Phillip Alday

The general impression I'm getting from the literature I've been skimming the last few days is that an anayltic solution may not be known. I'll keep looking for one. Maybe...

The failure on nightly is an upstream bug (https://github.com/JuliaLang/julia/issues/40609).

Ugh, I'm conflicted on that front. Because that would be breaking, but technically a user hitting overflow on their custom narrow types is expected behavior. @nalimilan do we have enough...

Okay, so on the breaking / non breaking front: 1. Did the API ever specify the return type for `mean` and the various moments? If not, this is not breaking...

> It would be good to prepare at least one implementation in a package before merging this, just to ensure the result suits our needs before committing to this API...

@nalimilan I have a pretty general implementation over in MixedModelsExtras (https://github.com/palday/MixedModelsExtras.jl/pull/11) that I want to use as an API testing ground. Let me know if that API works for you....

@nalimilan I would say put the `[g]vif` stub in wherever `RegressionModel` now lives and a general implementation in StatsModels.jl that other packages can optionally overload. Mostly I just want it...

The underlying problem seems to come from `var`, because `sqrt` preserves type and `std` is defined as `sqrt.(var(...)))`

Oh, and here's the problem: ```julia function var(A::RealArray, w::AbstractWeights, dim::Int; mean=nothing, corrected::DepBool=nothing) corrected = depcheck(:var, corrected) var!(similar(A, Float64, Base.reduced_indices(axes(A), dim)), A, w, dim; mean=mean, corrected=corrected) end ```

Same seems to hold for `Scatter`.