[R-sig-ME] p-value absense in output lmer
bbolker at gmail.com
Mon Dec 26 19:52:37 CET 2011
Simon Blomberg <s.blomberg1 at ...> writes:
> > Q2: Does the fact that many of those numbers are above 0.95 worry
> > anybody else? I would not have expected such high correlations in
> > "real biological data".
> Not really. If the data are not centred, then you expect big
> correlations between the slopes and intercepts. It can help to centre
> the data for numerical reasons, but it shouldn't affect the inferences.
> Or am I wrong? Please correct me, someone!
I believe that's correct. I don't always centre by default,
but it's worth doing (1) for computational purposes, especially
when getting convergence warnings etc.; (2) for inferential
purposes, especially when fitting models with interactions
(i.e. with interactions present, the main effects of parameters will
be estimated and reported at the 'zero' level of any continuous
So far, I haven't seen an example where the actual fitting went
wrong (silently) because of undue correlations among the input
variables. Very large correlations (|rho|>0.99) might suggest
identifiability problems. There is a whole literature on dealing
with collinear predictors (which is a subset of those that
can give rise to correlated parameters -- see Zuur et al 2009
doi: 10.1111/j.2041-210X.2009.00001.x, but it's a delicate subject
(in my opinion) depending on the goal of your analysis and the
kinds of errors you're willing to subject yourself to.
More information about the R-sig-mixed-models