Certainty of Evidence vs Strength of Recommendation

Nucleus of the evidence-appraisal cluster. How confident you are in an effect estimate and how strongly you should recommend acting on it are two separate judgments, and neither determines the other. GRADE calls the separation “the defining feature” of its approach and attributes the confusion in earlier grading systems to their failure to make it. [EXTRACTED (GRADE - Handbook) §1.2]

Ordinary framing

Better evidence -> stronger recommendation. Grading systems that produce a single letter or star rating collapse the two, so a reader cannot tell whether a weak recommendation reflects thin evidence or a genuinely close call on well-established effects.

The reframing — certainty is ONE of four inputs to strength, and the mapping is asymmetric

The separation is not merely record two numbers: certainty is one of four determinants of strength, not a proxy for it (the other three are set out below, under What sets strength, once certainty is fixed). But the two off-diagonal cases are not symmetric, and the page must not flatten them:

• High certainty -> weak (conditional) recommendation is common: panels facing important variability in values and preferences “are likely to offer a weak recommendation despite high quality evidence.” [EXTRACTED (GRADE - Handbook) §1.2]
• Low or very low certainty -> strong recommendation is allowed but rare. GRADE “allows for” it where “the benefits of antibiotic therapy clearly outweigh the downsides in most patients independent of the quality assessment” — a clause about one worked example, not a general licensing condition — while separately discouraging panels from making strong recommendations on low or very low confidence, confining the legitimate cases to a small set of paradigmatic situations. [EXTRACTED (GRADE - Handbook) §1.2, §6.3.2]

Mechanism — what drives strength once certainty is fixed

Strength is set by the balance of desirable and undesirable consequences, the variability of values and preferences among those affected, and resource use. Where patients with differing values would predictably make different choices, panels are likely to offer a weak recommendation even on high quality evidence — because a strong recommendation asserts that most or all would choose the same way. [EXTRACTED (GRADE - Handbook) §1.2]

Worked cases the handbook gives:

• High certainty, weak recommendation. Chemotherapy plus radiotherapy versus radiotherapy alone in locally advanced non-small-cell lung cancer: high quality evidence of a survival gain “of a few months” against added harm and burden — so values determine the choice.
• High certainty, weak recommendation. Continuing long-term warfarin after a first unprovoked deep venous thrombosis: high quality trials show reduced recurrence at the cost of bleeding and inconvenience; “because patients with varying values and preferences will make different choices,” the recommendation is weak.
• Low certainty, strong recommendation (GRADE’s counterfactual illustration, not a reported case). Rapid antibiotics in severe infection or sepsis has never been randomized against delay; on observational evidence rated low, panels “would be very likely to” make a strong recommendation. [EXTRACTED (GRADE - Handbook) §1.2]

Decision relevance

• A weak recommendation is not a hedge about the science. It is a statement that the right action depends on the person’s values. Reading it as the evidence is poor inverts its meaning — and in two of GRADE’s three worked examples in §1.2 the evidence was high quality. (Not to be confused with the five paradigmatic situations of §6.3.2, which are the strong-on-low-certainty cases — none of those is high quality.)
• Conversely, low certainty does not license inaction — but the exit is narrow. A strong recommendation on low-certainty evidence is legitimate where the balance is lopsided enough, and GRADE otherwise discourages it, so treating it as freely available inverts the rule as surely as ignoring it does.
• The practical test when reading any recommendation: ask which of the two judgments produced its strength. If the source does not let you tell them apart, you cannot use its strength rating.

What sets strength, once certainty is fixed (chunk 03)

Strength is “the extent to which a guideline panel is confident that desirable effects of an intervention outweigh undesirable effects, or vice versa, across the range of patients for whom the recommendation is intended.” Four factors drive direction and strength: the balance of desirable and undesirable consequences, certainty in the effect estimates, confidence in values and preferences, and resource use. [EXTRACTED (GRADE - Handbook) §6.1, §6.2]

