Non-linear mendelian randomization: detection of biases using negative controls with a focus on BMI, Vitamin D and LDL cholesterol

Abstract

Mendelian randomisation (MR) is an established technique in epidemiological investigation, using the principle of random allocation of genetic variants at conception to generate predicted causal effects of an exposure on an outcome. Extensions to this technique include non-linear approaches, that allow for differential effects of the exposure on the outcome depending on the level of the exposure. One popular method is the residual approach, which allows for stratification of the exposure via the level of the exposure and generation of causal estimates within each stratum. These “local” causal estimates are then used to make inferences about non-linear associations. Recent work has identified that under certain situations this method can lead to estimates that are seriously biased, and a new method - the doubly-ranked method – has been introduced as a possibly more robust approach. In this paper, we perform an analysis using negative control outcomes using both methods and find biased estimates in certain situations. We suggest caution in use of these methods and, if used, the performance of a number of sanity checks such as negative control outcomes in their reporting.

Body mass index and all-cause mortality in HUNT and UK Biobank studies: revised non-linear Mendelian randomization analyses

“…Associations with traits that are competing risk factors in strata of UK Biobank are displayed in Figure 1 . As previously observed in the literature for other exposures 26 , there are clear patterns in genetic associations with age and sex across strata. While there is no overall association of the genetic score with age or sex, there are associations in several of the strata.…”

“…Further, recruitment into UK Biobank is dependent on age, sex, and other factors, which leads to bias in Mendelian randomization estimates. While the effect of moderate selection on Mendelian randomization estimates is often slight 33 , the extent of selection bias may differ between strata, leading to differential bias across stratum-specific estimates 26 . It is likely that age and sex are the strongest predictors of study participants, and differential genetic associations with these traits were observed in UK Biobank.…”

Extensive co-regulation of neighbouring genes complicates the use of eQTLs in target gene prioritisation

“…Rather, we prefer to use MR to exclude exposures that are clearly inconsistent with a causal effect in the tested cell type or tissue. Finally, our results reinforce the need to include both positive and negative controls in MR analysis and use visualisation approaches to assess model fit [28].…”

Investigating the relationship of serum vitamin D levels with blood pressure and hypertension risk in The HUNT Study: Using traditional observational and Mendelian randomization approaches

“…The non-linear association was also suggested in recent meta-analyses of observational studies 6,44 . Nonetheless, there is a concern that the residual method may violate the assumption of constant genetic effects on serum 25(OH)D within each stratum 48 . To address this, the doubly-ranked method was developed to provide less biased estimate even in facing the violation of the constant assumption 42 .…”

“…To address this, the doubly-ranked method was developed to provide less biased estimate even in facing the violation of the constant assumption 42 . However, a recent study suggested that both methods may have yielded biased estimates because the authors found non-null associations between genetic instrument for serum 25(OH)D and age or sex as negative control outcomes within the UK Biobank data, in which the expected results should be null 48 . Hence, the observed inverse association between serum 25(OH)D and SBP within the lowest stratum using both methods in our study might also be due to unknown biases.…”