The causal model behind the lipid axis. Most of the wiki’s cardiovascular reasoning runs on LDL-C (cholesterol mass). The EAS Consensus (Ference et al. 2017) supplies the causal framework underneath it — and it reframes three things: what causes the disease (apoB particles, not cholesterol mass per se), how the dose works (cumulative exposure, not current level), and which number to trust (apoB over LDL-C in the metabolically-impaired).
LDL/apoB CAUSES ASCVD — a causality verdict, not an association
The consensus assessed the LDL-ASCVD link against Bradford Hill-style causality criteria across four independent method families — genetic studies, prospective cohorts, Mendelian randomization, and LDL-lowering RCTs — and reached an unusually strong verdict:
«Consistent evidence from numerous and multiple different types of clinical and genetic studies unequivocally establishes that LDL causes ASCVD.» (Ference et al., 2017)
The causal agent is not LDL-cholesterol as such but the apoB-containing particles that carry it:
«cholesterol-rich LDL and other apolipoprotein B (apoB)-containing lipoproteins, including very low-density lipoproteins (VLDL) and their remnants, intermediate density lipoproteins (IDL), and lipoprotein(a) [Lp(a)], are directly implicated in the development of ASCVD.» (Ference et al., 2017)
This is the strongest form of the surrogate-vs-outcome question resolved in the validated direction: LDL/apoB is the exemplar of a surrogate whose causal transmission to the hard outcome is itself evidenced — the opposite of the marker moved, patient did worse cases -> Surrogate Outcomes. The verdict rests on the concordance between the naturally-randomized genetic/MR evidence (unconfounded, lifelong exposure) and the LDL-lowering RCTs (intervene on LDL, the outcome moves — the source’s most compelling causal evidence): the natural experiment and the intervention agreeing, the top of the mechanism-strength gradient, not a mechanistic story.
The dose is CUMULATIVE — magnitude x duration, not current level
The dose-response is log-linear across >2 million participants and >150,000 events, and it compounds with time:
«any mechanism of lowering plasma LDL particle concentration should reduce the risk of ASCVD events proportional to the absolute reduction in LDL-C and the cumulative duration of exposure to lower LDL-C, provided that the achieved reduction in LDL-C is concordant with the reduction in LDL particle number and that there are no competing deleterious off-target effects.» (Ference et al., 2017)
Two decision consequences follow:
- Lower for longer beats lower later. Because risk tracks the area under the LDL/apoB curve over a lifetime, a modest reduction sustained for decades can outweigh a larger one started late — the cumulative-exposure frame, which a single current LDL-C snapshot cannot capture.
- Mechanism-agnostic, if concordant. Any route that lowers apoB particle concentration (diet, statins, PCSK9 inhibitors) reduces risk in proportion to the reduction achieved — provided the LDL-C drop reflects a real particle-number drop and carries no off-target harm. This is what licenses reading a diet’s LDL effect (Saturated Fat Intake and Replacement) and a drug’s LDL effect (Statins for Primary Prevention and the Power of Zero CAC) on the same causal scale. The Portfolio Dietary Pattern and LDL Cholesterol is a worked diet-route instance (LDL-C −17%, apoB −15%) whose event benefit is read off this scale — a borrowing, since no Portfolio trial has measured events. (Same-school source, not independent corroboration of the causal scale; confidence here unchanged.)
Measure apoB, not just LDL-C, in the metabolically-impaired — they DISCORD
The concordance proviso above is not academic — it breaks exactly where the wiki’s drifting-median stratum sits:
«in certain conditions (e.g. the metabolic syndrome, diabetes, and hypertriglyceridaemia), plasma LDL-C and LDL particle concentration can become discordant as a result of the predominance of small, dense cholesterol-poor LDL, and therefore plasma LDL-C may not accurately reflect LDL particle concentration or its effect on cardiovascular risk. Under these conditions, direct measurement of LDL particle number or apoB concentration (recognizing that each LDL particle contains a single apoB molecule) may more accurately reflect the causal effect of LDL on ASCVD.» (Ference et al., 2017)
So for an insulin-resistant, hypertriglyceridemic adult, LDL-C can under-state the atherogenic particle burden — the small-dense-LDL pattern packs more particles (more apoB) into a given cholesterol mass. This is the resolution of the carnivore critique’s open question (Axis 1 reasoned on LDL-C; the causal quantity is apoB, and in this stratum they diverge — apoB is the number to trust).
