Open on the substitution, not the fat
The lay question is what fats should I eat? The evidence answers a narrower one: what you swap a fat for — and what food carries it, and who is eating it. Appraise the swap, never the fat in a vacuum.
- What you replace a fat with decides its effect. A saturated-fat cut into more refined carbohydrate is a different exposure from the same cut into polyunsaturated oil, and the two do not carry the same sign.
- Type, not total amount, is the lever the evidence supports. The single largest RCT of cutting total fat found no cardiovascular benefit; how the fat is composed is what moves the outcome.
- Replacing saturated fat with unsaturated fat modestly cuts cardiovascular events and does not change how long you live — the evidence is strongest for the polyunsaturated swap, thinner and mostly observational for the monounsaturated one.
- The events benefit is real but modest, and sensitive to trial quality — said plainly here, not buried.
- Removing industrial trans fat is the single clearest, least-contested action in the whole picture.
- The loud controversies collapse into one quiet move — seed-oil panic and saturated-fat villainy both resolve to shift the mix toward whole-food unsaturated fats; the seed-oil-specific case is handled in its own deliverable.
- For a low-risk eater this is a low-cost hedge, not a promised heart-attack reduction — the honest ceiling.
The most load-bearing of these is the first. So start where the sign is set: which fat replaces which.
Replace saturated with unsaturated to set the sign
The replacement sets the outcome, not saturated fat on its own. Rank the swaps two ways and they agree. By how much they lower LDL cholesterol per 1% of energy exchanged: polyunsaturated -0.055, monounsaturated -0.042, carbohydrate -0.033 mmol/L, high certainty, and the lowering held down to a saturated-fat intake of 2% of energy — below the studied range there is no data (World Health Organization, 2023). And by guideline strength: replacing with polyunsaturated fat is a strong recommendation; with plant monounsaturated fat, or with fibre-bearing carbohydrate, only conditional (World Health Organization, 2023). The polyunsaturated swap wins on both counts — largest lipid effect and the firmer recommendation.
Cut saturated fat without naming the replacement is half an instruction — Better than What is the load-bearing question, not the fat in isolation.
One caveat keeps the swap honest: saturated fat is never removed into a vacuum. WHO scopes its replacement advice to energy balance, and notes that in positive energy balance «SFA intake may be reduced in part or entirely without the need for a replacement nutrient» (World Health Organization, 2023). So the comparator shifts with the eater’s calorie state — saturated fat versus what in balance, saturated fat versus nothing in surplus.
And two swaps are not one intervention measured in two groups. On hard cardiovascular events the two swaps are close in relative terms — polyunsaturated RR 0.79 (95% CI 0.62-1.00), carbohydrate RR 0.84 (0.67-1.06), both Low certainty, both touching or crossing the null (World Health Organization, 2023). Their absolute gap looks far wider — 50 fewer events per 1000 for the polyunsaturated swap against 12 fewer for the carbohydrate one — but only because the polyunsaturated trials ran at a 23.8% control event rate against 7.6% for the carbohydrate trials, roughly three times the baseline risk, in disjoint trial sets.
That makes saturated-fat-to-polyunsaturated and saturated-fat-to-carbohydrate genuinely different exposures, not one relative effect read across a clean baseline-risk split. Reading 50 fewer per 1000 against 12 fewer as polyunsaturated fat beating carbohydrate four-fold compares populations, not nutrients.
The swap that most moves the sign is therefore saturated fat to polyunsaturated fat. So: what does the hard-outcome evidence actually show for that move?
Read the SFA->events evidence as modest and RoB-sensitive
One hard outcome clears the null. Reducing saturated fat cut combined cardiovascular events to RR 0.83 (95% CI 0.70-0.98) — 15 fewer per 1000, Moderate certainty, 13 RCTs (World Health Organization, 2023). In person-count terms that is a number needed to treat of 56 in primary prevention and 53 in secondary, over about four years (Hooper et al., 2020). The Cochrane review supplying these RCT numbers is Hooper 2020 — WHO’s trial estimates are Hooper’s, one evidence base graded twice, so the two do not count as independent witnesses.
