Interpreting odds ratios in PROC LOGISTIC

noetsi

Fortran must die
#1
I ran this code

PROC LOGISTIC DATA=SASUSER.MODS5
PLOTS(ONLY)=ALL
;
CLASS Race_W (ref ="1") HISP (REF= "1") PRIVPUB1 (REF="0") EDDUMMY (REF="1") FEMDUMMY (REF="1")/PARAM=REF ;
MODEL LOGDUM (Event = '1')=Age Race_W HISP PRIVPUB1 EDDUMMY FEMDUMMY /
SELECTION=NONE
LINK=LOGIT
;
RUN;
QUIT;

The odds ratio table generates in part

EDDUMMY 0 VS 1 1.427
PRIVDUMMY 1 vs 0 2.199

Do you interpret this (the SAS results confuse me) that for EDDUMMY if the predictor takes on a value of 0 the event is 1.427 times more likely. And for PrivDummy if the predictor takes on the value of 1 the event is 2.199 times more likely?

In terms of interpreting the substantive impact I am less clear what is a large effect for logistic regression than linear (what is a large effect is more intuitively obvious to me for linear regression). When I see effects near 1 (for example race has a value of 1.062) I see this as a small effect size (since 1 is similar in logistic regression to 0 in linear regression as showing no effect size).
 

hlsmith

Omega Contributor
#2
EDDUMMY "0" group has a 1.427 times greater odds of Event = "1" compared to the EDDUMMY "1" group. You need to also include the CI to understand how precise these measures are, which in turn kind of fills you in on the sampling distribution.


So you get effect direction from OR, but magnitude is contextual and interpretation not always intuitive on the multiplicative scale. A possible alternative in the situation of rare outcome is looking at risk difference equivalences. So the results are on the additive scale. You just have to get used to OR in your area to grasp whether they are large or not.


Keep asking questions if you want!!