Why it matters
Outcomes people care about are often rare or slow, so investigators measure something faster and commoner instead — a marker standing in for the outcome. Substitution generally costs certainty, and GRADE restricts its use rather than merely asking that it be recorded: surrogates are for cases where evidence on population-important outcomes is lacking. (Schünemann et al., n.d.)
The rule
GRADE’s conditions, stated tightly:
- Consider surrogates only when evidence about population-important outcomes is lacking — not as a first-line convenience.
- When one is used, specify the population-important outcome and, separately, the surrogate substituting for it.
- “Guideline developers should not list the surrogates themselves as their measures of outcome.” The outcome remains the thing that matters; the surrogate is a route to estimating it.
- Substituting “may ultimately lead to rating down the quality of the evidence because of the indirectness”; §5.2.3 puts it more firmly — “in general, the use of a surrogate outcome requires rating down … by one, or even two, levels.” (Schünemann et al., n.d.)
How GRADE grades a surrogate — causal-pathway proximity
The first of two criteria the handbook supplies, and the most decision-relevant thing it says about surrogates: the penalty scales with distance along the putative causal pathway to the patient-important outcome.
- Far from the endpoint -> rate down two levels. GRADE’s case is calcium and phosphate metabolism standing in for patient-important outcomes in renal disease.
- Close to the endpoint -> rate down one level. Coronary artery calcification for myocardial infarction; bone mineral density for fractures; soft-tissue calcification for pain.
- The judgment draws on “consideration of the biology, mechanism, and natural history of the disease.” (Schünemann et al., n.d.)
So a marker’s standing is not binary. How far down the chain it sits is the question, and it is answered biologically rather than statistically.
The second criterion — does the chain actually transmit?
Proximity is not the only test the handbook supplies. Immediately after the proximity passage it adds a validation criterion, addressing whether the surrogate predicts the outcome at all:
“Example 9: Uncertainty in the relationship between surrogate and Surrogate outcomes (Downgraded by One or Two Levels) — Investigators examined the ‘validity’ of progression-free survival as a surrogate for overall survival… They found a statistically significant association between progression-free and overall survival in the randomized trials they analyzed, but predicting overall survival using progression-free survival remained uncertain. Rating down quality by one level for indirectness would be appropriate in this situation.”
This is the criterion that does the work the proximity rule cannot. A significant marker-outcome association was not sufficient — what GRADE asks is whether the outcome can be predicted from the marker, which is a strictly stronger demand than correlation. A marker can be proximate, correlated, and still fail here.
Two further treatments elsewhere in the handbook point the same way: that “even if well measured surrogates are available, confidence in estimates of effects on patient-important outcomes is very likely to be low”, and a dedicated list of “Key questions when using test accuracy as a surrogate”.
Tests / indicators
- Is the outcome named separately from the marker? A recommendation phrased directly in terms of a marker (lower X to below Y) has collapsed the distinction and cannot be checked.
- Was certainty rated down for indirectness — and if not, is a reason on the record? GRADE’s rule is “in general”, with an explicit judgment clause, so a non-downgrade is not automatically a defect. Three cases, and only the first two are: the surrogate was silently treated as the outcome; the downgrade was skipped with nothing said; or a stated, biology-based judgment was made that the chain transmits. The third is what GRADE invites — WHO’s LDL case below is a worked instance. The test is whether a reason was given, not whether the downgrade was taken.
- Is the marker -> outcome link itself evidenced? GRADE places surrogate use under indirectness, which presumes the link is an inference rather than an observation. The strength of that link is the whole question, and it is a separate evidential claim from the intervention -> marker effect.
Prognostic marker vs modifiable lever — the same distinction, one step earlier
The tests above assume you are already intervening on the marker. A prior version of the error appears whenever a measure merely predicts an outcome: a strong predictor is a marker to stratify with, not automatically a lever to pull. Three ways the predict->intervene step fails, and a measure that predicts can fail any of them:
- Reverse causation makes it a marker OF the outcome, not a cause of it — low grip strength, low VO2max, and low muscle mass all predict mortality partly because occult ill-health lowers them; the measure sits downstream of the thing you fear, so acting on it need not move the outcome (The U-Shaped Association Artifact carries the sick-quitter / reverse-causation machinery).
- The marker -> outcome link may be unevidenced (the surrogate test above).
- The association may not survive intervention (The Observational-Trial Discordance).
So a measure earns the “modifiable lever” label only on intervention evidence that acting on it changes a patient-important outcome; absent that, it is a prognostic metric — genuinely useful for placing someone in a stratum (the decision it does serve), never assumed to be a treatment target. Worked instances: Grip Strength and Mortality and Cardiorespiratory Fitness and Mortality (both cheap, strong predictors held explicitly as metrics-not-levers); Low Muscle Mass and Mortality (mass predicts, but strength out-predicts and neither is a proven target).
The composite-index case — allostatic load predicts mortality, no lever shown [2026-08-09, Parker]
The prognostic-marker-vs-lever cases above are single markers (grip, VO2max, muscle mass, CRP). Allostatic load (AL) is the same rule on a composite — a summed index of dysregulated biomarkers across four systems (cardiovascular, metabolic, neuroendocrine, immune). Parker’s gold SR-MA finds high AL predicts all-cause mortality (pooled HR 1.22, 95% CI 1.14-1.30, n=10) and CVD mortality (1.31, 1.10-1.57), and that «total AL scores better predict mortality outcomes than any individual AL biomarker» (Parker et al., 2022). Yet «Though no intervention studies were included» (Parker et al., 2022) — no evidence here that reducing AL reduces mortality.
The composite twist sharpens, not softens, the rule. Aggregating dysregulation makes AL a better predictor — that is the whole reason the index exists — but better prediction is not transmission: the Example-9 test (can the outcome be predicted? is strictly weaker than does-it-cause) is passed by construction and the causal question is untouched. A composite index is doubly a marker: it predicts well and sits partly downstream of occult illness (an already-sick body accumulates dysregulation), so a raised AL is partly a consequence. Prognostic (route (a) stratifier) yes; validated target no — the same line held for CRP and for muscle mass -> Allostatic Load and Mortality. (inferred from Parker et al., 2022)
The archetypal predicts-but-does-not-cause marker — CRP, settled by Mendelian randomization [2026-08-08]
C-reactive protein is a clean worked instance of the section above: a strong, log-linear predictor of coronary heart disease that a genetic natural experiment shows is not a cause — so it is a marker to stratify with, never a target to steer toward. Two large IPD meta-analyses, and the point is that they do not conflict:
| Parameter | ERFC 2010 (observational) | CCGC 2011 (Mendelian randomization) | Same quantity? |
|---|---|---|---|
| Exposure contrast | circulating CRP, per 1-SD higher usual ln CRP | genetically-raised CRP, per 1-SD higher genetic ln CRP | No — same scale, different source of variation |
| Circulating CRP -> CHD (adjusted) | RR 1.37 (95% CI 1.27-1.48) (Kaptoge et al., 2010) | RR 1.33 (1.23 to 1.43) ((CCGC) et al., 2011) | Yes — both observational -> they AGREE |
| Genetic CRP -> CHD | not estimated | RR 1.00 (0.90 to 1.13) ((CCGC) et al., 2011) | N/A — a gap on ERFC’s side |
Where the two measure the same quantity they agree; CCGC adds the genetic arm (null) that reveals the observational association is confounded — «C reactive protein concentration itself is unlikely to be even a modest causal factor in coronary heart disease» ((CCGC) et al., 2011). This is a refinement/disambiguation, not a tension: CCGC itself preserves the prediction claim — «Our findings also do not address the separate issue of the value of measurement of circulating C reactive protein in prediction of long term vascular risk» ((CCGC) et al., 2011). The Example-9 test this page centres (can the outcome be predicted? is a strictly weaker demand than causes) is exactly what separates the two: CRP passes prediction, fails causation. Full parameter table, independence caveat, and the treat-the-pathway-not-the-molecule consequence: Inflammation as a Modifiable Lever. ((CCGC) et al., 2011; inferred from Kaptoge et al., 2010)
An insulin-resistance readout on the same line — TyG predicts events, is not a target [2026-08-09, Liu]
The triglyceride-glucose (TyG) index — a cheap fasting lab proxy for insulin resistance, «considered a reliable surrogate marker of insulin resistance» (Liu et al., 2022) — is a fresh worked case of the prognostic-marker-vs-lever rule, and it sharpens two points the CRP case does not.
