[R-sig-eco] Random or repeated?

Dixon, Philip M [STAT] pdixon at iastate.edu
Fri Dec 9 18:23:55 CET 2011


Kyle,

The eco-stat list seems the wrong place to ask this question.  The people to ask are the quantitative plant breeders, e.g. at Texas A&M.

Some general advice about random effects and correlated observations, for either trait data or ecological data:
1) random effects provide a way to model correlations among errors.  So do 'repeated measures'.
2) what matters are the correlations among ERRORS.  If I remember the traditional QTL model correctly, errors are independent conditional on the state of the flanking markers.  So, you don't need random effects for RIL's, if you're analyzing a single trait at a time.  Your observation is correct: the random effect and the marker fixed effects are both trying to model the same thing.  The same trait measured in multiple environments is often considered as a multi-trait analysis, to estimate the correlation between pairs of environments.  It could also be analyzed as a single trait, with a model that includes fixed effects of environments.  Different models answering different questions.
3) yes, pedigrees (or relatedness coefficients) can be used to parameterize a correlation matrix.  This provides one way (not the only way) to account for phylogenetic relatedness.  Animal breeders use pedigree-based correlation matrices all the time.

Philip Dixon



More information about the R-sig-ecology mailing list