The same review attenuates its own headline, and the attenuation belongs in the same breath as the number. Restricting to trials at low summary risk of bias weakens the combined-events benefit to «more marginal protection» (Hooper’s Analysis 1.36; the exact figure is not narratively reported) (Hooper et al., 2020). For myocardial infarction specifically, the low-risk-of-bias restriction moves the estimate to a frank null, RR 0.93 (95% CI 0.81-1.08) (Hooper et al., 2020).
Read this precisely. The RR 0.93 (0.81-1.08) figure is the myocardial-infarction sensitivity analysis (Analysis 2.2), not the combined-events one — a critique that pins 0.93 on combined cardiovascular events has swapped the two outcomes. And the attenuation carries its scope: the primary RR 0.83 stands and survives most other sensitivity checks — trials that aimed to reduce saturated fat, that achieved a cholesterol reduction, or excluding the largest trial (Hooper et al., 2020). The risk-of-bias restriction is the exception, and the Moderate grade already encodes that fragility — the one-step GRADE downgrade tracks exactly these internal-validity limits. So the benefit is neither a settled large win nor debunked — it is real, modest, and trial-quality-sensitive.
Whether the harm is saturated fat’s or the benefit is polyunsaturated fat’s stays genuinely open — Hooper’s own replacement subgrouping cannot separate them (Astrup et al., 2020). That is why the robust action is the substitution — replace saturated fat with polyunsaturated or whole-food unsaturated sources — which is correct under either reading. It licenses neither avoid all saturated fat nor eat more seed oil.
The events benefit is real but modest. The mortality picture is different — and, unusually here, both camps agree on it.
Expect no mortality change from cutting saturated fat
Cutting saturated fat does not measurably change how long people live. All-cause mortality sits at RR 0.96 (95% CI 0.90-1.03) — 2 fewer per 1000, Moderate certainty (World Health Organization, 2023). The cause-specific mortality rows tell the same story: cardiovascular mortality RR 0.94 (0.78-1.13) and coronary mortality RR 0.97 (0.82-1.16), both spanning no effect (World Health Organization, 2023). Every mortality interval crosses the null.
This is where the reassessment critics and the guideline bodies converge rather than clash: the heterodox pole itself finds «no beneficial effects of reducing SFA intake on cardiovascular disease (CVD) and total mortality» (Astrup et al., 2020) — for mortality, exactly the Cochrane result.
A Moderate-certainty null is a result, not missing data. It sits in the no-meaningful-effect state, distinct from insufficient evidence — the trials were run, the estimate is reasonably graded, and it lands on no effect. So reducing saturated fat is not a longevity move, and should not be sold as one; anyone reading strong recommendation as this measurably saves lives is reading past the evidence.
Type of fat therefore moves cardiovascular events modestly and mortality not at all. Which raises the next question: does the amount of fat — the total, rather than the composition — move anything?
Target the type, not the total amount
The lever the evidence supports is which fat, not how much fat — and the largest randomized test of cutting total fat says so directly. The WHI Dietary Modification trial randomized 48,835 postmenopausal women to a low-fat pattern — total fat cut toward 20% of energy, the displaced calories replaced mainly by carbohydrate (more grains, fruit, and vegetables) at constant body weight — and over a mean 8.1 years found no significant effect on coronary heart disease (HR 0.97, 95% CI 0.90-1.06), stroke (HR 1.02, 0.90-1.15), or total cardiovascular disease (HR 0.98, 0.92-1.05) (Howard et al., 2006). This is a no-meaningful-effect on the tested contrast, a result and not missing data.
Read the null precisely — it bounds the amount axis, not fat quality. WHI moved total-fat quantity and replaced it with refined starch; it was not a saturated-fat-to-unsaturated swap, and its authors say so: «The trial is not a test of the dietary guidelines currently recommended for prevention of CVD … [that specify] replacement of saturated and trans fat with monounsaturated and polyunsaturated fat» (Howard et al., 2006). So the one large RCT of the quantity lever lands on no cardiovascular benefit, while the fabric’s graded evidence on composition stands untouched — replacing saturated fat with unsaturated fat modestly moves events and does not move mortality (see Saturated Fat Intake and Replacement, Low-Fat Dietary Pattern and Cardiovascular Disease).
