American Economic Association

Do the Rich Get Richer and the Poor Poorer? Experimental Tests of a Model of Power
Author(s): Yvonne Durham, Jack Hirshleifer and Vernon L. Smith
Source: The American Economic Review, Vol. 88, No. 4 (Sep., 1998), pp. 970-983
Published by: American Economic Association
Stable URL: http://www.jstor.org/stable/117014 .
Accessed: 16/01/2015 03:34

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 support@jstor.org.

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

http://www.jstor.org

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions
 Do the Rich Get Richer and the Poor Poorer?
Experimental Tests of a Model of Power

By YVONNE DURHAM, JACK HIRSHLEIFER,AND VERNON L. SMITH*

In a market economy, there is no clear im- As for the Paradoxof Power, the theoretical
plication as to whether economic activities explanation is that initially weaker or poorer
will tend to reduce or else to widen initial contenders are typically motivated to fight
wealth disparities. When it comes to political harder, that is, to devote relatively more effort
or military struggles, in contrast, it might be to appropriative (conflictual) effort. Put an-
expected that initially stronger or richer con- other way, the marginal payoff of appropria-
tenders would grow ever stronger and richer tive effort relative to productive effort is
still. What has been termed the "Paradox of typically greater at low levels of income.
Power" (POP) (Hirshleifer, 1991) is the ob- (When agriculturalprices fell to extraordinar-
servation that very often the reverse occurs: ily low levels in the great depression of the
poorer or weaker contestantsimprove theirpo- 1980's, Kansas farmers were urged by their
sition relative to richer or strongeropponents. leaders to "raise less corn and raise more
In warfare, small nations have often defeated hell.") Looking at it the other way, while the
larger ones, as notably occurred in Vietnam. rich may have the capability of exploiting the
Or consider political clashes over income re- poor, it might not pay them to do so.
distribution. Although citizens in the upper Nevertheless, in some social contexts, ini-
half of the income spectrum surely have more tially richer and more powerful contestants do
political strength than those in the lower half, exploit weaker rivals. Affluent aristocracies
modern governments have systematically often use their power to extort even more re-
been transferring income from the former sources from the lower classes. So the question
(stronger) to the latter (weaker) group.' is, when does and when does not the Paradox
Individuals, groups, or nations-if rational of Power hold? In the model, the governing
and self-interested-will equalize the mar- factor is a parameterm reflecting the decisive-
ginal returns of two main ways of generating ness of conflictual effort. When decisiveness
income: (1) production and mutually advan- is low, the rich are content to concentrateupon
tageous exchange, versus (2) "appropriative" producing a larger social pie of income even
efforts designed to redistributeincome or cap- though the poor will be gaining an improved
ture resources previously controlled by other share thereof. But when conflictual prepon-
parties (or to defend against the latter's at- derance makes a sufficiently weighty differ-
tempts to do the same). Management and la- ence for achieved income-at the extreme,
borjointly generate the aggregateoutputof the when the battle is "winner take all' '-the rich
firm, for example, yet at the same time contend cannot afford to let the poor win the contest
with one another over distribution of the over distributive shares.
proceeds. The balance between production and strug-
gle, as two ways of making a living, has been
examined in a number of theoretical studies,
among them Trygve Haavelmo ( 1954), Goran
*
Durham: Department of Economics, University of
Skogh and Charles Stuart (1982), Hirshleifer
Arkansas, Fayetteville, AR 72703; Hirshleifer: Depart- (1989, 1991), Stergios Skaperdas(1992), and
ment of Economics, UCLA, Los Angeles, CA 90095; Herschel 1. Grossman and Minseong Kim
Smith, Economic Science Laboratory,University of Ari- (1994). But how decision makers choose be-
zona, Tucson, AZ 85721. We thank two anonymous ref- tween productive and conflictual activities has
erees of this Joumrnal
for extremely valuable comments and
suggestions. not heretofore, so far as we could determine,
' Edgar K. Browning and Jacquelene M. Browning been addressed experimentally. That was the
(1994 pp. 259-61). first object of the study reportedin this paper.
970

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions
VOL. 88 NO. 4 DURHAM ET AL.: EXPERIMENTALTESTS OF A MODEL OF POWER 971

A second aim was to consider the degree to and P2 (where PI + P2 = 1) are determinedin
which subjects ended up at the theoreticalnon- a conflictual process. In particular,the Contest
cooperative Nash solution, as opposed to more Success Function (CSF) takes the fighting ef-
cooperative outcomes generating larger in- forts Fi as inputs, yielding the distributive
come for the group as a whole. In the experi- shares as outputs:
mental literaturethe extent of cooperation has
been found to be sensitive to, among other (3) Pl F'1/(F7 + F2')
things, the number of iterations of the game
and whetherpartnersare held fixed or else var- P2- FI(F' + F').
ied from round to round. Our experimentalin-
vestigation was designed to address these Here m is a "decisiveness parameter" con-
questions as well. trolling the mapping of the input ratio F1IF2
Section I below outlines the analytic model. into the success ratiop P2. For m 1 the CSF
Section II explains our implementationof tests is characterized by diminishing marginal re-
of the model. Section III describes the exper- turns as F1 increases with given F2, or vice
imental procedures and outcomes. Section IV versa. However, for m > 1 there will be an
discusses the results and summarizes. initial range of increasing returns before di-
minishing marginal returns set in.3'4
I. The Model As a simplifying assumption, we postulate
that conflict is nondestructive, i.e., there is no
Each of two contenders i = 1, 2 must divide "battle damage." Choosing fighting activity
his/her exogenously given resource endow- over productive activity involves some oppor-
ment Ri between productive effort Ei and ap- tunity loss of potential output, but the struggle
propriative ("fighting") effort Fi: does not itself damage the resource base or
otherwise reduce the aggregate of income
(1) El + F= RI attainable.
Finally, the incomes accruing to the contes-
E2 + F2 R2- tants are:

