multiple-testing-correction
License: restricted — no clear open-source license detected for the underlying tool; verify licensing before commercial use or redistribution.
Summary
Apply multiple-testing correction (Benjamini–Hochberg or other family-wise error rate control) to p-values derived from pathway enrichment tests to control false discovery rate across many simultaneous statistical tests. This skill prevents inflated Type I error when testing hundreds of pathways or metabolite categories against a single input metabolite set.
When to use
Apply this skill whenever you have performed Fisher's exact test or another statistical enrichment test on multiple pathways, lipid categories, or metabolite sets simultaneously (typically ≥2 tests, often 50–100+ tests in practice). The raw p-values from each individual test do not account for the multiple comparisons problem; correction is required before reporting or filtering significant results.
When NOT to use
- Input is a single hypothesis test (n=1 pathway or category) — no correction needed.
- P-values have already been corrected by the upstream software — applying correction twice will introduce bias.
- Analysis is exploratory and raw p-values are acceptable — though not recommended for publication without adjustment disclosure.
Inputs
- vector or data.frame column of raw p-values from Fisher's exact test (one p-value per pathway or metabolite category)
- number of tests performed (inferred from p-value vector length)
Outputs
- data.frame or vector of adjusted p-values (same length as input)
- enrichment results table with columns: pathway/category name, raw p-value, adjusted p-value, odds ratio or effect size, metabolite count
How to apply
After computing raw p-values from Fisher's exact test for each pathway or category, apply Benjamini–Hochberg (BH) correction to the full set of p-values using a standard method (e.g., R's p.adjust() function with method='BH'). The BH procedure controls false discovery rate (FDR) while preserving statistical power better than strict Bonferroni correction. Store both raw and adjusted p-values in the enrichment results table; use adjusted p-values (typically with threshold p_adj ≤ 0.05) for downstream filtering, visualization, and reporting. Document the correction method and cutoff in all output tables and figures.
Related tools
- R p.adjust() function (Performs Benjamini–Hochberg and other multiple-testing corrections on p-value vectors)
- enrichmet (Integrates Benjamini–Hochberg correction as a built-in step in pathway enrichment workflow; automatically computes and returns adjusted p-values in enrichment results table) — https://github.com/biodatalab/enrichmet
- stats::p.adjust (R base) (General-purpose multiple-testing correction for any set of p-values)
Examples
results <- enrichmet(inputMetabolites = inputMetabolites, PathwayVsMetabolites = PathwayVsMetabolites, p_value_cutoff = 0.05, min_pathway_occurrence = 2); # Benjamini–Hochberg correction is applied internally; view corrected p-values in results$pathway_enrichment_all
Evaluation signals
- Adjusted p-values are monotonically non-decreasing when sorted by raw p-value (verification of correct BH ranking).
- Adjusted p-value ≥ corresponding raw p-value for all tests (BH correction always inflates p-values to be conservative).
- Number of significant pathways after adjustment (adj_p ≤ 0.05) is ≤ number before adjustment.
- Enrichment results table contains both 'P_value' (raw) and adjusted p-value columns with clear labeling (e.g., 'Adjusted_P_value' or 'FDR').
- Reported significant findings cite the corrected p-value threshold and correction method (e.g., 'Benjamini–Hochberg corrected p ≤ 0.05').
Limitations
- BH correction assumes tests are independent or positively dependent; if pathways share metabolites (common in real data), the assumption may be violated, though BH remains valid.
- With very large numbers of tests (e.g., >10,000), BH correction may be overly conservative and reduce power; more sophisticated methods (e.g., Storey's q-value) may be preferred in those cases.
- Correction cannot recover signal from inherently noisy or underpowered individual tests; if raw p-values are all close to 1, correction will not rescue significance.
- Choice of FDR threshold (0.05 vs. 0.1 vs. 0.01) is arbitrary and should be stated a priori; post-hoc threshold selection risks p-hacking.
Evidence
- [intro] Compute adjusted p-values using Benjamini–Hochberg correction: "Compute adjusted p-values using Benjamini–Hochberg correction."
