Logistic regression problem

kati

New Member
#1
Hi,

I'm experiencing difficulties with understanding my results despite active googling, and i'd be really grateful, if someone could help me.

I have a data set of around 1000 cases. The outcome i'm interested in (yes, no) occurs in around 35 %. In order to identify independent predictors for this outcome, i tried to construct a logistic regression analysis. In the logistic regression model, a dichotomous variable had a negative B coefficient value against my expectations (in crosstabulation, the variable seemed to be significantly associated with outcome occurring, p <0.001). The variable's CI for the OR was below 1 and p<0.001. Also, the SE was 0.2. After doing some reading, i assumed the problem of the wrong sign would be related to correlation between that another independent variable, so i began eliminating them in different combinations with no effect on the sign of the B coefficient. Moreover, no significant correlation between the variables seemed to exist (VIF 2 at most). In the end,i tried a model that included only that one variable, and still the negative b coefficient persisted. I also ran a linear regression (same y and x) but this yielded a positive b coefficient.


I'm really having trouble figuring this out and would very much appreciate any help!

K
 

Link

Ninja say what!?!
#3
If cross tabulation shows that the relation is supposed to be positive and you are not including any covariates, I might look at whether the logistic regression model is modelling the occurance correctly. It may be modelling the probability that the outcome = 0. This happens in SAS and you have to specify "descending" in order for the model to show results for P(Y=1).

HTH
 

kati

New Member
#5
Thanks for your reply! I'm using SPSS and I went through all the options I could find but nothing there seemed to be nothing like this. Also, all in all the original mode made sense and the problem only concerned this one variable.

K
 

jrai

New Member
#6
How is your goodness of fit test coming out? It'd be good to see that the hypothesis is not being rejected. And when you ran a linear regression then did you do it y on that single x or did you use all the IVs used in logistic regression.