AMD is the second age-related eye disease this domain opens on (paired with Cataract and Modifiable Exposures). Unlike cataract, late AMD is not routinely reversible — so the patient-important outcome is irreversible central-vision loss, function, and independence, weighting a preventive lever more heavily if one exists. Three arms are now assessed and they part ways: physical activity is insufficient-evidence (no clear lever); the Mediterranean-diet pattern shows a protective association with AMD progression on the stronger observational designs; and the cardiometabolic/vascular risk factors (smoking, hypertension, CVD, diabetes) show significant prognostic associations with AMD occurrence, while BMI, cerebrovascular disease and blood lipids do not. All three rest on observational, gold-by-design-only-or-weaker evidence that cannot carry a causal reading.
This page is built with per-exposure sections so each exposure sits as a sibling arm; the diet arm (Mediterranean diet -> AMD) and the vascular-risk-factor arm (Babaker 2025) have now landed beside the physical-activity arm.
Physical activity — insufficient evidence, not a demonstrated null
Aune 2026 (gold-tier SR-MA of prospective cohorts) is the sole source. AMD is the weaker arm of the same paper whose cataract arm reaches WCRF probable: for AMD the authors reach «limited, and no conclusion could be drawn» — the insufficient-evidence state (The Insufficient-Evidence Statement), not no meaningful effect. (Aune et al., 2026)
- High vs low PA: RR 0.92 (95% CI 0.84–1.01), I2 = 60%, n = 14 cohorts — the CI crosses 1.00, so the association is not statistically significant. A point estimate near the cataract arm’s, but the interval admits no effect. (Aune et al., 2026)
- Dose-response: RR 0.92 per 20 MET-hours/week (0.74–1.13), I2 = 48%, n = 3 — again crossing 1; no nonlinearity (pnonlinearity = 0.34). Only 3 cohorts had usable leisure-time dose data. (Aune et al., 2026)
- Authors’ own summary: «This meta-analysis provides further support for an inverse association between physical activity and risk of cataract, but an association with age-related macular degeneration was less evident.» (Aune et al., 2026)
Why this is insufficient, not null — apply the expectancy test. With a wide CI (0.84–1.01), substantial heterogeneity (I2 = 60%), and a dose-response resting on 3 cohorts, the data cannot distinguish a real ~8% reduction from no effect; a point estimate below 1.00 with an interval that barely crosses it is under-powered, not exonerating. The same cohort-design measurement error that (in the cataract arm) argues attenuation-toward-the-null applies here too — so a true small effect could be masked. Hold AMD-PA as candidate/gap, neither asserted nor dismissed.
Mediterranean diet — protective for progression, insufficient for development
Marques-Couto 2025 is a PROSPERO-registered SR-MA of observational studies only — 8 studies, no RCT — of Mediterranean-diet (MD) adherence versus AMD development and progression. It pools separately by study design (different effect measures, different exposure to reverse causation), and the three pools split cleanly. (Marques-Couto et al., 2025)
- Prospective cohort (3 studies — AREDS / Rotterdam / Alienor), the strongest design: pooled «hazard ratio (HR) of 0.77 (95% CI: 0.67–0.88; p < 0.0001; I2 = 0%), corresponding to a 23% reduction in AMD progression risk». Significant, zero heterogeneity, temporally ordered (diet measured before progression). (Marques-Couto et al., 2025)
- Case–control (3 Portuguese Coimbra-Eye-Study cohorts): pooled «OR of 0.66 (95% CI: 0.54–0.81; p < 0.0001; I2 = 41.5%), corresponding to a 34% reduction in AMD risk». Significant; the heterogeneity traces to one study (Barreto 2023, the strongest single effect, OR 0.41) — a leave-one-out «sensitivity analysis showed that removing Barreto et al. (2023)» eliminated the heterogeneity (the sentence — and its residual-I2 figure — is cut at the chunk boundary, so “eliminated” is the source’s word and no specific I2=0% is quoted). (Marques-Couto et al., 2025)
- Cross-sectional (2 studies), NULL: pooled «OR of 0.96 (95% CI: 0.83–1.11; p = 0.6243; I2 = 0%), indicating no significant association». (Marques-Couto et al., 2025) But under inverse-variance pooling this null is effectively a single-study result: the large cross-sectional arm (Hogg / EUREYE, 4753 participants) dominates the weight against a much smaller second study, so the pool is not an independent synthesis.
