samuelklee
commented
4 years ago Just for posterity:
jhess 1:55 PM just to clarify: what is the logic behind approximating σ(τ/min/maj) ≈ (post90 - post10)/2? (edited) 1:55 what is the scale factor of 1/2 for? 1:58 one other thing — how come σ(min/maj) is the sum of the total CR segment’s variance (i.e. σ_τ) and the allelic segment’s variance? 2:00 this would imply that the allelic segments are actually a sum of the random variables corresponding to the allelic and total segmentation, which I’m not sure is the case?
slee 5:32 PM Sorry, just now seeing your questions! 5:33 The scale factor of 1/2 is pretty arbitrary. Just trying to give an estimate of posterior width when the credible interval might be skewed. A better approach would be to refit posteriors with Gaussians/Betas as mentioned previously. 5:35 However, I'm not actually convinced that these credible intervals are what we want to pass to the sigmas. As I also mentioned above, if sigma.tau is supposed to be a global quantity, probably the posterior median of the parameter that controls the global variance (given in the .cr.params file) might be a better thing to use. However, I never got a straight answer from anybody about whether this was a segment-level or global quantity---any idea?
slee 5:41 PM As for using the sum of the CR variance and the MAF variance, you're right---we should be propagating error for the product of the two random variables. Not sure what I was thinking...probably just a brain fart. Not sure if it will make a difference for ABSOLUTE, but thanks for catching that!
As pointed out by Julian, I incorrectly propagated errors in one part of the model translation. Not sure if this has any effect on ABSOLUTE results, but we can fix it up. Probably should re-examine some of the other expressions as well---would be great to finally get feedback on whether these are at all sensible.