MVMR vs silx

MVMR implements multivariable Mendelian randomization — conditional instrument strength, pleiotropy tests and heterogeneity-robust effect estimation from GWAS summary data. The package has been releasing steadily through 2026, and the substantive releases are all corrections rather than features. Two core routines were found to be computing the wrong quantity outright: qhet_mvmr() built weights from the minimised objective value instead of the minimiser, and strhet_mvmr() never minimised the Q-statistic at all.

This is a sustained audit, not a maintenance drift. Each release since February has fixed a specific analytical defect — omitted intercepts in the exposure-on-genotype regressions, a division by zero when a gencov list held exactly two variants, covariance matrices computed wrongly for matrix inputs, a spurious covariance warning — and several explicitly warn that reported values will differ from previous versions. The strhet_mvmr() rewrite to iteratively reweighted least squares also removes a combinatorial grid that could exhaust memory past three exposures, so the function is now usable as well as correct.

The corrections have been walking through the package function by function, and the ones with published fixes so far are the heterogeneity and covariance routines; the remaining untouched estimators are the natural next stop if the audit continues at this pace.