grs.filter.Qrs: Filter SNPs for inclusion in genetic risk score using...
In tobyjohnson/gtx: Genetics ToolboX

Description Usage Arguments Details Value Author(s) Examples

Performs a stepwise downward “model selection” in which SNPs are iteratively removed from the risk score until the heterogeneity test is no longer significant at the specified threshold.

1	grs.filter.Qrs(w, b, s, p.thresh = 0.05)

`w`	coefficients for the risk score
`b`	aligned beta coefficients in the testing dataset
`s`	standard errors
`p.thresh`	P-value threshold

A critical bug in this function was corrected in gtx package version 0.0.10. Output from previous versions should be treated as incorrect. Even with the bug fixed, the function implements a method with no well defined statistical properties.

When there are m SNPs in the risk score, the heterogeneity test is an (m-1) d.f. LRT comparing the 1 d.f. risk score model against the unconstrained m d.f. model, as reported by the Qrs element of grs.summary. At each iteration, if the risk score model has a significant heterogeneity test (at p.thresh), the SNP is removed that gives the greatest decrease in the heterogeneity test statistic.

No guarantee is given about the performance of this procedure. Intuitively, it is expected to work when the majority of SNPs in the risk score only affect the outcome with effects proportional to their weights in the score (including the possibility of zero effects), and a minority of SNPs affect the outcome with non-proportional effects. The working definition of ‘minority’ is unclear.

When using a risk score to make causal inference, the application of this procedure (and indeed of any use of the heterogeneity test) should not be viewed as a replacement for detailed biological knowledge about the mechanisms of action of the causal genetic variants tagged by the SNPs used in the risk score.

Compared with iteratively calculating the heterogeneity test using grs.summary and removing SNPs based on inspection of grs.plot, the grs.filter.Qrs procedure offers only convenience and (a modicum of) objectivity, and nothing more.

Since this heterogeneity test was initially proposed, a number of alternative methods have been described that are likely to be superior, including the MR-Egger approach of:

Bowden, J, Davey Smith, G, and Burgess, S (2015) Mendelian randomization with invalid instruments: Effect estimation and bias detection through Egger regression. Int. J. Epidemiol. 44:512-525 https://dx.doi.org/10.1093/ije/dyv080

A logical vector of the same length as w, set TRUE or FALSE respectively for SNPs included or excluded at the end of the model selection procedure.

Toby Johnson Toby.x.Johnson@gsk.com

  data(magic.scores)
  score1 <- subset(magic.scores, score == "FG2010")
  score1 <- within(score1, okay <- grs.filter.Qrs(coef, beta_TG, se_TG))
  with(score1, {grs.plot(coef, beta_TG, se_TG, locus);
    title(xlab = "FG effect", ylab = "TG effect")})
  with(score1, locus[!okay]) # loci removed
  with(subset(score1, okay), {grs.plot(coef, beta_TG, se_TG, locus);
    title(xlab = "FG effect", ylab = "TG effect")})