Block-aware two-stage BH-FDR for specimen-shared cohort clusters

Naïve BH-FDR within a modality family treats specimen-shared cohorts (e.g., Toray clusters with \(\sim 94.5\%\) specimen reuse) as independent tests, which is anti-conservative due to strongly correlated test statistics. This function implements a two-stage correction:

1. Within each provenance block, combine cohort-level p-values via Fisher's chi-square statistic. Two reference distributions are reported side by side:

   • conservative reference: \(S/c\) against \(\chi^2_{2k}\); intentionally conservative; used for the frozen benchmark's headline counts.

   • textbook reference (Brown 1975 / Kost-McDermott 2002): \(S/c\) against \(\chi^2_{2k/c}\) with \(df\) rescaled by the same inflation factor.

2. Apply BH-FDR across block-level p-values, separately for each reference.

The inflation factor is \(c(k) = 1 + (k - 1)\rho\) with \(k\) the number of cohorts in the block and \(\rho\) the assumed within-block correlation (default \(\rho = 0.25\) reproduces the frozen benchmark's headline configuration).

Usage

singlesample_bh_fdr_correct_blocked(results, block_id, rho = 0.25)

Arguments

results

List of singlesample_matched_null_benchmark return objects.

block_id

Character vector — provenance block identifier per cohort (e.g., "Toray-cluster", "Affy-singleton").

rho

Numeric in [0, 1]. Assumed within-block correlation; controls the inflation factor \(c(k) = 1 + (k - 1)\rho\). Default 0.25 reproduces the frozen benchmark's headline configuration. Set to 0 to recover vanilla Fisher.

Value

A data.frame with columns: cohort_idx, block_id, p_emp, p_block (conservative), p_block_textbook, q_block_BH (conservative), q_block_BH_textbook, q_cohort_within_block.