Study design and data sources

We used a two-sample MR design, analyzing summarylevel GWAS data for the exposure (olive oil consumption) and outcome (Alzheimer’s disease) from non-overlapping samples of European descent. The genetic instruments for olive oil consumption were obtained from the UK Biobank, a large population-based cohort. Specifically, we utilized genome-wide association results for the phenotype “Type of fat/oil used in cooking: Olive oil” (UKB field ID 1329), as made available by the MRC-IEU OpenGWAS project (GWAS ID: ukb-b-3875). This phenotype reflects participants’ self-reported primary cooking oil (with olive oil as one option) and can be considered a proxy for regular olive oil use in diet. The GWAS was conducted on 64,949 individuals of European ancestry in UK Biobank, with genotype imputation and association analysis performed by Neale et al. [16], where genetic effects are expressed per unit increase in olive oil use. We clumped the summary statistics to select independent top genetic variants for olive oil consumption. Variants were eligible as instruments if they reached a relaxed genome-wide significance threshold of p of less than 5 × 10^-6. SNPs were clumped at linkage disequilibrium (LD) threshold r² < 0.001 within a 10,000 kb window using the 1000 Genomes European reference panel. For each SNP, the effect size (β), standard error, and p-value were extracted, and F-statistics were calculated to assess instrument strength (F > 10 considered adequate). The proportion of variance explained (R²) was computed using standard formulas for continuous exposures. For each instrument, we recorded the effect allele and effect on olive oil use (β_exposure), standard error, and p-value. We also calculated the proportion of variance in the exposure explained (R2) and the Fstatistic for each SNP to evaluate instrument strength. The resulting instrument set consisted of 9 independent SNPs associated with olive oil use at P < 5×10^-6. All 9 SNPs had F-statistics > 10 (range from 23 to 27), indicating they are sufficiently strong instruments to minimize weak instrument bias. Supplementary Table 1 lists the SNP IDs, their olive oil association statistics, and F-statistics. Notably, none of these SNPs are located in or near the APOE gene region on chromosome 19 (the nearest genome-wide hits for olive oil were on other chromosomes), which reduces concern that our instruments could proxy APOE ε4 status. For the outcome, we obtained genetic association estimates for Alzheimer’s disease from the largest available GWAS meta-analysis of AD. We used summary statistics corresponding to GWAS ID ieu-b-5067 from the IEU OpenGWAS database, which corresponds to a 2022 metaanalysis of late-onset AD. This dataset includes 488,285 individuals of European ancestry (combining several cohorts), with roughly 71,880 clinically diagnosed AD cases and the remainder controls or proxy-cases [17]. The majority of cases came from the IGAP consortium and UK Biobank (using parental dementia history as proxy cases), as reported by Jansen et al. (2019) and updated by later analyses [18]. The AD GWAS was not conditioned on APOE genotype (i.e. the strong effect of APOE ε4 is present in the summary data), and results were reported as log-odds ratios (log OR) for AD per effect allele. We checked that none of the 9 olive oil instrument SNPs were palindromic with intermediate allele frequencies or had mismatched alleles between the exposure and outcome datasets. In initial design, if any issues had arisen, we planned to harmonize alleles and exclude ambiguous variants. In the end, all 9 SNPs were available in the AD GWAS and could be harmonized unambiguously.

Mendelian randomization

The primary MR analysis was carried out using the inverse-variance weighted (IVW) method under a randomeffects model (Figure 1), with SNP-specific weights proportional to the inverse of the squared standard error of their outcome association. Statistical significance was defined as two-sided P < 0.05, and results were expressed as β coefficients and odds ratios (OR = e^β) per 1 SD increase in genetically predicted olive oil consumption. The IVW estimate can be interpreted as the weighted regression slope of SNP-AD associations on SNP-olive associations, representing the causal log OR for AD per unit increase in olive oil consumption, assuming all SNPs are valid instruments. A fixed-effect IVW was also calculated given the relatively small number of instruments, but since there was no evidence of heterogeneity, the randomand fixedeffect estimates were nearly identical. We exponentiated the log OR to obtain an odds ratio (OR) for interpretability. The statistical significance threshold was set at P < 0.05 for the main IVW analysis, as we tested a single primary hypothesis.

