Hi,
I specified a multilevel longitudinal model in MixedModels.jl. I am having a heteroscedasticity issue, for which I’d like to try to weight observations.
So, the unweighted model:
fit(MixedModel,@formula(maxtrop~1+time*glucose+(time|mr/anest)),d,REML=false)
Unit weighting works (3577 is the row count of the dataset):
fit(MixedModel,@formula(maxtrop~1+time*glucose+(time|mr/anest)),d,wts=uweights(3577),REML=false)
… but, I can’t make it further than that. Any other spec results in the following error
Initial objective evaluation failed, rescaling initial guess and trying again.
Failure of the initial evaluation is often indicative of a model specification that is not well supported by the data and/or a poorly scaled model.
I tried, among others:
wts=aweights(float.(rand([1,2,3],3577)))
wts=aweights(rand(Float64,3577))
I don’t know if it might be of importance, but there is almost no signal in the data. The unweighted model results:
Linear mixed model fit by maximum likelihood
maxtrop ~ 1 + time + glucose + time & glucose + (1 + time | mr) + (1 + time | mr & anest)
logLik -2 logLik AIC AICc BIC
23762.1125 -47524.2251 -47502.2251 -47502.1510 -47434.2200
Variance components:
Column Variance Std.Dev. Corr.
mr & anest (Intercept) 0.05686690 0.23846781
time 0.00000000 0.00000000 -1.00
mr (Intercept) 0.07487753 0.27363758
time 0.00000000 0.00000000 +1.00
Residual 0.00000001 0.00009071
Number of obs: 3577; levels of grouping factors: 482, 477
Fixed-effects parameters:
─────────────────────────────────────────────────────────
Coef. Std. Error z Pr(>|z|)
─────────────────────────────────────────────────────────
(Intercept) -0.150825 0.0165952 -9.09 <1e-18
time -2.40061e-9 5.61142e-6 -0.00 0.9997
glucose 8.79342e-10 8.81333e-7 0.00 0.9992
time & glucose 2.37489e-10 6.75191e-7 0.00 0.9997
─────────────────────────────────────────────────────────
Any help would be greatly appreciated. Thanks!