Benjamini–Hochberg Procedure¶

A step-up rule that sorts m p-values and rejects through the largest rank k where p(k) ≤ (k/m)·α, bounding the false discovery rate — the expected proportion of false rejections among discoveries — rather than the probability of any false positive.

Core Idea¶

The Benjamini–Hochberg (BH) procedure is the canonical method for controlling the false discovery rate (FDR) — the expected proportion of rejected null hypotheses that are false rejections — when testing many hypotheses simultaneously. Given m tests producing p-values sorted in ascending order p(1) ≤ p(2) ≤ … ≤ p(m), the procedure finds the largest index k such that p(k) ≤ (k/m)·α, then rejects all hypotheses at ranks 1 through k. Under independence of the test statistics (or positive regression dependence), the resulting FDR is bounded by α·m₀/m ≤ α, where m₀ is the unknown number of true null hypotheses. Introduced by Benjamini and Hochberg in 1995, the procedure became foundational to the high-throughput revolution in genomics, neuroimaging, and large-scale clinical research.

The structural commitment the BH procedure makes is a shift in what error should be controlled. Classical single-test inference targets the familywise error rate (FWER) — the probability of any false discovery at all — which the Bonferroni correction enforces by setting the per-test threshold to α/m. At the scales of modern high-throughput science, FWER control at m = 20,000 gene tests or m = 100,000 brain voxels forces the per-test threshold so low that essentially no true signals survive: the correction eliminates discovery. The BH procedure instead controls the proportion of false discoveries among all discoveries, accepting that some fraction of reported positives will be false in exchange for recovering a tractable set of true signals. The step-up structure of the algorithm — sorting p-values and applying a threshold that relaxes linearly with rank — makes the procedure both computationally simple and interpretable: each researcher can reconstruct which tests clear the FDR threshold and at what rate false discoveries are expected among those that do.

Structural Signature¶

Sig role-phrases:

the test population — \(m\) hypothesis tests evaluated together, the batch over which multiplicity must be managed
the sorted p-values — the ascending list \(p_{(1)} \le \cdots \le p_{(m)}\), the ordered input the rank-relative rule walks
the error-notion choice — the prior decision of which error to control, set by the scientific stakes: familywise (any false positive) for confirmatory work, false discovery rate (proportion of false positives among hits) for screening
the step-up threshold — the rank-relaxing cutoff \(p_{(k)} \le (k/m)\cdot\alpha\) and the largest-\(k\) rule, with rejection of ranks 1 through \(k\)
the FDR guarantee — the engineered warrant that the expected false-discovery proportion among rejections is bounded by \(\alpha m_0/m \le \alpha\)
the dependence precondition — the condition the guarantee rests on: independence or positive regression dependence (PRDS) of the test statistics
the q-value read-out — the per-test summary, the smallest FDR at which a given test would be declared significant
the dependence-conditioned variant — the companion extension (Benjamini–Yekutieli's harmonic-number factor) restoring validity under arbitrary dependence at the cost of power

What It Is Not¶

Not Bonferroni or familywise-error control. Bonferroni bounds the probability of any false discovery by setting the per-test threshold to α/m; BH bounds the expected proportion of false discoveries among the rejections. The two control different quantities suited to different stakes — BH deliberately accepts some false positives to recover a usable yield where FWER control at scale would eliminate nearly every true signal.
Not a guarantee that the rejected hypotheses are true. Controlling the false discovery rate means a known fraction α of the reported positives are expected to be false; the output is a shortlist for follow-up, not a set of confirmed truths. Reading every BH rejection as established is the central over-read the rate-based framing exists to forbid.
Not control of a count. FDR bounds a proportion of false discoveries among discoveries, not their absolute number. A large rejection set can contain many false positives while still satisfying the rate; conflating the controlled proportion with a guaranteed small count misstates what the procedure delivers.
Not valid under arbitrary dependence. The bound αm₀/m ≤ α holds under independence or positive regression dependence (PRDS) of the test statistics. Applying plain BH to arbitrarily — perhaps negatively — dependent tests and trusting the stated bound is unwarranted; the Benjamini–Yekutieli variant restores validity under any dependence at the cost of power.
Not the false-discovery-rate concept itself. BH is one specific, reconstructable recipe for enforcing FDR control; the portable idea — control the rate of false positives among positives rather than the probability of any — is the broader concept, of which Storey's q-value, Benjamini–Yekutieli, and knockoffs are sibling enforcers. The step-up algorithm is not synonymous with the error notion it implements.
Not appropriate for confirmatory single-claim work. When even one false positive is costly and a single hypothesis carries the claim, the right target is the familywise error rate, not FDR. BH's relaxation is built for exploratory screening and discovery at scale; using it where a definitive single result is at stake controls the wrong error.

