Tests failure on rocker about tern HOT 8 CLOSED

Hi @anajens and @wwojciech , so the above error is coming from test-tabulate_mmrm.R However when I run the tests, I got different error messages.

from tern.

@yli110-stat697 what environment are you testing this in?

from tern.

@yli110-stat697 what environment are you testing this in?

In the latest rocker docker-registry.nest.roche.com/nest/r/rocker/nest:devel-latest

from tern.

So it seems like when arm is taken out of the vars in fit_mmrm, all the 7 optimizers are having converging problems during model fit.

from tern.

I also can't reproduce the initial error above and see the 3 errors related to model not converging when testing with rstudio_4.1.0_bioc_3.13 and staged.dependencies which need to be fixed.

I checked the results of the first test that fails above vs our last release and the estimates from the model that is returned by fit_lme4_single_optimizer are identical. However, there is now a new warning about a singular fit.

@danielinteractive any suggestions on how to proceed? We can easily update the tests to use different covariates but I'm wondering if there are any additional actions needed to update how messages are handled byfit_lme4?

Result from 2021_07_07 release (tested with 3.6.3 BEE):

Result from main branch (tested docker rstudio_4.1.0_bioc_3.13 with staged.dependencies):

Here's the code I used from test-tabulate_mmrm.R:

anl <- get_anl() %>%
  mutate(
    ARM = factor(ARM, levels = c("B: Placebo", "A: Drug X", "C: Combination")),
    AVISIT = factor(AVISIT)
  )

debugonce(fit_mmrm)

mmrm <- fit_mmrm(
  vars = list(
    response = "AVAL",
    visit = "AVISIT",
    covariates = c("BMRKR2"),
    id = "USUBJID"
  ),
  data = anl,
  cor_struct = "random-quadratic"
)

from tern.

@anajens Yeah this convergence stuff is unfortunately pretty fragile between different machines, R package versions etc. To me your example result above on main actually looks better than on latest release. I don't think we need to hide this additional line printed in the bottom. So I would go with your suggestion to update the models being tested. I hope it's not those comparing to SAS results though?

from tern.

Thanks Daniel. It's not the SAS tests that are failing.

I still don't understand how the model summary can look identical across systems but in one case isSingular returns FALSE and TRUE in the latter.

from tern.

@anajens I would guess that it is just the interpretation of model results into these singularity flags that changed.

from tern.