Multiregression Dynamic Models Author(s): Catriona M. Queen and Jim Q. Smith Source: Journal of the Royal Statistical Society. Series B (Methodological), Vol. 55, No. 4 (1993), pp. 849-870 Published by: Wiley for the Royal Statistical Society Stable URL: http://www.jstor.org/stable/2345998 . Accessed: 22/10/2013 11:53 Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp . JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected]. . Wiley and Royal Statistical Society are collaborating with JSTOR to digitize, preserve and extend access to Journal of the Royal Statistical Society. Series B (Methodological). http://www.jstor.org This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions J. R. Statist.Soc. B (1993) 55, No. 4, pp. 849-870 Multiregression DynamicModels By CATRIONA M. QUEENt and University of Nottingham,UK JIM Q. SMITH University of Warwick,Coventry, UK [ReceivedJuly1990. Final revisionJune19921 SUMMARY Multiregression dynamicmodels are definedto preservecertainconditionalindependence structures overtimeacross a multivariate timeseries.Theyare non-Gaussianand yetthey can oftenbe updatedin closedform.The firsttwomomentsof theirone-step-aheadforecast distributioncan be easily calculated. Furthermore, theycan be built to contain all the featuresof theunivariatedynamiclinearmodeland promisemoreefficient identification of causal structures in a timeseriesthanhas been possiblein thepast. Keywords: DYNAMIC LINEAR MODELS; GRANGER CAUSALITY; INFLUENCE DIAGRAMS; MULTIPROCESS MODELS; NON-LINEAR TIME SERIES; RECURSIVE SIMULTANEOUS EQUATION MODELS 1. INTRODUCTION Thispaperintroduces a classof multivariate statespacetimeseriesmodels,called multiregression dynamicmodels(MDMs), definedon a multivariate timeseries {Yt t>1.Thesemodelshavebeendevelopedtoforecast timeserieswhosecomponents arehypothesized to exhibit certainconditional in anyonetimeframe independences andthereis believedtobe a causaldriving mechanism within thesystem. To illustrate thetypeof timeserieswhichmayhavetheseproperties, Section5.1 considers the problemof forecasting themarketshareand relativemarketpriceof a particular brandina productmarket.Thecausaldrivethrough thisparticular is clearly system seenas thebrand'srelative pricetendsto dictateitsmarket share. MDMs definea class of non-Gaussian timeseriesmodelswhichdecomposethe forecasting systemintocomponents whoseconditional distributions are univariate Bayesiandynamic linearregression models(DLMs) (HarrisonandStevens,1976).As in simultaneous equationmodels,thecomponents ofthevectorYtareregressed on othercontemporaneous of thatvector.Sincetheforecasting components modelis expressedin termsof univariateDLMs, facilities availableforthe modellingof univariateseries,such as intervention, trendsand seasonals,can be transferred on to thesemodelswithno appealto stringent directly symmetry conditions being necessary(as in Harvey(1986) and Quintanaand West (1987)). Althoughthe modelledprocessescan be highly non-linear theyareofteneasilyupdatedin closed form.Thismakestheprocessespecially becausethemodelsareamenable interesting to analyticalinvestigation. or numerical Approximate methods,whichare often requiredfornon-Gaussian series(forexample,Kitagawa(1987)and Pole and West (1990)),withall therobustness issuesthatsurround them,arelargely unnecessary. Thepaperis organizedas follows.TheMDM is firstsetup in Section2 andthen tAddress for correspondence:Mathematics Department, Universityof Nottingham,.University Park, Nottingham,NG7 2RD, UK. ? 1993 RoyalStatisticalSociety 0035-9246/93/55849 This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions 850 QUEEN AND SMITH [No.4, defined inSection3. Resultspresented inthelatter sectionareformally provedin fully theproperties oftwointeresting Section4. Section5 examines specialcasesofMDMs and it is shownthat,despitebeingnon-linear, thefirsttwomoments of theirjoint can be calculatedalgebraically in closedform.Section5.1 forecastdistributions howanMDM hasbeenusedtoforecast illustrates a brand'spricerelative totherestof shareandtheformofthenon-linearity withitsmarket themarket together thatthey a discussion isstudied.Finally,Section6 offers ofsomeoftheproperties exhibit ofan MDM. It is shownhow an MDM, represented by a givengraphof an influence a specific conditional diagram (defined later),exhibits 'Granger causal'structure. Itis howsuchcausalstructures canbe testedbyusingmultiprocess illustrated modelling techniques. Throughoutthe paper the followingresultsfromthe theoryon conditional andgraphswillbe used.Anyrandomvectors independence X, Y, Z andW willhave conditional thefollowing independence properties: (a) X lI Y I (Y, Z), (b) XlIYIZo* YlIXIZ, X IIYI(Z,W) with (c) XI1(Y, Z) IW X* together XIIzIW ofY givenZ (seeDawid(1979)). whereX 1IY IZ readsX is independent Let (X1, . . ., Xm)be an orderedsetof randomvectorson which*1 1 *is defined. m- 1 conditional statements aregiven: Supposethatthefollowing independence XrliP'(Xr) I P(Xr) 2 < r< m (1.1) whereP(Xr) C {XI, . . ., Xr I} and P'(Xr) = {Xl,. . ., Xr I} \ P(Xr) where, forany in B whicharenotin A. A setof twosetsA and B, B\ A denotesthoseelements such as thesecan be usefullyrepresented statements conditionalindependence A graphG(X, E) consistsof a set of nodesX and edgesE, wherea graphically. directed edgefromXj to Xk is denotedby (Xj, Xk) e E. A directed graphwitha directed edgefromXitoXrifandonlyifXie P(Xr) iscalledthegraphofan influence diagram-seeHowardandMatheson(1981),Shachter (1986)andSmith(1989,1990). m fromXito Xk is a ThesetP(Xr) is knownas theparentset ofXr. A path oflength sequenceof nodes,Xi = Xi,0,. . ., Xi m = Xk, where(Xi,(j-1), Xi j) e E foreachj = 1, withpathsto Xk is calledtheancestorset of Xk and is m. The setof vertices withtheconditional denotedbyan(Xk). Thegraphofan influence diagramtogether statements independence (1.1) is calledan influence diagram. 