• Strong — the panel is confident the desirable effects outweigh the undesirable (or the reverse); “most or all individuals will be best served by the recommended course of action.”
• Weak — the desirable effects probably outweigh the undesirable “but appreciable uncertainty exists”; “not all individuals will be best served,” so the individual’s circumstances and preferences need more careful weighing than usual.
• Default under uncertainty is weak: where the panel is unsure whether the balance is clear, or the information bearing on the four factors is unavailable, “it would opt to make a weak recommendation.” [EXTRACTED (GRADE - Handbook) §6.1, §6.1.1, §6.1.2]

On the balance itself: large relative effects pointing the same way favour a strong recommendation; large desirable effects accompanied by large undesirable ones favour a weak one. Large absolute effects favour strength more than small ones — and since baseline risk drives absolute effect, “the strength of recommendations and its direction will likely differ in high- and low-risk groups” (Baseline Risk and the Relative-Absolute Split). [EXTRACTED (GRADE - Handbook) §6.2.1.1]

Two cautions GRADE states about its own output

• “Strong recommendations are not necessarily high priority recommendations.” [EXTRACTED (GRADE - Handbook) §6.1.1] Strength expresses confidence in the direction of net benefit, not the size of the benefit or its rank against other things a person could do. An intervention hierarchy therefore cannot be read off recommendation strength — a confidently-correct small effect and a confidently-correct large one are both “strong.” [INFERRED (GRADE - Handbook) — entailment: strength is defined on direction-confidence, priority on magnitude; the handbook flags the non-identity without developing it.]
• Recommendations are not dictates. “Even strong recommendations based on high-quality evidence will not apply to all circumstances and all patients,” and users “may reasonably conclude that following some strong recommendations based on the high quality evidence will be a mistake for some patients,” because no guideline “can take into account all of the often compelling unique features of individual patients.” Applying recommendations “by rote or in a blanket fashion” is explicitly disclaimed. [EXTRACTED (GRADE - Handbook) §6.1]

Applied — WHO SFA 2023, and how the four determinants show up in a real rationale

WHO’s 2023 SFA guideline is GRADE applied by an adopting body, and it exhibits the certainty->strength linkage cleanly (Saturated Fat Intake and Replacement):

Recommendation	Certainty	Strength
Reduce SFA to 10% energy	moderate	strong
Reduce below 10%	very low	conditional
Replace with PUFA	moderate	strong
Replace with MUFA / carbohydrate	low	conditional

The instructive case is the sub-10% recommendation. WHO issued it as conditional while recording that “no undesirable effects or other mitigating factors were identified that would argue against” it — so the conditionality is not a balance judgment at all. It is driven by certainty alone: “a conservative approach was therefore taken, leading to a conditional recommendation.” [EXTRACTED (WHO - Saturated and Trans Fatty Acid Intake 2023) Rationale for SFA recommendations 1 and 2]

• CORRECTED 2026-07-26 — this is GRADE’s default rule, not a new pattern. It was previously written up here as a fifth pattern alongside GRADE’s own paradigm cases. That claim is refuted by a quotation this same page carries above: where a panel “is uncertain whether the balance is clear” or the information bearing on the four factors “is not available,” it “would opt to make a weak recommendation.” WHO rated very low certainty and issued a conditional recommendation — the default executing normally. Retained below as a worked illustration of that rule, which is what it is good for; the paradigm-case framing is struck.
• This is a fifth pattern alongside GRADE’s own paradigm cases: not high-certainty-weak (values vary) nor low-certainty-strong (balance lopsided), but low-certainty-conditional-despite-no-known-harm — where the panel could identify nothing against the action and still declined to recommend it strongly. Under GRADE’s four factors this is certainty operating almost alone, which the handbook permits but does not illustrate. [INFERRED (GRADE - Handbook; WHO - Saturated and Trans Fatty Acid Intake 2023) — entailment from the four-factor rule applied to WHO’s stated rationale.]
• Reading consequence: a conditional recommendation cannot be assumed to mean “the benefit is contested.” Here it means “we could not rule it out and could not confirm it” — a different message with a different action implication.

