Adjustment for index event bias in genome-wide association studies of subsequent events, 2019, Dudbridge et al

T

tralfamadorian97

Established Member (Voting Rights)
Citation: Dudbridge, Frank, et al. "Adjustment for index event bias in genome-wide association studies of subsequent events." Nature communications 10.1 (2019): 1561.
Authors: Frank Dudbridge, Richard J. Allen, Nuala A. Sheehan, A. Floriaan Schmidt, James C. Lee,. Gisli Jenkins , Louise V. Wain , Aroon D. Hingorani & Riyaz S. Patel
Abstract:
Following numerous genome-wide association studies of disease susceptibility, there is increasing interest in genetic associations with prognosis, survival or other subsequent events. Such associations are vulnerable to index event bias, by which selection of subjects according to disease status creates biased associations if common causes of incidence and prognosis are not accounted for. We propose an adjustment for index event bias using the residuals from the regression of genetic effects on prognosis on genetic effects on incidence. Our approach eliminates this bias when direct genetic effects on incidence and prognosis are independent, and otherwise reduces bias in realistic situations. In a study of idiopathic pulmonary fibrosis, we reverse a paradoxical association of the strong susceptibility gene MUC5B with increased survival, suggesting instead a significant association with decreased survival. In re-analysis of a study of Crohn’s disease prognosis, four regions remain associated at genome-wide significance but with increased standard errors.

Link
 
My Summary

This paper discusses two topics:

Topic 1: The risk of collider bias in conditional GWAS.

Topic 2: A method to mitigate collider bias.

While most of the paper is devoted to Topic 2, I am more interested in Topic 1.

What is a conditional GWAS?

A conditional GWAS is a GWAS in which we

  • First select participants according to some phenotypic criteria called I
  • Among these participants meeting criteria I, we run a GWAS of a second phenotype P.
Here are some examples of conditional GWAS:

  • A GWAS of prognosis among participants with idiopathic pulmonary fibrosis (IPF). Here I is IPF incidence, and P is IPF prognosis.
  • A GWAS of breast cancer mortality among participants with breast cancer. Here I is breast cancer incidence, and P is breast cancer mortality. (This example comes from a follow-up paper)
What is collider bias?

The term “collider bias” comes from the causal inference literature. Roughly speaking, it refers to spurious associations produced by selection or conditioning.

A classic toy example found in many causal inference textbooks and on wikipedia is the following:

Suppose that acting talent and attractiveness are uncorrelated in the general population. Suppose that to be a celebrity, one needs either acting talent or attractiveness. If we focus on only celebrities, acting talent will appear to be negatively correlated with attractiveness. The first graph below shows the lack of correlation in the general population. The second graph shows the negative correlation among celebrities.

1783366810978.webp

How does collider bias arise in conditional GWAS?

Roughly speaking, the selection criteria I may produce distorted associations between genetic variants and the phenotype P.

How is collider bias dealt with?

Special statistical techniques are used to estimate the effect of collider bias and correct for it. However, this correction is not straightforward. There are competing correction approaches that make different assumptions. Several later papers cite the current paper and claim to improve on its correction method.

How large is the typical distortion produced by collider bias in conditional GWAS?

The paper claims that in most GWAS, collider bias distortion will typically be small. However, there are exceptions.

What are some examples of collider bias in conditional GWAS?

  • In the idiopathic pulmonary fibrosis example, a variant affecting the gene MUC5B appears to be a risk factor for incidence, but protective against worse prognosis. After correction for collider bias, the effects become concordant: the variant is a risk factor for both incidence and for worse prognosis.
  • In the breast cancer example above, the variant rs35054928 is strongly protective against breast cancer, but appears to be a risk factor for breast cancer mortality in the conditional GWAS. After correction for collider bias, rs35054928 is protective both against breast cancer and breast cancer mortality.


Note that in both these examples the variant of interest has a weak to moderate effect in the uncorrected conditional GWAS. I have not seen an example in which a variant with a strong, significant effect in the uncorrected conditional GWAS has its sign reversed by collider-bias correction.
 
Last edited:
tralfamadorian97 said:
Note, however, that in both these example the variant of interest has a moderate or high p value in the uncorrected conditional GWAS. I have not seen an example in which a variant with a small uncorrected conditional p value has its sign reversed by collider-bias correction.
Click to expand...
This is very interesting.

Just to make sure I’m understanding this correctly: a small p value would be a value further away from 1 (i.e. closer to 0), and a moderate or high p value would be something closer to 1, relatively speaking of course.
 
You must log in or register to reply here.
Back
Top Bottom