Dynamic Foreign Policy Behavior

Author(s): Matthew J. Lebo and Will H. Moore

Source: The Journal of Conflict Resolution, Vol. 47, No. 1, Building a Science of World
Politics (Feb., 2003), pp. 13-32
Published by: Sage Publications, Inc.
Stable URL: http://www.jstor.org/stable/3176180
Accessed: 03/03/2010 13:36

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=sage.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

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.

Sage Publications, Inc. is collaborating with JSTOR to digitize, preserve and extend access to The Journal of
Conflict Resolution.

http://www.jstor.org
Dynamic ForeignPolicy Behavior

MATTHEWJ. LEBO
Departmentof Political Science
State Universityof New York,StonyBrook
WILL H. MOORE
Departmentof Political Science
Florida State University

How best to classify eventcountsof directeddyadicforeignpolicy behaviorandhow best to model them

arepointsof disagreementamongresearchers.Shouldsuch series be modeledas unitroots ("perfect"mem-
ory) or as stationary("short"memory)?It is demonstratedthat the dichotomouschoice between unit root
(I(1)) and level stationarity(I(0)) is overly restrictive.The intermediate(and more general)possibility of
fractionalintegration(0 < I < 1), a conceptprovenuseful in studiesof aggregateopinion,is applied.Results
show thatfractionalintegrationis extremelycommonandthaterrorcorrectionmechanisms(ECMs)can still
be appropriatein the absence of unit-rootseries. FractionalECMs are used in action-reactionmodels of
bilateralrelationshipsto demonstratethis. Given the frequencyof fractionalintegration,its flexibility, and
the problemsencounteredwhen ignoring it, scholars should incorporatefractionalintegrationtechniques
into theirmodels.

Keywords:fractionalintegration;errorcorrectionmechanisms;foreignpolicy; action-reaction;ARFIMA

Action-reaction models have dominatedthe statistical analysis of foreign policy

behavior(e.g., Ward1982; Dixon 1983, 1986; Goldstein 1991; Goldstein and Free-
man 1990, 1991; Rajmairaand Ward 1990; Goldstein and Pevehouse 1997). These
models areevaluatedin a time-seriessetting, and scholarsinterestedin estimatingthe
parametersof action-reactionmodels have generallyturnedto the econometriclitera-
tureto identifyuseful techniquesfor estimatingparameters.We continuethattradition
by introducingand then using a relatively recent innovationin the analysis of time
series: fractionalintegration.'
In brief, the concernfor the level of integrationof a time series mattersboth for our
understandingof individualvariablesand for the modeling choices we make when
investigatingthe relationshipsbetween variables.Aside from wantingto understand
how the past of a series affects the present,we also wish to understandthatpartof the

1. Our presentationis not technical, but we provide references for readers who want to use the
techniques.

AUTHORS'NOTE:An earlierversion of this article was presentedat the 2001 annualmeeting of the
Peace Science Society in Atlanta.We would like to thankDave Clark,Jon Pevehouse, and Pat Regan for
helpful comments on the article.
JOURNAL OFCONFLICT Vol.47 No. 1,February
RESOLUTION, 2003 13-32
DOI:10.1177/0022002702239509
? 2003SagePublications

13
14 JOURNALOF CONFLICTRESOLUTION

variablenot explainedby internaltrends(the stochasticcomponent)using independ-

ent variables.The standardsolution is to de-trenda series before runningregressions,
and for a numberof years,the standardapproachto de-trendingseries was to takethe
firstdifferenceof the series anduse these differencedseries in regressions.2Failureto
deal with the trendsof each of our series leads to the problemof spuriousregression.
That is, it is surprisinglyeasy to conclude falsely that two series are causally related
when they sharea common trend.
The literatureon fractionalintegrationshows thatone need not be restrictedto the
choice of using the series in levels (i.e., the rawdata)or takingthe firstdifference:One
can take what is known as the fractionaldifference(describedbelow). This is useful
because the fractional integrationliteraturealso shows that over-differencingand
under-differencingproducebiased parameterestimates and bias summarystatistics,
such as the Durbin-Watsonscore andR2.More specifically,the bias affects the size of
the coefficientandstandarderrorsandcan lead one to inferthatrelationshipsarestron-
ger thanthey actuallyare.Overall,fractionaltechniquestakemuchmoreprecisemea-
surementsof the empiricallyobservedbehaviorof time series andallow the specifica-
tion of more exact models.
We study a numberof directed-dyadicforeign policy series to evaluate whether
they are fractionally integrated.We also reportthe results of estimating an action-
reaction model using data that we differenced,fractionallydifferenced, and left in
level form.Ourresultsindicatethatscholarswho studyforeignpolicy series wouldbe
well advisedto familiarizethemselveswiththe fractionalintegrationliteratureandadd
these diagnostic tests to their statisticaltoolkit.
The studyproceedsin five sections anda conclusion.In the firstsection, we briefly
review the literatureon action-reactionmodels of foreign policy behavior.We then
introducethe literatureon fractionalintegrationandfollow with a section in which we
discuss the processes that generate foreign policy series and some reasons why we
expect them to be fractionallyintegrated.In the section that follows, we discuss our
researchdesign andthe datathatwe use to createthe seriesthatwe study.Wethenpres-
ent our results in the fifth section and conclude the articlein a final section.

ACTION-REACTION MODELS
OF FOREIGN POLICY BEHAVIOR

One of the more common theoriesof behaviorin the social sciences is the action-
reactiontheory (Sandberg1978). It is expressed in the notion that actors respondin
kind to eitherhostile or cooperativebehaviorfrom anotheractor,and the theory has
receivedwide applicationin manyareasof inquiry.The causalargumentis simple:the
behaviorof actorA will be driven,at least in part,by the behaviorof actorB. Some-

2. The firstdifferenceis the value of an observationminusthe precedingobservation(i.e., Yt- Yt- 1).

Ratherthanusing the level values of the series (i.e., y,) in regressions,one uses the first differences.
 Lebo, Moore/DYNAMICFOREIGNPOLICYBEHAVIOR 15

times, the causalmechanismis understoodas contemporaneous(i.e., actorA responds

to actorB's presentbehavior),butit is generallyspecifiedas a retrospectiveresponse.3
Ward(1982) introducedthe action-reactionframeworkto the studyof foreignpol-
icy behavior, and Dixon (1983, 1986) contributedadditional analyses. Goldstein
(1991, 1995) and his colleagues (Goldstein and Freeman 1990, 1991; Goldstein and
Pevehouse 1997; Pevehouse and Goldstein 1999) set the presentstandardfor action-
reactionmodels in the studyof foreignpolicy behavior.These models canbe described
generally using the following two equations:

Yt= yY,_-+ Xt+ a (1)

X = eX,_I + rYt+ (2)