2. SETTING UP THE MODEL howtheheuristic causalrelationships Thissectiondescribes betweencomponent ofa multivariate timeseries{Yk }k < t candefinean appropriate processes forecasting MDM. modelwhichformsthebasisofthegraphical in recentyears. Causationbetween processeshas beenstudiedquiteextensively of causalityrequired Granger's (1969)originaldefinition that,forall timet, a best linearestimateof V,based on { Uk} k<t and { Vk}k<t need onlydependon { Vk}k< t. A of Grangercausalityis givenby Lauritzen(1989). Here, graphicalrepresentation This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions 1993] MULTIREGRESSIONDYNAMIC MODELS 851 definition of Florensand Mouchart(1985)is used-theysay thestronger however, thatU is a non-causeof V,ifforall timet Vt iI (Uk}k< { tVk}k < t Thus the originalidea of Grangeris thenreplacedwitha definition based on itselfandnowthebestestimate conditional independence thanthebestlinear (rather estimate)of Vtbased on { Uk }k I tand { Vk}k < t needonlydependon { Vk)k < t. Byusing a setofconditional thisdefinition statements relatedtocausality independence canbe definedacross processes.Althoughcausalitydefinedin termsof conditional independence does notquitecapturewhatis usuallymeantbycausality(Holland, 1986),non-causality ofXand Ycan oftenreasonably be considered as a consequence oftherebeinga lackofcausalrelationfromXto Y. Wermuth andLauritzen (1990) heuristically arguedthat,whendealingwithgraphicalmodels,variableswhichare to be causallylinkedshouldbe connected hypothesized bydirected edgesconsistent of causality.Thusa heuristic withthedirection influence diagramcan be created theconditional representing independence relatedtocausality between thevariables at a fixedtimeperiodt. Forexample,consider threebrands{X, Y,Z} ina business so thatthesales market ofbrandXcan be considered as a causalfactorindetermining Y's salesandthesales ofbrandY can be considered as a causalfactorin determining Z's sales.LetXt, Yt thesalesat timet ofbrandsX, Y and Z respectively. andZt represent Theheuristic withthecausalrelationships graphoftheinfluence diagramconsistent inthismarket is glvenin Fig. 1. Supposethatthesameinfluence thevariablesatalltime diagramalwaysrepresents points so that P{ Yt} = Xt and P{Zt} = Yt for t a 1. Suppose furtherthat the processesoverall timepointsup to and including timet can be represented bythe graphoftheinfluence diagramofFig. 2 suchthat C {xt-l}, P{Xt} P{yt} P{Zt} CZ IXtAt-1), C {Yt,zt-'} thecausal relationships betweenthesales Fig. 1. Heuristicgraphof theinfluencediagramrepresenting at timet of brandsX, Y and Z Fig. 2. {Xk}k the causal relationshipsbetweenthe processes Graph of the influencediagramrepresenting t, { Yk}kS< and {Zk}k< t This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions [No. 4, QUEEN AND SMITH 852 where Wt = { W1, . . ., W'I}T for any time series { Wk }k , 1 such that AT denotes the transpose of any matrixA. It is clear from Fig. 2 that the conditional independences related to causality are such that YtH {Zk}k<tI {Yk}k<t, {Xk}k?t. This leads naturally to a general definition of conditional causality which extends Florens and Mouchart's (1985) definitionso that Uis a non-cause of Vgiven Wif for all points in time t Vtl {Vk}k<t, {Uk}k<tI {Wk}k6t. A best estimate of Vtnot only now depends on the past series { Vk }k < tand { W }k<t but also on the currentvalue Wt. So, in particular, Fig. 2 implies that Z is a non-cause of Y givenX. Thus thebestestimateof Yt based on {Xk }k t, { Yk }k <t and {Zk }k t and {Xk }k < t. An appropriate forecastingmodel for need only depend on {Yk}k<t each variable at time t is simply a functionof its own past series, the past series of its parents and the value of its parents at time t. For example, the general form of an appropriate forecastingmodel for the series { Yk}k t iSgiven by Yt =f(xt yt-l ot) + Vt, wheref( ) is some function and vthas some probability distribution.This reasoning can be generalized directlyto the case where there are n series. The MDM uses this methodology for deriving its observation equations and through the special form of the system equation breaks a complex multivariate problem into n univariate problems. Notice how the causal relationships between the variables are accommodated and that the stringentsymmetryconditions imposed on the variables in the models of Harvey (1986) and Quintana and West (1987) are not required. The MDM will now be formallydefined. 3. MULTIREGRESSIONDYNAMIC MODEL Let yT = { Yt(1), Yt(2), . ., Yt(n)} denote the general n-dimensional multivariate timeseries at time t. The observed value of Yt(r) is denoted byyt(r); let yt(r)T= { yl(r), .,y(r)} and set Xt (r)T= Zt(r)T {Yt(1), Yt(2),. . ., I + 1), ... . {Yt(r Yt(r- 1)} Yt(n) 2 < r< n, 2 < r < n- 1. Suppose that for any fixed time t the structureof the series Y, can be heuristically representedby an influence diagram I such that Pt YA61) ': Xt(r). Suppose furtherthat the process influence diagram I* such that P{Yt(r), Let Et = {t(j)T9 ot(2)T, {Yk }k t can be heuristically represented by an CZ{Xt(r)bYtht(r)et Ot. f(n)T } be the state vectors determining the This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions 1993] 853 MULTIREGRESSIONDYNAMIC MODELS and lets, be thedimensionofthe of Y,(1), Y,(2), . . ., Y,(n)respectively, distributions set > convenience notational For 1. t vector0,(r), t(r)T = {Ot(j)T,. t(r)T = t(r+ j)T9 2 < r< . ., Ot(r- 1)T} . 0.. t(n)T 2 < r < nX n-1. Call {Yt}t> I an MDM if it is governedby the followingn observationequations, equations andsystem wheretheobservation equationandinitialinformation system on theseequationsgivenbelowhold on thepastandtherestrictions areconditional forall pointsintime: (a) observation equations1< r vt(r) - (O, Vt(r)) Yt(r) = Ft(r)T Ot(r) + vt(r) < n; (3. 