Mental health as a mediator of the association between educational inequality and cardiovascular disease: A Mendelian randomisation study

Background: Education is inversely associated with cardiovascular disease. Several mediators for this association have been established but a significant proportion of the protective effect remains unaccounted for. Mental health is a proposed mediator, but current evidence is mixed and subject to bias from confounding factors and reverse causation. Mendelian randomisation (MR) is an instrumental variable technique that uses genetic proxies for exposures and mediators to reduce such bias. Methods and Results: We used logistic regression and two-step MR analyses to investigate whether educational attainment affects risk of mental health disorders. We then performed observational and MR mediation analyses to explore whether mental health disorders mediate the association between educational attainment and risk of cardiovascular disease. Higher levels of educational attainment were associated with reduced depression, anxiety and cardiovascular disease in observational analyses [Odds Ratio (95% Confidence interval) 0.79 (0.77-0.81), 0.76 (0.73-0.79) and 0.79 (0.78-0.81) respectively], and MR analyses provided support for these reflecting causal effects [OR (95% CI) 0.72 (0.67-0.77), 0.50 (0.42-0.59) and 0.62 (0.58-0.66) respectively]. Both anxiety and depression were associated with cardiovascular disease in observational analyses [OR (95% CI) 1.63 (1.49-1.79) and OR (95% CI) 1.70 (1.59-1.82) respectively] but only depression was associated in the MR analyses [OR (95% CI) 1.09 (1.03-1.15)]. Roughly 6% of the total protective effect of education on cardiovascular disease was mediated by depression. Conclusions: Higher levels of educational attainment protect against mental health disorders and reduced depression accounts for a small proportion of the total protective effect of education on cardiovascular disease.


Abstract (249 out of 250 words)
Background: Education is inversely associated with cardiovascular disease. Several mediators for this association have been established but a significant proportion of the protective effect remains unaccounted for. Mental health is a proposed mediator, but current evidence is mixed and subject to bias from confounding factors and reverse causation. Mendelian randomisation (MR) is an instrumental variable technique that uses genetic proxies for exposures and mediators to reduce such bias. association between socio-economic inequality and cardiovascular disease is well-7 established, with those living in deprived areas typically having much higher cardio-8 vascular mortality than those in the less deprived areas 2 . More specifically, dispari-9 ties in educational attainment have recently been shown to affect cardiovascular 10 disease using both conventional and genetic epidemiological techniques 3 In the lat-11 ter, cardiovascular risk was estimated to decrease by a third for every 3.6 years of 12 additional full-time education past the age of eleven 4 . 13 14 A number of well-recognised cardiovascular risk factors appear to act as mediators 15 for education 5 . For example, low educational attainment is associated with in-16 creased tobacco smoking, higher BMI, and higher blood pressure which in turn are 17 linked to increased risk of cardiovascular disease 6 . However, even taking these me-18 diators into account, there is a significant proportion of the association between ed-19 ucation and cardiovascular disease that remains unaccounted for. 20 21 Existing evidence suggests an association between mental health and cardiovascu-22 lar disease as well as an association between education and mental health 7,8 . Con-23 sequently, mental health may be another potential mediator between education and 24 . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.

Methods and Results
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09. 10.20191825 doi: medRxiv preprint Here, we specifically used two-step MR using summary data for each step. Two-1 step MR measures the exposure-mediator, exposure-outcome and mediator-out-2 come effects separately and is particularly useful when binary mediators are being 3 measured individually. We utilised separate samples for the instrument and out-4 come summary data, thus providing greater sample sizes in some cases.  is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09.10.20191825 doi: medRxiv preprint 1 match between reported and chromosomal sex. We then applied the Medical Re-2 search Council Integrative Epidemiology Unit quality control procedure in order to 3 restrict the sample to those of European ancestry 18 . The final sample size with com-4 plete phenotype measures was 333,525 with a mean age of 56.9 years (standard 5 deviation [SD] = 8) and 54% females.  Supplementary Table S1). Mean educational 12 attainment was 13.9 years (SD = 5.1). 13 problem if they reported "Yes" to either of the routine survey questions "Have you 16 ever seen a general practitioner (GP) for nerves, anxiety, tension or depression?" or 17 "Have you ever seen a psychiatrist for nerves, anxiety, tension or depression?". is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09. 10.20191825 doi: medRxiv preprint The prevalence of these mediators in our sample was 34% for mental health prob-1 lems, 2.9% for depression, and 1.4% for anxiety. if their health records showed at least one diagnosis of angina or myocardial infarc-5 tion prior to recruitment classified by ICD codes I20 and I21 respectively. The prev-6 alence of cardiovascular disease in the sample at the time of recruitment was 6%.