6 and 5 are the action-reactioncoefficients: if they are positive, then Y andX covary
positively with the contemporaneousor pastvalues of X and Y.The y and Ecoefficients
representthe extentto which YandX areresponsiveto theirown past values. They are
generally understoodto representbureaucraticprocesses such as standardoperating
proceduresor budget maximization(e.g., Goldstein and Freeman 1990 specify that
they representpolicy inertia).As such, they areexpectedto be positive. Finally,a and
Pare constantsthatrepresentthe level of hostility or cooperationthe countrywould
exhibit towardthe other in the absence of action-reaction and policy inertia. The
model has proveditself to be very useful in a varietyof contexts:cooperativeas well as
hostile foreignpolicy behavior,greatpowerandsmallpowerinteractions,andso forth.
Rajmairaand Ward (1990), Ward and Rajmaira(1992), and Rajmaira(1997)
develop a prospectivemodification of the action-reaction model. Building on the
work of Axelrod (1984, 1986), Keohane (1986), and Axelrod and Keohane (1985),
they arguethatthe directed-dyadicforeign policy interactionsbetween rivals will be
characterizedby short-runreciprocitywithinthe contextof a long-memoriedequilib-
rium process that is drivenby each actor'sexpectationsof the other'sbehavior.
All of these studies estimatethe parametersof the models they propose and do so
using directed-dyadictime series. In this study,we areespecially interestedin making
use of relativelyrecentadvancesin the study of time series and asking whetherextant
studiesmay sufferfromeitherspuriouscorrelationor biasedparameterestimates.The
concernarisesfromthe fact thatthe action-reactionmodelersgenerallydo not address
the issue of integrationof the series they study.
Recentadvancesin time-serieseconometricscast a shadowoverall publishedtime-
series analyses of reciprocityin foreign policy. To make an explicit case, at least six
publishedstudiesestimateparametersusing datareportedin levels withouttesting for
the level of integration(Ward1982;Dixon 1983, 1986;Goldstein 1991;andGoldstein
and Freeman 1990, 1991), two studies test for the level of integrationand use data
reportedin levels (Goldstein and Pevehouse 1997; Pevehouse and Goldstein 1999),
andthreestudiestest for the level of integrationanduse the firstdifferenceof the series

3. In fact, most contemporaneousspecifications are driven more by the level of aggregationused

ratherthanreferenceto an explicit conceptualizationof simultaneousinteraction.
16 JOURNALOF CONFLICTRESOLUTION

Q: Do they Cointegrate? Q: Are they 1(0)?

Tests: ADF, VR KPSS, d Tests: ADF, V, KPSS, d.
for residual series.

Q: Do they fractionally Use Y,

, ; ~ ~ cointegrate? No ECM
Use: AY, AX Use: AY, AX Test: is d' < ? Y-
Yt +
p,X +
ECM No ECM
AYt-f O-PIA- t AY -o + PAX .,
- ci
aI(CMI.l

Use: (1-B)d X,, (1-B)dY, Use: (1-B)d Xt, (1-B)Yt

FECM No ECM
FAY,= po+ PiFAX,- aFECM,.I+ , FAY,= t piF:AX,

Figure 1: When to Use a Fractional ErrorCorrection Mechanism

(RajmairaandWard1990;WardandRajmaira1992;Rajmaira1997). We arguebelow
thatthereis reasonto believe thatthe type of foreignpolicy series used in these studies
is unlikely to be integratedof either order0 or 1. We describe a series of steps that
scholarsstudyingthese series ought to take to determinethe best methodfor estimat-
ing parameters.Those steps are outlinedin Figure 1.
We describethe issues sketchedin Figure 1 in some detail below. Here, we briefly
review the steps one shouldtake.The firstquestionis, do the individualseries contain
unit roots (i.e., are they integratedof order 1)? A stationaryseries forgets changes
(shocks) andreturnsto a constantmeanlevel, whereasa series with a unitrootremem-
berschangesperfectlyandmeanderswith no constantmean.Because the tests for unit
rootshavelow power,it is importantto use severalof them,includinga test to estimate
the value of d, the fractionaldifferencingparameter.If the series contain unit roots,
then Granger (1980) suggests that a linear combination of the series might be station-
ary, which is to say that the two series may be cointegrated. Two unit root series are said
to be cointegrated if a linear combination of the two is mean reverting (i.e., stationary).
If the series contain unit roots, then the next step is to determine whether they are
cointegrated, and several tests are available to make such a determination.
If the series are not cointegrated, then one should take the first difference of each
and use the differenced variables in one's action-reaction model. However, if the
series are cointegrated, then an error correction model is appropriate (again, using first
differences). If one rejects the presence of a unit root, one must also determine whether
to reject or embrace the conclusion of zero order integration. In Figure 1, this decision
is represented as answering "no" to the questions, "Do the series contain unit roots?"
 Lebo, Moore / DYNAMICFOREIGNPOLICYBEHAVIOR 17

and "Arethe series I(0)?" If the individualseries are fractionallyintegrated,then they

may be fractionallycointegrated(which is to say they may containa long-runrelation-
ship). Again, a diagnostictest is available.If one rejectsfractionalcointegration,then
one should estimate one's action-reactionmodel using fractionallydifferencedvari-
ables. However,if one acceptsthe presenceof fractionalcointegration,thenone would
want to estimate a fractionalerrorcorrectionmodel.
Our figure suggests that scholars using time series to study reciprocityin foreign
policy will wantto estimateparametersusing one of five models. To date,people have
selectedamongthreeof those models. We show thattwo otheruseful models areavail-
able. Furthermore,althoughthereareundoubtedlycases in which the threemodels that
people have been using will serve us well, we presentthe argumentthatit is likely the
case thatmanyforeignpolicy seriesarefractionallyintegratedand,given theprevalence
of reciprocityin foreignaffairs,thatmanypairsof foreignpolicy series are likely frac-
tionallycointegrated.To make our case, we begin with a discussionof integration.

INTEGRATION: 0, 1, OR
SOMEWHERE IN BETWEEN?

Traditionally,researchersusing time-series data have concernedthemselves with

the questionof whethertheirdataare stationaryor nonstationary.A stationaryseries
(not integratedor integratedof order 0) is one that, regardlessof short-termshifts,
maintainsa constantmean (or equilibrium)level (see series A in Figure2). As events
thatmove the series from its mean value arequickly forgotten,the series is character-
ized as having "short-memory"only. Such a series may oscillate aroundits equilib-
riumlevel, butit will be characterizedby meanreversionas well as stationaryvariance
and covariance.
In contrast,for an integratedseries (also known as a nonstationary,randomwalk,
unitroot, or I(1) series), shocks arerememberedperfectlyfrom one time periodto the
next, thusaccumulatingovertime withoutanydiscountingtowardan equilibriumlevel
(see series B in Figure2). The series will meanderwith a nonstationarymeanand will
exhibit infinite varianceand infinite covariance.
Parameterestimationwith integratedseries can be problematicas spuriousregres-
sions arecommon (Yule 1926; GrangerandNewbold 1974; Lebo, Walker,andClarke
2000).4To gain an intuitivesense of theproblem,considerthe fictitiousseriesin Figure
3, each of which grows overtime. In general,as one seriesgrows, the otherdeclines, so
they appearto sharea long-runequilibrium.But is the relationshipspurious?It turns
out thatlots of series grow overtime, andbecausethey do, they sharea common trend.
Series thatsharea commontrendbutno causalrelationshipwill producea statistically
significantrelationship.The key, then, is to remove the common trendbefore testing
for a relationship.
4. Spuriousregressions occur when estimationtechniquesfalsely reject the null hypothesis of no
relationshipbetween variables.With time-seriesdata,this is a particularproblembecause one can mistak-
enly conclude thata patternvisible in one variable(such as perfectmemory)explains a similarpatternin a
second variable.
18 JOURNALOF CONFLICTRESOLUTION

Net
Cooperation
Series B, 1(1) _

/
~/ **, : I/

~--.- -]* **4'

/ - /*Series C, I(d)
^' -/ I
-*

Series A,I(0)

