<p>Inphenotypeannotationscuratedfromthebiolog-icalandmedicalliterature,considerablehumaneffortmustbeinvestedtoselectontologicalclassesthatcapturetheexpressivityoftheoriginalnaturallanguagedescriptions,andannotationgranularitycanalsoentailhighercomputationalcostsforpartic-ularreasoningtasks.Docoarseannotationsforcertainapplications?Here,wemeasurehowannotationgranularityaffectsthestatisticalbehaviorofsemanticsimilaritymetrics.Weusearandomizeddatasetofphenotypeprdrawnfrom57,051taxon-phenotypeannotationsinthePhenoscapeKnowledgebase.WecomparedqueryprhavingvariableproportionsofmatchingphenotypestosubjectdatabaseprusingbothpairwiseandgroupwiseJaccard(edge-based)andResnik(node-based)semanticsimilaritymetrics,andcomparedstatisticalperformanceforthreedifferentlevelsofannotationgranularity:entitiesalone,entitiesplusattributes,andentitiesplusqualities(withimplicitattributes).Allfourmetricsexaminedshowedmoreextremevaluesthanexpectedbychancewhenapproximatelyhalftheannotationsmatchedbetweenthequeryandsubjectprwithamoresuddendeclineforpairwisestatisticsandamoregradualoneforthegroupwisestatistics.Annotationgranularityhadanegligibleeffectonthepositionofthethresholdatwhichmatchescouldbediscriminatedfromnoise.Theseresultssuggestthatcoarseannotationsofphenotypes,atthelevelofentitieswithorwithoutattributes,maybetoidentifyphenotypeprwithstatisticallysemanticsimilarity.</p>