Nucleus of the organic cluster. This page holds the health appraisal of the “organic” label:
what the measured differences are, how large they are against total diet, and — the load-bearing move —
whether the certification or an underlying exposure the label only partly tracks is doing the work.
It appraises the health axis only; the environmental / animal-welfare / farming-system reasons many
people buy organic are legitimate and are the buyer’s to weigh at layer 3 (the one-axis rule) — this
page does not price them.
Evidence-tier note (confidence: low). The health-outcome evidence for the organic label is
uniformly non-gold — nutrient/residue composition meta-analyses (Baranski, Srednicka-Tober) plus one
self-selected observational cohort (Baudry/NutriNet), now joined by a pooled SR/MA of hard cancer
endpoints (Theodoridis 2025) that is itself non-gold — three observational cohorts, GRADE certainty
very low, MDPI-Life. (Supersedes the earlier no systematic review of hard endpoints statement,
2026-08-20: an SR/MA now exists, is null, and does not lift the tier.) There is still no outcome RCT,
and none is likely (a certification cannot be blinded). The low confidence is a
structural feature of the field, not a gap acquisition can close — the honest finding is thin
evidence, labelled as such rather than upgraded.
The frame: “organic” is a provenance label, not an exposure
Provenance argues health in neither direction (the appeal-to-nature is symmetric rule): “natural / organic = healthier” is as unwarranted as “processed = harmful”. The health question does not have one answer because “organic” is a certification bundle — no synthetic pesticides, no synthetic fertiliser, no routine antibiotics, welfare rules, and often but not always pasture/outdoor access — and the bundle decomposes into separable sub-questions with different evidence and different effect sizes (a type-B disambiguation). Answer them separately; never as one verdict on the word. -> Is the Food Category Doing Any Work
The measured differences that survive appraisal are: a lower pesticide-residue incidence (of unproven health consequence at real-world levels), small nutrient differences (that run in both directions), and a pasture-driven fatty-acid profile in animal products — which is a feed effect, not a label effect. The hard-outcome case is the weakest link and is heavily confounded.
Sub-question 1 — pesticide / toxin residues: lower on organic, health consequence unproven
Baranski (meta-analyses over 343 publications) found “the frequency of occurrence of pesticide residues was found to be four times higher in conventional crops, which also contained significantly higher concentrations of the toxic metal Cd.” (Barański et al., 2014) The cadmium point estimate is “The on average 48 % lower Cd concentrations found in organic crops” — but the weighted percentage-difference CI on Cd was wide (−48%, 95% CI −112, 16), so the significant Cd difference rests on the standardized-mean-difference analysis, not the percentage metric. (Barański et al., 2014)
The decision question is not whether a difference is detectable but whether it reaches a patient-important outcome. Smith-Spangler (Annals SR) supplies the discipline: organic produce carried a lower residue-contamination risk (risk difference 30%), but «differences in risk for exceeding maximum allowed limits were small.» (Smith-Spangler et al., 2012) Detectable-difference is not health-difference when both arms sit largely below regulatory limits — the surrogate-vs-outcome gap. -> Surrogate Outcomes
Smith-Spangler’s one clean safety signal was elsewhere: bacteria resistant to 3+ antibiotics were commoner on conventional chicken and pork (risk difference 33%), while Escherichia coli contamination did not differ by farming method. (Smith-Spangler et al., 2012)
Sub-question 2 — nutrient content: small differences, running BOTH ways
Baranski found antioxidants/(poly)phenolics substantially higher in organic, with the phenolic classes “an estimated 19 (95 % CI 5, 33) %, 69 (95 % CI 13, 125) %, 28 (95 % CI 12, 44) %, 26 (95 % CI 3, 48) %, 50 (95 % CI 28, 72) % and 51 (95 % CI 17, 86) % higher, respectively” (phenolic acids, flavanones, stilbenes, flavones, flavonols, anthocyanins). (Barański et al., 2014)
But the SR that weighted outcomes and tested clinical significance found almost nothing. Smith-Spangler: “The published literature lacks strong evidence that organic foods are significantly more nutritious than conventional foods” — and of all nutrients “the estimate for phosphorus” alone reached significance, a difference the authors call not clinically significant. (Smith-Spangler et al., 2012)
Symmetric standards: the differences are not uniformly pro-organic. The milk meta-analysis found organic milk higher in alpha-tocopherol and Fe but lower in iodine and selenium — organic loses on two micronutrients. Uniformly favourable treatment would be a halo tell; the honest reading is a mixed, small-magnitude picture against total-diet intake. (Givens & Lovegrove, 2016)
Sub-question 3 — the cattle case: the label tracks the FEED, not the certificate
This is the worked example of the label and the causal exposure coming apart. Organic milk and meat carry a more desirable fatty-acid profile — and the driver is pasture/forage, which the organic certificate does not guarantee.