And WHI is no longer the only randomized test of the quantity lever. Hooper’s 2012 Cochrane review pooled the reduce-and-modify-fat RCTs (24 comparisons, 65,508 people) and contains WHI — so its reduced-fat subgroup is WHI generalized, not a second independent witness (type-F refinement, not independent backing). That subgroup is flatly null: «There was no suggestion of an effect on cardiovascular events in studies that compared reduced fat vs usual intake (RR 0.97, 95% CI 0.87 to 1.08 … 50,655 participants)» (Hooper et al., 2012). The whole class of total-fat-reduction trials lands where WHI did, not one trial alone. What moved events was the modification subgroup (saturated fat -> unsaturated: RR 0.82, fixed-effects 0.83), and the review’s overall 14% events reduction (RR 0.86, 0.77-0.96, moderate GRADE) is carried by the modifying arms, not the reducing ones (Hooper et al., 2012). Hooper’s own conclusion is this deliverable’s thesis at pooled-RCT scale — a benefit «on modification of dietary fat, but not reduction of total fat, in longer trials» (Hooper et al., 2012).
So a low-fat-vs-higher-fat quantity target is not what the evidence recommends chasing; the swap is. If type matters, it matters through a mechanism — and for saturated fat that mechanism runs through LDL and apoB. How far does that surrogate carry the weight of an averted event?
Trust the LDL/apoB direction, not the diet-to-drug magnitude
The surrogate is unusually well-founded in direction. LDL and the apoB-containing particles that carry it cause atherosclerotic cardiovascular disease, and the dose is cumulative — risk tracks magnitude times duration, the area under the LDL/apoB curve over a lifetime, not the current snapshot -> LDL ApoB and Cumulative Exposure. Each 1.0 mmol/L of LDL-C lowered cuts major vascular events by about a fifth (RR 0.78, 0.76-0.80) with no threshold in the range studied (Cholesterol Treatment Trialists’ Collaboration, 2010). So the saturated-fat -> LDL step is a directionally meaningful one: it points the right way, on a validated-surrogate exemplar rather than a marker-of-convenience.
A controlled-feeding factorial trial adds internal-validity weight to that saturated-fat -> LDL step. The APPROACH trial crossed two background saturated-fat levels against a meat-versus-plant protein swap, and found the two act as separate, additive levers: background saturated fat raised LDL-C and apoB regardless of the protein source, while «LDL cholesterol and apoB were higher with red and white meat than with nonmeat, independent of SFA content» (Bergeron et al., 2019). Isolating the saturated-fat effect from the protein-source effect inside one trial firms the direction the swap-the-SFA guidance already asserts — it does not change it. This is a surrogate (lipid) endpoint; no cardiovascular events were measured, so it strengthens the SFA -> LDL direction without adding a hard-outcome claim.
But direction is not magnitude, and a diet-induced LDL-C change is not interchangeable with a drug-induced apoB change. A double-blind trial demonstrates this most clearly — the surrogate moved and the outcome did not: replacing saturated fat with corn-oil linoleic acid lowered serum cholesterol -13.8% versus -1.0% in controls, yet produced no mortality benefit (CHD mortality 1.13, all-cause 1.07) (Ramsden et al., 2016). The causal model already names why — benefit follows only when the cholesterol drop is concordant with a real fall in particle number and carries no competing off-target harm -> Surrogate Outcomes. A statin’s LDL drop meets that proviso; a dietary one is not guaranteed to.