The Ei efforts are inputs to a joint produc- (4) I =PII

tion function. A convenient form for this func-
tion, characterizedby constant returnsto scale I2 = P2I'
and constant elasticity of substitution,is:
For each level of fighting effort by con-
(2) I = A(Ells + El/s)s, tender 2, there is a corresponding optimal ef-
fort for contender 1 (and vice versa). Thus,
where A is an index of total productivity and l's optimization problem is to choose F1 2 0
s is an index of complementarity between El so as to solve:
and E2 2 However, for utmost simplicity here
we have assumed A = s = 1, so that (2) re- Max I, =pi(FlIF2) X I(EjIE2)
duces to the simple additive equation:
subject to El + F1 =R
(2a) I= El + E2.
and similarly for side 2.
Thus, the parties can cooperate by combin-
ing their productive efforts so as to generate a
common pool of income available to the two 3 To allow for possible differences in conflictual ability,
of them jointly. But the respective shares Pi the inputs F1 and F2 could be mu.ltipliedby "fighting ef-
ficiency coefficients" f, andf2 . This type of asymmetry is
also ruled out here.
4 Altemativepossible forms of what are called here Con-
2 To allow for differentialproductivity,El and E2could test Success Functions are discussed in Gordon Tullock
be multiplied by productive "efficiency coefficients" el (1980), Avinash Dixit (1987), Hirshleifer (1989), and
and e2?. We do not explore this kind of asymmetry here. Skaperdas(1996).

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions
972 THE AMERICANECONOMICREVIEW SEPTEMBER1998

Assuming neither party's resource con- equal) we have the "strong form" of the Para-
straint is binding, and using the simplified dox of Power. As already noted, for any inte-
production function (2a), the Nash-Cournot rior solution (that is, when the poorerside does
reaction functions are: not run into its resource constraint) we must
have F1 = F2, so that the strong form of the
(5) F1 m(E1 + E2) paradox necessarily applies.5 It can be shown
Fm Fn + Fin that there will be interior solutions up to some
critical value p of the resource ratio:
F2_ m(Ej + E2)
Fm Fm + F21 (7) p =(2 + m)l/M.

The right-hand sides being identical, F1 = F2 Thus specifically, in our experiments em-
is always a solution of these equations. That ploying the low value mi 1 for the decisive-
is, the reaction curves intersect along the 45- ness parameter,the predictionis thatthe strong
degree line between the F1 and F2 axes. In fact, form of the POP will hold for low resource
this is the sole intersection in the positive ratios, specifically for R1IR2 ? 3. For resource
quadrant. ratios larger than p = 3, only the weak form,
If however the boundary constraintis bind- i.e., the strict inequality on the right of equa-
ing for the poorer side (which we always take tion (6), is predicted.
to be contender 2), the second equation would (iii) Conflict as an Inequality-Aggravating
be replaced by: Process.-The model also indicates that for
sufficiently high values of the decisiveness co-
(5a) F2= R2 efficient mnand the resource ratio RI1R2, the
Paradox of Power will not apply. The rich
In that case, at equilibrium F1 and F2 are in would get richer and the poor poorer. Specif-
general unequal, but the intersection of the re- ically, for our experiments using the high de-
action functions still determines the Nash- cisiveness coefficient m= 4, the critical value
Cournot equilibrium values of the fighting X of the resource ratio for this condition is ap-
efforts. proximately 2.18.6 Also, from (7), when m =
As indicated above, the experiments were 4 the critical p separating the weak from the
intended in part to challenge a numberof spe- strong forms of the Paradox of Power equals
cific predictions derived from the model. In 1.5. Thus in our experiments using the low
particular: resource ratio 25/15 11.67 we expect the
(i) Fighting Intensities.-As the decisive- weak form of the Paradox of Power to hold,
ness parameter m exogenously increases, it since 1.67 lies between p and T. However, for
pays both sides to "fight harder," i.e., the the experiments with RIIR2 32/8 = 4 >
equilibrium fighting efforts Fi will rise. (Im- r = 2.18, the prediction is that the initially
plying, of course, that the ultimate achieved better-endowed party will improve its relative
incomes Ii must fall.) position compared to the less well-endowed
(ii) Conflict as an Equalizing Process side:
(Paradox of Power), Strong vs. Weak
Form. -For sufficiently low values of the (8) 1I112 (F1/F2)"' > RI2R20
decisiveness parameter m, disparities in
achieved income will-owing to the Paradox
of Power-be smaller than the initial dispar-
ities in resource endowments. Letting con-
tender 1 be the initially better-endowed side: T
This result can come about only using the simplified
production function (2a), where the productive comple-
(6) RI/R2 > 1I/12 ! 1. mentaritycoefficient is set at s = 1. For the more general
CES production function (2), with s > 1, at equilibrium
only the weak form of the paradox holds.
When the equality on the right holds (i.e., 6 The value of T was obtained by finding the resource
when the achieved incomes of the initially ratio at which the condition I, II = RI/R2 was met for
richer and initially poorer sides end up exactly m = 4.