LDL-P vs apoB vs LDL-C (Challenge #18). Three ways to count the same causal thing, and the ranking is apoB >= LDL-P > LDL-C for reflecting the atherogenic-particle burden:
- LDL-C is a good particle surrogate MOST of the time — «Under most conditions, LDL-C concentration and LDL particle number are highly correlated, and therefore plasma LDL-C is a good surrogate»; it only fails in the discordant (metabolic-syndrome / diabetic / hypertriglyceridemic) state. So LDL-P/apoB add little for the concordant, lean person and a lot for this stratum.
- LDL-P (particle number, by NMR) approximates apoB far better than LDL-C does — both are counts, and «each LDL particle contains a single apoB molecule», so LDL-P ≈ apoB for the LDL fraction.
- But apoB is the more COMPLETE measure, and that gap widens exactly where it matters. apoB counts
all apoB-containing particles — LDL plus VLDL and their remnants, IDL, and Lp(a) — whereas LDL-P
counts only LDL. Those remnant/IDL particles are elevated precisely in the hypertriglyceridemic /
metabolic-syndrome state, so LDL-P misses the extra atherogenic particles apoB captures right where
discordance arises. apoB is also the more standardised, widely-available assay. So LDL-P is a good
LDL-only proxy for apoB; apoB is the target.
[EXTRACTED for the correlation + single-apoB-per-particle + the particle list; INFERRED for the remnant-coverage ranking, which follows from apoB's particle set vs LDL-P's] - Lp(a) is one of those apoB-containing particles, but it behaves as a separate, largely genetic risk axis — stable over life and not lifestyle-modifiable — so its decision role (a baseline-risk multiplier that up-weights this same lowering lever) is treated on its own page: Lipoprotein(a) and Cardiovascular Risk.
A cheap flag for the discordant stratum — the TyG index [2026-08-09, Liu]
How do you find the insulin-resistant, hypertriglyceridemic person for whom LDL-C under-states apoB, without measuring apoB on everyone? A fasting insulin-resistance readout marks the stratum. The triglyceride-glucose (TyG) index — «considered a reliable surrogate marker of insulin resistance» (Liu et al., 2022) — is computed from the same two labs (triglycerides and fasting glucose) that define the discordant state, and Liu’s mechanism is precisely this atherogenic-dyslipidemia pathway: «Insulin resistance in liver and adipose tissues drives the development of atheroscle- rotic dyslipidemia, generates a low-grade inflammatory state, and increases release of inflammatory markers» (Liu et al., 2022). So a raised TyG is a prompt to measure apoB in that person — the marker points at the discordance; apoB, not TyG, is the causal quantity and the target. Reading TyG (or the TG/HDL ratio) as a rival to LDL/apoB confuses a downstream readout of the atherogenic state for the causal driver within it -> Insulin Resistance Surrogates and Cardiovascular Risk, Surrogate Outcomes. (This bridges the insulin-resistance and apoB axes that the wiki had held apart; the IR state itself originates in ectopic-fat overflow -> Ectopic Fat and Depot-Specific Risk, where “metabolic status matters” is causal for the glycemic axis but the CVD lipid target within it is still apoB.) (Ference et al., 2017; inferred from Liu et al., 2022)
Second outcome — high LDL-C is a new (2024) dementia risk factor
The 2024 Lancet Commission added high LDL-C as one of its 14 modifiable dementia risk factors, the midlife-specific case being the strongest -> Dementia Prevention and Modifiable Risk Factors. A newer meta-analysis (3 UK cohorts, n=1,138,488) found «each 1mmol/l increase in LDL-C was associated with 8% increased incidence of all cause dementia» (1.08, 1.03-1.14); a 1.19M-participant study put high LDL (>3 mmol/L) at HR 1.33 (1.26-1.41). (Livingston et al., 2024) The risk is greater in midlife than late life — consistent with this page’s cumulative-exposure thesis (the brain-vascular / amyloid mechanism accrues with duration). Note the outcome asymmetry: an IPD analysis at older baseline age found no LDL/HDL association with cognitive decline, so the causal read is a midlife-exposure one, not a late-life-level one. This is a decision-relevant addition to LDL’s outcome menu (cognition alongside ASCVD), not a change to the ASCVD verdict above.
