I am encountering convergence problems with some models after updating to lme4 1.1-20 that I did not encounter with earlier versions of lme4 (in particular, lme4 1.1-15). I am encountering these new convergence issues for some, but not all models, but the models with convergence issues are generally simple models. Therefore I do not understand why they are not converging.

Here is an example:


d <- as.data.frame(
  fread("curl http://www.intensivelongitudinal.com/ch4/ch4R.zip | tar -xf- --to-stdout *time.csv")


fit <- lmer(intimacy ~ time * treatment + (time | id), data = d)

This throws a warning, which reads:

Model failed to converge with max|grad| = 0.00398684 (tol = 0.002, component 1)

This model previously converged without issue when using lme4 1.1-15. In addition, it is able to converge when using a rescaled version of the time variable (time01 in the dataset, which changes the range from 0 to 1 instead of 1 to 15).

Any insight into what lme4 1.1-20 is doing differently?


Session Info:

R version 3.5.2 (2018-12-20)
Platform: x86_64-apple-darwin15.6.0 (64-bit)
Running under: OS X El Capitan 10.11.6

Matrix products: default
BLAS: /System/Library/Frameworks/Accelerate.framework/Versions/A/Frameworks/vecLib.framework/Versions/A/libBLAS.dylib
LAPACK: /Library/Frameworks/R.framework/Versions/3.5/Resources/lib/libRlapack.dylib

[1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] lme4_1.1-20       Matrix_1.2-15     data.table_1.12.0

loaded via a namespace (and not attached):
 [1] Rcpp_1.0.0      lattice_0.20-38 digest_0.6.18   MASS_7.3-51.1   grid_3.5.2      nlme_3.1-137   
 [7] evaluate_0.13   minqa_1.2.4     nloptr_1.2.1    rmarkdown_1.12  splines_3.5.2   tools_3.5.2    
[13] xfun_0.5        yaml_2.2.0      compiler_3.5.2  htmltools_0.3.6 knitr_1.22  

1 Answer 1


tl;dr this is a false positive; you can fix it by switching optimizers back to the old default or tightening tolerances slightly.

The lme4 NEWS file shows for version 1.1-20

  • changed default optimizer to "nloptwrap" (BOBYQA implementation from the nloptr package) for lmer models. To revert to the old default, use control=lmerControl(optimizer="bobyqa")

(this is the only change relevant to this question).

The ?convergence man page points out that for the new nloptwrap optimizer,

  • Changing ftol_abs and xtol_abs to stricter values (e.g. 1e-8) is a good first step for resolving convergence problems, at the cost of slowing down model fits.

Let's try these two approaches to check that this is a false-positive warning:

## default settings (1.1-20)
fit <- lmer(intimacy ~ time * treatment + (time | id), data = d)
## switch back to pre-1.1-20 default optimizer
fit2 <- update(fit, control=lmerControl(optimizer="bobyqa"))
## tighten tolerances
fit3 <- update(fit, control=lmerControl(optCtrl=list(ftol_abs=1e-8,xtol_abs=1e-8)))

Neither of the latter two examples give convergence warnings.

How much do the estimates differ? Fixed effects:

# [,1]          [,2]
# (Intercept)    -9.325873e-15 -4.618528e-14
# time            5.294376e-15  1.942890e-15
# treatment       1.398881e-14  2.489675e-14
# time:treatment -6.196432e-15 -6.106227e-16

The RE standard deviations and correlations differ much more, but probably still not enough to worry about:

ff <- function(x) {as.data.frame(VarCorr(x))$sdcor}
# [,1]          [,2]
# [1,] -4.871044e-05 -2.082594e-05
# [2,]  8.250232e-06  5.907501e-06
# [3,] -5.482109e-05 -4.694410e-05
# [4,]  9.775125e-07  2.717170e-08

PS See ?troubleshooting, ?convergence, and ?allFit for more suggestions about dealing with lme4 problems.

PPS There's a good chance that we will tighten the default nloptwrap tolerances in a future release of lme4.

  • 1
    $\begingroup$ Thank you so much, very helpful! I was able to reproduce these fixes no problem on my end. Thanks again. $\endgroup$
    – kz1
    Apr 1, 2019 at 0:34

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.