proc logistic with too many zeroes

[Reeza]

I suspect that's the one that's causing the quasi separation issue - because it does separate the data quite cleanly.

[igforek]

It may be so.
Also, when I try to run genmod with the factors Bact Host and their interaction, the warning about the Hessian matrix not being positive definite may be caused by that many zeroes.
The model runs ok with no interaction.

[igforek]

I took a look at this page: https://support.sas.com/rnd/app/stat/examples/GENMODZIP/roots.htm

but I am not sure if this will apply to my data, because I have categorical independent variables. And binomial.

[igforek]

HostB is naturally infected by BactB
HostC is naturally infected by BactC
and so on
BactE is naturally infected by BactE

[igforek]

Mr. Ksharp,

I thought that unbalanced data meant that there were too many observations in one class, as compared to another class in an analysis.

Are you suggesting that too many zeros in one class can be a special case of unbalanced data?

Regards,

[Ksharp]

Yes. unbalanced data is way too biased from 1:1. Like  good:bad =99: 1.

You could oversample it into good:bad=30:70 and base that data to build a logistic model.

About your question, I am not able to follow , what does that ' special case ' mean ?

[Reeza]

This isn't a case of unbalanced data exactly - it's a case of extreme multicollinearity when two variables provide the same almost the exact same information. In that case, depending on context, one variable is usually dropped.

[igforek]

What do you think about Dr. Allison insight into dummy variables?
https://statisticalhorizons.com/multicollinearity

[Reeza]

Which of those cases do you think applies here?

Have you tried changing your reference level? Are you interested in estimates on the categorical variables - I think you are so they're not just control variables here. 

---

@igforek wrote:
What do you think about Dr. Allison insight into dummy variables?
https://statisticalhorizons.com/multicollinearity

---

[igforek]

Unfortunately, I am right mow working on a computer that does not run SAS
or any other statistical software. I will try changing my reference this
afternoon, after 5 pm central time.

I am thinking case 3 may apply to my analysis. And yes, I am interested in
estimates on the categorical variables

[Reeza]

---

@igforek wrote:
Unfortunately, I am right mow working on a computer that does not run SAS
or any other statistical software. I will try changing my reference this
afternoon, after 5 pm central time.

I am thinking case 3 may apply to my analysis. And yes, I am interested in
estimates on the categorical variables

---

Then his comments don't apply because it  only does if it's a control variable or you're interested in the overall effects. So you need to find some other way to deal with the collinearity or re-frame your problem/experiment somehow. If you knew this was likely to happen, host and bacteria would match primarily, a different design should have been choosen .... unfortunately it's likely too late to change that. 

[igforek]

To me personally, this look like a "Hindsight is 20/20" situation.
Other researchers in my field used this type of analysis successfully.
In my case, there are too many zeroes and some host do not even get infected by their own bacteria ("0" self-infection). Who knew?

[Ksharp]

What I mean is Y variable , NOT X variable , multicollinearity is for X variables .

[igforek]

Thank you for your comments.
I am still working on solving the issues with the analysis.
I will let all of you know how I goes.