The design gradient — and why the null is not the headline. The two designs able to order diet before disease (cohort, and partly case–control) both show protection; the design that cannot (cross-sectional) shows the null — and the authors read that null as a temporality artifact: «Cross-sectional and case–control studies assess diet and disease progression simultaneously, making it difficult to determine whether the MD influences AMD progression or whether AMD progression influences adherence to the diet». (Marques-Couto et al., 2025)
Note the direction of the gradient: a protective signal strongest in the temporally-ordered design and absent in the simultaneous one is not the classic reverse-causation signature (reverse causation would inflate the simultaneous designs, not the cohorts) — so here reverse causation is a weaker worry than healthy-user confounding, which the cohort adjustments (age, sex, smoking, BMI, education, energy intake) attenuate but cannot remove at root.
Development vs progression, and the confidence discount. The protective pools are for AMD progression; for initial development the review is explicitly weaker — it «showed a potential protective effect of the MD against AMD, particularly in terms of slowing its progression … although the evidence for preventing its initial development is less conclusive». (Marques-Couto et al., 2025)
Confidence is held low for the whole arm — gold-by-design-only: the venue is MDPI Nutrients; the evidence is observational with no trial; n = 8 studies; the exposure is self-reported MD adherence (an FFQ, usually measured once), so dietary measurement error (Measurement Error in Dietary Assessment) is the binding constraint and healthy-user confounding is uncontrolled at root. A significant progression OR/HR here is a consistent association, not a demonstrated causal effect — do not read the 23–34% figures as what MD would deliver, nor blend the prevalence-odds (case–control) and progression-hazard (cohort) reductions into one number.
Cardiometabolic and vascular risk factors — significant prognostic associations, no causal claim
Babaker 2025 is a PRISMA/PROSPERO-registered (CRD42023479050) updated SR-MA of observational studies only — «Eighteen of the 2640 identified studies met the inclusion criteria» (8 cohort, 7 cross-sectional, 3 case–control), «a total of 44,440 patients» across studies, quality graded on the Newcastle-Ottawa scale (studies scoring low were excluded; the kept set is a mix of high and moderate). The outcome is late AMD (neovascular + geographic atrophy). ORs were pooled by random-effects generic inverse variance. (Babaker et al., 2025)
NB the pooled n: 2640 is the number of screened records, not participants — the 18 included studies carry ~44,440 participants (the source’s «diagnosed with AMD» phrasing is loose: Table 2’s sample sizes are total cohort participants, of which AMD cases are a fraction). So this is not a tiny-sample MA; the binding limits are the small number of studies per factor and the observational designs, not sample size.
