This is a general, population-level appraisal of what the evidence says about soy for anyone the studies speak to. It is not advice for one person; whether and which soy you should eat depends on your stratum — sex, menopausal status, baseline LDL, iodine status, and the realistic alternative — and belongs at the end, not here.
Stop treating “soy” as one thing — that is the single most useful move. The isolated protein, the whole bean, the fermented paste, and the soy-sauce condiment are different exposures carrying different evidence — each question below was tested on a different one -> Is the Food Category Doing Any Work. When a benefit or a risk is pinned on “soy” without saying which soy, that is the tell it has outrun the evidence.
The bottom line, per outcome
Read the certainty claim-by-claim, not as one verdict on “soy”. Five gold-tier meta-analyses anchor the page — one each for the hormone, cancer and cognition outcomes, and two on LDL (the newer revising the older’s magnitude) — each with its own grade:
| Outcome | Direction | Certainty |
|---|---|---|
| Isoflavone -> male testosterone | null (no effect) | high (TT) |
| Isoflavone -> free testosterone / estradiol | consistent-with-null, less bounded | moderate |
| Soy isoflavone -> breast cancer | neutral-to-protective, never raised | low-moderate |
| Isolated soy protein -> LDL | lowering (~3%), baseline-dependent | moderate |
| Isoflavone -> cognition (memory) | small benefit on test scores | low (surrogate) |
| Thyroid (iodine-conditional) | unknown | insufficient (gap) |
| Natto / K2 (MK-7) -> bone / vascular | unknown | insufficient (gap) |
| Soy in ER+ breast-cancer survivors | unknown | insufficient (gap) |
- Male reproductive hormones are unmoved. In pooled controlled trials, neither soy nor its isoflavones shift testosterone — a well-powered null, firmest for total testosterone.
- Breast-cancer risk is neutral-to-protective, never raised in any stratum — but the protective signal is design-confounded and concentrates in Asian, lifelong-intake populations.
- What soy actually does is modest and form-specific: isolated soy protein lowers LDL cholesterol, mostly in people who already have high cholesterol, and the honest modern magnitude is a few percent.
Two facts bind the whole page. Each outcome rests on effectively one meta-analysis — the two on LDL share trials and one cites the other, so they are a revised single estimate, not independent confirmations — which is why overall confidence is medium despite five gold sources. And the headline effects are all surrogates — hormone levels, LDL, and cognitive-test scores, not events — so their transmission to what a person feels is a further, separate claim.
How good is the evidence — read this first
Soy is one food, but its evidence is four separate literatures, and they are not equally strong. Read the certainty per outcome, not for “soy” as a whole. Each outcome rests on effectively one meta-analysis (the two on LDL share trials and one cites the other), so overall confidence is medium.
- Hormones — pooled clinical trials, the strongest design here and the firmest finding. Reed 2021 gathered 41 controlled studies; certainty is high for the total-testosterone null specifically — two model specifications both land on no effect (treatment-vs-control I2 72%; change-over-time I2 29%), a 2010 predecessor agreed, and there is no funnel-plot asymmetry — and softer for the estrogen arms. (Reed et al., 2021)
- Breast cancer — entirely observational, and confounded by study design: the favourable pooled figure is carried by the weaker retrospective studies. Certainty is low-to-moderate. (Chen et al., 2014)
- LDL cholesterol — controlled trials, the direction well supported across two meta-analyses. The landmark 1995 analysis is dated and partly industry-funded; the newer Blanco Mejia 2019 pools the 46 trials the FDA reassessed and revises the intrinsic magnitude down to ~3%. Certainty is moderate — the two share trials and one cites the other, so this is a revised single estimate, not two independent confirmations. (Anderson et al., 1995) (Blanco Mejia et al., 2019)
- Cognition — pooled short trials, but on a test-score surrogate, not on dementia. Cui 2020 pooled 16 RCTs of isoflavone supplements; the composite improves and the effect is carried by memory. Certainty is low: the endpoint is a cognitive test score over weeks-to-2-years, and whether that transmits to the outcome anyone cares about — dementia — is untested. (Cui et al., 2019)
Thyroid (goitrogen) and antinutrient (phytate) fears were not searched. They are named gaps, not outcomes scored safe. For how these certainty grades are assigned, see Rating Certainty of Evidence and Upgrading Observational Evidence.