The gap widens where metabolism is impaired. In the insulin-resistant, hypertriglyceridemic state, LDL-C under-states the atherogenic particle burden. Small, dense particles pack more apoB into a given cholesterol mass, so apoB is the number to measure, and a raised triglyceride-glucose (TyG) index is a cheap prompt to draw it -> LDL ApoB and Cumulative Exposure. Genetic evidence sharpens the point: entered together, only apoB retains a robust effect on coronary disease (OR 1.92, 1.31-2.81) while the LDL-C estimate reverses to null (Richardson et al., 2020). Read this as a caution against over-reading the marker: the surrogate is firmer than the hard-event evidence it stands in for, so a moving LDL number is a well-warranted signal, not a proven averted event.
One exposure, though, needs no surrogate argument at all — its harm is direct and uncontested.
Cut industrial trans fat — the one unambiguous harm
Trans fat carries the least contested recommendation in the whole dietary-fat picture. WHO’s advice follows the saturated-fat shape at a lower threshold: reduce trans-fat intake to 1% of total energy (strong), reduce further below 1% (conditional), and replace it with polyunsaturated or plant-source monounsaturated fat (conditional) (World Health Organization, 2023). The 1%E figure marks the edge of the evidence, not a knee in the curve — read it as the intake below which trials thin out, not as a demonstrated point of diminishing return.
WHO also merged the two trans-fat sources rather than splitting them. A boundary that looks load-bearing — industrial (partially-hydrogenated-oil) and ruminant trans differ in origin, isomer profile, and how a consumer meets them — was tested and dropped: «Based on the evidence review for TFA, the WHO NUGAG Subgroup on Diet and Health concluded that industrially produced and ruminant TFA behaved in a similar manner with respect to effects on health and therefore formulated recommendations for total TFA» (World Health Organization, 2023). The recommendation is for total intake from both.
One feature sets trans fat apart from every other fat in this cut: the body cannot synthesise it. Trans fat is non-endogenous, so its biomarker is a clean intake measure — a within-fat boundary that carries real information -> Is the Food Category Doing Any Work. That measurability is part of why the trans-fat signal is the firmest fat finding held here.
Treat MUFA benefit as thin and pattern-level
Monounsaturated fat has no dedicated page in the fabric, and the reason is the evidence. WHO’s plant-MUFA-for-saturated-fat replacement is a conditional recommendation, but its randomised cardiovascular evidence is a single trial — 52 participants, 4 events, RR 3.00 (0.33-26.99), rated Very low: «only one small trial with olive oil as an intervention was included in the monounsaturated fatty acids subgroup» (World Health Organization, 2023). WHO’s moderate MUFA certainty rests entirely on observational data (cardiovascular disease RR 0.90, 0.84-0.96, 7 fewer per 1000, 3 studies) (World Health Organization, 2023). The RCT cell that exists points the other way, on numbers too small to mean anything.
The only hard-outcome MUFA signal of any weight is PREDIMED — and it is a dietary pattern, not isolated MUFA. The Mediterranean diet cut cardiovascular events roughly 30% (HR 0.70), an absolute ~1.7-2.1 percentage points over 5 years (Estruch et al., 2018) — but in a high-risk population (~49% type-2 diabetes, ~82% hypertensive), whose authors say generalisation «to persons at lower risk requires further research». That is an olive-oil-rich whole-diet effect at high baseline risk, not a class-level trial of MUFA. Label the MUFA evidence observational and pattern-level; where the olive-oil outcome data actually lives is Mediterranean Diet and Cardiovascular Events and Fish and Seafood Consumption.
The other polyunsaturated leg, n-6 linoleic acid, is where a live public controversy sits — and it is the next section’s subject.
Read n-6 linoleic acid as neutral-to-protective as a class
Linoleic acid is the main omega-6 fat in bottled seed oils, and the popular claim that it drives heart disease is the one place where the evidence is loud and the data are quiet. Every arm the fabric holds points away from harm. An objective-biomarker cohort pool ties higher tissue linoleic acid to lower total cardiovascular disease (HR 0.93, 0.88-0.99) and lower cardiovascular mortality (HR 0.78, 0.70-0.85) (Marklund et al., 2019). The randomized-trial arm is more equivocal but never adverse: increasing omega-6 fats is little-or-no-effect on all-cause mortality (RR 1.00) and on cardiovascular events (RR 0.97), and may modestly cut myocardial infarction (RR 0.88) (Hooper et al., 2018).