- Predicts incidence, not mortality — the split is the finding. Liu’s gold MA (12 cohorts, 6,354,990 general-population participants) finds highest-vs-lowest TyG raises CAD (HR 2.01, 95% CI 1.68–2.40), MI (1.36, 1.18–1.56) and composite CVD incidence (1.46, 1.23–1.74), but «there was no association between the TyG index and mortality» (CV mortality 1.10, 0.82–1.47; all-cause 1.08, 0.92–1.27 — both CIs cross 1) (Liu et al., 2022). A predictor’s signal is outcome-specific: TyG passes prediction on events and is silent on the mortality outcomes people weight most.
- A predictor that adds nothing over the incumbent score. Even the prediction value is bounded: «addition of the TyG index to the Framingham Risk Score (FRS) did not lead to improvement in its predictive power» (Liu et al., 2022). A marker can predict in isolation yet carry no incremental decision value once the standard factors are in hand — a distinct failure from CRP (whose incremental prediction survives even as its causation falls).
- Causation disclaimed, target-status denied. All 12 studies are observational — «causation cannot be proven» (Liu et al., 2022) — and Liu’s strongest claim is that TyG «may be considered an independent predictor for CVD incidence» (Liu et al., 2022), never a target. Like CRP, it sits downstream of the causal atherogenic state (apoB-bearing dyslipidemia), so it flags the stratum rather than naming the lever -> LDL ApoB and Cumulative Exposure. Full node: Insulin Resistance Surrogates and Cardiovascular Risk. (inferred from Liu et al., 2022)
Red flags
- A marker used because it is measurable, where the patient-important outcome was never named (the failure Rating Outcome Importance guards with the empty-row rule).
- A target expressed as a threshold on a marker, with no statement of which outcome it is a route to.
- Certainty ratings that look high for a body of evidence entirely composed of marker studies.
Green flags
- The evidence profile names the patient-important outcome, shows the surrogate used for it, and carries an explicit indirectness downgrade.
- The direction and magnitude of the marker -> outcome transmission is cited, not assumed.
Decision relevance
A surrogate may legitimately serve as a target to steer toward even where it is a weak stand-in for the outcome in the evidence — but only if the marker -> outcome transmission is itself an evidenced claim. Where it is assumed, the recommendation inherits the assumption silently, and markers have moved in the intended direction while the outcomes did not.
Two consumer-facing interventions that live entirely on this line: glucose tracking (Continuous Glucose Monitoring as a Health Intervention — moves HbA1c modestly, no hard outcome, almost no data in the healthy) and low-AGE / low-heat cooking (Dietary AGEs and Cooking Method — moves some lipid/inflammation markers, null on weight/glucose/BP, no hard outcome). Both are marketed on a marker whose transmission to a patient-important outcome is unevidenced.
The validated counter-exemplar — LDL / apoB. The rule cuts both ways: a surrogate whose causal transmission is evidenced is a legitimate target. LDL/apoB-particle burden is a well-evidenced such case — genetic, Mendelian-randomization and RCT evidence together establish it causes ASCVD, so lowering it (by diet or drug) reduces events in proportion to the reduction achieved -> LDL ApoB and Cumulative Exposure. It is the opposite of the marker moved, patient did worse cases: not all surrogates are equal, and this one has earned target status. (The proviso still bites — the transmission holds only where the LDL-C drop reflects a real particle-number drop and carries no off-target harm.)
And the transmission is quantified, not just asserted (CTT 2010). The strongest form of the surrogate moved AND the patient-important outcome followed: in the IPD meta-analysis of 26 statin RCTs, a 1.0 mmol/L LDL-C reduction cut major vascular events (RR 0.78) and all-cause mortality — «all-cause mortality was reduced by 10% per 1·0 mmol/L LDL reduction (RR 0·90, 95% CI 0·87-0·93)» (Cholesterol Treatment Trialists’ Collaboration, 2010). A surrogate reaching all-cause death in randomised evidence is the rare validated case -> LDL Lowering and Cardiovascular Events. But the validation is route-specific: it is statin (drug) lowering. The same marker lowered by diet did not transmit (Ramsden MCE/Sydney, below) — so even a well-validated surrogate must be re-checked for the agent actually used.
Applied — WHO downgrades strength because the evidence is a surrogate
WHO’s 2023 fat guideline supplies a worked instance of a surrogate costing strength, not just certainty. For replacing trans-fatty acids, the evidence runs through LDL cholesterol, and WHO reasons: while LDL “is a well-established biomarker for measuring the effects of interventions on CVD risk, and is considered by many to be a causal factor for atherosclerosis and coronary heart disease, it is not a physical manifestation or confirmation of disease. Therefore, a conservative approach was taken, leading to a conditional recommendation.” (World Health Organization, 2023)
Two things worth carrying:
- The marker was granted almost every credential a surrogate can have — well-established biomarker, widely held to be causal — and still produced a weaker recommendation. Surrogate status is not cancelled by a plausible mechanism.
- WHO explicitly declined the indirectness downgrade — and that is a live tension. It records LDL
as a critical outcome and states it “was not downgraded for indirectness when determining the
certainty in the evidence within the GRADE framework”. The surrogate discount was taken at the
strength step instead.
(World Health Organization, 2023)
- WHO’s warrant is the MARKER-to-OUTCOME link, not the exposure-to-marker one, and the direction matters because it is what criterion 2 above asks for. Annex 6 fn 14: “LDL cholesterol is an indirect measure of patient-important CVD outcomes. However, LDL cholesterol is a well-established biomarker for assessing the effects of interventions on CVD risk…, and is considered by many to be a causal factor for atherosclerosis and coronary heart disease. Therefore not downgraded for indirectness.” WHO is asserting that the chain transmits — the exact question Example 9 poses — not that SFA reliably moves LDL.
- So this is NOT the tension it looks like, and an earlier reading of it as one does not survive. GRADE’s rule is “in general, the use of a surrogate outcome requires rating down”, and it attaches an explicit judgment clause (“Consideration of the biology, mechanism, and natural history of the disease can be helpful in making a decision about indirectness”). WHO exercised that judgment, on the transmission question, with its reason on the record. A reasoned, documented, biology-based decision not to downgrade is the judgment GRADE invites, not a departure from a mandate — the not-joined check fires, and what is recorded here is a distinction, not a tension.
