[R-sig-Geo] your mail

[Roger Bivand]

On Sun, 8 Aug 2010, Sabira el Messlaki wrote:

> Dear Roger,
>
> What do you mean by using correlation= argument to lme() to fit a model
> including within-group correlations?

You do need to read the help page to look for the specification og 
"argument". Reading the book refered to on the help page may be a good 
idea too.

>
> I tried it like this: model3.fit <- lme(y ~ 
> x3+x5+x7+x9+LnBankt500+Perceelkwaliteit+LnNS_Afstand+PGebouwd+LogoBedrijf+Bouwjaar_cat4+LnBanIndDistr500, 
> random = ~1|GebID ,correlation=argument, datashab, method = "REML")
>
> But this does not seem to work. What is the reason that nobody developed 
> a moran's I for an lme object?
>

You have specified a random group effect. Consequently, the standard 
OLS relationship between the projection matrix of the included variables 
and the residuals does not hold. You can calculate Moran's I, but there is 
no theory to tell you whether it is significant or not. The formulae for 
the expectation and variance of the statistic need changing, and if they 
are not changed, the results will be misleading.

Roger

>
>
>
> -----Oorspronkelijk bericht-----
> Van: Roger Bivand [mailto:Roger.Bivand at nhh.no]
> Verzonden: vr 6-8-2010 15:30
> Aan: Sabira el Messlaki
> Onderwerp: Re: your mail
>
> On Fri, 6 Aug 2010, Sabira el Messlaki wrote:
>
>>
>> Dear Roger,
>
>> How do I test the Moran's I in residuals of an lme object. I have done a
>> multilevel analysis and now I want to know if there is still spatial
>> autocorrelation.
>
> Briefly, you have the opportunty to develop a test from the ground up
> yourself, as there are only tests for linear model residuals available. If
> you want to try to use a regular Moran's I on the residuals data frame
> columns from the fitted model, the significance tests will most likely be
> wrong.
>
> Could you try using the correlation= argument to lme() to fit a model
> including within-group correlations, typically based on inter-observation
> distance? If the fit of lme() with and without this argument is the same,
> it would suggest that there is no residual within-group autocorrelation
> for your correlation scheme.
>
> Hope this helps,
>
> Roger
>
>>
>> Thank you in advance
>>
>> Sabira
>>
>
>

-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Helleveien 30, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 95 43
e-mail: Roger.Bivand at nhh.no