Statisticalapproachesforevaluatingcausaleffectsandfordiscoveringcausalnetworksarediscussedinthispaper.Acausalrelationbetweentwovariablesisdifferentfromanassociationorcorrelationbetweenthem.Anassociationmeasurementbetweentwovariablesandmaybechangeddramaticallyfrompositivetonegativebyomittingathirdvariable,whichiscalledYule-Simpsonparadox.WeshalldiscusshowtoevaluatethecausaleffectofatreatmentorexposureonanoutcometoavoidthephenomenaofYule-Simpsonparadox.Surrogatesandintermediatevariablesareoftenusedtoreducemeasurementcostsordurationwhenmeasurementofendpointvariablesisexpensive,inconvenient,infeasibleorunobservableinpractice.Therehavebeenmanycriteriaforsurrogates.However,itispossiblethatforasurrogatesatisfyingthesecriteria,atreatmenthasapositiveeffectonthesurrogate,whichinturnhasapositiveeffectontheoutcome,butthetreatmenthasanegativeeffectontheoutcome,whichiscalledthesurrogateparadox.Weshalldiscusscriteriaforsurrogatestoavoidthephenomenaofthesurrogateparadox.Causalnetworkswhichdescribethecausalrelationshipsamongalargenumberofvariableshavebeenappliedtomanyresearchfields.Itisimportanttodiscoverstructuresofcausalnetworksfromobserveddata.Weproposearecursiveapproachfordiscoveringacausalnetworkinwhichastructurallearningofalargenetworkisdecomposedrecursivelyintolearningofsmallnetworks.Furthertodiscovercausalrelationships,wepresentanactivelearningapproachintermsofexternalinterventionsonsomevariables.Whenwefocusonthecausesofaninterestoutcome,insteadofdiscoveringawholenetwork,weproposealocallearningapproachtodiscoverthesecausesthataffecttheoutcome.
