[R-sig-ME] Questions about design and convergence warnings
Dube, Umber
udube at wustl.edu
Thu Sep 28 16:58:12 CEST 2017
Thanks for your continued development of lme4 and all the support you've provided.
This is my first mixed model analysis. I've done my best to read over the past messages and think I've found a proper method for performing it, but I would like to verify that is correct.
I'm interested in performing a generalized linear mixed model analysis on RNA-sequencing data from different tissues derived from the same organ (I have 4 different tissues from ~160 diseased organs, ~60 healthy organs).
I have been modelling the following as fixed effects:
RNA Integrity Number (RIN) - quality of the total RNA extracted from each tissues (continuous)
Post-mortem Interval (PMI) - how much time elapsed following death until the tissue was frozen (continuous)
Sex - genetic sex of the organ (categorical)
Age at death (AOD) - age of organ at death (continuous)
Species - species of organ (categorical)
Gene - normalized count data of gene expression (continuous)
I have been modelling the following as random effects:
Batch (categorical)
I understand that I have a nested model Organ/Tissue, but after reading (https://stackoverflow.com/questions/19414336/using-glmer-for-nested-data), I modeled tissue as a fixed effect due to the small numbers (4 tissues).
Altogether, my model is:
glmer(Disease ~ RIN + SEX + AOD + PMI + Species + (1|batch) + Tissue + (1|Tissue:Organ) + Gene, data=NormDEGenes, family=binomial(), control=glmerControl(optCtrl=list(maxfun=1e9) ) )
I unfortunately get convergence warnings with these models, but after going through the ?convergence documentation I hope they are false positives.
Warning messages:
1: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, :
unable to evaluate scaled gradient
2: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, :
Model failed to converge: degenerate Hessian with 1 negative eigenvalues
To address this:
1) I've centered and scaled my continous predictors
2) I've checked for singularity #False
3) I've printed and compared with internal calculations
4) I've tried all available optimizers. I believe all of them have failed to converge, but they all end up with approximately the same log-liklihoods.
> ss$ fixef ## extract fixed effects
(Intercept) RIN_scale SEXF AOD_scale PMI_scale Species1 Species2 Species3 Tissue1 Tissue2 Tissue3 Gene
bobyqa 0.11463 -0.41005 0.409846 -0.07834 -0.65198 14.79972 13.30085 14.88501 -0.35383 -0.31039 0.666657 -2.42662
Nelder_Mead -0.10894 -0.41005 0.40988 -0.07834 -0.652 15.02298 13.52411 15.10837 -0.35371 -0.31033 0.666808 -2.42659
nlminbw 0.067 -0.41005 0.409846 -0.07834 -0.65198 14.84728 13.34841 14.93257 -0.35383 -0.3104 0.666653 -2.4266
optimx.L-BFGS-B 0.066515 -0.40992 0.40951 -0.07822 -0.65189 14.84686 13.34846 14.93227 -0.35369 -0.3102 0.666332 -2.42644
nloptwrap.NLOPT_LN_NELDERMEAD -0.19423 -0.40082 0.401142 -0.07745 -0.63582 14.74898 13.28662 14.83122 -0.3464 -0.30534 0.646181 -2.36802
nloptwrap.NLOPT_LN_BOBYQA -0.19423 -0.40082 0.401142 -0.07745 -0.63582 14.74898 13.28662 14.83122 -0.3464 -0.30534 0.646181 -2.36802
> ss$ llik ## log-likelihoods
bobyqa Nelder_Mead nlminbw optimx.L-BFGS-B nloptwrap.NLOPT_LN_NELDERMEAD nloptwrap.NLOPT_LN_BOBYQA
-327.896 -327.896 -327.896 -327.896 -327.933 -327.933
> ss$ sdcor ## SDs and correlations
Organ:Tissue.(Intercept) batch.(Intercept)
bobyqa 3.68E-05 0.520826
Nelder_Mead 5.56E-03 0.52084
nlminbw 3.10E-08 0.520826
optimx.L-BFGS-B 0.00E+00 0.52084
nloptwrap.NLOPT_LN_NELDERMEAD 4.56E-08 0.517555
nloptwrap.NLOPT_LN_BOBYQA 4.56E-08 0.517555
> ss$ theta ## Cholesky factors
Organ:Tissue.(Intercept) batch.(Intercept)
bobyqa 3.68E-05 0.520826
Nelder_Mead 5.56E-03 0.52084
nlminbw 3.10E-08 0.520826
optimx.L-BFGS-B 0.00E+00 0.52084
nloptwrap.NLOPT_LN_NELDERMEAD 4.56E-08 0.517555
nloptwrap.NLOPT_LN_BOBYQA 4.56E-08 0.517555
> ss$ which.OK ## which fits worked
bobyqa Nelder_Mead nlminbw nmkbw optimx.L-BFGS-B nloptwrap.NLOPT_LN_NELDERMEAD nloptwrap.NLOPT_LN_BOBYQA
TRUE TRUE TRUE FALSE TRUE TRUE TRUE
I would appreciate comment on my design, convergence warnings, and troubleshooting results.
Thanks,
Umber
[[alternative HTML version deleted]]
More information about the R-sig-mixed-models
mailing list