search

conroylau / lpinfer

lpinfer: An R Package for Inference in Linear Programs
GNU General Public License v3.0
3 stars 5 forks source link

Change to fsst range problem #67

Closed a-torgovitsky closed 4 years ago

a-torgovitsky commented 4 years ago

We have changed the range problem to something that appears to work better and is also much simpler computationally. Please see the attached note, in particular Sections 6.1 and 8.1 Let me know if something doesn't make sense. Also, note that the "\hat{\Omega}_{n}^{e}" matrix is no longer needed with this new approach. fsst-module-updated.pdf

conroylau commented 4 years ago

Done! I have updated the fsst module with the updated range problem. I have also removed the parts that are no longer used, such as the $\hat{\Omega}_{n}^{e}$ matrix and its regularization parameter.

The following table shows the p-values and the components of the four DGPs in issue #24 (the updated DGPs). I am using the same set-up and the same parameters as in issue #24. DGP p-value (New) p-value (as in #24) Range component (New) Range component (as in #24) Cone component (unchanged)
1 0.96 0.90 8.67591 0.23877 0.61354
2 0.37 0.40 60.69488 2.00311 0.49790
3 0.83 0.83 0 0 44.53229
4 0.66 0.66 0 0 49.19598

The results for DGP 3 and 4 are the same as before because their Range components are set to 0 in the procedure.

For the output message, the change that I made is to remove the line that shows the regularization parameter for the Range studentization matrix in the summary message. This is also reflected in the most updated README.

Thanks!