[R-sig-ME] Difference of lme4 between the CRAN version and the development version
Joehanes, Roby (NIH/NHLBI) [F]
roby.joehanes at nih.gov
Thu Dec 6 22:48:01 CET 2012
Yes, I do. Thanks for the suggestion. Unfortunately, the options do not help---the error persists:
In updateXwts(Xwts) :
Cholmod warning 'not positive definite' at file:../Cholesky/t_cholmod_rowfac.c, line 431
The model I am running is:
y ~ Sex + Age + (...other fixed covariates...) + (1|patient_id)
where there is one observation per patient. I use pedigreemm to attach the pedigree information to the patient_id.
I understand that this model is an over-specified model because lme4 will normally throw an error. I bypassed that error by commenting the if statement that checks for it. The CRAN version of lme4 seems to run this model no problem, but the development lme4 seems to choke with CholMod complaint. Strangely enough, if I modify the model to:
y ~ Sex + Age + (...other fixed covariates...) + (1|Batch) + (1|patient_id)
then the model runs fine in both the CRAN version and the development version of lme4, to my surprise. To me, this suggests that the development version would not simply choke on over-specified model. I am thinking if there are any underlying technical problems behind the errors that I could somehow "fix" or get around.
I have read the theory and computational method for the development version of lme4. I found chapter 5 to be particularly helpful---the development lme4 also accommodates learners like me by having the parameters and variables named closely to the theory. I am in the process of reconstructing the deviance function in pure R for my learning purposes. I am digging for similar information in the CRAN version to do the same, but I have not been able to find some similar tutorial just yet. I hope that understanding the subtle nuances of the two implementations would shed lights toward the "fix".
On 12/6/12 4:18 PM, "laurent stephane" <laurent_step at yahoo.fr> wrote:
With the development version do you use the options optimizer="bobyqa" and control=list(restart=TRUE) when running lmer() ? Sometimes it solves some troubles.
De : "Joehanes, Roby (NIH/NHLBI) [F]" <roby.joehanes at nih.gov>
À : "r-sig-mixed-models at r-project.org" <r-sig-mixed-models at r-project.org>
Envoyé le : Jeudi 6 décembre 2012 19h09
Objet : [R-sig-ME] Difference of lme4 between the CRAN version and the development version
I am wondering if anyone could explain to me the difference of lme4 that is in the CRAN and of the development version. The mathematics and the method to compute the estimates for the development version are covered in the following document:
However, I do not know any such manual for the CRAN version. Could anyone point to me to the right direction? Are the vignette documents in CRAN the right ones ( http://cran.r-project.org/web/packages/lme4/index.html )?
With this information, I hope I could troubleshoot the errors I am encountering with it and pedigreemm, primarily the dreaded Cholmod error of "Downdated X'X is not positive definite" that I am getting with the development version, but not with the CRAN version).
More information about the R-sig-mixed-models