Soy does not shift male reproductive hormones
Reed 2021 meta-analyzed 41 controlled studies of whether soy or isoflavone intake shifts male reproductive hormones — isoflavones being the plant estrogen-like compounds in soy, chiefly genistein and daidzein. It covered total testosterone (TT), free testosterone (FT), estradiol (E2), estrone (E1), and sex hormone-binding globulin (SHBG). TT and FT were measured in 1753 and 752 men. (Reed et al., 2021)
The estimates cluster on zero:
- Total testosterone: standardized mean difference (SMD) -0.06 (95% CI -0.29 to 0.17; P=0.59). Free testosterone: 0.01 (-0.33 to 0.32). SHBG: -0.03 (-0.45 to 0.38). Estradiol and estrone: no significant effect on any model. There was no dose or duration effect, and none of the four studies running 12 months or longer found a testosterone effect. (Reed et al., 2021)
This is a null, not a gap — most firmly for testosterone. Apply the expectancy test: if soy lowered testosterone, a 41-study analysis powered to detect it in 1753 men would have. It did not, across every model, dose, and duration. That is evidence of no meaningful effect on the very endpoint the feminization worry is about — not “unproven”.
The certainty splits by hormone, and the TT “high” grade is not a homogeneity claim: the primary treatment-vs-control estimate carries substantial heterogeneity (I2 72%). It earns “high” because two independent model specifications both land on the null — change-over-time 0.09 (95% CI -0.02 to 0.20, I2 29%) alongside the -0.06 above — a 2010 predecessor agreed, and the interval excludes a meaningful lowering. The FT estimate is only moderate — its interval (-0.33 to 0.32) does not exclude a small effect in either direction, and its heterogeneity is higher still (I2 76%). Read the author’s summary that «neither soy protein nor isoflavone exposure affects TT, FT, E2 or E1 levels in men» as carried by the TT arm. (Reed et al., 2021)
The mechanism predicts this null. Isoflavones preferentially bind and activate ERbeta, one of the two estrogen-receptor subtypes, whereas estrogen binds ERalpha and ERbeta roughly equally. Because those receptors sit in different tissues and can act in opposite directions, isoflavones behave tissue-selectively and are classed as selective estrogen-receptor modulators (SERMs) — the same drug category as tamoxifen, not estrogen itself. Mechanism carries directional weight, but the trials, not the mechanism, settle it. (Reed et al., 2021)
Two boundary conditions. The case reports of gynecomastia and hypogonadism involved intakes near 360 mg/d isoflavones — roughly nine times a high native-Japanese dietary intake, and not reachable from food. And symmetric standards apply to the source: the corresponding author works in soy-industry-adjacent consultancy, so the paper is flagged rather than exempt. The null survives the flag — it holds across both model specifications, a matching 2010 result, no funnel-plot asymmetry, and it runs toward what the SERM mechanism predicts. Hormone levels are themselves a surrogate for the endpoints of fertility and feminization -> Surrogate Outcomes. (Reed et al., 2021)
Breast cancer: neutral-to-protective, never raised — but design-confounded
Chen 2014 pooled 35 epidemiological studies of high-versus-low isoflavone intake, stratified by menopausal status and by country:
| Stratum | Premenopausal OR (95% CI) | Postmenopausal OR (95% CI) |
|---|---|---|
| Summary | 0.74 (0.64–0.85) | 0.75 (0.63–0.86) |
| Asian countries | 0.59 (0.48–0.69) | 0.59 (0.44–0.74) |
| Western countries | 0.90 (0.77–1.04), NS | 0.92 (0.83–1.00), marginal |
Read the design before the headline number. The pooled ~25% reduction is confounded by study design. The stronger prospective cohort and nested case-control studies showed no significant protection — premenopausal cohorts OR 0.94 (0.74–1.14), postmenopausal cohorts 0.86 (0.73–1.00), the upper bound touching the null. The inverse signal came from retrospective case-control studies, whose recall and selection biases systematically inflate a diet-disease association. So the favourable figure tracks the weaker designs. This is a weaker-design-inflates-the-favourable-arm pattern — a cousin of The U-Shaped Association Artifact, though the mechanism here is recall and selection bias rather than unequal reporting precision. (Chen et al., 2014)
