codecov-commenter
commented
3 months ago Codecov Report
Attention: Patch coverage is 62.96296% with 10 lines in your changes missing coverage. Please review.
Project coverage is 64.73%. Comparing base (
6d00378) to head (1390806).
| Files | Patch % | Lines |
|---|---|---|
| src/ghostbasil_parallel.jl | 61.53% | 10 Missing :warning: |
Additional details and impacted files
```diff @@ Coverage Diff @@ ## main #35 +/- ## ========================================== - Coverage 65.09% 64.73% -0.37% ========================================== Files 5 5 Lines 679 689 +10 ========================================== + Hits 442 446 +4 - Misses 237 243 +6 ```:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
biona001
Motivation
When a LD matrix is too large (e.g. >5000 variables), the function
find_optimal_shrinkagebecomes too slow.Proposal
Rather than computing shrinkage on the entire correlation matrix, we can do so on just the group representative variables, since these variables are primary drivers for the dependencies across groups. This often reduce the LD matrix size by a factor of 10.
Potential issue
Often there are few representatives in a block (e.g. < 10). This can lead to large shrinkage values (e.g. >0.5) even though the LD matrix matches the Z-score exactly.
this PR
Now we compute shrinkage on the original matrix if the size is manageable (<1000 variables), otherwise we change to computing shrinkage value on the representatives.
Before merging
To check this approach is reasonable, we should re-run the Alzheimer's disease study to see if the shrinkage value is comparable to previously (which I think we around 0.01 to 0.02).