# Thread: Logistic Regression & Interaction Term

1. ## Logistic Regression & Interaction Term

Hi there, I'm trying to predict whether or not a patient has a simultaneous prescription order with various variables. To make it more clear...

DV
-Simultaneous Order (0 = No, 1 = Yes)

Predictors:
-Age (Continuous)
-Race (0 = Black, 1 = White, 2 = Unknown, 3 = Hispanic)
-Training Level (0 = Trainee, 1 = NP/PA, 2 = Attending)
-Prescribing Service (0 = Medical, 1 = Surgical, 2 = Pain Service)

After adding an interaction term of "Training Level" and "Prescribing Service"....the odds ratio BLOWS up. I attached a document showing before and after....please help me make sense of this...

Thank you!

2. ## Re: Logistic Regression & Interaction Term

Looking at your odds ratio they don't seem to get that much larger to me after interaction. For example Standingpoioid(1) goes from 1,865 to 1.951 which does not seem to be a huge change (although I am not a biological researcher so maybe that is a huge change for your analysis). Nor do your statistical significance change that dramatically.

3. ## The Following User Says Thank You to noetsi For This Useful Post:

gene2420 (06-12-2015)

4. ## Re: Logistic Regression & Interaction Term

Originally Posted by noetsi
Looking at your odds ratio they don't seem to get that much larger to me after interaction. For example Standingpoioid(1) goes from 1,865 to 1.951 which does not seem to be a huge change (although I am not a biological researcher so maybe that is a huge change for your analysis). Nor do your statistical significance change that dramatically.
Hmm you're right about that. The one thing that did change dramatically was Rxservice_Int(2). Also, the "Rxservice_Int(2) by Training_Int(1)" and "Rxservice_Int(2) by Training_Int(2)" have blown up OR's. Is that a cause for concern? Or should I just be reading into the Rxservice_Int*Training_Int? I'm sorry, I'm a little new to interactions...Thanks for you all your help though, Noetsi.

5. ## Re: Logistic Regression & Interaction Term

I thought initially they were in the original model (by blown up I thought you meant increased) but now I realize they have not. To begin with, I doubt that odds ratios that high have any substantive meaning (they are absurdly high compared to any I have seen or your main effects). Also there SE are absurdly high for the same variables.

One issue that might be causing this is what is called quasi-complete separation. The tell tale signs of this are very high slope and very high standard error. This is often caused by a dummy variable for example that at one level of the dummy the DV takes on one value always and at the other level it takes on another value of the DV always (or almost always). One thing to look for to suggest this might be an issue is the the test of the model (not the slopes the model itself). If you see the Likelihood ratio and Wald Chi Square Test (which normally generate similar results) are very far apart this suggests partial separation may be occuring. Wald chi square test are doubtful when you have this type of problem.

This is caused formally because you can perfectly, or near perfectly, predict the DV with a linear combination of the predictors. I have not seen it addressed in the context of interaction, but I assume the same issues apply as apply to main effects in this regard.

I am not a statistician nor a biologist (I am a data analyst who uses statistics). Thus I am cautious about suggesting solutions. However, you might want to see Paul Alison's book "Logistic Regression Using SAS" 2nd ed 46-55 which includes some suggestions.

6. ## The Following User Says Thank You to noetsi For This Useful Post:

gene2420 (06-12-2015)

7. ## Re: Logistic Regression & Interaction Term

Originally Posted by noetsi
I thought initially they were in the original model (by blown up I thought you meant increased) but now I realize they have not. To begin with, I doubt that odds ratios that high have any substantive meaning (they are absurdly high compared to any I have seen or your main effects). Also there SE are absurdly high for the same variables.

One issue that might be causing this is what is called quasi-complete separation. The tell tale signs of this are very high slope and very high standard error. This is often caused by a dummy variable for example that at one level of the dummy the DV takes on one value always and at the other level it takes on another value of the DV always (or almost always). One thing to look for to suggest this might be an issue is the the test of the model (not the slopes the model itself). If you see the Likelihood ratio and Wald Chi Square Test (which normally generate similar results) are very far apart this suggests partial separation may be occuring. Wald chi square test are doubtful when you have this type of problem.

This is caused formally because you can perfectly, or near perfectly, predict the DV with a linear combination of the predictors. I have not seen it addressed in the context of interaction, but I assume the same issues apply as apply to main effects in this regard.

I am not a statistician nor a biologist (I am a data analyst who uses statistics). Thus I am cautious about suggesting solutions. However, you might want to see Paul Alison's book "Logistic Regression Using SAS" 2nd ed 46-55 which includes some suggestions.
Thanks for you help Noetsi, I'll keep investigating this further with the quasi-complete separation idea.

8. ## Re: Logistic Regression & Interaction Term

I asked our true statisticians to comment. Hopefully one will

9. ## The Following User Says Thank You to noetsi For This Useful Post:

gene2420 (06-12-2015)

10. ## Re: Logistic Regression & Interaction Term

Originally Posted by noetsi
I asked our true statisticians to comment. Hopefully one will
Awesome! I would love to see the insight of others as well

 Tweet

#### Posting Permissions

• You may not post new threads
• You may not post replies
• You may not post attachments
• You may not edit your posts