[ADMB Users] difference between ADMB-RE and R/mgcv in SEs for smoother coefficients in a GAM fitted by maximum likelihood

[H. Skaug]

In my view this is empirical Bayes where x is treated in the
frequentist way and u in the Bayesian. That is straight forward
for point estimation, but creates some confusion (as we
now see) for the uncertainty part. I have always viewed your
variance formula as resulting from "going Bayesian
also about x", which I thought was a nice solution. If
people want something more accurate than the formula
they can use -mcmc, so then it becomes important
to be philosofically consistent within. (I just noticed
the email about the sign problem).

Hans

On Thu, Nov 29, 2012 at 10:21 PM, dave fournier <davef at otter-rsch.com> wrote:
> On 12-11-29 01:16 PM, H. Skaug wrote:
>
> Not if you are a frequentist.  Then there is no u just uhat(x).  I don't
> belong to churches.
> However the frequentist approach produces numbers much closer to what the OP
> had.