1) (b) system equationOt = (c) wt - (0, Wt); 1 + wt GtOtO (3.2) initial information(00 IDo) - (mo. co). (3.3) but known columnvectorFt(r)is allowedto be an arbitrary The sr-dimensional functionof xt(r)and yt- 1(r),but not zt(r)and yt(r); Vt(1),. . ., Vt(n)are theknown scalar observationvariances;thes x s matricesGt = blockdiag{Gt(1), . . ., Gt(n)}, Wt = blockdiag{Wt(1), . . ., Wt(n)} and CO = blockdiag{CO(1),. . ., CO(n)} are assumedknownand are suchthatGt(r), Wt(r)and CO(r) ares, x sr squarematrices of pastvectorsxt-1(r)and yt-1(r)butnothingelse. Thus whichmaybe functions {Yt(r) Vt(r). Iy 1, withmeanFt(r)T Ot(r)and variance Ft(r),Ot(r)}followssomedistribution The errorvectorsvT = {vt(l), . . ., vt(n)} and wT = {wt(1)T, . . ., wt(n)T}, where errorvectorfor system errorand wt(r)is thesr-dimensional vt(r)is theobservation . and . . ., variables 1 that r such are wt(1), ., wt(n)areall vt(n) vt(l), < n, Yt(r), < with independent mutually are vectors the and mutuallyindependent {vt,wt}t,> time. the fora vectortimeseries,consider howan MDM wouldbe defined To illustrate time fixed t, at any where, . = { that YT ., Yt(1),. followingexample.Suppose Yt(4)} MDM The 3. in influence the of the diagram given Fig. Ytcanberepresented by graph forYtwouldhaveobservationequationsinwhichFt(1) is a functionofyt-1(l); Ft(2) is a function of {yt(l), yt-1(2)}; Ft(3) is a function of {yt(l), yt(2), yt-1(3)}; Ft(4) is a functionof {yt(2), yt(3), yt-'(4)}. X 2I Fig. 3. Graphof theinfluencediagramfor Yt(1),..., Y(4) This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions [No.4, Thereare twoimportant resultswhichare centralto thetheoryof MDMs, the will be of which giveninthenextsection.It is shownthatif proofs 854 QUEEN AND SMITH ,1= I et- l(r) IYt- musthold. thenthefollowing conditional independence statements Result 1. ,1=1 Iet(r) I Yt. (3.4) Inwordsthismeansthatif{Ot-1(r)}aremutually independent giventhedatayt- then that, undertheDLM {Ot(r)}arealsoindependent givenYt.Itwillfollowbyinduction that{00(r)} areinitially independent for provided independent, theparameters remain availableinformation. all timegiventhecurrent Result2. Ot(r) l zt(r)I xt(r), yt(r). (3.5) intermsofthecomponents ofytthisreads Written Ot(r)liyt(r+ 1). yt(r+ 2), ... .,yt(n)Iyt(I)q. . .,9yt(r). In wordsthismeansthat,giventhepastobservation vectorsof thefirstr indexed oftherestofthepastdata. series,Ot(r)is independent theMDM, COis setto be blockdiagonalandso theparameters Whendefining for each variableare initiallymutuallyindependent. Therefore,by result1, the associatedwitheach variableare updatedindependently aftereach parameters and remainindependent at each timepoint.Thus, as each variable observation univariate followsa conditional distribuBayesiandynamic model,theconditional andconditional tionforeachvariablecanbe updatedindependently forecasts canbe suchas intervention, foundseparately. trendsand Bayesianforecasting techniques on to thesemodels.A complexmultivariate transferred seasonalscan be directly beendecomposed inton univariate problemhas therefore problems. of normality has beenmadeforeitherresults1 and 2 or when No assumption theMDM. Itfollowsthatlargeclassesofmultivariate suchas those defining processes ofthedecomposition. discussedbyKitagawa(1987)can be usedas thecomponents function of {xt(r),yt-'(r)},becauseit is However,evenwhenFt(r)is a non-linear fixedattimet,a normalMDMcanbedefined suchthat assumedknownandtherefore vtand wtare Gaussian so that,conditionalon xt(r), Yt(r) is Gaussian. The MDM is particularly simpleto workwithinthiscase. Each variablefollowsa normalDLM are distributions and, as such,updatingand forecastsof conditionalunivariate identicalwiththosegivenbyHarrisonand Stevens(1976).Section5 discussestwo dynamicmodel specialcases of normalMDMs calledthe linearmultiregression (LMDM) and the correctedlinearmultiregression dynamicmodel(CLMDM) in which Ft(r) is a linear functionof {xt(r), yt- '(r)}. In reality, theseriesareobservedsimultaneously so thatxt(r)willnotbe observed beforea forecastfor Yt(r)is made. The marginalforecastdistributions foreach variableare therefore willnotbe required.In generalthesemarginaldistributions ofYtformanyMDMs canbe derived Gaussian.However,themoments fairly easily. ofthefirst twomoments ofYtfortheCLMDM andthefirst moment Thederivation in Section5. oftheLMDM areillustrated This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions 1993] MULTIREGRESSION 855 DYNAMIC MODELS Results1 and2 combinetoenablethejointone-step-ahead of forecast distribution Considertheforecast distribution forn brands: Y, to be simplified. P(Yt IYt 1= | p{ytIt,yt1}p{tIyt1} dOet. statements of result1, together The conditionalindependence withthe structure of thevectorFt(r)specified bytheMDM, ensurethatthejointdistribution can be of yt(r)with expressedas the productof the individualforecastdistributions regressors xt(r).So P(ytIyt-1) = r = r p { yt(r)Ixt(r),yt-(r)} p I yt(r)Ixt(r),yt-l(r),0t(r)}p 0t(r)| yt-1 } dOt(r). dt,(,) statements of result2 and thespecified However,bytheconditional independence of Gt(r),p(Ot(r)Iyt- 1) onlydependson xt-1(r)andyt- 1(r)andthuscan be structure as rewritten p{Ot(r)Iyt1} =p1{t(r) Xt-(r),yt-l(r)} Thenextsectionformally provesresults1 and2. 4. FORMAL PRESENTATIONOF RESULTS To provetheresultsofthelastsection,thefollowing theorem givenbyPearland Verma(1987), Pearl (1988) and in its presentformby Lauritzenet al. (1990) is all the conditionalindependences required.It identifies definedby an influence diagramand gives an algorithmfor decidingwhetherany givenconditional can be logicallydeducedfroman influence independences diagram. Theorem1. Supposethatconditionalindependence statements for a set of inan influence variablesarerepresented whosegraphisI, andthatU, Vand diagram, Wdenotesetsofvariableson it.Adapttheinfluence diagraminthefollowing way. (a) Formthe directedsubgraphI, of I whosenodes are in the ancestorset an(U, V, W) andwhosedirected edgesarethoseinI whichliebetween these nodes. (b) Joinallpairsofnodes(X, Y) e P(Z) byan undirected edge,whereP(Z) isthe thegraphsince parentsetofZ andZ e II. Thisprocessis knownas moralizing all parentsofa singlevariablearejoinedbyan arc.Call thismixedgraphI2. all directed (c) Forman undirected graphJbyreplacing edgesinI2 byundirected edges. Then UllVIw ifall undirected a nodeR e U andS e Vmustpassthrough pathsinJbetween a node W. Te This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions [No.4, if thisconditionis violated,thenthereis a probabilistic Furthermore, influence diagram theconditional respecting independence statements oftheinfluence diagram forwhichUIl VI Wis nottrue. Theorem1 can nowbe directly usedto provethefollowing theorem. 