Genetically determined low income modifies Alzheimer's disease risk.

BACKGROUND: Socioeconomic status (SES) is considered to be associated with the prevalence of Alzheimer's disease (AD). However, the causal association remain unclear. Here, we determining whether income has a causal protective effect on the risk of developing AD using Mendelian randomization (MR).
METHODS: Single-nucleotide polymorphisms (SNPs) that are strongly associated with household income levels (P<5×10-8) from the UK Biobank (UKB) (n=286,301) were selected as instrumental variables for this study. Confounding instruments were removed through data set browsing. Selected SNPs were then harmonized with results from an AD genome-wide meta-analysis (71,880 cases, 383,378 controls) including both case-control and proxy cases. The analysis was conducted using MR methods, and multiple sensitivity analyses were applied for testing of potential bias.
RESULTS: After confounding instrument removal and clumping, 9 SNPs associated with household income level identified by the UKB were left for the MR analysis. Our results demonstrated that higher household income level was causally related with a lower risk of AD (odds ratio 0.78, 95% confidence interval: 0.69-0.89; P<0.001). Multiple sensitivity analyses suggested no obvious evidence for heterogeneity or pleiotropy of the results.
CONCLUSIONS: Under MR assumptions, our results suggest robust evidence of a causal association between household income and AD risk, which may provide potential prevention strategies for this devastating disease. 2021 Annals of Translational Medicine. All rights reserved.

Introduction

Alzheimer’s disease (AD) is the most common form of dementia, and it has major implications for worldwide health and social services (1). No effective intervention is currently available for this devastating disease (2). The identification of modifiable risk factors that prevent the disease may be one possible avenue to relieve the burden of AD. Socioeconomic status (SES) corresponds to a complex bundle of social and economic factors including education, income, and occupational status (3) and is considered to be associated with the prevalence of AD (4,5). Studies have demonstrated that higher education level, increasing income, and higher occupational status are associated with decreasing risk of AD (6-9). As an important part of SES, balancing income inequality is expected to reduce the prevalence rate of AD. However, previous observational studies may be biased due to measurement error, confounding, and reverse causation and thus fail to demonstrate the causal association between income and AD (10,11). Under these circumstances, Mendelian randomization (MR) can be used to determine a causal effect.

MR uses genetic variants that are associated with risk factors as instruments for causality analysis. The results can determine whether an observational association between a risk factor and an outcome is consistent with a causal effect (11). The genetic variants must not be associated with any confounding factor or be directly associated with the outcome (12). MR has already demonstrated causal effects of potential risk factors in the context of multiple diseases and has revealed effects of body mass index, education attainment, and alcohol assumption on AD (13-15). Compared to observational methods, MR is less biased, since it uses single-nucleotide polymorphisms (SNPs) that are randomly allocated at conception as instruments to conduct instrumental variable analysis (16). In this study, we used SNPs selected from a recent genome-wide association study (GWAS) of household income as instruments to perform a 2-sample univariable MR analysis and identify the causal effect of income on the risk of developing AD. We present the following article in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting checklist (available at https://dx.doi.org/10.21037/atm-21-344).

Methods

The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013).

Income data and instruments

SNPs associated with income were obtained from a recently published GWAS on household income in a sample of 286,301 UK Biobank (UKB) participants with White British ancestry (17). There were 138,425 male participants included, and the mean age was 56.5 (SD =8.0 years). Total household income before tax was originally collected using a 5-point scale through the UKB, where 1 denoted less than £18,000, 2 £18,000–£29,999, 3 £30,000–£51,999, 4 £52,000–£100,000, and 5 greater than £100,000. The level of household income was subjected to a regression analysis using income as the outcome. The 40 genetic principal components, genotyping array, batch, age, and sex were used as predictors. Among 30 income-related loci identified in this study, 18 SNPs at a genome-wide significance level (P<5×10-8) were extracted for further analysis.

AD

AD data were obtained from a recently published genome-wide meta-analysis study, which used both traditional clinically diagnosed AD and AD-by-proxy participants (based on parental diagnoses reported by children of parents) from the UKB (18). The AD-by-proxy phenotype showed a strong genetic correlation with AD (rg=0.81). A sample of 71,880 cases consisting of both AD and AD-by-proxy participants and 383,378 controls was included in the meta-analysis, and the test statistics per SNP per GWAS were used to calculate the P value. The results were used as the outcome data set in this MR analysis.

Statistical analysis

All statistical analyses were conducted using R software (version 3.6.3, R Foundation for Statistical Computing), the MR analysis was performed using the “TwoSampleMR” package (version 0.5.4) (19).

