Question: Possible to incorporate the sign or direction of an effect in the IHW covariate?
gravatar for dhibar
3.5 years ago by
dhibar50 wrote:

I am wondering if it is possible to incorporate information about the direction of an effect in the test distribution into a covariate used in IHW?

More specifically, I have a set of DESeq2 results: P-values and Beta statistics for my predictor of interest. Separately I obtained DESeq2 results from a RNAseq experiment with a similar predictor of interest. I want to use the results from the second experiment as a covariate in IHW such that a p-value in the first distribution is weighted differently in the IHW calculation based on its significance in the second distribution. That part seems simple enough following the vignette.

However, it would be interesting to base the IHW weighting on whether or not the direction of effect (upregulated or downregulated) was the same in both experiments (in addition to weighting based on the significance). Is that possible to do in the current IHW implementation?

ihw • 674 views
ADD COMMENTlink modified 3.5 years ago by Nikos Ignatiadis160 • written 3.5 years ago by dhibar50
Answer: Possible to incorporate the sign or direction of an effect in the IHW covariate?
gravatar for Nikos Ignatiadis
3.5 years ago by
Nikos Ignatiadis160 wrote:


this is the first IHW question on the support forums and a good one! Actually there are two distinct parts to your question:

1. Can the direction of the effect be used as a covariate?

It depends on the application, but in many cases you can. In particular, it should be valid (i.e. conditional independence under the null holds) if you are using a symmetric two-sided test for the null hypothesis H:β=0. This is the case for the (two-sided) Wald test employed by DESeq2 (since asymptotically the estimator has a standard normal distribution with mean 0).  So you could use IHW as follows:

# lfc1, pval1, lfc2, pval2: estimated lfc and pvalue for experiment 1, respectively 2
covariate <- as.factor(lfc1*lfc2 > 0) 
ihw_res <- ihw(pval2, covariate, alpha)


2. Can both the "directionality" statistic and the p-values from the previous experiment be used for IHW weighting?

Yes, they can! There's a natural way to combine these to create a new ordered covariate, for which the smoothness (in terms of the total variation penalty employed) is still a reasonable assumption! In particular, you could just mirror the p-values from the first experiment (i.e. calculate 1-pvalue) and the multiply the ones for which the directionality changed between experiments by a -1. I'd expect the weight~covariate curve to be approximately convex with a greater rise on the right side.

covariate <- pmax(1-pval1,0) * sign(lfc1 * lfc2)
ihw_res <- ihw(pval2, covariate, alpha)


Hope this helps!







ADD COMMENTlink modified 3.5 years ago • written 3.5 years ago by Nikos Ignatiadis160

Thanks, this seems to be exactly what I am looking for!

ADD REPLYlink written 3.5 years ago by dhibar50
Please log in to add an answer.


Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
Powered by Biostar version 16.09
Traffic: 393 users visited in the last hour