Domain opener — the pro-LA biomarker pole of the seed-oils question. Linoleic acid (LA, 18:2n-6)
is the major omega-6 PUFA and the dominant fat in bottled seed oils (soybean, corn, sunflower,
safflower). The lay seed-oils-are-inflammatory / drive-heart-disease claim runs through omega-6:
LA is the precursor of arachidonic acid (AA), whose eicosanoids are cast as proinflammatory and
prothrombotic. This page holds four evidence arms that contest that harm claim: an observational biomarker
pool (Marklund — clear benefit, CVD), the RCT arm (Hooper 2018 Cochrane — little-or-no-effect to
uncertain, more equivocal), a mortality-outcome SR+MA (Li 2020 — dietary + biomarker LA lower
all-cause, CVD and cancer mortality; observational, and overlapping Marklund’s cohorts, so a
refinement not an independent check — see The mortality outcome menu), and the direct
inflammation-mechanism test (Johnson 2012 — RCT SR: feeding LA does not raise inflammatory markers,
refuting the harm pathway on its own endpoint — see The inflammation surrogate, tested directly).
None shows the predicted harm;
the observational and RCT arms disagree on the strength of benefit, which is a design-difference
refinement, not a filed tension (see The RCT arm). The recovered-RCT contrarian pole is now held on BOTH its trials
(Ramsden’s Minnesota Coronary Experiment reanalysis — randomized mortality null; and the Sydney
Diet-Heart companion — randomized mortality adverse in a secondary-prevention stratum; see The
recovered-RCT contrarian pole and The Sydney secondary-prevention arm). Neither delivers a
general-population harm finding: MCE is null, and Sydney’s harm — though genuinely randomized and
stronger than MCE’s — is one small single-blind high-dose n-6-selective trial that pools to
non-significance and is already absorbed into the benefit-netting Hooper/Li pools. What Sydney does
add is a secondary-prevention / high-dose n-6-selective stratum flag (possible harm, insufficient
evidence) and the n-6-selective-vs-mixed-PUFA distinction. A facet of the
dietary-fat cluster, linking up to Saturated Fat Intake and Replacement (LA is the PUFA in the
SFA->PUFA replacement question).
The effect estimate — objective biomarker, hard outcomes
FORCE-consortium individual-level pooled analysis of 30 prospective cohorts, 13 countries, 68 659 participants, 15 198 CV events, follow-up medians 2.5-31.9 y. Exposure is the in vivo circulating or adipose-tissue LA level (% of total fatty acids) — an objective biomarker, not self-reported intake. Effect is stated per interquintile range (the 10th-to-90th-percentile span of LA status), pooled by inverse-variance meta-analysis.
| Outcome | HR per interquintile range (95% CI) | Direction |
|---|---|---|
| Total CVD | 0.93 (0.88-0.99) | lower risk (7% per IQR) |
| CVD mortality | 0.78 (0.70-0.85) | lower risk (22% per IQR) |
| Ischemic stroke | 0.88 (0.79-0.98) | lower risk (12% per IQR) |
| Total CHD | 0.94 (0.88-1.00) | lower, non-significant (P=0.065) |
| AA — total CVD | 0.95 (0.90-1.01) | non-significant; NOT elevated |
- Categorical (quintile) confirmation: highest vs lowest LA quintile, CVD mortality HR 0.77 (0.69-0.86); the linear and categorical readings agree, and no nonlinearity was detected (P-nonlinearity >0.05 each) — a monotone gradient over the studied range, not a U or a plateau -> The U-Shaped Association Artifact. (Marklund et al., 2019)
- Absolute magnitude is not cleanly derivable here. The pooled analysis reports relative hazards per IQR against a distributional contrast, not a single control event rate per 1000, so an absolute per-1000 translation (as WHO Annex 6 gives for SFA) would require importing a baseline the source does not supply. The IQR framing itself resists a per-serving dose statement: it is a status-span, not an intake dose. (inferred from Marklund et al., 2019)
- No consistent effect modification. Associations did not differ by age, sex, race, n-3 level, diabetes, statin or aspirin use (Bonferroni-corrected). So this is a route-(a) exposure — absolute benefit scales with baseline risk, not a claimed effect-modifier — with one exception: the LA-stroke inverse association was stronger in FADS1 rs174547 T-allele carriers (P-interaction 0.002), which the authors flag as pleiotropic and «should therefore be viewed cautiously.» (Marklund et al., 2019)
The RCT arm — Hooper 2018 Cochrane, and it is more equivocal [2026-08-04]
The [AWAITS] is now cashed. Hooper’s Cochrane review is the RCT test of the same question:
19 RCTs, 6461 participants, 1-8 years, of increasing omega-6 fats (12 trials LA, 7 GLA, none
DGLA/AA; omega-6 usually displacing SFA or MUFA), only 3 at low risk of bias, GRADE-rated (WHO NUGAG
co-graded). Where Marklund’s biomarker pool finds a clear benefit, the RCTs find little or no effect
to uncertain:
| Outcome | Hooper RCT — RR (95% CI) | Certainty | Evidence state |
|---|---|---|---|
| All-cause mortality | 1.00 (0.88-1.12) | Low | little or no effect |
| CVD events (any) | 0.97 (0.81-1.15) | Low | little or no effect |
| CVD mortality | 1.09 (0.76-1.55) | Very low | uncertain (insufficient) |
| CHD events | 0.88 (0.66-1.17) | Very low | uncertain (insufficient) |
| Stroke | 1.36 (0.45-4.11) | Very low | uncertain (insufficient) |
| MI | 0.88 (0.76-1.02) | Low | may reduce — NNTB 53 |
| Serum total cholesterol | MD -0.33 mmol/L (-0.50 to -0.16) | High | reduces (dose-related) |
«We found no evidence that increasing omega-6 fats reduces cardiovascular outcomes other than MI, where 53 people may need to increase omega-6 fat intake to prevent 1 person from experiencing MI. Although benefits of omega-6 fats remain to be proven, increasing omega-6 fats may be of benefit in people at high risk of MI.» (Hooper et al., 2018)
Four evidence states, named precisely (the whole point of the distinction). Increasing omega-6 is little-or-no-effect on all-cause mortality and CVD events — at Low certainty, which is neither a confidently-demonstrated null nor “insufficient evidence.” It is insufficient (uncertain) on CVD mortality, CHD events and stroke — Very low certainty, wide intervals. It is a probable small benefit on MI (Low) and a definite benefit on serum total cholesterol (High). So the honest RCT-side reading is: not shown to reduce hard CVD or death, probably lowers MI a little, certainly lowers cholesterol.