The magnitude and the metric, now held directly — CTT + Marston [2026-08-05]
Ference asserts two things this page rested on but did not hold the primary evidence for: that lowering
apoB reduces risk in proportion to the reduction, and that apoB is the number to measure. Two sources now
supply that evidence directly (both extract the causal magnitude / metric, so both are sources:).
CTT 2010 — the per-mmol magnitude and the monotone shape (IPD meta-analysis, 26 statin RCTs, 169,138 participants). Per 1.0 mmol/L LDL-C reduction: major vascular events RR 0.78 (0.76-0.80), all-cause mortality RR 0.90 (0.87-0.93), with «no evidence of any threshold within the cholesterol range studied» — benefit persisting «even among those reaching 1·8 mmol/L (70 mg/dL) or lower». (Cholesterol Treatment Trialists’ Collaboration, 2010) This is the per-unit face of the cumulative thesis: because successive mmol multiply (0.78×0.78≈0.6), more and longer both pay, which is what lower for longer means quantified. Full magnitude home: LDL Lowering and Cardiovascular Events. Note this is statin-lowering — the concordant, no-off-target case Ference’s proviso privileges; the MCE counter-cases above show why a dietary LDL-C change does not inherit it.
Marston 2022 — apoB beats LDL-C, directly, at scale (UK Biobank n=389,529 + FOURIER/IMPROVE-IT n=40,430, plus cited Mendelian randomization). Entering the lipids simultaneously, «only apoB was associated» with MI (aHR 1.27 per SD, 1.15-1.40); non-HDL-C and TG fell to non-significant once apoB was held constant (Marston et al., 2022). This upgrades this page’s measure apoB claim from Ference’s assertion to a large-cohort + MR demonstration. Two refinements:
- Particle NUMBER, not TYPE or CONTENT, carries the risk. Adjusting for apoB, the TG/LDL-C ratio was flat — «for a given concentration of apoB-containing lipoproteins, the relative proportions of particle subpopulations may no longer be a predictor of risk». So the small-dense-vs-large-LDL distinction Astrup leaned on collapses into count the particles — apoB — which is this page’s held view, now with a direct test behind it. CONTESTED (the number-vs-content leg only). An independent MR (Helgadottir 2022, deCODE/Danish; no Ference/Ala-Korpela/Marston authors) separates apoB from non-HDL-C using 82/235 discordant variants and finds the opposite: CAD risk is proportional to non-HDL-C (cholesterol content), not to apoB particle number, and it attributes Marston’s observational apoB signal to confounding of two ~0.9-correlated traits. This does not touch the apoB-over-LDL-C verdict above (both agree there); it contests only Marston’s finer independent-of-content claim -> full joined issue: ApoB Particle Number vs Cholesterol Content. Confidence held (one MR vs Marston + the number-camp lineage; the leg is filed contested, not overturned).
- It confirms this page’s remnant-coverage ranking. The section above inferred apoB > LDL-P > LDL-C because apoB captures remnant/IDL/VLDL particles LDL-C misses. Marston states it directly: «non-HDL-C in particular is the preferred surrogate for apoB, as it incorporates TG-rich lipoproteins in addition to LDL» — the inference now has a source.
Not independent-E. Ference co-authors Marston, and CTT is the RCT evidence family Ference’s consensus
already synthesizes — same research programme. These are F/refinement + primary-evidence grounding (the
magnitude and metric the consensus asserted, now quoted from the underlying studies), not an independent
convergence that would earn [E-independent].
