Subscribe Bookmark RSS Feed

How to perform a conditional logistic regression using JMP?

jennifer_n_blak

Community Trekker

Joined:

Jan 25, 2016

Hello!

  I am interested in performing a conditional logistic regression in JMP, but have been unable to find any documentation regarding how to go about this except in model choice type of scenario.  Can anyone advise?

 

Briefly, here is what I am trying to do.  

I have three categorical nominal groupings which are each divided into a control and treatment group.  My outcomes fall into 5 non-ordered categories.  I am interested in comparing the treatment to the control within each group and then see whether the treatment has similar effects across groups. My sample size is small (14-30) per initial group, equaling 70 individuals overall (35 treated, 35 control).
 
After researching various types of logistic regressions, it seems that a conditional logistic approach might be best.  Matched-pairs might also be a contender, but the individuals are not necessarily matched, so much as the overall characteristics of the group.
 
Do you have a recommendation for how to perform an analysis such as this? 
I would be interested in both a discussion on the appropriateness of technique (or alternate recommendations) as well as how-to using the JMP platform. I do not have experience using JSL.
 
Thanks so much for your assistance!
Jennifer
 
 
2 ACCEPTED SOLUTIONS

Accepted Solutions
Dan_Obermiller

Joined:

Apr 3, 2013

Solution

Although this approach can work, it would have some difficulties. Let me propose a different path.

Rather than making three models (one for each grouping) and then trying to figure out if the treatment effect is the same across the groups, make one model.

 

I see a data table with three columns: Grouping, Trt/Control, and Outcome. All columns are nominal.

In Fit Model, specify a model of Grouping, Trt/Control, and Grouping*Trt/Control. Outcome is the Y. This will give you a nominal logistic regression model. To see if there is an association between the treatment and the outcome, look at the Trt/Control part of the Effect Likelihood Ratio Tests report. To see if there are differences in the groupings, look at the Grouping line of that report. Is the treatment effect the same across groups? That is answered by the interaction term. A significant interaction says that the treatment effect is different across the groups. Lots of other good stuff to look at, but those are the basics for the questions you asked.

 

If you truly want a separate analysis for each group, then specify the model as just Trt/Control with Outcome as the Y and put the Grouping variable in the BY role. It is more difficult then to determine if the treatment effect is the same across the groups because each analysis is separate and would require either hand analysis or doing something like I first outlined.

Dan_Obermiller

Joined:

Apr 3, 2013

Solution

The only recommendations are to: 1) do what you are doing -- can you combine some of the levels of the categorical variables? Fewer levels means fewer parameters in the model. 2) Can you live with not answering some of the questions. For example, do you need to know if the treatment effect is the same for each group? If not, you can remove the interaction term from the model and that would probably help tremendously.

 

Not great options, but that is all I see since getting more data is not an option.

8 REPLIES
Dan_Obermiller

Joined:

Apr 3, 2013

Solution

Although this approach can work, it would have some difficulties. Let me propose a different path.

Rather than making three models (one for each grouping) and then trying to figure out if the treatment effect is the same across the groups, make one model.

 

I see a data table with three columns: Grouping, Trt/Control, and Outcome. All columns are nominal.

In Fit Model, specify a model of Grouping, Trt/Control, and Grouping*Trt/Control. Outcome is the Y. This will give you a nominal logistic regression model. To see if there is an association between the treatment and the outcome, look at the Trt/Control part of the Effect Likelihood Ratio Tests report. To see if there are differences in the groupings, look at the Grouping line of that report. Is the treatment effect the same across groups? That is answered by the interaction term. A significant interaction says that the treatment effect is different across the groups. Lots of other good stuff to look at, but those are the basics for the questions you asked.

 

If you truly want a separate analysis for each group, then specify the model as just Trt/Control with Outcome as the Y and put the Grouping variable in the BY role. It is more difficult then to determine if the treatment effect is the same across the groups because each analysis is separate and would require either hand analysis or doing something like I first outlined.

jennifer_n_blak

Community Trekker

Joined:

Jan 25, 2016

Dear Dan,

  Thank you very much!  Your instructions were clear and the effect likelihood showed that both the treatment and the grouping were significant, but not the interaction of the two.  

Surprisingly, though, when I look at the parameter estimates they all have ~.99 as the P value.  How is it possible to have many parameter estimates all be non-significant and yet get significance for the overall effects of treatment and grouping in the model?

(I have included a screen shot of the output below. 

Grouping = sex_2015;  Trt/control = treatment; Outcome = sex_2016)

 

Thanks Again!

Jennifer

 

PRU_plus_control_15-16_-_Fit_Nominal_Logistic.jpg

Dan_Obermiller

Joined:

Apr 3, 2013

Ah, yes. This was something I feared. You do not have much data. Only 70 observations. You need to have those 70 observations distributed appropriately through the various combinations to help with the fit results. You are getting "unstable" messages which is JMP's way of telling you that the parameter estimates are not very reliable due to a lack of data. At the overall level, there is enough to consider significance for each term, but within each model (remember that you actually have a model for each level of the response variable), there is not enough data to accurately estimate the parameters.

jennifer_n_blak

Community Trekker

Joined:

Jan 25, 2016

I was afraid of that! I will play around with condensing down some of the categories to see if I can beef up the sample size in a combined group

I will play around with condensing down some of the categories to see if I can beef up the sample size in a combined group. Aside from getting more data (which is not really an option for this experiment at this point, sadly) would you have any other recommendations?

Thank you so much for your help!

Dan_Obermiller

Joined:

Apr 3, 2013

Solution

The only recommendations are to: 1) do what you are doing -- can you combine some of the levels of the categorical variables? Fewer levels means fewer parameters in the model. 2) Can you live with not answering some of the questions. For example, do you need to know if the treatment effect is the same for each group? If not, you can remove the interaction term from the model and that would probably help tremendously.

 

Not great options, but that is all I see since getting more data is not an option.

jennifer_n_blak

Community Trekker

Joined:

Jan 25, 2016

Understood! Thanks for all your help!

 

markbailey

Staff

Joined:

Jun 23, 2011

What is the meaning of the five nominal response levels: A, DEAD, F, M, and N?

jennifer_n_blak

Community Trekker

Joined:

Jan 25, 2016

Dear Mark,

  I am working with a tree that can change sex from year to year. So trees in the experiment started out as male(M), female(F), or ambivalent sex (A).  After the treatment those same trees ended up as male(M), female(F), non-reproductive (N), ambivalent sex (A), or dead.  These are nominal categorical responses because we don't yet really understand how sex changes normally in these species.  They may, in fact, turn out to be ordinal with more information.