[R-sig-ME] Bleeding edge lme4 (or lme4a) plus DF estimation

[Joehanes, Roby (NIH/NHLBI) [F]]

Hi Ben,

On Mar 1, 2012, at 2:05 PM, Ben Bolker wrote:

> Joehanes, Roby (NIH/NHLBI) [F] <roby.joehanes at ...> writes:
> 
>> I discovered a bug on your lme4Eigen's refit function. 
>> This is on version 0.9996875-9 (Description
>> revision 169). I hope I got this right. If the original data
>> matrix has some missing data in it, somehow the X
>> and Z matrices (and y column) are correctly trimmed (i.e., the rows
>> with missing data are removed).
>> However, if I fit it with another y column, it reports error due 
>> to length mismatch. The error is as follows:
>> Error: length(newresp <- as.numeric(as.vector(newresp))) == length(rr$y) 
>> is not TRUE
>> 
>> Should I try the latest version and see if this bug has been fixed?
> 
>  It probably hasn't.  It would be very helpful if you could send a small
> self-contained example to lme4-authors <at> r-forge.wu-wien.ac.at --
> we could probably make one up ourselves, but it would be quicker/
> more motivational if you did it.

Unfortunately, my data is considered classified. I am not authorized to give one out. I hope I can compose an example real soon.

>> 
>> Also, will update method speed up computation
>> if I changed the X (or Z) matrix a little bit by swapping or
>> adding up to three columns (from about 40+ columns)? 
> 
>  Probably not -- refit saves time by (1) starting from previous
> starting values and (2) not having to rebuild X and Z components.
> You could do #1 yourself by doing something like setting
> 
> update(...,start=getME(prevfit,"theta"),...)
> 
> (I think).

Thank you. I will try it out.

Sincerely,
Roby