Time
Figure 2: 1(0),1(1),and l(d) Series

Toremovethe trend,I(1) series shouldbe differencedpriorto modeling.Only when

all series arerenderedI(0) can one have confidencein the inferencestakenfrommod-
els that include them. Thus, time-series modelers have come to rely on tests of
stationaritysuch as the Dickey-Fullertest (Dickey and Fuller 1979, 1981) to guard
againstthese threats.
Substantively,conclusions aboutwhethera series is integratedimply a greatdeal
aboutthe variableof interest.For variablesmeasuringa government'sforeign policy
behavior,concludingstationarityimplies anequilibriumpolicy level thatremainscon-
stant despite short-termactions or long-termeffects. On the other hand, concluding
that such a variableis nonstationaryimplies the belief that actorscontinuallyupdate
their actions based on new informationand thatno equilibriumbehaviorexists. Nei-
ther of these characterizationsseems to fit our intuitivenotions aboutforeign policy
behavior,a point to which we returnbelow.
Recently,researchersin political science have questionedthe strictdichotomy of
classifying andtreatingvariablesas havingintegerlevels of integrationandhaveintro-
duced the concept of fractionalintegrationfirst describedby Granger(1980; also see
Box-Steffensmeierand Smith 1996, 1998; Lebo and Clarke2000; Clarkeand Lebo
forthcoming).Ratherthan making the strict choice between I(0) or I(1), fractional
integration generalizes the traditional autoregressive integrated moving average
(ARIMA)model to become an autoregressivefractionallyintegratedmoving average
(ARFIMA) model. A fractionallyintegratedseries will have characteristicsof both
 Lebo, Moore/ DYNAMICFOREIGNPOLICYBEHAVIOR 19

Net
Cooperation Series B

Series A

Time
Figure 3: A Spurious Relationship?

stationaryand nonstationaryseries and possess long, ratherthan short or perfect,

memory.Thatis, it will be meanrevertingbutonly overa muchlongerperiodthanwith
a stationaryseries (Box-Steffensmeier and Smith 1996, see series C in Figure 2).
Equation(3) specifies the ARFIMA model for variableY,:

(B)t --1
(1 - B)d Yt=
O(B) (3)

In equation(3), (B) representsthe "backshiftoperator"such that(B)X,= X, _ . Fur-

thermore,4 and0 representstationaryautoregressiveandmoving averageparameters,
and?, is the errorterm,which is distributednormallywith mean 0 and variance02. Of
centralimportancehere is d, the fractionaldifferencingparameter.When d = 0, equa-
tion (3) reducesto a stationaryARMA process.When d = 1, equation(3) reducesto an
unit-rootARIMA process that will requiredifferencingto renderit I(0). Studies of
popularpolitical variables,such as approvalratings,have foundthatmost series fall in
the range 0 < d < 1 (Box-Steffensmeierand Smith 1996; Lebo, Walker,and Clarke
2000; Byers, Davidson,andPeel 2000; De Boef 2000). Fractionallevels of integration
are common because these variablesare constructedby aggregatingthe behaviorof
many individuals,each of whom may behave differentlyin terms of his reliance on
past opinionsor actions to makecurrentdecisions. We contendthatbecause the series
20 JOURNALOF CONFLICTRESOLUTION

of country-levelbehavior are createdby aggregationin several respects (discussed

below), they may well be fractionallyintegrated.To accountfor fractionalintegration,
fractionaldifferencingmust be employedto guardagainstthreatsto inferencesimilar
to those encounteredwhen dealing with wholly integrateddata (Lebo 2000; Clarke
and Lebo forthcoming).
Given thatARIMAmodels can be generalizedto ARFIMAmodels (to accountfor
fractionalintegration),it is not surprisingthatcointegrationmethodscan also be gen-
eralized to allow for fractional cointegration (Abadir and Taylor 1998; Box-
Steffensmeier and Tomlinson 2000; Clarke and Lebo forthcoming). Traditional
cointegrationspecifies that a cointegratingrelationshipbetween variables requires
each seriesto be I(1) andsome linearcombinationof the variablesto be I(0) (Banerjee
et al. 1993). Generalizingthis, fractionalcointegrationrequiresthe original(or "par-
ent")seriesto be I(d), where0 < d < 1, andthe combinationof variablesto havea lower
level of integration, d', such that d' < d. When two series are (fractionally)
cointegrated,they will have an equilibriumrelationship,moving very close together
over the long term. A shock that moves the series apartwill quickly dissipate in the
case of traditionalcointegration,but re-equilibrationwill be much slower where
cointegrationis fractional.
For scholars interested in action-reaction processes, fractional cointegrationis
importantbecause action-reactionprocesses will generatea long-runequilibratingor
error-correctionrelationship.If the series are not I(1), then action-reactionmodels
imply that they should be cointegrated:I(1) series that have a long-runequilibrium
relationshipwill be cointegrated.However,if the series are I(d), where 0 < d < 1, and
have a long-runequilibratingor error-correctionrelationship,then they will be frac-
tionallycointegrated.As we describein moredetailin the following section,the preva-
lent use of action-reactiontheoriesto describedirected-dyadicforeign policy behav-
ior providesgood reasonto investigatethe possibility thatthese series arefractionally
cointegrated.

INTEGRATION AND
FOREIGN POLICY SERIES

Because of the threatto inference, it is importantto test for fractionalintegration

andcointegration.However,doing so may well be morethanan "econometricfix,"for
there is reason to believe that some individualforeign policy series are fractionally
integratedandthatsome pairsof them arefractionallycointegrated.As Crescenziand
Enterline (2001) have recently emphasized, most students of internationalpolitics
believe thatthe past matters.The orderof integrationof a series has implicationsfor
the mannerin which past events influence presentbehavior.First,considerthe argu-
mentwe raisedbriefly above.If a foreignpolicy series is stationary,it alwaysrevertsto
its mean,which suggeststhat(1) policy is constantovertime and(2) eventsdo not have
an impactfor long. This strikesus as a descriptionthatis unlikelyto comportwith most
scholars'sense of foreign policy behavior:althoughforeignpolicy is certainlysticky,
it does change over time and sometimes quite dramatically.Furthermore,events may
 Lebo, Moore / DYNAMICFOREIGNPOLICYBEHAVIOR 21

well have a lastingeffect: theirimpactwill likely decay overtime, butthey areunlikely

to be quickly forgotten.
Let us now considera unit-rootforeignpolicy series:here,the impactof pastevents
never decays-they have just as much of an effect on policy today as they did when
they first occurred.This description strikes us as even less likely to comport with
scholars' intuition of foreign policy behavior.The third option, a fractionallyinte-
gratedseries, is one in which memoryis long but not perfect.In the contextof foreign
policy series, this suggests that events from the past have an impact long into the
future-long enough thatwe cannotmodel them with autoregressiveor moving aver-
age terms-but that the impact of the past decays over time. Furthermore,different
countries (and pairs of countries)might exhibit differentpatterns.Fractionallyinte-
grated series can account for such variationand seem to bettermatch our intuitions
aboutforeign policy thaneither stationaryor unit-rootseries.
As noted above, if the series are generatedin such a way thatthe researcheraggre-
gates across heterogeneousdata-generatingprocesses, the series will likely be frac-
tionallyintegrated.Thus,in additionto makingourcase basedon intuition,we also put
forth an argumentabout sources of heterogeneitythat might exist in foreign policy
series. We identify two such sources below: differentactors and differentissues.
The most likely source of short-memoriedand long-memoriedheterogeneity is
aggregationacross differentactors.One mannerin which researchersmight studythe
behaviorof multiple actors is to aggregateacross multiplecountries.Most studies of
foreign policy examine directeddyads that are pairs of countries,but in a few cases
researchershave aggregatedthe behaviorof severalcountriesto createa single series
(Davis and Ward1990; Goldsteinand Pevehouse 1997). For example, Goldstein and
Pevehouse(1997) studythe behaviorof the UnitedNations (UN) andNATOcountries
towardSerbiaduringthe Bosnian conflict of the 1990s. They thus createdan interna-
tional actorsseries thatwas the sum of the behaviorof each actor(i.e., the UN and the
NATOcountries).Because thereis little reasonto believe thatthe UN and each of the
NATOcountriesusedthe pastin the same way whenrespondingto the Serbs,this inter-
nationalactors series is an aggregationacross heterogeneousactors. Thus, we argue
thatwheneverresearcherscreatean aggregateseries thatcombinesthe behaviorof two
or more countriestowardanother,that series is likely to be fractionallyintegrated.5
The second reasonwhy series mightbe understoodas an aggregationof the behav-
ior of multipleactorsis thatforeignpolicy is not the behaviorof a unitarystate:execu-
tives, bureaucrats,and militaryofficials all contributebehaviorto the foreign policy
series we study,and it is ratherlikely that some of them will use the past differently

