P26 Can genetic instrumental variables improve causal estimates of the income-health relationship? A two-sample Mendelian randomisation study

Background People with lower incomes tend to have poorer health outcomes, but the extent to which this reflects a causal relationship between income and health is unclear. This relationship has policy relevance, but is challenging to ascertain in observational data due to substantial risk of confounding. Mendelian randomisation (MR) exploits natural genetic variation as a random assignment mechanism to test causal hypotheses. MR has previously been used for ‘biologically distal’ exposures such as educational attainment, and recent MR studies have incorporated income as a secondary exposure or mediator. However, it is unclear how the income-health associations of these studies should be interpreted and what mechanisms and possible biases are involved. Methods We used two-sample MR and published summary genetic data from multiple European-ancestry samples to estimate the effect of log earned income on mortality, mental health (depression, anxiety, and subjective wellbeing), and other risk factors (smoking, alcohol consumption, and body mass index). We investigated sources of bias such as competing pathways between genetic instruments and health outcomes (i.e. pleiotropy), confounding by demographic or intergenerational factors that influence genetics, and reverse causation (i.e., the effect of health on income). Results In the main MR analysis, 10% higher income was associated with lower risk of death (OR 0.93 [0.90, 0.96]), depression (OR 0.96 [95% CI 0.93, 0.98]), anxiety (OR 0.96 [0.92, 0.99]), ever-smoking (OR 0.94 [0.92, 0.95]), and lower BMI (−0.03 SD [−0.04, −0.01]). We found no association with alcohol consumption (−0.001 SD [−0.01, 0.01]) or subjective wellbeing (0.005 SD [−0.01, 0.02]). Results were robust to conventional tests of pleiotropy and reverse causation, and most remained after controlling for education in multivariable MR, except death and smoking. However, non-null results for negative control outcomes (birthweight, 0.02 SD [0.00, 0.03], and childhood asthma, OR 0.90 [0.83, 0.98]) indicate likely residual bias. Within-family analyses were inconclusive due to lack of power. Conclusion MR estimates of the effect of income on many but not all health outcomes capture a pathway that is at least partly distinct from the effects of education. However, it is unclear how much of the apparent effect is due to bias, particularly confounding by demographic and intergenerational factors, which our underpowered within-family analyses were unable to fully capture. Assuming biases are mostly away from the null, our results may reflect an upper bound. Therefore, our findings suggest causation explains only a small part of the income-health gradient.