If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Findings from randomized controlled trials have yielded conflicting results on the association between blood pressure (BP) and dementia traits. We tested the hypothesis that a causal relationship exists between systolic BP (SBP) and/or diastolic BP (DBP) and risk of Alzheimer’s disease (AD).
We performed a generalized summary Mendelian randomization (GSMR) analysis using summary statistics of a genome-wide association study meta-analysis of 299,024 individuals of SBP or DBP as exposure variables against three different outcomes: 1) AD diagnosis (International Genomics of Alzheimer's Project), 2) maternal family history of AD (UK Biobank), and 3) paternal family history of AD (UK Biobank). Finally, a combined meta-analysis of 368,440 individuals that included these three summary statistics was used as final outcome.
GSMR applied to the International Genomics of Alzheimer's Project dataset revealed a significant effect of high SBP lowering the risk of AD (βGSMR = −0.19, p = .04). GSMR applied to the maternal family history of AD UK Biobank dataset (SBP [βGSMR = −0.12, p = .02], DBP [βGSMR = −0.10, p = .05]) and to the paternal family history of AD UK Biobank dataset (SBP [βGSMR = −0.16, p = .02], DBP [βGSMR = −0.24, p = 7.4 × 10−4]) showed the same effect. A subsequent combined meta-analysis confirmed the overall significant effect for the other SBP analyses (βGSMR = −0.14, p = .03). The DBP analysis in the combined meta-analysis also confirmed a DBP effect on AD (βGSMR = −0.14, p = .03).
A causal effect exists between high BP and a reduced late-life risk of AD. The results were obtained through careful consideration of confounding factors and the application of complementary MR methods on independent cohorts.
). Hypertension is one of the most common conditions that degrade cerebral circulation, and it has been found that prolonged high blood pressure (BP) is a cause of vascular dementia in individuals under 85 years of age (
). This evidence points to a potential link between BP and AD, which would make BP an ideal target for early AD therapy, given that medications able to manipulate BP exist. However, despite several published studies including meta-analyses of randomized controlled trials (RCTs), this association between high BP and AD remains unclear, with many studies reporting discordant results (
). However, even RCTs, integrally designed to tackle study confounding, have failed to provide consistent results on the dementia outcome, either because of the long follow-up between hypertension manifestation and dementia onset or because of an early termination of the study (
Conversely, a recent longitudinal study on ∼2.6 million individuals in the United Kingdom confirmed a moderate positive association between long-term high BP and vascular dementia, though, paradoxically, the study reported a weaker inverse association with AD (
) was reported as a causal association, but such inference should be interpreted cautiously. These inconsistencies in the literature call for the use of alternative methods, for instance those that target causal effects and use complementary data sources, rather than single cohorts. Mendelian randomization (MR) is a causal inference methodology that can test the effect of a modifiable exposure on a disease by using genetic variants to provide evidence of robust associations and incorporates the advantage of summary statistics from large-cohort genome-wide association studies (GWASs) (
In this study, two-sample MR approaches were used to assess the potential causal relationship between high BP (systolic BP [SBP]/diastolic BP [DBP]) and AD risk. Data were obtained from a collection of complementary sources, including large-scale GWAS analyses on blood pressure (BP) (
This study relied on de-identified summary-level data that have been made publicly available (the International Consortium for Blood Pressure [ICBP] GWAS and the International Genomics of Alzheimer's Project [IGAP] GWAS) and on the patient-level UKBB phenotypic and genetic data; ethical approval was obtained in all original studies and for the UKBB data usage (application No. 43309). Data were analyzed for this study from January 1, 2019, to June 30, 2019.