Scope of Application¶

Because the Benjamini–Hochberg procedure is a named statistical algorithm — a step-up rule mapping a sorted list of p-values to a rejection set with bounded false-discovery proportion — not a mechanism, it is not bounded to a subject matter: it applies wherever its precondition holds, namely a population of many hypothesis tests evaluated together with statistics independent or positively dependent (PRDS). The applications below are real uses of the identical algorithm on the same \((k/m)\cdot\alpha\) threshold, not analogies; the boundary to respect is instrument-reach versus over-reading (FDR controls a proportion, not a count, and the bound fails under arbitrary dependence — switch to Benjamini–Yekutieli). The genuine scope is the multiple-testing subfield of high-throughput statistics.

Genomics and bioinformatics — the enabling application: differential-expression analysis across tens of thousands of genes routinely reports "significant at FDR 0.05" via BH-adjusted p-values.
Neuroimaging — voxel- and cluster-level fMRI inference uses BH and its descendants to manage hundred-thousand-test multiplicity.
Mass-screening clinical and epidemiological studies — prioritising among many candidate biomarkers or risk factors tested together.
Astronomy and physics — source detection, transit detection, and signal-candidate searches across many parallel tests.
At-scale A/B testing — tech platforms running thousands of parallel experiments adopt FDR control for their reporting layer.

Clarity¶

The BH procedure makes legible a distinction that single-test thinking hides: familywise error rate — the probability of even one false discovery — versus false discovery rate — the expected proportion of false discoveries among the discoveries actually reported. Before this distinction is drawn, "control your false positives at multiple testing" sounds like a single demand with a single right answer (Bonferroni), and a researcher who watches every true signal vanish under α/m at twenty thousand genes can only conclude that rigorous inference and discovery are simply incompatible at scale. Naming FDR dissolves that false dilemma: it reveals that FWER was the wrong target for screening, not that screening is hopeless. The sharper question a practitioner can now ask is not "how do I correct for multiplicity?" but "which error am I willing to control here — the chance of any false positive, or the rate of them among my hits?" — a choice that should track the scientific stakes (confirmatory single-claim work versus exploratory discovery) rather than be settled by reflex.

It also makes the rank-dependent logic of the threshold inspectable. By sorting the p-values and applying a cutoff that relaxes linearly with rank, \(p_{(k)} \le (k/m)\cdot\alpha\), the procedure exposes exactly why a given test clears the bar: not because its p-value beat a fixed constant, but because of where it sits in the ordered list and how much evidence accumulated beneath it. This turns multiplicity from an opaque penalty into a reconstructable accounting — every researcher can walk the same list, find the same step-up index, and read off both which hypotheses survive and the rate at which false discoveries are expected among them. The q-value (the smallest FDR at which a test would be declared significant) becomes a directly interpretable per-test summary of that accounting.

Manages Complexity¶

Before the FDR framing, the multiplicity problem at scale presents as an open-ended bargaining question that has to be re-litigated for every study. A genomics screen of twenty thousand genes, an fMRI map of a hundred thousand voxels, a panel of candidate biomarkers, a layer of thousands of parallel A/B experiments — each tempts the analyst into an ad hoc negotiation over how harshly to penalize for the number of tests, how many real signals can be sacrificed to keep false ones out, whether the Bonferroni threshold that left twelve genes standing is "too conservative" and by how much it might be loosened without losing rigor. That negotiation has no principled stopping point; it invites a different fudge factor for every dataset and every reviewer, and it offers no way to state, afterward, what the reported hit list actually guarantees. The BH procedure dissolves the negotiation by fixing the quantity that gets controlled. Instead of an unbounded argument about thresholds, the analyst commits in advance to one number — the tolerable false discovery rate α — and the procedure deterministically returns the rejection set whose expected false-discovery proportion is bounded by it. The sprawl of per-study haggling collapses to a single declared parameter and a single sorted-and-stepped computation.

