[R] Question on WLS (gls vs lm)

[Stats Wolf]

Indeed, they should give the same results, and hence I was worried to
see that the results were not that same. Suffice it to look at
standard errors and p-values: they do differ, and the differences are
not really that small.


Thanks,
Stats Wolf



On Thu, Jun 24, 2010 at 2:39 PM, Joris Meys <jorismeys at gmail.com> wrote:
> Indeed, WLS is a special case of GLS, where the error covariance
> matrix is a diagonal matrix. OLS is a special case of GLS, where the
> error is considered homoscedastic and all weights are equal to 1. And
> I now realized that the varIdent() indeed makes a diagonal covariance
> matrix, so the results should be the same in fact. Sorry for missing
> that one.
>
> A closer inspection shows that the results don't differ too much. The
> fitting method differs between both functions; lm.wfit uses the QR
> decomposition, whereas gls() uses restricted maximum likelihood. In
> Asymptopia, they should give the same result.
>
> Cheers
> Joris
>
>
> On Thu, Jun 24, 2010 at 12:54 PM, Stats Wolf <stats.wolf at gmail.com> wrote:
>> Thanks for reply.
>>
>> Yes, they do differ, but does not gls() with the weights argument
>> (correlation being unchanged) make the special version of GLS, as this
>> sentence from the page you provided says: "The method leading to this
>> result is called Generalized Least Squares estimation (GLS), of which
>> WLS is just a special case"?
>>
>> Best,
>> Stats Wolf
>>
>>
>>
>> On Thu, Jun 24, 2010 at 12:49 PM, Joris Meys <jorismeys at gmail.com> wrote:
>>> Isn't that exactly what you would expect when using a _generalized_
>>> least squares compared to a normal least squares? GLS is not the same
>>> as WLS.
>>>
>>> http://www.aiaccess.net/English/Glossaries/GlosMod/e_gm_least_squares_generalized.htm
>>>
>>> Cheers
>>> Joris
>>>
>>> On Thu, Jun 24, 2010 at 9:16 AM, Stats Wolf <stats.wolf at gmail.com> wrote:
>>>> Hi all,
>>>>
>>>> I understand that gls() uses generalized least squares, but I thought
>>>> that maybe optimum weights from gls might be used as weights in lm (as
>>>> shown below), but apparently this is not the case. See:
>>>>
>>>> library(nlme)
>>>> f1 <- gls(Petal.Width ~ Species / Petal.Length, data = iris,  weights
>>>> = varIdent(form = ~ 1 | Species))
>>>> aa <- attributes(summary(f1)$modelStruct$varStruct)$weights
>>>> f2 <- lm(Petal.Width ~ Species / Petal.Length, data = iris, weights = aa)
>>>>
>>>> summary(f1)$tTable; summary(f2)
>>>>
>>>> So, the two models with the very same weights do differ (in terms of
>>>> standard errors). Could you please explain why? Are these different
>>>> types of weights?
>>>>
>>>> Many thanks in advance,
>>>> Stats Wolf
>>>>
>>>> ______________________________________________
>>>> 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.
>>>>
>>>
>>>
>>>
>>> --
>>> Joris Meys
>>> Statistical consultant
>>>
>>> Ghent University
>>> Faculty of Bioscience Engineering
>>> Department of Applied mathematics, biometrics and process control
>>>
>>> tel : +32 9 264 59 87
>>> Joris.Meys at Ugent.be
>>> -------------------------------
>>> Disclaimer : http://helpdesk.ugent.be/e-maildisclaimer.php
>>>
>>
>
>
>
> --
> Joris Meys
> Statistical consultant
>
> Ghent University
> Faculty of Bioscience Engineering
> Department of Applied mathematics, biometrics and process control
>
> tel : +32 9 264 59 87
> Joris.Meys at Ugent.be
> -------------------------------
> Disclaimer : http://helpdesk.ugent.be/e-maildisclaimer.php
>