The analyses were performed using genetic instruments associated with BP selected from summary statistics of a recently published Stage 1 GWAS meta-analysis as part of the ICBP (
). The IGAP cohort uses AD diagnosis as an outcome measure. The UKBB cohort uses self-reported family history of AD/dementia as a proxy for AD diagnosis. The accuracy of this measure has been shown using genetic correlation analysis and reported by Marioni et al. (
). In order to assess whether any genetic instrument was previously associated with AD as an outcome and remove possible confounding, we searched the PhenoScanner GWAS database (version 2; http://phenoscanner.medschl.cam.ac.uk) (
) and removed SNPs that reached genome-wide significance (p < 5 × 10−8) in published dementia GWASs. As part of the data harmonization process, we checked that all effect/reference alleles were in agreement with the dbSNP151 reference in all used summary statistics (
) and in the UKBB family history results from the paternal and maternal logistic regression analyses described below. All nonmatching alleles were switched (A1=A2 and A2=A1) and the sign of the beta estimates changed. For quality control, we checked for reciprocal strand alleles (e.g., C=G and T=A), excluding alleles with ambiguous SNPs and allele frequencies between 0.4 and 0.6. All instrument SNPs present in the UKBB were in Hardy-Weinberg equilibrium (p > 1 × 10−10), and they also had a minor allele frequency >1% and imputation quality info score >0.9. A list of genetic instruments is reported in Table S1. We retained a total of 81 genome-wide-significant SNP instruments from the ICBP summary statistics (DBP, n = 61; SBP, n = 64) (
), and a description of the SNP exclusion is reported in the Supplement.
We tested the hypothesis that high BP may have a causal effect on AD. We performed a two-sample MR analysis using beta estimates of the ICBP Stage 1 meta-analysis (Supplement) as exposure for BP in the MR design, and the following datasets as outcome for AD: 1) β estimates and standard errors (SEs) of AD from the IGAP (
) (see Supplement); 2) log-transformed odds ratios (ORs) and SEs obtained from analysis of MFH-UKBB and PFH-UKBB data (Supplement); and 3) β estimates and SEs of a combined meta-analysis summary statistics (MA-AD) (
), which excludes SNPs that show evidence of pleiotropic effects by the heterogeneity in dependent instruments outlier analysis (HEIDI-outlier test < 0.01). We therefore estimated a causal association between the exposure and the outcomes using the GSMR method (
We used 10,000 randomly selected unrelated samples from the UKBB dataset as a reference to determine linkage disequilibrium patterns, and we clumped SNPs for independence if correlated at r2 > .001. Only one representative SNP was retained.
We subsequently performed a sensitivity analysis for each dataset combination using the TwoSampleMR R (
) package to conduct MR analysis using inverse variance–weighted (IVW) meta-analysis with fixed/random effect and MR-Egger regression. The IVW is the equivalent of a weighted regression of exposure on outcome effects with the intercept constrained to zero. Owing to this constraint, the results can be biased if instrument SNPs show horizontal pleiotropy. This can be caused by the influence on the outcome through causal pathways other than the exposure (
). We therefore compared the IVW results with the MR-Egger regression, testing whether the MR-Egger intercept was different from zero. This was used as an indicator of average pleiotropic bias because this regression intercept is not constrained at the origin. Although the MR-Egger method’s estimate is known to be relatively robust to the presence of pleiotropy, it is also affected by a reduced statistical power (
). Therefore, in presence of pleiotropy, an MR-Egger regression estimate was used; in case of no pleiotropy, an IVW-MR estimate was preferred. Moreover, in order to detect heterogeneous outcomes, we further included leave-one-SNP-out analyses and the modified Cochran Q statistic, also implemented in the TwoSampleMR R package (
). The coefficient of determination (R2) of exposure on genetic variants was calculated on the total allele score using the TwoSampleMR R package.
SNPs were used as instruments in GSMR on three different AD exposures: 1) AD risk from the IGAP, 2) MFH of AD (MFH-UKBB), and 3) PFH of AD (PFH-UKBB). Finally, subsequent meta-analysis (MA-AD) confirmed results by combining the IGAP, MFH-UKBB, and PFH-UKBB outcomes. We used sensitivity analyses to measure the effect of employing different types of MR methods, IVW-MR and MR-Egger, and of pleiotropy.
Results using the IGAP dataset showed a statistically significant protective causal effect of SBP on AD (βGSMR = −0.19, p = .04) (Figure 1A). IVW-MR with random effect and MR-Egger confirmed the GSMR results for SBP (xy-IVW = −0.19, p = .03; xy-MR-Egger = −0.72, p = .03). No pleiotropy was detected in this case. The DBP analysis in the IGAP did not show evidence of a causal relationship (βGSMR = −0.06, p = .55) (Figure 1B). We additionally performed the same analyses in IGAP excluding the ambiguous-strand SNPs that were included in the previous analysis. The alternative results did not deviate significantly from the original DBP analysis in the IGAP (data not shown). Sensitivity analysis detected a source of pleiotropy (MR-Egger intercept, p < .01), and the MR-Egger regression estimates confirmed an inverse causation in the DBP (xy-MR-Egger = −0.95, p = 3.9 × 10−3) analysis. No evidence of confounding heterogeneity of effect sizes (leave-one-out; Cochran Q statistic, p > .10) was observed in either of these two analyses.