Crucially, the observational and trial evidence here do not diverge into a tension — the direction agrees. A mortality meta-analysis puts higher linoleic acid at lower all-cause mortality via both its dietary arm (RR 0.87) and its biomarker arm (RR 0.91) (Li et al., 2020), so the cohort and biomarker readings land together rather than at odds. And the proposed mechanism fails on its own endpoint: feeding linoleic acid does not raise any commonly measured inflammatory marker (CRP, IL-6, TNF-alpha, fibrinogen) across randomized trials (Johnson & Fritsche, 2012). The toxic seed oil harm thesis therefore fails on observational, trial, and mechanistic evidence alike — three independent chances to appear, none taken.
One within-class nuance earns a mention without becoming the verdict. The Sydney Diet-Heart trial — a secondary-prevention, post-heart-attack population fed a concentrated, n-6-selective safflower-oil dose with no n-3 — found the treated arm died more (all-cause HR 1.62, cardiovascular 1.70, coronary 1.74, all borderline) (Ramsden et al., 2013). This is a real signal, but a narrow one: one small single-blind trial at an extreme dose in established coronary disease, already pooled into the benefit-netting reviews above, and answering a different question than the general-population class. It flags a possible harm in a specific high-dose n-6-selective stratum; it is not the class verdict, which stays neutral-to-protective.
The louder seed-oil arguments — oxidation and aldehydes on heating, hexane extraction, the n-6:n-3 ratio, FADS genotype — are a separate question about a processed food, not about the fatty acid, and are appraised in Seed Oils. This section holds only the class-level finding.
Split n-3 by form and hold the class evidence thin
n-3 as a single class is under-anchored in the fabric, and for a good reason: the omega-3 label spans a benefit and a null depending on the form, the dose, and the population. The evidence lives on the forms, not the class, so the honest move is to split it rather than average it.
Marine EPA and DHA as food — oily fish — are appraised as a food exposure in Fish and Seafood Consumption, not here. As a capsule or purified isolate, the trials openly disagree: a 1 g/day EPA+DHA supplement in a general primary-prevention population was null (HR 0.92) (Manson et al., 2019), while 4 g/day of purified EPA cut events by a quarter (Bhatt et al., 2019). But that benefit sits in a high-risk, statin-treated, high-triglyceride stratum against a contested comparator, so it is not a general low-risk lever (Cardiometabolic Interventions and Hard CV Outcomes in Low-Risk People; Dietary Supplements). Different compound, different dose, different population: the divergent results are a distinction, not a contradiction (Is the Food Category Doing Any Work). Plant-source ALA is barely covered in the fabric — a genuine gap; no direction is inferred here.
Across every fat class in this cut, two dependencies keep resurfacing: the food that carries the fat, and the metabolic state of the person eating it.
Check the food matrix and the eater’s metabolic state
The label saturated fat sorts foods that behave differently once you eat them. Processed and unprocessed red meat carry much the same saturated fatty acids, yet processed meat tracks with coronary and diabetes risk while unprocessed red meat does not — so whatever carries the risk, it is not the saturated fat: «the SFA content of meat is unlikely to be responsible for this association» (Astrup et al., 2020). Cheese and yogurt, both saturated-fat-rich, associate inversely with cardiovascular risk; dark chocolate’s stearic acid is neutral. The nutrient predicts harm; the whole food carrying it often predicts the opposite -> Is the Food Category Doing Any Work. A recommendation phrased at the nutrient — cut saturated fat — is aimed at an exposure that describes no single food on the plate.
The effect also turns on who is eating. Saturated fat’s harm, and the lipid marker that stands in for it, both depend on adiposity and metabolic state. In an insulin-resistant, hypertriglyceridemic adult, small dense particles pack more atherogenic apoB into a given cholesterol mass, so LDL-C under-states the real particle burden and apoB becomes the number to trust (Ference et al., 2017) -> LDL ApoB and Cumulative Exposure. A raised triglyceride-glucose (TyG) index — «a reliable surrogate marker of insulin resistance» (Liu et al., 2022) — flags that stratum cheaply.
