[R-sig-ME] Redirect from R-help Digest, Vol 103, Issue 15: [R] MCMCglmm heteroscedasticity dependent on predictor

[Atle Torvik Kristiansen]

Hello,

1) Good. I'll try a proper prior and see if it makes a difference.

2) The trace plots look fine for the residuals, and they're practically
zero. The only model where the plots looked awful were for the first
order random regression (us(1+time):units). This specification was
troublesome in lme4 also, and gave peculiar error messages. The third
order random regression plots were fine and with exceptional predictive
precision. The plots for the residuals were quite similar with or
without population size (hmeanpop) in the rcov specification:

R-structure:  ~us(hmeanpop):units

                        post.mean  l-95% CI u-95% CI eff.samp
hmeanpop:hmeanpop.units 0.0005212 9.088e-05 0.001236     2900

R-structure:  ~units

      			post.mean  l-95% CI u-95% CI eff.samp
units 			0.0005273 9.229e-05 0.001254     2900

So perhaps I can drop the ~us(hmeanpop):units specification in MCMCglmm?

4) From what you're saying it seems I've run into the same problem, and
troublesome solution, I did with comparing predictors in nlme and lme4.

Gelman et al. looks appealing, although not straightforward to
implement. However, they seem to be comparing models, and indirectly
inferring the relative importance of the predictors, rather than
comparing predictors 'within' one model. Not quite what I want, but it
might prove itself a better option after a good read. Thanks for the
literature tip:)

Yes, I'm also glad your help is helpful:)

Kind regards,

Atle Torvik Kristiansen