(apologies - I should have written coast * MBL not ML) I'm not sure of my ground here, but surely do lose something - you wouldn't retain coast:MBL if it's not significant, as you lose degrees of freedom, and this gets worse the more terms and the more interactions you consider. I think it's a different issue with the random effects, I can see a case for retaining a random effect on design grounds even though it technically might not look significant, but I'm not so sure for fixed effects. On that basis wouldn't we always be fitting the indecipherable A*B*C*D instead of A+B+C+D, even if the additive effects are adequate? Mike
"hadley wickham" <h.wickham at gmail.com> 10/11/2008 12:59 >>>
The coast * ML term tests for HSP high/low dependent on coast. To test this fit the full model with method = ML and compare it to lme(HSP~coast+MBL, random= ~1|site, method ="ML") using anova(model1, model2). There are alot of technical issues with testing both fixed and random effects in mixed models, for details see past posts on the R-sig-ME list and also on the R Wiki (http://wiki.r-project.org/rwiki/doku.php?id=guides:lmer-tests). But lets ignore those, this should do OK. The if coast*ML is significant then no need to go any further. If it isn't then repeat from the coast+MBL, deleting one of those fixed terms and repeating.
I thought that this was generally a bad idea. You don't lose anything by keeping the non-significant terms in the model, but if you drop them out you can falsely inflate the significance of other terms. Haley -- http://had.co.nz/
This message (and any attachments) is for the recipient ...{{dropped:6}}