- What remains genuinely notable is narrower and still worth holding: the same marker was judged direct enough to escape the certainty downgrade and indirect enough to cost strength. WHO’s stated reason for the second is that LDL “is not a physical manifestation or confirmation of disease” — which is true of it at the certainty step as well. WHO invokes this consideration only at the strength step and does not explain why the same fact weakens strength yet not certainty.
- The evidence on the surrogate can be stronger than the evidence on the outcome and still yield a weaker recommendation: in the same guideline, replacing SFA with PUFA, MUFA or carbohydrate lowers LDL at high certainty, while the hard-outcome certainty behind those replacements is moderate to low (Saturated Fat Intake and Replacement). High certainty about the marker is not high certainty about the person.
The certainty profile of a surrogate vs its own hard outcomes (2026-07-26)
A scoped observation, worth carrying because it is visible only in the annex. In WHO’s evidence profile for replacing SFA with PUFA, every disease outcome is rated Low or Very low, and the one High-certainty row in the profile is LDL cholesterol (-0.055 mmol/L per 1% energy exchange; all-cause mortality Low, CVD mortality Very low, CVDs Low, CHD Low, stroke Low, type 2 diabetes Very low). (World Health Organization, 2023)
The disciplined reading, and the trap beside it. The useful habit is to ask which outcome carried a headline certainty label. The trap is to answer that question from the profile alone: WHO states elsewhere that its overall certainty “was based on disease and mortality outcomes”, and draws its moderate rating for the PUFA replacement from an RCT subgroup analysis (Hooper) that sits outside this profile. A first attempt at this page’s neighbourhood asserted the surrogate had silently carried the roll-up; a blind critique falsified it against that sentence, and the claim is withdrawn. What survives is the observation above and the habit — not an allegation. (World Health Organization, 2023)
The contrast case — where surrogate and hard outcome AGREE. The SFA profile above shows surrogate and outcome diverging in certainty (LDL High, every disease outcome Low). The opposite configuration is on Linoleic Acid and Cardiovascular Disease: the same nutrient (LA) lowers LDL/apoB and its objective biomarker tracks lower hard CVD, mortality and stroke — surrogate and patient-important outcome pointing the same way. The lesson is symmetric, not reversed: agreement is added warrant (here the apoB-mediated pathway, LDL ApoB and Cumulative Exposure) but still does not license reading the surrogate as the outcome — the LA outcome evidence stands on its own biomarker-cohort footing. The disciplined habit (ask which outcome carried the certainty; do not let the marker stand in) is the same whether the two agree or diverge.
Worked instance — the surrogate moved, the outcome unmeasured (FINGER, 2026-08-05)
A clear illustration of the trap in an RCT (not just in guidance): FINGER (Ngandu 2015) is a 2-year multidomain-lifestyle RCT whose primary endpoint is a cognitive-test composite (NTB Z-score), a surrogate for dementia. The composite moved significantly (between-group 0.022/yr, p=0.030), yet a 7-year extended follow-up is (only) planned to assess intervention effects on incidence of dementia and Alzheimer’s disease — i.e. the patient-important outcome was not measured here. (Ngandu et al., 2015) -> Multidomain Lifestyle Intervention and Cognitive Decline
Two disciplines this instance enforces. (i) A significant surrogate result is not a demonstrated outcome effect — the trial itself keeps them apart, reading its own result as public-health significance that is not easily translated into clinical or personal significance. (Ngandu et al., 2015) (ii) Whether a 2-year cognitive-composite change is a validated surrogate for dementia incidence is the open transmission question this page centres — legitimacy needs the causal link to the outcome to be evidenced, and here it is exactly what the follow-up is meant to test, not assumed. (inferred from Ngandu et al., 2015)
The hard-endpoint companion — preDIVA chose the outcome and got a null [2026-08-07]. The sharpest
version of the discipline is a trial that deliberately declined the surrogate. preDIVA (Moll van Charante
2016), a 6.7-yr multidomain vascular-care RCT (n=3526), measured clinical dementia incidence directly
and found nothing (HR 0.92, 95% CI 0.71-1.19). Its authors state the choice explicitly: «rather than
exploring effects on surrogate endpoints, we chose a clinical diagnosis of dementia as the outcome to draw
conclusions on prevention of dementia with unequivocal clinical relevance» — accepting a higher type-II-error
risk to buy an unambiguous endpoint. (van Charante et al., 2016) So within one cluster (multidomain dementia prevention) the field shows both halves of the
surrogate gap side by side: the surrogate moved (FINGER cognitive composite, 2 yr) and, on a thinner
intervention in a lower-risk group, the hard outcome showed no reduction (preDIVA dementia incidence,
6.7 yr). The two are not the same quantity (different endpoint, intervention, population, duration — a
distinction, not a contradiction -> Multidomain Lifestyle Intervention and Cognitive Decline). Note
what preDIVA does and does not do here: it is not FINGER’s own outcome trial (different intervention,
no cognitive training; different population), so it cannot test whether FINGER’s cognitive-composite
transmits to dementia — FINGER’s own extended follow-up is that test. What it illustrates is the general
discipline: a positive surrogate does not carry the outcome, and a separate, differently-designed outcome
trial in the same cluster read differently. (Ngandu et al., 2015; inferred from van Charante et al., 2016)
Limits
- GRADE’s two criteria are stated but not operationalized. Proximity is coarse and biologically judged; the Example 9 validation criterion asks the right question — can the outcome be predicted from the marker — but is delivered as a worked case rather than a standard, with no threshold for how much predictive uncertainty triggers one level versus two. A quantitative surrogate-validation framework (trial-level and individual-level association measures) would convert both into checkable standards; whether one exists that guideline bodies actually apply is unprobed here.
- Source currency: §3 is flagged in-source as rewritten in the 2024 GRADE Book.
A guidance-level version of the same critique [2026-07-28, Willett ch.16]
This page asks when a recommendation may rest on a marker. Willett’s policy chapter makes the same objection about how nutrition guidance has historically been built:
«Even until now, many dietary recommendations and policies have been based on short-term feeding studies with different levels of an essential nutrient and a single physiological variable as the outcome. While these can be important, such studies examine the effect of the dietary factor through a very narrow window because they do not consider the many pathways by which a dietary factor can influence health and the many steps in a pathway from dietary intake to occurrence of disease.» (Willett, 2012)
Note the concessive clause — «While these can be important» — is Willett’s, not a hedge added here. He is not rejecting feeding studies; he is bounding what a single physiological endpoint can carry.
The structure of the objection is precisely this page’s: a short-term feeding study measures one step in a causal chain, and a recommendation asserts something about the endpoint. The gap is the number of unmeasured steps and unconsidered pathways, which is why a marker can move correctly while the outcome does not.
Where this lands in the corpus. The wiki already holds WHO’s LDL non-downgrade as a worked case of a body reasoning across that gap explicitly. Willett’s claim is the general version, from a methods textbook rather than a guideline — so it is a second, differently-situated statement of the same concern, and unlike the WHO case it is not one body assessing its own process.
Bounded, because Willett is a participant in these debates and not a neutral observer of them. The
same chapter asserts that «for years dietary guidelines have emphasized low-fat, high-carbohydrate
diets when there was little evidence of benefit» — a substantive and contested position on a question
this wiki has not adjudicated. Recorded as Willett’s stated position, not adopted; it bears on
SC-12, and the handle is noted here without scoring it, per the ingest-scope rule. The surrogate
point above stands on its own reasoning and does not depend on that claim being right.