The direction is inverse or neutral — never a raised risk in any stratum. But the finding does not transport well. The protection concentrates in Asian populations (OR ~0.59); the Western signal vanished once stratified by design, and Chen attributes the Asian result to early-life and higher lifelong intake rather than to adult supplementation. The certainty is low-to-moderate: observational, heterogeneous across subgroups, with publication bias detected on Egger’s test, and the strongest designs are null. The defensible claim is “not a risk, and plausibly protective under lifelong high intake” — not “soy prevents breast cancer”. (Chen et al., 2014)
Isolated soy protein modestly lowers LDL
Anderson 1995 is the landmark meta-analysis: 38 controlled trials substituting soy protein (mean 47 g/d) for animal protein. Net of the control diet, it found total cholesterol -23.2 mg/dL (-9.3%), LDL -21.7 mg/dL (-12.9%), triglycerides -13.3 mg/dL (-10.5%), and a non-significant HDL change (+2.4%). Across the soy-diet-alone model, 25, 50, and 75 g/d mapped to cholesterol falls of 8.9, 17.4, and 26.3 mg/dL — roughly monotone over the studied range, with no knee located. (Anderson et al., 1995)
The effect is baseline-dependent, and this is the decisive nuance. Initial cholesterol accounted for about 77% of the between-study variance. In people with normal cholesterol (under 200 mg/dL) the change was a non-significant -3.3%; in severe hypercholesterolemia (over 335 mg/dL) it reached -19.6%. This is a route-(a) baseline-risk effect: the lever is real mainly for those who already have high cholesterol. LDL is a surrogate for atherosclerotic events, and its causal transmission runs through LDL ApoB and Cumulative Exposure -> Surrogate Outcomes, not assumed here. (Anderson et al., 1995)
Two caveats bound this figure. First, it is the isolated form only: Anderson excluded whole soybeans, so the evidence is for isolated or textured soy protein replacing animal protein — not for tofu, edamame, or soy sauce -> Is the Food Category Doing Any Work. Second, symmetric standards: the 1995 trial was supported in part by a soy-protein manufacturer, with the lead author on its advisory group, so a favourable industry-funded result earns more scrutiny, not less. (Anderson et al., 1995)
A newer, larger meta-analysis now supersedes the 1995 magnitude as the best current estimate. Blanco Mejia 2019 pooled the 46 trials the FDA used to reassess soy’s heart-health claim: «Soy protein intake in 50 trial comparisons demonstrated a significant reduction in LDL cholesterol (MD: −4.76 mg/dL; 95% CI: −6.71, −2.80 mg/dL), equivalent to −3.2% (95% CI: −4.5, −1.9%; P < 0.0001)» at a median 25 g/d, with no dose-response detected. (Blanco Mejia et al., 2019)
The −3.2% sits far below Anderson’s −12.9%, but the two are not the same quantity: Anderson dosed higher (47 vs 25 g/d) in higher-baseline populations, so dose and baseline explain most of the gap — Anderson’s own normal-cholesterol arm (a non-significant −3.3%) already reconciles with it. Hold the direction; the honest modern magnitude is a small intrinsic ~3% drop on a surrogate, not the 1995 ~13%. Symmetric standards cut the other way this time: a co-author sits with the Soy Nutrition Institute, yet the finding revises the effect down, against the sponsor’s interest. (Anderson et al., 1995) (Blanco Mejia et al., 2019)
On memory, the benefit is small and sits on a test score, not on dementia
Cui 2020 pooled 16 randomized trials of isoflavone supplements — roughly 80 to 160 mg/day, six weeks to two years, mostly postmenopausal women — and found the composite cognitive score improved: SMD 0.19 (95% CI 0.07 to 0.32, I2 18.1%). The composite is carried by memory (SMD 0.15, 95% CI 0.03 to 0.26, I2 0.0%); the other domains, tested singly, do not reach significance. So the honest reading is narrow: a small average gain, concentrated in one domain. (Cui et al., 2019)
The endpoint is the limit. These trials measured a cognitive test score over two years or less — a surrogate. None counted a case of dementia. A test-score gain of this size need not transmit to the outcome a person actually fears, and no trial here shows that it does. This is the same surrogate-vs-event gap that sinks certainty on LDL, applied to cognition: hold the direction, discount the magnitude, and do not read a memory-test point as a dementia result -> Surrogate Outcomes.