Significant associations (all relative ORs; the source gives no absolute AMD risk by stratum, so absolute benefit of removing a factor is not stated): (Babaker et al., 2025)
- Smoking: OR 1.86 (95% CI 1.33–2.6; P = .0003), I2 = 81% (high heterogeneity; a leave-one-out «heterogeneity in the smoking outcome was resolved by the removal of studies by Park et al and Hogg et al»). The one factor here with a strong external prior and a specific mechanism, and a Layer-1 big rock in its own right. (Babaker et al., 2025)
- Hypertension: OR 1.24 (95% CI 1.09–1.4; P = .0007), I2 = 0%. (Babaker et al., 2025)
- Cardiovascular disease: OR 1.44 (95% CI 1.11–1.87; P = .006), I2 = 67%. (Babaker et al., 2025)
- Diabetes: OR 1.44 (95% CI 1.3–1.6; P < .00001), I2 = 0%. (Babaker et al., 2025)
For context (non-modifiable, not levers): advanced age OR 1.11 (1.06–1.15) and male gender OR 1.63 (1.13–2.35) were also significant. (Babaker et al., 2025)
Not significantly associated — the insufficient / no-demonstrated-association state, not a proven null: «There was no statistically significant association between AMD and other risk factors, such as BMI, cerebrovascular disorders, cholesterol, or triglyceride levels». On a base of few studies per factor these NS arms are under-powered, not exonerating; the authors read the lipid null as unsettled («our findings regarding the nonsignificant association of cholesterol and triglycerides with AMD add to the ongoing debate about the role of lipids in AMD pathogenesis»). (Babaker et al., 2025)
These are route-(a) baseline-risk / prognostic associations, NOT intervention effects and NOT effect-modification. They say who is more likely to develop AMD, not that removing the factor lowers AMD risk by the OR shown. No trial, no MR, no dose-response is reported; the implied levers (quit smoking, treat HTN, manage diabetes) are inferences from association, not demonstrated AMD-specific effects.
Quality caveats — why the tier is high, not gold, and confidence stays low. All 18 studies are observational and «the observational nature of these study designs marks a major limitation»; the pooled ORs mix crude and adjusted estimates — «the differences in adjustments of the calculated OR in the included studies may account for heterogeneity in some outcomes» — so residual confounding is uncontrolled and the point estimates are not confounder-adjusted throughout. The venue is a mega-journal (Medicine, Wolters Kluwer, open-access CC-BY-NC); each factor pools only a handful of studies; and smoking’s I2 = 81% is high. A significant OR here is a consistent association, not a demonstrated causal effect. (Babaker et al., 2025)
Directionality is genuinely contested for the CVD arm. The MA frames CVD as a predictor of AMD, but its own discussion corroborates the link with AMD -> CVD studies (AMD raising myocardial- infarction and subclinical-CVD risk). AMD and CVD share a vascular/atherosclerotic substrate, so which precedes which is unresolved — the safe reading is co-occurrence / shared risk substrate, not CVD-causes-AMD.
Decision relevance (Layer 1)
The honest holding is a G-gap: for AMD the fabric currently cannot recommend physical activity as a preventive lever, because the evidence is insufficient — and it must not borrow the cataract arm’s significance to say otherwise (they are distinct diseases with distinct evidence strength graded separately in the same paper). This does not argue against activity — PA remains a big rock on mortality/cardiometabolic grounds (Physical Activity Dose and Mortality) — it says the AMD-specific case is unproven. The reportable finding here is the absence of a clear lever, which is itself a decision-relevant result (it licenses not optimizing activity for the eyes specifically).
The diet arm is the AMD page’s first lever with a signal — but a weak-warrant one. Unlike PA, MD adherence does reach significance (on progression), so for someone already at AMD risk the diet lever is the better-evidenced of the two eye-specific bets. But the warrant is low (observational, n = 8, MDPI, self-reported), the benefit is for progression not development, and MD is already a big-rock lever on cardiometabolic and mortality grounds — so the AMD-specific case adds little at the margin for a person who would adopt an MD pattern anyway. Within one organ the two levers diverge — PA insufficient, diet protective-for-progression — the same “the shared lever does not transfer uniformly” caution the Shared Modifiable Levers Across Age-Related Diseases page makes for PA across cataract vs AMD.