Limits

• GRADE does not remove judgment, and says so. “Applying the GRADE approach does not minimize the importance of judgment or [suggest] that quality can always be objectively determined”; there will be cases of “legitimate disagreement about the interpretation of evidence.” Its claim is that judgments become explicit rather than implicit. [EXTRACTED (GRADE - Handbook) §1.1]
• The framework governs guideline panels offering “recommendations for the management of typical patients.” [EXTRACTED (GRADE - Handbook) §1.1] Carrying a strength rating to a specific individual requires a transportability step the rating itself does not perform. [INFERRED (GRADE - Handbook) — entailment from the population-level framing.]
• Source currency: this handbook (updated October 2013) is being phased out in favour of the GRADE Book (2024); §§1-3 are flagged in-source as already rewritten there. [EXTRACTED (GRADE - Handbook) front matter]

A real process where appraisal and recommendation are done by DIFFERENT PEOPLE [2026-07-27]

This page treats certainty and strength as two judgements a guideline panel makes. The US Dietary Guidelines process separates them institutionally — and NASEM records that as an atypical inherited feature, not a design it endorses:

«Multiple sets of standards exist that could serve as models, with the understanding that the process for developing DGA recommendations does not follow typical guidelines development processes because the experts assessing the evidence do not write the guidelines and recommendations themselves.» [EXTRACTED (NASEM - Redesigning the DGA Process) ch.3, footnote 6]

Why this matters beyond the DGA. The certainty/strength split is normally a distinction within one panel’s reasoning; here it is a handoff between two bodies, which creates a failure mode the framework above does not describe: the recommendation can depart from the appraisal silently, because no one person owns both.

NASEM’s response is not to close the seam but to make it auditable — Recommendation 2 asks that «the secretaries of USDA and HHS should provide the public with a clear explanation when the DGA omit or accept only parts of conclusions from the scientific report.» [EXTRACTED (NASEM - Redesigning the DGA Process) ch.3]

That the safeguard is requested is the finding. A recommendation to explain departures implies departures occur and are not currently explained. What this does NOT establish is that any specific recommendation was distorted — NASEM asks for transparency about the handoff, it does not document a case. [INFERRED (NASEM - Redesigning the DGA Process) — the silent-modification failure mode is the wiki's reading; NASEM assembles the parts without naming it] -> [[Which Objective Moved This Recommendation]]

A third axis the two-axis frame does not hold: how PRECISE is the number [2026-07-28]

NNR - Nordic Nutrition Recommendations 2023 attaches no strength or certainty grade to any recommendation it issues — not one, across 374 pages. [searched: "strong recommendation"/"conditional recommendation"/"GRADE approach"/"certainty of evidence" across all 10 NNR chunks — 1 hit total, and it is NNR describing WHO's grading of WHO's own sugars guideline, not grading its own]

That alone would only make NNR a third ungraded body. What makes it worth a section is the reason NNR gives, which introduces a distinction this page did not hold:

«FBDGs are formulated more generally than the DRVs for nutrients, although the causal associations of foods and health outcomes can be stronger than for nutrients and health outcomes. As with DRVs, there are seldom precise calculations behind the quantitative FBDGs. The precise FBDGs are based on best scientific knowledge and most often decided as consensus among expert groups.» [EXTRACTED (NNR - Nordic Nutrition Recommendations 2023) chunk 02]

Read the first clause before the third, because it reverses the obvious reading. This is not a body confessing that its evidence is weak. NNR says the causal association behind a food-based guideline can be stronger than the nutrient-level one — and then says the number attached to it is a consensus rather than a calculation. Those are claims about two different things.

Axis	Question	Held by
Certainty	how confident are we that the effect is real and about this big?	GRADE, WHO
Strength	how strongly should we recommend acting on it?	GRADE, WHO
Precision of the threshold	is the specific number derived, or agreed?	NNR, explicitly

The third is independent of the first two, and that is the finding. A body can be highly certain that whole-grain intake lowers risk, recommend it strongly, and still have picked 90 g/day by expert consensus rather than derived it from a curve. Certainty about a relationship does not transmit to precision about a cutpoint — and a reader who sees only a strong recommendation attached to a round number will assume it does. [INFERRED (NNR - Nordic Nutrition Recommendations 2023) — the three-axis decomposition is this page's; the certainty/precision split is stated by NNR in the passage above]

This sharpens a rule the corpus already carries rather than contradicting it. The dose-response section of the telos holds that where a threshold appears in guidance, the first hypothesis is that it marks the edge of the evidence, not a feature of the curve. NNR supplies the mechanism for that hypothesis in a guidance body’s own voice: the thresholds are «most often decided as consensus among expert groups». That is a guidance body stating the general case, where SACN’s 30 g was one worked instance.

