Closed
Description
Discovered in #431, if you try to extrapolate quantiles for predictions which have been thresholded, you can get quantiles outside of those thresholds. For example, if you have a quantile level of .1
which has value 0
, then .05
will be negative, even if the quantile is negative.
Ways I can think to deal with this:
- include the support in the distribution, probably as an interval, e.g.
c(0,Inf)
, and use that inquantile_extrapolate
. Unsure if the types will play nicely with this one. - Add limit options to get passed to
quantile_extrapolate
. Seems like it will require the user to know that they have thresholds, whereas the first option will build it into the result as part oflayer_threshold
. - Just let it be and let the user figure it out by mentally thresholding. Seems not ideal.