Different results when changing the reference group in logistic regression

Hi everyone,

I have a (maybe) basic question. I have this binominal logistic regression model:

Dependent var: Work full-time (dummy var: 1=work full-time; 0= work part-time)

Independent var: categorial var (1= the first generation; 2=the 2nd generation; 3=3rd+ generation) --> recoded into 3 dummy variables.

Controls for several indicators (sex, age, education,...).

I ran this model twice. The first time with the first generation as the reference group. The second time with the second generation as the reference group).

However, they gave me two different results (same B value (of course, one negative and one positive), but different exp(B)). The first result shows that the 2nd generation is 14% less likely than the 1s generation to work full-time . The second result shows that the first generation is 16% more likely than the 2nd generation to work full-time.

Table 1.PNG
Table 2.PNG

I attach the screenshots of the 2 result tables for you to investigate. My question is why is there this difference? Which number should I take given the fact that I want to see the full-time employment gap between the 1st and 2nd generations (14% or 16%)?

Thank you for your responses.


Note: for this dependent variable (full-time), the difference between 2 results is only 2%. but for other variables, the difference is much bigger.
Last edited:
I can't see the attachments so others probably can not either. Did you look at the SPSS documentation?

In SAS dummy variables are tied to effect coding not reference coding and the OR are tied always to reference coding. SPSS, which I don't know, may do the same. So that is a place to start. But without seeing the code and the results it will be hard to comment.

Almost no one on this board uses SPSS so it is not going to be easy to get answers. Different software work differently and that is critical in answers.

Thanks for your reply.

Sorry, I've just attached the screenshots again. I think that this question is more about the nature of logistic regression than about the software. SPSS or SAS/R/Stata would give me similar results. I don't understand why the result changes when I change the reference. Exp(B)s of Gen 3 are different in 2 regression tables of course, but the exp (B) of "Gen 1 compared to Gen 2" should be the same as the exp(B) of "Gen 2 compared to Gen 1".

Thanks again,
Well they might not. People get confused in SAS on this issue because they don't realize when they build dummy variables sas defaults to effect coding for the slopes, but uses reference coding for the odds ratios.

Dummy variables are being compared to the intercept so if the intercept changes the dummies will as well. In other words for 3 it will be 1 compared to 0 relative to what the intercept is.