What the analyst tracks reduces to three things: which error notion the scientific stakes call for, the chosen rate α, and the ordered list of p-values with its length m. From those, everything downstream reads off mechanically. The stakes set the branch — confirmatory single-claim work, where even one false positive is costly, stays on the familywise branch and pays the Bonferroni α/m price; exploratory, screening, or discovery work takes the FDR branch and accepts a known proportion of false hits in exchange for a usable yield. Once the branch is FDR, α and the sorted list fully determine the outcome: walk the ranks, find the largest k with p(k) ≤ (k/m)·α, reject through k, and the guarantee that no more than a fraction α of the rejections are expected to be false comes for free, holding under independence or positive dependence. No bespoke modeling of the particular genes, voxels, or experiments is required; the substrate enters only as a list of p-values and a count. A high-dimensional, judgment-laden "how do I correct for multiplicity in this study" problem is compressed to a low-dimensional one — pick the error notion the stakes warrant, set one rate, run one step-up scan — with the qualitative outcome (which hypotheses survive, and what their collective reliability is) read directly off the sorted list rather than re-derived case by case.

Abstract Reasoning¶

The BH procedure's first move is boundary-drawing on the error notion itself — a decision that runs before any computation. Facing a batch of \(m\) tests, the analyst reasons from the scientific stakes to which error to control: if a single false claim would be costly and the work is confirmatory, the situation is on the familywise branch and pays the Bonferroni \(\alpha/m\) price; if the work is exploratory screening where the point is to harvest a usable shortlist of candidates for follow-up, it is on the FDR branch and accepts a known proportion of false positives. The characteristic inference is from the purpose of the testing to the quantity worth bounding — and the procedure's standing contribution is to make that a deliberate choice rather than a reflex, dissolving the apparent dilemma in which rigor and discovery look incompatible at twenty thousand genes by relocating the problem to the wrong-target diagnosis: FWER was never the right thing to control for a screen.

Once on the FDR branch, the move is deterministic construction: from the sorted p-values \(p_{(1)} \le \cdots \le p_{(m)}\), the length \(m\), and the chosen rate \(\alpha\), the analyst walks the ranks, finds the largest \(k\) with \(p_{(k)} \le (k/m)\cdot\alpha\), and rejects through \(k\) — and reasons forward to a guarantee that comes attached: no more than a fraction \(\alpha\) of the rejections are expected to be false. The reasoning is rank-relative, not threshold-absolute: a test clears the bar not because its p-value beat a fixed constant but because of where it sits in the ordered list and how much evidence piled up beneath it, so the same p-value can be rejected in one batch and retained in another depending on its company. This makes the verdict reconstructable — any analyst can walk the same list to the same step-up index — and lets the q-value serve as a per-test summary of the accounting: the smallest FDR at which that test would be declared significant.

The diagnostic move runs in the validation direction and concerns the guarantee's fine print. The FDR bound holds under independence or positive regression dependence (PRDS); so when the test statistics are arbitrarily — perhaps negatively — dependent, the analyst infers that the plain BH guarantee may not hold and reasons to a correction: the Benjamini–Yekutieli variant, which multiplies the threshold by a harmonic-number factor to buy validity under any dependence at the cost of power. The inference is from the dependence structure of the tests to whether the bound is trustworthy as stated, and from there to which member of the procedure's family to deploy.

Finally there is a predictive yield move that lets the analyst forecast the consequence of the choice before running it: because the threshold relaxes linearly with rank rather than collapsing to \(\alpha/m\), the FDR branch is predicted to return more discoveries than Bonferroni whenever real signals are present and densely ranked near the top — the gap between the twelve genes Bonferroni leaves standing and the larger set BH recovers is itself read off the shape of the sorted list. The reasoning runs from the abundance and ranking of true effects forward to an expected yield, so the analyst can anticipate, not merely observe, how much discovery the relaxation buys and at what known false-discovery cost.