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions
VOL. 88 NO. 4 DURHAM ET AL.: EXPERIMENTALTESTS OF A MODEL OF POWER 973

II. ImplementingTestsof the Model pate that the condition of fixed partners will
favor somewhat more cooperative behavior.
Certain game-theoretic and implementa- However, we will be implementing a normal-
tional concerns are also addressed in our ex- form game (simultaneous choice of strategies
perimentaltest of the above model. In the strict presentedin a payoff matrixeach round) rather
game-theoretic sense, the noncooperative than an extensive-form game (sequential
equilibrium is about strangerswho meet once, choice by the playerseach round). McCabe and
interact strategically in their self-interest, and Smith (1997) show that the extensive form fa-
will never meet again. Such conditions con- vors cooperationrelative to the normal form of
trol for repeated-gameeffects, since the antag- two theoretically "equivalent" games. This is
onists have no history or future. Yet in many because cooperativeintentionscan be signalled
contexts individuals interact in repeated by one player and the second player can recip-
games, where they can signal, punish, and rocate (not defect) within the same round.
build reputations. In the particularly simple Hence, the normal form of the experimentsre-
version where the one-shot game is iterated ported below is expected to make cooperation
with the same payoffs each round, we have a (i.e., reduced levels of fighting) difficult even
supergame. The study of such games has been in repeat interactions.
motivated by the intuition or "folk theorem"
that repetition makes cooperation possible III. ExperimentalProceduresand Outcomes
(Jean-Fran9ois Mertens, 1984). But formal
theorems to this effect for finite horizons have A. ExperimentalDesign
not been forthcoming, and interest has settled
on experimental studies of both single-play We conducted 24 experirnentsusing a total
games and supergames, and on variations in of 278 subjects. No subject participated in
the protocol for matching players in repeat more than one experiment.There were six bar-
play. gaining pairs in each experiment, except for a
Kevin A. McCabe et al. (1996) studied a few cases with only four or five pairs. Each
class of extensive-form games in which the experiment involved repeated play, the pay-
partiesmove sequentially in a series of rounds. offs being constant in each round. Within each
In any round the first-mover can forward sig- round, each subject pair chose simultaneously
nal the desire to cooperate, but the otherplayer a (row, column) in a matrix displaying the
can defect. In one game the first player can payoff's of each. Subjects were not informed
punish such defections. In the other he/she has how many rounds would take place; in fact, in
no such recourse. If pairs are matched at ran- each experiment there were 16 or 17 rounds
dom for each play, in a repeated sequence of before termination.Subjects were recruitedfor
unknown length, subjects gradually learn to two-hour sessions but the experiments took
cooperate when the punishment option is much less time, making credible the condition
available; when this option is not available of an unknown horizon.7
they tend to play noncooperatively. If instead In every round each subject allocated his/
the same pairs remain matched for the entire her initial endowment of tokens between an
length of the supergame, they tend to achieve "Investment Account" (IA) and a "Ra-
cooperation whether or not the opportunityfor tioning Account" (RA). (We deliberately
direct punishment or defection is available. avoided using any terminology suggestive of
Consequently,in additionto testing the sub- "fighting.")8 Tokens contributed to the IA
stantivepredictionsassociatedwith the Paradox corresponded to productive effort E; in the
of Power, we will be addressingsome of these theoretical model: the paired IA contributions
issues that have arisen in the experimentaland generated an aggregate pool of income (in the
game-theoretictraditions.Specifically, we will
be comparing the results of experiments in
which the partnersare randomlyvaried in each
'In McCabe et al. ( 1996) this technique was found to
round with experiments in which the partners be effective in leading to cooperation, even on the "last"
are fixed throughoutthe supergame. As sug- repetition.
gested by the preceding discussion, we antici- 8 Copies of the instructions are available upon request.

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions
974 THE AMERICANECONOMICREVIEW SEPTEMBER1998

form of "experimentalpesos") in accordance noncooperative equilibrium.However, the ex-

with equation (2a) above. Funds put into the perimental literature has intensively investi-
RA correspondedto fighting effort Fi and de- gated conditions under which subjects might
termined the respective distributive shares Pi arrive at a more cooperative outcome. This is
and P2 in accordance with equations (3). For the first issue addressed:
simplicity, only integer choices were permit-
ted. (More precisely, each subject could allo-
H,: the null hypothesis is that (Fl, F2)
cate, within his/her resource constraint,
amounts in integral hundreds of tokens to in-
vest in the IA-the remainder,of course, go- - (C1, C2)
ing into the RA.) The totals of pesos ultimately
achieved were converted into actual dollars at Ha: the alternative hypothesis is that (Fl, F2)
the end of the experiment, so subjects had a
substantialmotivation to make well-calculated (N1,N2).
choices. (The payoffs ranged from $0.25 to
$75.25, not including the $5 fee for showing Here the Ni signify the respective fighting ef-
up. The average payoff was $17.66.) forts Fi under the Nash solution, while the Ci
To challenge the implications of the model, are the fighting efforts under the Cooperative
we manipulated the resource endowments RI solution.
and R2 and also the decisiveness coefficient m. The theoretical Nash solution is generated
Four experiments were run with each of the by the intersection of the reaction functions of
three endowment vectors (RI, R2)= (20, 20), equations (5) above for an interioroutcomie-
(25, 15), and (32, 8)-the first series using or in the case of a boundary outcome (where
a low value m = 1 of the decisiveness param- the poorer contender 2's resource constraintis
eter, and the next using a high value m = 4. binding), substituting (5a) for the second
Thus there were 24 experiments in all. equation in (5). The Cooperative solutioniwas
Also, in view of the McCabe et al. (1996) defined as (Cl, C2) = ( 1 I ). That is, each side
result thatcooperationis promotedby repeated devotes the minimal allowable positive
play with the same partners,each group of four amount to fighting effort.9 Table 2 is a sum-
experiments was furthersubdivided into alter- mary that allows a comparison between the
native matching protocols. In the first ("vary- Nash equilibrium and the average of the six
ing partners" ) protocol, partners were (or, in a few cases, four or five) observations
randomly changed each round. Under the sec- on round 16 in each of the 24 experiments.
ond ("fixed partners") protocol, subjects Note that almost all the contesting pairs'
were randomly paired at the beginning of the choices are much closer to the Nash prediction
experiment but played repeatedly with the than the Cooperative ( 1, 1)9 but are biased on
same partnerthroughout. the low side of Nash.
Overall there were eight experiments under As an illustration,Figure 1 depicts the first
each of the three endowment conditions. Four two of the 24 experiments.Each * symbol plots
of the eight involved varying partners, and the fighting efforts F1 and F2 chosen by one of
four fixed partners. There was an analogous the 12 bargainingpairsin the sixteenth(the last
subdivision between experiments conducted or next-to-last) round. Also shown are the
using m =1 and using m = 4. The upshot is
that there were exactly two experiments for
each of the 12 sets of experimentalconditions
or "treatments." The treatmentdesign is sum-
marized in Table 1.
I equations (3), the relative shares p, and P2 are in-
In
B. Results-Nash versus determinate when F, = F2 = 0. To remedy the indeter-
Cooperative Solutions minacy, the Profit Table in the Instructions provided for
zero payoff to a player whenever he or she put zero into
the RA, i.e., whenever Fi = 0 was chosen. So the coop-
The theoretical model described in the erative combination maximizing the aggregate payoff, un-
previous section derived the Nash-Cournot der the integer constraint, is (Cl, C2) = ( 1, 1 ).