Milk (Srednicka-Tober, 170 studies): “There were no significant differences in total SFA and MUFA concentrations between organic and conventional milk.” n-3 PUFA were higher in organic «by an estimated 7 (95 % CI −1, 15) %» for total PUFA and 56% (CI 38, 74) for n-3 PUFA; alpha-linolenic acid +69% (CI 53, 84), very-long-chain n-3 (EPA+DPA+DHA) +57% (CI 27, 87), conjugated linoleic acid +41% (CI 14, 68). (Givens & Lovegrove, 2016)
Meat (Srednicka-Tober, 67 studies): SFA “similar”, MUFA “slightly lower” in organic; total PUFA and “n-3 PUFA, which were an estimated 23 (95 % CI 11, 35) % and 47 (95 % CI 10, 84) % higher in organic meat, respectively.” Heterogeneity was high and “could be explained by differences between animal species/meat types” — the boundary is doing little work across species. (Średnicka-Tober et al., 2016)
Both meta-analyses attribute the difference to feed, not certification — stated in their own voice:
- Milk: “Redundancy analysis of data from a large cross-European milk quality survey indicates that the higher grazing/conserved forage intakes in organic systems were the main reason for milk composition differences.” (Givens & Lovegrove, 2016)
- Meat: “Evidence from controlled experimental studies indicates that the high grazing/forage-based diets prescribed under organic farming standards may be the main reason for differences in FA profiles.” (Średnicka-Tober et al., 2016)
Consequence for the decision. The causal lever is grass vs grain, and organic certification only correlates with pasture (it mandates outdoor access but not a forage-dominated diet). A pasture-raised conventional animal can therefore beat an organic grain-fed one on the nutrient that actually reaches the milk/meat. Specify the exposure (grass-fed / pasture), not the word (organic) — the label is a partial proxy for the thing that matters. This is Is the Food Category Doing Any Work Test 3 at the production level: the presumed mechanism (n-3/CLA) is carried by feed, and “organic” is the wrong exposure to steer by. (inferred from Givens & Lovegrove, 2016; Średnicka-Tober et al., 2016)
Sub-question 4 — hard outcomes: one cohort, and it exists here to DEMONSTRATE the confound
Baudry (NutriNet-Sante, 68,946 French adults, 1340 incident cancers) reports “High organic food scores were inversely associated with the overall risk of cancer (hazard ratio for quartile 4 vs quartile 1, 0.75; 95% CI, 0.63-0.88; P for trend = .001; absolute risk reduction, 0.6%; hazard ratio for a 5-point increase, 0.92; 95% CI, 0.88-0.96).” (Baudry et al., 2018)
This is ingested to demonstrate the healthy-user confound, NOT to assert organic prevents cancer — the observed-healthy-population-is-not-evidence-for-a-component trap, and the paper’s own Table 1 makes it vivid. Higher organic score tracked higher income, postsecondary education, more physical activity, and former (not current) smoking; the diet-quality score rose (mPNNS-GS 7.41 -> 8.19 across Q1->Q4), BMI fell (24.46 -> 22.92), fibre rose (17.9 -> 22.6 g/d) and processed meat (23.7 -> 15.1 g/d) and red meat (48.7 -> 31.4 g/d) fell. (Baudry et al., 2018) Organic buyers differ from non-buyers on nearly every established cancer risk factor at once — the exposure is a marker of overall diet quality and socioeconomic position, both of which predict the outcome independently. The authors say so: “Multiple studies have reported a strong positive association between regular organic food consumption and healthy dietary habits and other lifestyles. Hence, these factors should be carefully accounted for in etiological studies in this research field.” (Baudry et al., 2018)
Two facts that keep the confound reading honest (symmetric standards).