The seed-oil-harm thesis also fails on the RCT side (symmetric standards). A reader hunting harm will note the point estimates >1 — CVD mortality 1.09, stroke 1.36 — but both are Very low certainty on tiny event counts (472 CVD deaths; 54 strokes, CI 0.45-4.11 spanning large benefit and large harm). Hooper states the inflammation route only as the untested counter-hypothesis («omega-6 fats may theoretically worsen cardiovascular risk»). So the RCTs neither confirm the observational benefit nor the mechanistic harm on hard outcomes — they are underpowered for both. (Hooper et al., 2018)
RCT vs observational — a REFINEMENT, not a filed tension
The instinct is to file Marklund-benefit vs Hooper-little-effect as a tension. The parameter
table forbids it — the two are not the same quantity:
| Parameter | Marklund (observational) | Hooper 2018 (RCT) | Same quantity? |
|---|---|---|---|
| Exposure | in vivo LA status (biomarker %, decades of habitual intake) | increasing omega-6 intake vs lower, within-trial change | NO — status level vs intervention change |
| Analyte | LA specifically (AA analysed separately) | omega-6 broadly — 12 LA + 7 GLA pooled | NO — LA-specific vs LA+GLA mix |
| Design / N | 30 prospective cohorts, IPD pool, 68 659 | 19 RCTs, 6461 (3 low-RoB) | NO — observational vs randomised |
| Horizon | follow-up medians 2.5-31.9 y | 1-8 y | NO — decades vs a few years |
| CVD mortality | HR 0.78 (0.70-0.85) benefit | RR 1.09 (0.76-1.55) Very low | opposed points; RCT CI barely spans 0.78 |
| CVD / total events | HR 0.93 (0.88-0.99) benefit | RR 0.97 (0.81-1.15) Low | both near-null; obs excludes 1, RCT includes it |
Same-quantity? = NO on exposure construct, analyte, design and horizon — so the not-joined check (ii) (different scope/unit, consistent once matched) fires, and the artifact is a refinement (F), not a joined tension. The resolver is the design difference, and it cuts both ways: (a) the RCTs are short (1-8 y vs decades), underpowered (few events, only 3 low-RoB, dietary-adherence-diluted), and mix in GLA — so they «remain to be proven» rather than «shown null»; (b) equally, Marklund’s strong CVD-mortality benefit (HR 0.78) is not reproduced by randomisation, so it must carry the standing observational discounts (residual confounding, reverse causation) that the RCTs are built to remove. Triangulation weakens the strong observational arm — it does not overturn it. What survives across both designs is narrower and firmer than either alone: LA is not harmful on hard CVD outcomes, and it lowers atherogenic lipids (Marklund’s LDL/apoB mechanism; Hooper’s High-certainty TC). The dramatic HR-0.78 CV-mortality figure is the fragile, single-design part. (Hooper et al., 2018; inferred from Marklund et al., 2019)
Two RCT-side refinements that ground Marklund’s own caveats
- The benefit concentrates where baseline omega-6 is low (< 5% E) — grounding the LA-depleted concern. Hooper: «across outcomes there is a suggestion of benefit from increasing omega-6 fats where baseline omega-6 fat intake is below 5% E, and this effect is lost where omega-6 fat intake is at least 5% E» — heavily caveated («only 7 of the 17 included trials reported baseline», underpowered). This is the essential-nutrient dose logic: adding an essential fat helps the deficient, not the replete — the same decision-relevant risk (too little LA) Marklund flagged for high-oleic LA-depleted oils, now with a rough threshold. (Hooper et al., 2018)
- The Ramsden contrarian pole is partly captured here. Hooper cites «Ramsden 2010 … found no significant effects on all-cause mortality (RR 1.16 …), non-fatal MI (RR 1.03 …), CHD death (RR 1.17 …) or CVD events RR 1.13 …), and these results have not altered substantially in their most recent update … (Ramsden 2016)» — Ramsden’s recovered-RCT nulls are consistent with Hooper’s pooled little-or-no-effect, not a suppressed harm. And Hooper is more conservative than the AHA advisory (Sacks 2017, RR 0.71): it excluded 2 of Sacks’ 4 core trials (Oslo = multifactorial + multivitamin; Finnish Mental Hosp = 2-cluster cross-over) — the same trial-selection dispute the Astrup/Hamley thread raises on Saturated Fat Intake and Replacement and Does Reducing Saturated Fat Reduce Cardiovascular Events. So when the staged Ramsden reanalysis lands, the joined issue is narrower than “benefit vs harm”: it is a within-RCT trial-selection dispute over a set of trials Hooper has already pooled to a null. (Hooper et al., 2018)
Self-critique (Hooper 2018 weave) [run 2026-08-04, before commit]. Not-joined: parameter table
built before prose, «same quantity?» = NO on exposure construct (status-level vs intervention-change),
analyte (LA-only vs LA+GLA), design and horizon — filed F/refinement, no tension. Laundered-E:
explicitly NOT claimed — the two arms diverge (obs benefit vs RCT little-effect), so no
convergent-independent-E; and Hooper-2018/Hooper-2020 share a team (method-shared, marked not-E).
Overclaim / symmetric standards: the RCT arm is stated at its true certainty (Low/Very-low), «little
or no effect» distinguished from «no effect» and «insufficient»; the seed-oil-harm thesis is refuted
symmetrically — the >1 point estimates (CVD mortality 1.09, stroke 1.36) are reported AND flagged
Very-low/imprecise, not buried. Triangulation: the claim is weakens, does not overturn — the strong
observational CV-mortality benefit is demoted to observational-only, not erased. Counter-passage:
Hooper’s own Discussion cross-comparisons (Hooper 2015, Ramsden, Sacks, Mensink, Farvid) read end-to-end.
The trial-selection contrarian pole — Hamley 2017 [2026-08-04]
The is now cashed. Hamley is the reanalysis that operationalises the recovered-RCT/trial-selection critique for the SFA->n-6-PUFA-replacement diet-heart trials. He splits the 11 diet-heart trials by whether the arms differed in anything other than the fat swap (trans fats in control-arm margarines, multifactorial advice, control-arm vitamin-E deficiency, cardiotoxic medication) and finds the one significant benefit (all-trials total CHD events RR 0.80, 0.65-0.98) vanishes in the confounder-free “adequately controlled” subset (RR 1.02, 0.84-1.23), with a significant subgroup difference (P=0.002). His verdict: replacing SFA with mostly n-6 PUFA is «unlikely to have either a beneficial or an adverse effect on CHD events, CHD mortality and total mortality.» (Hamley, 2017)
On this page’s question (does n-6 PUFA harm the heart?), Hamley is null-on-BOTH — it cuts against the harm thesis too. Even in the confounder-free subset the estimates straddle 1.0 (total CHD 1.02, major CHD 1.06, CHD mortality 1.13 — all CIs spanning benefit and null, none a significant harm). So the loud seed-oil-harm claim finds no support in Hamley’s cleaner subset either; his finding is absence of a clear effect, not harm.
Where it converges and where it clashes (same-quantity care). Hamley’s adequately-controlled hard-CHD
null (1.02) points the same way as this page’s Hooper 2018 omega-6 result (increasing omega-6 -> CVD
events RR 0.97, Low): both say isolating the n-6 PUFA intervention removes the hard-events signal.
Shared-evidence-base convergence (overlapping diet-heart trials, both RCT-MAs), NOT [E-independent].
The residual clash is the SFA-replacement subgroup — the SFA Cochrane review’s PUFA-replacing-SFA
subgroup RR 0.73 (0.58-0.92, reported inside Hooper 2018’s discussion, NOT Hooper 2018’s own
omega-6-increase estimate of 0.97) and AHA-Sacks 0.71, vs Hamley’s adequate-only 1.02 — and it turns on
which trials are adequately controlled, the same trial-selection dispute already flagged in Two
RCT-side refinements above. (The mainstream side does not hold 0.73 firmly either — the SFA-review
lineage’s replacement subgroups are underpowered; see Saturated Fat Intake and Replacement.) Filed on Does Reducing Saturated Fat Reduce Cardiovascular Events (Hamley grounds its
joined-issue 1), not re-litigated here.