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions
VOL. 88 NO. 4 DURHAM ET AL.: EXPERIMENTALTESTS OF A MODEL OF POWER 975

TABLE 1-TREATMENTS

Number of experiments (numberof subjects)*

Endowments
(RI, R2) Decisiveness (m) Variable pairing Fixed pairing Totals
20, 20 1 2 (24) 2 (24) 4 (48)
25, 15 1 2 (24) 2 (24) 4 (48)
32,8 1 2 (22) 2 (24) 4 (46)
20, 20 4 2 (22) 2 (22) 4 (44)
25, 15 4 2 (20) 2 (24) 4 (44)
32,8 4 2(24) 2 (24) 4(48)
TOTALS 12 (136) 12 (142) 24 (278)

Notes: * Owing to some recruiting problems, a few experiments were run using only eight or ten subjects (four or five
pairs). Each experiment was run for either 16 or 17 rounds.

reaction (step) functions, the computed Nash

noncooperativeequilibriumat the intersectionof
these functions,and the postulatedCooperative
solution at ( 1, 1). At a glance, the observations
tend to fall between the Nash and the Coopera- exp[- 2S2 (Fi - N)2
tive solutions, but much nearer to the fonner.
This was in fact typical; errordeviations from
Nash tended to the Cooperativeside. = exP{2S2 [2tj (-N)
We used the likelihood ratio to test the
alternative Nash hypothesis Ha against the
cooperative null hypothesis H0. 0 It was as-
sumed that the observations for the "fighting + Tn(Nf2-C2)]} .
efforts" Fi are normally distributed, with
mean ,a = C under Ho or mean ,a = N under
Ha, and variance V = S2 (the sample vari- (9b) in \ = TnF( -
ance)." Then for any given treatment the
likelihood ratio is: + (N -2 _)(Ih2S2).

Here the t subscriptindexes the rounds from 1

to T, while the i subscriptindexes the individ-
' Computing the likelihood ratio allows the analysis to
include Bayesian updating of prior beliefs as well as tra- ual pair observations from 1 to n.
ditional significance tests. We consider both avenues, with A X < 1 would indicate that, for this partic-
similar results. The likelihood ratio is particularlyconve- ular treatment, the observed choices had a
nient for Bayesian conversion of prior beliefs p' into pos- higher probability of occurring under the al-
terior beliefs p" in light of the expenmental evidence. The
relevant version of Bayes' Theorem is:
ternative (Nash) hypothesis than under the
null (Cooperative) hypothesis. The 12 rows of
pO Likelihood of evidence underH,, p, Table 3 list the A's for all the treatments,
p't Likelihood of evidence underH,, p,

As a technical qualification, a strict Bayesian would eter." However, we have taken the liberty of simply em-
want to deal with the fact that the true normal variance V ploying the observed sample variance S2 for V. Doing so
is unknown. In principle one ought to specify priorbeliefs provides an enormous computational saving without sub-
about the variance and deal with it as a "nuisance param- stantially affecting the results.

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions
976 THE AMERICANECONOMICREVIEW SEPTEMBER1998

TABLE 2-EXPERIMENTAL RESULTS

Average results (sixteenth

Treatmentparameters Nash solution observations)
Experiment
no. m Pairing RI, R2 R1/R2 N1, N2 NIIN2 F,, F2 F1/F2 I/12
1 1 v 20, 20 1 10, 10 1 7.83, 6.83 1.15 1.15
2 1 V 20,20 1 10, 10 1 8,9 0.89 0.89
3 1 F 20, 20 1 10, 10 1 8.67, 6.67 1.30 1.30
4 1 F 20, 20 1 10, 10 1 4,5 0.8 0.8
5 1 V 25, 15 1.67 10, 10 1 10.83, 8.5 1.27 1.27
6 1 v 25, 15 1.67 10, 10 1 9,9.17 0.98 0.98
7 1 F 25, 15 1.67 10, 10 1 10.17, 9 1.13 1.13
8 1 F 25, 15 1.67 10, 10 1 7.67, 6.83 1.12 1.12
9 1 V 32, 8 4 10, 8 1.25 11.83, 7.67 1.54 1.54
10 1 V 32, 8 4 10, 8 1.25 10.33, 7.5 1.38 1.38
11 1 F 32, 8 4 10, 8 1.25 5.17, 3.17 1.63 1.63
12 1 F 32, 8 4 10, 8 1.25 5.4,4.6 1.17 1.17

