[R] Weird SEs with effect()

[Prof Brian Ripley]

On Sun, 17 Feb 2008, Gustaf Granath wrote:

> Hi John,
>
> In fact I am still a little bit confused because I had read the
> ?effect help and the archives.
>
> ?effect says that the confidence intervals are on the linear predictor
> scale as well. Using exp() on the untransformed confidence intervals
> gives me the same values as summary(eff). My confidence intervals
> seems to be correct and reflects the results from my glm models.
>
> But when I use exp() to get the correct SEs on the response scale I
> get SEs that sometimes do not make sense at all. Interestingly I have

What exactly are you doing here?  I suspect you are not using the correct 
formula to transform the SEs (you do not just exponeniate them), but 
without the reproducible example asked for we cannot tell.

> found a trend. For my model with adjusted means ~ 0.5-1.5 I get huge
> SEs (SEs > 1, but my glm model shows significant differences between
> level 1 = 0.55 and level 2 = 1.15). Models with means around 10-20 my
> SEs are fine with exp(). Models with means around 75-125 my SEs get
> way too small with exp().
>
> Something is not right here (or maybe they are but I don not
> understand it) so I think my best option will be to use the confidence
> intervals instead of SEs in my plot.

If you want confidence intervals, you are better off computing those on a 
reasonable scale and transforming then.  Or using a profile likelihood to 
compute them (which will be equivariant under monotone scale 
transformations).

> Regards,
>
> Gustaf
>
>
>> Quoting John Fox <jfox at mcmaster.ca>:
>>
>> Dear Gustaf,
>>
>> From ?effect, "se: a vector of standard errors for the effect, on the scale
>> of the linear predictor." Does that help?
>>
>> Regards,
>>  John
>>
>> --------------------------------
>> John Fox, Professor
>> Department of Sociology
>> McMaster University
>> Hamilton, Ontario, Canada L8S 4M4
>> 905-525-9140x23604
>> http://socserv.mcmaster.ca/jfox
>>
>>
>>> -----Original Message-----
>>> From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-
>>> project.org] On Behalf Of Gustaf Granath
>>> Sent: February-16-08 11:43 AM
>>> To: r-help at r-project.org
>>> Subject: [R] Weird SEs with effect()
>>>
>>> Hi all,
>>>
>>> Im a little bit confused concerning the effect() command, effects
>>> package.
>>> I have done several glm models with family=quasipoisson:
>>>
>>> model <-glm(Y~X+Q+Z,family=quasipoisson)
>>>
>>> and then used
>>>
>>> results.effects <-effect("X",model,se=TRUE)
>>>
>>> to get the "adjusted means". I am aware about the debate concerning
>>> adjusted means, but you guys just have to trust me - it makes sense
>>> for me.
>>> Now I want standard error for these means.
>>>
>>> results.effects$se
>>>
>>> gives me standard error, but it is now it starts to get confusing. The
>>> given standard errors are very very very small - not realistic. I
>>> thought that maybe these standard errors are not back transformed so I
>>> used exp() and then the standard errors became realistic. However, for
>>> one of my glm models with quasipoisson the standard errors make kind
>>> of sense without using exp() and gets way to big if I use exp(). To be
>>> honest, I get the feeling that Im on the wrong track here.
>>>
>>> Basically, I want to know how SE is calculated in effect() (all I know
>>> is that the reported standard errors are for the fitted values) and if
>>> anyone knows what is going on here.
>>>
>>> Regards,
>>>
>>> Gustaf Granath
>>>
>>> ______________________________________________
>>> R-help at r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-help
>>> PLEASE do read the posting guide http://www.R-project.org/posting-
>>> guide.html
>>> and provide commented, minimal, self-contained, reproducible code.
>>
>>
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595