(inferred from Hamley, 2017; Hooper et al., 2018)
Symmetric standards. Hamley is single-author moderate, its adequacy filter post-hoc
(author-conceded: «the categorisation is difficult to be based on criteria developed prior to the
litera-ture review»), and the adequate subset is only 5 trials — so the null is «unlikely to reduce»,
not «shown null», and the exclusion move can bias toward the null as the pro-benefit MAs’ inclusion move
biases away from it. But it declares no funding and no competing interests — a cleaner COI than this
page’s Unilever-supported biomarker pole — so it earns the same scrutiny, no more, no less. Its scope is
also narrow: extreme trial PUFA doses (>=14.3 %E) from vegetable-oil supplementation, and Hamley states
«these results should not be generalised to other foods high in mostly n-6 PUFA such as nuts and
seeds.» (Hamley, 2017)
The mortality outcome menu — Li 2020 widens it beyond CVD [2026-08-04]
Marklund and Hooper both answer a CVD question. Li 2020 (SR+MA, 38 studies / 44 prospective cohorts) extends the outcome menu to the broadest patient-important endpoint — all-cause mortality — plus cancer mortality, via both a dietary arm (811,069 participants, self-report) and a biomarker arm (65,411 participants). Higher LA is associated with lower mortality on all three outcomes:
| Outcome | Dietary — RR extreme categories (95% CI) | Biomarker — RR per SD (95% CI) |
|---|---|---|
| All-cause mortality | 0.87 (0.81, 0.94) | 0.91 (0.87, 0.95) |
| CVD mortality | 0.87 (0.82, 0.92) | 0.89 (0.85, 0.94) |
| Cancer mortality | 0.89 (0.85, 0.93) | 0.91 (0.84, 0.98) |
- This is an F-refinement, NOT independent-E corroboration of Marklund (the guard, stated). Li’s
biomarker arm literally contains Marklund 2019 — Li’s Table 2 attributes nine cohorts (CHS,
EPIC-Norfolk, MCCS, MESA, SHHEC, AGES-Reykjavik, HS-Japan, ARIC, CCCC) to
Marklund, 2019 (7). Both are observational cohort pools with overlapping cohorts and the same design class, so their agreement is shared-evidence-base, not two independent routes — it widens the outcome coverage, it does not add a second independent backing on the CVD claim. (inferred from Li et al., 2020) - A usable dose statement on CVD mortality. Where Marklund’s IQR framing resisted a per-%E dose, Li’s dietary dose-response is linear for CVD mortality (P<0.001): RR 0.93 (0.91, 0.95) per 5% energy increment of LA. For total and cancer mortality the curve is nonlinear but with «no clear threshold effect» — monotone-declining over the studied 1.1-11.6% E range, no knee located (total mortality 0.97 at 5% E → 0.88 at 10% E; cancer 0.96 → 0.83). This is another curve that is monotone or not-estimable, no plateau on the decision-relevant range -> The U-Shaped Association Artifact, and Li states the burden the corpus expects: a threshold is asserted-absent, not located. (Li et al., 2020)
- The RCT/observational split persists on all-cause mortality (same refinement, not a new tension). Hooper’s RCTs give all-cause mortality RR 1.00 (0.88-1.12), Low, little-or-no-effect; Li’s observational pools give 0.87 / 0.91. This is the identical design-divergence already adjudicated for CVD mortality in RCT vs observational — a refinement: not-joined check (ii) fires (short randomised intervention-change vs decades of habitual status), so it is an F-refinement, and the strong mortality benefit is observational-only. The RCTs are underpowered for mortality (few events, decades-old, chronically-ill participants); Li argues cohorts are «of high importance» here precisely because a powered mortality RCT is impractical. (Hooper et al., 2018; inferred from Li et al., 2020)
- A reverse-causation stratum flag. The protective total-mortality signal is a general-population finding: «the association between LA and total mortality among cancer patients was nonsignificantly positive», and excluding baseline-cancer cohorts strengthened the protective estimate (0.81 at 10% E). Among the already-sick the benefit does not hold — a sick-subgroup marker, consistent with the prospective-only design reducing but not removing reverse causation. (Li et al., 2020)
Self-critique (Li 2020 weave) [run 2026-08-04, before commit]. Laundered-E: explicitly blocked —
Li’s biomarker arm overlaps Marklund’s cohorts (nine cohorts shared, read off Table 2), so filed as
F/shared-evidence, [E-independent] NOT claimed. Not-joined: the all-cause RCT-vs-observational split
is mapped to the existing CVD-mortality parameter row (same quantity mismatch), no new tension.
Overclaim / symmetric standards: stated at true certainty — observational, residual confounding not
excluded, COI toward the finding recorded (California Walnut Commission); the benefit is «modest» (RRs
0.87-0.91), not oversold, and the mortality benefit is flagged observational-only. Counter-passage:
Li’s Discussion (RCT limits, dietary-vs-biomarker strengths/limits, guideline spread) read end-to-end.
The recovered-RCT contrarian pole, now held — Ramsden MCE 2016 [2026-08-04]
The Minnesota Coronary Experiment (MCE, 1968-73) is the recovered-data pole Marklund pre-empted: a double-blind, parallel RCT (n=9570; the largest diet-heart trial, the only one in large cohorts of women and >=65s, the only one with autopsy endpoints), in which corn-oil LA (13.2%E, ~2x the average US diet) replaced SFA. It is the cleanest available test of the diet-heart surrogate logic — and it comes apart at the surrogate->outcome step: (Ramsden et al., 2016)
- The surrogate moved, hard: serum cholesterol -13.8% (SD 13.0) intervention vs -1.0% control (P<0.001), in every prespecified subgroup, as the Keys equation predicted.
- The outcome did not: «Kaplan Meier graphs showed no mortality benefit for the intervention group in the full randomized cohort or for any prespecified subgroup.» The updated 5-trial meta-analysis (MCE + SDHS + RCOT + LA-Vet + MRC-Soy, n=10 808) 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)
- Two harm-suggestive signals, both weaker than the headline (see the symmetric-standards read below): a within-trial cholesterol-change -> death association (per 30 mg/dL drop, HR 1.22 (1.14-1.32), >=65-driven), and a provisional autopsy finding — intervention 41% (31/76) vs control 22% (16/73) with >=1 MI, IRR 1.90 (1.01-3.72), no less atherosclerosis (149/295 files). (Ramsden et al., 2016)
Attempt the contradiction — is this the seed-oil HARM the page’s harm thesis wanted? Parameter table first (same-quantity discipline). Marklund’s biomarker benefit vs Ramsden’s recovered RCT:
| Parameter | Marklund (biomarker cohort) | Ramsden MCE (recovered RCT) | Same quantity? |
|---|---|---|---|
| Exposure | in vivo LA status (biomarker %), habitual | corn-oil LA 13.2%E replacing SFA, high-dose intervention | NO — status vs high-dose intervention |
| Population | 68 659 general-population adults, 13 countries | institutionalized psychiatric/nursing-home; mean age 52, 25% >=65 | NO — general vs institutionalized elderly |
| Design / horizon | 30 prospective cohorts, medians 2.5-31.9 y | 1 RCT, mean 2.9 y in-hospital | NO — observational decades vs short RCT |
| Outcome tested | CVD mortality HR 0.78 (0.70-0.85) | all-cause mortality: no benefit (randomized); CHD-mortality MA 1.13 (0.83-1.54) | NO — CVD-mortality biomarker-contrast vs all-cause randomised-null |
Same-quantity? = NO on every row — so the loud «biomarker benefit vs recovered-RCT harm» clash the page anticipated does NOT hold as a joined tension (not-joined check (ii): different exposure construct, population, design and outcome, consistent once matched). The honest reading, even-handed:
- The randomized mortality contrast is NULL, not harm — and consistent with Hooper. The clean randomized MCE outcome (all-cause mortality, full cohort) shows no benefit, and Ramsden’s 5-trial MA (1.13 / 1.07, both spanning 1) is the same little-or-no-effect Hooper’s 19-RCT Cochrane pool reaches. Hooper already pools MCE and Ramsden’s nulls (see Two RCT-side refinements above: «these results have not altered substantially in their most recent update … (Ramsden 2016)»). So the recovered data do not establish a suppressed harm on hard randomized outcomes — they confirm the RCT-side null, the same design-divergence-from-observational already filed as an F-refinement.
- The harm-suggestive signals do not survive symmetric standards as harm. (a) The cholesterol-death HR 1.22 is, by Ramsden’s own concession, «observational in nature» — a within-trial cholesterol-change association, >=65-driven, and exactly the frailty/reverse-causation shape The U-Shaped Association Artifact warns of (low cholesterol marks the frail/dying); Ramsden’s frailty adjustment is crude (weight/BP change) and cannot fully remove it. (b) The autopsy MI IRR 1.90 is a randomized between-group contrast and the strongest harm hint, but it rests on half the autopsy files (149/295) and Ramsden calls it provisional. Neither is a demonstrated hard-outcome harm.