2. Let {Yt}t, I be governed Theorem byan MDM and,usingthenotationofthe assume that previoussection, Ot- i(r) llyt-l(r), zt- l(r) Ixt- l(r) r=2, . . n. (4.1) 856 QUEEN AND SMITH t-l(r) Ixt-l(r), yt-1 et-l(r)zt-l(r), Ot-i(r)ot-i(r)Ot- l(r)Iy t-I r= 1,. . . n, r=2, . . n- 1. (4.2) (4.3) Thenthefollowing conditional independence statements mustalso be true: Ot (r) Ot(r) r=2, ........ ,n, ........ r=l 1......... ....... ,n, 1y t(r) .z t(r) Ix t(r) 1z t(r) . t(r) Ix t(r).y t(r) t Ot(r) 1 t(r) . t(r) IY ,n- 1. r=2, ............. . (4.4) (4.5) (4.6) Proof. To provethisresult, itis first thattheconditional necessary independence intheinductive statements contained hypotheses (4.1)-(4.3)andtheMDM itselfare represented 1itis thensimpletoshowthatconditional graphically. Byusingtheorem statements independence must also hold. (4.4)-(4.6) For clarityof presentation, theinductive hypotheses (4.1), (4.2) and (4.3) are represented arcsintheseparate bythedirected graphsofinfluence ofFigs4, diagrams 5 and 6 respectively. In each diagram,variableshavebeenallowedto sharenodes whenit is convenient to do so and whenthereis no loss ofinformation. Thus,for example, in Fig. 4 zt- l(r- 1) = {yt- l(r), Zt- 1(r)} and At-1(r- 1) = {Ot-l(r), At-1(r)} Thejustification forthemissing directed arcsinthesegraphsofinfluence diagrams willnowbe presented all undirected (forthemoment arcsareignored). B I ( 1)r I_ _V _( _ _ Fig. 4. The directedarcsof thisgraphgivethegraphof theinfluencediagramforinductivehypothesis (4.1); thedirectedand undirectedarcs make up themoralizedgraph This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions 1993] MULTIREGRESSION DYNAMIC MODELS 857 B 4~~~(r) I - - -4-(--) - A~~~~~0()0r arcsofthisgraphgivethegraphoftheinfluence Fig. 5. Thedirected forinductive diagram hypothesis andundirected arcsmakeup themoralized (4.2); thedirected graph B A IC arcsofthisgraphgivethegraphoftheinfluence Fig. 6. Thedirected forinductive diagram hypothesis andundirected (4.3); thedirected arcsmakeup themoralized graph This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions [No.4, statedbythe (4.2) canbe equivalently (c) ofSection1,assumption Usingproperty pairof assertions 858 QUEEN AND SMITH i(r)11Ot_i(r) IY- l,.7 Ot_ et- l(r) (4.8) zZt- l(r) Ixt- l(r), yt- 1(r). yt- 1 nodesofsubvectors (4.8), arcsbetween (4.1) andassertion Now,byassumption of thanrtothosenodesofsubvectors allhaveanindexinYtgreater whosecomponents In haveindiceslessthanorequaltorareallowedtobe omitted. Otwhosecomponents vectors parameter (4.3) allowarcsbetween (4.7) andassumption assertion contrast, to be omitted.Thisjustifiestheimplied whichdo notsharethesamecomponents intheboxeslabelledA in Figs4-6. statements independence conditional TheblockdiagonalformsofGtand WtintheMDM ensurethat Ot(r) 1et- I \ {t- l(r)} IYt- 19 et- 1(r) 1 < r < n. of setsofcomponents diagramanyarcsbetween Thuson thegraphoftheinfluence This be omitted. can index common no with of of sets and Ot etcomponents diagramboxes. theomissionof arcsinbox B oftheinfluence justifies argument implicitin the observation statements Finallythe conditionalindependence equationsoftheMDM meanthat Mtr) l et\ fot(r)) IYt- l(r), Xt(r),Ot(r)- diagram justifytheomissionof arcsin box C of theinfluence Thesestatements graphs. wheneach arcsinFigs4-6 arethoseextraarcswhichareintroduced Theundirected Itisnowsimpletocheckthatthesetofnodes inmoralized. diagrams oftheinfluence thenodesassociated variablesblockall pathsbetween theconditioning representing 1the So bytheorem setsUand Vclaimedtobeindependent. indifferent withelements D resultis proved. statedin thefollowing in Section3 can nowbe formally The resultsintroduced 2. oftheorem corollary i.e. if11fl 100(r),then 1. Ifinan MDM theinitialstatesareindependent, Corollary forall timest l1r= I Ot(r) I Yt and Ot(r) 11y t(r + 1),9. . . , yt(n) Iyt(I)q yt(r). 2. Fromthehypotheses Proof. To provetheresultfort= 1proceedas intheorem equations(setC) thegraphof equation(setB) andtheobservation (setA), thesystem arcsofthegraphin Fig. 7. bythedirected diagramI is represented theinfluence arcsof 2, again,theundirected as intheorem thesameargument Byusingexactly whenthegraphis moralized,and so theconditional Fig. 7 are thoseintroduced 2 cannowbe deducedfortimet= 1. statements (4.4)-(4.6)oftheorem independence 2 mustbetruefort= 1,andso theorem sinceHrLI 00(t)underthecorollary, Therefore, 2 mustbe trueforall timest. oftheorem theassertions byinduction from statements independence 2 holdsforalltimest,theconditional Sincetheorem can be combined: thetheorem This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions 19931 MULTIREGRESSIONDYNAMIC MODELS B C~( ) ID~~~~~ 859 I E()j1 arcsof thisgraphgivethegraphsof theinfluence diagramfortheproofof Fig. 7. The directed 1; thedirectedand undirectedarcs make up themoralizedgraph corollary 1I{.ot(r),CO~()},zt(r)Ixt(r),yt(r). Ot(r) (4.9) Now, by property(c) (Section 1) statement4.9 impliesthat Ot(r)I ,I{t(r), ,t(r)} zt 1(r),xt i(r),,yt) (r); therefore, rn 1at(r Iyt. Statement(4.9) also impliesthat Ot(r)llzt(r) Ixt(r),yt(r) and so Ot(r) llyt(r + 1), . . . , yt(n) Iyt(I)q yt(r); the corollaryand hencetheresultsof Section3 are proved. therefore D Theseresultsapplyto a verywideclass of models.However,itis helpfulinitiallyto considertheimplicationsfortheLMDM and CLMDM becausethejoint distribution of the variablesin thesemodelscan be calculatedexplicitlyand the one-step-ahead forecastshave a particularlysimpleform. 