[R-sig-ME] Generalized Linerar Model vs Logistic regression

[Ben Bolker]

  (Please keep r-sig-mixed-models at r-project.org in the Cc: list ...)

On 16-11-22 06:57 PM, Cleber Iack wrote:
> Thank you very much for the return.
> 
> Yes, school is a categorical predictor with 4 levels and Posto is
> categorical with 5 levels
> 
> I have 16000 individuals, being that predictive variables are in media
> with 5 longitudinal observations for individual. I have 76000 lines.
> This base would be enough?  I apologize for not having understood very well.

  That sounds like enough provided that the response is not extremely
rare or extremely common.  With this kind of design you do have to be a
little bit careful with another aspect, which is that the Laplace
approximation may be a bit questionable in this case (low effective
sample size *per cluster*) - it would be a good idea, if you can, to
bump up the value of the nAGQ argument to, say, nAGQ=10 (the fit will be
slower, and this won't work if you have complex random effect structures).
> 
> The intraclass correlation coefficient (ICC) what I mean would be to
> assess how much of the total variance of the population is explained by
> the difference between schools.
> 
> Each one of my schools have a very different identity, but I thought of
> a test in order to be able to demonstrate to an audience in a
> presentation for example.

  I personally wouldn't need any convincing that you should be using a
mixed model in this case.  Quoting the ICC along with its confidence
interval would be useful if the among-individual variability is actually
of practical interest.  Other significance tests are available, as I
suggested in my previous message.

> 
> Once again thank you very much and I thank others comments as you can do.
> 
> 2016-11-22 20:24 GMT+00:00 Ben Bolker <bbolker at gmail.com
> <mailto:bbolker at gmail.com>>:
> 
> 
> 
>     On 16-11-22 12:44 PM, Cleber Iack wrote:
>     > Dear,
>     >
>     > Good night.
>     >
>     > I am Phd student, but I have a model with reply Binaria, and I own 23
>     > predictive variables, among them school (4) and Posto (5)
> 
> 
>       I'm guessing that what you mean by this is that school is a
>     categorical predictor with 4 levels and Posto is categorical with 5
>     levels?
> 
>       According to Frank Harrell's (_Regression Modeling Strategies_) rules
>     of thumb, you need about 20 times as many effective observations as the
>     number of parameters in the model you're trying to fit; in the case of
>     binary data, 'effective' observations means the minimum of (number of
>     zeros, number of ones) in your response, i.e. the number of the
>     less-common response. So I hope you have a large data set (at least
>     hundreds and preferably thousands of observations ...)
> 
>     >
>     > a) My ICC on the school is showing 0.23, can I use this information to
>     > corroborate the use of Generalized Linerar Model instead of a Logistic
>     > regression?
>     >
>     > b) If the letter a) is not true, I can verify through the AIC and BIC?
> 
>       I don't know about ICC. You can in principle use AIC or BIC, or a
>     likelihood ratio test (LRT), to justify the use of a mixed model. My
>     personal preference would be to use the mixed model if it makes sense in
>     the context of your observational/experimental design (e.g. you have a
>     number of discrete groups in your population that can be thought of as
>     exchangeable, i.e. changing the identity of the groups wouldn't change
>     your conclusions), and not to try to use quantitative tests for this
>     purpose.  You can read more about the use and caveats of AIC, BIC, LRT,
>     etc. at http://tinyurl.com/glmmFAQ
> 
>     >
>     >
>     > I appreciate any feedback.
>     >
>     > Thank you
>     >
>     > Cleber
>     >
>     >       [[alternative HTML version deleted]]
>     >
>     > _______________________________________________
>     > R-sig-mixed-models at r-project.org
>     <mailto:R-sig-mixed-models at r-project.org> mailing list
>     > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>     <https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models>
>     >
> 
>     _______________________________________________
>     R-sig-mixed-models at r-project.org
>     <mailto:R-sig-mixed-models at r-project.org> mailing list
>     https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>     <https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models>
> 
>