Multiple linear reg (some obs. not independent)

[StatsClue]

Building a multiple linear reg. model. Some patients (less than 10% I suspect) underwent the same procedure twice and are listed as two seperate obs.
Been using proc glm.
>Are regression assumptions about independence violated?
>Do I need to account for these few non-independent obs. in my regression?
>Do I need to change to proc genmod (from glm)? (Genmod constitutes robust regression I think and allows you to ignore the independence assumption, unless I'm mistaken).
>So, if I should change to genmod, do I still need to account for this non-independence? If yes,....know the syntax? Does glm, by any chance, allow you to account for non-independence?

THANKS.

[StatsClue]

Folks.....?

[Dason]

>Are regression assumptions about independence violated?

Yes.

>Do I need to account for these few non-independent obs. in my regression?

It couldn't hurt and you probably should. It's hard to say how big of an impact the repeated measures will have but it would be a better, more honest analysis if you accounted for them if you can. With that said if you don't think you can account for them one thing you could do is simulate some data according to the ~10% repeated observations that you think is going on and then analyze what this does to the analysis - do you increase power? Do you keep your nominal alpha? Those are questions you could ask and answer through simulation

>Do I need to change to proc genmod (from glm)? (Genmod constitutes robust regression I think and allows you to ignore the independence assumption, unless I'm mistaken).

Like I said in the chatbox a few times; genmod is for generalized linear models - not robust regression. proc robustreg takes care of robust regression but I don't think that's what you want. If you don't know what a generalized linear model really is then I would suggest reading up on them and hopefully you'll see that just converted to a "glm" doesn't fix your problem.

[Karabiner]

Building a multiple linear reg. model. Some patients (less than 10% I suspect) underwent the same procedure twice and are listed as two seperate obs.

Pehaps you have to define your inclusion/exclusion criteria more thoroughly.

Kind regards

K.

[StatsClue]

Pehaps you have to define your inclusion/exclusion criteria more thoroughly.

Kind regards

K.

Repeat procedures on the same patients aren't meant to be excluded. I was only asking about how this was to be taken into account.

[jrai]

Actually proc Genmod can take care of this because repeated statement in Genmod invokes GEE (generalized estimating equations) which is used for modelling for correlation within the subjects.

If you're modelling fixed effects (you won't know this if you don't understand GLMM) then you can manage the correlations within proc GLM itself by using the Absorb statement. If you're modelling random or mixed effects then you'd need proc Mixed.

But GEE in Genmod is probably the easiest option.

[Dason]

But you would need to know which observations are repeated right?

My point was mainly that just using genmod doesn't fix the problem.

[Karabiner]

Repeat procedures on the same patients aren't meant to be excluded.
I was only asking about how this was to be taken into account.

How it should be taken into account depends on what is actually
studied, and on the study objectives. Unfortunately all this was
not described here. In many instances, deleting the second occurence
of the same subject is a simple consequence of the definition of the
study population.

Kind regards

K.