To assess the robustness of the result, we implemented several sensitivity MR methods that relax the standard IVW assumptions. MR-Egger regression provides an estimate of the causal effect while allowing for an intercept term that captures average directional pleiotropy. The MREgger point estimate and its 95% confidence interval were computed. More importantly, the MR-Egger intercept for deviation from zero was tested. A non-zero intercept (with P < 0.05) would indicate overall horizontal pleiotropy among the instruments, whereas a non-significant intercept suggests no strong directional pleiotropy [19]. Weighted median MR was also performed, which could yield a consistent causal estimate even if up to 50% of the weight comes from invalid instruments, as reported previously [20]. Additionally, mode-based estimates (simple and weighted) were explored, which group SNPs by the similarity of their causal effects and derive the overall effect from the largest cluster, providing consistency under different patterns of invalid instruments [21].

Pleiotropy and heterogeneity diagnostics

Heterogeneity in the SNP effect estimates was evaluated using Cochran’s Q statistic for IVW and for MR-Egger. Cochran’s Q assesses whether there is greater dispersion in individual SNP causal estimates than expected by chance; a high Q (with P < 0.05) would suggest that effect sizes differ beyond sampling error, potentially due to heterogeneity or outliers. If significant heterogeneity were found, we planned to report the random-effects IVW result (which accounts for between-SNP variance) and investigate outliers. We also constructed a funnel plot of each SNP’s MR estimate (β_SNP = β_AD/β_olive) against its precision; symmetry of the funnel indicates no small-study or directional bias. MR-PRESSO global test was run to detect any outlier SNPs with pleiotropic distortion; if MRPRESSO identified outliers, we intended to repeat IVW after removing them [22]. Furthermore, a leave-one-out analysis was conducted by re-computing the IVW estimate 9 times, each leaving one SNP out.

Two-step Mendelian randomization mediation analyses

We hypothesized that any causal effect of olive oil on AD might operate via cardiovascular or metabolic improvements (since olive oil is known to improve lipid profiles, reduce blood pressure, etc., which in turn might influence dementia risk). To test this, we performed two-step MR for multiple mediators: six lipid traits (low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, total cholesterol, triglycerides, apolipoprotein B, apolipoprotein A1), two blood pressure traits (systolic and diastolic blood pressure), two stroke subtypes (ischemic stroke overall, and the small-vessel stroke subtype, as AD could be influenced by cerebrovascular disease), three metabolic traits (body mass index, waist-hip ratio adjusted for BMI, and type 2 diabetes mellitus), and one inflammation marker (C-reactive protein, CRP). These mediators were chosen a priori to represent major pathways of interest: lipid-related, vascular, metabolic, and inflammatory pathways. For each putative mediator, we obtained genetic association estimates for that mediator from corresponding GWAS (all largely European ancestry, taken from IEU OpenGWAS or published sources).
We then carried out two MR analyses: (a) olive-oil → mediator, using our 9 olive SNPs against the mediator outcome; and (b) mediator → AD, using SNPs associated with the mediator as instruments against AD risk. The framework for the two-step MR performed in this study is presented in Figure 2. Where possible, we used the same SNPs for both steps (i.e., if a SNP was a valid instrument for both olive and mediator), or else separate instrument sets were used and combined via the product method. The indirect effect of olive oil on AD via a given mediator was calculated as the product of the MR estimate for olive oil on the mediator and the estimate for the mediator on AD. Standard errors for indirect effects were derived using the delta method, and Z-tests were used to assess the significance of mediation. We also estimated the proportion of the total effect mediated, although given the total effect was close to 0, this proportion is not meaningful in practice. The two-step MR analyses were implemented for each mediator separately; a Bonferroni correction could be applied within each category if needed, but since none approached significance, we report unadjusted p-values for clarity.

Multivariable Mendelian randomization

While two-step MR evaluates potential mediating pathways separately, multivariable Mendelian randomization (MVMR) allows assessment of the direct effect of olive oil consumption while simultaneously controlling for multiple correlated risk factors. We included four traits that could lie on the causal pathway: apolipoprotein B (ApoB), systolic blood pressure (SBP), diastolic blood pressure (DBP), body mass index (BMI), and C-reactive protein (CRP). These were chosen because olive oil intake in observational studies is linked to improved lipid profiles, lower blood pressure, and reduced inflammation, all recognized midlife risk factors for dementia. SNP instruments for ApoB, SBP, BMI, and CRP were obtained from large published GWAS (> 250,000 samples each) and combined with olive oil SNPs after clumping (r² < 0.001; 10,000 kb) to ensure independence. Genetic associations for each SNP with all exposures and with Alzheimer’s disease were harmonized. Multivariable IVW regression was then applied to estimate the direct causal effects of each exposure on AD, accounting for correlations between instruments [23]. The model used inverse-variance weighting and was fitted in R (version 4.2.2) using the TwoSampleMR and MVMR packages [24, 25]. Conditional Fstatistics were used to assess instrument strength, and all tests were two-sided with significance set at P < 0.05. Analyses followed the STROBE-MR reporting guidelines, and no multiple comparison adjustment was applied to the primary hypothesis (olive oil association with AD).

