[R-sig-ME] R mixed effects problem

Douglas Bates bates at stat.wisc.edu
Mon Aug 27 15:45:22 CEST 2007 

On 8/26/07, Iasonas Lamprianou <lamprianou at yahoo.com> wrote:
> Deat friends, I still need some answers on the following questions. I posted this email but I got no response, it would be great if any one of you could give me some hints. Also, I noticed that the resid function "is not yet implemented" in lmer for the bionomial model. Has this been solved in lme2?

> jason

> ----- Forwarded Message ----
> From: Iasonas Lamprianou <lamprianou at yahoo.com>
> To: r-sig-mixed-models at r-project.org
> Sent: Friday, 24 August, 2007 12:55:10 PM
> Subject: Re: R-sig-mixed-models Digest, Vol 8, Issue 13

> Dear friends, in the following example, I am testing teacher effect. So, the teacher is affecting the score of the pupils, and the items are also random effects. The ID is the id of the pupils. Year (pupils are either year6 ot year7) and age and gender may play some role on the test score.

> Question 1: In the following example, I am not sure what the Correlation of Fixed Effects is. Could someone please explain it to me? How do I use these numbers?

The estimators of the fixed-effects parameters are random variables
and can be correlated.  This table shows the correlation of the
estimators, evaluated at the estimates.

It would be hard to go into more detail without explaining some
elementary statistics, which would raise the question of why you are
trying to fit very sophisticated statistical models if you are unaware
of basic properties of statistical models.  We can answer questions
about the software but it is not realistic to expect that we will
provide a tutorial on statistics via an email list.

> Question 2: Also, what is the meaning of this statment? Estimated scale (compare to  1 )  0.865952 . How do I interpet this?

The log-likelihood for a generalized linear mixed model cannot be
evaluated explicitly - it must be approximated.  One of the
approximations, called penalized quasi-likelihood (PQL), provides an
estimate of a scale parameter that should be fixed at 1 for a GLMM
model with binomial or Poisson responses.  The apparent value of this
parameter is returned.  Values much larger than 1 indicate
over-dispersion relative to the assumed conditional distribution of
the responses.  Values less than 1 indicate under-dispersion.

It is not uncommon to have moderate under-dispersion in GLMMs because
the variability in the data can be modeled by random effects or by the
per-observation noise.