Using the GWAS catalog (20) and the PhenoScanner GWAS database (21,22), we excluded those SNPs that were associated with confounding factors, like educational attainment, or that were directly associated with AD. After removal of SNPs that reached genome-wide significance (P<5×10-8) based on an assessment of confounding factors, 9 SNPs associated with household income level remained. The strength of the instruments was estimated on the basis of the F statistic, with an F statistic >10 commonly cited as a value that can avoid bias in MR analysis (12) and a lower mean F statistic corresponding to a greater bias (23). The selected SNPs in the exposure-outcome data set as mentioned above were then harmonized. The coefficients of SNP-exposure and SNP-outcome were combined in a fixed-effects meta-analysis, and the inverse-variance-weighted (IVW) approach was used to detect the overall estimated causal effect (24). This method is equal to a weighted regression of SNP-outcome coefficients on the SNP-exposure coefficients, and the intercept is assumed to be zero. The result is unbiased if there is no horizontal pleiotropy or if the horizontal pleiotropy is balanced. Power calculations were performed using the mRnd power calculation tool to test whether the sample size was adequate to reject the null hypothesis (25). Other MR methods including MR-Egger, weighted median, simple mode-based, and weighted mode-based were conducted to control potential bias attributable to violation of the assumptions of the IVW MR analysis. These methods are more robust for horizontal pleiotropy at the cost of reduced statistical power (26-28).

Pleiotropic genetic variants may affect the outcome independently and bias the result. Leave-one-out sensitivity analysis was used to explore SNP heterogeneity, where the random-effects IVW was performed again, but each SNP was left out in turn. This analysis determined if 1 or more SNPs were invalid instrumental variables or if the causal association was driven by a single SNP.

Furthermore, the Cochran’s Q test, MR-Egger intercept test, MR Pleiotropy RESidual Sum and Outlier (MR-PRESSO) global test were used to investigate potential pleiotropy. Cochran’s Q was employed as a test for the heterogeneity of the MR estimate, and a Cochran’s Q P value<0.05 was considered an indicator of heterogeneity (29). The MR-Egger method assumes no intercept term in the model, and the intercept is considered to be zero. When the P value is larger than 0.05 in the MR-Egger intercept test, it can provide evidence for absence of pleiotropic bias (26). MR Pleiotropy RESidual Sum and Outlier (MR-PRESSO) compares the difference between the residuals for each genetic variant in the variable, and thus pleiotropic effects can be detected and outliers identified with this method. MR-PRESSO re-analyzes the association without the outliers, correcting for possible pleiotropic effects (30).

Results

Causal effect of household income on AD risk

After instrument selection and SNP exposure-outcome data harmonization were performed as described above, 9 SNPs were left for MR analysis (). All SNPs passed the clumping with a clumping window of 10,000 kb and a linkage disequilibrium (LD) r2>0.001. All SNPs had an F statistic value greater than 10, and the mean F statistic value was 156.54, which demonstrated a low risk of instrument bias.

Table 1

Instrument single-nucleotide polymorphisms for Mendelian randomization analysis and their effects on household income

SNP	Chromosome	Effect allele	Other allele	β value	Standard error	P value
rs159365	5	G	A	0.016	0.002	9.04E-15
rs2332719	3	G	A	−0.013	0.002	1.16E-10
rs537160	6	G	A	−0.013	0.002	2.34E-10
rs12119149	1	T	C	0.013	0.002	5.24E-10
rs12954483	18	G	A	−0.013	0.002	5.80E-10
rs7597007	2	T	G	−0.012	0.002	1.46E-09
rs37976	7	T	C	0.012	0.002	3.14E-09
rs306755	20	C	T	0.012	0.002	3.82E-09
rs2563332	5	T	C	−0.012	0.002	9.90E-09

SNP, single-nucleotide polymorphism.

Using an IVW approach, we found that the AD and AD-by-proxy cohort demonstrated a causal association between genetically predicted household income and AD: the odds ratio (OR) was 0.78 with a 95% confidence interval (CI) between 0.69 and 0.89 (P<0.001; ). Moreover, the effect of household income on the risk of developing AD in weighted median (OR 0.80, 95% CI, 0.68–0.93; P=0.005), simple mode-based (OR 0.71, 95% CI, 0.56–0.91; P=0.027), and weighted mode-based analyses (OR 0.76, 95% CI, 0.59–0.96; P=0.053) demonstrated a similar causal effect as that found with IVW methods, while the result of MR-Egger analysis was in opposition to other methods, maintaining a positive coefficient value ().

