v1.2.0

What's Changed

Bug Fixes

• Fix missing factor of 2 in variance formula: produit_reciproque_asymetrique() was summing only the upper triangle of the symmetric correlation matrix without doubling, causing the corrected z-score and beta standard errors to be underestimated. This inflated type 1 error rates under sample overlap.
• Fix Digby tetrachoric exponent: Changed default from 0.75 to pi/4 (0.7854), reducing systematic bias in correlation estimates (e.g. bias at rho=0.5 drops from -0.02 to -0.003).
• Always use sign-of-beta dichotomization: The old p-value transform method severely underestimated the tetrachoric correlation. The sign-of-beta method (Southam et al. 2017) is now always used. The --use-beta-sign flag has been removed.

New Features

• --digby-exponent CLI option: Override the Digby approximation exponent (default: pi/4 = 0.7854). Use --digby-exponent 0.75 to restore the old behaviour.

Breaking Changes

• --use-beta-sign flag removed (this behaviour is now the default and only option)

Other

• Removed dead code (b_transform, add_minimum, legacy meta-analysis prototype)
• Added simulation notebook (test_data/sim_compare_fixes.ipynb)

Assets

• metacarpa_static: Statically linked Linux x86_64 binary (no dependencies required)

Fixed dichotomization

Fixed dichotomization (use --use-beta-sign). For full changelog look here.

v. 1.0.1

When the input p-values are too small (<e-200) METACARPA now falls back to cpp_dec_float<400> instead of cpp_dec_float<200>.

High precision version

This has been compiled by setting cpp_dec_float<1000> for increased precision in p-values.