- A real intra-field friction, recorded (not adjudicated in an ingest). Marklund/Hooper dismiss the old corn-oil trials as trans-fat-confounded; Ramsden argues the opposite for MCE — the trans-fat-rich margarines were in the control/baseline diets, «(but not the intervention diet). Thus, confounding by dietary trans fat is an exceedingly unlikely explanation for the lack of benefit of the intervention diet.» So the standard pro-LA rebuttal to MCE (trans-fat confound) is itself contested by MCE’s own reanalysts -> Does Reducing Saturated Fat Reduce Cardiovascular Events (joined-issue 1, where this is filed). (Ramsden et al., 2016)
Symmetric standards — Ramsden gets the same bar, both directions. For it: NIH-funded, no
financial conflicts — a cleaner COI than this page’s Unilever-supported biomarker pole, so it is not
discounted for being contrarian; it is a real double-blind RCT with objective (all-meals-provided)
exposure. Against it: one trial, 1968-73, a very high LA dose in an institutionalized psychiatric/elderly
population (generalizes poorly, and Ramsden says explicitly the results «should not be generalized to nuts
or other unprocessed foods»), the full randomized dataset was never recovered (so no clean RCT-contrast
on mortality), heavy censoring, and the harm arms are observational/provisional. Its net contribution here
is not a harm finding: it firms the RCT-side null (LA replacement lowers cholesterol/LDL but has not
been shown to reduce hard outcomes) and supplies the archetypal surrogate->outcome disconnect
-> Surrogate Outcomes. No tension filed — the contradiction was attempted and did not hold as a
clean benefit-vs-harm clash.
(Hooper et al., 2018; Marklund et al., 2019; inferred from Ramsden et al., 2016)
Self-critique (Ramsden MCE weave) [run 2026-08-04, before commit]. Not-joined / attempt-
contradiction: the parameter table was built before prose, «same quantity?» = NO on all four rows, so the
anticipated benefit-vs-harm tension is NOT filed — the RCT arm is null-consistent-with-Hooper, an existing
F-refinement, not a new clash. Overclaim / symmetric standards: the two harm-suggestive signals are
stated at their true weight (cholesterol-death = observational/frailty, author-conceded; autopsy =
provisional, half files), not read as demonstrated harm; the randomized null is the load-bearing outcome.
Laundered-E: Ramsden shares the diet-heart RCT pool with Hooper (MCE is in Hooper), so explicitly
NOT independent-E of anything — marked shared-evidence. Counter-passage: Ramsden’s own trans-fat
rebuttal (against the pro-LA dismissal) and his observational-in-nature concession both read end-to-end
and represented, so neither side is straw-manned.
The Sydney secondary-prevention arm — the harm signal that does NOT fold the way Minnesota did [2026-08-04]
The Sydney Diet Heart Study (SDHS) is the second recovered-data Ramsden trial: a single-blind, parallel-group RCT (1966-73), 458 men aged 30-59 with a recent coronary event (86% acute MI) — i.e. secondary prevention in established CHD. The intervention (n=221) replaced SFA with safflower-oil LA + safflower PUFA margarine: a concentrated, n-6-SELECTIVE exposure (~74.6 g LA/100 g, no n-3), raising LA to ~15% E. Median follow-up 39 months. Unlike Minnesota, the harm signal here is the randomized ITT primary outcome itself, not an observational sub-analysis:
- Serum total cholesterol fell MORE in the LA arm (-13.3% v -5.5%, P<0.001) — the Keys-predicted surrogate move — yet the LA arm died more, three concordant borderline-significant endpoints: all-cause 17.6% v 11.8%, HR 1.62 (1.00-2.64) P=0.051; CVD 17.2% v 11.0%, 1.70 (1.03-2.80) P=0.037; CHD 16.3% v 10.1%, 1.74 (1.04-2.92) P=0.036. (Ramsden et al., 2013)
- Within-intervention dose-response points to LA specifically: «an increase of 5% of food energy from n-6 LA predicted 35% and 29% higher risk of cardiovascular death and all cause mortality» while «the reduction in SFA was not significantly related to any mortality outcome» — the LA rise, not the SFA fall, tracked death. And the cholesterol drop: «these reductions were not associated with mortality outcomes» — a surrogate->outcome disconnect even cleaner than Minnesota’s (adverse, not merely null) -> Surrogate Outcomes. (Ramsden et al., 2013)
Attempt the contradiction — is this the secondary-prevention HARM the harm thesis wanted? Same-quantity table first, against the pro-LA nucleus AND against Minnesota.
| Parameter | Sydney (Ramsden 2013) | Nucleus benefit (Marklund / Li) | Minnesota (Ramsden 2016) | Same quantity? |
|---|---|---|---|---|
| Population | secondary-prevention post-MI men 30-59 | general-population adults | institutionalized psychiatric/elderly | NO — post-MI vs general |
| Exposure | high-dose (15%E) n-6-selective LA, no n-3 | LA biomarker status / dietary LA | corn-oil LA 13.2%E replacing SFA | NO — selective high-dose intervention vs status |
| Design / blinding | single-blind RCT, n=458, 63 deaths | 30-cohort IPD / 38-study SR | double-blind RCT, n=9570 | Sydney weaker blinding, MCE larger |
| Randomized mortality | adverse (HR 1.62-1.74, borderline) | benefit (obs) | null (5-trial MA 1.07-1.13) | opposed — but obs vs RCT NOT same quantity |
Verdict: NO tension filed against the nucleus — but the signal is REAL and stronger than Minnesota’s, so it lands as a stratum flag, not a dismissal. The even-handed read:
- FOR a real signal (why it does NOT fold like Minnesota): the harm sits in the randomized ITT primary outcome, not an observational sub-analysis (Minnesota’s randomized mortality was null; its harm was an observational cholesterol-death HR and a half-file autopsy count). Three concordant endpoints, a within-arm dose-response, and the trans-fat confound runs the wrong way to explain it: the intervention removed common margarines/shortenings (major TFA sources), so «Restriction … in the intervention group would be expected to substantially reduce consumption of trans fatty acids compared with the control group», and MUFA-adjusted sensitivity «did not noticeably alter» the LA-death relation. (Ramsden et al., 2013) So the naive intervention-margarine-TFA-manufactures-harm hypothesis does not survive the paper’s own evidence (caveat: Ramsden is the interested party, MUFA is an imperfect TFA proxy, TFA was never measured).
- AGAINST it being a general LA-harm finding (symmetric standards): one small (n=458, 63 deaths), single-blind, old trial at an extreme LA dose Ramsden himself will not generalize — «adverse effects of increasing n-6 LA from 6% of food energy to 15% … are not necessarily generalizable to lower LA intakes … [or] populations without established coronary heart disease.» (Ramsden et al., 2013) The all-cause CI touches 1.00; the intervention also cut MUFA and cholesterol. Ramsden’s OWN updated MA pools it to non-significance: LA-selective trials CHD mortality 1.33 (0.99-1.79) P=0.06, CVD 1.27 (0.98-1.65) P=0.07 — trends, not effects. Only the secondary-prevention subset reaches significance («1.84 (1.11 to 3.04); P=0.02»), and it is dominated by Sydney itself. (Ramsden et al., 2013)
- Not-joined against the nucleus (ii + F-absorption). Sydney answers a different question
(secondary-prevention, high-dose, n-6-selective) than Marklund/Li (general-population status/intake),
so not-joined check (ii) fires — different population and intervention, a distinction not a clash.