> Thanks
>
>
>
> Generalized linear mixed model fit using Laplace
> Formula: SCORE ~ YEAR + AGE + GENDER + (1 | ITEM) + (1 | ID) + (1 | TEACHER)
>    Data: data2
> Family: binomial(logit link)
>   AIC  BIC logLik deviance
> 4979 5027  -2482     4965
> Random effects:
> Groups  Name Variance Std.Dev.
> ID           0.508153 0.71285
> TEACHER      0.039018 0.19753
> ITEM         3.874608 1.96840
> number of obs: 7043, groups: ID, 754; TEACHER, 29; ITEM, 10
> Estimated scale (compare to  1 )  0.865952
> Fixed effects:
>                  Estimate Std. Error z value Pr(>|z|)
> (Intercept)    -1.6334597  0.6457887 -2.5294   0.0114 *
> YEAR[T.Year 7]  0.0495903  0.1262880  0.3927   0.6946
> AGE             0.0004811  0.0009579  0.5023   0.6155
> GENDER[T.girl] -0.0546018  0.0924347 -0.5907   0.5547
> ---
> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
> Correlation of Fixed Effects:
>             (Intr) YEAR[7 AGE
> YEAR[T.Yr7] -0.101
> AGE         -0.203 -0.095
> GENDER[T.g] -0.063 -0.067 -0.008
>
>
>
> Dr. Iasonas Lamprianou
> Department of Education
> The University of Manchester
> Oxford Road, Manchester M13 9PL, UK
> Tel. 0044 161 275 3485
> iasonas.lamprianou at manchester.ac.uk
>
>
> ----- Original Message ----
> From: "r-sig-mixed-models-request at r-project.org" <r-sig-mixed-models-request at r-project.org>
> To: r-sig-mixed-models at r-project.org
> Sent: Wednesday, 22 August, 2007 1:00:17 PM
> Subject: R-sig-mixed-models Digest, Vol 8, Issue 13
>
>
> Send R-sig-mixed-models mailing list submissions to
>     r-sig-mixed-models at r-project.org
>
> To subscribe or unsubscribe via the World Wide Web, visit
>     https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> or, via email, send a message with subject or body 'help' to
>     r-sig-mixed-models-request at r-project.org
>
> You can reach the person managing the list at
>     r-sig-mixed-models-owner at r-project.org
>
> When replying, please edit your Subject line so it is more specific
> than "Re: Contents of R-sig-mixed-models digest..."
>
>
> Today's Topics:
>
>    1. Re: R-sig-mixed-models Digest, Vol 8, Issue 10
>       (Iasonas Lamprianou)
>
>
> ----------------------------------------------------------------------
>
> Message: 1
> Date: Tue, 21 Aug 2007 03:35:05 -0700 (PDT)
>
> Subject: Re: [R-sig-ME] R-sig-mixed-models Digest, Vol 8, Issue 10
> To: r-sig-mixed-models at r-project.org
> Message-ID: <607384.7411.qm at web54103.mail.re2.yahoo.com>
> Content-Type: text/plain; charset=iso-8859-1
>
> Hi all, I have a question regarding the paper
> Estimating the Multilevel Rasch Model: With the
> lme4 Package
> by Doran, Bates etc al.
> I would be greateful if someone could answer these questions, I am still a beginner.
>
> In page 10, they fitted a model with no intercept. (1) What might be the motivation for this?  (1) How would the interpretation be different if they had included am intercept? (3) How may I derive the table of residuals (observed-predicted)? (4) How can I estimate/evaluate the fit of individual items, subjects or companies?
>
> I know that I have many questions, but I really love lme4 and I would like to routinely use it in my research. I have a great education dataset which I would love to analyse. Any experts wishing a joint publication? My motivation is to learn as much as possible during the collaboration for this paper.
>
> Dr. Iasonas Lamprianou
> Department of Education
> The University of Manchester
> Oxford Road, Manchester M13 9PL, UK
> Tel. 0044 161 275 3485
> iasonas.lamprianou at manchester.ac.uk
>
>
> ----- Original Message ----
> From: "r-sig-mixed-models-request at r-project.org" <r-sig-mixed-models-request at r-project.org>
> To: r-sig-mixed-models at r-project.org
> Sent: Wednesday, 15 August, 2007 1:00:14 PM
> Subject: R-sig-mixed-models Digest, Vol 8, Issue 10
>
>
> Send R-sig-mixed-models mailing list submissions to
>     r-sig-mixed-models at r-project.org
>
> To subscribe or unsubscribe via the World Wide Web, visit
>     https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> or, via email, send a message with subject or body 'help' to
>     r-sig-mixed-models-request at r-project.org
>
> You can reach the person managing the list at
>     r-sig-mixed-models-owner at r-project.org
>
> When replying, please edit your Subject line so it is more specific
> than "Re: Contents of R-sig-mixed-models digest..."
>
>
> Today's Topics:
>
>    1. Re: lmer, intercepts and offsets (Daniel Farewell)
>
>
> ----------------------------------------------------------------------
>
> Message: 1
> Date: Tue, 14 Aug 2007 16:39:02 +0000 (GMT)
> From: Daniel Farewell <farewelld at cf.ac.uk>
> Subject: Re: [R-sig-ME] lmer, intercepts and offsets
> To: r-sig-mixed-models at r-project.org
> Message-ID: <523651.84987.qm at web27101.mail.ukl.yahoo.com>
> Content-Type: text/plain; charset=iso-8859-1
>
> This is a follow-up to a thread from back in March ("lmer, intercepts and offsets"). I'm hoping (at least) to better understand how lmer works.
>
> I'd like to "trick" lmer into thinking it has converged to certain parameter estimates. This is straightforward for variance components, making use of the 'start' parameter and using 'control' to set the number of the various kinds of iteration to zero.
>
> My ultimate goal is to extract posterior second moments from a model "fit" where I have specified both the variance components and the fixed effects.
>
> Obviously it is possible to dig inside the fitted model and manually alter the fixed effects, but this has no impact on the result of a call to ranef, presumably because the posterior means (and variances?) have already been calculated, and are sitting in the ranef slot of the fitted model.
>
> My question is this: at what stage do the random effects get calculated? Simplifying greatly, at some stage lmer must calculate betahat(Omega) (the fixed effects) and ranef(Omega) (the estimated random effects) for the converged value of Omega. Does ranef(Omega) depend on the result of betahat(Omega)? If so, then presumably tinkering with the betahat(Omega) results at the appropriate point inside lmer would result in what I want. If not (that is, if the dependence on the fixed effects is indirect) what needs tinkering with?
>
> Is there a good reason why lmer does not allow models with no fixed effects at all? With the right offset, this would be another way to achieve the same result.
>
> [[replacing trailing spam]]
>
> Daniel Farewell
>
>
>
> ------------------------------
>
> _______________________________________________
> R-sig-mixed-models mailing list
> R-sig-mixed-models at r-project.org
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>
>
> End of R-sig-mixed-models Digest, Vol 8, Issue 10
>
>
>
> ------------------------------
>
> _______________________________________________
> R-sig-mixed-models mailing list
> R-sig-mixed-models at r-project.org
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>
>
> End of R-sig-mixed-models Digest, Vol 8, Issue 13
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>

More information about the R-sig-mixed-models mailing list