资源描述:
《the r book generalized linear models》由会员上传分享,免费在线阅读,更多相关内容在工程资料-天天文库。
1、13GeneralizedLinearModelsWecanusegeneralizedlinearmodels(GLMs)–pronounced‘glims’–whenthevarianceisnotconstant,and/orwhentheerrorsarenotnormallydistributed.Certainkindsofresponsevariablesinvariablysufferfromthesetwoimportantcontraventionsofthestandardassu
2、mptions,andGLMsareexcellentatdealingwiththem.Specifically,wemightconsiderusingGLMswhentheresponsevariableis:countdataexpressedasproportions(e.g.logisticregressions);countdatathatarenotproportions(e.g.log-linearmodelsofcounts);binaryresponsevariables(e.
3、g.deadoralive);dataontimetodeathwherethevarianceincreasesfasterthanlinearlywiththemean(e.g.timedatawithgammaerrors).VarianceVariance41302002468102468100246810MeanMeanVarianceVariance0.01.01.52.02.53.00.502040608010002468100246810MeanMeanTheRBook,SecondE
4、dition.MichaelJ.Crawley.©2013JohnWiley&Sons,Ltd.Published2013byJohnWiley&Sons,Ltd.558THERBOOKThecentralassumptionthatwehavemadeuptothispointisthatvariancewasconstant(topleft-handgraph).Incountdata,however,wheretheresponsevariableisanintegerandthereareoft
5、enlotsofzerosinthedataframe,thevariancemayincreaselinearlywiththemean(toptight).Withproportiondata,wherewehaveacountofthenumberoffailuresofaneventaswellasthenumberofsuccesses,thevariancewillbeaninvertedU-shapedfunctionofthemean(bottomleft).Wheretherespon
6、sevariablefollowsagammadistribution(asintime-to-deathdata)thevarianceincreasesfasterthanlinearlywiththemean(bottomright).ManyofthebasicstatisticalmethodssuchasregressionandStudent’sttestassumethatvarianceisconstant,butinmanyapplicationsthisassumptionisun
7、tenable.HencethegreatutilityofGLMs.AGLMhasthreeimportantproperties:theerrorstructure;thelinearpredictor;thelinkfunction.Thesearealllikelytobeunfamiliarconcepts.Theideasbehindthemarestraightforward,however,anditisworthlearningwhateachoftheconceptsinvol
8、ves.13.1ErrorstructureUptothispoint,wehavedealtwiththestatisticalanalysisofdatawithnormalerrors.Inpractice,however,manykindsofdatahavenon-normalerrors:forexample:errorsthatarestronglyskewed;errorsthatarekurtotic;errorst