13 4 v 20, 20 1 16, 16 1 10.33, 12.83 0.81 0.42

14 4 V 20, 20 1 16, 16 1 14.67, 15.33 0.96 0.84
15 4 F 20, 20 1 16, 16 1 11.67, 13.83 0.84 0.51
16 4 F 20, 20 1 16, 16 1 10, 9.2 1.09 1.40
17 4 V 25, 15 1.67 16, 15 1.07 15.5, 11.83 1.31 2.95
18 4 V 25, 15 1.67 16, 15 1.07 16.5, 12.5 1.32 3.04
19 4 F 25, 15 1.67 16, 15 1.07 16.17, 13.83 1.17 1.87
20 4 F 25, 15 1.67 16, 15 1.07 13.5, 12.33 1.10 1.44
21 4 V 32, 8 4 12, 8 1.5 11.67, 7.33 1.59 6.42
22 4 V 32, 8 4 12, 8 1.5 1-1.67,7 1.67 7.72
23 4 F 32, 8 4 12, 8 1.5 10.5,7.5 1.4 3.84
24 4 F 32, 8 4 12, 8 1.5 11, 4.67 2.36 30.76

expressed for convenience in terns of logs Round" columns, equations (9a) and (9b) are
(the log-likelihood ratios) as in (9b). Positive modified by simply dropping the indexing
entries in the table represent results favoring over t.] From the statistical point of view, the
the null hypothesis while negative entries fa- "All Rounds" reportsprovide a larger sample
vor the altemative hypothesis. size (though not independent) and thus are
The columns towardthe left of Table 3 iden- less influenced by random fluctuations.On the
tify the conditions for each of the 12 treat- other hand, the "Sixteenth Round" reportsare
ments. The remaining columns show the more likely to isolate the mature behavior of
results for "All Rounds" and also for the the experimental subjects. Finally, F1 refers to
"Sixteenth Round" (that is, the last or next- the subject having the larger, and F2 the
to-last round) separately. [For the "Sixteenth smaller, resource endowment. (In the equal-

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions
VOL 88 NO. 4 DURHAM ET AL.: EXPERIMENTALTESTS OF A MODEL OF POWER 977

F2
20 _ _

15 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -- - - - - - - - - - - - - -

Nash: 1
1 0 - -- ----
- --- -- - - - -- -- -- -- - --

Cooperative: 0

00

0 5 10 15 20
Fl
FIGURE 1. EXPERIMENTS Al AND BI
(RI, R2) - (20, 20); m = 1; VARYING PARTNERS

endowment cases, the assignment of F, versus likelihood ratios for the "Sixteenth Round"
F2 was random.) columns are somewhat less negative than in
The results summarizedin Tables 2 and 3 the corresponding "All Rounds" columns.
overwhelminglysupportthe Nash as opposedto This is in partthe consequence of smaller sam-
the Cooperativesolution.In Table 3 the predom- ple size, but that is evidently not the entire
inantly negative values of the log-likelihood story- since the only two instances of
ratios(46 of the 48 tabulatedentries)correspond positive values both fall under the "Sixteenth
of course to likelihood ratios less than unity in Round" headings. So, for these cases there is
equation (9a) above.'2Using a likelihood ratio the suggestion thatparticipantswere "learning
test, 45 of the 48 entries in Table 3 unambigu- to cooperate" by the sixteenth round of inter-
ously-at significancelevel a = 0.005-imply action. Second, again for all 24 comparisons,
rejectingthe Cooperativehypothesisin favor of the results under the "fixed partners" (F)
Nash. Only 1 of the 48 (the single entrymarked condition are noticeably less negative thanthe
*) unambiguouslydoes the reverse. For the re- corresponding "varying partners" (V) results.
mainingtwo entriesmarked?, significancetest- Since the "fixed partners" condition facili-
ing using a = 0.001 indicates that whichever tates the development of mutualunderstanding
hypothesis is taken to be the null is rejectedin ing, we examine the dynamrrics of their inter-
favor of the other!3 action below for evidence of best-response
Apart from the generally negative signs for moves.
the log X's, two features of Table 3 stand out. It is particularlysignificatntthat the only ex-
First, in all 24 possible comparisons the log- ceptions to the observed tendency to converge
to near the Nash equilibrium occur under the
treatmentin which the Nash equilibrium lies
'2 Under a Bayesian interpretation,any observer, re-
at the boundaryof the constraintset for one of
gardless of prior beliefs, should revise those beliefs so as the bargainers.As shown in Smith and James
to attach greaterconfidence to the Nash hypothesis. M. Walker (1993), in such cases any devia-
i3 In Bayesian terms, for these two cases the likeli- tion or slippage from the predicted outcome is
hoods are about equal under the null and alternative necessarily biased, and changes in variance
hypotheses, so no great revision of prior beliefs is in-
dicated. The evidence, while improbable either way, is will change the mean. We should also note
not much more improbable under one hypothesis than that, with fixed partners,if bargainersdeviate
under the other. from Nash, either to signal cooperation or to

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions
978 THE AMERICANECONOMICREVIEW SEPTEMBER1998

RATIOS
TABLE3-LOG-LIKELIHOOD

Nash versus Cooperative Solution"

All Rounds
(T n = 198) Sixteenth Round (n = 12)

