Page 14 - Regression Guideline

"MulJlinear"regression"with"stepwise"
Step&4:&Stepwise&mulDvariable®ression&
Sidebar:&What&is&stepwise®ression&and&why&was&this&method&of®ression&
selected?&
" An"important"part"of"regression"modeling"is"determining"which"set"of"predictors"do"the"best" job"of"predicJng"the"dependent"variable."SAVVI"AnalyJcs"uses"a"method"called"stepwise" regression."This"method"examines"each"individual"property"characterisJc"as"a"predictor"to" determine"the"extent"to"which"knowing"this"specific"characterisJc"was"useful"for"gecng"a" more"accurate"esJmate"of"the"sales"price"for"the"properJes"examined.""
"
SelecJon"of"variables"in"a"stepwise"regression"can"be"either"forward"or"backward."In" forward"selecJon,"the"model"starts"with"zero"predictors"and"adds"one"new"predictor"at"a" Jme."Variables"conJnue"to"be"added"to"the"model"unJl"no"remaining"variables"improves" the"value"of"R2"beyond"a"minimum"threshold"of"staJsJcal"significance."(The"default" minimum"“FPPPtoPPPenter”"in"SAVVI"is"0.15.)"Backward"eliminaJon"is"similar"except"that"the" model"begins"with"all"predictors"included"and"then"eliminates"those"that"do"not"maintain"a" staJsJcal"threshold"of"significance.""
"
The"result"of"stepwise"regression"modeling"is"the"idenJficaJon"of"the"best"subset"of" predictors"for"esJmaJng"a"sales"price"for"a"given"property."One"common"outcome"of"this" process"that"can"be"confusing"at"first"is"when"the"subset"of"best"ficng"predictors"does"not" include"all"of"the"property"characterisJcs"considered,"excluding"even"some"the"appraiser" was"certain"were"important."This"occurs"because"the"excluded"characterisJcs"were"found" to"be"staJsJcally"unrelated"to"sales"price"for"the"properJes"used"to"develop"the"regression" model."These"excluded"characterisJcs"do"not"improve"price"esJmaJon"even"though" intuiJvely,"it"might"seem"as"if"they"should.""
"
This"can"be"for"two"reasons:""1)"the"property"characterisJc"simply"was"not"associated"with" variaJon"in"the"sales"prices"of"the"properJes"examined;"or"2)"the"variaJon"in"sales"price" accounted"for"by"the"excluded"predictor"was"beser"accounted"for"by"other"property" characterisJcs"already"included"as"predictors"in"the"model"(see"“nonPindependence"of" property"characterisJcs”"below)."
"
13"