2
$\begingroup$

I have a model with occurrence of a disease represented by a binary dependent variable (DV) and 8 independent variables (IVs) at different levels. I need to create a multi-level model, in which the treatment is placed in the lower order and the demographics are in the higher order.

However, I am not familiar with the multilevel model for logistic regression. Please give me some names of necessary multilevel analyses for doing a multilevel binary logistic regression (and any hints you think are useful). I wonder if GEE (generalized estimating equation) is the answer, because I have correlations between the IVs? What about "conditional (fixed-effects) binary logistic regression" again for the paired data among my IVs? Or else? ...

$\endgroup$
  • 1
    $\begingroup$ You can also do it with version 13 of Stata using gsem. $\endgroup$ – user32173 Nov 1 '13 at 8:55
  • 1
    $\begingroup$ Without more detail, this is more like a comment than an answer. Could you add some syntax? $\endgroup$ – Peter Flom - Reinstate Monica Nov 1 '13 at 9:47
6
$\begingroup$

R allows what are called generalized linear mixed effects models. In these, the response variable is allowed to be from a few different families, including binomial (which, if coded as 0 and 1, gives logistic regression).

The function used to be called glmer(). I'm pretty sure that now more recent versions of the regular mixed effects models function lmer() allows you to specify a family (e.g. 'binomial') and a link function (e.g. 'logit'). lmer() allows the specification of random effects and nesting. You can find more info on Doug Bates' slides, in particular the very last one, here . He wrote lmer(), so I believe him when he says it works.

Keep in mind that you need numerous (more than 6 or so) different 'subjects' to be able to estimate random effects efficiently.

$\endgroup$
  • $\begingroup$ Many thanks. Yes what I was looking for was genelarized lieanr mixed effect models and I saw later in SPSS 19 and above they are available too. I guess I would try R's version since SPSS is not at all handy in this case. $\endgroup$ – Vic Aug 23 '13 at 8:54
  • $\begingroup$ Also thank you very much for the hints. :) I think I have enough sample size then (regardless of power calculations). $\endgroup$ – Vic Aug 23 '13 at 9:01
  • 2
    $\begingroup$ The link to the Doug Bates slides is broken. $\endgroup$ – coip Mar 29 '17 at 19:16
3
$\begingroup$

Is this question related to your question here? Do you want to assess the effect of treatment on disease, with matched (paired) individuals?

Anyway, the difference between conditional logistic regression and GEE is the interpretation. If you want to get subject specific estimate, you can use conditional logistic regression (e.g. clogit in R), otherwise for population average estimate, you can use GEE (e.g. R package gee). Note that the reason to use multilevel models is the correlation within paired data.

$\endgroup$
  • $\begingroup$ Many thanks for the clarifications of the differences and for the relevant R packages. I think I would go with a mixed-model GLZ instead of GEE and conditional logistic regression since I have both paired and unpaired IVs among my variables. And yes that is it (the link you provided) with a little difference. In that thread, I had not mentioned that I need two different levels (thus a multilevel model). $\endgroup$ – Vic Aug 23 '13 at 8:58
  • $\begingroup$ (of course if you had any other hints regarding GEE and colgit versus mixed model GLZ, I would be so grateful to hear). Thanks. $\endgroup$ – Vic Aug 23 '13 at 9:04
  • 1
    $\begingroup$ Do you mean mixed generalized linear model (GLM)? What do you mean by "paired and unpaired IVs"? I guess it is matched individuals rather than matched independent variables, as discussed in your last thread. $\endgroup$ – Randel Aug 25 '13 at 3:13
  • 2
    $\begingroup$ As to the choice among GEE, clogit and mixed GLM for binary data, some people are in favor of clogit since some of the consistency properties break down, especially with small within-subject sample size; some prefer GEE to obtain population average estimate as there is attenuation of the effect of the covariates for subject specific model (e.g. mixed GLM). You can have your own choice based on your data format and analysis goals. $\endgroup$ – Randel Aug 25 '13 at 3:25
  • $\begingroup$ yes by GLZ I mean generalized linear model (I didn't use GLM since it is confused by general). Yes your comment on the "matched individuals not IVs" was great. I saw it after I wrote the above comment. But my individuals are matched in certain aspects, but not in other aspects. So I have two levels of analysis. My setup is elaborated on in here: talkstats.com/showthread.php/… $\endgroup$ – Vic Aug 25 '13 at 9:08

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.