Simultaneous estimation of bi-directional causal effects and heritable confounding from GWAS summary statistics

Abstract

Mendelian Randomisation (MR), an increasingly popular method that estimates the causal effects of risk factors on complex human traits, has seen several extensions that relax its basic assumptions. However, most of these extensions suffer from two major limitations; their under-exploitation of genome-wide markers, and sensitivity to the presence of a heritable confounder of the exposure-outcome relationship. To overcome these limitations, we propose a Latent Heritable Confounder MR (LHC-MR) method applicable to association summary statistics, which estimates bi-directional causal effects, direct heritability, and confounder effects while accounting for sample overlap. We demonstrate that LHC-MR outperforms several existing MR methods in a wide range of simulation settings and apply it to summary statistics of 13 complex traits. Besides several concordant results, LHC-MR unravelled new mechanisms (how being diagnosed for certain diseases might lead to improved lifestyle) and revealed potential false positive findings of standard MR methods (apparent causal effect of body mass index on educational attainment may be driven by a strong ignored confounder). Phenome-wide search to identify LHC-implied heritable confounders showed remarkable agreement between the LHC-estimated causal effects of the latent confounder and those for the potentially identified ones. Finally, LHC-MR naturally decomposes genetic correlation to causal effect-driven and confounder-driven contributions, demonstrating that the genetic correlation between systolic blood pressure and diabetes is predominantly confounder-driven.

Composite trait Mendelian Randomization reveals distinct metabolic and lifestyle consequences of differences in body shape

“…The decreased consumption of wine and alcohol taken with meals is also consistent with lower SES 41 . These results are consistent with the well-known observational association between obesity and low SES 42 , however recent results suggest that this is due to residual population stratification and non-genetic familial effects 43,44 .…”

“…The lack of change in smoking habits supports the hypothesis of the increased tobacco smoking observed for PC2 being used as a means to lose weight 37 . In contrast, the changes in alcohol consumption and SES are directionally consistent with those observed for PC2, suggesting these may be affected by similar causal pathways or subject to the same biases from population stratification or non-genetic familial effects 43,44 .…”

“…GWAS effect estimates themselves have been shown to be biased due to population effects, such as parental effects, population stratification, and assortative mating 52 . Although such biases in the exposure would not violate the assumptions of MR, their presence in the outcome and any non-genetic effects shared between exposure and outcome may affect MR causal effect estimates 43,44 .…”

Guidelines for performing Mendelian randomization investigations

“…If the genetic associations with the outcome do not follow this pattern, then a non-causal explanation would be preferred. Emerging methods that take this approach include the Causal Analyses Using Summary Effect Estimates (CAUSE) 76 and Latent Heritable Confounder Mendelian randomization (LHC-MR) 77 methods.…”

Genetic insights into the causal relationship between physical activity and cognitive functioning

“…The present study applies the Latent Heritable Confounder MR (LHC-MR) method (41), which addresses the aforementioned limitations, to simultaneously estimate the bidirectional causal effects between physical activity and cognitive functioning, while accounting for possible heritable confounders of their relationship (e.g., body mass index, educational attainment, level of physical activity at work, and material deprivation). Unlike standard MR, LHC-MR accounts for sample overlap from genome-wide genetic instruments, thereby allowing the exposure and the outcome to originate either from overlapping datasets or the same dataset.…”