- The model adjusted for a large confounder set (income, education, physical activity, smoking, alcohol, BMI, energy, diet-quality score, fibre, processed and red meat), and the association survived. So this is not an unadjusted artifact — it is a residual-confounding concern, which is weaker but not nothing. Self-reported diet carries error large enough that adjustment is incomplete by construction -> Measurement Error in Dietary Assessment.
- The paper’s own internal check points away from organic-per-se: “Combining both a high-quality diet and a high frequency of organic food consumption did not seem to be associated with a reduced risk of overall cancer compared with a low-quality diet and a low frequency of organic food consumption.” (Baudry et al., 2018) The signal was carried where diet quality was low, not stacked on top of an already-good diet — consistent with diet quality, not the organic label, doing the work. Site-specific results were narrow (postmenopausal breast cancer, non-Hodgkin lymphoma and lymphomas; no association at other sites).
The pooled SR/MA firms the null — Baudry’s protective signal does not survive pooling (type-F)
Theodoridis (SR + random-effects MA, three observational cohorts, 733,954 individuals, PROSPERO-registered) found «There was no difference between the two interventions regarding overall cancer (HR = 0.93, 95% CI: 0.78–1.12), breast cancer (HR = 1.01, 95% CI: 0.81–1.26), colorectal cancer (HR = 1.01, 95% CI: 0.93–1.10), and non-Hodgkin lymphoma risks (HR = 0.70, 95% CI: 0.17– 2.94).» (Theodoridis et al., 2025) The overall pool carried I2=84% and a prediction interval of 0.10-8.57 — a null point estimate over enormous between-study heterogeneity.
The MA CONTAINS Baudry, so this is NOT an independent confirmation — type-F, not type-E. Its three constituents’ own overall-cancer HRs (Table 3): Andersen 0.99 (0.91-1.08), Baudry 0.76 (0.64-0.90), Bradbury 1.03 (1.00-1.06). (Theodoridis et al., 2025) Baudry’s lone protective HR — the single cohort this page already held — is diluted to a null pool by Andersen (null) and Bradbury (the 623k Million Women Study, highest weight, marginally above 1). The protective signal was one cohort’s, not the literature’s.