Knowledge Transfer¶

The BH procedure is a named statistical algorithm — a step-up rule that maps a sorted list of p-values to a rejection set with a bounded false-discovery proportion — not a causal mechanism, so "mechanism within / metaphor beyond" does not apply: there is nothing to recognise in a system or to analogise across domains, only a computation to run. The construct transfers literally wherever its precondition holds — a population of many hypothesis tests evaluated together, producing p-values, with test statistics independent or positively dependent (PRDS) — and that precondition is substrate-blind, which is why the identical algorithm serves genomics and bioinformatics (FDR-0.05 differential-expression across tens of thousands of genes), neuroimaging (voxel- and cluster-level fMRI inference at hundred-thousand-test multiplicity), mass-screening clinical and epidemiological studies (candidate biomarkers and risk factors), astronomy and physics (source, transit, and signal-candidate detection), and at-scale A/B testing (thousands of parallel experiments). In every case the substrate enters only as a list of p-values and a count \(m\): sort ascending, find the largest \(k\) with \(p_{(k)} \le (k/m)\cdot\alpha\), reject through \(k\), and read off the same guarantee. There is no substrate-specific tailoring — the domain happens to be "hundreds-to-millions of parallel tests" each time, and the breadth is recurrence of one piece of machinery, not of a structural pattern under different vocabulary.

Because this is an instrument, the boundary to mark is instrument-reach versus over-reading, and it has two edges. The reach edge: the procedure applies to any well-calibrated batch of p-values regardless of what was tested, so its portability is genuinely wide but it is the portability of a tool, not of an insight about the world. The over-reading edge is in the guarantee's fine print: the FDR bound \(\alpha m_0/m \le \alpha\) holds under independence or positive regression dependence, so applying plain BH to arbitrarily — perhaps negatively — dependent statistics and trusting the stated bound is an over-read, corrected by switching to the Benjamini–Yekutieli variant (a harmonic-number factor buying validity under any dependence at the cost of power). A second over-read is forgetting that FDR controls a proportion, not a count: the output is a shortlist in which a known fraction are expected false, not a set of confirmed truths — appropriate for screening and discovery, not for a single confirmatory claim, which belongs on the familywise branch. The instrument reaches exactly as far as calibrated p-values and the assumed dependence structure, and no further.

Where a genuinely cross-domain idea is wanted, it is not the BH algorithm but the more general notions it instantiates that travel, and those are housed at the parent level. The portable concept is false_discovery_rate itself — controlling the rate of false positives among reported positives rather than the probability of any false positive — which, independent of BH specifically, has been adopted as the default at-scale error-control framework in data mining, machine learning, and the tech-industry reporting layer; alongside it sit multiple_comparisons_correction (the family of which BH is the canonical FDR member, with Bonferroni, Holm, Hochberg, BY, Storey's q-value, and knockoffs as siblings), hypothesis_testing_null_vs_alternative (the per-test framework BH wraps), and type_i_type_ii_errors (the conceptual atoms whose rate BH bounds). The portable lesson — that at scale the right thing to control may be a proportion of errors rather than their existence, chosen by the scientific stakes — belongs to the FDR concept and the multiple-comparisons family. "Benjamini–Hochberg" is the specific reconstructable recipe that enforces it, and its rank-relative step-up machinery, its q-values, and its dependence-conditioned variants stay within statistics (see Structural Core vs. Domain Accent).

Examples¶

Canonical¶

Benjamini and Hochberg's 1995 paper introduced the step-up rule; a small worked example shows it. Test m = 5 hypotheses at α = 0.05 and sort the p-values ascending: 0.008, 0.012, 0.020, 0.040, 0.200. The rank-relative thresholds (k/m)·α are 0.010, 0.020, 0.030, 0.040, 0.050. Scanning for the largest k with p(k) ≤ (k/m)·α: p(1)=0.008≤0.010, p(2)=0.012≤0.020, p(3)=0.020≤0.030, and p(4)=0.040≤0.040 all pass, but p(5)=0.200>0.050 fails. The largest passing rank is k = 4, so reject the four smallest-p hypotheses. Bonferroni, by contrast, compares every p-value to α/m = 0.010 and rejects only the single test at 0.008 — recovering one discovery where BH recovers four.

Mapped back: The five hypotheses are the test population; the ascending list is the sorted p-values. Finding k = 4 by the p(k) ≤ (k/m)·α rule is the step-up threshold in action. Choosing to bound the false-discovery proportion rather than the chance of any false positive is the error-notion choice, and the promise that ≤5% of the four rejections are expected false is the FDR guarantee — visibly looser, and higher-yield, than Bonferroni's familywise control.