(inferred from Cholesterol Treatment Trialists’ Collaboration, 2010; Marston et al., 2022)
The genetic disentanglement — apoB is the NECESSARY trait (Richardson MVMR 2020) [2026-08-06]
The measure apoB claim above rested on Ference’s assertion and Marston’s observational multivariable analysis. Richardson supplies the same disentanglement by a genetic natural experiment — multivariable Mendelian randomisation — the design that removes reverse causation, confounding, and (via genetic instruments) much measurement error. A de-novo UK Biobank GWAS (up to 441,016 participants) built instruments for LDL-C, triglycerides, and apoB; these were carried into MR against CARDIoGRAMplusC4D (60,801 CHD cases). (Richardson et al., 2020)
Univariable, each atherogenic; multivariable, only apoB survives. Assessed one-at-a-time, LDL-C (OR 1.66 per SD), TG (1.34) and apoB (1.73) each raised CHD risk. Entered together in multivariable MR:
| Trait (per 1 SD, multivariable MR) | Direct effect on CHD |
|---|---|
| apoB | OR 1.92 (95% CI 1.31-2.81); P<0.001 — retained |
| LDL-C | OR 0.85 (0.57-1.27); P=0.44 — reversed to null |
| Triglycerides | OR 1.12 (1.02-1.23); P=0.01 — weakened |
“In multivariable MR, only apolipoprotein B … retained a robust effect, with the estimate for LDL cholesterol … reversing and that of triglycerides … becoming weaker.” (Richardson et al., 2020)
“These findings suggest that apolipoprotein B is the predominant trait that accounts for the aetiological relationship of lipoprotein lipids with risk of CHD.” (Richardson et al., 2020)
Read it correctly — apoB is necessary, LDL/TG are not discredited. The LDL-C null is NOT a claim that cholesterol is causally inert; Richardson is explicit that the findings «do not discredit the causal roles that LDL cholesterol or triglycerides play», because apoB does not occur in physiological isolation but is always accompanied by cholesterol and triglycerides. The source’s own verdict is necessity, in its words: «apolipoprotein B is the necessary element in order for lipoprotein lipids to exert their causal effect on risk of CHD—in other words, apolipoprotein B is a critical entity that underlies the relationship of lipid traits and risk of CHD.» (Richardson et al., 2020) The probabilistic corollary the paper draws: «changes in cholesterol or triglycerides that are not accompanied by commensurate changes in apolipoprotein B may not lead to altered risks of CHD.» (Richardson et al., 2020) This is the genetic-natural-experiment shadow of Ference’s concordance proviso and of the MCE/Sydney worked cases below: a cholesterol change discordant with apoB does not transmit to the outcome — now shown in genes, not just in a diet trial.
New extract — the HDL “protection” is apoB-mediated. Univariable MR made HDL-C (OR 0.80) and apoA-I (0.83) look protective; adjusted for apoB both «attenuated substantially to the null» (HDL-C 0.91, P=0.36; apoA-I 0.94, P=0.59), while apoB stayed causal (1.68). So HDL-raising therapies «will only have beneficial effects if they also lower apolipoprotein B» — the genetic rationale for the failed CETP/HDL-raising programme, and a reason not to treat HDL-C as an independent target. (Richardson et al., 2020)
Decision refinement — target the particle count. «the primary focus of lipid-modifying therapies ought to be the reduction in number of atherogenic lipoproteins (as measured by apolipoprotein B) rather than the reduction in cholesterol or triglycerides», especially where drugs have discrepant effects across the lipid traits. (Richardson et al., 2020) This grounds the page’s held measure apoB view with a genetic design, and sharpens it from measure to target and dose on apoB.
NOT independent-E — the citation chase kills it (the reportable finding). Richardson looks like the
independent second school the nucleus lacked (Bristol / MRC-IEU / Davey Smith; a genetic method-class,
not a consensus). It is not, on three grounds, any one sufficient: (i) Brian A. Ference is a
co-author of Richardson (author 5) — the author of the EAS Consensus this nucleus rests on and a
co-author of Marston; (ii) Richardson cites the CTT LDL-lowering RCT meta-analyses as antecedent
(“conclusively shown that lowering of cholesterol in atherogenic lipoproteins … reduction in risk”,
its refs 1-5 / the CTT collaboration) and explicitly builds on prior Ference multivariable-MR work — a
prior investigation using a form of multivariable MR that «obtained similar findings to those we report in
the present study» (ref 26 = Ference 2019); (iii) the whole apoB-particle-number / response-to-retention thesis
is one research programme (Sniderman / Ference / Ala-Korpela). So MR is not a new independent route — it
is a sharper instance of one of the four pillars Ference’s consensus already synthesizes, co-authored by
Ference. Verdict: F (refinement) + primary genetic grounding, NOT [E-independent]. The confidence
cap therefore holds — see Limits.
(Ference et al., 2017; inferred from Richardson et al., 2020)
The non-Ference second opinion — an AGNOSTIC MR reaches apoB-supreme, from a distinct group (Zuber MR-BMA 2021) [2026-08-07]
The three lipid sources above (CTT, Marston, Richardson) all share the Ference lineage, which is why the
causal case — however triangulated across RCT / observational / genetic designs — has stayed
confidence: medium. Zuber is the non-Ference second opinion the Limits section has been awaiting:
a multivariable MR from the Burgess/Cambridge group (no Ference co-author) using MR-BMA — an
agnostic Bayesian model-averaging algorithm that scores every combination of 30 lipoprotein
measures / metabolites and lets the data pick, rather than testing a curated LDL-C/TG/apoB triple.
