Skip to content

Frequently asked questions

Summary statistics

What columns does viprs_fit require?

The summary statistics need four kinds of information:

  • A variant identifier: SNP (usually an rsID), or the pair CHR and POS.
  • Sample size: a per-variant N column, or a study-wide sample size passed with --gwas-sample-size.
  • Alleles: A1 for the effect/tested allele and A2 for the reference or other allele.
  • A signed measure from which the standardized marginal effect can be obtained: Z, BETA with SE, or PVAL/CHISQ accompanied by a signed statistic such as BETA, Z, or OR.

A p-value on its own does not encode effect direction. See the viprs_fit summary-statistics guide for the canonical names and a custom-format example.

The common mappings are #CHROM to CHR, ID to SNP, OBS_CT to N, T_STAT/Z_STAT to Z, and P to PVAL. In a simple biallelic table where the alternate allele was tested, ALT maps to A1 and REF maps to A2.

For native PLINK 2 output, use --sumstats-format plink2. The parser uses PLINK 2's A1 column as the tested/effect allele and infers A2 from REF and ALT (or ALT1), which also handles rows where the tested allele is not ALT.

Does VIPRS use A1 as the effect allele?

Yes. BETA, Z, and other signed association statistics must be expressed with respect to A1. A2 is the reference or other allele. During harmonization, VIPRS checks the alleles against the LD reference and changes the sign of signed statistics when it detects an allele swap.

How are variants matched between summary statistics and the LD reference?

The underlying merge_snp_tables function first matches on SNP when both tables contain variant IDs. Otherwise, it matches on the CHR/POS pair. It then checks A1 and A2, corrects valid allele swaps, and excludes incompatible variants.

Which genome build should I use?

The precomputed LD panels currently distributed for VIPRS use GRCh37/hg19 positions. Summary statistics matched by CHR and POS must therefore use GRCh37/hg19 as well. Lift over coordinates from another build before fitting. Even when matching by rsID, make sure the variant definitions and alleles agree with the LD panel.

Models

What is VIPRSMix, and how do I select it?

VIPRSMix uses a sparse mixture prior with several non-null Gaussian components of different scales and a spike at zero, conceptually similar to SBayesR. The --n-components value counts the non-null components only. The default value of one fits VIPRS, while any value greater than one automatically fits VIPRSMix; for example, --n-components 4 selects a four-component VIPRSMix model.

The Supplementary Material of Zabad et al. (2023) reports experiments with this prior.