5. LINEAR MULTIREGRESSIONDYNAMIC MODELS Considerthe MDM definedby equations(3.1)-(3.3). This sectionintroducestwo special cases of the MDM-the LMDM and the CLMDM-which are especially This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions 860 [No. 4, QUEEN AND SMITH simpleto workwith. of yt- 1 and Supposethat{v,(r),w,(r)}) I arejointlyGaussian,areindependent Ft(1)doesnotdependon Yt. Now,let 2 Ft(r)T = {xt(r)T,it(r)T}, wherext(r)T = { yt(1), . < r n . . yt(r- 1)} and it(r) is a setof knownexogenousvariables supposethat onxt(r).ThisprocessisanLMDM. Alternatively, thatarenotdependent Ft(r)T= [{xt(r) - -ft(r)}T,it(r)T], where (5.1) E{ Yt(r- 1) I yt- 1}] xt-1(r)andyt-1(r). and,fromresult2 ofSection3, ft(r)onlydependsonyt- 1through ft = [E Yt(l) Iyt- 1}, . . ., onxt(r). thatarenotdependent variables Onceagainit(r)isa setofknownexogenous Thisprocessis a CLMDM. Thus,giventhecomponents wt(r)whichprecedeit,each containedin the as a univariateDLM withregressors component Yt(r)is described above. given vectors Ft(r) simultaneous equations versionofa recursive NotethattheLMDM is a stochastic variableslisted againsta subsetof contemporary modelwhereYt(r)is regressed specialcase of this.Unlikein beforeYt(r).Harvey(1989) considersa degenerate buton the estimation theemphasishereis noton parameter economics,however, continued thatYtcanexhibit giventheinevitable jointdistribution typesofpredictive stateintheprocess. abouttheregression uncertainty between brandsmodelledbyan LMDM havea different Thecausalrelationships fromthosemodelledbya CLMDM. Forexample,supposethatthere interpretation to be a arejusttwovariables,{ Yk(1)) 1 and { Yk(2)}k,, I, suchthatY(1) is thought changein causal factorof Y(2). If {Yk}k I followsan LMDM, thena long-term Y(1)'s levelwouldcause a sustainedlevelchangein Y(2). However,if {Yk }k levelchangein Y(1) wouldhave a followsa CLMDM, thenthesamelong-term on Y(2). In thiscase thelevelchangein Y(2) wouldbe short causaleffect different term,followedbya driftbackto theoriginallevel.Thisis due to thefactthatthe unlikethe LMDM whichuses the actual CLMDM uses residualsas regressors observation.Thus the CLMDM essentiallyrelatescausalitythroughforecast of distribution theforecast of Yt(1)helpsinadjusting so thatmisforecasting residuals would in level a sudden Y(1) change true. not Therefore, the reverse is whereas Yt(2) leadto a largeresidualvalueinthemodelforY(2), thusleadingto a levelchangein intheregression Y(2). As themodelforY(1) adaptstothelevelchange,sotheresidual level. backtoitsoriginal and Y(2) willdrift termin Y(2)'s modelwillbecomesmaller aniceinwhicha CLMDM wouldbe a usefulmodel,consider a situation To illustrate and Yt(2)isthemarket creammarket. SupposethatYt(1)isthetotalsalesofice-cream shareofice-cream typeA. Nowiftherewerea heatwavethetotaldemandYt(1) oficeincrease.If brandA hadextrastockandcouldcopewiththe creammightsuddenly shareYt(2)wouldbe brandB, say,thenthemarket thananother extrademandbetter be modelledby a This situation could wave. the heat over increase to expected CLMDM suchthat Yt(2) = Ot+ 0t Yt(1)-E[Yt(1)Jyt-1]} + Et, This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions MULTIREGRESSIONDYNAMIC MODELS 19931 861 whereEtis someerror termand0, is someparameter associated withA's ability tocope withextrademand. For boththeLMDM and theCLMDM, theseparatecomponents { Yt(r)Ixt(r)} haveeithera Gaussianor Studentt-distribution depending on whether theforecast distribution is assumedknownorestimated. Theforecast distributions andupdating relationships associatedwiththeseconditional univariate modelsareidenticalwith thoseofWestandHarrison(1989).Itthenfollowsfromtheorem 2 thattheone-stepaheadforecast ofYtis simply density theproductoftheunivariate conditional onestep-aheadforecastdensitiesof { Yt(r)Ixt(r)}, forr= 1, . . ., n. Althoughthe joint forecast distribution ofYtwillnot,ingeneral, be Gaussian,itsmeanandcovariance matrixtakea relatively simpleformand thesewillnow be derivedhere.Assume thisderivation throughout thatall meansand variancesarefoundconditionally on xt(r). The meanofthemarginal forecast distribution forYtwhenitfollowsan LMDM willfirstly bederived. As hasalreadybeenmentioned, eachvariableundertheLMDM followsa univariateregression DLM. Usingthe forecastdistributions for the regression DLMs givenby Westand Harrison(1989),p. 111,it is clearthatthe oftheconditional expectation forecast distribution of{ Yt(r)Ixt(r)}giventhepastcan thenbe rewritten as = E{ Yt(l) Iyt- '(1)} a-(?)( E{ Yt(r) Iyt-l(r), xt(r)} = at(?(r) + I r-1 i E at()(r) Yt(i) i=l where at(0)(r)is a function of {it(r), yt- 1(r),xtt(r)} 2 < r< n (5.2) only. The marginalforecastmeansforYtgiventhepastcan thenbe easilycalculated fromtheseconditional meansbyusingtheidentity E{ Yt(r) Iyt-l(r), xt- 1(r)} = E[E{ Yt(r) Iyt-l(r), xt(r)}]. (5.3) Writea (O)T = {at(?(1), . . ., a(?)(n)} and A as then x n lowertriangularmatrixwhose { j, k)thelement ajkis givenby ajk= k <j, at(k)(j) (0 otherwise. Then,byusingidentity (5.3),equation(5.2)acrossallr= 1,..., n canbeexpressed by E(YtIyt-1) = a(?) + A E(YtIyt-1) and so E(Yt|yt-1) = (I-A)-1a"0. for the CLMDM, the expectation Similarly, of the forecastdistribution for { Yt(r)Ixt(r)}giventhepastcanbe rewritten as E{ Yt(1) yt- 1(1)} = - E{t Y(r) Iyt- 1(r),xt(r)} = ()(09 r-1 jt(O)(r)+ E at(i)(r){Yt(i) - ft(i)} t=1 This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions 2 < r< n (5.4) 862 [No. 4, QUEEN AND SMITH toa different valueoftheanalogousa(i)(r)ofequation(5.2) whereja(i)(r)corresponds and ft(r)is thesame as in equation (5.1). By usingidentity(5.3) it is obviousthatthe theCLMDM aresimply meansforYt(r)following marginal forecast E{ Y (r) Iyt- l(r), xt- l(r)} = jt()(r). of Ytgiventhepastis foundfromthe Although themarginal covariancematrix in boththeLMDM and theCLMDM, thecovariance samerecursive relationships formandso onlythederivation ofthismatrix oftheCLMDM takesa simpler matrix to findthevarianceofthe itis first willbe shownhere.To findthismatrix necessary distribution of Yt(r),forr = 1, . . ., n. forecast marginal = {IO*(r)T, isthesetofparameters contained 0E*(r) SupposethatOt(r)T Gt(r)T} where is thesetassociatedwithit(r). If Rt(r)is the in Ot(r)associatedwithxt(r)and #t(r) of {0t(r) Iyt- } thenRt(r)can be expressed covariancematrixof thepriordistribution by = Rt(r) R Rt*(r) Rt'(r)) \R'(r)T At r)! where cov{IO*(r),et*(r)Iyt- 1}, *(r) Rt(r) = cov{Ot(r), #t(r)y 1}, R'(r) = cov {IO*(r), t(r) yt- 1}. forecast distribution of { Yt(r)Ixt(r)}given thevarianceoftheconditional Therefore, as byWestand Harrison(1989),p. 111,can be rewritten var{ Yt(1) Iyt- var{ Yt(r) yt- l(r), xt(r)} 1(1)} = rt2(r)+ = 2(1), h {xt(r)} 2 Sr < n where h {xt(r)} = trace[St(r)T{xt(r) - ft(r)} {xt(r) - ft(r)}TSt(r)] whereeach scalar'r2(r) 00, r= 1,. . ., n, is a functionof {it(r), yt- l(r), R'(r), Vt(r)}and St(r)St(r)T= R *(r). xt- l(r), Rt(r), forecast distribution ofYt(r)giventhepast,it To findthevarianceofthemarginal holds: notedthatforanytwovariablesZ(1) andZ(2) thefollowing identity is first (5.6) var{Z(2)} = E[var{Z(2)I Z(1)}] + var[E{Z(2)I Z(1)}]. Suppose thatEt is theforecastcovariancematrixforYt suchthat {2t }jk= rt(j, k) = cov{ Yt(j), Yt(k) Iyt- 1} j, k= 1, . . ., n and letEt(r)be theforecastcovariancematrixof{ Yt(1),. . ., Yt(r- 1)} forr= 2,. . .. withequations(5.4) and (5.5) itcan be n. Therefore byusingidentity (5.6) together shownthat at(I, 1) = r2(1), ot(r, r) = E[var{ Yt(r) Iyt- l(r), xt(r)}] + var[E{ Yt(r) Iyt- l(r), xt(r)}] This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions MULTIREGRESSIONDYNAMIC MODELS 1993] 863 2 = 2(r) + trace{S,(r)TE,(r)S,(r)} + a (r) at(r) whereit(r)T = {d(l)(r).. 1)(r)} a(, r< n To findthemarginalcovariancebetweenYt(r)and Yt(k),notethatforthetwo variablesZ(1) and Z(2) E{Z(1)Z(2)} = E[E{Z(2)Z(1)|Z(2)}] Therefore supposethatZ(1) thatforr > 2 = Yt(r)andZ(2) ort(r)T= {ft(r, 1),. . ., = E[Z(2)E{Z(1)fZ(2)}]. = Xt(r) andletot(r)be a vectorsuch o(r, r-1)} = cov{Yt(r),Xt(r)Iyt} Then ot(r) = E[Xt(r) E{Yt(r) I Xt(r)} Iyt-11. Byequation(5.4) thisbecomes ot(r) = d(?)(r)E{Xt(r) Iyt-1} + r-1 Z d(i)(r)E[Xt(r){ Yt(i)- ft(i)]Iyt 1 i=l It is clearthat Et(r+ 1) ENt(r) at,(r)) (r) Tort(r, r) \rt So ifEt(2) = a(1, 1) is given,thenEt(r+ 1) canbe calculatedas a simplefunction of Et(r), ot(r) and ot(r, r). Hence themarginalforecastcovariancematrixoftheCLMDM andexpectations is simply calculatedfromthevariances oftheupdateddistributions fortheseparateconditional DLMs. component regression Thesemodelsarebestillustrated bya simpleexample. 5. 1. Example A simpleexamplewhichillustrates thisnewclassofmodelscanbe constructed as follows.Suppose thatyT = { Y,(1), Yt(2)} is to be modelledbyan LMDM whereYt(1) thepriceof a certainbrandrelative to therestof themarketand Yt(2) represents thebrand'smarket share.Supposethatat anyfixedtimethetwovariables represents inFig. 8(a). Supposefurther canberepresented bythegraphoftheinfluence diagram thattheprocesses{ Yk(l )}k < tand { Yk(2)}k < tcan be represented bythegraphofthe influence diagramofFig. 8(b). The generalLMDM forthissystemis givenby thefollowing observation and system equations: Yt(1) Yt- (2) (a) Fig. 8. Graphof theinfluencediagramsfor(a) Y,(1) and Y,(2) forfixedtimetand (b) { Yk(2)}k < t This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions < { Yk(l)}k kand 864 QUEEN AND SMITH equation(a) observation vt(l) - N(O, Vt(1)), Yt(1)= Ft(1)TOt(1) + vt(j), Yt(2) = Ft(2)TOt(2) + vt(2), equation(b) system I + wt, Ot= GtOtO (c) initialinformation- vt(2)- N(0, Vt(2)); wt - N(0, Wt); (0oIDo) - N(mo,CO) [No. 4, (5.7) (5.8) (5.9) (5.10) whereFt(1) = !t(1), Ft(2)T = {l,yt(l)},7 T = {Ot(l)T, 0(o)(2),0(l)(2)}, mT = {mO(l)T, m60)(2),m6l)(2)},Gtis thes, + 2 square identitymatrix(wheres, is the dimensionof Ot(l)}, Wtand COare block diagonal and vt(l), vt(2),wt(l) and wt(2) are mutually independentof each otherthroughtime. Using exactlythe same notationintroducedto definethemeansand variancesof theLMDM, and notingthatin thisexamplea(?)(2) = m(t)1(2)and a(l)(2) = m(l)1(2), to therequiredconditionalforecastmeansand equations(5.2) and (5.5) lead directly varianceswhichare givenby 1, yt(1)} = at()(2) + yt(1) at(1)(2) EYt (2)I yt- (5.11) and var{Yt(1)Iyt-1(1)} = t2(j), var{ Yt(2) I yt- 1,yt(l)} = yt(1)2R *(2) + r2(2) where'r2(2)is a functionof Rt(2), R'(2) and Vt(2)and in thiscase R *(2) is simplya scalar. are simply The firsttwo momentsof the forecastdistribution E[Y,Iy ] Y 1 / a(O)(1) (a,?)(2)+ a(O)(1)a(l)(2)) (5.13) and covariancematrix (2,9I) ort(2,92))(.4 ort when ot(1(,2) = rt(21) at(0)(2) E{Yt (1) yt-1} + at(1)(2)EtY(1)21yt-1} and 2) rt(2, = 'r7(2)+ R *(2)E{ Yt(1)2I yt- 1} + a(1)(2)2r2(N1). A timeseriesplot of { Yk(2)}k,O is givenin Fig. 9 togetherwithvariousone-stepahead forecastsof theseries.The timeseriesitselfis denotedbythedotson thegraph. The one-step-aheadconditionaland marginalforecastsof Y(2) derivedfromthe MDM are representedby the thin and thickcurvesrespectively.The conditional forecastsobtainedfromtheMDM are fairlyaccurate,indicatingthattherelativeprice ofthebrandappearsto be a good predictorofthebrand'smarketshare.The one-step- This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions 1993J MULTIREGRESSIONDYNAMIC MODELS 865 - 0.25 0.24 0.230.220.21 0.200.19 0.187 O. 19 0.176 0.16 1 2 3 4 5 6 7 8 9 10 11 Time (Years) Fig. 9. Time seriesplot of Y(2) withone-step-aheadforecasts:..., , one-steptimeseries; ahead forecastconditionalon relativeprice; -, one-step-aheadmarginalforecastderivedfromthe MDM; ---, forecasts from the steady model ahead forecastsof the simplesteadymodel for Y(2) are representedby the broken curvein Fig. 9. If these are compareddirectlywiththe marginalforecastsof the MDM, thenclearlythe forecastsobtainedfromtheMDM are farsuperiorto those obtainedfromthe steadymodel. This illustrateshow the forecastperformanceof a model for Y(2) can be improvedby using Y(1) as a regressorand eventhoughY(1) cannot be observed before Yt(2) the MDM can still use Y,(1) to obtain realistic forecastsof Yt(2). Clearly, both { Yt(1) Iyt- 1(1)} and { Yt(2) Iyt- 1, yt(1)} have normal distributions. However,thejointforecastdistribution of { Yt(1), Yt(2) 1yt 1} is notbivariatenormal because of theappearanceofyt(1)in thevariancetermof { Yt(2) yt y1, t(I)} . Indeed thisjoint distributioncan be verynon-normal.This is demonstratedin Fig. 10 in whichthecontoursof thejointdistribution of { Yt(1), Yt(2)} afterthemarginof Yt(1) has been normalizedand forvariousparametervalues are given. FromFig. 