A well-evidenced surrogate-to-target argument (LDL/apoB) [2026-07-28, ESC]
The telos distinguishes a surrogate (stands in for an outcome in the evidence) from a target (something to steer toward), and licenses a target only if its causal transmission to a named patient-important outcome is itself an evidenced claim. ESC makes exactly that argument for LDL-C, and its structure is worth extracting whatever one concludes about the conclusion.
ESC’s preamble — its own confidence claim, recorded as ESC’s and not adopted: the causal role of LDL-C in ASCVD «is demonstrated beyond any doubt by genetic, observational, and interventional studies».
Then the load-bearing attribute:
«The relative reduction in CVD risk is proportional to the absolute size of the change in LDL-C, irrespective of the drug(s) used to achieve such change.» (European Society of Cardiology, 2021)
Why «irrespective of the drug» is the whole argument. A marker is a mere surrogate when the benefit might belong to the intervention rather than to the marker’s movement. If several drugs with different mechanisms produce outcome benefit proportional to the LDL-C change they achieve, the common factor is the marker, not any drug’s off-target effects. That is a dose-response test across causal routes — the marker earns target status by behaving the same way no matter how it is moved.
This is the same discriminator Willett’s case studies identified, arriving in a different domain: convergence across independent method classes (here genetic, observational, interventional) is what separated folic acid from vitamin A and dietary fat -> Upgrading Observational Evidence.
Three bounds, because this page’s job is scepticism about markers.
- The wiki has NOT verified the underlying claim. ESC asserts drug-independence with one reference; this corpus holds none of the trials, no Mendelian-randomisation source, and no contrary analysis. What is recorded is the argument’s structure and ESC’s assertion, not a verified fact about LDL.
- «Beyond any doubt» is a guideline’s rhetoric, not a certainty rating. ESC attaches no GRADE-style certainty to it, and this corpus has repeatedly found strength language and evidence grading to move independently -> Certainty of Evidence vs Strength of Recommendation.
- A validated target for one outcome is not a validated target generally. The claim is about ASCVD. It licenses nothing about LDL-C and all-cause mortality, cancer, or any other endpoint, and the telos’s warning stands: markers have moved the right way while patients did worse.
Contrast worth keeping, and it is what makes this page’s distinction operational. WHO’s LDL
non-downgrade (above) was a judgement that indirectness did not apply. ESC’s argument is a
different kind of thing — an empirical claim about invariance across interventions, which is
checkable. Two bodies, two routes to treating LDL as more than a surrogate: one by reasoned
non-downgrade, one by asserted mechanism-independence. The second is the stronger form, and the
corpus should hold the sources that test it. AWAITS a Mendelian-randomisation or trial-pooling source
on LDL-lowering across drug classes.
The certainty gradient runs opposite to outcome importance — measured twice [2026-07-28]
WHO’s SFA Annex 6 grades every outcome for one exposure. The ordering is the finding:
| Outcome class | Best certainty in the annex |
|---|---|
| LDL cholesterol (surrogate) | High — the only High row |
| Cardiovascular events (hard, composite) | Moderate |
| All-cause mortality (hard) | Moderate |
| CVD / CHD mortality, type 2 diabetes | Low |
| Stroke, CHD events | Very low |
(World Health Organization, 2023)
The best-known quantity is the one nobody cares about directly, and the outcomes a person would actually choose on are known least well. This is not a defect in WHO’s grading — it reflects that a lipid response can be randomised, measured in weeks, in small samples, while an event outcome cannot. The certainty gradient tracks measurability, not importance.
And it is now measured twice, on two exposures, in two guidelines. Sodium Intake and Blood Pressure has the identical shape: High certainty on blood pressure, very low on hard outcomes. Two independent exposures showing the same inversion is the beginning of a general claim about this literature — that wherever a nutrient acts through a measurable intermediate, the intermediate will be better graded than the endpoint, by construction wherever it holds, because the designs that grade well are the ones the intermediate admits. (World Health Organization, 2012, inferred from 2023)
Why this is the practical core of the surrogate problem. The pressure to act on the marker is not irrationality — it is a rational response to the marker being the better-evidenced quantity. A reader following “act on the strongest evidence” is led toward LDL and blood pressure by the grading itself. The discipline this page asks for therefore runs against the grain of the evidence hierarchy, not with it, which is why it needs stating rather than assuming.
The guard that keeps this honest: two exposures is two, and both are WHO’s. A third body grading the same inversion on a different exposure would make this a property of the literature; two WHO guidelines make it a hypothesis with two instances.
The mirror case — a surrogate WHO DECLINED to credit [2026-07-29, WHO NSS 2023]
The LDL and ESC cases above are surrogates argued up to target status. Non-Sugar Sweeteners is the opposite move by the same body: a moving surrogate WHO refused to bank. In the NSS trials, short-term body weight falls (MD -0.71 kg pooled), yet WHO ruled that «evidence of minor weight loss or reduced BMI over several months or less … does not represent a health benefit», because weight «must be sustained over the long term» to matter — and the long-term cohorts point the other way. (World Health Organization, 2023b)
The same second-criterion test, opposite verdict — read through this page’s lens. This page’s load-bearing test (Example 9) is whether the outcome can be predicted from the surrogate — whether the chain transmits. WHO did not cite Example 9 for either exposure, but its stated reasons map cleanly onto that test, and land oppositely for two of its own surrogates:
| Case | Surrogate | WHO’s stated reason (mapped to transmission) | WHO’s verdict | Same quantity? |
|---|---|---|---|---|
| SFA/TFA 2023 | LDL cholesterol | «considered by many to be a causal factor for atherosclerosis» — transmission asserted (genetic/MR/RCT elsewhere) | credited — declined indirectness downgrade | — |
| NSS 2023 | short-term body weight | short-term loss «does not represent a health benefit» absent evidence it is «sustained over the long term» — no evidence the marker predicts the outcome | declined — not banked as benefit | No — different surrogates, one rule |
Why this is a refinement, not a tension (not-joined check (ii): different surrogates). The two are
different quantities judged under the same rule — credit a surrogate only where its transmission to the
patient-important outcome is evidenced. LDL cleared it; short-term weight did not. So the NSS case is the
negative worked instance of this page’s decision rule: the discipline the page asks for, applied by a
guideline body against itself. NON-independent (both WHO) — this corroborates the rule, not via an
independent field, so no [E-independent] is claimed.
The sharper point: the surrogate here was not merely weak — it was moving in the good direction while the hard-outcome cohorts moved the other way (weight down in RCTs; obesity/T2D/CVD/mortality associations up in cohorts). That is the textbook marker moved the right way while patients did worse warning made concrete — and it is why a moving surrogate is bankable only when transmission is shown, not assumed. The cohort direction is itself unadjudicated (-> The U-Shaped Association Artifact).