Applied / In Practice¶

In genomics the procedure is the workhorse that made high-throughput discovery statistically respectable. A differential-expression study might test roughly 20,000 genes at once for a difference between diseased and healthy tissue, producing 20,000 p-values. Bonferroni control at this scale sets the per-gene threshold to 0.05/20,000 = 0.0000025, so severe that almost every real signal is lost. Standard pipelines (limma, DESeq2, edgeR) instead apply BH, reporting genes "significant at FDR 0.05." If the resulting list has 200 genes, the FDR guarantee says roughly 5% of them — about 10 — are expected to be false discoveries, an acceptable contamination rate for a candidate list bound for follow-up validation. The procedure converts an impossible multiplicity problem into a usable, quantified shortlist.

Mapped back: The 20,000 genes are the test population; choosing FDR over Bonferroni's familywise rate is the error-notion choice forced by the scale. Reporting "significant at FDR 0.05" via BH-adjusted p-values is the step-up threshold and its per-test q-value read-out. That ~10 of 200 reported genes are expected false is the FDR guarantee — a controlled proportion, not a promise every hit is real.

Structural Tensions¶

T1: Familywise error versus false discovery rate (which error the stakes warrant). The procedure's foundational move is to choose which error to control, and neither choice is universally right. Familywise control (Bonferroni's α/m) bounds the probability of any false positive — indispensable when a single false claim is costly and one hypothesis carries the conclusion, but at twenty thousand genes it drives the per-test threshold so low that essentially every true signal vanishes, eliminating discovery. FDR control accepts a known proportion of false hits to recover a usable yield — indispensable for screening, but the wrong target when a definitive single result is at stake, because a "shortlist with 5% expected false" is not a confirmed finding. The tension is that the two guarantees are not better and worse versions of one thing but answers to different questions, and picking by reflex (Bonferroni because it is "rigorous," BH because it is "modern") controls the wrong error for the actual scientific purpose. Diagnostic: Is a single false positive costly here (confirmatory — familywise), or is the goal a usable candidate list for follow-up (screening — FDR)?

T2: Discovery yield versus contamination (power bought with a known fraction of falsehoods). BH's linear-in-rank threshold recovers far more discoveries than Bonferroni — four instead of one in the worked example, hundreds instead of a dozen in a genomic screen — and it does so by deliberately admitting false positives up to the rate α. Raising α widens the rejection set and the yield, but also raises the expected fraction of the reported hits that are false; lowering it purifies the list but discards real signals back toward the Bonferroni desert. The tension is that yield and purity move together in the wrong direction: the procedure cannot give both a rich discovery set and a near-clean one, and the chosen α is a standing bet on how much downstream validation can absorb wasted follow-up. There is no setting that maximizes discovery and minimizes contamination at once; α buys one with the other. Diagnostic: Can the downstream validation pipeline absorb the false-positive fraction a higher α admits, or does the cost of chasing false leads demand a stricter rate and a smaller yield?

T3: A controlled proportion versus an uncontrolled count (5% of many is still many). FDR bounds the expected proportion of false discoveries among discoveries — and pointedly not their absolute number. The same guarantee that reassures ("only 5% expected false") permits a large rejection set to contain a large count of false positives while fully satisfying the rate: 5% of 2,000 hits is 100 false leads. The tension is that the reassurance is proportional but the follow-up cost is absolute — a lab validating candidates pays per false positive, not per percentage point — so a rate that sounds tight can deliver an operationally heavy load of wrong hits whenever the discovery set is large. Reading the controlled proportion as if it were a guaranteed small count is the central over-read the rate-based framing is built to forbid, and it is easiest to commit exactly when the procedure has been most successful (many discoveries). Diagnostic: Is the tolerable quantity of false leads a proportion of the hit list or an absolute number the follow-up can afford — and does the expected count (α times the rejection set size) fit the latter?