10itcan be shownthatthemodesand antimodeslie on a quintic(and so exhibita butterfly catastrophe;Zeeman(1977)). Thejointforecastdistribution is only whenthe varianceof { Yt(1) yt-} does not appear in the varianceof symmetrical {Yt(2) 1yt }, i.e. when m(1)1(2)= 0. For any non-degenerate values of the parameters,it can be concluded,aftera littlealgebra,thatthereis a value ofyt(2)suchthat theconditionalpredictivedensityof { Yt(1)IYt(2)=yt(2)} is bimodal. The worstcase occurswhenOt(2)is uncertainwherethedistribution becomesverynon-Gaussian.As - 0, thentheprocesstendsto a bivariatenormaldistribution. However,unlike Rt*(2) the analogous simultaneousequationsmodelsthe assumedstochasticdrifton Ot(2) preventsthislimitfromeverbeingreached. FromFig. 10 it is clearthatthepointE[Yt(1) Iy't 1(1)] (in thiscase Yt(1)= 0) takes on special significanceas it is thepointabout whichthecontoursare symmetrical or asymmetrical.Regressionon the forecastresidual [Yt(1) - E{ Yt(1)Iyt-1(1)}] will oftenthereforeseem more natural. This gives the correspondingCLMDM whose meansand covariancematrixare givenby This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions Y2 [No. 4, QUEEN AND SMITH 866 II 13 2 V2 10~~~~~~~~~~~~S1 -U~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~t S to -LU ~~~~~~~~~~~-to -10 -IS~~~~~~~~~~~~~R -4 -3 -U -L O ? -to U 4 3 -3 -4 -U -to 4 -a Y2 Y2 -U -3 U L o -, S Y2 / a -tOS Lo -to -US -U0S -Uso -a.a -a -l. -I -0.13 5 , 20 .5 L2L.U IUS-.s I5 1L .I U-. I-0.5 a.a.1 U U.U no a. a5- -S-Ur1 ha-IS-L ennraiedt0aeOma -2 a Yt VII 5 2 an unto vaiac 2 and mt ( 4 U L a 3 Vi -lU aa-I -5.5 i.a a 5 . yi Yi *0 O -LsU ,()= 2 0i h is -US -LU -ISo -S a -U.S-U U.S -4 -3 i1 2 3 YL -3 -U I S 3 VI LU tl,2 I~~~~~S t(1 of Y,(1) of and the when contour Various 10. marginaldistribution plots Y,(1) asymmetric Fig. Yt(2) has been normalizedto have 0 mean and unitvarianceand m- ~(2) * 0; Rt*(2) = T2;(2) = 10 in thefirst row,R *(2) = 10, -r?(2)= 1 in thesecondrowand R *(2) - 1 r(2) = lO0inthethirdrow; mt- (2) variesby column,takingthe values 1 in thefirstcolumn,5 in thesecondand 10 in thethird IS E[Y~Iyt~'I (a)(1)) = = where lt(2boat2,2)) This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions 1993] MULTIREGRESSION 867 DYNAMIC MODELS rt(2,1) at(l)(2)E[{ Yt(1)- ft(1)}2I yt rt(2,2) = rt2(2)+ R *(2) E[{ Yt(1) t(1)}21yt-I + a(l)(2)2 var[{ rt(1,2) = = ], Yt(1) ft(1)} Iyt-I rt2(2) + rt2(1){R *(2)+a (1)(2)2}. The graphsin Fig. 10 willstillbe appropriateforthiscorrectedmodel. 6. DISCUSSION OF MULTIREGRESSIONDYNAMIC MODELS Some of the interesting theoreticalaspects of MDMs will be discussed in this section. MDMs definea class of non-Gaussiantimeseriesmodels whichdecompose the forecastingsysteminto componentswhose conditionaldistributions are univariate Bayesian dynamicregressionmodels. As was mentionedin Sections3 and 5, these univariatemodelscan be normal.In particular,ifFt(r)is a linearfunctionof {xt(r), yt- l(r)}, theneach variablewould followa generalizationof one of Priestley'sstatedependentmodels(Priestley,1980). Alternatively, each modelcould be a non-linear dynamicmodel (West et al., 1985; Migon and Harrison,1985; Pole et al., 1988). If each componentfollowsa univariateBayesianlinearor non-linearmodel,thenit is possibleto use theupdatingrelationships directlyfromunivariateBayesiandynamic models,eventhoughthemodelregresseson contemporary componentsof Ytand can be highlynon-linear.This maketheprocessespeciallyinteresting because themodels are amenableto analyticalinvestigation. Approximateor numericalmethods,withall therobustnessissuesthatsurroundthem,are largelyunnecesary.However,as can be seen fromFig. 10, thevectorYt can have a joint forecastdistribution whichis very non-Gaussian, even when Ft is a linear function of (Yt(1), . . ., Yt(r- 1)). In this respectthisclass of modelsis verysimilarto themodelsof Wermuthand Lauritzen (1990) in the factthat the conditionaldistributions are fairlysimplebut the joint model is farmorecomplex. It is interestingthat, unlike conditional independencewithin time frames, conditionalcausality of the type describedhere depends on the orderingof the variablesin theMDM. Furthermore each influencediagramcorrespondsto a unique LMDM or CLMDM structure. For example, suppose that for five variables modelled with an LMDM the followingregressionequationshold: YM(I) = O(O)(1)+ vM(1); Yt(2) = a(?)(2) + 0(l)(2)yt(1) + vt(2); Yt(3) = O(o)(3)+ O(l)(3)yt(1)+ O(2) (3)yt(2)+ vt(3); (6.1) Yt(4) = 0O()(4) + O (3)(4)yt(3) + vt(4); Yt(5) = O()(5) + 6 (3)(5)yt(3) + vt(5). The graphof theinfluencediagramconsistentwiththeseequationsis givenbygraph (a) in Fig. 11. It is easyto checkthattheLMDM on theseregression equationscould alternatively be representedby the LMDM on componentsof Y takenin the order{ Y(1), Y(2), This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions 868 [No. 4, QUEEN AND SMITH (a) (b) Fig. 11. Two graphswiththesame impliedconditionalindependences,but withdifferent LMDMs Y(3), Y(5), Y(4)}. The graph of the influencediagram,Fig. 11(a), would remain unchanged. In generalany new orderingof thevariablesthatis compatiblewiththeinfluence diagramof a singletimeframeof an LMDM or CLMDM, giventhepast,willproduce equationsthatare algebraicallyidenticalwiththeoriginal. Howeverbythedecompositiontheorem(Smith,1989),foranygiventimeframet, theimpliedsetof conditionalindependencestatements bygraph(b) in Fig. 11, given the past, are identical with those representedby graph (a), but the conditional independencesrelated to causalityin the two diagramscorrespondto two quite different LMDMs. For example,fromgraph(a) it is clearthat Yt(2) Iyt 1 = Yt(2) Iyt1(2), yt-1(1) whereasgraph(b) meansthat Yt(2) I yt- = Yt(2) I y"-(2), yt-'(I), yt-1(3). Because ofthedifferent covariancestructure on thesystemerrorwtimpliedbyeach of theseinfluencediagrams,thereis no guaranteethatthesetwo statements could hold simultaneously. Suppose that the contextof the model is such that the time frameconditional independencesare logicallydetermined.However,suppose thatthereis uncertainty thecausal structure regarding acrossthevariablesintheproblem,althoughthiscausal structure is assumedconsistentovertime.In thiscases a class I multiprocess MDM (Harrisonand Stevens,1976) can modelthesystem.A multiprocess MDM considers m models {M(1), . . ., M(m)} simultaneouslywherethereis one model foreach of the m possible causal structuresfor the