Constant value in logistic regression

arb

New Member
#1
Hi,
My question concerns the constant represents in logistic regression analysis (which can be included/excluded in SPSS).

I am trying to understand the effect of binary variable (risk factor) DP on the dependent variable RP, a symptom.

If I were to assume that the incidence of RP in the population W/O risk factor DP is 50%, then I believe I could leave the constant out of the model.

However if I do not "know" the incidence of RP in the population W/O risk factor DP, then I should leave the constant in. The model will then then include a negative constant to correct for a baseline incidence of <50%, or it will include a positive constant to correct for a baseline incidence of >50%. For example if the baseline incidence is 25%, the constant becomes -1.1.

The constant corrects for the probability/incidence of the dependent variable w/o ANY of the dependent variables (in my case I just have one dependent variabl)

Is this reasoning correct?

Thanks in advance
 
#2
Pardon me if I am missing something very basic here, but: why are you doing logistic regression if your indepdent variable is categorical (RP present or not present)? Why not just a 2X2 contingency table?

But to answer your question: the constant term in a logistic regression represents is the log-odds of DV=1 vs. DV=0, if all IV values are zero. For example: if you logit-regress having a heart attack in the next year vs. the concentration of some marker in the blood, then the constant term is the natural log of the odds of having a heart attack when there is no marker in your blood.
 

arb

New Member
#3
thank you for your help

I had been including several independent variables in my model so I had been using logistic regression, but in this case I do only have 1 binary IV and a binary DV, i dont have a good reason for choosing log regression in this case.

is there a difference in the odds ratio that are computed between a 2x2 table versus from a simple logistic regression?
 
#4
I haven't worked out whether they are mathematically guaranteed equal, but I would be very surprised if they (and their confidence intervals) were not very close.