And decisively: Hooper 2018 and Li 2020 already POOL Sydney (Li explicitly lists it as a divergent
adverse trial) and still net to null/benefit — so at the meta-analytic level Sydney is a component
of the benefit-netting pools, not a contradictor of them (F-absorption). No
tension. - The beyond-summary move — the n-6-selective vs mixed-PUFA distinction (type B/A). Ramsden’s MA identifies PUFA composition as «a major source of heterogeneity for all mortality outcomes»: n-6-selective trials (SDHS, Rose corn-oil, Minnesota Survey) trend adverse, while the four trials that raised n-3 alongside n-6 show reduced CV mortality (0.79 (0.63-0.99) P=0.04). So «benefits previously attributed to greater intake of total PUFA might be specifically attributable to n-3 PUFAs.» This refines the whole LA question: much of the “PUFA benefit” evidence base mixes n-3, and the isolated n-6 signal (in these old high-dose secondary-prevention trials) is at best null, at worst adverse. (Ramsden et al., 2013)
The decision-relevant residue — a route-(b/c) stratum flag, INSUFFICIENT evidence, possible harm. For a post-MI patient weighing a high-dose, n-6-selective SFA->LA replacement (safflower/sunflower as the near-sole fat, no n-3), the isolated evidence cell is small and its sign is adverse (Sydney’s randomized harm + the 1.84 secondary-prevention pooled subset + the OXLAM oxidative-stress mechanism, which was strongest in Sydney’s smokers/drinkers). This is not established harm and not the general-population LA question (where the nucleus holds not-harmful + lipid benefit). It is a genuine insufficient-evidence, possible-harm stratum the population-level benefit sources do not isolate — the honest four-state reading, weighted by the person’s stratum at layer 3. (Hooper et al., 2018; Li et al., 2020; inferred from Ramsden et al., 2013, 2016)
Type-E guard — Sydney is NOT independent corroboration of Minnesota. Both are Ramsden recovered-RCT
reanalyses (same author, same method program); Sydney’s updated MA literally contains the Minnesota
Coronary Survey; both feed the diet-heart RCT pool Hooper/Li already meta-analyse. So “two recovered
contrarian trials” is one program, not two independent backings — [E-independent] explicitly NOT
claimed. The two are complementary in design position (MCE double-blind large/null; Sydney
single-blind small/adverse), which is informative, but not independent evidence.
Self-critique (Ramsden Sydney weave) [run 2026-08-04, before commit]. Not-joined / attempt-
contradiction: parameter table built before prose, «same quantity?» = NO on population + exposure vs the
nucleus, and Hooper/Li already pool Sydney to null/benefit (F-absorption) — so no tension filed, matching
the Minnesota precedent. Symmetric standards: the harm signal is graded as genuinely stronger than
Minnesota’s (randomized ITT vs observational) AND bounded (small, single-blind, extreme dose, pools to
non-significance in Ramsden’s own MA) — not privileged for being contrarian nor dismissed for it; the
trans-fat confound is examined in the direction the evidence actually runs, with the interested-author
caveat kept. Overclaim: the residue is stated as insufficient evidence / possible harm in a narrow
stratum, explicitly not general-population harm and not established. Laundered-E: Sydney↔Minnesota
non-independence stated outright (shared author/method, MA-containment). Counter-passage: Ramsden’s own
generalizability limits, non-significant pooled MA, and trans-fat sensitivity all read end-to-end and
represented against his headline.
The biomarker design sidesteps the binding constraint — but only because LA is essential
The core methodological move: LA status is measured biochemically, not from an FFQ. Marklund’s justification is exactly the mirror image of why the same trick fails for saturated fat:
«Because LA cannot be produced endogenously (making tissue levels reasonable markers of intake), biomarker (circulating and adipose tissue) levels correlate with dietary consumption. Such objective biomarkers allow the evaluation of dietary exposure of LA status independent of self-reported food habits and estimated nutrient composition of different foods.» (Marklund et al., 2019)
This is the decisive intersection with Measurement Error in Dietary Assessment. Self-reported intake is the binding constraint on the whole diet-fat literature (34% under-reporting of energy; flattened gradients). A tissue biomarker removes the self-report/recall error — for a fatty acid the body cannot synthesise. Saturated Fat Intake and Replacement records the converse (Van Dam & Hunter): for SFA and MUFA the biomarker route is closed, because de novo lipogenesis makes tissue SFA a function of intake and carbohydrate/insulin status, so the marker cannot separate the terms. LA is essential — no endogenous source — so its tissue level is a genuine (not perfect) intake marker. Marklund still notes that circulating LA levels can be influenced by metabolism (the FADS desaturase genotype above is one such determinant) even while calling them established, useful markers of diet (Marklund et al., 2019) — so the biomarker removes the FFQ error, it does not deliver an error-free intake measure. The point is comparative: the same instrument that is a dead end for SFA is usable for LA. That within-fat-type boundary is real information, not a technicality -> Is the Food Category Doing Any Work. (inferred from Marklund et al., 2019)
- Where this design sits in Upgrading Observational Evidence. It is still observational (biomarker cohorts, not RCTs), so it starts at low certainty. What it buys is not a GRADE upgrade factor but the removal of the exposure-measurement error that downgrades ordinary FFQ-based nutrition cohorts — an IPD pooled analysis with an objective exposure and centrally-adjudicated outcomes. It is the strongest observational form the LA question can take, and it is the reason a biomarker cohort here outranks a bigger FFQ cohort. But «residual confounding attributable to unmeasured or imprecisely measured covariates» remains the authors’ own stated limit, and reverse causation is argued-against (prospective-only design, adipose long half-life) rather than excluded. (Marklund et al., 2019)
LA vs AA — the omega-6 -> inflammation via arachidonic acid claim, tested directly
The seed-oil-harm mechanism is: LA converts to AA -> proinflammatory/prothrombotic eicosanoids -> CVD. Marklund breaks the chain at two points:
- Conversion is limited. «stable isotope studies suggest very limited conversion of LA to AA in humans, and trials show limited effects of increasing dietary LA on plasma and adipose tissue AA levels» — so raising dietary LA does not straightforwardly raise AA. (Marklund et al., 2019)
- AA itself is not associated with higher CVD risk (HR 0.95, 0.90-1.01; in extreme-quintile and total-plasma analyses lower risk, 0.92 / 0.81). The eicosanoid story is one-sided: AA is «also the main precursor to key anti-inflammatory metabolites, such as epoxyeicosatrienoic acids and prostaglandin E2, and other mediators that actively resolve inflammation, such as lipoxin A4», and gives rise to prostacyclin (antiaggregatory, vasodilatory). «These complex biological effects preclude simplistic inference on the health effects of AA metabolites.» (Marklund et al., 2019)
Decision consequence: the mechanistic omega-6-is-proinflammatory argument is a mechanism claim whose predicted outcome (higher CVD) does not appear when measured — the net-effect-not-intended guard firing in the pro-seed-oil direction. The dedicated inflammation-marker RCT evidence that closes whether LA moves inflammatory markers at all is now held — see the next section (Johnson 2012).