Pesticide-residue level as a surrogate for harm [2026-07-29, organic-food cluster]
The organic-food case is this page’s rule applied to a contaminant marker rather than a physiological one. Organic produce carries measurably fewer synthetic-pesticide residues — Smith-Spangler - Organic Foods Safer or Healthier Systematic Review 2012 found a residue-contamination risk difference of 30%, and Baranski (Barański et al., 2014) a fourfold higher residue incidence in conventional crops. But residue presence is a surrogate, and the patient-important outcome is harm at real-world exposure. Smith-Spangler supplies the transmission gap directly: differences in risk for exceeding maximum allowed limits were small. (Smith-Spangler et al., 2012)
When both arms sit largely below the regulatory limit, a difference in detectable residues does not transmit to a difference in harm — the marker moved, but the chain to the outcome is unevidenced at those levels (the expectancy test: decades of residue-difference data, still no outcome trial). Organic’s residue advantage is real and is a legitimate reason to prefer it on a precautionary axis, but it is not a demonstrated health benefit — the same discipline this page asks of LDL and of short-term weight. -> Organic vs Conventional Food
Meal-timing that moves markers but nothing patient-important [2026-07-29, Sutton eTRF]
The eTRF trial (Time-Restricted Eating) is a clean worked case of surrogates moving with no weight change and no hard outcome. Under supervised, weight-matched controlled feeding, shifting eating earlier lowered morning systolic/diastolic BP «by 11 ± 4 mm Hg … and 10 ± 4 mm Hg», cut mean insulin «by 26 ± 9 mU/L», improved insulin resistance «by 36 ± 10 U/mg (p = 0.005)», and reduced an oxidative- stress marker — while fasting glucose (a pre-specified glycaemic outcome) was null. (Sutton et al., 2018)
Everything credited to eTRF is a surrogate, and the transmission to a patient-important outcome is unevidenced here (n=8, 5 weeks, prediabetic men, no events, no trajectory). Worse for the marker story, two surrogates moved in opposite directions — insulin/BP improved while fasting triglycerides rose 57 ± 13 mg/dL — and the trial cannot say which, if any, tracks the outcome a person cares about. So eTRF’s numbers are a legitimate target candidate only if the marker->outcome chain is later shown to transmit; on this evidence they are markers, not a benefit. Note the direction-of-effect asymmetry: a short-fast-before-testing artifact plausibly manufactured the triglyceride rise, which is exactly why a single marker in isolation is not diagnostic.
And it is not just eTRF’s n=8 — the whole intermittent-fasting literature lives on this line. The pooled IF network-MA (99 RCTs, 6582 adults, median 12-week follow-up) reports only intermediate cardiometabolic outcomes — weight, BMI, lipids, glucose markers, BP — with no hard events, no mortality, no trajectory (Semnani-Azad et al., 2025). So even the best-powered summary of the field cannot say whether any IF regimen changes a patient-important outcome; it can only rank the markers — this page’s rule instantiated at whole-field scale -> Time-Restricted Eating.
The surrogate discount as a NON-GRADE guideline body builds it in [2026-07-31, USPSTF]
USPSTF operationalizes the surrogate penalty structurally, in its own vocabulary — intermediate outcomes vs health outcomes. Health outcomes are «symptoms, functional levels, and conditions that patients can feel or experience»; intermediate outcomes (blood pressure, serum cholesterol, weight, dietary intake) «are not health outcomes in and of themselves». The rule:
«The USPSTF gives greater weight to evidence of an effect on health outcomes than evidence of an effect on risk factors or intermediate outcomes. The fact that a preventive service has a proven effect on an intermediate outcome does not necessarily establish that it can improve outcomes that are perceptible to patients.» (US Preventive Services Task Force, 2022)
Where USPSTF is HARSHER than GRADE’s one-or-two-level discount. In its analytic framework the intermediate->health-outcome step is a distinct key question (KQ6), and USPSTF near-forecloses banking benefit through it: «Due to the inherent limitations of the evidence that will be used to link intermediate and health outcomes, it is very unlikely that the evidence for this key question will be deemed convincing. This is due to the likely need to depend on observational evidence and the high potential for confounding.» (US Preventive Services Task Force, 2022)
This is the same transmission test this page’s Example-9 criterion asks, made near-structural. GRADE downgrades one-or-two levels and lets a reasoned biology judgement escape the penalty (WHO’s LDL case above). USPSTF instead makes the intermediate->health linkage presumptively unconvincing, so a recommendation resting on a marker is pushed toward its indirect-evidence pathway and the certainty of net benefit is capped accordingly. Two instruments, same concern, different severity of the default. (inferred from US Preventive Services Task Force, 2022) -> GRADE vs USPSTF - Two Appraisal Systems
And the mortality corollary is stated plainly: «few preventive interventions have a measurable effect on all-cause mortality» — so USPSTF routinely grades on a health outcome one step short of death, and names the three reasons all-cause and cause-specific mortality diverge (competing harms; rare condition; competing mortality at old age). A marker moving is not a life saved; a cause-specific death averted is not an all-cause death averted. -> Rating Outcome Importance (US Preventive Services Task Force, 2022)
The surrogate that INVERTED — bone density up, fractures up (testosterone) [2026-08-01, Snyder TRAVERSE]
A striking worked refutation of a surrogate — and a rare one, because it lands on GRADE’s own named example. This page’s proximity rule (above) lists «bone mineral density for fractures» as a surrogate close to the endpoint — rate down only one level (Schünemann et al., n.d.). Testosterone therapy is the case where that proximate, credentialed surrogate moved the intended way while the patient-important outcome moved the opposite way.
Prior trials established the surrogate: «Testosterone treatment in men with hypogonadism improves bone density and quality, but trials with a sufficiently large sample and a sufficiently long duration to determine the effect of testosterone on the incidence of fractures are needed.» (Snyder et al., 2024). The TRAVERSE fracture RCT (n=5204, powered 80% to detect a 30% fracture reduction) then measured the outcome:
«A total of 91 of 2601 participants (3.50%) in the testosterone group and 64 of 2603 participants (2.46%) in the placebo group had one or more clinical fractures, excluding fractures of the ster- num, fingers, toes, facial bones, and skull (haz- ard ratio, 1.43; 95% confidence interval [CI], 1.04 to 1.97)» (Snyder et al., 2024)
The authors state the surprise directly: «We did not expect these results, because most previous studies showed that testosterone im- proved many measures of bone structure and quality» (Snyder et al., 2024).
The parameter comparison — same surrogate, opposite verdict from the outcome:
| Parameter | GRADE Handbook | Snyder 2024 (TRAVERSE fracture RCT) | Same quantity? |
|---|---|---|---|
| The surrogate | bone mineral density | areal + volumetric BMD (raised by T in prior trials) | Yes — both BMD |
| The patient-important outcome | fractures | clinical fractures | Yes |
| GRADE’s stance | BMD is proximate -> rate down only one level | — | — |
| Assumed direction of transmission | BMD up -> fewer fractures | BMD not measured in TRAVERSE (up in prior trials), fractures up (HR 1.43) | the outcome contradicts the surrogate |
Why this instance is distinct on the page. The LDL case is a surrogate argued up to target status (transmission evidenced); the NSS and organic cases are surrogates declined (transmission unevidenced). This is a third kind: a surrogate whose transmission was assumed on GRADE’s own proximity grounds and then empirically inverted by an adequately powered RCT. Scope caveat: the inversion is inferred across studies, not observed within one — the BMD-up leg is transported from prior trials in partly-different hypogonadal populations, and TRAVERSE (CVD-enriched men, osteoporosis not an entry criterion) did not itself measure BMD. That the surrogate was never measured in the population where the outcome inverted is itself part of the lesson, not a hole in it. Proximity plus a plausible mechanism plus a consistently-moving marker were still not enough — exactly Example 9’s point (correlation is not prediction), made concrete on a patient-important outcome. It is the textbook marker moved the right way while patients did worse warning, realised. (Bound it: one RCT substudy, HR 1.43 with a CI that barely excludes 1 [1.04-1.97], mechanism unknown — a striking instance, not a settled law.) -> Testosterone Adiposity and Muscle
(inferred from Schünemann et al., n.d.; Snyder et al., 2024)
A guideline body rewrites its CASE DEFINITION to demote a surrogate [2026-08-04, EWGSOP2]
The prior instances are surrogates argued up, declined, or inverted inside an evidence appraisal. This is a fourth, structurally different form: a guideline body reorganising its own diagnostic definition around the surrogate discipline. For three decades sarcopenia was defined by low muscle mass. The EWGSOP2 (2019) consensus demotes mass to a confirmatory role and promotes muscle strength to the primary criterion, on exactly this page’s grounds: «muscle strength comes to the forefront, as it is recognised that strength is better than mass in predicting adverse outcomes» (Cruz-Jentoft et al., 2018).