Genetic instruments for olive oil consumption

We identified 9 independent SNPs associated with olive oil use in UK Biobank at P < 5 × 10^-6 (none reached the conventional 5 × 10^-8 threshold). The lead variants were located in several loci across the genome (on chromosomes 1, 3, 5, 7, 8, 9, 10, and 11, Supplementary Table 1). The strongest instrument was rs6966854 on chromosome 7 (P = 2.2 × 10^-7 for association with olive oil use), and all 9 SNPs had P < 5 × 10^-6 by design. The effect sizes for these SNPs on the olive oil phenotype were small (e.g. > the largest β was 0.055 on an SD scale), reflecting the polygenic and weakly heritable nature of dietary preferences. Cumulatively, the 9 SNPs explained an estimated R2 of 0.074% of the variance in olive oil use. The mean F-statistic was 24, with each SNP having F > 10, indicating that the genetic instrument collectively had sufficient strength for MR (reducing risk of weak-instrument bias towards the null). None of the SNPs showed genome-wide associations with obvious confounders (we cross-checked against known trait associations in GWAS Catalog), though a few were in or near genes of interest (for example, one SNP near ANKRD55 has been associated with autoimmune conditions, and another near KLF10 related to metabolic traits). Palindromic variants were aligned by allele frequency, and there were no ambiguities in harmonization.

Primary analysis of olive oil and Alzheimer’s disease

The IVW MR analysis revealed no evidence of a causal effect of olive oil consumption on Alzheimer’s disease risk (Figure 3). The scatter plot shows that all nine genetic instruments cluster tightly around the null effect line. The regression slopes for IVW, MR-Egger, weighted median, and mode-based estimators are nearly flat and overlapping, indicating consistency across methods. The narrow dispersion of SNP estimates around the origin suggests no outlier variants with disproportionately large effects, reinforcing the robustness of the null finding.

The IVW estimate (random-effects) was a log OR of 0.0017 (95% CI: -0.0045 to 0.0079) per SD increase in olive oil consumption (Table 1). This corresponds to an OR of 1.0017 (approximately 1.002) for AD per SD higher olive oil use, with a 95% confidence interval ranging from 0.9955 to 1.0079. The association was not statistically significant (P = 0.583). Using a fixed-effects IVW model (appropriate given low heterogeneity), the estimate was identical (β = 0.0017) with a slightly smaller standard error (SE = 0.0024) and P = 0.48. For interpretation, an SD increase in the olive oil usage phenotype (which might be akin to comparing people who primarily use olive oil versus those who rarely do) was associated with essentially no change in AD risk. The point estimate was very close to zero, and the confidence intervals excluded any effect of material importance. For instance, we can infer that a protective effect larger than roughly 1% reduction in odds of AD per SD of olive oil use is statistically incompatible with our data (since the 95% CI lower bound was OR 0.995). Similarly, there was no evidence of a harmful effect (upper CI: 1.008, just a 0.8% possible increase) (Table 1).

All sensitivity MR methods concurred with the null finding (Figure 4). The forest plot illustrates individual SNP effects on Alzheimer’s disease risk, along with pooled estimates from IVW and MR-Egger. Each SNP-specific estimate is centered close to zero, with 95% confidence intervals overlapping the null. The combined IVW and MR-Egger estimates are similarly centered around the null line, demonstrating that neither individual instruments nor aggregate estimates provide evidence for a causal association. This uniformity across SNPs rules out the possibility that a subset of instruments might be driving the results.

