About this document

This document forms part of the analysis used in the paper:

Collider bias undermines our understanding of COVID-19 disease risk and severity. Gareth Griffith, Tim T Morris, Matt Tudball, Annie Herbert, Giulia Mancano, Lindsey Pike, Gemma C Sharp, Tom M Palmer, George Davey Smith, Kate Tilling, Luisa Zuccolo, Neil M Davies, Gibran Hemani

It is hosted at https://github.com/MRCIEU/ukbb-covid-collider.

Here we show a set of analyses to illustrate collider bias induced by non-random testing of Covid-19 status amongst the UK Biobank participants, and some approaches to adjust for the bias. The methods are described in further detail in Griffith et al. (2020).

The following variables from the UK biobank phenotype data are used:

  • 34-0.0 - Year of birth (converted into age for this analysis)
  • 31-0.0 - Sex (male = 1, female = 0)
  • 23104-0.0 - Body mass index (BMI)

Also, the linked Covid-19 freeze from 2020-06-05 is used to identify which individuals have been tested and tested positive.

In the analysis that follows, we will be estimating the association between testing positive for Covid-19 and the risk factors age, sex and BMI. The key concern with such an analysis is that we only observe test results among individuals who have received a test. SARS-CoV-2 infection and the risk factors themselves will influence the likelihood of receiving a test, which could induce spurious associations among them when we condition on receiving a test. We will explore inverse probability weighting and sensitivity analyses to address the potential collider bias.

Read in the data


knitr::opts_chunk$set(warning=FALSE, message=FALSE, echo=TRUE, cache=TRUE)
dat <- dat[complete.cases(dat[,c("age","sex","bmi","tested")]), ]