Seite - 331 - in Differential Geometrical Theory of Statistics

Entropy2016,18, 421 anddiscretegraphicalmodels. Testing isoftenusedtochecktheconsistencyofaparametricmodel withgivendata, andtocheckdependencyassumptionssuchas independencebetweencategorical variables.However,wenotean importantcaveat: aspointedoutby[14,15], the fact thataparametric model“passes”agoodness-of-fit testonlyweaklyconstrains the resulting inference. Theessential point here is that goodness-of-fit is a necessary, but not sufficient, condition for model choice, since—ingeneral—manymodelswillbeempiricallysupported. This issuehasrecentlybeenexplored geometrically in [16]usingCIG. Therehavebeenmanypossible test statisticsproposed forgoodness-of-fit testing, andoneof theattractionsof thePower-Divergence family,defined in (11), is that themost importantonesare included in the family and indexedbya single scalar λ. Of course,when there is a choice of test statistic, different inferences can result fromdifferent choices. Oneof themain themesof [5] is to give theanalyst insightabout selectingaparticularλ. Keyconsiderations formaking theselection ofλ include the tractabilityof the samplingdistribution, its poweragainst important alternatives, andinterpretationwhenhypothesesarerejected. Thefirstorder,asymptotic inN,χ2-samplingdistributionforallmembersof thePower-Divergence family,which is appropriatewhenall observedcounts are “large enough”, is themost commonly usedtool, andaveryattractive featureof the family.However, thiscanfailbadly in the“sparse”case andwhen themodel is close to theboundary. Elementary,momentbasedcorrections, to improve small sampleperformance,arediscussedin[5] (Chapter5).Moreformalasymptoticapproaches to these issues includethedoublyasymptotic, inNandk, approachof [17],discussedinSection2and similarnormalapproximation ideas in [18]. Seealso [19]. Extensivesimulationexperimentshavebeen undertakento learn inpracticewhat ‘largeenough’means, see [5,20,21]. Whentherearenuisanceparameters tobeestimated(as iscommon), [22]pointsout that it is the samplingdistribution conditionalupontheseestimateswhichneeds tobeapproximated,andproposes higher ordermethodsbasedon theEdgeworth expansion. Simulation approaches are oftenused in the conditional context due to the common intractability of the conditional distribution [23,24], and importancesamplingmethodsplayan important role—see [25–27].Otherapproachesused to investigatethesamplingdistributionincludejackknifing[28], theChen–Steinmethod[29],anddetailed asymptoticanalysis in [30–32]. Inveryhighdimensionalmodel spaces, considerationsof thepowerof tests rarelygenerates uniformly best procedures but,we feel, geometry can be an important tool in understanding the choices thatneedtobemade. Further, [5], states thesituation is“complicated”, showingthis through simulationexperiments.Oneof thereasons forReadandCressie’spreferredchoiceofλ=2/3is its goodpoweragainst someimportant typesofalternative–theso-calledbumpordipcases–aswellas therelative tractabilityof its samplingdistributionunder thenull.Otherconsiderationsaboutpower canbefoundin[33]which looksspecificallyatmixturemodelbasedalternatives. 3.3. Linkswith InformationGeometry At the time that the Power-Divergence family was being examined, there was a parallel development in InformationGeometry; oddly,however, it seemedtohave takensometimebefore the links between the two areas were fully recognised. A good treatment of these links can be found in [6] (Chapter 9). Since it is important to understand the extreme values of divergence functions, considerationsofconvexitycanclearlyplayanimportantrole. ThegeneralclassofBregman divergences, [6,34] (page240), and[35] (page13) isveryusefulhere. ForeachBregmandivergence, therewill existaffineparametersof theexponential family inwhich thedivergence function isconvex. In theclassofproductPoissonmodels—whichare thekeybuildingblocksof log–linearmodels—all membersof thePower-Divergence familyhavetheBregmanproperty. Theseare thenα-divergences, capableofgenerating thecomplete InformationGeometryof themodel [35],with the linkbetweenα andλgiven inTable1. Theα-representationhighlights thedualityproperties,whichareacornerstone of InformationGeometry,butwhich is ratherhidden in theλ representation. TheBregmandivergence 331