make quantreg run with statistical uncertainty #143
Conversation
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #143 +/- ##
==========================================
+ Coverage 65.12% 65.56% +0.44%
==========================================
Files 17 17
Lines 1832 1838 +6
==========================================
+ Hits 1193 1205 +12
+ Misses 639 633 -6 ☔ View full report in Codecov by Sentry. |
|
@kemccusker The labor SCCs are now replicating and this PR is ready for review again. |
| @@ -318,6 +318,8 @@ def scc(): | |||
| self.logger.info("Processing SCC calculation ...") | |||
| if self.fit_type == "quantreg": | |||
| self.full_uncertainty_iqr | |||
| self.calculate_scc | |||
There was a problem hiding this comment.
Not clear to me why we call full_uncertainty_iqr and stat_uncertainty_iqr here. Are we assuming we'll want to calculate both anything we have a "quantreg" fit type? I don't think that's necessarily true.
There was a problem hiding this comment.
I wasn't sure exactly how to work with this since it's kind of a weird quirk of the order_plate function. In the original scc function which is a "course" of the order_plate function, it will only calculate full_uncertainty_iqr (if quantreg). This meant that if I wanted statistical uncertainty, I couldn't go through the order_plate function. There are a few potential solutions I can think of:
- Add a new attribute to dscim that specifies uncertainty range (i.e. stat or full)
- Replace this scc function with a function that calls recipe functions according to which
save_filesyou have specified - Replace the two uncertainty functions with a single function that feeds the scc (collapsed for stat uncertainty or uncollapsed for full uncertainty) into it based on
fair_aggregation - Add a few more "courses" that allow the user to select which type of scc they want
There was a problem hiding this comment.
Let's try the first option "Add a new attribute to dscim that specifies uncertainty range (i.e. stat or full)" in a later PR
src/dscim/menu/main_recipe.py
Outdated
| @save(name="stat_uncertainty_iqr") | ||
| def stat_uncertainty_iqr(self): | ||
| """Calculate the distribution of quantile-weighted SCCs produced from | ||
| quantile regressions collapsed across pulse dimension. |
There was a problem hiding this comment.
"collapsed across pulse dimension" is a bit confusing. This IQR is taken across only the quantreg quantile dimension because there is no climate uncertainty dimension. This is trickier with RFF-SPs and we need to test it.
There was a problem hiding this comment.
Let's reword "quantile regressions that have already been collapsed across other dimensions to give statistical-only uncertainty".
In a separate PR, we should nicely fail if someone tries to run quantreg with RFF-SPs.
No description provided.