headers
Thomas Sattler | 1 Sep 18:19
Picon

Re: glmer, quasipoisson and standard errors of the coefficients

Dear Albert Romero

I am referring to SE-part of your question:
Lately I had the same problem as you had: SE of my lme-models overlap zero 
to some extent. Of course, it could be that my variables are just bad. But 
from other analysis I knew this could not be. On a bypass (I actually 
wanted to find out something else) I had to find out that the SE-output of 
R corresponds to the usually-called Standard Deviation. (Actually R assumes 
that you are taking multiple data sets from the same population and you are 
comparing their means --> in this case the output SE is correct). Two 
senior bio-statisticians confirmed this. Look at:

G <- lm(rnorm(1000) ~rnorm(1000))
             summary(G)
             vcov(G)

vcov corresponds to the variances of the parameters, in the summary you 
find the standard errors . You will find out that standard error^2 = 
variances, thus SE ~ SD. So, with

SE/sqrt(n)

=> you obtain the SE that you want. I know, it sounds a bit strange that 
this is the problem, but step-by-step calculations convinced me...

Maybe somebody can explain it in a statistically more comprehensive way. 
Nevertheless, the outcome will be the same.

Best wishes
(Continue reading)

Permalink | Reply | 0 comments |
headers
Ben Bolker | 2 Sep 00:41
Picon

Re: glmer , quasipoisson and standard errors of the coefficients

Albert Romero wrote:
> Hello,
> I am trying to simplify backwards a mixed effects model, using lmer 
> function from lme4 package. As my data are species numbers and there 
> exists overdisperison, I think appropriate to use glmer function with 
> error family quasipoisson. I compare one model with its simplification 
> through log-likelihood ratio tests.
> Nevertheless, once I have selected a simplified model, I find in the 
> summary of this 'significant' model that estimated coefficients  are  
> associated  to  very big  standard errors, to the point that no one of 
> the coefficients seem to be significantly different from zero.
>
> Here come my questions:
> Can anybody explain this contradiction among standard errors of the 
> estimated coefficients and the significance of the model?
> Is unappropriated to use Log-likelihood backwards simplification with 
> quasipoisson errors?
>

   There are several issues here (and you should think about asking this 
question on r-sig-mixed-models , where there
is more expertise).

  1. glmer with a quasipoisson link does not provide a likelihood 
(rather, a quasilikelihood),
so you shouldn't necessarily assume that you can do *any* 
likelihood-based inference with
the results from this analysis.  The most conservative approach is to 
use only the estimated
standard errors or Z statistics on the parameters (this is a Wald test)
(Continue reading)

Permalink | Reply | 0 comments |
headers
Philip Dixon | 2 Sep 15:39
Favicon

Re: large s.e.'s for coefficients

Albert,

1) Have you looked at the correlation among your X variables?  

The situation you describe, overall model highly significant but large se's on 
each individual coefficient, often arises with multicollinearity.  Variance 
Inflation Factors (VIF's) are one common diagnostic.   If multicollinearity is 
the issue, there should be an even simpler model that fits almost as well.

2) I haven't checked how glmer computes the quasi-likelihood.  If done 
correctly, that should account for the overdispersion.  If not, it may be that 
the lnL is computed without adjusting for overdispersion but the s.e.s of the 
coefficients include overdispersion.  That will also create the behaviour you 
have encountered.

There is one theoretical issue with the use of a quasi-likelihood ratio test.  
 The quasilikelihoods for both models (full and reduced) should be computed 
using the same estimate of overdispersion, taken from the full model.  The 
usual calculations use different estimates for the two models.  That is not a 
practical problem if the two estimates are similar, as they often are.

Best wishes,
Philip Dixon
Permalink | Reply | 0 comments |
headers
Albert Romero | 4 Sep 15:47
Favicon

Re: glmer , quasipoisson and standard errors of the coefficients

Thank you all for the answers. I'll be careful with this kind of models 
till I learn how they work...

By the way, I have realized that, comparing a Poisson model  with a 
quasi-Poisson model,  the estimated  SE of the coefficients in the 
quasi-Poisson model are the SE of the Poisson model times the scale 
parameter of the quasi-Poisson model. I guess that this is related with 
Bolker's statement of dividing the LRT by the scale parameter to obtain 
the quasi-likelihood.

Kind regards.

--

-- 
Albert Romero Puente
Departament de Biologia Vegetal-Botànica
Universitat de Barcelona
Facultat de Biologia
3a Planta
Av. Diagonal, 645. (08028) Barcelona
Tel. 0034 93 402 14 71
Permalink | Reply | 0 comments |
headers
Krzysztof Sakrejda-Leavitt | 5 Sep 22:17
Picon

Multi-state models

Dear colleagues,

I am am starting work on a project using multi-state models based on 
capture-mark-recapture data to create parameter estimates for matrix 
population models.  The software the group is currently using is MSurge, 
but I am interested in packages within R which would substitute for that 
function.  I am aware of pomp and msm, but i have experience with 
neither.  Any comments on functionality and robustness in real 
applications (or in testing) would be appreciated, especially coming 
from anyone who has also worked with MSurge/MARK. 