The inflammation surrogate, tested directly — Johnson 2012 RCT SR [2026-08-04]
The [AWAITS] is now cashed. Marklund breaks the LA -> AA -> CVD chain at its outcome end (AA not
associated with CVD). Johnson 2012 (systematic review of RCTs, J Acad Nutr Diet) tests the surrogate
end directly: does feeding LA raise inflammatory markers in humans? 15 RCTs (8 parallel, 7 crossover),
18 LA-vs-control comparisons, healthy non-infant free-living subjects, LA doses spanning control arms
of ~5-10 g/day up to high arms of 27-64 g/day (Junker sunflower-oil arm 64.3 g/day; Adam liquid diet
43.5 g/2 200 kcal). Qualitative synthesis (marked heterogeneity precluded meta-analysis). The result is
a direct null on the harm mechanism’s own endpoint:
- No effect on any commonly-measured marker. Across the trials there were no significant LA-vs-control differences in CRP, fibrinogen, PAI-1, IL-6, TNF-alpha, or soluble adhesion molecules (ICAM-1, P-/L-selectin). (Johnson & Fritsche, 2012)
- The two significant signals were eicosanoid metabolites, and their own authors called them non-inflammatory (Blair: urinary PGE2 up / 2,3-dinor-TXB2 down, read as favorable vascular effects; Adam: tetranorprostanedioic acid, attributed to LA oxidation not eicosanoid production). (Johnson & Fritsche, 2012)
«We conclude that 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)
This mechanizes Marklund’s AA-null with systematic RCT backing (an F — mechanism-for-the-null). Marklund stated LA->AA conversion is «very limited» as a background assertion; Johnson supplies the pooled trial evidence underneath it, and breaks the harm chain at three successive links:
- Link 1 — LA does not reliably raise AA. «There is now consistent evidence that dietary LA does not unconditionally increase circulating AA concentrations»; a systematic review of 36 human clinical trials found «no effect on the phospholipid pool of plasma/serum AA concentrations of decreasing dietary LA by up to 90%» (P=0.39) «or of increasing LA intakes by up to 600%» (P=0.72) — the LA->AA step is saturated at low intake, so dietary swings do not move tissue AA. (Johnson & Fritsche, 2012)
- Link 2 — even preformed AA does not raise cytokines. Directly supplementing AA (Thies 700 mg/day for 12 wk; Kelley 1 500 mg/day for 49 d) left TNF-alpha, IL-1, IL-6 and adhesion molecules unaffected (Kelley raised only in vitro LTB4/PGE2 in stimulated cells, not secreted cytokines) — so «there is little evidence to suggest that consumption of preformed AA contributes directly to increased concentration of markers of chronic inflammation.» (Johnson & Fritsche, 2012)
- Link 3 — AA-derived eicosanoids are not one-sidedly proinflammatory (PGE2 also inhibits TNF-alpha and IL-1 and induces anti-inflammatory lipoxins; AA is precursor to vasodilatory epoxyeicosatrienoic acids). This restates Marklund’s own dual-eicosanoid point and rests on the same underlying reviews (Calder, Serhan) — a shared-literature echo, NOT an independent-E convergence. (Resolvins are derived from the n-3 fatty acids EPA/DHA, not AA — that competition is the other, secondary harm route, not this one.) (Johnson & Fritsche, 2012)
Surrogate discipline — a null on the surrogate rebuts the mechanism, it does not prove benefit. Inflammatory markers (CRP, IL-6, TNF-alpha) are surrogates, not patient-important outcomes -> Surrogate Outcomes. Johnson’s null does exactly one thing: it removes the proposed harm pathway (LA -> inflammation -> disease) — it is not evidence that LA lowers inflammation or improves any outcome. Johnson himself claims only «a measure of reassurance regarding current dietary recommendations», not benefit. (inferred from Johnson & Fritsche, 2012)
Type-E guard — this is NOT independent corroboration of the outcome benefit. Johnson tests a
different endpoint (the inflammation surrogate) via a different design (RCTs) answering a different
question (does the harm mechanism fire?) than Marklund/Li (observational cohorts, CVD/mortality
outcome). So Johnson does not add a second independent route to the benefit claim — it refutes the
harm claim. Classified F/A (mechanism-refutation of the harm arm), [E-independent] NOT claimed.
(inferred from Johnson & Fritsche, 2012; Marklund et al., 2019)
Symmetric standards — the null is bounded, and stated so. Johnson does not overclaim: small samples (largest 60 completers; three metabolic-ward studies 6-9 subjects), short durations (2 wk to 40 d), marker variability so that «the possibility of false negative outcome measures cannot be dismissed», and ~half the trials did not explicitly prohibit NSAIDs. The conclusion is a bounded absence — «the possibility that large intakes of LA increase markers of inflammation cannot be eliminated» — calling for «larger, longer-term, dose-response studies». This is a no-meaningful-effect (bounded) reading of the harm mechanism, distinct from insufficient-evidence: RCTs exist and are consistently null, but they are underpowered for a subtle effect. (Johnson & Fritsche, 2012)
- A genetic stratum where the null may not hold (route-b/route-c flag). Johnson notes genotype may account for «a quarter of the variation in circulating/tissue AA»: Martinelli found a genetic profile enhancing LA->AA conversion associated with higher CRP and elevated CVD risk, and Mathias found most African Americans carry a FADS variant tied to elevated circulating AA (10% vs 7% of total fatty acids) — speculated to partly explain higher CVD incidence. This is the mirror of Marklund’s FADS1 rs174547 stroke-interaction: the population-average inflammation null may not transport to fast-converter genotypes. A candidate effect-modification stratum, not an established one (mechanism + association, no in-stratum RCT). (Johnson & Fritsche, 2012)
- Bounded by date — but not superseded within the corpus. Johnson is 2012; the loud seed-oil
discourse is 2023-2025, and no later dedicated inflammation-marker RCT SR is held here, so the direct
test is a decade old.
[UPDATE 2026-08-23: Su 2017 (SR-MA) is now the later, quantitative direct test — see the Su subsection below; it pools Johnson's-era + newer RCTs to the same null with CIs.]Its direction is concordant with everything the corpus has ingested since (Marklund 2019, Li 2020, Hooper 2018 all post-date it and none find the predicted harm), so it is a bounded, not stale finding: a newer direct inflammation-RCT SR would be the source that could update it. (inferred from Johnson & Fritsche, 2012)
Self-critique (Johnson 2012 weave) [run 2026-08-04, before commit]. Laundered-E: explicitly
blocked — Johnson tests a different endpoint/design/question than the outcome sources, and its
dual-eicosanoid point shares Calder/Serhan with Marklund; filed F/A mechanism-refutation, [E-independent]
NOT claimed. Overclaim / surrogate discipline: the null is a surrogate null that rebuts the harm
mechanism, stated explicitly as NOT proof of benefit; Johnson’s own reassurance-not-benefit stance
preserved (his verbatim «reassurance» wording quoted, the contrast label is this page’s). Symmetric standards: the bounded-null limits (small n, short, false-negatives possible,
«cannot be eliminated») are reported, not buried, and the genetic fast-converter stratum where the null
may fail is flagged. Counter-passage: Johnson’s Limitations and Conclusions read end-to-end; the two
«significant» eicosanoid findings surfaced with their authors’ non-inflammatory interpretations.
Su 2017 — the quantitative RCT-MA that bounds Johnson’s null [2026-08-23]
Johnson’s null was qualitative (heterogeneity precluded pooling) and healthy-only, and the section above flagged that no later dedicated inflammation-RCT SR was held. Su 2017 (SR-MA, Food & Function, gold) is now that source: 30 RCTs (26 trials), 1377 adults, higher-LA vs low-LA control, pooled as standardized mean differences. It supplies the magnitudes + CIs Johnson could not, and every pooled marker is a null (all 95% CIs cross 0):
- CRP SMD 0.09 (-0.05 to 0.24) · TNF -0.01 (-0.19 to 0.17) · IL-6 0.11 (-0.07 to 0.29) ·
fibrinogen -0.01 (-0.21 to 0.19) · PAI-1 antigen -0.16 (-0.51 to 0.18) · sICAM-1
-0.05 (-0.23 to 0.13) · sVCAM-1 0.17 (-0.04 to 0.38) · sE-selectin -0.02 (-0.23 to 0.20) ·
sP-selectin -0.08 (-0.37 to 0.22) · adiponectin 0.17 (-0.17 to 0.50) · MCP-1
0.14 (-0.33 to 0.60). Heterogeneity near zero (I2 ~0% for most), no publication bias (Egger’s P all
0.18). (Su et al., 2017)
«Our meta-analysis suggested that increasing dietary LA intake does not have a significant effect on the blood concentrations of inflammatory markers. However, the extent of change in dietary LA intake might affect the effect of LA supplementation on CRP.» (Su et al., 2017)
Parameter table — the BLOCKING same-quantity check (why this is F, not a tension). Johnson and Su answer the same question about the same markers; they differ only on synthesis mode, precision, and population — which is exactly the claim-refinement (F) axis, not a joined-issue clash.
| Parameter | Johnson 2012 | Su 2017 | Same quantity? |
|---|---|---|---|
| Question | does feeding LA raise blood inflammatory markers in adults? | identical | YES |
| Design | qualitative SR, 15 RCTs / 18 comparisons | quantitative SR-MA, 30 RCTs (26 trials), 1377 subjects | same class (RCT SR); Su pools |
| Population | healthy non-infant free-living ONLY | adults incl. unhealthy (obesity, dyslipidemia, HTN, MetS, CVD-risk, PCOS, PAOD) | NO — Su broader (the refinement) |
| Endpoint metric | narrative no significant difference | pooled SMD (95% CI) per marker | Su quantifies the same endpoints |
| CRP result | no significant effect (narrative) | SMD 0.09 (-0.05 to 0.24), overall null | YES marker; Su bounds it |
| TNF / IL-6 / fibrinogen / adhesion mol. | no significant effect | all SMD CIs cross 0 | YES |
| Verdict | qualitative null | pooled null + CRP-at-high-dose caveat | composite F beats either alone |
The one honest non-null — CRP at large LA increments (symmetric standards). The overall CRP pool is null, but subgroup + meta-regression surface a dose signal Johnson could not see: above-median LA-intake-difference studies showed CRP rising significantly (subgroup SMD 0.20 [0.02, 0.39], P=0.034).