Instruments from the Global Lipids Genetics Consortium; outcome from CARDIoGRAMplusC4D 2017 + UK Biobank
(453,595 individuals, 113,937 CAD cases). (Zuber et al., 2020)
apoB wins the agnostic search, and survives its own removal. The top-ranked model is apoB alone (model posterior probability 0.464); apoB has the strongest marginal evidence (marginal inclusion probability 0.868, FDR <0.005), and no other cholesterol or triglyceride measure is consistently selected. The load-bearing robustness test — the one Richardson’s curated design structurally could not run — is the leave-apoB-out: remove apoB from the candidate set and
«No alternative risk fac-tor had similar strength of evidence, suggesting that ApoB is indeed the most important risk factor and not just a rep-resentative of a group of highly correlated lipoprotein measures with similar evidence.» (Zuber et al., 2020)
apoB is also selected in every sensitivity analysis (an earlier CARDIoGRAMplusC4D release without UKBB; a 55-variant NMR-GWAS instrument set), with one honest exception: in UK-Biobank-only outcome data apoB ranked second behind very-small-VLDL triglyceride content, which the authors flag «should therefore be interpreted with some caution». The conclusion tracks the nucleus exactly — «ApoB, representing the number of hepatic-derived lipopro-tein particles, is the key determinant of CAD risk among lipid-related measurements» — and, like Richardson, it does not discredit LDL-C: «These results do not invalidate LDL-cholesterol as a causal risk factor for CAD risk. Indeed, LDL particles con-tain an apolipoprotein B molecule.» (Zuber et al., 2020)
Endogenous, not dietary — a transfer caveat that reinforces the MCE/Sydney/Astrup line. Zuber’s fasting-derived instruments mean «our analysis is well placed to answer causal questions about endogenous lipid pathways, but is less able to answer questions about lipoproteins from die-tary sources.» (Zuber et al., 2020) So this genetic apoB-supremacy speaks to hepatic/endogenous metabolism and drug targets — not directly to a dietary lipoprotein change, consistent with the diet-vs-drug non-interchangeability the proviso sections below establish.
Independence adjudication — the parameter table (why this is PARTIAL-E, not clean [E-independent]).
The independence claim is a cross-source claim, so it is built parameter-by-parameter against the closest
comparator (Richardson), not asserted:
| Parameter | Zuber 2021 | Richardson 2020 | Same quantity? / independence bearing |
|---|---|---|---|
| Analytical group | Burgess — Cambridge MRC-BSU + Imperial | Davey Smith — Bristol MRC-IEU | Different — independence gain |
| Ference a co-author? | No | Yes (author 5) | Differ — Zuber removes the cap’s lineage |
| Shared author | Ala-Korpela (author 3) | Ala-Korpela (author 4) | SAME — shared programme figure |
| Method | Agnostic MR-BMA over 30 measures | Curated univariable + MVMR of LDL-C/TG/apoB | Different — independence gain |
| Instrument / exposure GWAS | Kettunen NMR metabolomics, 24,925 (mainly Finnish) | de-novo UK Biobank GWAS (up to 441,016) | Different — independence gain |
| Outcome GWAS | CARDIoGRAMplusC4D 2017 + UK Biobank | CARDIoGRAMplusC4D | SHARED lineage — NOT independent |
| Estimand | Marginal inclusion prob 0.868; log-OR 0.392/SD | Frequentist OR 1.92/SD (MVMR) | Different metric — convergence is on the conclusion, not a shared number |
| Conclusion | apoB top-ranked across all datasets; LDL-C/TG not independently selected | apoB retained; LDL-C reverses to null | SAME — apoB-supreme |
The table shows the two agree on the conclusion (apoB is THE causal lipid) via different group, method, and exposure data — a real independence gain over Richardson — but share the outcome-data lineage (CARDIoGRAMplusC4D/UKBB) and one programme author (Ala-Korpela), and Zuber cites Ference 2019 + Richardson 2020 as the «two recent studies» it builds on. So the witnesses are not fully independent. Zuber even states the principle for its own internal analyses: «we refer to them as sensitivity analyses rather than replication analyses … rather than providing an independent replication of the findings.» (Zuber et al., 2020)
Verdict: partial-E / strong-F — the AWAITS is PARTIALLY cashed, the cap loosens but does not lift.