Best,

Krzysztof

--------------------------------------
Krzysztof Sakrejda-Leavitt

Organismic and Evolutionary Biology
University of Massachusetts, Amherst
319 Morrill Science Center South
611 N. Pleasant Street
Amherst, MA 01003

work (cell): 413-325-6555
email: sakrejda@...
Permalink | Reply | 1 comment |
headers
Bret Collier | 5 Sep 23:48
Picon

Re: Multi-state models

Krzysztof,
I probably just beat Jeff to it, but you should look into RMark, it is 
an interface to MARK using R, I use it all the time and it works great, 
help is very detailed, and multi-strata models are implemented in it.

http://www.phidot.org/software/mark/rmark/

Bret

Krzysztof Sakrejda-Leavitt wrote:
> Dear colleagues,
> 
> I am am starting work on a project using multi-state models based on 
> capture-mark-recapture data to create parameter estimates for matrix 
> population models.  The software the group is currently using is MSurge, 
> but I am interested in packages within R which would substitute for that 
> function.  I am aware of pomp and msm, but i have experience with 
> neither.  Any comments on functionality and robustness in real 
> applications (or in testing) would be appreciated, especially coming 
> from anyone who has also worked with MSurge/MARK.
> Best,
> 
> Krzysztof
> 
> --------------------------------------
> Krzysztof Sakrejda-Leavitt
> 
> Organismic and Evolutionary Biology
> University of Massachusetts, Amherst
> 319 Morrill Science Center South
(Continue reading)

Permalink | Reply | 0 comments |
headers
John_Gross | 6 Sep 18:01
Favicon

John Gross/FTCOLLINS/NPS will return March 13, 2006


I will be out of the office starting  08/30/2008 and will not return until
09/08/2008.
Permalink | Reply | 0 comments |
headers
Grant Humphries | 10 Sep 20:28
Favicon

Spatial overlays

Hi,

I am wondering if anyone knows how to program spatial overlays into R, and
if anyone knows how to take that method and then automate it to run say,
1000 times for example.

Thanks,
Grant Humphries

University of Alaska Fairbanks
-EWHALE-lab Irving I  Rm 419
902 Koyukuk Dr.
Fairbanks, Alaska
99775
907-474-7959
ftgrh@...
Permalink | Reply | 2 comments |
headers
Glen A Sargeant | 10 Sep 22:18
Picon
Favicon

Re: Spatial overlays

Can you describe your data, your objective, and the issues that are 
causing trouble?

*************************************************
Glen A. Sargeant, Ph.D.
Research Wildlife Biologist/Statistician
Northern Prairie Wildlife Research Center
8711 37th Street SE
Jamestown, ND  58401

Phone: (701) 253-5528
E-mail:  glen_sargeant@...
FAX:     (701) 253-5553
*************************************************

"Grant Humphries" <ftgrh@...> 
Sent by: r-sig-ecology-bounces@...
09/10/2008 01:28 PM

To
r-sig-ecology@...
cc

Subject
[R-sig-eco] Spatial overlays

Hi,

I am wondering if anyone knows how to program spatial overlays into R, and
if anyone knows how to take that method and then automate it to run say,
(Continue reading)

Permalink | Reply | 1 comment |
headers
Kingsford Jones | 11 Sep 18:14
Picon

Re: Spatial overlays

Grant,

I'm not sure what you a looking to do, but this thread may be of use:

https://stat.ethz.ch/pipermail/r-sig-ecology/2008-May/000121.html

Something to add to that thread is this recent book which covers many
of the spatial tools available in R:

http://www.amazon.com/Applied-Spatial-Data-Analysis-Use/dp/0387781706/ref=pd_bbs_sr_1?ie=UTF8&s=books&qid=1221148754&sr=8-1

Also, this is a useful guide for finding answers to R questions and/or
eliciting useful responses on R mailing lists:

http://www.r-project.org/posting-guide.html

best,

Kingsford Jones

On Wed, Sep 10, 2008 at 2:18 PM, Glen A Sargeant <gsargeant@...> wrote:
> Can you describe your data, your objective, and the issues that are
> causing trouble?
>
> *************************************************
> Glen A. Sargeant, Ph.D.
> Research Wildlife Biologist/Statistician
> Northern Prairie Wildlife Research Center
> 8711 37th Street SE
> Jamestown, ND  58401
(Continue reading)

Permalink | Reply | 0 comments |

Gmane