[R-sig-ME] Equivalent option to 'nobound' in SAS

Ben Bolker bbolker at gmail.com
Fri Oct 22 15:30:13 CEST 2010

On 10-10-22 09:02 AM, Carina Salt wrote:
> Hi everyone
> I'm trying to analyse a dataset (in nlme) where - within group -
> measurements are taken at two different times.  Each measurement occasion
> yields two values (one at each level of a 2-level factor, called Type) that
> are negatively correlated.  I have been handling this negative correlation
> by using
> correlation = corCompSymm (form = ~ Type | Group / Time)
> However, I've been advised by a SAS-using colleague that in SAS he would
> simply include Time as a 2-level random factor (presumably nested or crossed
> with Group) then use the the NOBOUND option in PROCMIXED to remove the
> boundary constraints on estimates and allow estimated variance parameters to
> be negative - this would handle the negative variance in a way that
> including Time as a random factor in the usual way would not.
> So my question is whether there is an equivalent of NOBOUND in R (in nlme or
> lme4 - or in any other library that does linear mixed models)?  I have
> looked at the help files but can't see anything.  Also, if so, which is
> better - the NOBOUND approach or my current approach of specifying a
> correlation structure?
> Any help would be much appreciated!  And if this question has been asked
> before I apologise, but I couldn't find anything when I searched.
> Regards
> Carrie

   I don't think this is an option in nlme or lme4.

   At the risk of sounding like one of those cranky R guys, it seems as
the solution you're using in nlme is reasonable/appropriate/mimics
the way that you actually think about the problem ("the two measurements
within a group at a particular time are negatively correlated"), while the
SAS approach is a kluge ("if I pretend one of these variances is
negative ..."). 
What are the advantages of the SAS-style negative variance approach? 
What do you mean by "...handle the negative variance in a way that
Time as a random factor in the usual way would not"?

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