Logistic regression with interactions dilemma

Moderators: statman, Analyst Techy, andris, Fierce, GerineL, Smash

allencornelius
Posts: 2
Joined: Sun Dec 10, 2017 6:00 pm

Logistic regression with interactions dilemma

Postby allencornelius » Sun Dec 10, 2017 8:03 pm

SPSSers,

I am getting some confusing results from different approaches with logistic regression. Here is the situation.

I have a question about how SPSS creates categorical contrasts and then interaction terms in logistic regression. The data in question has four dichotomous variables – verdict, race, SES of defendant, and SES of participant. Verdict (guilty or not) is predicted from the three other predictors, and two interaction terms (Race* SES of defendant and SES of defendant*SES of participant.). Since all variables were coded as 0 and 1, I thought there would be no need to have SPSS create contrasts, and the results were this -

B S.E. Wald df Sig. Exp(B)
Race .330 .333 .986 1 .321 1.392
SESofDef -.861 .407 4.467 1 .035 .423
SESofPart -.111 .333 .110 1 .740 .895
Race by SESofDef -1.283 .486 6.975 1 .008 .277
SESofDef by SESofPart 1.583 .486 10.612 1 .001 4.872
Constant .418 .285 2.156 1 .142 1.520

However, if you use the SPSS option to create categorical variables, which I think in essence changes the 0 and 1 to indicator variables of 1 and 0, you get vastly different results, and I can’t figure out why –

B S.E. Wald df Sig. Exp(B)
Race(1) .953 .354 7.243 1 .007 2.593
SESofDef(1) .561 .408 1.893 1 .169 1.752
SESofPart(1) -1.473 .354 17.263 1 .000 .229
Race(1) by SESofDef(1) -1.283 .486 6.975 1 .008 .277
SESofDef(1) by SESofPart(1)1.583 .486 10.612 1 .001 4.872
Constant .078 .286 .074 1 .786 1.081

The values for the interaction terms are identical, but the values for the main effect terms are vastly different.

If you just examine a model with just the main effects and not the interactions, the results of the two methods are identical (as they should be), but when you add the interaction terms, the results for the main effects are different, but the results for the interactions are identical. Something about how indicator interaction variables are related to the main effect variables, but I don't know why reversing the coding on the variables would have such an effect. And which is the correct interpretation?

Any idea what is going on here? Any help to untangle this would be appreciated.

Thanks.

Allen
allencornelius
Posts: 2
Joined: Sun Dec 10, 2017 6:00 pm

Re: Logistic regression with interactions dilemma

Postby allencornelius » Sun Dec 10, 2017 11:21 pm

Sorry about the formatting of the output.

Who is online

Users browsing this forum: No registered users and 1 guest

cron