Genetic instruments 15
Educational attainment: 1271 independent genome-wide significant loci (p<5x10 -8 ) 16 were identified in a GWAS meta-analysis of educational attainment by Lee and col-17 leagues (N = 1,131,881) 19 . Educational attainment was measured as years in educa-  is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09.10.20191825 doi: medRxiv preprint Mental health problems: 14 independent genome-wide significant loci were identi-1 fied in a recent GWAS of a phenotype described as "broad depression" in UK Bi-2 obank data (113,769 cases and 208,811) 20 . Here, we report this phenotype as "Hav-3 ing had a mental health problem" instead. Individuals were classified as having had 4 a mental health problem if they reported "Yes" to either of the routine survey ques-5 tions "Have you ever seen a general practitioner (GP) for nerves, anxiety, tension or 6 depression?" or "Have you ever seen a psychiatrist for nerves, anxiety, tension or 7 depression?". Cases with diagnosed bipolar disorder or schizophrenia or who were 8 taking anti-psychotic medications were excluded. A number of participants were 9 also excluded by inter-relatedness filtering. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09.10.20191825 doi: medRxiv preprint and identified 108 independent significant loci following LD clumping (r2<0.001, dis-1 tance>10000 kb). 2 3 Cardiovascular disease: We used summary data from a recent GWAS meta-analysis 4 from the CardioGRAMplusC4D consortium 23 . This included 48 cohorts with a total 5 of 60,801 cases and 123,504 controls. The cohort studies typically defined a case 6 as having a diagnosis of myocardial infarction or angina with several also requiring 7 confirmation from angiographic evidence. 8 9

III. Statistical analysis 10
All analyses were performed using R (Version 3.6.2) 24 . A directed acyclic graph for 11 the proposed causal relationships investigated is shown in Figure 1. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09.10.20191825 doi: medRxiv preprint The proportion mediated statistic was calculated using the product of coefficients 1 method as detailed in previous work 25 . The effect sizes were calculated using abso-2 lute risk differences as estimated from the odds ratios from our regression analysis 3 and the known prevalence figures in UK Biobank. Standard errors were calculated 4 using the delta method. 5 6 MR univariable analysis: The univariable analysis consisted of using two-sample MR 7 methods applied in the TwoSample MR package 0.4.18 26 . We first ran our analysis 8 using educational attainment as an exposure and each of our proposed mediators 9 and cardiovascular disease as outcomes. We then repeated the analysis using each 10 mediator as an exposure and cardiovascular disease as an outcome. When using the anxiety disorder instrument, we performed MR Robust Associated 18 Profile Score (MR RAPS) instead of IVW to improve estimation whilst using a re-19 laxed p-value threshold for the instrument 31 . MR RAPS offers a greater robustness 20 to pleiotropy when using many weak instruments, as we do here. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . Two-step MR analysis: We next performed two-step MR using the TwoSample MR 1 package 0.4.18. We used the results of our univariable MR analysis for the expo-2 sure-mediator step. For the mediator-outcome step we then used the residual mul-3 tivariable MR method to calculate the effect of each mediator on cardiovascular dis-4 ease having adjusted for the effects of educational attainment 32 . The estimates for 5 each step were multiplied together to produce an indirect effect estimate and the 6 proportion mediated statistic was then calculated as for the observational analysis 7 (detailed above). is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. We found strong evidence to suggest that educational attainment has an inverse 7 association with cardiovascular disease (see Figure 2). In our observational analysis, 8 for every standard deviation increase in educational attainment there was a 24% 9 (95% CI 21-27%) reduction in cardiovascular disease. Our MR analysis suggested a 10 causal, protective effect for education against cardiovascular disease. 11 12

Association of education with mental health 13
We also found that education has an inverse association with anxiety and depres-14 sion and that this is consistent across both adjusted observational and MR analyses 15 (see Figure 2). For each standard deviation increase in educational attainment there 16 were 25% (95% CI 24-26%) and 21% (95% CI 19-23%) reductions in depression 17 and anxiety respectively. Our MR analysis suggested that education has a causal, 18 protective effect against anxiety and depression. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020.  is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09.10.20191825 doi: medRxiv preprint 5.6%-7.5%) account for a small proportion of the negative association between ed-1 ucation and cardiovascular disease (see Figure 4). The proportion mediated by anxi-2 ety disorder was very small in the observational analysis, and we found little evi-3 dence of mediation in our MR analysis.

