regression interaction effects

[David Greenberg]

Barbara, Jaccard is misinterpreting the results of the logistic regression. Before explaining why they are misinterpreted let me point out that the odds ratio you give for the effect of Male*Part-time is 2.5886, which is not exactly the same as what you say in the paragrapah below the output results, where you say it is  2.8688. I haven't checked which is right. Anyway, you caan't look at the coefficient of the interaction term in isolation from the terms involving the variables that make up the interaction term. Your equation has the logit equaling
 a + b_1*Male +b_2 *FT +b_3*PT + b_4*Male*FT +b_5*M*PT. We can now evaluate this for different combinations of sex and work-status of the mother.
For a mother of a girl who is not employed, the only contribution is from the constant term, a. For a boy whose mother is unemployed, the coefficient is a+ b_1.
For a mother of a boy, the mother working part time. the contributions are a +b_1 + b_3 + b_5.  If working with the odds ratios, you have to take the product of .4257 and 2.5886 or 2.8688. To compare with the boys whose mothers don't work, divide by 1.1068. You will see that the result is an odds ratio that is only a little greater than 1. This is because the effects of PT and PT*Male are in the opposite direction. I am curious as to the source. Who is Jaccard? Can you give me a complete citation? I have been thinking of writing an article on this misinterpretation of interaction terms, and this would be  a good example to use.
     - David Greenberg, Sociology Department, NYU

[Judith Saebel]

David,
 
I think the reference could be "Interaction Effects in Logistic Regression"; paperback
ISBN: 9780761922070; Sage Publications, Inc; Pub Date: 27-03-2001; Pages: 80.
(see also
http://www.sagepub.co.uk/booksProdDesc.nav?contribId=501580&prodId=Book11268 <http://www.sagepub.co.uk/booksProdDesc.nav?contribId=501580&prodId=Book11268> )
 
Judith Saebel

________________________________

From: SPSSX(r) Discussion on behalf of David Greenberg
Sent: Sat 22/09/2007 9:44 AM
To: [hidden email]
Subject: regression interaction effects



Barbara, Jaccard is misinterpreting the results of the logistic regression. Before explaining why they are misinterpreted let me point out that the odds ratio you give for the effect of Male*Part-time is 2.5886, which is not exactly the same as what you say in the paragrapah below the output results, where you say it is  2.8688. I haven't checked which is right. Anyway, you caan't look at the coefficient of the interaction term in isolation from the terms involving the variables that make up the interaction term. Your equation has the logit equaling
 a + b_1*Male +b_2 *FT +b_3*PT + b_4*Male*FT +b_5*M*PT. We can now evaluate this for different combinations of sex and work-status of the mother.
For a mother of a girl who is not employed, the only contribution is from the constant term, a. For a boy whose mother is unemployed, the coefficient is a+ b_1.
For a mother of a boy, the mother working part time. the contributions are a +b_1 + b_3 + b_5.  If working with the odds ratios, you have to take the product of .4257 and 2.5886 or 2.8688. To compare with the boys whose mothers don't work, divide by 1.1068. You will see that the result is an odds ratio that is only a little greater than 1. This is because the effects of PT and PT*Male are in the opposite direction. I am curious as to the source. Who is Jaccard? Can you give me a complete citation? I have been thinking of writing an article on this misinterpretation of interaction terms, and this would be  a good example to use.
     - David Greenberg, Sociology Department, NYU