given conditionalindependencestructure.They allow thepredictionof complexserieswithoutanya prioriassumptionsabout causal Thisis becausetheforecastsare foundbyusingp(YtIyt- 1,M(i)) mixedwith structures. probabilitiesp(M(') Iyt-) to give appropriatepredictivedensities. Multiprocess modelscan also be combinedwithsome decisionrulewhichcomparesthe forecast of thevariousmodelsso thatthemostappropriatecausal structure performance can be selectedfromthen alternatives. AlthoughGrangercausalityhas attractedacademic interest,in practiceit has proveddifficultto discriminatebetweendifferent causal structures by usinglinear systems.However, by embeddingcausalityin the non-linearMDM, multiprocess MDMs can givean on-lineassessmentof hypothetical causal relationships acrossthe variables.The MDM corresponding to a givencausal structure can thenbe made at This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions 1993] MULTIREGRESSION DYNAMIC MODELS 869 the conditionalcomponentscan be as complicatedas least plausible. Furthermore, necessary,containingtrends,regressionterms,seasonal factors,and so on. It looks promisingthattheMDMs willenabletheselectionofcausal structures therefore across practicalmodels. Noticethatfromthegeometriesof thejoint densitieson themodel of Section5.1 most informationabout causal relationshipsseems to come whenthe relationship betweenthe variablesis uncertain(in thisexamplethiscorrespondsto R *(2) being large). This willoccur earlyin the seriesand afterexternalintervention. This might explain whyZellner(1984) findsit so difficultto discriminatebetweentwo causal structures-themodels thathe considersare not dynamic;theywould assume that is notconsidered. Rt*(2)- 0 and externalintervention 7. CONCLUSION MDMs have been developedto modelmultivariate timeserieswhosecomponents exhibitconditionalindependencesrelatedto causality.Althoughhighlynon-linear, the models are relativelysimpleto implement.Preliminaryuse of thesemodels on serieswitha small numberof componentsseems promising.Use of the MDM to model a largebusinessmarketis documentedin Queen (1992). ACKNOWLEDGEMENTS We wouldliketo thanktheScienceand Engineering ResearchCounciland Unilever Researchforthe financialsupportgivenwhileundertaking thisresearch. REFERENCES Dawid, A. P. (1979) Conditionalindependencein statisticaltheory(withdiscussion).J. R. Statist.Soc. B, 41, 1-31. Florens,J. P. and Mouchart,M. (1985) A lineartheoryfornon-causality.Econometrica,53, 157-175. Granger,C. W. J. (1969) Investigatingcausal relationsby econometricmodels and cross-spectral methods.Econometrica,37, 424-438. Harrison,P. J.and Stevens,C. F. (1976) Bayesianforecasting (withdiscussion).J.R. Statist.Soc. B, 38, 205-217. Harvey, A. C. (1986) Analysis and generalisationof a multivariateexponentialsmoothingmodel. MangmntSci., 32, 374-380. (1989) ForecastingStructuralTimeSeriesModels and theKalmanFilter.Cambridge:Cambridge Press. University Holland, P. W. (1986) Statisticsand causal inference.J. Am. Statist.Ass., 81, 945-970. Howard, R. A. and Matheson, J. E. (1981) Influencediagrams.In Readings on the Principlesand Applicationsof Decision Analysis (eds R. A. Howard and J. E. Matheson), vol. II, pp. 719-762. Menlo Park: StrategicDecision Group. Kitagawa,G. (1987) Non-Gaussianstate-spacemodellingof nonstationary timeseries(withdiscussion). J. Am. Statist.Ass., 82, 1032-1063. Lauritzen,S. L. (1989) Mixed graphicalassociationmodels. Scand. J. Statist.,16, 273-306. Lauritzen,S. L., Dawid, A. P., Larsen, B. N. and Leimer,H. G. (1990) Independencepropertiesof directedMarkov fields.Networks,20, 491-505. Migon,H. S. and Harrison,P. J. (1985) An applicationof non-linearBayesianforecasting to television and advertising.In Bayesian Statistics2 (eds J. M. Bernardo,M. H. DeGroot, D. V. Lindleyand A. F. M. Smith).Amsterdam:North-Holland. Pearl, J. (1988) ProbabilisticReasoningin IntelligentSystems:Networksof Plausible Inference.San Mateo: Morgan Kaufmann. This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions 870 QUEEN AND SMITH [No. 4, Pearl,J.and Verma,T. S. (1987) The logicofrepresenting dependenciesbydirectedgraphs.In Proc. 6th Natn. Conf. ArtificialIntelligence,Seattle,pp. 374-379. Los Altos: Morgan Kaufmann. Pole, A. and West,M. (1990) EfficientBayesianlearningin non-lineardynamicmodels.J.Forecast.,9, 119-136. Pole, A., West,M. and Harrison,P. J.(1988) Non-normaland non-lineardynamicBayesianmodelling. In BayesianAnalysisof TimeSeriesand DynamicModels (ed. J. C. Spall). New York: Dekker. models:a generalapproachto non-lineartimeseriesanalysis.J. Priestley,M. B. (1980) State-dependent TimeSer. Anal., 1, 47-71. Queen, C. M. (1992) Using the multiregression dynamic model to forecastthe brand sales in a competitiveproductmarket.Research Report 92-10. NottinghamStatisticsGroup, Universityof Nottingham,Nottingham. Quintana,J. M. and West,M. (1987) Multivariatetimeseriesanalysis:newtechniquesapplied to internationalexchangeratedata. Statistician,36, 275-281. probabilisticinference.In Uncertainty and ArtificialIntelligence(eds Shachter,R. D. (1986) Intelligent L. N. Kanal and J. Lemmer),pp. 371-382. Amsterdam:North-Holland. Smith,J. Q. (1989) Influencediagramsforstatisticalmodelling.Ann. Statist.,17, 654-672. (1990) Statisticalprinciplesof graphs.In InfluenceDiagrams,BeliefNets and Decision Analysis (eds R. M. Oliverand J. Q. Smith),pp. 89-120. New York: Wiley. Wermuth,N. and Lauritzen,S. L. (1990) On substantive researchhypotheses, conditionalindependence graphsand graphicalchain models. J. R. Statist.Soc. B, 52, 21-50. West,M. and Harrison,P. J. (1989) BayesianForecastingand DynamicModels. New York: Springer. West, M., Harrison,P. J. and Migon, H. J. (1985) Dynamicgeneralisedlinearmodels and Bayesian forecasting(withdiscussion).J. Am. Statist.Ass., 80, 73-97. Zeeman, E. C. (1977) CatastropheTheory:SelectedPapers (1972-1977). Reading:Addison-Wesley. Zellner,A. (1984) Basic Issues in Econometrics.Chicago: University of Chicago Press. This content downloaded from 86.30.149.44 on Tue, 22 Oct 2013 11:53:08 AM All use subject to JSTOR Terms and Conditions
© Copyright 2026 Paperzz