Treatmentparametersand matching protocol Types Types

RI R2 m Pairing F, F2 F, F2

20 20 1 V -873 -654 -87 -30

20 20 1 F -545 -169 -3.8 -0.7?
25 15 1 V -939 -1302 -31 -133
25 15 1 F -262 -218 -20 -23
32 8 1 V -879 -2583 -31 -111
32 8 1 F -123 -73 0.8? 8.5*
20 20 4 V -1260 -1376 -114 -78
20 20 4 F -68 -92 -6.3 -10
25 15 4 V -1475 -541 -1001 -115
25 15 4 F -860 -796 -48 -69
32 8 4 V -2756 - 1643 -1008 -114
32 8 4 F -373 -332 -43 -23

a B. W. Lindgren (1962), for example, derives a most powerful test (among the class of tests where a errors are not

smaller, none has a larger power 1 - f) for a simple H, against a simple H, using the likelihood ratio. The best critical
region is X = N,,(x)/N,,(x)< K, where N(x) is the normal density evaluated for HOor He,,and K is a constant chosen to
set the Type I error (a) at the desired level. Ho is then rejected in favor of H, if ) < K.
Setting a = 0.001, KE,for each experimentaltreatment,E, was computed from the following [for the Sixteenth Round
case, T = 1 in (9)]:

PROB >
CI W,(FE KE)= 1 - N( j 0.001,
SF/;n

where FEis the sample mean level of observed fighting, and SF is the varianceacross all n pairs in treatmentE. The results
from this likelihood ratio test allow us to reject the hypothesis ,u = C = 1 (cooperation)in all cases except for the * entry.
The entriesmarked? indicate cases where cooperationis rejectedin favor of Nash, but when Ho and Ha are interchangedso
thatcooperationbecomes the null hypothesis,Nash is rejectedin favor of cooperation.This illustratesthe inherentambiguity
of "classical tests": the outcome need not be independentof which hypothesis is chosen as the null!

punish failures to reciprocate, a bargainer direction of cooperation. A negative number

whose Nash outcome is on a boundarycan sig- indicates slippage in the direction of conflict
nal cooperation without constraint,but punish- beyond that called for by the Nash solution.
ment is asymmetrically restricted. As expected, the positive numbers far out-
We can quantify the average percent devi- weigh the negative numbers, and the positive
ations in the direction of cooperation by defin- numbers predominate more under the "fixed
ing "slippage fractions" S, and S2: partners" (F) condition. Finally, there is a
noticeable positive correlation between the
SI and S2 numbers on each row of the table:
(10) S _ N; F, when one subject behaves cooperatively,
N- C, his/her partner is likely to do so as well.
Once again, as expected, this positive cor-
In Table 4, a positive number in the two relation holds particularly for the "fixed
right-hand columns indicates slippage in the partners" condition. And in addition, it

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions
VOL. 88 NO. 4 DURHAM ET AL: EXPERIMENTALTESTSOF A MODEL OF POWER 979

TABLE 4-SLIPPAGE TOWARD COOPERATION

Average
Treatmentparameters Nash solution Average results slippage*
Experiment no. m Pairing RI, R2 R1/R2 N,, N2 F, F2 SI S2
1 1 V 20, 20 1 10,10 7.83 6.83 0.24 0.35
2 1 V 20,20 1 10,10 8 9 0.22 0.11
3 1 F 20, 20 1 10, 10 8.67 6.67 0.15 0.37
4 1 F 20,20 1 10,10 4 5 0.67 0.56
5 1 V 25, 15 1.67 10, 10 10.83 8.5 -0.09 0.17
6 1 V 25, 15 1.67 10, 10 9 9.17 0.11 0.09
7 1 F 25, 15 1.67 10,10 10.17 9 -0.02 0.11
8 1 F 25, 15 1.67 10, 10 7.67 6.83 0.26 0.41
9 1 V 32,8 4 10,8 11.83 7.67 -0.20 0.05
10 1 V 32, 8 4 10,8 10.33 7.5 -0.04 0.07
11 1 F 32,8 4 10,8 5.17 3.17 0.54 0.69
12 1 F 32, 8 4 10,8 5.4 4.6 0.51 0.46

13 4 V 20, 20 1 16,16 10.33 12.83 0.38 0.21

14 4 V 20, 20 1 16,16 14.67 15.33 0.09 0.04
15 4 F 20, 20 1 16, 16 11.67 13.83 0.29 0.15
16 4 F 20, 20 1 16, 16 10 9.2 0.40 0.45
17 4 V 25, 15 1.67 16, 15 15.5 11.83 0.33 0.23
18 4 V 25, 15 1.67 16, 15 16.5 12.5 -0.03 0.18
19 4 F 25, 15 1.67 16, 15 16.17 13.83 -0.01 0.08
20 4 F 25, 15 1.67 16, 15 13.5 12.33 0.17 0.19
21 4 V 32,8 4 12,8 11.67 7.33 0.03 0.10
22 4 V 32,8 4 12,8 11.67 7 0.03 0.40
23 4 F 32,8 4 12, 8 10.5 7.5 0.14 0.07
24 4 F 32,8 4 12,8 11 4.67 0.09 0.48

Notes: * Si = (Ni - Fi)(Ni - C).

holds noticeably more strongly for the cases chooses a profit-maximizingstrategybased on

with equal resource endowments (RI, R2) = the other player's choice in period t - 1 (a
(20, 20). shortsightedbest-reply strategy) then they will
converge to Nash. In fact subjects' choices
C. Dynamics of Interactions may have some inertia, and may involve co-
operative signalling. One way of modelling
for Fixed Partners
these dynamic interactions, and obtaining a
measure of the propensity to choose best re-
A fuller treatment of how individual pairs plies, is to estimate the following equation for
interact requires analysis of their interactive Fit (the fighting effort chosen by subject i in
choices over time. If each player in period t period t):

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions
980 THE AMERICANECONOMICREVIEW SEPTEMBER1998

# of Subjects
25rX
20
20

15 K?,%% - 15
15 10| 14 ..
18~~~~~~~~~~~

5144- --
15 T 13 j01?1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 >1
Delta
FIGURE 2. FREQUENCIES OF ESTIMATED DELTAS
N 140

Fit I- F *. For the fixed-pairs condition the

data substantially support the adaptive best-
=FIt + ( 1-6i )(Fit- ,- F*) + eit, response Cournot dynamic.

