Detecting and adjusting for selection bias within a Mendelian randomization study

PhD project (3/4 yr research project leading to independent research at the doctorate level)

Kate Tilling, Jack Bowden, George Davey Smith


Return to list

Rationale

A major challenge in using Mendelian randomization to analyse data from large cohorts is the issue of non-random selection into the analytical sample. For example, the response rate for recruitment into the UK Biobank study was particularly low (6%) and it is known that individuals with higher than average levels of educational attainment and general health are over-represented in the study [1].
Selection into (or dropout from) a study has the potential to distort the findings of future MR analyses due to collider bias [2]. For example, if the exposure of interest causes dropout then subsequent SNP-exposure estimates used to derive causal effects will will also be biased [3]. In standard observational analyses, non-random sampling is usually adjusted for via inverse probability weighting, and dropout from a study by either inverse probability weighting or multiple imputation. Approaches to selection bias within the Mendelian randomization framework are currently under-developed.

Aims & objectives

This PhD will develop methods for detecting and adjusting for selection and other associated biases within the framework of an MR study. The methods developed will be applied to Mendelian randomization analyses which have external information on selection bias, either via comparison with routine data or via baseline data on subjects who subsequently drop out of the study.

Methods

Specific methodologies to be explored will include multiple imputation, inverse probability weighting, conditional logistic regression [4] and data augmentation [5]. Simulations will be used to compare the different methods, and the situations under which each might be preferable in practice. The methods developed will be applied to the UK Biobank study, and other studies with plausibly different selection mechanisms. The project would suit someone with a strong background in biostatistics or other quantitative discipline, who is interested in developing practical methodological solutions to aid the analysis of epidemiological data.

References

[1] Fry A, et al. AJE 2017; 186(9):1026–1034.
[2] Munafò M, et al. IJE 2017 doi: 10.1093/ije/dyx206.
[3] Hughes R, et al. BioRxiv 2017 https://www.biorxiv.org/content/biorxiv/ea />[4] Breslow NE et al. AJE 1978:108(4): 299–307
[5] Clayton D. Biometrika 2003. 90: 976-981


Created on Nov. 2, 2017, 3:20 p.m.

Edit this page