In the UKBB cohort, estimates confirmed a protective causal relationship between both SBP and DBP with AD family history: MFH-UKBB (SBP [βGSMR = −0.12, p = .02], DBP [βGSMR = −0.10, p = .05], PFH-UKBB (SBP [βGSMR = −0.16, p = .02], DBP [βGSMR = −0.24, p = 7.4 × 10−4]) (Figure 1C–F). Sensitivity analyses found no evidence of confounding heterogeneity of effect sizes (leave-one-out; Cochran Q statistic, p > .10) or from pleiotropy (MR-Egger intercept, p > .05). Furthermore, IVW-MR with random effect confirmed the GSMR results: SBP (MFH-UKBB [xy-IVW = −0.12, p = .02], PFH-UKBB [xy-IVW = −0.17, p = .03]), DBP (MFH-UKBB [xy-IVW = −0.10, p = .03], PFH-UKBB [xy-IVW = −0.24, p = 3.1 × 10−4]) (Table 1).
Table 1Summary of the MR-Based Analysis of Blood Pressure and Dementia
In order to compare the results between the GSMR analysis and the other MR analysis, the same SNPs were used in each MR method. MR-Egger meta-regression intercept different from zero (p value < .05) was tested as an indication of directional horizontal pleiotropy driving the results. An MR-Egger test intercept different from zero was observed in the DBP-IGAP analysis. In this case, MR-Egger was chosen to estimate the causal relationship between BP and AD. In all the other analyses, no horizontal pleiotropy was detected (pleiotropy p > .05) and an IVW method was used to estimate a causal effect. A Cochran’s Q heterogeneity p value >.10 was observed in all analyses. All analyses used the ICBP GWAS summary statistics (
). A total of 64 SNPs for SBP and 59 SNPs for DBP overlapped between the UKBB and IGAP datasets. For SBP, we confirmed the overall significant causal effect (βGSMR = −0.14, p = .03; xy-IVW = −0.14, p = .03) (Figure 1G; Table 1). For DBP, we also confirmed a protective effect of high DBP and AD (, p = .03; xy-IVW = −0.14, p = .03; xy-MR-Egger = −0.50, p = .02) (Figure 1H; Table 1). No evidence of pleiotropy (MR-Egger intercept, p > .05) or heterogeneity of effect sizes was found in either of these two analyses (leave-one-out; Cochran Q statistic, p > .10).
Both the MFH-UKBB and PFH-UKBB datasets reached a moderate power to detect a causal effect both in SBP analyses (MFH-UKBB, 51.6%; PFH-UKBB, 53.3%) and in DBP analyses (MFH-UKBB, 46.7%; PFH-UKBB, 86.7%). Also, statistical power of 63.1% was reached using the IGAP dataset in the SBP analysis, but reduced power of 12.1% was evident when using DBP genetic instruments. The overall analysis using the MA-AD results reached a minimum of 94.2% power in the SBP analyses and a maximum of 95.6% power in the DBP analyses (Table S2).
Observational analyses of the relationship between BP measurements and family history of AD were estimated in the UKBB dataset (methods reported in the Supplement; cohort summary statistics in Table S3). Logistic regression analyses adjusting for age, sex, body mass index, smoking, hypertension medication, 40 principal components, genotype array, and genetic batch were performed. Results showed that SBP was associated with a protective effect on family history of AD (Figure 2; Supplement), which was detectable after both causal and observational estimates were scaled to a genetically predicted 10-mm Hg difference in BP. This protective effect was unchanged when the observational model was adjusted for antihypertensive drugs (Figure 2).