5. We are less sanguinethatan aggregateseries of countriestowarda single actorwill be fractionally

integrated.Ourthinkingis thataggregatingthe behaviorof a single actor-whether an individualcountryor
a corporateactor,such as NATO-toward all countrieswill aggregateover the heterogeneousbehaviorof
thatactortowardothers.Thatis, we suspect thatany given actorwill have varianceover the importanceof
past events in its relations with differentcountries. Yet, we are less confident that this will be so for a
noncorporategroupof actors,such as the behaviorof all countriesin the worldtowarda given country.It is
certainlyplausiblethattherewill be heterogeneityin the importanceof memory,butit also seems to us plau-
sible thattherewill be ratherlimited heteroegeneity.This mightespecially be so with respectto a hegemon
such as the United States.Thatis, most third-worldcountriesmighthave rathersimilarresponsepatternsto
U.S. behavior,thus limiting the heteroegeneityacross such actors.
22 JOURNALOF CONFLICTRESOLUTION

when formingtheirjudgments.Tothe extentthis is so, actorswith somewhatheteroge-

neous informationgeneratethe foreign policy behaviorof a given country.
The second point concerns the claim that countries interactover specific issues
(Vasquez and Mansbach 1984; Dixon 1986; Vasquez 1993; Hensel 2001). By this
argument,the foreignpolicy series thatscholarsstudyareaggregatesof foreignpolicy
interactionsacrossdifferentissue areas.Forexample,Dixon (1986) arguesthatdiffer-
ent action-reaction processes describedU.S.-Soviet interactionsin political, mili-
tary,and culturalissue areas. A similar argumentis raised by scholarswho contend
thatcooperativebehaviorandhostile behaviorshouldbe measuredon two dimensions,
insteadof a single cooperation-hostilitydimension.To the extent that foreign policy
behavioris somewhatdistinctacrossissue areas,actorsmightwell have differenttime
horizonsfor differentissues. To the extentthatthis is so, a single series thataggregates
across issue areaswill be fractionallyintegrated.
This discussionimplies a few expectations.First,we conjecturethatforeignpolicy
series createdfromeventsdataarelikely to be fractionallyintegrated.In some series, a
consistent foreign policy across multiple issues may well produce stationaryseries.
We are not ruling those out. Similarly,there may be a few series that exhibit perfect
memory(i.e., a unitroot),althoughwe arecircumspectaboutthe likelihoodof observ-
ing such series. However,we suspect that at a minimum,many,and likely most, for-
eign policy series are fractionallyintegrated.Second, and more specifically, we sus-
pect that foreign policy series that are aggregatedacross actors will be fractionally
integratedand exhibit longer-memoriedprocesses than directed-dyadicseries com-
posed of one country'sbehavior.Third,given the prevalenceof reciprocitymodels of
foreign policy in the literature,we suspect that some, and perhapsmany,dyadic for-
eign policy series arefractionallycointegrated.In the following section, we specify a
reciprocitymodel and describe the data and researchdesign we use to explore these
expectations.

RESEARCH DESIGN

Because we are interestedin reciprocityin foreignpolicy, we specify and estimate

an action-reactionmodel. As noted above, if the series have a long-runequilibrium
relationship,then an errorcorrectionmodel will help us model that relationship.An
error-correctionrelationshipis capturedvia the differenceof the lagged values of two
series. If we have two series, say the foreignpolicy behaviorof countryx towardcoun-
tryy (call thatseries X,) andthe foreignpolicy behaviorof countryy towardcountryx
(call thatseries Yt),thenthe error-correction mechanism(ECM)canbe writtenas (X_,
-_ _ ).6
To see why the ECMhelps one model a long-runequilibriumrelationship,consider
what happensto the change in X, when the difference between X,_ and Y, , is zero:
6. This specificationof the errorcorrectionmechanismimposes the simplifyingassumptionthatboth
countriesrespondto one anotherin a perfectlyreciprocalfashion.A generalizederrorcorrectionmechanism
(GECM)can be writtenas AX,= -a(X, _ I - kY,- i). We assume thatk = 1. The virtueof this assumptionis
thatit simplifiesthe analysis.See Banerjeeet al. ( 1993, 50-61) for a discussionof how to estimatethe GECM
and DeBoef (2001) for an applied model.
 Lebo, Moore / DYNAMICFOREIGNPOLICYBEHAVIOR 23

AX, = - (X,_ I - Y,_ ). In such a circumstance, there will be no change in X,. Now imag-
ine thatthe differenceis positive (i.e., X,_ > Y,_ ). Inthatsituation,AX,will havea neg-
ative value as it is revised downwardto bring it back in line with Y,.If the difference
between X,_ , and Y, is negative, then AX,will have a positive value.
Of course, accordingto action-reactionmodels, Y,will also be influenced by the
difference of the lagged value of the two terms,andwe can write thatas (AY,= (X,_ -
Y,_-). Note the absence of a negative sign in frontof the ECM in this equation.This
ensuresthatwhen the differencebetweenX, 1and Y, , is positive (i.e., Y,_ <X,_ I),AlY,
will be positive, bringing Ytback in line with X,. Furthermore,when the difference
betweenX,_ 1 and Y,_ is negative,AY,will be negative.Thus, the ECM is a convenient
termfor modelinga long-runequilibriumrelationshipin a regression.Because action-
reactiontheoryleads us to anticipatesuch a relationship,we addanECMtermto equa-
tion (1) above.
There are, of course, an enormous numberof pairs of countries (or dyads) one
might study. We selected Egypt and Israel, the United States and China, the United
States and Russia, the United States and France,and the United States and Israel in
honor of Ward's(1982) initial study in this area, which focused on those countries.
This spatialdomaindoes not permitus to generalizebeyondthe sample-it is not a
randomsampleof directeddyads. However,we have no cause to believe thatit suffers
from any specific sample selection bias and thus cannot argue that the results are
unrepresentativeof the populationof directeddyads.We studymonthlyforeignpolicy
behavior,and the temporaldomain of the study varies in accordwith the data set we
used: 1948:1 to 1978:12 (Cooperationand Peace Databank[COPDAB], 1966:1 to
1991:12 (WorldEvent InteractionSurvey [WEIS]), and 1987:1 to 1997:6 (Kansas
Events Data System [KEDS]). We study several data sets to see whetherour results
were dependenton a specific coding scheme. All directeddyads are availablein the
COPDAB and WEIS data, but the KEDS projecthas a more limited set of options.
Because the U.S.-Chinadirecteddyad was availablein the KEDS data,we selected it
as the directeddyad to study across all threedata sets.7
All three data sets take the same general approachto measuringforeign policy
behavior:they code news reportsof events using a scheme in which an actorandtarget
areidentified,the type of behaviorsent by the actorto the targetis coded on an ordinal
scale that measures a cooperativeto hostile dimension, and the date of the event is
recorded.The COPDAB and WEIS data were createdusing humancoders, and the
KEDS data were createdwith machine-assistedcoding.
Weuse the net scoreof the weightedevents,wherenet score equalsthe differenceof
cooperative and hostile events aggregatedover the month. The COPDAB weights
were createdby surveys of expertsby Azar (1993) and are reportedin the COPDAB
manual.Goldstein (1992) createda set of weights for the WEIS data, and the KEDS
projecthas produceda slightly modified set of Goldsteinweights for its data.8These
weights are used to create interval-likedata from the ordinal COPDAB and WEIS

