Open dmbates opened 5 months ago
If you use a contrast like EffectsCoding()
for Source
then the conditional means of the random effects for the term (1 + Source | Lot & Wafer)
end up being close to multiples of [1, -1]
for the first level of Source
and [1, 1]
for the second level.
julia> first(m2.b)
2×24 Matrix{Float64}:
3.70686 -5.46891 2.67088 -0.979013 -1.71899 -2.75497 … -0.997343 -0.39882 -0.0995575 -0.178385 0.56977 -1.67469
-3.72693 5.49852 -2.68534 0.984313 1.7283 2.76988 -1.00222 -0.400768 -0.100044 -0.179257 0.572553 -1.68288
I'm not sure if that is a consequence of an unstable model or of only having 3 observations for each Lot & Wafer
combination.
One of the tests in
test/pls.jl
uses the:oxide
models defined intest/modelcache.jl
The second model is notoriously hard to fit. Different optimizers give very different parameter estimates but with similar values of the objective. I think this is because the model is ill-defined as
Source
is constant within eachLot
.To me this means that you can't expect to fit a random-effects term like
(1 + Source|Lot)
.Am I confusing myself?