D. Results-The Paradox of Power

where F * is the best reply in period t to i's
choice of strategy in period t - 1. In this my- The experiments tested a numberof specific
opic Cournotdynamic, the deterministiccom- predictions of the analytic model.
ponent of Fitis distributedbetween an adaptive
weight element 6 related to i's currentbest re- Prediction 1: Higher values of the decisive-
ply F* (to the opponent's previous-round ness parameterm will lead to larger fighting
choice) and an inertial element 1 - 6 related efforts on both sides.
to i's previous choice Fit -. Or, in the second
form of (11 ) the choice of Fit can be inter- So the fighting efforts F, and F2 should both
preted as a best reply, plus an imperfect adap- be greater at the higher decisiveness level
tive adjustmentbased on the error difference m = 4 than at m - 1. The upper half of Table
between last period's choice Fit - I and this pe- 2 shows the results for m 1, and the lower
riod's best reply. Figure 2 provides a histo- half for m = 4. There are 48 comparisons, of
gram of the frequency distribution of 140 which a remarkable 45 are in the direction
individual estimates of 6i over all decision tri- predicted.
als for each i. Overwhelmingly, the individual
8i (and therefore the 1 - 6i) values are in the Prediction 2a: At the low value m 1 for
unit interval indicating some weight being the decisiveness parameter, the initially
given to i's previous choice and some to i's poorer side will always end up improving its
best reply. They are also overwhelmingly sig- position.
nificantly different from zero. Note that over
half the subjects exhibit values of (1 - 6i) of At m = 1, the attained income ratio I1/12
at least 0.5, indicating considerable weight (which for m = 1 simply equals the ratio of
being attachedto correctingthe errordeviation fighting efforts F1IF2) should exceed the re-

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions
VOL. 88 NO. 4 DURHAM ET AL.: EXPERIMENTALTESTSOF A MODEL OF POWER 981

source ratio RI/R2. The requirement of un- IV. Discussionand Summary

equal initial endowments limits the relevant
data to rows 5 through 12 of Table 2. Here all This experimental investigation deals with
eight of the eight comparisons showed the pre- a mixed-incentive, iterated-play, bilateral in-
dicted relative improvement-that is, I, /I2 < teraction. In each of some 16 rounds, paired
RI/R2-and almost always by quite a wide individuals had to strike a balance between
margin. productionand appropriation:more explicitly,
between investing resources in joint produc-
Prediction 2b: For m = 1 the poorer side tion versus engaging in a distributive struggle
should attain approximateequality of income over the respective shares.
(strong form of the POP) for initial resource We tested two main kinds of predictions:
ratios RI/R2 < 3, but only some relative im- ( 1 ) The first group dealt 'withissues com-
provement- 1 < I,/12 < RI/R2-for larger mon to much of the game-theoretic and ex-
resource ratios (weak form of the POP). perimental literature. Of these, the major
question was the degree to which the experi-
Looking once again only at the unequal en- mental outcomes approximatedthe noncoop-
dowments cases, rows 5 through 12 of Table erative Nash solution, as opposed to a more
2, the averageof the tabulatedresultsis 1,/I2 = cooperative outcome generating a larger in-
1.125, on the high side of the predictedI,/12 = come for the group as a whole. We also
1. By way of comparison, for rows 9 through compared protocols with randomly varying
12 where only the weak form I, /I2> 1 is pre- partnerseach round as opposed to fixed part-
dicted, the average outcome is I,/I2 = 1.43. ners over the entire sequence of play.
So, at least relatively, the predicted compari- (2) The second group of predictions dealt
son of the strong-form versus weak-form pre- with inferences from the specific model of
dictions is supported. conflict in Hirshleifer ( 1991 ), and specifically
those associated with the Paradox of Power.
Prediction 2c: At the high value m = 4 for the The paradox is that, in many situations, an in-
decisiveness coefficient, the Paradoxof Power itially poorer side will end up gaining in rel-
should continue to hold (in its weak form) for ative position in comparison with an initially
p < R1/R2<Kr, where p = 1.5 and r = 2.18. richer and thus strongeropponent.
But for higher resource ratios the richer side With regardto the first group of predictions,
should end up actually improving on its rela- the experimental observations overwhelm-
tive position. That is, in this range I,12
I= (Fl / ingly supported the Nash as opposed to the
F2)4 shouldexceedRI/R2. Cooperative solution. tiowever, while the
Nash solution is much better supported in a
For the unequal-endowments rows 17 dichotomous comparisonbetween the two, the
through20 of Table 2, the resource ratio is RI/ experimental results typically displayed some
iY - 25/15 = 1.67, lying between p and T. degree of slippage in the direction of cooper-
So the Paradox of Power is predicted in these ation. The convergence toward Nash was
cases. However, for rows 21 through 24 the weaker under the fixed partners as opposed
resource ratio is R IR2 = 32/8 = 4 > to the varying partners protocol, and also
2.18 = r, so we expect the rich to become was weaker in the mature (sixteenth round)
richer still. choices than the overall behavior. Together
Taking up the latter group first, three of the with an observed tendency toward positive
four cases supportthe prediction I,/I2 (Fl / correlation of the deviations from the Nash
F2)4 > 4. In fact, the average of the observed equilibrium,these results are consistent with a
results was a much higher I, /I2= 12.19. Turn- "leaming to cooperate" interpretation.Fixed
ing to the first group, however, all four cases partnersover multiple rounds of interactionfa-
violate the prediction! Quantitatively,the pre- vor the development of mutual understanding
dicted Nash outcome (N1, N2) = (16, 15) im- relative to varying partners. Still, we must
plies I, /I2 = (16/15 )4 = 1.29 < 1.67, while reemphasize, overall the results were domi-
the averageof the observed results was I, I2 - nated by noncooperative (Nash) behavior. A
2.32 > 1.67. dynamic analysis of fixed partner interaction