T4: Rank-relative verdict versus absolute significance (a test's fate depends on its company). A test clears the BH bar not because its p-value beat a fixed constant but because of where it sits in the ordered list and how much evidence piled up beneath it — so the very same p-value can be rejected in one batch and retained in another depending on which other tests it was run with. This rank-relativity is what makes the procedure powerful and reconstructable, but it also means significance is a property of the batch, not of the individual test: adding or dropping hypotheses, or changing m, can flip a borderline verdict without any change to the test itself. The tension is that a researcher's decision about which tests to include in the family — often itself a judgment call — silently determines outcomes, and two analysts partitioning the same experiments into different families can reach different conclusions about the identical hypothesis, each correctly applying BH. Diagnostic: Is the hypothesis being judged on its own evidence, or on its rank within a family whose membership was itself a discretionary choice?

T5: Plain-BH power versus dependence-robust validity (the guarantee's fine print). The clean FDR bound α·m₀/m ≤ α holds only under independence or positive regression dependence (PRDS); under arbitrary — possibly negative — dependence the stated guarantee is not assured. The Benjamini–Yekutieli variant restores validity under any dependence by multiplying the threshold by a harmonic-number factor, but that factor sharply shrinks the rejection set and sacrifices the very power that motivated leaving Bonferroni behind. The tension is that robustness to unknown dependence and discovery power trade directly: trust plain BH and you may be over-rejecting when the statistics are adversely correlated; switch to BY and you may be throwing away real signals to insure against a dependence structure that may not obtain. Since the true dependence among tens of thousands of test statistics is usually unknown, the choice is made under uncertainty about whether the plain guarantee even applies. Diagnostic: Are the test statistics credibly independent or positively dependent (plain BH holds), or possibly negatively dependent (the bound needs BY, at a power cost) — and is the dependence structure actually known?

T6: Autonomy versus reduction (a named algorithm or the FDR concept and multiple-comparisons family it enforces). "Benjamini–Hochberg" is a specific, reconstructable recipe — the rank-relative step-up scan, the q-value read-out, the PRDS-conditioned validity, the BY variant — and that machinery stays within statistics; it is a tool to run, portable to any batch of calibrated p-values but carrying no insight about the world with it. What genuinely travels as an idea is the parent it instantiates: false_discovery_rate — controlling the rate of false positives among positives rather than the probability of any — now the default at-scale error framework across data mining, machine learning, and industry reporting, independent of BH specifically. Alongside sit multiple_comparisons_correction (the family, with Bonferroni, Holm, BY, Storey's q-value, and knockoffs as siblings), hypothesis_testing_null_vs_alternative, and type_i_type_ii_errors. The tension is between a standalone canonical algorithm that earns its own name and worked examples, and the recognition that its portable lesson — at scale the right thing to control may be a proportion of errors, chosen by the stakes — belongs to the FDR concept and the correction family, of which BH is one enforcer. Diagnostic: Resolve toward the parents (false_discovery_rate, multiple_comparisons_correction) when stating the portable lesson; toward "Benjamini–Hochberg" when specifying the exact recipe that maps a sorted p-value list to a rejection set.

Structural–Framed Character¶

The Benjamini–Hochberg procedure sits at the mixed position on the structural–framed spectrum, leaning a touch more framed than a discovered regularity like Benford's law because it is an invented recipe rather than an observed fact — yet held back from framed-leaning by two genuine structural marks. On evaluative_weight it is essentially neutral: the procedure renders no verdict on the world; it is a determinate computation that maps a sorted p-value list to a rejection set, praising and convicting nothing. It does carry a designed purpose (control the false-discovery rate) the way any instrument does, but that is instrumental fitness, not the normative classification a fallacy label pronounces. On its determinacy it also patterns structural: it is a substrate-blind algorithm that runs identically on genes, voxels, transits, or A/B experiments — the domain enters only as a list of p-values and a count — which is why, like Benford's test, it "transfers literally wherever its precondition holds." Those two marks keep it off the framed pole.

