# Extremely weird logistic regression results

I am getting very weird results in my logistic regression analysis and am confused about why. Any help would be greatly appreciated. Here are the details:

1. The illustration below represents the results of the 2nd Model in a logistic regression that includes 2 interaction terms. I am getting some very weird standard error, B coefficients, and odds-ratio results that are clearly not reliable for the 2nd model. But the weird thing to me is that when these variables are included separately in the first model, the results are normal.

2. All variables are binary, coded as 0 and 1. So essentially they are all dummy variables. There don't appear to be any outliers and I am certain they are coded correctly. There are sufficient cases in each category when I check crosstabs for my interaction terms and independent variables with the dependent variable.

3. The results go back to normal if I remove variable 1 from the 1st Model. But if I do than I cannot look at the direct effects of that variable. I also get correct results if I run separate analyses for the effects of each interaction term. Is it OK to just do that instead?

4. From my understanding it is appropriate to conduct a logistic regression with all binary variables. Is my understanding wrong or is it not appropriate to create an interaction term with binary variables or something?

Illustration:

FYI there are 4 variables: Variable 1 (independent, coded 0/1) Variable 2 (independent, coded 0/1) Variable 3 (independent, coded 0/1) & the Dependent (coded 0/1)

Model 1 Logistic regression including direct effects of all independents on the dependent. Normal results

Model 2 Logistic regression including direct effects of all independents on the dependent AND the effects of 2 interaction terms on the dependent. The interaction terms are the product of variable 1 times variable 2 and the product of variable 1 times variable 3. Results include extremely high B and odds-ratio coefficients and standard error values. For example: B=-20.799, SE=40192.876, odds-ratio= 1153680239

*I have also found that if I do NOT include the direct effects of Variable 1 in Model 1 OR if I only include 1 interaction term in Model 2, the effects go back to normal.

Thank you so much!!!

• Although you refer repeatedly to "illustrations" and "results," there are none of the former and only incomplete accounts of the latter. Have you had problems including this information in your post? – whuber Aug 11 '17 at 18:19
• To investigate an extreme coefficient in logistic regression, it's wise to begin by checking for linearly dependent variables and complete separation. – Kodiologist Aug 11 '17 at 18:41
• Yes, my pictures didn't load. :-( Thank you! I figured out the problem. You're correct, it was a collinearity problem. – alinamars Aug 11 '17 at 20:54