«However, in studies with a higher (than median) difference in LA intake, blood CRP was significantly increased when LA consumption was increased (P = 0.034). Moreover, meta-regression analysis revealed a positive relationship between the effect size of CRP and difference in LA intake (coefficient: 0.020, 95% CI: 0.003 to 0.038; P = 0.023) (Fig. 5).» (Su et al., 2017)
This is a within-surrogate dose caveat, not a hard-outcome harm: the increments driving it (up to 37.7 g/day LA difference between arms) far exceed typical intake swings, CRP is itself a surrogate -> Surrogate Outcomes, and every other subgroup (duration, LA source, age, BMI, health status) stayed null. Reported here rather than buried — the harm-refutation arm carries its own boundary. (Su et al., 2017)
Population extension — the null now holds in unhealthy adults too (an F-amplify). Johnson tested healthy subjects only; Su’s health-status subgroup shows the null holds in BOTH strata (CRP: healthy 0.10 [-0.17, 0.37] vs unhealthy 0.09 [-0.10, 0.27]), widening the harm-mechanism refutation to the obesity / dyslipidemia / MetS strata that carry higher baseline inflammation — the strata where the proinflammatory-LA claim is loudest. (Su et al., 2017)
Type-F, NOT independent-E (author-diff done first, then citation). Author lists are disjoint (Su,
Liu, Chang, Huang, Wang [Jiangnan University] vs Johnson & Fritsche) — so the cheapest independence test
passes. But independence fails on the decisive check: Su cites Johnson 2012 as its antecedent (ref
21), framing itself as the quantitative + population-extended successor
(«a systematic review (without meta-analysis)… only studies conducted in healthy subjects were
included… results were not quantitatively synthesized»), and the two review overlapping RCT sets
(Su spans 1998-2016, containing Johnson’s pre-2012 trials). A meta-analysis that re-pools and cites an
earlier SR is laundered-E, not genuine convergence — so [E-independent] is NOT claimed. Filed F
(claim-refinement): Su bounds Johnson’s qualitative null with pooled effect sizes, adds the CRP dose
caveat, and extends the population.
(Johnson & Fritsche, 2012; inferred from Su et al., 2017)
What this changes for the page (confidence unchanged, arm strengthened). Page confidence: stays
low — that grades the benefit direction (no hard-outcome RCT shows LA lowers CVD; the case rests
on observational biomarkers + null surrogates). Su does not touch that; it strengthens the harm-
refutation arm specifically, converting Johnson’s qualitative decade-old null into a quantitative,
bounded, population-extended one. The proposed LA -> inflammation -> disease pathway is now refuted at
its surrogate endpoint with pooled CIs, in healthy and unhealthy adults alike — with a single honest
caveat (CRP at extreme LA doses). Cashes the omega-6-inflammation belief signal: the lay
seed-oils-are-inflammatory claim does not survive direct RCT-MA testing of the inflammatory markers it
names. (inferred from Su et al., 2017)
Self-critique (Su 2017 weave) [run 2026-08-23, before commit]. Laundered-E: explicitly blocked
— author-diff done first, then the decisive ref-21 citation + overlapping RCT sets; filed F, [E-independent]
NOT claimed. Parameter table: built before the prose; same-quantity YES on question + markers (the F
warrant), NO on population + precision (the refinement axis) — no cell left empty. Overclaim / surrogate
discipline: null is a surrogate null rebutting the harm mechanism, stated as NOT proof of benefit;
page confidence held at low. Symmetric standards: the one non-null (CRP at high dose, P=0.034; meta-
regression coefficient) is surfaced, not buried, with its extreme-dose boundary. Counter-passage: Su’s
discussion + subgroup tables read end-to-end; the CRP dose caveat and the health-status subgroup both
surfaced from Su’s own analysis, not asserted.
Mechanism — a case where surrogate and hard outcome AGREE
The proposed cardioprotective mechanism runs through the lipid surrogates:
«In randomized controlled feeding trials, dietary PUFA (primarily LA) as a replacement for either carbohydrates or saturated fat lowers low-density lipoprotein cholesterol, triglycerides, and apolipoprotein B levels, and raises high-density lipoprotein cholesterol» (Marklund et al., 2019)
plus stated effects on HbA1c/insulin resistance, blood pressure, and visceral/liver fat.
Contrast with the SFA story — this is the instructive intersection with Surrogate Outcomes. On saturated fat, LDL (the surrogate) carries High certainty while every hard CVD/mortality outcome is Moderate-or-below — surrogate and outcome diverge in certainty, and a reader who takes the LDL result as the outcome result is over-reading. Here the alignment is the other way: LA lowers LDL/apoB and the same LA biomarker tracks lower hard CVD, mortality and stroke — the surrogate and the patient-important outcome point the same way in the same nutrient. That agreement is same-pathway coherence — surrogate and outcome sit on the one apoB-mediated chain (LDL ApoB and Cumulative Exposure), NOT an independent-route (type-E) convergence, which a single source cannot supply — and it still does not license reading the surrogate as the outcome; the outcome evidence stands on its own biomarker-cohort footing, observational, and it is what carries the claim. (inferred from Marklund et al., 2019)
- The RCT arm confirms the lipid surrogate at High certainty — but note which fraction. Hooper’s RCTs give total cholesterol MD -0.33 mmol/L (High certainty), dose-related. The LDL-specific estimate was null (MD -0.04, 2 trials, 244 participants) — but Hooper reads that as underpowering, not absence: the highly-controlled Mensink 2016 trials show omega-6 replacing SFA lowers LDL (-0.058 mmol/L per 1% E), and Hooper concludes «we were underpowered to see effects on other lipids.» So the surrogate agrees across observational and RCT arms; the hard-outcome benefit is where they part. (Hooper et al., 2018)
Decision relevance
- Four evidence states, named — and the arms differ by design. Observational (Marklund + Li): LA status/intake -> lower CVD, CV mortality, ischemic stroke (Marklund) and all-cause + cancer mortality (Li 2020, dietary + biomarker concordant): benefit (gold/high pools, but overlapping cohorts — one observational body of evidence, not two). RCT (Hooper 2018): increasing omega-6 -> little or no effect on all-cause mortality and CVD events (Low), insufficient/uncertain on CVD mortality, CHD and stroke (Very low), probable small benefit on MI (Low, NNTB 53), definite total-cholesterol reduction (High). AA -> higher CVD: no meaningful effect (not the mechanism’s predicted harm). Inflammation mechanism (Johnson 2012 RCT SR): feeding LA -> no meaningful effect on inflammatory markers (CRP/IL-6/TNF-alpha) — a surrogate null that rebuts the harm pathway, not proof of benefit. What holds across all arms: not harmful on hard outcomes or the inflammation surrogate + lowers atherogenic lipids; the strong CV-mortality benefit is observational-only. Whether bottled seed oils as consumed (heated, in ultraprocessed foods) move an outcome vs the realistic alternative: insufficient evidence either arm — this is a nutrient-level (LA-status / omega-6-dose) finding, not a food-level or cooking-method one.
- Frame as substitution, not addition. The mechanism evidence is for LA replacing saturated fat or refined carbohydrate. The cardioprotective signal is a replacement effect; it says nothing about adding seed-oil calories on top of an existing diet. Judge against the realistic alternative fat.
- A stratum the finding actually flags: LA-depleted seed oils. Marklund notes «current trends in oil production are leading to increased use of high-oleic, LA-depleted seed oils, which can increase the risk of insufficient PUFA consumption in population subgroups» — i.e. the decision-relevant risk the data support is too little LA, not too much. (Marklund et al., 2019)
- Big-rock ranking: this is a fat-type / substitution lever, not a big rock. It ranks alongside the SFA-replacement question, well below smoking/adiposity/inactivity. For someone already lean, active and eating LA-containing foods, the marginal seed-oil decision is a small, contested lever — attention is an anti-signal applies (the seed-oil controversy is loud and the effect is modest).