Figure 1

Summary Mendelian randomization (MR) estimates of association between household income and Alzheimer’s disease (AD). (A) The causal association between household income and AD was derived from the inverse-variance-weighted, MR-Egger, weighted median, simple mode-based, and weighted mode-based methods. Inverse-variance-weighted, number, odds ratio, and confidence interval. (B) Scatterplot demonstrated single-nucleotide polymorphisms’ potential effects of household income on AD. The slope of each line corresponds to the estimated MR effect per method, and the black dots correspond to the intersection points of the effects of exposure and outcome.

Sensitivity analysis

The results of leave‐one‐out sensitivity analysis are displayed in a forest plot (). All error bars of the meta-analysis with 1 SNP removed are on the left side of the zero line, demonstrating there was no single SNP driving the causal link, and the conclusion was stable. The Cochran’s Q P value was equal to 0.663, suggesting no remarkable heterogeneity of instruments. The MR-Egger intercept test P value was 0.321, and MR-PRESSO identified no outliers with a global test P value of 0.484. These tests provided evidence for the absence of heterogeneity or pleiotropy. The power was 1 with a sample size of 455,258, and the proportion of cases was equal to 0.158.

Figure 2

Leave‐one‐out sensitivity analysis. Forest plot of leave‐one‐out analysis to investigate whether the causal association was driven by a unique single nucleotide polymorphism. All error bars of the meta-analysis with 1 single-nucleotide polymorphism (SNP) removed are on the left side of the zero line.

Discussion

In this study, we used MR analysis with selected genetic instruments to demonstrate that increasing household income had a protective role on AD risk. The results were robust to heterogeneity and pleiotropy. The causal association estimates using weighted median, simple mode-based, and weighted mode-based analyses were concordant with IVW analysis results, which confirmed the result’s robustness to violation of MR assumptions.

Our results were similar to the conclusions of other previous studies in relation to income, where lower household income and occupational social class were associated with higher risk of AD and dementia mortality (8,31,32). However, all prior study results were limited by selection bias and the heterogeneous nature of the comparison groups. The strength of using MR to study the causal associations between income and AD is that instruments are randomly allocated genotypes from birth, which reduces confounding and allows for a more robust causal effect. Moreover, as an unbiased, imprecise estimate, MR is preferable to assess causal association compared with a precise, biased estimate generated by observational studies (33).

The causal associations between income and AD suggest that social status has an important impact on cognitive disease. One possible mechanism underlying this causal effect may be the cognitive reserve hypothesis, according to which individuals with higher education and income levels have learned how to compensate for the disease process and can stave off the effects of brain disease longer (34). This study suggests that people with low SES are more vulnerable to dementia and that more attention should be paid to this vulnerable group to prevent dementia.

The significant results in IVW and other analyses attest to our findings’ robustness, while the adverse result of the MR-Egger approach raise some concerns. Although the MR-Egger method is more sensitive to detecting violations of the instrumental variable assumptions, the influence of strong variants might have biased the results and have inflated type 1 error rates (35). In that case, the results of IVW analysis are more reliable.

Our study has several limitations. First, income was measured at the household level instead of the individual level. However, previous GWAS have shown that household income has a genetic correlation of 0.90 (SE =0.04) with educational attainment measured on an individual level, indicating that household-level effects are likely to be generalizable to individuals (36). Second, participants of AD-by-proxy group are from the UKB, meaning an overlap was present between exposure and outcome. Using the overlap calculation tool for 2-sample MR, bias caused by overlap was equal to zero, and the type I error rate was equal to 0.05 with an overlap rate of 0.8, suggesting no significant bias of the result (37). Third, although we had retrieved literature and applied heterogeneity and pleiotropy tests, the biological effects of these significant SNPs are not available to fully filter out pleiotropy. Fourth, the GWAS data set used was mainly derived from populations of European ancestry to avoid confounding due to population stratification; thus, the present results may not be applicable to other ethnic groups, and further research is needed to better understand how these findings may generalize to other populations. Finally, nonrandom selection into the analytical cohorts of the UKB (38), where individuals with higher income levels and general health are more likely to participate in the follow-up visit, might have biased the results.

Conclusions

In this study, we found robust evidence for the causal association between household income and AD risk. The higher income group demonstrated a reduced risk of developing AD, which is consistent with previous studies. This finding clarifies the effects of income on AD occurrence and may provide potential prevention strategies for vulnerable people.

The article’s supplementary files as