do we want to stick with `oldsvyquantile` for the foreseeable future?

[ajdamico]

it would be great if you could just glance at the list of places that we use oldsvyquantile and double-check whether it makes sense to keep using the deprecated version of the function?

https://github.com/ajdamico/convey/commit/5592eb7a28e00f39fb7fc489139fd25949dc1ae9

i doubt it matters much, but may be worth a quick look?

[guilhermejacob]

I would stick with that for now. Unless you can set up the exact same estimator in oldsvyquantile using the new svyquantile, that would mean changing estimators, which would likely require other variance estimator(s).

[guilhermejacob]

Let's look at the estimators. As far as I know, svyoldquantile estimates the quantile using a linear interpolation on the CDF; the new svyquantile has a qrule parameter that implements several different estimators. The new ones might work just fine, but we would need some to (at least) run some simulations before making that change; that is why I suggest keeping it for now.

In fact, Langel & Tillé (2011) use another estimator, based on the 4th definition of Hyndman and Fan (1996), and another variance estimator -- that, unlike ours, does not rely on density estimation. I don't know if it matches Lumley's implementation of the H&F estimator; it might not.

Additionally, we need to make sure we can replicate results in vardpoor. As far as I know, its estimator is not the one in old oldsvyquantile, as it does not rely on that linear interpolation.

[ajdamico]

makes sense, thank you!!