[adegenet-forum] Using PCA of SPCA in linear models with environmental data.

[Jombart, Thibaut]

Dear Hanan, 

this is a tricky question, and I don't think there is a single universal answer. Technically speaking, the only requirement is that your residuals are independent, so you need to make sure there is no spatial autocorrelation left there. Otherwise minimizing the sum of squared residuals is no longer the ML solution.

The real problem relates to the interpretation, and the assumption implicitly made by the model. There is several reasons why spatial genetic patterns can occur. Your model has the form:
genetic pattern = lat+lon + environment + residuals

Which means that beyond linear trends, genetic patterns are due to the environment. It makes sense to treat spatial autocorrelation as a confounding factor first removed from the analysis. But lat+lon is often not enough to capture all spatial structures. In this respect, using PCs from PCA on the left side is probably better than sPCA (no need to seek spatial structures to remove them afterwards).

Cheers

Thibaut

________________________________________
From: adegenet-forum-bounces at lists.r-forge.r-project.org [adegenet-forum-bounces at lists.r-forge.r-project.org] on behalf of Hanan Sela [dooshra at gmail.com]
Sent: 12 July 2012 07:34
To: adegenet-forum at lists.r-forge.r-project.org
Subject: [adegenet-forum] Using PCA of SPCA in linear models with       environmental data.

Hello all
I am trying to estimate the major factors affecting the spatial distribution of wild wheat genotypes.  I am using a linear model where the PCA or the SPCA   first and second axis are the dependent variables and the environmental variables are the predictors. Additionally I am using the longitude and the latitude as predictors.   Since there is a spatial reference on the left side of the formula, I was wondering if using SPCA on the right side will not be a problem.
Thank you
Hanan