yeah uv is a real delight. I wish it worked for R too

Unfortunately the methods literature hasn't really solved this problem yet AFAIK.

This very recent paper tried to extend the Hyndman&Fan quantile estimators to work with weights and also creates estimators that depend on the arbitrary sort order in the data.

academic.oup.com/jssam/advanc...

Yeah, I think NSE is a huge reason why R can be more ergonomic. That aside, for a long time I thought I hated Python when it turned out I just hated pandas and matplotlib. I have gripes with polars but it’s so much better than pandas.

😬 what a mess. And yeah, in the short term when there are weights it makes me want to just settle and use type 2

The !! operator can stay as is, but when genzplyr is loaded it has to be pronounced “gang gang”

Thanks Matt! That’s good to hear- I’ll update the post to note that

Thanks for pointing that out, I’ll read up on it. It looks from the function documentation like it wouldn’t have the same reproducibility issue as the others: “the functions correctly combine weights of observations having duplicate values of x before computing estimates.”

It should be fine if you’re doing interpolation of the ECDF that corresponds to the Hyndman and Fan types 1-3 (and if that works for your problem domain). It’s just tricky once you get into types 4-9, like these various R packages do.

Please let me know if you find any clear docs about how SPSS computes weighted quantiles. I suspect SPSS only allows integer frequency weights, which is a much easier scenario than the general problem of arbitrary weights that the R packages are trying to deal with.

Here's a simple R example. Two R packages yield different results from each other, and both R packages will surprisingly return different results depending on the arbitrary way that the data are sorted. So you can easily wind up getting a different estimate on Tuesday than you got on Monday.