7. Due to considerationsof space and ease of presentation,we focus the bulk of our analysis on one
data set, COPDAB.
8. These weights are availableonline at http://www.ukans.edu/-keds/KEDS.WEIS.code.html.
24 JOURNALOF CONFLICTRESOLUTION

scales. To aggregatethe series into monthly intervals,we calculatedthe mean of all

weighted events duringthe month.

ESTIMATION AND RESULTS

We examinethe propertiesof severalpopulareventstime seriesused in the interna-

tional relations literature.We begin with a univariateexaminationof the individual
series to determinewhetherthey are (1) stationary,or I(0); (2) fractionallyintegrated,
or I(d), where0 < d < 1; or (3) integratedI(1). We find thatwith one exception,thereis
strong empirical evidence that these foreign policy series are all fractionally inte-
grated.Afteroutliningourseriesof tests, we presentthe resultsof a multivariateanaly-
sis thatestimatesparametersfor an action-reactionmodel in a fractionalcointegration
framework.9
Table 1 shows the results of multipletests of stationarity.10 The Dickey Fullerand
AugmentedDickey-Fuller each tests a null hypothesisof a unitroot,makingit suscep-
tible to false acceptanceof unitrootswhen dataareinsufficientfor significantt scores.
Diebold and Rudebusch(1991) furtherdemonstratethe low power of Dickey Fuller
tests when faced with fractionalalternatives.For every variablein Table 1, we can
rejectthe null hypothesisof unit-rootbehavior.However,given the limitationsof these
tests and the possibility of fractionalintegration,we do not stop at this point and use
additionaltests to furtherdiagnose our series.
The varianceratiotest tests a null hypothesisof a randomwalk with driftversusan
alternativehypothesis of fractionalnoise (Diebold 1989). Again, we reject the null
hypothesisof a randomwalk for every series of interesthere. Rejectionof unit-root/
random-walkhypothesesdoes not necessarilyimply thata series is I(0), andthusaddi-
tional tests with null hypothesesof stationarybehaviorare useful.
Two such tests aredevelopedby Kwiatkowski,Phillips, Schmidt,andShin (KPSS)
(1992). Each of the KPSS tests tests a null hypothesis of a stationary,strongmixing
process. The KPSS ml,test includes an interceptand linear time trend, whereas the
KPSS r, test does not. Table 1 shows thatin most of our series, the null hypothesisof
stationarityis rejectedby one or both of these tests.
The last column of Table 1 summarizesthe findings of these four tests. In 7 of 13
series, the tests agreethatwe shouldrejectbothextremesandconcludefractionalinte-
gration.For anothertwo series (United States to Chinaand Israelto Egypt), the tests
are contradictory,suggesting eitherfractionalintegrationor level stationary.Tests of
the remaining4 series indicate level stationarity.Given this sometimes contradictory
evidence and tests based on dichotomous choices, formal estimates of d become
extremelyuseful at this point to diagnose the precise level of integration.

9. In the interestof conserving space, Table 3 presentsonly one set of parameterestimates, but we
estimatedparametersusing all of the COPDAB series. Those results are similar,althoughnot always as
strong,and will be made availablein a replicationdata set.
10. For a brief comparativesummaryof the tests to follow, including estimatorsfor d, see Lebo,
Walker,and Clarke(2000).
 Lebo, Moore / DYNAMICFOREIGNPOLICYBEHAVIOR 25

TABLE 1
Summaryof StationarityDiagnostics
Data Set, Year;
VariableName DF/ADF VarianceRatio KPSS(Tri) KPSS(rj ) Diagnosis

KEDS 87:1-97:6
China to United States Reject d = 1 Reject d = 1 d= 0 d= 0 d=O
United States to China Reject d = 1 Reject d = d=O d=0 d=0
WEIS 66:1-91:12
Chinato United States Reject d = 1 Reject d = 1 Reject d = 0 Reject d = 0 0<d< 1
United States to China Reject d = 1 Reject d = 1 Reject d = 0 d= 00 d< 1
COPDAB 48:1-78:12
China to United States Reject d = 1 Reject d = 1 Reject d = 0 Reject d = 0 0<d< 1
United States to China Reject d = 1 Reject d = 1 Reject d = 0 Reject d = 0 0<d< 1
United States
aggregatesent Reject d = 1 Reject d = 1 Reject d = 0 Reject d = 0 <d < 1
United States
aggregatereceived Reject d = 1 Reject d = 1 d = 0a d=0 d=0
Egypt to Israel Reject d = 1 Reject d = 1 Reject d = 0 Reject d = 0 0 <d< 1
Israel to Egypt Reject d = 1 Reject d = 1 Reject d = 0 d = Oa 0d< 1
United States to USSR Reject d = 1 Reject d = 1 Reject d = 0 Reject d = 0 0<d< 1
United States to France Reject d = 1 Reject d = 1 Reject d = 0 Reject d = 0 0<d< 1
United States to Israel Reject d = 1 Reject d = 1 d=0 d=0 d=0

NOTE:DF = DickeyFuller;ADF = AugmentedDickey-Fuller;KEDS= KansasEventsDataSystem;WEIS=

WorldEvent InteractionSurvey;COPDAB = Cooperationand Peace Databank.
a. Reject at .1 level.

Table2 shows the resultsof ourd estimates.Eachestimateof d is testedagainstboth

the possibilitythatthe series is I(0) andthatit is I( 1)-t values shownarebasedon how
manystandarderrorsawayfrom0 and 1 estimatesarethusgiving a muchbetterability
to pinpointthe level of integrationthananyor all of the tests Table1 provides.Based on
the d estimates,we can conclude thatd does not equal0 for 12 of our 13 series:behav-
ior between the United States and Chinain each of the KEDS, WEIS, and COPDAB
data sets; between Egypt and Israelin the COPDAB data set; and for aggregatemea-
suresof UnitedStatesbehaviorin the COPDABdata.One possible exceptionis United
Statesto Chinain the KEDS data.AmericanbehaviortowardIsrael,for which ouresti-
mateof d is 0, is the only series thatappearsto be level stationary.Note thatthe 4 series
with the lowest estimatesof d (UnitedStatesto Israeld = 0, UnitedStatesto Chinad =
.10, United States received, and China to United States d = .15) are the 4 series for
which our stationaritytests cannotrejectthe possibility thatd = 0. The generalagree-
ment between the unit-root/stationarity tests and the d estimates demonstrates the
value of employing multiple tests of stationarity to first home in on the correct order of
integration and then pinpoint it using the direct estimates.
Furthermore, Table 2 tells us that we can be very confident that d does not equal 1
for any of these series. This result indicates that using first differencing would over-
difference the series.l Thus, fractional integration characterizes all but one (or per-
11. For the consequences of over-differencing,see DeBoef and Granato(1997) and Lebo (2000).
26 JOURNALOF CONFLICTRESOLUTION

