Dear Prof. Ripley,
Thank you for your quick response.
(A)
> link-sqrt is a name and not accepted. link="sqrt" is a literal character
> string, and is.
I am not entirely sure whether I understand that statement but this is what I
found out. If I specify family=gaussian(link=sqrt), the glm() fails to run
because it is not a default link (so, I understand this part). Following
Venables and Ripley (2002):
> summary(glm(cnt~herbc+herbht,data=sotr,family=gaussian(link="sqrt"),start=c(0.1,-0.004,0.01)))
Call:
glm(formula = cnt ~ herbc + herbht, family = gaussian(link = "sqrt"),
data = sotr, start = c(0.1, -0.004, 0.01))
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.462211 0.043475 10.632 < 2e-16 ***
herbc -0.003315 0.001661 -1.996 0.0461 *
herbht 0.010241 0.001291 7.935 4.86e-15 ***
AIC: 3235.0
> summary(glm(cnt~herbc+herbht,data=sotr,family=quasi(link=power(0.5),variance=constant),start=c(0.1,-0.004,0.01)))
Call:
glm(formula = cnt ~ herbc + herbht, family = quasi(link = power(0.5)),
data = sotr, start = c(0.1, -0.004, 0.01))
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.462211 0.043475 10.632 < 2e-16 ***
herbc -0.003315 0.001661 -1.996 0.0461 *
herbht 0.010241 0.001291 7.935 4.86e-15 ***
AIC: NA
Notice that the parameter estimates and corresponding standard errors are
identical. So, my interpretation is that family=gaussian(link="sqrt") is
identical as specify family=quasi(link=power(0.5)) in glm(). The exception is
that AIC (and thus maximized log-likelihood values) can be computed for
family=gaussian(link="sqrt").
The questions are:
(A.1) Is this interpretation correct?
(A.2) If (A.1) is true, does family=gaussian(link="sqrt") implies that I am
doing a Generalized Linear Model with normal distribution and the link function
is: sqrt(mu) = b0+b1(herbc)+b2(herbht)?
(B)
> In less technical terms, in model 1 you compute the likelihood from
> probabilities
> and in model 2 from probability densities, and the latter depend on the
> units of measurement.
Yes, you are correct and I understand it now. Although not as common these
days, some small mammal studies still use sqrt transformation of count as
response variable and carry out a linear model fitting with predictors (via
least squares). So, the exercise that I got into is to compare performances of
linear model with sqrt transformation of count and GLM with Poisson. However,
knowing that we can't compare logLik or AIC based on different measures of
responses. So, I thought that comparison under GLM framework might be an
approach closer to the intention.
Thank again for your quick respond and advices. I appreciate it very much.
Best regards,
TzengYih Lam
-----------------------------------
Ph.D. student
College of Forestry
Oregon State University
-----Original Message-----
From: Prof Brian Ripley [mailto:[EMAIL PROTECTED]
Sent: Wed 12/10/2008 11:45 PM
To: Lam, Tzeng Yih
Cc: r-help@r-project.org
Subject: Re: [R] Validity of GLM using Gaussian family with sqrt link
a) There is a difference between link=sqrt and link="sqrt".
link: a specification for the model link function. This can be a
name/expression, a literal character string, a length-one
character vector or an object of class '"link-glm"' (such as
generated by 'make.link') provided it is not specified _via_
one of the standard names given next.
link-sqrt is a name and not accepted. link="sqrt" is a literal character
string, and is.
b) Your first model is a model for integer observations, the second for
continuous observations. As such, the log-likleihoods are computed with
respect to different reference measures and are not comparable. In less
technical terms, in model 1 you compute the likelihood from probabilities
and in model 2 from probability densities, and the latter depend on the
units of measurement.
On Wed, 10 Dec 2008, Lam, Tzeng Yih wrote:
> Dear all,
>
> I have the following dataset: each row corresponds to count of forest floor
> small mammal captured in a plot and vegetation characteristics measured at
> that plot
>
>> sotr
> plot cnt herbc herbht
> 1 1A1 0 37.08 53.54
> 2 1A3 1 36.27 26.67
> 3 1A5 0 32.50 30.62
> 4 1A7 0 56.54 45.63
> 5 1B2 0 41.66 38.13
> 6 1B4 0 32.08 37.79
> 7 1B6 0 33.71 30.62
> ...
>
> I am interested in comparing fit of different specification of
> Generalized Linear Models (although there are some issues with using AIC
> or BIC for comparison, but this is the question that I like to post
> here). Here are two of the several models that I am interested in:
>
> (1) Poission log-linear model
>> pois<-glm(cnt~herbc+herbht,family=poisson,data=sotr)
>> summary(pois)
> Call:
> glm(formula = cnt ~ herbc + herbht, family = poisson, data = sotr)
>
> Coefficients:
> Estimate Std. Error z value Pr(>|z|)
> (Intercept) -1.341254 0.089969 -14.908 <2e-16 ***
> herbc -0.007303 0.003469 -2.105 0.0353 *
> herbht 0.024064 0.002659 9.051 <2e-16 ***
> ---
> Null deviance: 1699.0 on 1180 degrees of freedom
> Residual deviance: 1569.8 on 1178 degrees of freedom
> AIC: 2311.4
>
>
> (2) Gaussian with sqrt link model
>> gaus.sqrt<-glm(cnt~herbc+herbht,family=gaussian(link="sqrt"),data=sotr,start=c(0.1,-0.004,0.01))
>> summary(gaus.sqrt)
> Call:
> glm(formula = cnt ~ herbc + herbht, family = gaussian(link = "sqrt"),
> data = sotr, start = c(0.1, -0.004, 0.01))
>
> Coefficients:
> Estimate Std. Error t value Pr(>|t|)
> (Intercept) 0.462211 0.043475 10.632 < 2e-16 ***
> herbc -0.003315 0.001661 -1.996 0.0461 *
> herbht 0.010241 0.001291 7.935 4.86e-15 ***
> ---
> Null deviance: 1144.6 on 1180 degrees of freedom
> Residual deviance: 1062.9 on 1178 degrees of freedom
> AIC: 3235.0
>
>> logLik(gaus.sqrt)
> 'log Lik.' -1613.524 (df=4)
>
>> From the glm() help file that I read, family=gaussian() accepts the links
>> "identity", "log" and "inverse". There is no mentioning of gaussian()
>> accepting "sqrt" link. Although "sqrt" link is available for family=poisson()
>
> A. Therefore, is the code in (2) actually computing Maximum Likelihood
> Estimates (MLE) of the coefficients using Gaussian family with "sqrt"
> link or is it computing MLE of something else?
>
> B. If the code in (2) is computing the MLE with gaussian(link="sqrt"),
> then will the maximized value of log-likelihood function using logLik()
> be valid (other than the issue that the dispersion parameter is counted
> as a parameter in aic() within glm())?
>
> Thank you in advance and I appreciate it very much for any advices that are
> offered.
>
> Best regards,
> TzengYih Lam
>
>
> TzengYih Lam, PhD Student
> College of Forestry
> Oregon State University
>
>
>
>
>
>
>
>
> [[alternative HTML version deleted]]
>
> ______________________________________________
> R-help@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, [EMAIL PROTECTED]
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
[[alternative HTML version deleted]]
______________________________________________
R-help@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.
