search

palday / MixedModelsPermutations.jl

Permutation tests for MixedModels.jl
https://palday.github.io/MixedModelsPermutations.jl/stable
MIT License
4 stars 2 forks source link

inflation factor: PivotedCholesky is not unpivoted #38

Closed behinger closed 2 years ago

behinger commented 2 years ago

We are running into huge inflationfactor issues, because the pivotedcholesky is not correctly unpivoted, similar to e.g. https://github.com/JuliaLang/julia/issues/33704

Unfortunately, unpivoting like this:

L = chol.L[invperm(chol.p),:]

results in a non-triangular matrix which we'd need LowerTriangular by force (but the L*L' covariance matrices look great ;-))

Alternatively, we could just do everything in Covariance space, divide elementwise. But we then have a rank-deficient inflation matrix, which imho requires changes downstream in https://github.com/palday/MixedModelsPermutations.jl/blob/90c678007e1586009629b452700eeae3ded710eb/src/permutation.jl#L271 the multiplication with the RE; or we'd need to convert it back to cholesky running into the inital problem

What do you think @palday?

35 is not dealt with then either; but not sure we actually still have to do it then?

behinger commented 2 years ago

this was solved by #39

What is still unclear, whether we even need the intermediate step of adding noise to a rank incomplet pivoted cholesky factorization (#35)