Several publications suggest that high BP manifested in midlife may contribute to an increased risk of AD in late life (
). However, a meta-analysis published by the Alzheimer’s Research Forum suggested that no association exists between midlife high SBP or high DBP and the incidence of AD but suggested that an inverse association between late-life hypertension and AD does exist (
). Similar contradictory results have been reported by other meta-analyses including observational studies or meta-analyses of RCT studies that had not been able to determine a clear effect of midlife use of antihypertensive drugs and late-life risk of dementia (Supplement and Table S4). A possible cause of these contradictory results might be that observational studies analyze correlation, rather than causation, and that RCT studies did not reach statistical power. In this study, we performed two-sample MR analysis, using publicly available GWAS summary statistics and the UKBB dataset, to test for a causal association between higher levels of BP due to genetic predisposition and lower AD risk. Using family history of AD as a proxy for AD status, we report evidence of a protective effect of genetically inherited higher levels of BP and a lower chance of having PFH or MFH of AD. A logistic regression analysis of family history of AD and BP was also performed in the UKBB, reporting an association between high BP levels and reduced risk of AD. Findings from the MR and logistic regression analyses showed similar standardized ORs (Figure 2). Our data confirm previous findings (
), and by selecting larger sample sets we showed how increased power can provide more conclusive results when using MR methods. For this study, we used systolic and diastolic measures of BP that have a shared genetic component. Even though two-thirds of the SNPs used in this study were associated with both SBP and DBP (Table S1), a slightly stronger association was detected with DBP (Table 1).
This study included only genetic instruments that showed the strongest association with high BP and that were recently published in a large meta-analysis (∼1 million individuals). We applied stringent filters to include only SNPs that confirmed previous findings to avoid weaknesses often found in instrument selection (
) reported a protective causal relationship of SBP with AD (OR of 0.83 [95% confidence interval (CI), 0.73–0.94] per genetically predicted 10-mm Hg increase in SBP) and included 24 SNPs using the IGAP summary statistics. Second, Larsson et al. (
) found no association between SBP (OR of 0.94 [95% CI, 0.77–1.14] per genetically predicted 10-mm Hg increase in SBP) or DBP (OR of 0.96 [95% CI, 0.79–1.16] per genetically predicted 10-mm Hg increase in DBP) with AD, retaining 100 SNPs and using, also, the IGAP summary statistics. Third, Andrews et al. (
) also found a moderate protective effect of both SBP (OR of 0.99 [95% CI, 0.99–1.00] per genetically predicted 10-mm Hg increase in SBP) and DBP (OR of 0.99 [95% CI, 0.98–0.99] per genetically predicted 10-mm Hg increase in SBP) using both novel and replicated SNPs associated with BP (435 SNPs for SBP analyses and 450 SNPs for DBP analyses) published by Evangelou et al. (
), giving a greater power to detect a larger proportion of variance, hence the detection of a true null association. However, there is also a possibility of a dilution of the association due to a greater pleiotropy caused by the higher number of instruments. Our study also used the IGAP results, but we included a more powerful combined meta-analysis including the IGAP and UKBB together (
), which reached a significant statistical power to detect a causal effect of ∼96% in the DBP analyses and 94% in the SBP analyses (Table S2). Our results were consistent both at the single-study level (PFH/MFH; IGAP) and when using summary statistics from the AD meta-analysis published by Marioni et al. (
). Moreover, the uncertainty in the SNP-exposure and SNP-outcome association estimates is independent, and if weak instrument bias were an issue, effects would be in the direction toward the null and, therefore, would not be the cause of overestimation of the MR estimates (
). Our analyses did not reach enough statistical power for either SBP (OR, 0.97 [95% CI, 0.89–1.05]; 13.3% power) or DBP (OR, 0.96 [95% CI, 0.84–1.10]; 8.4% power) to detect a true causal effect, whereas Andrew et al. (
) reported the presence of heterogeneity in both MR models, while we did not find any evidence that our analyses required further corrections when using only established SNPs associated with BP. These sensitivity analyses’ results are reported in the Supplement (MR power analyses in Table S2 and MR analyses in Table S5) and have not been used in our main analyses owing to a lack of statistical power in the MR method.
Self-reported parental family history of AD and dementia from UKBB participants was employed as a proxy phenotype for assignment of cases and controls in this study. We accept that this information could be affected by a self-reporting bias; however, along with variable accuracy, this was already evaluated through genetic correlation analysis in a global meta-analysis showing that self-reported information of parental AD can indeed reflect an accurate proxy for clinical diagnosis (
), which limits the applicability of our findings to more diverse ethnic populations.
