You are on page 1of 3

Stepwise Discriminant Function Analysis SPSS will do stepwise DFA.

. You simply specify which method you wish to employ for selecting predictors. The most economical method is the Wilks lambda method, which selects predictors that minimize Wilks lambda. As with stepwise multiple regression, you may set the criteria for entry and removal (F criteria or p criteria), or you may take the defaults. magine that you are working as a statistician for the nternal !e"enue Ser"ice. You are told that another !S employee has de"eloped four composite scores (# $ % #&), easily computa'le from the information that ta(payers pro"ide on their income ta( returns and from other data'ases to which the !S has access. These composite scores were de"eloped in the hope that they would 'e useful for discriminating ta( cheaters from other persons. To see if these composite scores actually ha"e any predicti"e "alidity, the !S selects a random sample of ta(payers and audits their returns. )ased on this audit, each ta(payer is placed into one of three groups* +roup $ is persons who o"erpaid their ta(es 'y a considera'le amount, +roup , is persons who paid the correct amount, and +roup - is persons who underpaid their ta(es 'y a considera'le amount. # $ through #& are then computed for each of these ta(payers. You are gi"en a data file with group mem'ership, # $, #,, #-, and #& for each ta(payer, with an e.ual num'er of su'/ects in each group. Your /o' is to use discriminant function analysis to de"elop a pair of discriminant functions (weighted sums of #$ through #&) to predict group mem'ership. You use a fully stepwise selection procedure to de"elop a (may'e) reduced (less than four predictors) model. You employ the 0 12S method of selecting "aria'les to 'e entered or deleted, using the default p criterion for entering and remo"ing "aria'les. Your data file is 34A% ST5P.sa", which is a"aila'le on 2arl6s SPSS% 3ata page %% download it and then 'ring it into SPSS. To do the 34A, click Analy7e, 8lassify, and then put +roup into the +rouping 9aria'le 'o(, defining its range from $ to -. Put #$ through #& in the : ndependents 'o(, and select the stepwise method.

8opyright ,;;< 2arl 1. 0uensch % All rights reser"ed.


34A%Step.doc

Page , 8lick =ethod and select :0ilks6 lam'da and :>se pro'a'ility of F. 8lick 8ontinue.

>nder Statistics, ask for the group means. >nder 8lassify, ask for a territorial map. 8ontinue, ?2. 1ook at the output, :9aria'les @ot in the Analysis. At Step ; the ta( groups (o"erpaid, paid correct, underpaid) differ most on # - ( drops to .A-A if #- is entered) and :Sig. of F to enter is less than .;B, so that predictor is entered first. After entering # -, all remaining predictors are eligi'le for entry, 'ut # $ most reduces lam'da, so it enters. The 0ilks lam'da is reduced from .A-B to .$C$. ?n the ne(t step, only # , is eligi'le to enter, and it does, lowering 0ilks lam'da to .;B<. At this point no "aria'le already in meets the criterion for remo"al and no "aria'le out meets the criterion for entry, so the analysis stops. 1ook 'ack at the Step ; statistics. ?nly #, and #- were eligi'le for entry. @ote, howe"er, that after #- was entered, the p to enter dropped for all remaining predictors. 0hyD #- must suppress irrele"ant "ariance in the other predictors (and "ice "ersa). After # $ is added to #-, p to enter for #& rises, indicating redundancy of #& with #$. Interpretation of the Output from the E ample !ro"ram f you look at the standardized coefficients and loadin"s you will see that high scores on DF1 result from high #- and low #$. f you look 'ack at the "roup means you will see that those who underpaid are characteri7ed 'y ha"ing low # - and high #$, and thus low DF1. This suggests that DF1 is good for discriminating the cheaters (those who underpaid) from the others. The centroids confirm this. f you look at the standardi7ed coefficients and loadings for DF2 you will see that high DF2 scores come from ha"ing high #, and low #$. 4rom the group means you see that those who o"erpaid will ha"e low DF2 (since they ha"e a low #, and a high #$). DF2 seems to 'e

Page good for separating those who o"erpaid from the others, as confirmed 'y the centroids for DF2. n the territorial map the underpayers are on the left, ha"ing a low DF1 (high #$ and low #-). The o"erpayers are on the lower right, ha"ing a high DF1 and a low DF2 (low #,, high #-, high #$). Those who paid the correct amount are in the upper right, ha"ing a high DF1 and a high DF2 (low #$, high #,, high #-).

8opyright ,;;< 2arl 1. 0uensch % All rights reser"ed.

You might also like