Closed s3alfisc closed 2 years ago
Thanks for the excellent testing. I think these examples were all fixed by moving a )
in one line.
I am publishing version 0.6.5 now which I hope fixes the problems above and brings a modest but noticeable additional reduction in per-session latency, at least if you only doing one major kind of test, like WRE or WCR.
Yes, all tests pass with v0.6.5 & latency on first use is reduced by around 5s!
I have now managed to set up continuous integration of
fwildclusterboot
vswildboottestjlr
& have discovered one bug inwildboottestjlr
that I believe goes back to toWildBootTests.jl
. (I have also found one bug infwildclusterboot
that I need to address).You can find the test results of the CI tests here under run-lib-actions check-r-package. The Julia vs R tests are all in test_r_vs_julia.r. I will set up a similar workflow for
wildboottestjlr
, and I even think it might be possible to trigger the workflow every time you push toWildBootTests
- I will check this if you think it is of interest to you. Beyond that, I believe that you should have all rights to trigger the workflow manually.Bug:
For
rademacher weights and nonull and multi-variable hypotheses
, if the estimated p-value is 0, the confidence interval computed by WildBootTests is incorrect.Example:
Changing beta0 changes the severity of the error:
If
pval > 0
, the error does not occur:Update
There is a related error (only for rademacher weights & impose_null = FALSE & multivariable hypotheses) for one sided tests if the p-val is either 0 or 1, which leads to incorrect t-stats (and likely also confidence intervals):