The reverse worry — that soy harms the brain — does not survive its own evidence. It traces to one observational cohort (Hawaiian men, midlife tofu intake); the trial pool and later observational work run the other way. Cui states it plainly: «current evidence does not support the proposition that ISFs negatively affect cognition». (Cui et al., 2019)
Where this lands in the ranking: low. A small surrogate benefit from a supplement is a minor lever, not a big rock. For someone already lean, active, and eating a reasonable diet, the dementia levers that carry real evidence are elsewhere — the modifiable risk factors, not an isoflavone pill -> Dementia Prevention and Modifiable Risk Factors.
The forms are not interchangeable — specify the exposure
“Soy” names several exposures with different active fractions and different evidence. This is a terminological disambiguation with decision consequences, and it is Is the Food Category Doing Any Work instantiated: the category “soy” is too coarse to carry one verdict, and the better-designed evidence tracks the better-specified exposure.
- Isolated or textured soy protein — the LDL-lowering form, and Anderson’s exposure. A high-DIAAS complete plant protein -> Protein Quality and the DIAAS Score.
- Whole or minimally processed (tofu, edamame, soy milk) — the isoflavone-plus-protein workhorse, the form carrying the hormone and cancer evidence.
- Fermented (miso, tempeh, natto) — lower phytate, because fermentation defuses the antinutrient concern -> Antinutrients in Plant Foods. Natto is a notably rich source of vitamin K2 (MK-7, the long-chain menaquinone form); whether that MK-7 moves a bone or vascular outcome is a gap. See Fermented Foods and Health for the live-culture question; the soy matrix stays here.
- Soy sauce / tamari — a condiment and salt vehicle with a negligible isoflavone dose. Do not credit it with soy’s effects.
The synthesis across the three meta-analyses is that soy’s evidence is unusually form-dependent: each outcome was tested with a different soy exposure, so the composite answer is not “soy does X” but a form-indexed one — the isolated protein moves LDL, the isoflavone-bearing foods leave male hormones unmoved and sit neutral-to-protective on breast cancer, and the condiment does nothing.
Antinutrients and thyroid: named gaps, not findings
Two recurring soy concerns sit outside the evidence appraised above. Naming them as gaps is a result in itself: it keeps “unstudied” apart from “shown safe”, and neither is asserted nor dismissed here.
- Thyroid (goitrogen). No systematic review conditioned on iodine status is held, so this is a gap. The prior to test — not a finding — is that soy is a mild goitrogen mainly under iodine deficiency; in euthyroid, iodine-replete people it likely causes no clinically important harm. The one practical issue is that soy may raise levothyroxine dose needs, an absorption-and-timing effect rather than a thyroid injury.
- Antinutrients (phytate). Phytate is the mineral-binding compound in many plant seeds. Its handling lives at Antinutrients in Plant Foods: reduced by fermentation and soaking, and a concern only at the margin of mineral status, not a general harm. This page extracts no soy-specific outcome evidence.
