Date: Sat, 4 Dec 2010 07:38:47 -0500
Reply-To: Peter Flom <peterflomconsulting@mindspring.com>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Peter Flom <peterflomconsulting@MINDSPRING.COM>
Subject: Re: proc surveylogistic: why no interaction term odds ratios?
Content-Type: text/plain; charset=UTF-8
Robert Feyerharm <robertf@HEALTH.OK.GOV> wrote
>After running proc surveylogistic on complex survey data, my colleagues
>and I have noticed that SAS does not output odds ratio estimates for the
>interaction terms of categorical variables as well as any main effect
>terms which are interacting(however, proc rlogist in SUDAAN does output
>interaction odds ratios).
>
>Any ideas why SAS doesn't print this information? Is it related to the
>methodology (effect terms rather than dummy terms) SAS uses to code
>categoricals?
No, it's related to the nature of logistic regression; in fact, until recently, you couldn't get these for interaction in PROC LOGISTIC either.
You can see my paper: PROC LOGISTIC: Traps for the unwary (NESUG) for more details and a workaround. Briefly - an interaction effect with a continuous dependent variable means that the additive model for the two main effects is insufficient. This works because, in this case, main effects can be added; e.g. if the DV is income and you have IVs for sex and ethnicity, then the interaction will be (close to) 0 if the effect of sex plus the effect of ethnicity (almost) fully explains the variation in income. You can add the two effects. You can look at the average effect of sex (across ethnicities) and average effect of ethnicity (across sexes).
However, odds can't be added and averaged this way. Or, rather, they can be, but you get a mess. The average odds is not the average of the odds.
Peter
Peter L. Flom, PhD
Statistical Consultant
Website: http://www DOT statisticalanalysisconsulting DOT com/
Writing; http://www.associatedcontent.com/user/582880/peter_flom.html
Twitter: @peterflom
| 