MR-Egger regression yielded a causal estimate of β = 0.0059 (SE 0.0105), which, although positive in point estimate, was not significantly different from zero (P = 0.59). The MR-Egger point estimate has a wide CI (-0.0148 to 0.0266 on the log OR scale), but it overlaps with the IVW estimate. Importantly, the MR-Egger intercept was -0.00008 (SE 0.00019), which is very close to zero with P = 0.689. This indicates no evidence of directional pleiotropy on average, the genetic instruments do not appear to influence AD through pathways unrelated to olive oil (if they did, the intercept would likely deviate from zero). The intercept CI (-0.00045 to 0.00029) is narrow, suggesting any net pleiotropic effect is negligible. Weighted median MR gave an estimate of β = 0.0016 (SE 0.00345, P = 0.648), virtually identical to IVW. Simple mode and weighted mode estimates were also near zero (β ≈ 0.0015) and non-significant (P > 0.76). Thus, whether we assume that at least 50% of the weight comes from valid instruments (weighted median) or even that only the largest instrument cluster is valid (mode methods), the conclusion remains the same: no causal effect of olive oil on AD. Table 1 summarizes the MR estimates by method, all of which are null.

Pleiotropy and heterogeneity checks

We found no significant heterogeneity in the SNP-specific causal estimates. Cochran’s Q for IVW was Q = 13.11 on 8 degrees of freedom (P = 0.108), indicating that the variation in SNP effects could be due to chance alone. For MR-Egger, Q = 12.79 on 7 df > (P = 0.077); while slightly lower df >, this was also non-significant (and the borderline P = 0.077 likely reflects one less SNP or inherent noise rather than definite heterogeneity). Given the nonsignificant Q, we interpret that the instruments are fairly homogeneous in their estimated effect on AD which strengthens the validity of the IVW estimate. The scatterplot visually confirms that all SNPs cluster tightly around the null line, with no outliers far off the regression line (Figure 3).
The funnel plot (Figure 5) of SNP effects was symmetric and roughly inverted funnel-shaped, as expected under no directional pleiotropy. The funnel plot displays SNP estimates plotted against their precision. The distribution is symmetric around the vertical null line, indicating the absence of systematic bias. SNPs with higher standard errors (less precise estimates) scatter more widely, as expected, but remain evenly balanced across both sides of the null. The lack of skewness in the funnel supports the conclusion from MR-Egger’s intercept test that there is no directional pleiotropy, confirming that genetic instruments are valid.

SNPs with larger SE (less precise) showed wider scatter around the zero effect, but there was no tendency for those to lean to one side of the null. The MR-Egger intercept test, as noted, was null (P = 0.69), reinforcing that there is no evidence that the instruments as a group have pleiotropic biases. We also applied the MR-PRESSO global test, which did not detect any outlier SNP (global test P = 0.72). Therefore, we did not remove any SNPs as potential pleiotropic outliers all instruments were retained in the analysis.
The leave-one-out sensitivity analysis further demonstrated the robustness of the null result (Figure 6). The leaveone-out analysis demonstrates that removing any single SNP does not materially alter the overall causal estimate. Each recalculated IVW estimate lies close to zero, with confidence intervals consistently spanning the null. This stability across all iterations indicates that no single SNP exerts undue influence on the analysis. Such robustness enhances confidence that the null finding is not driven by outlier variants or instrument instability.

When excluding each SNP in turn, the IVW estimates remained very close to zero and none of those excludingone analyses yielded a significant association. For instance, leaving out the top SNP on chr7 (rs6966854) gave an OR of 1.0017 (P = 0.63); leaving out rs11137009 (chr8) gave OR: 0.9998 (P = 0.96); leaving out rs10970249 (chr9) gave OR: 1.0025 (P = 0.47); and so on (full leaveone-out results in Supplementary Table 2). No single variant’s removal produced a notable swing in effect size or confidence interval. We also note that none of the individual SNP-AD associations showed even a nominally significant effect on AD (all single-SNP Wald ratios had P > 0.4).

Two-step Mendelian randomization results

Given the null total effect of olive oil on AD, the mediation analyses also yielded null results. The two-step MR provided no significant indirect effects through any tested pathway, but the estimates still help clarify where an effect would have appeared if present (Table 2). For lipid traits, genetic proxies for olive oil showed no meaningful association with circulating lipids. For LDL-C, the IVW estimate for olive oil to LDL-C was β = +0.055 SD per SD of olive oil (95% CI -0.137 to +0.248, P = 0.58). Associations were likewise non-significant for HDL (β = +0.083, P = 0.24), triglycerides (β = -0.046, P = 0.38), total cholesterol (β = +0.0313, P = 0.14), apolipoprotein B (β = -0.046, P = 0.75), and apolipoprotein A1 (β = +0.059, P = 0.32). MR estimates for lipid traits to AD were mostly null as well: LDL-C β = -0.00046 per SD (P = 0.18), HDL-C β = -0.00026 (P = 0.28), triglycerides β = -0.00033 (P = 0.22), and ApoB β = -0.00020 (P = 0.48).