What this does NOT establish. NNR is describing how quantitative FBDGs are set in general, including its own; it is not conceding that any particular NNR number is arbitrary, and it explicitly grounds them in «best scientific knowledge». Consensus-derived is not evidence-free — it is un-graded, which is a different and narrower claim.

Where this leaves the three ungraded bodies the wiki now holds. SACN issues no strength axis; NASEM wants GRADE’s properties without the brand; NNR grades nothing and says why. None of these is evidence for the others — three bodies independently declining to grade is a fact about the field’s norms, not a converging judgment about grading’s value, and no [E-independent] is claimed.

Self-critique of the NNR addition [run 2026-07-28, before commit]

• Quote-truncation check: RUN, and it caught the error before the prose was written. The usable sentence is «there are seldom precise calculations behind the quantitative FBDGs». Quoted alone it reads as a confession of weak evidence. The sentence before it says the opposite — food-level causal associations «can be stronger» than nutrient-level ones. The full passage is quoted and the paragraph explicitly instructs the reader to take the first clause first. This is the same failure shape as the four defects logged in the 2026-07-27 session, caught by running the check rather than by noticing.
• Absence claim: SCOPED, not asserted. “NNR grades nothing” is backed by a stated four-term search across all ten chunks with the single hit explained (NNR describing WHO). The earlier form of this claim on the source page carried no search scope.
• Over-claim check on the third axis. The claim is that precision of the threshold is independent of certainty and strength. NNR states the certainty/precision split; the three-axis table and the independence claim are the wiki’s and are tagged. A weaker reading is available — that NNR is merely restating low certainty in other words — and the first clause of the quote is what rules it out, which is why the quote could not be truncated.
• Laundered independence: explicitly denied in the body. Three ungraded bodies are not three witnesses; the closing paragraph says so rather than leaving the count to imply corroboration.
• Residual: NNR’s passage is about quantitative FBDGs generally. Applying it to NNR’s own numbers is warranted by «As with DRVs» and by NNR issuing no grades, but NNR never says “our numbers are consensus” in the first person.

Does the separation actually DO anything? The corpus’s own pairings [2026-07-28]

This page holds GRADE’s claim that certainty is one of four determinants of strength and not a proxy for it, and that the off-diagonal case — high certainty, weak recommendation — is common. The corpus now holds enough graded recommendations to check whether that separation shows up in practice. It does not.

Read the sample bound FIRST, because it decides what this can mean. Every paired certainty-and-strength recommendation the wiki holds comes from two documents, both WHO. The PA evidence-profile annex carries per-outcome certainty but no recommendation strengths (it is the annex, not the guideline), so it contributes no pairs. This is one body’s house practice, not a finding about guideline-making.

Source	Recommendation	Strength	Overall certainty
WHO SFA 2023	reduce SFA to 10%E	strong	moderate
WHO SFA 2023	further reduce to <10%E	conditional	very low
WHO SFA 2023	replace with PUFA	strong	moderate
WHO SFA 2023	replace with MUFA (plant)	conditional	low
WHO SFA 2023	replace with carbohydrate from fibre-containing foods	conditional	low
WHO Sugars 2015	free sugars <10%E	strong	moderate (observational, caries)
WHO Sugars 2015	free sugars <5%E	conditional	very low (ecological)

[EXTRACTED (WHO - Saturated and Trans Fatty Acid Intake 2023) chunk 01 — «The overall certainty in the evidence for recommendation 1 was moderate, and for recommendation 2 was very low»] [EXTRACTED (WHO - Sugars Intake 2015) Remarks, p.16]

Seven recommendations, seven diagonal cells, zero off-diagonal. Strong pairs with moderate every time; conditional pairs with low or very low every time. The case GRADE calls common — high certainty, weak recommendation — occurs zero times, and so does the rare one it discourages.