- [intro] enrichment results table with pathways, metabolite counts, p-values, adjusted p-values, and effect sizes: "data.frame with pathways, metabolite counts, p-values, adjusted p-values, and effect sizes"
- [intro] p_value_cutoff parameter controls significance threshold for individual tests: "p_value_cutoff = 0.05"
- [intro] Fisher's exact test enrichment workflow executed on each category: "Execute Fisher's exact test on each lipid ontology category using the enrichmet workflow to test for significant association"
1---2name: multiple-testing-correction3description: Use when whenever you have performed Fisher's exact test or another statistical enrichment test on multiple pathways, lipid categories, or metabolite sets simultaneously (typically ≥2 tests, often 50–100+ tests in practice).4license: CC-BY-4.05---67# multiple-testing-correction89> **License: restricted** — no clear open-source license detected for the underlying tool; verify licensing before commercial use or redistribution. <!-- asb-license-banner -->10## Summary1112Apply multiple-testing correction (Benjamini–Hochberg or other family-wise error rate control) to p-values derived from pathway enrichment tests to control false discovery rate across many simultaneous statistical tests. This skill prevents inflated Type I error when testing hundreds of pathways or metabolite categories against a single input metabolite set.1314## When to use1516Apply this skill whenever you have performed Fisher's exact test or another statistical enrichment test on multiple pathways, lipid categories, or metabolite sets simultaneously (typically ≥2 tests, often 50–100+ tests in practice). The raw p-values from each individual test do not account for the multiple comparisons problem; correction is required before reporting or filtering significant results.1718## When NOT to use1920- Input is a single hypothesis test (n=1 pathway or category) — no correction needed.21- P-values have already been corrected by the upstream software — applying correction twice will introduce bias.22- Analysis is exploratory and raw p-values are acceptable — though not recommended for publication without adjustment disclosure.2324## Inputs2526- vector or data.frame column of raw p-values from Fisher's exact test (one p-value per pathway or metabolite category)27- number of tests performed (inferred from p-value vector length)2829## Outputs3031- data.frame or vector of adjusted p-values (same length as input)32- enrichment results table with columns: pathway/category name, raw p-value, adjusted p-value, odds ratio or effect size, metabolite count3334## How to apply3536After computing raw p-values from Fisher's exact test for each pathway or category, apply Benjamini–Hochberg (BH) correction to the full set of p-values using a standard method (e.g., R's p.adjust() function with method='BH'). The BH procedure controls false discovery rate (FDR) while preserving statistical power better than strict Bonferroni correction. Store both raw and adjusted p-values in the enrichment results table; use adjusted p-values (typically with threshold p_adj ≤ 0.05) for downstream filtering, visualization, and reporting. Document the correction method and cutoff in all output tables and figures.3738## Related tools3940- **R p.adjust() function** (Performs Benjamini–Hochberg and other multiple-testing corrections on p-value vectors)41- **enrichmet** (Integrates Benjamini–Hochberg correction as a built-in step in pathway enrichment workflow; automatically computes and returns adjusted p-values in enrichment results table) — https://github.com/biodatalab/enrichmet42- **stats::p.adjust (R base)** (General-purpose multiple-testing correction for any set of p-values)4344## Examples4546```47results <- enrichmet(inputMetabolites = inputMetabolites, PathwayVsMetabolites = PathwayVsMetabolites, p_value_cutoff = 0.05, min_pathway_occurrence = 2); # Benjamini–Hochberg correction is applied internally; view corrected p-values in results$pathway_enrichment_all48```4950## Evaluation signals5152- Adjusted p-values are monotonically non-decreasing when sorted by raw p-value (verification of correct BH ranking).53- Adjusted p-value ≥ corresponding raw p-value for all tests (BH correction always inflates p-values to be conservative).54- Number of significant pathways after adjustment (adj_p ≤ 0.05) is ≤ number before adjustment.55- Enrichment results table contains both 'P_value' (raw) and adjusted p-value columns with clear labeling (e.g., 'Adjusted_P_value' or 'FDR').56- Reported significant findings cite the corrected p-value threshold and correction method (e.g., 'Benjamini–Hochberg corrected p ≤ 0.05').5758## Limitations5960- BH correction assumes tests are independent or positively dependent; if pathways share metabolites (common in real data), the assumption may be violated, though BH remains valid.61- With very large numbers of tests (e.g., >10,000), BH correction may be overly conservative and reduce power; more sophisticated methods (e.g., Storey's q-value) may be preferred in those cases.62- Correction cannot recover signal from inherently noisy or underpowered individual tests; if raw p-values are all close to 1, correction will not rescue significance.63- Choice of FDR threshold (0.05 vs. 0.1 vs. 0.01) is arbitrary and should be stated a priori; post-hoc threshold selection risks p-hacking.6465## Evidence6667- [intro] Compute adjusted p-values using Benjamini–Hochberg correction: "Compute adjusted p-values using Benjamini–Hochberg correction."68- [intro] enrichment results table with pathways, metabolite counts, p-values, adjusted p-values, and effect sizes: "data.frame with pathways, metabolite counts, p-values, adjusted p-values, and effect sizes"69- [intro] p_value_cutoff parameter controls significance threshold for individual tests: "p_value_cutoff = 0.05"70- [intro] Fisher's exact test enrichment workflow executed on each category: "Execute Fisher's exact test on each lipid ontology category using the enrichmet workflow to test for significant association"
HolobiomicsLab