Alex Arslan

> I don't see why `B where B` above couldn't be kept as `Union{Int, Null}.` This would make a lot of sense if unions are made more efficient. Perhaps, though...

What I mean is that if you have `Array{Pair{Int64, B} where B, 1}`, `B 1.2)`, you won't be able to. I (completely naively) assume that Julia defensively makes the non-constant...

That's a good point. Don't mind me, then...

> We only use this field for computing the uncertainty after the estimation, right AFAICT the Fisher information field isn't meaningfully used anywhere, but the covariance matrix field is indeed...

> But with this PR it would fail, no? No, at least as written it would use a pivoted Cholesky factorization. > I'm not sure using pinv is sound when...

Would it be enough to add something like ```julia rank(M) < min(size(M)...) && throw(ArgumentError("model matrix is not full rank")) ``` to the `fit(CoxModel, ...)` method? Your point about only using...

Based on our discussion, ~~the title of the PR is now inaccurate and~~ (I've updated that and the OP) what's implemented is the following: - Throw an error if `X`...

Also note that there are other ways you can write it depending on the end goal. A direct replacement would be `map(!ismissing, x)`, but to index into a structure, filtering...

Currently it computes the pointwise standard error of the log of the survivor function using Greenwood's formula. I forget why exactly I opted for that but likely it was because...

A slightly nicer (though basically equivalent) workaround: ```julia if !applicable(JSON.lower, NA) JSON.lower(::NAtype) = nothing end ```