Parameter table (same-quantity check — BLOCKING on the cross-source claim):
| Parameter | Incumbent — Baudry 2018 (quoted + locus) | Theodoridis 2025 (quoted + locus) | Same quantity? |
|---|---|---|---|
| Baudry’s own overall-cancer HR | ”hazard ratio for quartile 4 vs quartile 1, 0.75; 95% CI, 0.63-0.88” (Baudry chunk 01) | «Overall cancer 0.99 (0.91–1.08) 0.76 (0.64–0.90) 1.03 (1.00–1.06)» Table 3, high-vs-low (Theodoridis chunk 01) | YES — same cohort, same overall-cancer HR (Q4-v-Q1 = high-v-low; 0.75/0.76 rounding). Theodoridis REPRODUCES Baudry; it does not independently confirm it. |
| Overall-cancer effect object | single-cohort quartile-extreme HR 0.75 (0.63-0.88) | 3-cohort random-effects POOL, HR 0.93 (0.78-1.12), I2=84% (Theodoridis chunk 01) | NO — one cohort’s extreme-quartile contrast vs a pool across three cohorts. Different objects; the pool is the new evidence, and it is null. |
| Exposure definition | 3x 24h recalls -> mPNNS-GS organic quartiles | heterogeneous across the three (FFQ / 24h recall / single yes-no question); «substantial heterogeneity in the definition and labeling of organic food products» (Theodoridis chunk 01) | NO — one operationalization vs a mix; a driver of the I2=84%. |
A null in designs biased toward benefit is stronger evidence against organic-prevents-cancer, not weaker. The paper names the residual-bias direction: healthy-user and self-selection push toward a spurious protective signal — «Because of this selection bias, the genuine connection between organic food consumption frequency and cancer risk may be overestimated.» (Theodoridis et al., 2025) So an observational pool that lands null, when its confounding should have manufactured a protective HR, withholds the benefit the bias predicts — at least as consistent with no effect as with a small real one. But it adds no causal-identification route: it pools the same confounded cohorts (GRADE certainty very low for every endpoint; Baudry and Bradbury high risk of bias on exposure measurement), so it firms no benefit without resolving the confound the incumbent frame rests on. (inferred from Theodoridis et al., 2025)
Site-specific stays INSUFFICIENT, not null, where the cohorts disagree. Pooled non-Hodgkin lymphoma HR 0.70 (0.17-2.94), I2=90% — the CI spans a ~17x range because the constituents point opposite ways (Andersen 1.97 [1.28-3.04, elevated and significant], Baudry 0.14, Bradbury 0.79 [protective]); breast is likewise discordant (Baudry postmenopausal 0.66 [0.45-0.96] vs Bradbury 1.09 [1.03-1.15, elevated]). Colorectal is the one coherent endpoint (pooled 1.01 [0.93-1.10], I2=0%). (Theodoridis et al., 2025) Read the NHL point estimate as insufficient evidence, never no effect — the expectancy test fails on a CI that wide over three discordant cohorts.
The design cannot cleanly attribute, and no RCT exists (you cannot blind or randomise a lifetime of organic eating — the streetlight problem). Expected answer for this sub-question: cannot cleanly attribute, exactly as the spec predicted — and now with the pooled hard-outcome estimate showing null, not merely one confounded protective cohort.
The net health verdict
- Residues: genuinely lower on organic (incidence ~4x lower in crops; contamination risk difference 30%), but conventional residues mostly sit below limits and no evidence ties the difference to a patient-important outcome. A surrogate, not a demonstrated benefit.
- Nutrients: small, mixed-direction differences (more polyphenols/antioxidants; less iodine/selenium in organic milk); the weighting SR found no clinically significant nutritional superiority.
- Animal-product fatty acids: a real +40-70% relative lift in n-3/CLA — but it is a feed (pasture)
effect the organic label only partly captures. The lift is reported by the sources as a relative
product-level difference; whether it is large in absolute terms against total-diet n-3 is a
G-gap — no held source quantifies dairy/meat’s share of total n-3 intake, so the absolute magnitude is unestablished here rather than known to be small. (inferred from Givens & Lovegrove, 2016; Średnicka-Tober et al., 2016) - Hard outcomes: the pooled observational evidence is null — a SR/MA of three cohorts (Theodoridis 2025) gives overall-cancer HR 0.93 (0.78-1.12), breast 1.01 (0.81-1.26), colorectal 1.01 (0.93-1.10), with Baudry’s lone protective signal (0.76) diluted by the two larger cohorts (Andersen 0.99, Bradbury 1.03). GRADE certainty very low, still confounded, no RCT; non-Hodgkin lymphoma (0.70, 0.17-2.94) is insufficient, not a null. Firms no established causal cancer benefit with pooled hard-outcome evidence — without adding a causal-identification route.