The other three criteria pull framed. Human_practice_bound points framed and is the decisive difference from Benford: there is no observer-free version of BH — it has no existence in nature the way lithospheric rebound or the leading-digit curve of river lengths does, because it is a piece of methodology that exists only inside the human practice of statistical hypothesis testing; remove that practice and there is no sorted p-value list, no error to control, nothing for the step-up rule to act on. Institutional_origin points framed in the same breath: the procedure was invented (Benjamini and Hochberg, 1995) as a deliberate artifact of a methodological tradition, not discovered as a fact the world already instantiated — even its very object, the false-discovery rate, is a constructed error notion chosen to fit the scientific stakes. Vocab_travels is pinned to statistics: p-values, FDR, PRDS dependence, q-values, the step-up threshold, the BY harmonic-number correction all carry their content only inside the multiple-testing frame; the algorithm's portability is the portability of a tool to any calibrated p-value batch, not of an insight that re-instantiates as mechanism under new vocabulary. And import_vs_recognize patterns as literal instrument-reuse within statistics rather than cross-domain analogy — but the genuinely cross-domain idea is not BH at all.

The portable structural skeleton is rate-control-rather-than-existence-control — bounding the expected proportion of false positives among reported positives instead of the probability of any false positive, a general error-management stance that recurs across data mining, machine learning, and industry reporting layers. That skeleton is genuinely portable, and it is exactly what BH instantiates from its umbrella primes — false_discovery_rate (the concept), multiple_comparisons_correction (the family), and the hypothesis_testing/type_i_type_ii_errors atoms beneath them — not what makes "Benjamini–Hochberg" itself travel: the cross-domain reach belongs to the FDR concept, while the procedure's distinctive content — the rank-relative step-up scan, the q-value read-out, the PRDS fine print, and the dependence-robust variant — is precisely the statistics-bound recipe that stays home. Its character: an evaluatively neutral, substrate-blind computational recipe, but an invented one constituted by and expressed in the vocabulary of statistical methodology, whose only cross-domain content is the rate-control idea already carried, in general form, by the false-discovery-rate and multiple-comparisons primes it enforces.

Structural Core vs. Domain Accent¶

This section decides why the Benjamini–Hochberg procedure is a domain-specific abstraction and not a prime, and it carries the case for its domain-specificity — there is no separate section for that.

What is skeletal (could lift toward a cross-domain prime). Strip the statistics and one portable stance survives: when many yes/no decisions are made at once, control the expected proportion of wrong "yes" calls among the calls made, rather than the probability of any wrong call at all. Rate-control-rather-than-existence-control — an error-management posture that trades a guarantee about the existence of any false positive for a guarantee about their fraction, with the choice between the two set by the stakes. That skeleton is genuinely portable: it recurs across data mining, machine learning, and industry reporting layers as the default at-scale error framework, which is exactly why it is housed in the catalog as the parents BH instantiates — false_discovery_rate (the concept), multiple_comparisons_correction (the family), with hypothesis_testing_null_vs_alternative and type_i_type_ii_errors as the atoms beneath. But it is the core BH shares, not what makes "Benjamini–Hochberg," the named entity, distinctive.

What is domain-bound. Everything that makes it the BH procedure in particular is statistical methodology, and none of it survives extraction. The specific recipe — sort the p-values ascending, find the largest rank \(k\) with \(p_{(k)} \le (k/m)\cdot\alpha\), reject through \(k\) — is a rank-relative step-up scan defined over an object (a calibrated p-value list of length \(m\)) that exists only inside hypothesis testing. The FDR guarantee \(\alpha m_0/m \le \alpha\), the PRDS dependence precondition it rests on, the q-value read-out, and the Benjamini–Yekutieli harmonic-number variant that restores validity under arbitrary dependence at a power cost are all discipline-internal machinery. Even the object it controls — the false-discovery rate — is a constructed error notion, not a fact the world instantiates. The decisive test: remove the p-values and the multiple-testing frame and there is no step-up rule to run — what remains is the bare idea of bounding a proportion of errors, a looser thing already named by its parents.

Why this does not clear the prime bar. A prime's vocabulary travels and its transfer is recognition of the same mechanism, not analogy; BH is a still-clearer non-prime because it is not even a mechanism but an invented computation, and its transfer is bimodal in the way the bar cares about. Within statistics the algorithm transfers literally wherever its precondition holds — the identical step-up scan serves genomics, neuroimaging, mass-screening clinical work, astronomy, and at-scale A/B testing, the substrate entering only as a list of p-values and a count \(m\); that breadth is the portability of a tool, the recurrence of one piece of machinery, not a structural pattern recognized under new vocabulary. Beyond the multiple-testing frame the recipe does not carry at all: there is no p-value list to sort, so any cross-domain use renames components or reaches only by analogy. When the genuinely portable idea is wanted — that at scale the right thing to control may be a proportion of errors rather than their existence, chosen by the stakes — it is already carried, in general form, by false_discovery_rate and multiple_comparisons_correction (with the hypothesis_testing_null_vs_alternative and type_i_type_ii_errors atoms beneath). The cross-domain reach belongs to those parents; "Benjamini–Hochberg," as named — the rank-relative scan, the q-values, the PRDS fine print, the BY variant — carries statistics-bound methodology that does not and should not travel.

