Transcription of Regression with a Binary Dependent Variable - Chapter 9
1 Regressionwitha Binary DependentVariableChapter9 MichaelAshCPPAL ecture22 CourseNotesIEndgameITake-home nalIDistributedFriday 19 MayIDueTuesday 23 May (Paper or emailedPDFok; no Word,Excel,etc.)IProblemSet7 IOptional,worth up to 2 percentagepointsof extracreditIDueFriday 19 MayIRegressionwitha Binary DependentVariableBinary DependentVariablesIOutcomecanbe coded1 or 0 (yes or no,approvedor denied,successor failure)Examples?IInterprettheregression as modelingtheprobability thatthedependentvariableequalsone(Y= 1).IRecallthatfor a Binary Variable ,E(Y) = Pr(Y= 1)HMDA exampleIOutcome:loandenialis coded1, loanapproval0 IKeyexplanatory Variable :blackIOtherexplanatory variables:P=I, credithistory, LTV, Probability Model(LPM)Yi= 0+ 1X1i+ 2X2i+ + kXki+ 1expressesthechangein probability thatY= 1 associatedwitha ^Yiexpressestheprobability thatYi= 1Pr(Y= 1jX1;X2.)
2 Xk) = 0+ 1X1+ 2X2+ + kXk=^YShortcomingsof theLPMI\NonconformingPredictedProbabilit ies"Probabilitiesmustlogicallybe between0 and1, construction(always userobuststandarderrors)ProbitandLogitRe gressionIAddressesnonconformingpredicted probabilitiesin theLPMIB asicstrategy:boundpredictedvaluesbetween 0 and1 bytransforminga linear index, 0+ 1X1+ 2X2+ + kXk,whichcanrangeover( 1;1) intosomethingthatrangesover[0;1]IWhenthe indexis bigandpositive,Pr(Y= 1)! bigandnegative,Pr(Y= 1)! to transform?Usea thecumulativestandard normaldistribution, .Theindex 0+ 1X1+ 2X2+ + kXkis treatedas (Y= 1jX1;X2; : : : ;Xk) = ( 0+ 1X1+ 2X2+ + kXk)InterpretingtheresultsPr(Y= 1jX1;X2; : : : ;Xk) = ( 0+ 1X1+ 2X2+ + kXk)I jpositive(negative)meansthatan increaseinXjincreases(decreases)theproba bility ofY= jreportshow theindexchangeswitha changeinX, buttheindexis onlyan inputto jis hard to interpretbecausethechangeinprobability for a changeinXjis non-linear, dependsonallX1;X2; : : : ; interpretationis computingthepredictedprobability^Yfor alternativevaluesofXISameinterpretationo f standard errors,hypothesistests,andcon denceintervalsas withOLSHMDA \Pr(deny= 1jP=I.
3 Black) = ( 2:26+2:74P=I+0:71black)(0:16)(0:44)(0:08 3)IWhiteapplicantwithP=I= 0:3:\Pr(deny= 1jP=I;black) = ( 2:26 + 2:74 0:3 + 0:71 0) = ( 1:44)= 7:5%IBlackapplicantwithP=I= 0:3:\Pr(deny= 1jP=I;black) = ( 2:26 + 2:74 0:3 + 0:71 1) = ( 0:71)= 23:3%Logitor LogisticRegressionLogit,or logisticregression,usesa slightlydi erentfunctionalformof theCDF(thelogisticfunction)insteadof thestandard cientsof theindexcanlook di erent,buttheprobabilityresultsare usuallyverysimilar to theLPM,thethreemodelsgeneratesimilar LogitandProbitModelsIOLS(andLPM,whichis an applicationof OLS)hasaclosed-formformulafor^ ILogitandProbitrequirenumericalmethods to nd^ 's thatbest t LeastSquaresOneapproachis to choosecoe cientsb0;b1; : : : ;bkthatminimizethesumof squaresof how far theactualoutcome,Yi, is fromtheprediction, (b0+b1X1i+ +bkXki).
4 NXi=i[Yi (b0+b1X1i+ +bkXki)]2 MaximumLikelihood EstimationIAnalternativeapproachis to choosecoe cientsb0;b1; : : : ;bkthatmake thecurrentsample,Y1; : : : ;Ynas likelyas possibleto example,if youobservedataf4;6;8g, thepredictedmeanthatwouldmake thissamplemostlikelyto occuris ^ MLE= (logit)commandsareunderStatistics! Binary OutcomesInferenceandMeasuresof FitIStandard errors,hypothesistests,andcon denceintervalsareexactlyas in OLS,buttheyreferto thecoe cientsandmustbe translatedintoprobabilitiesby cut-o , ,0:50,andcheckthefractioncorrectlypredic ted, .sensitivity/speci cityChoosea cut-o .Sensitivity is thefractionof observedpositive-outcomesthatare correctlyclassi city is thefractionof observednegativeoutcomesthatare correctlyspeci analogousto theR2 IExpressesthepredictivequality of themodelwithexplanatory variablesrelativeto thepredictivequality of thesampleproportionpofcaseswhereYi= 1 IAdjustsfor addingextraregressorsSensitivity andSpeci cutoffSensitivitySpecificityReviewingthe HMDA results( )ILPM,logit,probit(minor di erences)IFourprobitspeci ,controllingfor a widerangeof otherexplanatory ModelsLimitedDependentVariable(LDV)ICoun tData(discretenon-negativeintegers),Y20; 1;2.
5 , , , ,mode of ,chooser, probit,ICansometimesconvertto severalbinary sampleselectionmodels.