Hello everybody!
I was planning to do a conditional logistic regression on some experimental data i have not yet gathered. I do have some expectations on how the data will look like, however, and I created a dataset to play with and run some tests on in beforehand. In the end I will probably have data on around 150 individuals.
Each individual delivers 20 to 40 datapoints for the variables Y (binary, independent) and X (continuous, between 0 and 100, dependent) per treatment of which there are 3 in total (baseline, treat1, treat2)
I expect most of the individuals to have an individual switching point S at some level of X so that they exhibit Y=0 for values below and Y=1 for values above S. This switching point differs between subjects and I might also differ between treatments for each single individual.
I may have some misunderstanding regarding the conditional logistic regression, so this is what I assume:
A conditional logistic regression with Y as dependent and X as the only independent variable on the data of one single treatment subject-ID as a group variable would probably suffer from perfect separation, since lots of subjects switch from Y=0 to Y=1 for all X≥S.
Running it over a dataset with three different treatments per subject when some of the subjects do indeed have different switching points depending on the treatment and introducing two treatment dummies should in my eyes solve the separation problem.
My question is whether I am right in this assumption, because when I run
clogit Y X treat1 treat2, Group(subject_id)
with my made up data, I receive the "convergence not achieved" notification. In my dataset, individual switching points differ quite considerably between treatments, so I am a bit puzzled as to why I get this result.
I would understand getting this message if there is no difference between the treatments whatsoever for most of the subjects, but this is not the case here. What am I missing?
I was planning to do a conditional logistic regression on some experimental data i have not yet gathered. I do have some expectations on how the data will look like, however, and I created a dataset to play with and run some tests on in beforehand. In the end I will probably have data on around 150 individuals.
Each individual delivers 20 to 40 datapoints for the variables Y (binary, independent) and X (continuous, between 0 and 100, dependent) per treatment of which there are 3 in total (baseline, treat1, treat2)
I expect most of the individuals to have an individual switching point S at some level of X so that they exhibit Y=0 for values below and Y=1 for values above S. This switching point differs between subjects and I might also differ between treatments for each single individual.
I may have some misunderstanding regarding the conditional logistic regression, so this is what I assume:
A conditional logistic regression with Y as dependent and X as the only independent variable on the data of one single treatment subject-ID as a group variable would probably suffer from perfect separation, since lots of subjects switch from Y=0 to Y=1 for all X≥S.
Running it over a dataset with three different treatments per subject when some of the subjects do indeed have different switching points depending on the treatment and introducing two treatment dummies should in my eyes solve the separation problem.
My question is whether I am right in this assumption, because when I run
clogit Y X treat1 treat2, Group(subject_id)
with my made up data, I receive the "convergence not achieved" notification. In my dataset, individual switching points differ quite considerably between treatments, so I am a bit puzzled as to why I get this result.
I would understand getting this message if there is no difference between the treatments whatsoever for most of the subjects, but this is not the case here. What am I missing?
