36) and the Mean Value Theorem (MVT) yield a point x such that

(37) W(C,C) -

W(Qm) = [m -

C]G(m) + ? G(y) dy

= [m- C]G(m) + [C -

m]G(x)

= [m- C][G(m)

- G{x)} > 0,

since G is increasing and the number x guaranteed by the MVT is between m and C. Hence m = C is indeed a best response, and V* is indeed a symmetric BNE.

PROOF OF THEOREM 2: The RHS of ODE (5) is positive and continuous, so its solution V* is increasing and continu

ously differentiable. The Marshallian lower bound follows from the fact that V*(C) = C ensures

zero profit, which dominates the negative realized profits from setting the threshold V*(C) < C. To see that VM is an upper bound for the SBNE threshold, note that the investor seeks thresh old Vx to maximize [Vx

- C] [V0/Vx]? [(1

- H(V*-\Vx)))/(l

- H(C))}n. The optimum threshold

value V, =

V*(C) must satisfy the FOC

n = __1_J_ _ nh[V*-\Vx)]V^XVx) K } Vi-C Vx i _ H[V*-\VX)} '

By (19), at Vx =

VM the RHS of (38) reduces to

(39) - W < 0. V ; 1 - H[V*-l{Vx)} Since the RHS of (19) is negative for Vx > VM(C\ no value of Vx > VM can satisfy the FOC (38). On the other hand, since the first RHS term in (38) goes to +oo as Vx \ C while the other terms remain bounded, the continuity of the RHS in Vx guarantees a solution to (38) at some value

Vx G (C, VM(C)), i.e., VM is indeed an upper bound. A similar argument shows that the auction bid function V is also an upper bound. Evaluating

the RHS of (38) at Vx ?

V(C), the first and third terms disappear due to the FOC that character izes V, and the remaining term, -?/Vx, is negative. Again, higher values of Vx only make the RHS of (38) more negative, but we know that the RHS is positive for sufficiently small values of Vx. Hence the intermediate value theorem again guarantees a solution to (38) at some value

vx e (c, v(O). Recall that (33) showed that V*'(Q/)

= n/(n + 1), independent of ?. Since V7*

= V in the spe cial case ?

= 0, we have V(CV) =

n/(n + 1) =

V*'(Cn). Of course, V*(Q,) =

Cu= V(Cu), so V and V* are indeed tangent at Cv.

A substantially expanded version of the preceding material is available as online Appendix A.

REFERENCES

Anderson, Steven T. 2003. "Strategie Timing of Investment: Extensions of Real Options Theory and Tim

ing Games, with Private Information, and an Empirical Application to Foreign Market Entry." Ph.D.

Diss. University of California-Santa Cruz.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions

VOL. 100 NO. 4 ANDERSON ETAL.: PREEMPTION GAMES 1803

Anderson, Steven T., Daniel Friedman, and Ryan D. Oprea. 2008. "Preemption Games: Theory and

Experiment." University of California-Santa Cruz Working Paper 637.

Bouis, Romain, Kuno J. M. Huisman, and Peter M. Kort. 2009. "Investment in Oligopoly under Uncer

tainty: The Accordion Effect." International Journal of Industrial Organization, 27(2): 320-31.

Boyer, Marcel, ?ric Gravel, and Pierre Lasserre. 2004. "Real Options and Strategic Competition: A Sur

vey." Paper presented at the 8th Annual International Real Options Conference, Montreal, Canada.

Christensen, Clayton M. 1997. The Innovator's Dilemma: When New Technologies Cause Great Firms to

Fail. Boston: Harvard Business School Press.

Dixit, Avinash K., and Robert S. Pindyck. 1994. Investment under Uncertainty. Princeton, NJ: Princeton

University Press.

Dufwenberg, Martin, and Uri Gneezy. 2000. "Price Competition and Market Concentration: An Experi mental Study." International Journal of Industrial Organization, 18(1): 7-22.

Fudenberg, Drew, and Jean Tirole. 1985. "Preemption and Rent Equilization in the Adoption of New

Technology." Review of Economic Studies, 52(3): 383-401.

Grenadier, Steven R. 2002. "Option Exercise Games: An Application to the Equilibrium Investment Strat

egies of Firms." Review of Financial Studies, 15(3): 691-721. Henry, Claude. 1974. "Investment Decisions under Uncertainty: The 'Irreversibility Effect.'" American

Economic Review, 64(6): 1006-12.

Hirsch, Morris W., and Stephen Smale. 1974. Differential Equations, Dynamical Systems, and Linear

Algebra. New York: Academic Press.

Hoppe, Heidrun C, and Ulrich Lehmann-Grube. 2005. "Innovation Timing Games: A General Frame

work with Applications." Journal of Economic Theory, 121(1): 30-50.

Huck, Steffen, Hans-Theo Normann, and Jorg Oechssler. 2004. "Two Are Few and Four Are Many: Number

Effects in Experimental Oligopolies." Journal of Economic Behavior and Organization, 53(4): 435-46.

Kaplan, E.L., and Paul Meier. 1958. "Nonparametric Estimation from Incomplete Observations." Journal

of the American Statistical Association, 53(282): 457-81.

Karlin, Samuel, and Howard M. Taylor 1975. A First Course in Stochastic Processes. 2nd ed. New York:

Academic Press.

Krishna, Vijay. 2002. Auction Theory. San Diego, CA: Elsevier Science, Academic Press.

Lambrecht, Bart, and William Perraudin. 2003. "Real Options and Preemption under Incomplete Infor

mation." Journal of Economic Dynamics and Control, 27(4): 619-43.

Levin, Dan, and James Peck. 2003. "To Grab for the Market or to Bide One's Time: A Dynamic Model of

Entry." RAND Journal of Economics, 34(3): 536-56. McDonald, Robert, and Daniel Siegel. 1986. "The Value of Waiting to Invest." Quarterly Journal of Eco

nomics, 101(4): 707-27. Mikesell, Raymond F., William H. Bartsch, Jack N. Behrman, Phillip E. Church, Gertrud G. Edwards,

Henry Gomez, William G. Harris, Markos Mamalakis, Donald A. Wells, Miguel S. Wionczek, and

James E. Zinser. 1971. Foreign Investment in the Petroleum and Mineral Industries: Case Studies of Investor-Host Country Relations. Baltimore, MD: Johns Hopkins University Press.

Myerson, Roger B. 1981. "Optimal Auction Design." Mathematics of Operation Research, 6(1): 58-73.

Oprea, Ryan, Daniel Friedman, and Steven T. Anderson. 2009. "Learning to Wait: A Laboratory Investi

gation." Review of Economic Studies, 76(3): 1103-24.

Reinganum, Jennifer F. 1981. "Dynamic Games of Innovation." Journal of Economic Theory, 25(1): 21-41.

Rudin, Walter 1966. Real and Complex Analysis. New York: McGraw-Hill.

Vickrey, William. 1961. "Counterspeculation, Auctions, and Competitive Sealed Tenders." Journal of Finance, 16(1): 8-37.

Weeds, Helen. 2002. "Strategic Delay in a Real Options Model of R&D Competition." Review of Economic Studies, 69(3): 729-47.

This content downloaded from 140.211.65.82 on Thu, 04 Feb 2016 18:36:18 UTC All use subject to JSTOR Terms and Conditions