Two of the page’s own tells drove the demotion, stated by an independent body in a different domain
(muscle, not lipids/BMD) — a generality instance for the rule (type-E flavour, corroboration not pad):
- Outcome-proximity: strength predicts falls/fractures/disability/mortality better than mass — so the closer-to-outcome measure wins primacy. Reinforced by a rate asymmetry EWGSOP2 reports: after 50, strength declines faster than mass (1.5-5% vs 1-2%/year), so a mass number under-reads the functional loss.
- Measurement reliability compounds it: mass/quality «remain problematic as primary parameters» because they are technically hard to measure accurately, while grip strength is cheap and reliable — so the worse predictor was also the noisier one. Both considerations pointed the same way.
Bound it: this is a definitional/prognostic reorganisation by expert consensus, not an RCT showing that steering strength changes outcomes — the loop stays open, and the value here is the appraisal move (pick the marker closest to the outcome), independently enacted -> Sarcopenia Definition and Diagnosis, Testosterone Adiposity and Muscle.
A refinement worth carrying — a demoted surrogate can still independently predict the hard outcome. The demotion above is easy to over-read as mass is a worthless marker. de Santana’s SR-MA finds low appendicular muscle mass still predicts all-cause mortality «cannot be completely explained by differences in muscle strength» (ASMI SMD −0.18, 95% CI −0.23 to −0.12) (de Santana et al., 2021). So the marker EWGSOP2 pushed to a confirmatory role is not merely a proxy for the closer-to-outcome measure (strength) — it carries independent prognostic signal. The surrogate lesson is sharper for it: a marker demoted for being a worse/noisier predictor is still a predictor, not a validated target — de Santana is observational, and no RCT shows raising muscle mass reduces mortality. Predictor, not target: the same line this page holds for every moved marker -> Low Muscle Mass and Mortality.
A within-study surrogate ladder — mass moved, strength did not (leucine in the elderly) [2026-08-19, Komar]
The muscle cases above are prognostic markers (mass, strength as predictors). Komar 2015 is the interventional version on the same tissue, and it lays the surrogate ladder bare within one gold SR+MA (16 RCTs, 999 elderly, chronic leucine-rich protein 2-7.8 g/d). The rationale chain is acute muscle-protein-synthesis -> lean mass -> strength/function; Komar tests how far up the chain a chronic supplement actually reaches:
- The mass surrogate moved — lean body mass +0.99 kg [95% CI 0.43, 1.55; p=0.0005] (confined to the sarcopenic subgroup, +1.14 kg; null in healthy elderly).
- The closer-to-patient outcome did not — «neither hand grip strength nor knee extension strength were affected by leucine supplementation in a fashion significantly different from control interventions» (grip WMD +0.23 [-0.26, 0.73], p=0.36, I2=65%; knee +0.07 Nm/kg [-0.26, 0.40], p=0.68) (Komar et al., 2015).
So the intervention transmitted to the mass surrogate but not to the strength endpoint one step closer to what a person values — a within-study instance of a marker moving while the outcome it stands in for does not. Bound it (symmetric standards): the strength arm is under-powered, not a clean null — Komar attributes it to «a small number of trials potentially insufficient to yield significant results», and broader-inclusion SRs did find grip gains, so strength sits at insufficient evidence rather than confident no effect (the grip I2=65% is a disperse null) (Komar et al., 2015). Even the mass gain cannot isolate leucine from co-ingested protein/energy. The upstream acute-MPS mechanism is Anabolic Resistance; the mass->mortality prognostic link is Low Muscle Mass and Mortality; the intake target is Protein Intake for Older Adults. (inferred from Komar et al., 2015)
The inverse use — a surrogate NULL that rebuts a HARM mechanism [2026-08-04, Johnson LA-inflammation]
Every case above uses a surrogate to argue for something (a benefit, a target) or asks whether a body should credit a moved marker. Johnson 2012 is the mirror: a surrogate deployed to argue harm — the seed-oil thesis that dietary linoleic acid (LA) drives disease via inflammation, with inflammatory markers (CRP, IL-6, TNF-alpha) as the surrogate for the pathway. A systematic review of RCTs finds feeding LA does not raise those markers: «virtually no evidence is available from randomized, controlled intervention studies among healthy, noninfant human beings to show that addition of LA to the diet increases the concentration of inflammatory markers» (Johnson & Fritsche, 2012).
The asymmetry of a surrogate null — what it does and does not license [inferred from @johnson2012]:
- It rebuts the mechanism, not the outcome. A null on the surrogate breaks the proposed harm pathway (LA -> inflammation -> disease) on its own terms. That is a real move — the harm story was built on the surrogate, so a null on the surrogate collapses the story’s own logic.
- It is NOT proof of safety on hard outcomes, nor of benefit. The surrogate-to-outcome chain is unproven in either direction, so a marker that fails to rise says nothing about disease or death — the same gap that stops a marker that does move from proving harm. Johnson himself claims only «a measure of reassurance regarding current dietary recommendations», not benefit -> Linoleic Acid and Cardiovascular Disease.
- The direction of the argument does not relax the discipline. A reader sympathetic to seed oils is tempted to read the null as vindication; symmetric standards forbid it — the surrogate is exactly as weak a proxy when it helps your prior as when it hurts it. The correct output is the harm mechanism is not supported, not seed oils are safe — those are different claims resting on different (here, observational) evidence.
Generality for the rule: a surrogate is a two-edged proxy — its evidentiary weakness is invariant to whether it is wielded to establish an effect or to rebut one. Filing a marker-null as a mechanism-rebuttal (not an outcome verdict) is the surrogate discipline applied in the negative.
A body grades DISEASE certainty off surrogates — and its own incidence grades sit higher [2026-08-06, EFSA sugars]
EFSA 2022 assigns causal-relationship certainty to chronic-disease outcomes for added/free sugars — «moderate for obesity and dyslipidaemia (> 50-75% probability), low for NAFLD/NASH and T2DM (> 15-50% probability) and very low for hypertension (0-15% probability)» — but names the endpoints the RCTs actually measured: «based on data from RCTs which investigated the effect of ‘high’ vs. ‘low’ sugar intake on surrogate disease endpoints, i.e. body weight, liver fat, fasting glucose, fasting triglycerides and SBP.» (European Food Safety Authority, 2022) So the certainty in the disease relationship is transmitted through a surrogate — EFSA is transparent about the step, exactly this page’s discipline.