Zuber removes the Ference-lineage objection for this witness and adds method + group + exposure-data
independence, which is the strongest corroboration the nucleus has yet held — enough to move confidence
toward high. But shared outcome data + the shared Ala-Korpela authorship mean it is not the clean
independent-school replication that would license medium -> high on its own. It is therefore not marked
[E-independent]; the residual gap narrows to a non-Ference AND non-Ala-Korpela group on a distinct
outcome dataset (see Limits).
(Ference et al., 2017; Richardson et al., 2020; inferred from Zuber et al., 2020)
Self-critique (Zuber weave) [run 2026-08-07, before commit]. Laundered-E check (the R2 risk this
weave was built to avoid): the shared outcome-data lineage (CARDIoGRAMplusC4D/UKBB) and the shared
Ala-Korpela authorship are surfaced in the parameter table, not buried — the page explicitly refuses
the clean [E-independent] mark and downgrades to partial-E, so it does not launder shared-data agreement
as independence. Same-quantity: Zuber’s marginal-inclusion-probability / log-OR and Richardson’s
frequentist OR are named as different estimands; the convergence is filed on the conclusion
(apoB-supreme), not on an equated number. Overclaim: confidence stays medium (upper), not bumped to
high — the movement is stated as loosens the cap, does not lift it, with the narrowed residual AWAITS
kept explicit. Not a fake refinement: Zuber adds evidence Richardson structurally lacked (the agnostic
leave-apoB-out test + all-datasets robustness + different exposure data), so it is a real F, not a restate.
Current guidance has taken up measure apoB in the discordant — and cites this page’s own evidence [2026-08-06]
The measure apoB where LDL-C and particle number discord claim is no longer only the consensus’s; the 2026 US guideline operationalizes it into a clinical rule: «Apolipoprotein B (ApoB) testing can be useful to improve risk assessment and guide therapy once LDL-C and non-HDL-C goals are met, particularly in those with elevated triglycerides (TG) (≥150 mg/dL), diabetes, or low achieved LDL-C (<70 mg/dL)» (Blumenthal et al., 2026) — exactly this page’s discordant stratum (metabolic/CKM syndrome, diabetes, hypertriglyceridemia, treated-low LDL), where «LDL-C may appear at goal while apoB remains elevated, masking residual risk». The guideline also notes the Martin/Hopkins LDL-C estimator «markedly reduces discordance with apoB» — narrowing (not closing) the case for a separate apoB draw in the concordant majority.
This is guidance UPTAKE, not independent confirmation — the guideline reproduces the vault’s held
Marston evidence. ACC-AHA’s apoB stance cites the same data this page rests on: «only apoB remains
significant when assessed together (adjusted hazard ratio per 1 SD, 1.27 [95% CI, …»
(Blumenthal et al., 2026) — Marston 2022’s exact estimate. So the
guideline is the same Marston/CTT programme restated as a recommendation; it is F (a what-to-do
operationalization of the held claim), not [E-independent], and does not lift the confidence cap. Its
value is that the wiki’s held measure apoB view is now the standing US clinical rule, with a named
threshold set (TG ≥150 / diabetes / achieved LDL-C <70).
(inferred from Blumenthal et al., 2026; Marston et al., 2022)
A second causal axis beside apoB — residual inflammatory risk (CANTOS) [2026-08-08]
apoB is a causal axis, not the only one. CANTOS is the cleanest evidence that inflammation is a separate causal lever operating on top of well-controlled lipids: canakinumab cut cardiovascular events «independent of lipid-level lowering» (LDL/HDL unchanged), in statin-treated post-MI patients (Ridker et al., 2017), and Ridker names the stratum: «statin-treated patients with residual inflammatory risk as assessed by means of a high-sensitivity C-reactive protein level of 2 mg or more per liter at baseline have future event rates that are at least as high as, if not higher than, those among statin-treated patients with a residual risk due to LDL cholesterol level» (Ridker et al., 2017).