A boundary runs even inside a fat class: the body makes its own even-chain saturated and monounsaturated fat, so «Biomarkers generally perform poorly for fatty acids that can be produced endogenously, including even-chained saturated and monounsaturated fatty acids» (Willett, 2012), while trans fat and the omega-3s, which the body cannot synthesise, leave a clean intake signal — a within-category line that carries real information. These dependencies set up the harder point: some questions the evidence structurally cannot answer.
Name what the evidence structurally cannot show
Researchers measure diet by asking people, and the error is large enough to be the dominant fact about almost any fat dose-response. Self-report compresses the exposure range from both ends — «those who consumed considerably less than the average were more likely to overreport intake, while those who ate more than the average tended to underreport» (Willett, 2012) — which flattens any real gradient. The asymmetry is the usable part: flat-slope compression can hide a knee but never manufacture one, so a null or monotone fat curve is weak evidence of no gradient, not proof of one -> Measurement Error in Dietary Assessment.
An objective biomarker is the obvious escape, and it works only for fats the body cannot make. For linoleic acid, essential and non-synthesizable, the tissue level is a genuine intake marker, and the error-laden questionnaire and the biomarker agree (Li et al., 2020). For even-chain saturated and monounsaturated fat, de novo synthesis blends intake with carbohydrate and energy status, and the escape is shut. So the honest inventory: the total-fat dose-response shape is barely estimable (the one large RCT tested a single fat->carbohydrate reduction, not a curve), plant omega-3 (ALA) is a genuine named gap, and whether the cardiovascular signal is saturated-fat harm or polyunsaturated benefit stays an open attribution -> Is the Food Category Doing Any Work. Given all this, what does a low-risk eater actually do?
Act on the substitution a low-risk eater can sustain
A substitution survives every one of these uncertainties, not a target. Shift the fat mix from saturated toward whole-food unsaturated fats — nuts, seeds, olive oil, fish, fatty plants — and remove industrial trans fat. That decision is right whether the eventual cause is saturated-fat harm or unsaturated-fat benefit, and it depends on none of the open attributions above. Two over-reads to refuse: avoid all fat (the amount axis is a gap, not a lever) and drink more bottled seed oil (the class evidence supports the swap, not a supplement). The realistic comparator sets the sign, so judge the change against what the fat actually replaces -> Better than What, and count adherence as part of the effect — a modest sustained shift beats a strict abandoned one.
Then keep the size honest. For a genuinely low-risk person the proven hard-outcome benefit is small in absolute terms, because absolute benefit is the relative effect times a low baseline risk -> Baseline Risk and the Relative-Absolute Split. No cardiometabolic lever the evidence holds delivers a large absolute reduction in cardiovascular events at low risk. Even blood-pressure lowering, whose relative effect is proven in primary prevention (HR 0.91, 0.89-0.94, without prior cardiovascular disease) (Blood Pressure Lowering Treatment Trialists Collaboration, 2021), buys little on a small baseline -> Cardiometabolic Interventions and Hard CV Outcomes in Low-Risk People. So the fat swap is a low-cost hedge, not a promised heart-attack reduction. Reporting that ceiling is itself the finding, one that licenses a low-risk eater to stop chasing a large effect the evidence does not offer.
Evidence box
Question ’For an adult choosing what fats to eat: what is the effect of the amount and type of dietary fat (saturated, monounsaturated, n-6 and n-3 polyunsaturated, industrial trans) on each patient-important outcome (cardiovascular events, all-cause mortality), what is the dose-response shape, and on what does any effect depend — what the fat replaces, the food matrix carrying it, or the eater”s metabolic state?‘ Evidence included 21 sources — 9 gold, 10 high, 1 weak Overall certainty Medium (see Rating Certainty of Evidence) Source-selection note 1 source(s) below the gold evidence bar feed this page: Astrup (narrative review, weak). Each labelled by tier; none load-bearing for the core claims. Last updated 2026-09-05 · Independently reviewed: No · Full edit history