The within-source contrast is the sharp part. In the same opinion, EFSA grades SSBs «high for
obesity, T2DM, HTN and CVD (> 75-100% probability)» — and those come from prospective cohorts on actual
disease incidence, not surrogates. (European Food Safety Authority, 2022) So inside one document, the incidence-graded exposure (SSBs)
outscores the surrogate-graded one (added/free sugars %E) on the same diseases. That is the certainty-
follows-measurability gradient (above) seen within a single body’s own assessment, not just across two
guidelines — and it is not fully clean, because SSBs are also the additive-energy vehicle, so design and
exposure both differ; recorded as a corroborating instance of the gradient, not an independent
controlled contrast. NON-independent within the source; a third distinct body (after WHO and
USPSTF) transparently grading disease certainty off surrogates — corroborates the rule, no
[E-independent] claimed. -> Free Sugars Intake
(inferred from European Food Safety Authority, 2022)
The archetypal DISCONNECT — the surrogate moved, the outcome did not (Ramsden MCE) [2026-08-04]
The prior cases argue a surrogate up or down, or catch one inverting. Ramsden’s recovered Minnesota Coronary Experiment is the cleanest disconnect the corpus holds, and it is unusually strong because it is a double-blind RCT, not an observational comparison. Replacing saturated fat with corn-oil linoleic acid lowered serum cholesterol -13.8% vs -1.0% control (P<0.001) — a large, dose- related move in the intended direction — yet «Kaplan Meier graphs showed no mortality benefit for the intervention group in the full randomized cohort or for any prespecified subgroup», and the 5-trial meta-analysis found «no evidence of benefit on mortality from coronary heart disease (1.13, 0.83 to 1.54) or all cause mortality (1.07, 0.90 to 1.27).» (Ramsden et al., 2016) Ramsden’s summary: LA replacement «effectively lowers serum cholesterol but does not support the hypothesis that this translates to a lower risk of death.» (Ramsden et al., 2016)
Two grades of disconnect, kept distinct (symmetric standards).
- The clean, randomized disconnect: the surrogate moved and the randomized outcome did not. The intervention-vs-control mortality contrast is a null — cholesterol fell, death did not. This is the textbook surrogate failure: the marker responded exactly as the theory required, and the endpoint it was standing in for did not follow. It requires no causal claim about harm — only that the surrogate did not transmit.
- The stronger «patients did worse» claim is real but weaker-graded. Ramsden also reports that greater cholesterol reduction tracked higher death (HR 1.22 per 30 mg/dL) and more autopsy MI (IRR 1.90) — but he concedes the cholesterol-death arm is «observational in nature», it is >=65-driven and frailty-shaped (low cholesterol marks the frail/dying -> The U-Shaped Association Artifact), and the autopsy signal rests on half the files. So the corpus banks the disconnect as demonstrated and holds the worse-outcome overlay as suggestive, not proven.
Why this is the load-bearing counter to the validated-surrogate exemplar. The LDL/apoB case above is a surrogate whose transmission is evidenced — but only under Ference’s proviso: provided the LDL-C drop is concordant with the particle-number drop and there are no competing off-target effects -> LDL ApoB and Cumulative Exposure. MCE is the worked instance of why that proviso is not boilerplate: the same surrogate (serum cholesterol) lowered by a different agent (diet LA rather than a statin) did not reduce events, and Ramsden’s proposed reason is exactly an off-target escape hatch — «a decrease in low density lipoprotein can represent widely different biochemical phenomena», so «some agents that decrease low density lipoprotein have been shown to reduce the risk of coronary heart disease, while others have no clear effect, and still others might actually increase risk.» (Ramsden et al., 2016) The decision-relevant generalisation: a surrogate validated for one route of movement is not validated for every route — validate the marker->outcome transmission for the agent you are actually using, not for the marker in the abstract. This is the same discipline ESC’s «irrespective of the drug» invariance argument tries to establish for LDL-lowering drugs; MCE shows it can fail for a dietary LDL change. (inferred from Ramsden et al., 2016)
Sydney Diet Heart (Ramsden 2013) is the same disconnect, one grade sharper — the outcome moved the WRONG way, randomized. MCE’s randomized outcome was a null; the companion SDHS (single-blind RCT, 458 post-MI men, n-6-selective safflower-oil LA) had the surrogate fall MORE in the LA arm (total cholesterol -13.3% v -5.5%, P<0.001) yet the randomized ITT mortality go up: all-cause HR 1.62 (1.00-2.64), CVD 1.70, CHD 1.74 — and those cholesterol «reductions were not associated with mortality outcomes». (Ramsden et al., 2013) So SDHS is not just a marker that failed to transmit (MCE) but a marker that moved the “right” way while the endpoint moved the opposite way — a sharper grade of surrogate invalidation than MCE’s null (an adverse divergence, not merely an absent benefit). The same discipline it teaches, and the same caveat: the disconnect (surrogate-vs-outcome divergence) is demonstrated, but reading it as LA-causes-death is over-reading a small single-blind high-dose secondary-prevention trial that pools to non-significance -> Linoleic Acid and Cardiovascular Disease (The Sydney secondary-prevention arm). The surrogate lesson survives the appraisal even where the harm claim does not: a dietary agent lowering cholesterol is not thereby shown to lower death. (inferred from Ramsden et al., 2013)
A disease DIAGNOSIS as the surrogate — diabetes prevention did not transmit to CV events (DPPOS) [2026-08-20]
The Ramsden cases are lab-marker surrogates (serum cholesterol). DPPOS moves the disconnect one rung up the ladder: the endpoint standing in for hard outcomes is a whole disease diagnosis — incident type-2 diabetes — which the Diabetes Prevention Program’s lifestyle and metformin arms durably prevented (58% / 31% relative reduction, sustained ~15 years -> Lifestyle vs Metformin for Diabetes Prevention). Over a 21-year median follow-up, DPPOS asked whether preventing the diagnosis reduced hard cardiovascular events. It did not: «Neither metformin nor lifestyle reduced major cardiovascular events in DPPOS over 21 years despite long-term prevention of diabetes» — metformin vs placebo HR 1.03 (95% CI, 0.78–1.37; P=0.81), lifestyle vs placebo HR 1.14 (95% CI, 0.87–1.50; P=0.34), both null and both trending the wrong way for lifestyle, with «No effect of either intervention … on the extended cardiovascular outcome» either. (Goldberg et al., 2022)
Why this grade of disconnect earns its place beside the lab-marker cases. A disease diagnosis is a better surrogate than a serum marker — it is nearer the patient-important endpoint (diabetes is itself morbidity, and DPP named CVD as «the leading cause of death among patients with type 2 diabetes»). That a surrogate this good still failed to transmit over two decades is the lesson: proximity to the outcome does not guarantee transmission — a diagnosis-level surrogate can disconnect exactly as a lab marker can. (inferred from Goldberg et al., 2022)