Sensitivity analyses 8
The results of our sensitivity analyses are presented in Supplementary Tables S3-13  9 and included: testing for pleiotropy and reverse causation, testing for bidirectionality 10 in our observational data, and testing personality traits as additional potential medi-11 ators. The results provided little evidence of bias from directional pleiotropy but did 12 raise the possibility of reverse causation between cardiovascular disease and the 13 proposed mediators in our observational analyses. 14 15

Exploratory analysis of tobacco smoking as a mediator between depression and car-16 diovascular disease 17
In a further analysis, we sought to investigate a potential role for multiple mediators 18 in this pathway, specifically smoking as a mediator between adjusted depression 19 and cardiovascular disease. We found that depression is associated with increased 20 smoking activity and that lifetime smoking was associated with increased cardio-21 vascular risk (see Supplementary Tables S15-16). In Figure 5, we show that smok-22 ing accounts for a large proportion of the increased cardiovascular risk associated 23 with depression, 29% (95% CI 28.9-29.7%) according to our MR analysis. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. In this study, we used both observational and MR techniques to investigate mental 5 health as a mediator between education and cardiovascular disease. Our results 6 demonstrate that depression accounts for a proportion of the protective effect edu-7 cation has on cardiovascular disease. 8 9 We demonstrated, in observational analyses, that educational attainment has a 10 clear, inverse association with anxiety disorder, major depressive disorder, and car-11 diovascular disease. Our MR analyses suggested these relationships are causal. In is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09.10.20191825 doi: medRxiv preprint of additional mediators (see Figure 1). Our observational analyses showed that de-1 pression was positively associated with lifetime smoking which was, in turn, associ-2 ated with increased cardiovascular risk even after adjustment for depression. Our 3 MR analyses suggested these associations are causal. We found that lifetime smok-4 ing accounts for as much as 30% of the association between depression and cardi-5 ovascular disease. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. tional years in education may also help to offset exposure to stressors until one has 10 greater maturity and resources to cope with them. By considering these observa-11 tions, successful individual-level interventions in the education system might be de-12

veloped. 13 14
In an exploratory analysis, we found that smoking accounts for a large proportion of 15 the association between depression and cardiovascular disease. The link between 16 smoking and cardiovascular disease is well-known and depression and smoking are 17 also associated 42,43 . This link between smoking and depression has been commonly 18 attributed to a 'self-medication' hypothesis where smoking is used to cope with de-19 pressive symptoms. Other factors that are proposed to partially explain the remain-20 der of the association between depression and cardiovascular disease include 21 chronic inflammation and changes in lipid metabolism 44 . is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09.10.20191825 doi: medRxiv preprint

Implications 1
Our study has several important public health implications. First, it strengthens the 2 case for investment in education and interventions related to it by demonstrating 3 that educational attainment can reduce mental health burden, which is itself a sig-4 nificant source of morbidity and mortality 45 . Second, by illustrating the links between 5 education, mental health, and cardiovascular disease, we have reinforced the need 6 to recognise the potential interplay between social determinants, physical and men-7 tal health problems. Indeed, in this study, we show how a physical disease burden 8 could be partly alleviated by targeting a social determinant of mental health. Finally, 9 we have also illustrated how to further explore potential mediators between depres-10 sion and cardiovascular disease in order to elucidate additional intervention targets. 11 Here, we have highlighted tobacco smoking as such a target. 12 13

