[R-sig-ME] pedigreemm with one observation per individual

[David Duffy]

On Mon, 22 Feb 2010, Colin Millar wrote:

> My two pence worth: I would agree.  We have used a similar trick when 
> modelling overdispersed poisson counts.  We tried using the quasipoisson 
> family but it didn't give sensible results - it seems to return the mean 
> as the residual variance component.  On that note it would be useful if 
> the use of quasipoisson was not allowed or a warning was given for the 
> unsuspecting user.

Model identifiability is a hard problem ;) It is hard in the SEM context, 
where one has summary covariance matrices so you can count observed 
statistics and compare this to the number of model parameters, but it's 
harder still for "raw" model fitting.  But, yes, lmer() should have a 
"no checking" mode for the fool^H^H^H^Hexperienced user.

Maybe lmer needs an automatic random restart option, though such things 
often require so much specification of constraints on the permitted 
starting values that it is easier just to restart by hand.

> Aaron Rendahl wrote:
> Here is a somewhat crude fix; it adds an additional parameter
> (checklevels) to the lmer_finalize and glmer_finalize functions in
> lme4 that turns off the unwanted sanity check.  The default is
> checklevels=TRUE, which maintains the previous behavior.  The
> pedigreemm function is then changed to set checklevels=FALSE before
> calling the appropriate finalizing function.  I'm running these
> changes locally and they seem to work fine.

Another 2c, David Duffy.

-- 
| David Duffy (MBBS PhD)                                         ,-_|\
| email: davidD at qimr.edu.au  ph: INT+61+7+3362-0217 fax: -0101  /     *
| Epidemiology Unit, Queensland Institute of Medical Research   \_,-._/
| 300 Herston Rd, Brisbane, Queensland 4029, Australia  GPG 4D0B994A v