There are several potential confounders that were considered when planning this study. First, to overcome the issue of population stratification, we corrected the UKBB genetic associations for population relationship by including genetic principal components (
). Second, antihypertensive medication use has a modifying effect on the exposure, and this confounding might have induced the inverse causal effect detected by our MR analysis between AD risk and high BP (
). Meta-analyses that published the effect of antihypertensive medications on AD show contradictory results (either a protective effect or no effect) (Table S4). Our observational results remained unaffected after adjusting for this (Figure 2), and a recent study that used an MR approach to determine the effect of antihypertensive drugs on AD risk did not detect any causal effect by alleviating high BP (
). Specifically, the follow-up of individuals genetically predisposed to high BP (SBP/DBP) might be discontinued owing to the higher mortality rate (due to other clinical complications, such as stroke or heart attack) (
) in individuals with either PFH or MFH history of AD against controls with no family history. The comparison of the moderate-to-vigorous physical activity score means in the offspring groups with family history of AD and no family history of AD showed no statistically significant differences (Table S6). Therefore, a change in the offspring lifestyle, hence physical activity, does not seem an attributable confounding factor influencing results interpretability.
The UKBB included individuals with AD family history with age at assessment close to AD average age at onset of 65 years (MFH: median age 61 years; PFH: median age 59 years) compared with subjects with no AD family history (median age 57.5 years) (Table S6). On the one hand, when comparing UKBB BP measurements in these two groups (Table S7A, B) by subgrouping them in age groups, hypertension was detected neither in the group <50 years of age nor in the group with age ranging from 50 to 59 years, nor was hypotension detected in the group >60 years of age, as we could expect considering recent findings of midlife hypertension and late-life hypotension in AD subjects (
). On the other hand, hypertension between 140 and 150 mm Hg was observed in the group >60 years of age, and the family history of AD groups showed a slightly lower overall BP compared with the group no family history of AD. This could still imply a healthier volunteer composition of the UKBB (
However, the IGAP study, by including AD-diagnosed patients with a minimum age at onset of 68 years and older, presented similar MR results, included either as an MR outcome by itself or in combination with the UKBB in an overall meta-analysis (MA-AD) (
). It still needs to be addressed whether possible hypotension could explain the results obtained using the IGAP summary statistics, considering that no BP measurement was provided by the IGAP consortium.
In conclusion, this study leverages support for a causal association between high SBP and/or high DBP BP profiles and a reduced late-life risk of AD. Our findings were obtained by analysis of the estimates associated with genetic instruments, previously identified by GWASs within participants of European ancestry. This was followed by the application of multiple MR methods to minimize the extent of confounding factors and allow exclusion of genetic instrument outliers. Nevertheless, further research is required to confirm the impact of the use of antihypertensive drugs on dementia outcomes and address potential survival bias, in order to determine whether the protective causal effect detected in the present and previous MR studies (
) can describe the true nature of the relationship between BP and AD.
Acknowledgments and Disclosures
This work was supported by Janssen Research and Development, LLC (of Johnson & Johnson). The authors who received direct salary support were WS, LW, DN, and MF.
WS and LW had full access to the UK Biobank data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. WS, AJN-H, LW, and DN were involved in concept and design. All authors were involved in acquisition, analysis, or interpretation of data. WS and LW were involved in drafting of the manuscript. WS, LW, DN, MF, LS, SMG, AP-U, DPA, and AJN-H were involved in critical revision of the manuscript for important intellectual content. WS and LW were involved in statistical analysis. DN and AJN-H obtained funding. LW, DN, AJN-H, and NB were involved in administrative, technical, or material support. LW, AJN-H, and NB supervised.
We thank Prof. Simon Lovestone, Prof. John Isaac, Dr. Vaibhav A. Narayan, and Dr. Dai Wang for the helpful discussion on the project.
The authors report no biomedical financial interests or potential conflicts of interest.
As life expectancy increases, so does the number of people who are living with dementia, currently around 50 million worldwide (according to the World Alzheimer Report 2018). This number is projected to rise to 152 million by 2050, presenting an ever-increasing burden on society. In the absence of effective disease-modifying treatments, strategies have been focused on identifying potentially modifiable risk factors that could be targeted to delay or prevent dementia (1). One of these risk factors is hypertension.