Putting it together
- Male hormones are unmoved and breast-cancer risk is not raised. Soy does not lower testosterone — a well-powered null, firmest for total testosterone — and it does not raise breast-cancer risk in any stratum. On cancer it is neutral-to-protective, with the protection an Asian, lifelong-intake pattern, not a dose a Western adult can add later.
- The real effect is small and specific. Isolated soy protein modestly lowers LDL, mostly in people who already have high cholesterol, and the honest modern magnitude is a few percent — a minor lever, not a treatment, and on a surrogate rather than an event.
- Match the claim to the form. The isolated protein moves LDL; the whole-bean foods carry the hormone and cancer evidence; the fermented pastes shed the phytate concern; the soy-sauce condiment does nothing. A benefit or concern attached to “soy” with no form named has outrun its evidence -> Is the Food Category Doing Any Work.
- Judge it against what it replaces. As a complete plant protein displacing red or processed meat, soy’s case rests on the substitution as much as on anything intrinsic; your sex, menopausal status, baseline LDL, iodine status, and realistic alternative set the individual weighting, at layer 3.
What this appraisal cannot yet answer
The honest edge of this page. Each item is unstudied-here, not scored safe.
- Soy in ER-positive breast-cancer survivors. The prominent clinical worry that dietary isoflavones might stimulate estrogen-receptor-positive tumour growth or blunt tamoxifen is not adjudicated here — Chen covers incidence, not survival. Gap.
- Thyroid, iodine-conditional. No soy-thyroid systematic review conditioned on iodine status is held. Gap.
- Natto and vitamin K2 (MK-7) -> bone and vascular calcification. Whether natto’s MK-7 moves a bone or cardiovascular outcome is unresolved; cross-links Fermented Foods and Health. Gap.
- Prostate cancer, menopausal symptoms, bone density. Named in scope, no source held. Gap.
- Absolute breast-cancer risk reduction. Chen reports relative odds ratios only; the absolute benefit
depends on baseline incidence in the stratum, which the analysis does not supply. This needs pooling the
wiki cannot perform —
G (needs aggregation). - Soy -> hard cardiovascular events. The LDL effect is a surrogate; no trial measures whether isolated soy protein lowers heart-attack or mortality risk, so the event benefit is inferred through LDL ApoB and Cumulative Exposure, not measured. Gap.
Mandatory caveats
- Open loop. This wiki grades internal coherence and fidelity to its sources — never whether a recommendation actually improves outcomes in the world. A clean appraisal is not a validated result.
- Appraise, do not prescribe. This is a general, population-level appraisal, not medical advice; selecting, dosing, and screening for contraindications (thyroid-medication timing, survivorship) are prescriber acts requiring information this document does not hold.
- A general appraisal, applied per person. Sex, menopausal status, iodine and mineral status, baseline LDL and cardiovascular risk, and your realistic alternative decide the individual weighting.
- Health axis only. This cut weighs soy on hormonal, cancer, and cardiometabolic outcomes. Soy also carries environmental (low-footprint protein, but entangled with animal-feed and deforestation) and economic loads; that those trade-offs exist is noted here, never priced against the health finding.
- The target is not agreement or divergence with any guideline. It is showing how good the soy evidence is — clinical-trial-firm on male hormones, observational and design-confounded on cancer, dated and industry-funded on LDL — and letting the forms carry different verdicts.
Evidence box
Question ’What does the evidence show about soy products” effect on each patient-important outcome — direction, magnitude, for whom, how certain — do the forms (whole, fermented, isolated protein) differ, and what do the hormonal, breast-cancer, thyroid and antinutrient effects amount to?‘ Evidence included 5 sources — 5 gold Overall certainty Medium (see Rating Certainty of Evidence) Source-selection note All sources are gold or high tier. Last updated 2026-09-05 · Independently reviewed: No · Full edit history