Dear collegues,
Hear is the answer to my email... (I send it before, but we had anemail problem and it did not go. Sory by the delay) thank you all Best wishes Marta At 10:07 03/10/02 +0200, you wrote: >Dear collegues, > >I have another doubt.... >When we have several possible models (with different covariates as an >external trend), how doi we decide which one is the most proper one to use >(doing the variograms and kriging). >Do we simply look at the value of the minimising function (the lowest would >be the best)? >Do we look at percentage of the spatial variance explained by the model? >Do we look at the lowest C (sill+nugget)? > >Can you help me? >thank you >Marta Hi Marta, if you have several possible covariates it is worth you while selecting the best correlated. This can be a statistical measure or in the case of external drift a clear physical relationship (for example rainfall and elevation are often linearly related, but a clear statistical relationship may not be observable). Be parsimonious, use as few variables as possible. In any case the external drift (or trend) should have a spatial structure that is smoother than the variable of interest. Hans Wackernagels book, "Multivariate Geostatistics" provides a good overview, Then of course there are validation tests that can help you choose such as X-validation, but a rationale choice in the first instance can save a lot of time testing different models. Benjamin Warr Research Fellow Centre for the Management of Environmental Resource(CMER) INSEAD Boulevard de Constance, 77305 Fontainebleau Cedex, France Marta You can use the Cressie goodness of fit statistic to assess relative merits of various models. You should also look at the cross validation statistics to see which gives you closest to the ideal behaviour and best correlation between estimated and actual values. Better estimates (in the real sense) are achieved with models in which the nugget effect/sill ratio is a minimum and where the range of influence is longest (in that order of priority). The total sill is virtually irrelevant in determining your kriging weights and only affects the kriging variance as a constant factor. Unless, of course, you are doing lognormal kriging where the total sill is vital to the back transform. Isobel Clark http://uk.geocities.com/drisobelclark __ Estimada Marta: Te escribo fuera de la lista para hacerlo en espa�ol, lo cual creo que puede ser m�s �gil para ambos. Para la elecci�n de un modelo determinado no debes basarte en �ndices estad�sticos. La geoestad�stica no es la aplicaci�n de la estad�stica a diversos campos. Por ello, debes elegir un modelo que se ajuste bien a tus datos y que expliquen razonablemente la posible distribuci�n de los mismos. Para ello debes realizar consideraciones subjetivas (elegir un modelo que se ajuste m�s o menos a tus datos) y basarte en la experiencia previa (si existe). O sea, si en un trabajo anterior un modelo se ajust� bien a los datos, no hay raz�n para considerar otro modelo distinto. Otra cosa que puede ser �til es conocer si la distribuci�n de la variable estudiada es m�s o menos continua. En ese caso es conveniente ajustar un tipo de modelo u otro (gaussiano o esf�rico, por ejemplo). Si te interesa, tengo en prensa un libro sobre geoestad�stica lineal, el cual podr�a enviarte v�a mail. Un saludo. Dear Marta, Have a look at reduced mean error and reduced variance. In my opinion, these are indeed good indicators of approppriate models. By the way, I would like to learn if you have GS+ software. If you have, what is the version? Looking forward to hearing from you, soon. Yours Sincerely, Dr. Mahmut CETIN Hi marta If you fit the models usig maximum likelihood you can use the AIC or BIC as a criteria to choose the model. AIC: Akaike information criteria BIC: Baeysian information criteria The function likfit() in geoR returns them both Cheers P.J. I don't think there is going to be an absolute answer to your question. However I suggest that you look at the following: 1. Do you have data at the same number of locations for all of the possible covariates? (more data locations is better) 2. The match between data locations for possible covariates and data locations for the principal variable. 3. Standard regression diagnostics when fitting the principal variable to each of the possible covariates 4. Since you are going to be using the "residuals" for the principal variable to estimate and model the variogram (or covariance), how does the fitting compare using each of the different possible covariates? For example, there are multiple cross-validation statistics you can use 5. Of course you also want to consider which of the different possible covariates makes better sense , i.e., is the relationship between the possible covariate and the principal variable simply one of correlation or is there evidence or theory or reason to believe that there is something closer to a causual relationship. 6. Are the possible covariates interdependent? Donald E. Myers http://www.u.arizona.edu/~donaldm > ><((((�>`�.��.���`�.�.���`�...�><((((�>`�.��.���`�.�.���`�...�><((((�> `�.��.���`�.�.���`�...�><((((�>`�.��.���`�.�.���`�...�><((((�>`�.��.�� Marta Rufino Centre Mediterrani d'Investigacions Marines i Ambientals (CMIMA). CSIC Passeig Maritim 37-49 08003 BARCELONA Tfno:34 93 230 95 40 Tfax:34 93 230 95 55 ><((((�>`�.��.���`�.�.���`�...�><((((�>`�.��.���`�.�.���`�...�><((((�> `�.��.���`�.�.���`�...�><((((�>`�.��.���`�.�.���`�...�><((((�>`�.��.�� -- * To post a message to the list, send it to [EMAIL PROTECTED] * As a general service to the users, please remember to post a summary of any useful responses to your questions. * To unsubscribe, send an email to [EMAIL PROTECTED] with no subject and "unsubscribe ai-geostats" followed by "end" on the next line in the message body. DO NOT SEND Subscribe/Unsubscribe requests to the list * Support to the list is provided at http://www.ai-geostats.org