Strengths and limitations 14
The strengths of our study include the use of MR analysis to complement a conven-15 tional observational approach which allowed us to better overcome bias arising 16 from confounding, measurement error and reverse causation. The use of a 'triangu-17 lation approach' also allowed for greater confidence in causal inference when there 18 was consistency between the different approaches 46 . We were also able to limit bias 19 arising from self-reporting and misclassification by using strict case definition for 20 both our observational and MR data sources. Such bias can be particularly prob-21 lematic when using binary exposures in MR 47 . An additional strength of our study 22 was the use of two-step MR for a mediation analysis. This allowed us to specifically 23 . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09.10.20191825 doi: medRxiv preprint quantify the proportion of our exposure's effect attributable to our proposed media-1 tors. Finally, we extended our analysis in order to illustrate how our methodology 2 could explore multiple mediators in the proposed pathway. 3 4 In our study, the use of MR demonstrated a discrepancy when observing the effects 5 of mental health on cardiovascular disease. When adjusting for confounding in our 6 mediation analysis, we likely over-adjusted in order to confidently isolate the effects 7 of the mediator. To measure this, we conducted a sensitivity analysis without ad-8 justment although this suggested that the effect sizes we have found may actually 9 be underestimates (see Supplementary Table 17). Instead, the discrepancy may 10 have primarily been due to reverse causation, with cardiovascular disease leading 11 to mental health problems and exaggerating observed associations. In a further 12 sensitivity analysis, where cardiovascular disease was the exposure and depression 13 and then anxiety the outcomes, we did indeed find such bidirectionality. old; the null result should therefore be interpreted with caution. However, we utilised 20 the MR RAPS method to compensate for the relaxed threshold and we also con- is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . and a larger sample size but also produced a null result (see Supplementary Table  1 S8). 2 3 Despite adequate instrument strength, our 'mental health problems' instrument pro-4 duced wide confidence intervals, suggesting these analyses may have been under-5 powered. Unlike our other instruments, this one was constructed from a GWAS per-6 formed on a single sample, as the unusual phenotype measured is only available in 7 the UK Biobank. In future, we hope other cohorts will have this or a similar pheno-8 type measured, leading to larger sample sizes for analysis and, hopefully, the dis-9 covery of further significant variants. 10

11
When conducting two-sample MR, overlap between the samples can be problem-12 atic. The samples used for our exposure and mediators did feature partial overlap 13 due to the use of UK Biobank data in both sets. This introduces a greater risk of a 14 type I error in our two-sample univariable analyses of education and mental health 15 due to possible overfitting. However, we also saw an inverse association with edu-16 cation in our sensitivity analyses that utilised personality traits related to mental 17 health (see Supplementary Table S8). These featured substantively less overlap be-18 tween samples. In addition, this limitation did not apply to our two-step analysis 19 which pools the data for the exposure and mediators before analysis is conducted. 20 Another common limitation of MR studies is bias from directional pleiotropy 49 . We is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . Supplementary Tables S3 and S7), which 1 suggest our results were not biased by directional pleiotropy. 2 3 Finally, even with our approach, education may still be acting as a proxy for socio-4 economic position. A proposed strategy for investigating this would be to conduct 5 an additional analysis using a measure such as household income which would al-6

measurement of MR Egger intercepts (see
low for comparison as well as adjustment as part of a multivariable MR approach. A 7 recently published GWAS meta-analysis exploring household income 50 offers the 8 possibility of this in future studies. Natural experiment techniques could also be 9 used having previously suggested that the raising of school age reduced cardiovas-10 cular disease 51 . 11 12

Conclusions 13
In conclusion, we have demonstrated that educational attainment is inversely asso-14 ciated with mental health problems, anxiety, depression, and cardiovascular dis-15 ease. A reduction in depression accounts for part of the inverse association be-16 tween education and cardiovascular disease. Finally, we have shown that smoking 17 accounts for a large proportion of the association between depression and cardio-18 vascular disease. Our findings offer significant clinical and public health implications 19 by further demonstrating interplay between social determinants, mental and physi-20 cal disease and associated targets for intervention. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09.10.20191825 doi: medRxiv preprint Acknowledgements 1 We are grateful to the participants of the UK Biobank and the individuals who 2 contributed to each of the previous GWAS analyses conducted as well as all the 3 research staff who worked on the data collection. This research has been 4 conducted using the UK Biobank Resource under Application Number 9142. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. .

Supplementary Methods 2
Supplementary Tables S1-S15 3 Supplementary Figures S16-S17 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

Figures and Figure Legends
The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09.10.20191825 doi: medRxiv preprint Figure 2 The association of educational attainment with mental health and cardiovascular disease presented as odds ratios per standard deviation increase in education . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09.10.20191825 doi: medRxiv preprint Figure 3 The association of mental health with cardiovascular disease following adjustment for the effects of educational attainment . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09.10.20191825 doi: medRxiv preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09.10.20191825 doi: medRxiv preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 11, 2020. . https://doi.org/10.1101/2020.09.10.20191825 doi: medRxiv preprint