[R-sig-ME] Questions on migrating from nlme to lme4

Martin Maechler maechler at stat.math.ethz.ch
Fri Jun 22 18:05:29 CEST 2007

>>>>> "DM" == Dieter Menne <dieter.menne at menne-biomed.de>
>>>>>     on Fri, 22 Jun 2007 09:28:33 +0000 (UTC) writes:

    DM> Douglas Bates wrote:
    >> Generally I recommend using mcmcsamp to produce "confidence intervals"
    >> on the variance components.  Approximate (and symmetric) confidence
    >> intervals on the fixed effects are reasonably accurate but such
    >> intervals for the random effects can be poor approximations.

    DM> Problem is that referees who don't read the regular Douglas B columns tend to
    DM> say "mcmc ... ha?", and, after explanation, 'we do not publish poker games' (<-
    DM> slightly paraphrased from the original comment).

Hmm, we are getting off-topic here, but I think these referees
are not quite fit for the 21st century.

MANY modern statistical procedures depend on random numbers to
some extent:

- neural nets  solve a high dimensional minimization and the
  solution depends on the random starting values.
  {Some silly people would therefore always use the same random
  seed before starting the nnet}

- The good old K-Means algorithm very often starts with random
  centers {and again: people use versions of the algorithm that, e.g.,
  	  always start with same indices of observations as starting
	  values ==> their algorithm depends on the *ordering*
	  of the observations --- which I think is worse}
- All high-breakdown robust statistics procedures ...

- All K-fold cross-validation  ...

- All bootstrapping / bagging / bragging / ...

depend on random (sub)sampling.

If referees ask researchers to refrain from all such methods, 
a good journal editor should switch referees,
or a good author should switch to a better journal  :-)


More information about the R-sig-mixed-models mailing list