The vascular risk factors add little at the margin, and smoking is why. The significant factors — smoking, hypertension, CVD, diabetes — are the same exposures already ranked as big rocks on mortality and cardiometabolic grounds, so an AMD-specific association changes almost no decision that their own outcomes did not already drive. Smoking is the exception worth stating plainly: it is a Layer-1 big rock, non-substitutable, and AMD is simply one more organ-specific harm of the dominant exposure (Smoking and Mortality) — the association here reinforces the existing “stop smoking” ranking rather than opening a new lever. For hypertension the realistic comparator is already a mature, low-harm BP drug class (Blood Pressure Lowering and Cardiovascular Events), so any AMD increment is captured by treatment a high-BP person would plausibly be on anyway. Because the source gives only ORs (relative) with no absolute AMD risk by stratum, absolute benefit of modifying any factor cannot be stated — a modest relative association on a low-to-moderate baseline AMD incidence is a small absolute effect. And the four significant factors are prognostic markers, not proven causal levers (route (a), not (b)/(c)) — the honest holding is these strata carry higher AMD risk, not removing the factor delivers the OR.
Limits and misreadings — where this breaks
- Do not read the null as licence to be sedentary. Insufficient AMD evidence is a statement about the AMD-specific arm only; PA’s mortality/cardiometabolic case is untouched — the correct inference is no proven eye-specific bonus, not no reason to move.
- Do not read it as a demonstrated null either. The wide CI and 3-cohort dose data cannot exclude a real small effect; a future SR could move this to benefit or to no meaningful effect. It is candidate/gap, and it fires as a finding only in the weak sense that the current lever is unproven.
- All observational, self-reported PA — the cohort measurement error that (elsewhere) argues attenuation-toward-null also caps how much weight the borderline point estimate can carry.
- Do not read the vascular ORs as intervention effects. Smoking/HTN/CVD/diabetes are prognostic associations (route (a)) from a crude-and-adjusted-mixed observational pool; the OR is who develops AMD, not how much removing the factor lowers AMD risk, and the source states no absolute risk.
Held threads / gaps
- Insufficient PA evidence — needs more cohorts, objective PA measurement, and AMD-subtype resolution (early vs late/neovascular were not separately estimable in the extracted text).
- Diet arm landed — Marques-Couto 2025 (MD -> AMD), the paired SR, now populates the diet section above. It is a different exposure from Aune’s physical-activity arm on the same disease, so the two arms are F/refinement of the AMD exposure map, never type-E independent backing (different claims, not one claim reached twice). (Marques-Couto et al., 2025)
- AMD development (not progression) is the open cell — the protective pools are for progression;
initial development rests on the cross-sectional null (a single dominant study). A gold SR or trial on
MD -> incident AMD, and on AMD subtype (geographic-atrophy vs neovascular), would move this.
G-gap. - No trial / no MR for the diet arm — the whole MD arm is observational; healthy-user confounding is
unremoved. An MD RCT with an AMD endpoint or a genetic instrument is the missing warrant.
G-gap. - Vascular-risk-factor arm landed — Babaker 2025 populates the cardiometabolic section above with prognostic ORs for smoking / HTN / CVD / diabetes. It appraises a different exposure set from the PA and diet arms on the same disease, so it is F/refinement of the AMD exposure map, never type-E independent backing of the other arms (different claims, not one claim reached twice); its authors are all new to the page, so no same-author collision. (Babaker et al., 2025)
- CVD -> AMD vs AMD -> CVD direction is unresolved — Babaker pools CVD as an AMD predictor but
corroborates with AMD-predicts-CVD studies; a design that fixes temporal order (incident-AMD cohort
with baseline vascular status, or bidirectional MR) would settle it.
G-gap. - Lipids / BMI NS on a thin base — cholesterol, triglycerides, BMI and stroke were NS, but from
few studies each and with the lipid role called an «ongoing debate»; a better-powered SR, or lipid
subtype resolution, could move these from insufficient toward benefit/harm/null.
G-gap. - No absolute-risk layer for the vascular factors — Babaker reports only relative ORs; stratum-level
absolute AMD incidence (to convert an OR into absolute benefit) is the missing piece for a Layer-1
ranking.
G-gap. - Aune/Jayedi coreference — this and the cataract page share held multi-work authors (Aune, Jayedi); any later AMD source sharing them is F/attribution, never type-E independent corroboration.