Do not impute at GWAS-observed positions in RAISS

[Yining97]

Problem

RAISS chose imputation targets purely by variant_id: unknowns <- which(!refPanel$variant_id %in% knownsId). So a genomic position already typed in the GWAS still had its other panel forms imputed — a second allele or the opposite orientation of a site the GWAS actually measured. On the ADSP R5 panel this inflated the imputed set (43 observed multi-allelic positions → 2,969 imputed variants in one block) and manufactured redundant records at measured positions, contradicting the intent of imputation (fill un-observed variants).

Fix

In both RAISS entry points (raissSingleMatrix and raissSingleMatrixFromX), a panel variant is an imputation target only if its canonical (chrom, pos) is not present in the observed z-scores, on top of the existing variant_id check. Positions with no GWAS observation — including genuinely multi-allelic sites — still impute all their panel forms, so legitimately imputed variants remain available to fine-mapping and the credible set.

Validation

• Synthetic panel with a second allele at a typed position: the old selection imputes it; the guard does not. Un-observed positions still impute, and an un-observed multi-allelic site keeps both ALTs. Verified on both the LD-matrix and SVD (fromX) paths.
• Also fixed generate_dummy_data, which sampled the known set's variant_id and pos independently (placing a known id at a foreign position) — an inconsistent fixture the guard correctly flags; now drawn from the same panel rows.
• Full suite: this branch 3035 rows / 3 failed / 6 errors vs clean main 3033 / 3 / 6 (+2 new tests) — identical failure counts, zero new regressions. The 3 failures are pre-existing order-dependent joint-dispatcher tests (pass in isolation); the 6 errors are the pre-existing test_ctwasPipeline.R ones.

Tests added: test_sumstatsQc.R +2 (both RAISS paths).

🤖 Generated with Claude Code