What this does and does not do to this page. It does not touch the apoB causality verdict — it bounds it: lowering apoB is necessary-and-causal for the lipid channel, but a person at goal on apoB can still carry residual inflammatory risk that lipid-lowering does not address. So “target apoB” is the lipid-axis answer, not the whole cardiovascular answer. Full treatment (the marker-vs-lever crux, the no-mortality-benefit boundary): Inflammation as a Modifiable Lever. (inferred from Ridker et al., 2017)
Limits
- The Ference-lineage cap is now LOOSENED but not LIFTED by Zuber —
confidence: medium(upper), held. The causal case triangulates across three Ference-lineage method-classes — RCT (CTT), observational multivariable (Marston), genetic multivariable MR (Richardson) — all landing apoB-supreme. Zuber 2021 partially cashes the independence owed: a genuinely different group (Burgess/Cambridge, no Ference), a different agnostic method (MR-BMA over 30 measures), and different exposure data (Kettunen NMR), all reaching apoB-supreme — the strongest corroboration held. That is enough to push confidence to the upper end of medium. But it is not a clean independent replication: Zuber shares the outcome-data lineage (CARDIoGRAMplusC4D/UKBB) with Richardson/Marston, shares one programme author (Ala-Korpela, on both Zuber and Richardson), and cites Ference 2019 + Richardson 2020 as antecedent — so the witnesses are not fully independent, and medium -> high is not licensed.AWAITS <a non-Ference AND non-Ala-Korpela apoB-causal analysis on a distinct outcome dataset> — an apoB-over-LDL-C causal demonstration from a group with no Ference/Ala-Korpela co-authorship and an outcome dataset other than CARDIoGRAMplusC4D/UKBB would be the genuine[E-independent]lift and license medium -> high. Zuber narrowed this residual; it did not close it. - Causal ≠ the only lever. LDL/apoB causation does not make it the largest absolute lever for a given person — absolute benefit still scales with baseline risk (Baseline Risk and the Relative-Absolute Split), and the net of a diet or drug depends on the whole strategy, not the LDL number alone.
- Off-target caveat is load-bearing: the any-mechanism-works claim is conditioned on no competing deleterious off-target effects — a real diet or drug can lower LDL and still net-harm through another pathway, so this validates the lipid channel, not any intervention wholesale.
A contested refinement — is a DIET-induced LDL-C change a good apoB proxy? (Astrup et al. 2020) [2026-07-29]
Astrup presses the concordance proviso above in a specific direction: it argues a diet-induced LDL-C reduction from SFA restriction is an unusually poor proxy for the atherogenic-particle change, so CVD benefit inferred from it is overstated. The claim, and where it stands against this page:
- It CONCEDES the causal core — «LDL particles play a causal role in the development of CVD» and there is “a relationship between lowering of LDL cholesterol and CVD benefit.” So this is not LDL-denial; it is a claim about the diet-induced change specifically. (Astrup et al., 2020)
- The argument: SFA restriction lowers mainly “large LDL particle subspecies… which are much less strongly related to CVD risk,” not the small dense LDL, and also lowers HDL — so the total:HDL ratio barely moves and “the potential benefit of dietary restriction of saturated fat could be substantially overestimated by reliance on the change in LDL cholesterol levels alone.”
- PURE grounds the diet-lipid discordance with data (now a held source). Dehghan (via its companion Mente 2017 lipid analysis) reports higher SFA → higher LDL but higher HDL, lower triglycerides, and lower ApoB/ApoA1 ratio (the stronger predictor), while higher carbohydrate → lower LDL but higher ApoB/ApoA1 — concluding «predicting the net clinical effect based on considering only the effects of nutrient intake on LDL cholesterol is not reliable». (Dehghan et al., 2017) This is the discordance-in-the-diet-direction claim in cohort data — and note it is itself an apoB argument (the hazard tracked apoB/apoA1, not LDL-C), so it reinforces measure apoB, this page’s held view, rather than displacing it. Observational and confounded by income, so directional not decisive.
- Where it lands against this page’s model. This page already holds the discordance mechanism — in the metabolic-syndrome/diabetic/hypertriglyceridemic state, LDL-C under-states apoB particle number, so measure apoB. Astrup and Ference agree LDL-C is an imperfect proxy and apoB is the target; they diverge on direction for the SFA case (Astrup: SFA raises mostly the benign large fraction, so diet-LDL overstates harm). The large-vs-small-LDL distinction Astrup leans on is substantially superseded by apoB particle number as the summary causal quantity (this page’s held view) — apoB counts the particles regardless of size, and PURE’s apoB/apoA1 signal is itself an apoB argument. So the honest status: the diet-induced-LDL-C caveat is real and this page already carries its mechanism (measure apoB, not LDL-C); the further claim that SFA’s LDL rise is benign-by-particle-size is contested and dated, and does not overturn apoB causality. (inferred from Astrup et al., 2020; Ference et al., 2017) -> full joined issue: Does Reducing Saturated Fat Reduce Cardiovascular Events.