But this null is BOUNDED — weaker-graded than the clean MCE disconnect. MCE’s null was double-blind with no comparable dilution; DPPOS’s is confounded three ways the authors name: «Provision of group lifestyle intervention to all, extensive out-of-study use of statin and antihypertensive agents, and reduction in the use of study metformin together with out-of-study metformin use over time may have diluted the effects», in what was «a relatively low-risk cohort from the standpoint of the prevention of CVD» (statins 56–62%, antihypertensives 68–74%, mild hyperglycemia at 21y). (Goldberg et al., 2022) A longer 30-year Da Qing follow-up in a higher-risk cohort did show a lifestyle MACE benefit (HR 0.74 [0.59–0.92]). So DPPOS does not establish “diabetes prevention has no CV benefit”; it establishes that the diagnosis->events link is not automatic and is undetectable in a low-risk stratum over 21 years — a transmission that may be present-but-undetectable rather than absent, the baseline-risk reading worked on Baseline Risk and the Relative-Absolute Split. Banked as a disconnect instance carrying its own dilution caveat, not a clean randomized null. (inferred from Goldberg et al., 2022)
A TREATMENT’s symptom/severity target as the surrogate — CPAP fixes the apnea, not the heart [2026-08-21, Yu]
All the cases above are measured markers standing in for outcomes. This one moves the disconnect onto the treatment target itself: positive airway pressure (PAP/CPAP) is prescribed to abolish the apnea and relieve daytime sleepiness, and it does — pooled Epworth sleepiness improved -1.92 (95% CI -2.79 to -1.06) with QoL, anxiety and depression gains — yet across 10 RCTs (N=7266) it did not reduce major adverse cardiovascular events (RR 0.77 [0.53-1.13]), CV death (1.15 [0.88-1.50]), or all-cause death (1.13 [0.99-1.29]). Crucially the cardiovascular intermediate markers were null too: no effect on blood pressure, BMI, lipids, or glycemia -> Sleep Apnea Treatment and Cardiovascular Risk. (Yu et al., 2017)
Why it earns its place — the surrogate underwrote the outcome expectation. Prior reports that PAP gave «modest decreases in blood pressure» plus its «apparent beneficial effectsofPAPonintermediatebio- markers» were, in Yu’s account, the rationale for expecting a hard-outcome benefit, and PAP already sits in guidance for a CV indication (2014 AHA/ASA: consider PAP for ischemic stroke/TIA). The RCTs then found the BP surrogate itself null — so «The absence of any significant association of PAP with intermediate markers of vascular risk … may explain the null associations of PAP with hard vascular outcomes.» (Yu et al., 2017) The symptom benefit is real and patient-important on its own; it simply is not evidence for the CV outcome the surrogate was recruited to predict — the guidance-null form of this diagnostic -> Which Objective Moved This Recommendation. (inferred from Yu et al., 2017)
When the surrogate IS the disease definition — treating-to-target normalizes TSH, moves nothing (Feller) [2026-08-30]
The CPAP case above hit its symptomatic target but not the heart; the Feller thyroid case is one turn sharper, and it lands the surrogate discipline on a standing drug for a common diagnosis. Subclinical hypothyroidism is defined by the marker — mildly elevated thyrotropin (TSH) with normal free T4 — so levothyroxine is prescribed to normalize that number. Feller’s gold SR+MA (21 RCTs, 2192 nonpregnant adults) shows the drug does exactly that at the surrogate and nothing at the person: treated-arm TSH fell into range (0.5-3.7 mIU/L) while placebo stayed elevated (4.6-14.7 mIU/L), «indicating that treatment was associated with nor- malization of thyrotropin levels» (Feller et al., 2018) — yet «thyroid hormone therapy … was not associated with benefit regarding general quality of life (n = 796; SMD, −0.11; 95% CI, −0.25 to 0.03) or thyroid-related symptoms (n = 858; SMD, 0.01; 95% CI, −0.12 to 0.14) … the quality of evidence assessed with the GRADE tool was judged moderate to high.» (Feller et al., 2018)
Why this instance is distinct on the page. In most cases above the surrogate is measured alongside the outcome; here the surrogate is the diagnostic criterion itself, so treating-to-target is guaranteed to move it — the marker cannot fail to respond, which makes its non-transmission to any patient-important outcome (QoL, symptoms, mood, cognition, BP, BMI — all null, several at GRADE high) a structurally clean dissociation: the drug did what it targets and the person did not change. And unlike CPAP, even the symptoms themselves did not improve, so there is no residual patient-facing benefit to hold onto. This drives a de-escalation: the standing prescription for the modal SCH adult treats a definitional number, not the person -> Levothyroxine for Subclinical Hypothyroidism. Bounded to the studied stratum (mildly elevated TSH, age ≤~74, mild-to-moderate symptoms): the null may not transport to TSH >10 mIU/L or to a high-symptom-burden subgroup, which the trials could not test. (inferred from Feller et al., 2018)
The surrogate’s DIRECTION flips with the comparator — e-cig haemodynamics [2026-08-30, Skotsimara]
Most instances above ask does the moved marker transmit to the outcome? The e-cigarette CV case adds a prior question: the same surrogate moves in opposite directions depending on the comparator, so a bare «e-cig moves BP» carries no decision until the comparator is named. Skotsimara’s SR+MA (14 studies, N=441, moderate quality) finds e-cig use raises HR/SBP/DBP acutely vs baseline (HR +2.27 bpm, 95% CI 1.64 to 2.89; SBP +2.02; DBP +2.01) yet lowers SBP/DBP when a smoker switches from combustible tobacco (SBP -7.00 mmHg, -9.63 to -4.37) (Skotsimara et al., 2019). Neither figure is a hard outcome: the only CV-event datum is a single observational MI association the authors flag as «sensitive to non-random misclassification bias», and «there is no epidemiological data on the risk for stroke or heart failure incidence» (Skotsimara et al., 2019).
Why this instance is distinct. Elsewhere the failure is a marker that moves but does not transmit; here the marker’s very sign is comparator-dependent, so the surrogate cannot even be read as directionally harmful or beneficial without fixing the counterfactual (clean air vs continued smoking). The cell stays at insufficient hard-outcome evidence either way -> The Insufficient-Evidence Statement, and the decision lives in the stratum split, not the marker -> Electronic Cigarettes and Cardiovascular Risk. (inferred from Skotsimara et al., 2019)
A PREVALENCE statistic read as harm — biochemical B12 deficiency without symptoms (Pawlak) [2026-09-02]
A different failure mode: not a marker that fails to transmit over time, but a prevalence figure (X% of vegetarians are B12 deficient) silently read as X% harmed. The source that reports the high prevalence supplies the correction in its own voice — in the one study that checked, «none of the vegetarians included in their study had clinical symptoms despite the fact that about two-thirds of the sample had B12 depletion or deficiency, as indicated by both low holo-TCII and elevated MMA» (Pawlak et al., 2013). So a biochemical depletion/deficiency rate is a surrogate, not a patient-important-outcome rate, and the marker-to-outcome chain in this stratum is insufficiently evidenced, not demonstrated.
Both sides, because the source states both. Pawlak immediately guards the other direction — the biochemical marker must not be dismissed as harmless either: «one has to be careful with concluding that biochemical indicators of B12 defi- ciency in vegetarians are not associated with any adverse symptoms» (Pawlak et al., 2013) (mild symptoms under-detected; hematological signs maskable by concurrent iron/folate; enzyme function impaired at the biochemical stage). The instance is therefore a clean insufficient-evidence holding on transmission — the marker neither proven to harm nor proven benign -> Vitamin B12 Status in Vegetarian and Vegan Diets. (inferred from Pawlak et al., 2013)