TABLE2
Estimatesof d from (0, d, 0) Models

Data Set, Year,VariableName Estimateof d StandardError t Value,d ? 0 t Value,d # I

KEDS 87:1-97:6
Chinato United States 0.15 0.07 2.06* -11.65**
United States to China 0.10 0.07 1.37 -12.34**
WEIS 66:1-91:12
China to United States 0.24 0.05 4.75** -15.05**
United States to China 0.21 0.05 4.16** -15.64**
COPDAB 48:1-78:12
Chinato United States 0.21 0.05 4.46** -16.80**
United States to China 0.33 0.05 7.02** -14.24**
United States aggregatesent 0.53 0.05 11.27** -9.99**
United States aggregatereceived 0.14 0.05 2.98** -18.28**
Egypt to Israel 0.47 0.05 9.99** -11.27**
Israelto Egypt 0.55 0.05 11.69** -9.57**
United States to USSR 0.29 0.05 6.17** -15.09**
United States to France 0.17 0.05 3.61* -17.65**
United States to Israel 0.00 0.05 0.00 -21.26**

NOTE:KEDS = KansasEventsDataSystem;WEIS= WorldEventInteractionSurvey;COPDAB= Coop-

erationand Peace Databank.Estimatesof d arebased on Robinson's(1995) Gaussiansemiparametricesti-
mator using regression analysis of time series procedure (RGSER). RGSER is available at http://
www.estima.comor by request.To obtain the d estimate of the undifferencedseries, one must first differ-
ence the series creatinga (0, 1 + d, 0) series, thenestimated andfinally add 1 to the d estimateobtained(see
Lebo, Walker,and Clarke2000 for an explanationof these and otherestimatorsof d).
*p <.05. **p <.01.

hapstwo) of these series. Eachof these series will be meanrevertingin the long term,
but events will have far more enduringeffects than researcherswho have concluded
stationarityhave assumed. These results supportour expectationthat foreign policy
event count series are fractionally integrated. Finally, these results tell us that
cointegrationin the traditionalsense cannot be present, and thus traditionaluse of
ECMs is not appropriate. However, given fractional integration, fractional
cointegrationmay exist, allowingfor the use of fractionalerrorcorrectionmechanisms
(FECMs).
Given the frequencywith which series like these will be fractionallyintegrated,
new methods must be applied to effectively judge the relationshipsbetween two or
more such variables. Two majorareas are importanthere: first, the practice of frac-
tional differencing must be applied to ensure that variables are not under- or over-
differenced; second, because none of these variables is a unit root, traditional
cointegrationtechniqueswill not apply.We must acquaintourselves with how to test
for, and model with, the more generalfractionalcointegrationtechniques.'2

12. See De Boef (2001) for a discussionof errorcorrectionmodels in the contextof nearintegration,a
topic that is beyond the scope of this study.
 Lebo, Moore / DYNAMICFOREIGNPOLICYBEHAVIOR 27

As our example here, we estimateparametersfor an action-reactionequationthat

capturesEgyptianactionstowardIsraelas compiled for the 1948 to 1978 periodin the
COPDABdataset. As seen in Table2, boththeEgypt-IsraelandIsrael-Egyptseries are
fractionallyintegrated.Given the history between the two countries,we would cer-
tainlyexpect a close relationshipbetweenthe two series:actionsby one arelikely to be
reciprocatedby the other.Tojudge the closeness of thisrelationshipin a simple regres-
sion framework,we would need to first fractionallydifferenceeach series.13
Before estimating this model, however, we test for evidence of fractional
cointegration. Again, given that neither series is I(1), traditional approaches to
cointegrationwould not allow the possibility of testingfor the existence of a long-run
equilibriumrelationshipbetween the two. However, we can test for a fractionally
cointegratingrelationshipby relaxingthis qualificationandtestingto see if some com-
binationof two variableshas a lower level of integrationthaneitherof the two parent
series. Forthe Egypt-Israeldirecteddyad,the residualsof a simple regressionareinte-
grated of a significantly lower level (d' = .17) than are either of the parent series
(Egypt-Israeld = .47, Israel-Egyptd = .55). Thus, finding thatd' < d for both Egypt-
Israel and Israel-Egypt,we conclude that a fractionally cointegratingrelationship
exists betweenthe two series.Thatd' is morethansix standarderrorsbelow d for either
parentseries providesconfidencein this conclusion.Thus,an equilibriumrelationship
exists between the two variables such that an event that moves them apartwill be
impermanent(Clarkeand Stewart1995). Withfractionalcointegration,however,the
rateat which the two series returnto theirequilibriumfollowing such an event is much
slowerandnonconstantthanin a cointegratingrelationship;thatis, the errorcorrection
termitself is long memoried.Given this relationship,a FECMis appropriateto model
the equilibriumrelationship(Clarkeand Lebo forthcoming).
Having found evidence of fractionalcointegration,we now estimate equation(1)
with an ECM term added to capturedirectly the error-correctionprocess. We report
three sets of estimates in Table 3: a set in which we have fractionallydifferencedthe
series, a set in which we differencethe series as would be done if one determinedthat
the series were unit roots, and a set in which the parameterswere estimatedusing the
level series as would be done if one determinedthatthe series were I(0). Thus, we can
comparewhetherthe estimatesare influencedby these decisions. Althoughthe signs
and significanceof the parametersdo not vary across the threemodels, the size of the
parametersand some of the summarystatistics vary in precisely the fashion antici-
pated in the literatureon integration.
Column 1 of Table 3 shows our multivariatemodel of Egyptian actions toward
Israel. The dependentvariableis fractionallydifferencedby its own value of d (.47)
using the filter (1 - B) 47prior to estimation. Includedas independentvariablesare
IsraeliactionstowardEgypt(filteredby (1 -B) 55)anda FECM(filteredby (I -B) 17).
The close relationshipbetween Egyptianand Israeliactions is evident. The coeffi-
cient of .29 on the Israelto Egyptvariableand the t statisticof 6.75 indicatethatEgypt
reciprocatesIsrael'shostile behaviorwith behaviorthatwill on averagebe aboutone-
13. Whereasfirstdifferencingappliesthe filter(I - B) to a given series, fractionaldifferencingapplies
the more generalfilter (1 - B).. The regressionanalysis of time series (RATS)code for fractionallydiffer-
encing the series will be included in our replicationdata set.
28 JOURNALOF CONFLICTRESOLUTION

TABLE3
EgyptianBehaviortowardIsrael, 1948 to 1978:
ARFIMA,ARIMA, and ARMA Models

FractionalModela DifferencedModelb Level-FormModelc

IndependentVariable Coefficient SE Coefficient SE Coefficient SE

Constant 3.20** 0.83 0.05 0.83 -9.61** 1.08

Israelto Egypt 0.39** 0.04 0.36** 0.04 0.50** 0.04
t 6.75 9.65 13.04
(F)ECM{1} -0.39** 0.05 -0.88** 0.05
t -7.75 -16.90
Durbin-Watson 1.89 2.05 1.84
Sum of squaredresiduals 88,225 93,917 98,033
Standarderrorof estimate 15.50 15.98 16.28
CenteredR2 0.20 0.47 0.31

