Closed tobiasmuetze closed 5 months ago
hi @tobiasmuetze this issue should be fixed now, previously we report the fit incorrectly (we did not report the best successful one but some random one) and lead to your observed issue. now should be fine
@clarkliming - Does this only affect the error code or does it affect the coefficients as well ? We have been seeing issues with rbmi where mmrm occasionally returns different coefficients and I am wondering if that is related to this ?
@gowerc It also affects coefficients. But as I understand this behavior should still have been deterministic
@clarkliming - Does this only affect the error code or does it affect the coefficients as well ? We have been seeing issues with rbmi where mmrm occasionally returns different coefficients and I am wondering if that is related to this ?
yes @gowerc the coefficients got changed, and also all other things, vcov, theta, likelihood, etc. the fit object got changed. the termrandom
is not correct but it previously obtain the index among the successful fits and use that index to obtain the fit among all fits. However, if you see unstable coefficients on the same data, best if it can be reproduced and we can then investigate. if under rbmi, then might be true that we have some edge cases in bootstrapping?
Summary
The help for
mmrm()
states: "When optimizer is not set, first the default optimizer (L-BFGS-B) is used to fit the model. If that converges, this is returned. If not, the other available optimizers fromh_get_optimizers()
, including BFGS, CG and nlminb are tried." However, I am experiencing the case where if no optimizer is specified, the convergence status of 1 is returned, but when the optimizer nlminb is specified, the convergence status is 0. It is unclear how this behavior fits to what is stated in the help.Minimal example
A data set is simulated that has 300 subjects in the treatment group and 150 subjects in the control group. Subjects have one baseline and 12 post-baseline visits. The control group mean is set to zero for all visits. The treatment group mean is set to 0 for the baseline visit and 0.4 for the post-baseline visits. The standard deviation is set to 1 and the correlation matrix of a subject’s outcome across visit has an AR(1) structure There is no missing data.
The data set includes one row per subject and per post-baseline visit.
Session info