What this does and does not license

• It is NOT “WHO collapsed certainty into strength.” If the other three determinants — balance of benefits and harms, variability in values and preferences, resource use — happen not to vary much across population nutrient-threshold recommendations, then a monotone mapping is the correct output, not a defect. Population thresholds for a nutrient are close to a best case for stable values: no procedure, no burden, no cost to the individual of the recommendation itself.

• What it DOES establish is a reader-side problem, and that is the decision-relevant part. From the output alone you cannot tell which of the two reasons produced a conditional:

  • conditional because the evidence is thin -> more research could move it, and the reader’s own values are not what is in question;
  • conditional because values genuinely vary -> more research will NOT move it, and the reader’s values are exactly what decides it.

  Those license different actions, and the label is identical in both cases. Where the mapping is monotone, the strength axis carries no information the certainty axis did not already carry — so in this corpus, reading only the certainty would lose nothing.

• The rationale text is where the distinction survives, when it survives at all. WHO SFA’s strength rationale for the strong PUFA recommendation cites «moderate certainty overall» — i.e. it reaches for certainty to justify strength rather than for the other three determinants.

[INFERRED (WHO - Saturated and Trans Fatty Acid Intake 2023; WHO - Sugars Intake 2015) — every cell is quoted from its source; the monotonicity observation and the reader-side consequence are this page's]

The falsifier, stated in advance

A guideline whose recommendations vary in values/preferences should produce the off-diagonal — screening decisions, drugs carrying burden or side-effects, interventions with real opportunity cost. AWAITS ADA-EASD - Nutrition Therapy for Diabetes 2019 and NASEM - Clinical Practice Guidelines We Can Trust; if either shows high-certainty conditional recommendations, this section’s pattern is WHO-and-nutrition-specific and must be narrowed to say so. Recorded as a prediction rather than a conclusion, because a 7-cell single-body sample is exactly the size at which a pattern is most tempting and least supported.

Two bodies that do not have the axis at all

• SACN - Carbohydrates and Health 2015 issues no strength axis whatsoever — its bespoke instrument grades evidence and stops, so the certainty/strength question cannot even be posed of it.
• WHO - Sodium Intake 2012 issues strength without certainty — every recommendation is «strong», the word conditional does not appear in the document, and no certainty rating is attached to any recommendation line. Certainty exists only in its per-outcome profiles.

Note what the second one means for the table above: WHO’s own practice is not constant across its guidelines. The 2012 sodium guideline attaches no overall certainty; the 2023 SFA guideline states one per recommendation. So the monotone pattern is a property of WHO’s recent guidelines, and the corpus can see the practice changing — which is a stronger reason than sample size to refuse to generalise it.

Self-critique [run 2026-07-28, before commit]

• The sample bound is stated before the table, not after it. Seven cells from two documents by one body is the kind of evidence that reads as a law. The section is structured so a reader cannot reach the pattern without first reading what it rests on.
• Steelmanned the null reading, and it survives. “Monotone mapping is the correct output when the other three determinants do not vary” is not a hedge — it is a rival explanation that the evidence here cannot rule out, and it is given its own bullet rather than buried.
• Falsifier named in advance with two specific artifacts, so this cannot be quietly retained if contradicted. Per the AWAITS convention, both are linkable and will surface via reconcile.py awaits-arrived when ingested.
• Type check: A, and single-body. This is an emergent observation across sources, present in none of them — but the sources are not independent (one body, overlapping committees), so no [E-independent] and the claim is about a practice, not about the world.
• Nucleus-coverage discipline observed. This section was written because the missing sources had something to say on this page’s question; the facet sources that do not (Naude, Goldenberg, Moynihan, Te Morenga, Willett) were left out of sources: rather than rolled up — which is what the generator’s own guard requires.
• Residual: the strongest unverified step is that WHO’s five SFA strength ratings and two sugars ratings were each set independently. If a single committee applied one implicit rule across both, the seven cells are closer to one observation than to seven.