NOTE:ARFIMA= autoregressivefractionallyintegratedmoving average;ARIMA = autoregressiveinte-

grated moving average; ARMA = autoregressivemoving average; FECM = fractionalerrorcorrection
mechanism.
a. ARFIMA:Dependentvariabledifferencedby 0.47; Israelto Egypt differencedby 0.55; errorcorrec-
tion mechanismdifferencedby 0.17.
b. ARIMA:All variablesdifferencedby 1.
c. ARMA: All variablesleft in level form, no errorcorrectionmechanismestimated.
**p < .01, one-tailed test.

thirdas hostile as Israel'sbehavior.More specifically,if Israelwere to mobilize troops

or impose sanctions(a weighted score of 29 on the COPDAB scale), then on average
Egypt can be expected to respondwith an 8.4 level event, which is closest to a diplo-
maticstatementexpressingdissatisfactionwith the action(a weightedscoreof 6 on the
COPDAB scale). But this coefficient tells only partof the story.
The FECMparameteris also statisticallysignificant,indicatingthatan equilibrium
relationshipexists betweenthe two. Thus, actionsthatmove these variablesapartwill
eventuallydissipate as the two re-equilibrate39% in the first period and (1 - B)39in
subsequentperiods.Put differently,Egypt's initial response will be measured-only
one-thirdas hostile as Israel'sact. However,Egyptwill on averagecontinueto raise its
conflict towardIsrael in subsequentmonths:the effect of the troop mobilizationor
sanctionswouldnot dissipatewith Egypt'sinitialresponse.Thus,failureto includethe
FECMtermwould misconstruethe effect of Israel'sbehavioron Egypt's foreign pol-
icy towardIsrael(which we will explorebelow). How does the FECMmodel compare
with a traditionalECM model?
The second columnin Table3 shows thatusing a traditionalECMherewould over-
estimateEgypt's averageresponseto Israelandthe speed with which equilibriumwill
be regainedonce lost. In the differencedmodel, the variablesaredifferencedin a tradi-
tional (0, 1,0) format,andan ECM is used as thoughthe series cointegratein the tradi-
tional sense. First,note thatthe estimatedcoefficientfor the Israelto Egyptvariableis
.36. This also implies a roughly one-thirdaverageresponse rate by Egypt to Israeli
hostility, although .36 is larger than .29. The larger bias is found in the estimated
 Lebo, Moore /DYNAMICFOREIGNPOLICYBEHAVIOR 29

parameterof the ECM versus the FECM: .88 comparedwith .29. If we relied on the
differencedmodel and ECM, we would conclude that 88% of a shock thatmoved the
two variables apartwould be forgotten one month later, and another88% of what
remainedwouldbe forgottenin the monthafterthat.The FECMof the correctlyspeci-
fied model of the firstcolumn demonstratesthatre-equilibrationis a muchmoregrad-
ual process: Egypt "remembers"events far longer than the differencedmodel sug-
gests. If these results generalize to other dyads, then Rajmairaand Ward's (1990)
cointegrationfindingswith respectto the U.S.-USSR rivalryoverestimatethe speed at
which they returnto equilibrium.
Furthermore,the second column of Table3 demonstratesthe severalproblemswe
introduceby failing to properlyaccountfor fractionalintegrationin ourvariables.The
values of t statisticsarebiased, as is the R2statistic,which morethandoublesfrom .20
to .46, making evident the tendency of over-differencedmodels towardfinding rela-
tionships where none exist.
Last, the thirdcolumn of Table3 treatsthe variablesas thoughthey were level sta-
tionaryand uses them in theiroriginalform. The resultsarebiased in the estimatesof
coefficients and model fit.14 First, the coefficient for the Israelto Egypt variablesug-
gests thaton average,Egyptmetes out aboutone-halfof Israel'shostile behavior.Con-
tinuingwith the examplefrom above, if Israelmobilizedtroopsor imposed sanctions,
this suggests thatEgypt would respondwith a 14.5-level event, perhapsby a warning
retaliation(which is closest with a weighted COPDAB score of 16). However, that
event will not have an impact in subsequentmonths.
Most important,the fractionalmodel is trueto the natureof the data-generatingpro-
cess as tested, whereasthe differencedand level-formmodels are not. The fractional
model incorporatesthe memory of the variablesof interest as precisely estimated,
whereas the second and thirdmodels proceed from the false assumptionthat d must
equal 0 or 1 and guess work about which extreme should be appropriate.
Substantively,one goal of action-reactionstudiescan be to analyzethe responsive-
ness of eachpartyto thebehaviorof theotherparty(or,in morecomplexaction-reaction
models, the behaviorof a thirdparty).Althoughthe purposeof this articleis to demon-
stratenew tools, ourresultsdo have some substantiveimplications.First,the low esti-
mates of d suggest thatalthoughthe past matters,memoryis finite:the estimatessug-
gest thatthe series arelong memoried,but,relativelyspeaking,not exceptionallylong
memoried(eventsdo not continueto have an impactyearsor decadeslater).This find-
ing implies thathawkishperspectivesthatemphasizethe impossibilityof cooperation
between rivals are politicized claims rather than reality-based claims (see also
Goldstein and Freeman 1990). Second, the fractionallydifferencedestimates of the
error-correctionaction-reactionmodel indicatethatalthoughEgypt was unwilling to
driftapartfrom Israelfor too long duringthe Nasser and Sadatyears, it was willing to
drift apartfor longer than one would infer if one did not fractionallydifference the
series. Thus, with respect to the main purposeof the article,this example shows the
importanceof incorporatingfractionalintegrationin studiesof internationalbehavior.

14. These problemsare not solved by including a lagged endogenous variable.

30 JOURNALOF CONFLICTRESOLUTION

CONCLUSION

Extant time-series studies of foreign policy behavior are subject to criticism

becausethey haveconfinedtheirmodelingchoice to eitherstationaryor nonstationary
behavior.This oversimplificationleads us to misunderstandboththe natureof individ-
ual variablesand the relationshipsamong variables.We have arguedthatforeign pol-
icy series aggregateover heterogeneousdata-generationprocesses and arethus likely
to be fractionallyintegrated.These expectationsare met because our estimatesof the
fractionaldifferencingparameter,d, indicate that fractionalintegrationis prevalent
among these series. We furtherfind supportfor our contentionthat memory will be
longer (and d values larger)in relationshipsbetween friendly nations.
Choosinga level of integrationto describeone's variablesof interestis not simply a
matterof modelingchoice. Rather,it implies importantassumptionsaboutthe nature
of the variablesthemselves.Limitingourselvesto just two such choices is an oversim-
plificationthatleads to ourmisunderstandingthe natureof the variables.Furthermore,
using methodssuch as ARIMAthatallow for only integerlevels of integrationleads to
large problems when we estimate the relationships between variables. Using
ARFIMAmodels insteadallows us a much more precise explanationof the data and
increasesour ability to understandcomplex political relationships.
In our example above, we found thatthe relationshipbetween Israeland Egypt for
the period from 1948 to 1978 is fractionally cointegrated.Thus, as suggested by
RajmairaandWard(1990), cointegrationtechniquesareappropriate forthe studyof for-
eign policy in rivalries;however,these authorsmay well overstatethe closeness of the
cointegratingrelationship.We find evidence that althoughthese two series exist in a
long-termequilibrium,disruptionsto thisequilibriumwill takea long timeto dissipate.
Futurestudiesthatuse directed-dyadicforeign policy series and othermeasuresof
internationalbehaviorover time should carefully consider the constructionof their
series andbe awareof the particularproblemaggregationengenders.The use of frac-
tional integrationand fractional cointegrationtechniques should greatly enhance
researchers'abilities to specify the dynamics of internationalbehavior.