Relationships to Other Abstractions¶

Current abstraction Benjamini–Hochberg Procedure Domain-specific

Parents (2) — more general patterns this builds on

Benjamini–Hochberg Procedure is a kind of Multiple Comparisons Correction Prime

Benjamini-Hochberg is multiple-comparisons correction specialized to a rank-based step-up rule that controls false discovery rate rather than familywise error rate.
Benjamini–Hochberg Procedure presupposes False Discovery Rate Domain-specific

Benjamini Hochberg Procedure presupposes False Discovery Rate.

Hierarchy paths (12) — routes to 6 parentless roots

Benjamini–Hochberg Procedure → Multiple Comparisons Correction → Type I & Type II Errors → Hypothesis Testing (Null vs. Alternative) → Statistical Inference → Inductive Reasoning

Show alternative paths (11)

Not to Be Confused With¶

Bonferroni and the family-wise-error corrections (Holm, Šidák). Procedures that bound the probability of even one false positive across a family of tests — Bonferroni by comparing every p-value to α/m, Holm by a step-down refinement. They are siblings of BH in the multiple-comparisons family but target a different error: the familywise error rate, not the false-discovery proportion. At high-throughput scale their thresholds annihilate nearly every true signal, the regime BH was built for. Tell: Is the guarantee "almost surely no false positive at all" (Bonferroni family, FWER), or "at most a fraction α of the reported hits are false" (BH, FDR)?
Hochberg's step-up procedure (Y. Hochberg, 1988). A genuinely confusable near-namesake: also a step-up rule that walks sorted p-values from the largest down, but one that controls the familywise error rate, not the false-discovery rate. The shared "Hochberg" and shared step-up mechanics invite conflation, yet the error notion each bounds differs. Tell: Does the step-up rule guarantee no false positive at all (Hochberg's FWER procedure), or a bounded false-discovery proportion (Benjamini–Hochberg)?
The Benjamini–Yekutieli procedure. The dependence-robust member of the same family, which multiplies the threshold by a harmonic-number factor to keep the FDR bound valid under arbitrary (even negative) dependence, at a cost in power. It is a subtype-sibling of BH for the case where BH's independence/PRDS precondition fails, not a separate error notion. Tell: Are the test statistics independent or positively dependent, so plain BH's bound holds (BH), or possibly negatively dependent, requiring the harmonic-factor correction (Benjamini–Yekutieli)?
Storey's q-value / positive FDR. An alternative route to FDR control that estimates the proportion of true nulls (π₀) and reports, per test, the smallest FDR at which it is significant. BH produces adjusted p-values by a fixed rank-relative rule; Storey's approach estimates a quantity to gain power. They are sibling enforcers of the same FDR concept, easily confused because BH's own per-test read-out is also called a q-value. Tell: Is the significance threshold set by the deterministic \((k/m)\cdot\alpha\) step-up (BH), or by an estimate of the true-null fraction feeding a pFDR/q-value (Storey)?
The false-discovery-rate concept and the multiple_comparisons_correction family it enforces (parent umbrella). The portable idea — control the rate of false positives among positives rather than the probability of any, chosen by the scientific stakes — plus the family of which BH is the canonical FDR member. Treated more fully as those parents elsewhere; BH is the specific reconstructable recipe that enforces the concept, not the concept itself. Tell: Are you stating the portable lesson about which error to control at scale (false_discovery_rate / the correction family), or specifying the exact rule that maps a sorted p-value list to a rejection set (Benjamini–Hochberg)?

Neighborhood in Abstraction Space¶

Benjamini–Hochberg Procedure sits in a moderately populated region (54^th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.

Family — Statistical Inference & Model Failure Modes (16 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-07-12