Because neither step was significant, indirect effects were essentially zero, for example LDL-mediated β_indirect = 0.055 × (-0.00046) = -0.000025 with P = 0.61. Results were similar for HDL, triglycerides, total cholesterol, ApoB, and ApoA1 (all P > 0.4). These findings do not support lipid-mediated pathways.
Genetically higher blood pressure showed only weak, non-significant links to AD in this dataset; for example, a 10 mmHg higher SBP corresponded to an OR of 1.00 for AD with P = 0.97. Accordingly, indirect effects via SBP or diastolic blood pressure were non-significant (both P > 0.9). Using large stroke GWAS instruments, olive oil genetic scores were not associated with ischemic stroke risk (P = 0.56), and stroke liability did not show a significant effect on AD (P = 0.86). Small-vessel stroke behaved similarly. No mediation through blood pressure or stroke was detected.
For metabolic traits, genetically higher olive oil use did not reduce adiposity or diabetes risk. The point estimate for BMI was positive and non-significant (P = 0.53). Waist-hip ratio adjusted for BMI showed no association (β = -0.002 SD, P = 0.98), and type 2 diabetes liability was also null (OR: 1.03 per SD of olive oil, P = 0.41). In mediator-to-outcome analyses, higher genetically predicted BMI showed a small, non-significant inverse relation with AD (P = 0.49), and type 2 diabetes liability had no significant effect (P = 0.16). Indirect effects of olive oil via BMI and diabetes were not significant (P = 0.64 and 0.48). There is therefore no evidence for mediation through metabolic pathways. For inflammation, CRP did not appear to mediate any effect. Olive oil instruments were not associated with CRP (β = +0.027 SD of log-CRP per SD of olive oil, P = 0.63), and CRP showed no causal effect on AD (β = -0.00032 per SD, P = 0.36). The indirect effect via CRP was non-significant (P = 0.67).
Overall, across lipid, vascular, metabolic, and inflammatory domains, two-step MR showed no significant mediation (Figure 7). Neither the cardiometabolic effects often attributed to olive oil nor potential adverse metabolic effects translated into any detectable change in AD risk within the precision of this analysis. The analysis shows that genetically predicted olive oil consumption does not significantly alter lipid traits (LDL, HDL, triglycerides, ApoB, ApoA1), blood pressure (systolic, diastolic), metabolic traits (BMI, WHR, type 2 diabetes), or inflammation (C-reactive protein). Furthermore, none of these traits exhibit a significant causal relationship with AD in this dataset. Consequently, all indirect effect estimates are close to zero, with wide confidence intervals encompassing null.

Multivariable Mendelian randomization results

A multivariable MR was performed including olive oil and four risk factors (ApoB, SBP, BMI, CRP) to estimate the direct effect of each on AD, where the results are presented in Table 3. The direct effect of olive oil on AD, controlling for those factors, was null. The multivariable IVW yielded a coefficient β = -0.00244 for olive oil (SE = 0.00241), corresponding to OR of 0.9976 per SD increase in olive oil use, with P = 0.31.

The MVMR estimates for other exposures also did not reach a statistical significance. For ApoB, the direct effect was β = -0.00037 (SE = 0.00035) on AD (OR: 0.9996 per 1 SD increase), P = 0.29. SBP showed β = +0.00040 (SE = 0.00056) per SD increase (5 mmHg), P = 0.48. BMI had β = -0.00036 (SE = 0.00073) per SD increase (4.8 kg/ m2), P = 0.62. CRP had β = -0.00078 (SE = 0.00071) per SD increase, P = 0.27. None of these associations reached significance. In the multivariable analysis, olive oil also did not demonstrate a significant direct effect on AD. The point estimate was close to null (OR:1.00), with a 95% CI ranging approximately from 0.993 to 1.002. The MVMR models accounted for overlapping instruments across exposures, and the confidence intervals, although wider than in univariable MR, still excluded moderate effects.