Ranked (Layer 1), “organic” is a small, low-certainty lever at best — dominated for anyone by the big rocks (smoking, adiposity, activity, overall diet pattern) and, on the nutrient axis, by the grass-vs-grain choice that is a better-specified exposure than the certificate. Attention is an anti-signal applies: the topic is heavily discussed and mostly small-effect.
Value types
- type-B — decomposes the “organic” label into separable exposures (residues / nutrients / feed), each with its own effect size and evidence state.
- type-A — the label-vs-exposure decoupling: pasture, not certification, drives the animal-product difference. The induction is composition-attribution (the two same-group Srednicka-Tober MAs name feed as the driver) + certification-rule reasoning (organic mandates outdoor access, not a forage diet) + the Is the Food Category Doing Any Work Test-3 lens — NOT two independent supports (milk and meat share an author group).
- type-F — Smith-Spangler’s weighting SR bounds Baranski’s compositional differences (detectable != clinically significant); Baudry bounds the outcome claim (confounded); Theodoridis’s pooled SR/MA firms the outcome null — overall-cancer HR 0.93 (0.78-1.12) across three cohorts, showing Baudry’s single-cohort protective HR (0.76) does not survive pooling. NOT type-E: the MA CONTAINS Baudry (pooled-constituent double-count firewall — a study inside the pool is not an independent route), so it refines/firms rather than independently confirms.
- type-G — no RCT on hard outcomes; residue-to-outcome transmission unevidenced at real-world levels; a grass-fed-specific (vs organic-certified) outcome comparison is unheld.
Limits and gaps
- Open loop: this grades coherence and source-fidelity, never validity. It appraises, does not prescribe or budget.
- Guidance-null: this agrees with the mainstream position — buy organic if you value the non-health axes, but do not expect a documented population-level health benefit; the measured differences are real but their health consequence is unproven.
- The Baranski/Srednicka-Tober meta-analyses share an author group (Leifert et al.); Smith-Spangler is an independent group reaching the more conservative nutritional verdict — the two are not independent confirmations of each other, and where they diverge (compositional difference vs clinical significance) the divergence is the finding, not a vote to average.
G (needs aggregation)— partially cashed for cancer (2026-08-20): Theodoridis (2025) supplies the pooled organic->cancer magnitude the wiki could not itself compute (overall HR 0.93 [0.78-1.12], null, GRADE very low). The gap persists for non-cancer hard outcomes (CVD, all-cause mortality) and for causal identification — pooling confounded observational cohorts computes a magnitude, not an unconfounded effect; the trials still do not exist.AWAITSa grass-fed-vs-grain-fed animal-product outcome source that separates feed from certification; and any residue-level human-outcome source that could move residues from surrogate to outcome.
Appraising this observational evidence — the instrument [2026-07-31]
The cohort signal (organic → lower cancer, Baudry 2018) is observational; ROBINS-I (Risk of Bias Assessment Tools) is the appraisal tool, with domain 1 (confounding — organic buyers differ on many exposures) and domain 2 (selection) the likely caps. Flagged as a re-appraisal candidate there; not re-graded here.
”Free range” / “outdoor access” is a weak label - and that sharpens the exposure-vs-certification point (deliverable-critique, 2026-08-01)
A labelling-integrity note on the WELFARE / consumer axis (named, not priced - the wiki adjudicates only the health axis): “free range” and “outdoor access” are minimal regulatory standards (a nominal door or brief access), not a guarantee of genuine pasture - the fig-leaf gap between the term and consumer expectation is real. The health-relevant consequence reinforces this page’s frame: any compositional benefit that would actually track pasture (e.g. the grass-fed fat / omega-3 profile in animal products) is driven by the underlying exposure, which the label captures only loosely - so the certification is an even weaker proxy for the thing doing the work -> Is the Food Category Doing Any Work. Welfare adequacy itself is out of the health axis and not adjudicated here.