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions
982 THE AMERICANECONOMICREVIEW SEPTEMBER1998

predominantly supported a Cournot myopic predictedF2= 15. However, there is a bound-

adaptive best-reply strategy in which subjects' ary problem here: the Nash prediction for the
choices were best replies to their opponent's poorer side would require them to devote 100
previous choice but with a positive correction percent of their resources (R2 = 15) to fight-
for error in anticipating that previous best re- ing. Thus, any error whatsoever on their part
ply. This dynamic helps to explain the con- must necessarily lead to a deficiency of fight-
vergence tendencies to Nash. ing effort, which at least partially rationalizes
With regard to the underlying conflict the "anomalous" result found.
model, the central prediction (Prediction 1) To sum up: in this experimentalcontext our
was that larger fighting efforts would be ob- results supportthe Nash as opposed to the Co-
served for higher values of the "decisiveness operativesolution, though with some degree of
coefficient" m-a parameterthatindicates the slippage in the directionof the latter. And the
degree to which the fighting efforts as inputs theoreticalpredictionsas to when the Paradox
determine the relative shares of incomes at- of Power-that an initially weaker party will
tained. Prediction 1 was overwhelmingly con- improve its position relative to a strongerop-
firrmed:in 45 of 48 comparisons, when fighting ponent-will or will not be observed, are also
became a more decisive determinantof rela- broadlysupported.
tive income shares, both sides invested more
in the struggle.
REFERENCES
The evidence was more mixed regarding
when the Paradox of Power-that the poorer
side would improve its relative position- Browning, Edgar K. and Browning, Jacquelene
would hold. For the experiments employing a M. Public finance and the price system,
low value m = 1 for the decisiveness coeffi- 4th Ed. Englewood Cliffs, NJ: Prentice
cient, Prediction2a was that at least the "weak Hall, 1994.
form" of the paradox should always hold, that Dixit, Avinash. "Strategic Behavior in Con-
is, that I,/12 > R1 R2. In fact eight of eight tests." American Economic Review, De-
possible comparisons confirmedPrediction2a. cember 1987, 77(5), pp. 891-98.
Prediction 2b was more stringent, specifying Grossman, Herschel I. and Kim, Minseong.
that for the four cases where the initial re- "'Swords or Plowshares? A Theory of the
source ratio was sufficiently low the "strong Security of Claims to Property." Journal
form" should hold: I2 = I2. The observed av- of Political Economy, December 1995,
erage income ratio for these cases was I 12 = 103(6), pp. 1275-88.
1.125, not very far from the prediction. Haavelmo,Trygve. A study in the theory of
For the high value m = 4 of the decisiveness economic evolution. Amsterdam: North-
coefficient, Prediction 2c was that the Paradox Holland, 1954.
of Power would hold in its weak form for the Hirshleifer,Jack. "Conflict and Rent-Seeking
low-resource ratio RI/R2 = 1.67, but should Success Functions:Ratio vs. DifferenceMod-
be violated for the high resource ratio RI/ els of Relative Success." Public Choice, No-
2 = 4. The latter part of this prediction was vember 1989, 63(2), pp. 101-12.
substantially confirmed. For an already high _ "The Paradox of Power." Economics
resource ratio it was predicted that the rich and Politics, November 1991, 3(3), pp.
would get richer, and in fact they did so. But 177-200.
they also did so for the 1.67:1 resource ratio Lindgren,B. W. Statistical theory. New York:
where, according to the theory, the poor Macmillan, 1962.
should instead have improved their position. McCabe, Kevin A.; Rassenti, Stephen J. and
From the point of view of the theoretical pre- Smith,VernonL. "Game Theory and Reci-
diction, the richer contestants might be fight- procity in Some Extensive Form Experi-
ing too hard, or the poorer not hard enough. mental Games." Proceedings of the
Inspection of lines 17-20 of Table 4 indicates National Academy of Sciences, November
that the rich are on average close to F1 = 16, 12, 1996, 93(23), pp. 13421-28.
the predicted amount of resources devoted to McCabe,KevinandSmith,Vernon."Intentionality
fighting, but the poor are falling short of the Detection and Mindreading: Why Game

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions
VOL. 88 NO. 4 DURHAM ET AL.: EXPERIMENTALTESTS OF A MODEL OF POWER 983

Form Matters." Unpublished manuscript, Skogh,Goranand Stuart,Charles."A Contrac-

Economic Science Laboratory,University of tarian Theory of Property Rights and
Arizona,May 1997. Crime." Scandinavian Journal of Econom-
Mertens,Jean-Fransois."Repeated Games," in ics, 1982, 84(1), pp. 27-40.
John Eatwell, Murray Milgate, and Peter Smith, Vernon L. and Walker, James M.
Newman, eds., The new Palgrave: A dic- "Monetary Rewards and Decision Cost
tionary of economics. London: MacMillan, in Experimental Economics." Economic
1987, pp. 151-53. Inquiry, April 1993, 31(2), pp. 245-
Skaperdas, Stergios. "Cooperation, Conflict, 61.
and Power in the Absence of Property Tullock,Gordon."Efficient Rent Seeking," in
Rights." American Economic Review, Sep- James M. Buchanan, Robert D. Tollison,
tember 1992, 82(4), pp. 720-39. and Gordon Tullock, eds., Toward a theory
. "Contest Success Functions." Eco- of the rent-seeking society. College Station,
nomic Theory, February 1996, 7(2), pp. TX: Texas A&M University Press, 1980,
283-90. pp. 97-112.

This content downloaded from 128.235.251.160 on Fri, 16 Jan 2015 03:34:31 AM

All use subject to JSTOR Terms and Conditions