Skip to content

Tests and gene-level calibration¤

Choose a response model, a variant-level test, and a gene-level calibration method.

Variant-level tests¤

The score test is the cis default: it fits one covariate-only null model per phenotype and reuses it across variants. Wald tests fit each variant's effect and are the usual choice for nominal scans requiring full-model effect estimates.

Use robust standard errors only with Wald tests

Score and SPA inference do not support Huber–White sandwich errors. Use --test wald --robust-se when requesting them; this option does not make score or SPA inference robust to model misspecification.

Tail and gene-level calibration¤

--spa attempts a saddlepoint correction for score statistics with absolute normal z-score above 1.96 that pass the CGF's support checks. Smaller statistics use the normal tail. A failed root solve or invalid correction also falls back to the normal approximation.

There is no separate SPA-success flag: model_converged describes the null fit, not the tail calculation.

Faster cis scans with SPA and ACAT¤

--spa --acat combines variant p-values into one gene-level p-value without permutations. It is typically much faster than Beta permutation because it avoids permutation refits; the speedup depends on the data and permutation count. See the Quickstart for both commands.

Use SPA with score-test ACAT

ACAT can amplify inaccurate variant tail p-values into misleading gene-level results. We strongly recommend --spa --acat for score tests. SPA can still fall back to the normal approximation.

Permutation calibration¤

Without --acat, cis scans use a Beta approximation fitted to permutation maxima. Each shuffle contributes the maximum absolute score or Wald statistic across the window. --spa does not replace these statistics with saddlepoint p-values.

Beta permutation does not share ACAT's direct sensitivity to asymptotic variant p-values: it compares statistics computed by the same procedure in the observed and permuted data. Geometry shared across those tests can therefore be reflected in the reference distribution. This requires valid permutations and successful fitting; it cannot repair numerical failures or guarantee an accurate Beta approximation.

The methods can yield different discoveries. Both need result checks and FDR correction across genes.

For Python interfaces, see Variant-level tests and Gene-level aggregation.