REFERENCES

Abadir,K. M., and A. M. R. Taylor.1998. On the definitionsof (co-) integration.Unpublishedmanuscript,

Universityof York.
Axelrod, R. 1984. The evolutionof cooperation.New York:Basic Books.
.1986. An evolutionaryapproachto norms.AmericanPolitical Science Review 80:1095-111.
Axelrod, R., and R. O. Keohane. 1985. Achieving cooperationunderanarchy.WorldPolitics 38:226-54.
Azar, E. E. 1993. Conflict and Peace Databank (COPDAB), 1948-1978. ICPSR study no. 7767, second
release. Ann Arbor,MI: Inter-UniversityConsortiumfor Political and Social Research.
Banerjee,A., J. Dolado, J. W. Galbraith,and D. F. Hendry.1993. Cointegration,error correction,and the
econometricanalysis of non-stationarydata. New York:OxfordUniversityPress.
Box-Steffensmeier,J. M., andR. M. Smith. 1996.The dynamicsof aggregatepartisanship.AmericanPoliti-
cal Science Review90:567-80.
 Lebo, Moore /DYNAMICFOREIGNPOLICYBEHAVIOR 31

. 1998. Investigatingpolitical dynamicsusing fractionalintegrationmethods.AmericanJournalof

Political Science 42:661-89.
Box-Steffensmeier,J. M., and A. R. Tomlinson.2000. Fractionalintegrationmethods in political science.
Electoral Studies 19:63-76.
Byers, D., J. Davidson, and D. Peel. 2000. The dynamicsof aggregatepolitical popularity:Evidence from
eight countries.Electoral Studies 19:49-62.
Clarke,H. D., andM. J. Lebo. Forthcoming.Fractional(co)integrationandgoverningpartysupportin Brit-
ain: A methodologicalresearchnote. BritishJournal of Political Science.
Clarke,H. D., and M. C. Stewart. 1995. Economic evaluations,prime ministerialapprovaland governing
partysupport:Rival models reconsidered.BritishJournal of Political Science 25:145-70.
Crescenzi,M. J. C., andA. J. Enterline.2001. Timeremembered:A dynamicmodel of interstateinteraction.
InternationalStudies Quarterly45:409-31.
Davis, D. R., andM. D. Ward.1990. They dancealone:Deathsandthe disappearedin Chile.Journalof Con-
flict Resolution34:449-75.
De Boef, S. 2000. Persistence and aggregations of survey data over time: From microfoundationsto
macropersistence.Electoral Studies 19:9-29.
.2001. Modeling equilibriumrelationships:Errorcorrectionmodels with strongly autoregressive
data.Political Analysis 9:78-94.
De Boef, S., andJ. Granato.1997. Near-integrateddataandthe analysisof politicalrelationships.American
Journal of Political Science 41:619-40.
Dickey, D. A., andW. Fuller. 1979. Distributionof the estimatorsfor autoregressivetime series with a unit
root. Journal of the AmericanStatisticalAssociation 84:427-31.
.1981. Likelihood ratio statistics for autoregressivetime series with a unit root. Econometrica
49:1057-72.
Diebold, F. X. 1989. Randomwalks versus fractionalintegration:Power comparisonsof scalar andjoint
tests of the variance-timefunction.InAdvancesin econometricmodeling,edited by R. Raj. New York:
Kluwer.
Diebold, F. X., andG. D. Rudenbusch.1991. On the powerof Dickey-Fullertests againstfractionalalterna-
tives. EconomicsLetters35:155-60.
Dixon, W. J. 1983. Measuringinterstateaffect. AmericanJournal of Political Science 27:828-51.
.1986. Reciprocityin United States-Sovietrelations:Multiple symmetryor issue linkage?Ameri-
can Journal of Political Science 30:421-45.
Goldstein, J. S. 1991. Reciprocity in superpowerrelations:An empiricalanalysis. InternationalStudies
Quarterly35:195-209.
.1992. A conflict-cooperationscale for WEIS internationalevents data.Journalof ConflictResolu-
tion 36:369-85.
.1995. Great-powercooperationunderconditionsof limited reciprocity:Fromempiricalto formal
analysis. InternationalStudies Quarterly39:453-77.
Goldstein,J. S., andJ. R. Freeman.1990. Threeway street:Strategicreciprocityin worldpolitics. Chicago:
Universityof Chicago Press.
.1991. U.S.-Soviet-Chinese relations: Routine, reciprocity,or rational expectations?American
Political Science Review 85:17-36.
Goldstein, J. S., and J. C. Pevehouse. 1997. Reciprocity,bullying, and internationalcooperation:Time-
series analysis of the Bosnia conflict. AmericanPolitical Science Review91:515-29.
Granger,C. W. J. 1980. Long memory relationshipsand the aggregationof dynamic models. Journal of
Econometrics 14:227-38.
Granger,C. W. J., and P. Newbold. 1974. Spuriousregressionin econometrics.Journal of Econometrics
2:111-20.
Hensel,P.R. 2001. Contentiousissues andworldpolitics:The managementof territorialclaims in the Amer-
icas, 1816-1992. InternationalStudies Quarterly45:81-109.
Keohane,R. 0. 1986. Reciprocityin internationalrelations.InternationalOrganization40:1-27.
Kwiatkowski,D., P.C. B. Phillips, P. Schmidt,andY. Shin. 1992. Testingthe null hypothesisof stationarity
against the alternativeof a unit root. Journal of Econometrics54:159-78.
32 JOURNALOF CONFLICTRESOLUTION

Lebo, M. J. 2000. Fractionalintegrationand political modeling. Unpublishedmanuscript.

Lebo, M., andH. Clarke.2000. Modelling memoryandvolatility:Recent advancesin the analysisof politi-
cal time series. Electoral Studies 19:1-8.
Lebo, M. J., R. W. Walker,and H. D. Clarke.2000. Youmust rememberthis: Dealing with long memoryin
political analyses. Electoral Studies 19:31-48.
Pevehouse,J., andJ. S. Goldstein. 1999. Serbiancomplianceor defiancein Kosovo?Statisticalanalysis and
real-timepredictions.Journal of ConflictResolution43:538-46.
Rajmaira,S. 1997. Indo-Pakistanirelations:Reciprocity in long-termperspective.InternationalStudies
Quarterly41:547-60.
Rajmaira,S., and M. D. Ward.1990. Evolving foreign policy norms:Reciprocityin the superpowertriad.
InternationalStudies Quarterly34:457-75.
Sandberg,I. W. 1978. On the mathematicaltheory of social processes characterizedby weak reciprocity.
Journal of Peace Science 3:1-30.
Vasquez,J. A. 1993. The war puzzle. New York:CambridgeUniversityPress.
Vasquez,J. A., and R. W. Mansbach. 1984. The role of issues in global cooperationand conflict. British
Journalof Political Science 14:411-33.
Ward,M. D. 1982. Cooperationand conflict in foreign policy behavior.InternationalStudies Quarterly
26:87-126.
Ward,M. D., andS. Rajmaira.1992. Reciprocityandnormsin UnitedStates-Sovietforeignpolicy.Journal
of ConflictResolution36:342-68.
Yule,G. U. 1926. Why do we sometimesget nonsensecorrelationsbetweentime-series?A studyin the sam-
pling and the natureof time series (with discussion). Journalof the Royal StatisticalSociety 89:1-64.