Limits
- Five sources now held (observational biomarker + RCT MA + mortality SR+MA + inflammation RCT SR +
recovered RCT), confidence stays
low. Marklund (gold IPD cohort pool) + Hooper 2018 (gold Cochrane RCT MA) + Li 2020 (high SR+MA, dietary + biomarker mortality) + Johnson 2012 (high RCT SR, inflammation markers) + Ramsden 2016 (high recovered RCT + MA). Ramsden firms the RCT-side null on hard outcomes (LA replacement lowers cholesterol but has not been shown to reduce CHD/all-cause death) and supplies the surrogate->outcome disconnect, but it does not establish harm and does not move the confidence (which is about the CVD/mortality benefit). Johnson firms the harm-refutation to RCT level on the mechanism’s own endpoint (feeding LA does not raise inflammatory markers) — but that is a surrogate null rebutting the harm pathway, not support for the benefit claim, so it does not move the page’s confidence, which is about the CVD/mortality benefit. Li widens the outcome menu but is observational and overlaps Marklund’s cohorts (not an independent check), and the RCT arm still tempers the mortality benefit (Hooper all-cause RR 1.00 vs Li 0.87-0.91). So the composite does not earnmedium: what holds across all four is LA is not harmful on hard outcomes or the inflammation surrogate + lowers atherogenic lipids; the strong mortality/CV-mortality benefit remains observational-only. No Mendelian-randomisation triangulation is held here yet — that is what would move confidence, not another observational pool. - Industry funding — recorded, not netted (symmetric standards). «Unilever provided Tufts University with a restricted grant… to partly support this analysis», stated to have «no role in study design, study conduct, data analysis, manuscript preparation, or decision to submit»; two authors report Unilever research support, and the senior author reports extensive food-industry personal fees. Unilever is a major seller of LA-rich vegetable-oil spreads, so the COI runs toward the pro-LA finding. This does not refute the result (a large IPD pool with prespecified protocol resists sponsor steering), but it is exactly the halo-across-axes signal to keep visible, and a reason the contrarian pole must get the same appraisal bar when it lands. (Marklund et al., 2019)
- The contrarian pole is now HELD (Ramsden MCE 2016) — and no tension was filed, because the
contradiction did not hold. Marklund pre-empted the recovered corn-oil trials as «hampered by their
short duration, small numbers of events, substantial dropout, and confounding by industrial trans
fats.» (Marklund et al., 2019) With
Ramsden 2016 now ingested (The recovered-RCT contrarian pole above), the anticipated
biomarker-benefit-vs-recovered-RCT-harm tension fails the same-quantity check: MCE’s randomized
mortality contrast is a null (5-trial MA: CHD-mortality 1.13, all-cause 1.07 — the same
little-or-no-effect as Hooper’s pool, which already contains MCE), not a harm. The harm-suggestive
signals are observational (cholesterol-death HR 1.22, frailty-shaped) or provisional (autopsy IRR 1.90,
half files). Note Ramsden also rebuts Marklund’s own trans-fat-confound dismissal for MCE specifically
(trans fat was in the control, not the intervention). So the recovered-RCT pole firms the RCT-side
null, not harm.
- Ramsden Sydney Diet Heart — CASHED 2026-08-04 — woven above (The Sydney secondary-prevention arm). The check resolved: its harm signal is genuinely stronger than MCE’s (randomized ITT primary outcome, not observational), so it does NOT fold the way Minnesota did — but it is one small single-blind high-dose n-6-selective trial that pools to non-significance (Ramsden’s own MA) and is already absorbed into Hooper/Li’s benefit-netting pools. Lands as a secondary-prevention / n-6-selective stratum flag (insufficient evidence, possible harm) + the n-6-selective-vs-mixed-PUFA distinction. NOT independent-E of MCE (same Ramsden program). No tension.
- AWAITS — the cluster this opener anchors:
-
is now woven above (*The RCT arm — Hooper 2018 Cochrane*); it tempered rather than upgraded the observational benefit (F-refinement, design-divergence). -
biomarker LA vs *all-cause*, CVD and *cancer* mortality; woven above (*The mortality outcome menu*). F-refinement (widened outcome menu + dietary/biomarker concordance), not independent-E (overlaps Marklund's cohorts). -
LA does not raise inflammatory markers (CRP/IL-6/TNF-alpha) in healthy humans; woven above (*The inflammation surrogate, tested directly*). F/A mechanism-refutation of the harm arm, surrogate null (rebuts the mechanism, not proof of benefit), NOT independent-E of the outcome benefit. -
reanalysis; woven above (*The trial-selection contrarian pole*). Null-on-both (no benefit, no harm) in its adequately-controlled subset; converges with Hooper 2018's omega-6 hard-events null, clashes with the SFA-replacement subgroup (0.73/0.71) on trial-adequacy. Grounds joined-issue 1 on [[Does Reducing Saturated Fat Reduce Cardiovascular Events]]. - The recovered-RCT-harm arm (Ramsden Minnesota and Sydney reanalyses) is now fully ingested (both woven above). Net: MCE randomized-null; Sydney randomized-adverse but small/single-blind/ high-dose/n-6-selective and pooled to non-significance. No general-population harm signal survived (Hamley found none either); the surviving contrarian residue is a secondary-prevention / n-6-selective stratum flag, not a harm finding. The open lever now is Mendelian-randomisation (the design that would move the observational benefit to causal), not another recovered trial.
-
Food-level, on mortality — Zhang 2025, and the LA-rich oils do NOT move together [2026-09-02]
The arms above answer the LA question at the nutrient / biomarker level. Zhang 2025 (NHS/NHSII/HPFS, 221,054 adults, up to 33 y, 50,932 deaths, FFQ every 4 y) adds the bottled-oil food level on mortality — and the LA-rich oils split rather than tracking together:
- Soybean oil (the LA-rich workhorse of the US seed-oil supply) is inverse: total mortality per 5 g/d HR 0.94 (0.91-0.96), cancer mortality per 5 g/d 0.94 (0.89-0.99); the modelled butter->soybean swap gives total mortality 0.85 (0.80-0.91). Canola 0.85 (0.78-0.92) and olive 0.92 (0.91-0.94) are likewise inverse (per 5 g/d, total mortality). (Zhang et al., 2025)
- But corn and safflower — also LA-rich — are NULL. Corn per 5 g/d 1.07 (0.99-1.15) (borderline-positive, P-trend .09); safflower 0.90 (0.72-1.11), wide. Zhang attributes corn’s non-benefit to historical partial hydrogenation (trans fats in earlier follow-up), high-temperature commercial frying (oxidation), and minimal n-3; safflower to low consumption and wide CIs. So the food-level LA-oil signal is not uniform — a refinement the nutrient-level LA pools (which pool all LA together) structurally cannot see: which bottled oil, and how processed/cooked, matters. (Zhang et al., 2025)
F-refinement, NOT independent-E — same cohorts, instrument, and school. Zhang shares the
NHS/NHSII/HPFS cohorts, the Willett FFQ, and authors (Hu, Willett, Guasch-Ferré, Yanping Li) with
Li 2020 and much of the Harvard linoleic base already held here — the same shared-cohort
non-independence this page flags for Li (Li’s biomarker arm overlaps Marklund’s cohorts). Its
concordance with the LA-lowers-mortality direction is volume, not a second route, so it does not
move confidence (stays low; the open lever is still Mendelian-randomisation, not another observational
pool). What it adds is the food-level, culinary-resolved cut the nutrient pools omit — consistent with
this page’s standing caveat that whether bottled seed oils as consumed (heated, in UPF) move an outcome
is a food-level question the nutrient evidence does not settle. Full butter->oil substitution model +
symmetric-standards appraisal on Saturated Fat Intake and Replacement (Butter vs plant oils at the
food level).
(Li et al., 2020; inferred from Zhang et al., 2025)