The off-target/concordance proviso, worked — Ramsden MCE 2016 [2026-08-04]
Ference’s any-mechanism-works claim is conditioned twice: provided the LDL-C drop is concordant with the particle-number drop and there are no competing deleterious off-target effects. The recovered Minnesota Coronary Experiment is the corpus’s cleanest case of that proviso biting — a large cholesterol reduction that did not transmit to the outcome, in a double-blind RCT.
What MCE showed. Replacing SFA with corn-oil LA lowered serum cholesterol -13.8% vs -1.0% control, yet produced no mortality benefit in the full randomized cohort or the 5-trial MA (CHD mortality 1.13, all-cause 1.07) -> Linoleic Acid and Cardiovascular Disease. Ramsden’s own reading is an explicit off-target/concordance argument: «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)
Same-quantity discipline — this does NOT refute apoB causality; it lands inside the proviso. MCE measured total serum cholesterol only (Ramsden concedes LDL/HDL subfractions were not assayed), so it cannot show the cholesterol drop was concordant with a particle-number (apoB) drop — the first escape hatch. And LA plausibly adds a competing off-target effect (increased LDL-oxidation susceptibility) — the second. So MCE is consistent with Ference’s framework, not a counterexample to it: it is the worked demonstration that the concordance + no-off-target proviso is load-bearing, not boilerplate. The decision-relevant transfer: a diet-induced cholesterol change is not interchangeable with a drug-induced apoB change — validate the marker->outcome transmission for the agent actually used -> Surrogate Outcomes. This is the empirical shadow of ESC’s «irrespective of the drug» invariance claim: the invariance is evidenced across LDL-lowering drugs; MCE shows it is not automatic for a dietary LDL change. (Ference et al., 2017; inferred from Ramsden et al., 2016)
The Sydney companion (Ramsden 2013) is the same proviso, biting harder — the second escape hatch made concrete. SDHS lowered total cholesterol MORE in the LA arm (-13.3% v -5.5%) yet the arm had higher randomized mortality, and Ramsden proposes exactly the competing off-target effect the proviso names: oxidized-LA metabolites (OXLAMs) as an atherogenic route independent of the cholesterol drop, strongest in the trial’s smokers/drinkers. Like MCE it measured total cholesterol only (not apoB/particle number), so it too lands inside Ference’s proviso rather than refuting it — a diet-lowered cholesterol with a plausible off-target harm, not a clean apoB-concordant reduction. It sharpens the same decision transfer (a dietary cholesterol change is not interchangeable with a drug-induced apoB change) with a randomized adverse outcome instead of MCE’s null. Small single-blind high-dose n-6-selective secondary-prevention trial — the appraisal caveats are on Linoleic Acid and Cardiovascular Disease; not independent-E of MCE (same Ramsden program). (inferred from Ramsden et al., 2013)
Self-critique (Ramsden MCE weave) [run 2026-08-04, before commit]. Overclaim: the section
explicitly says MCE is consistent with, not a counterexample to, apoB causality — the two proviso
escape hatches (total-cholesterol-only measurement; LA-oxidation off-target) are named and both
source-grounded, so it does not launder a contrarian RCT into a refutation of Ference. Not-E: no
independence claimed — it is an F/worked-instance. Same-quantity: diet-total-cholesterol vs
drug-apoB-particle kept distinct throughout.
Self-critique [run 2026-07-29, before commit]
- Over-claim check: the causal verdict is quoted from the consensus, not asserted by the wiki; the
lower for longer and measure apoB consequences are tagged to the source’s own proviso (concordance
- no off-target). The single-source/
confidence: mediumlimit is stated.
- no off-target). The single-source/
- Surrogate framing: LDL/apoB is presented as the validated-surrogate exemplar, explicitly the counter-case to the wiki’s usual surrogate caution — not a licence to trust surrogates generally.
- Mechanism-gradient discipline (cold-audit fix): the causal weight is attributed to the concordance of the genetic/MR evidence AND the LDL-lowering RCTs (the source’s own pillar; the RCTs are its most compelling causal evidence), not to genetic/MR alone — the top of the mechanism-strength gradient, not a mechanistic plausibility story.