jgscott / fdrreg Goto Github PK

Hi,

I'm trying to run FDRreg with default parameters except for nulltype='empirical'. However, I get many warnings like the following:

f(z) misfit = -0.2. Rerun with increased df.

I think I have tracked this down to the efron function call. It seems to be an issue with the fitting of the empirical Z-score distribution. The warning suggests to run with an increased degrees of freedom, which is taken from the control input list (specifically, the densknots parameter, default value is 10). I tried to get advice for how much to increase this parameter from the documentation, but was a bit confused by the help file for FDRreg, which says:

nmids, densknots: Only used if method='efron'; this functionality is retained for legacy reasons and is not recommended for actual data analysis. nmids: How many histogram midpoints are used in Efron's spline-based estimator of f(z). Defaults to 150. densknots: how many spline knots. Defaults to 10.

Is the 'empirical' null type not recommended for actual data analysis? If it is, do you have any practical advice for how to set this densknots parameter?

Thanks in advance,
Keegan

Hi, I tried to install FDRreg v0.2 using devtools::install_github('jgscott/FDRreg', subdir="R_pkg/"), but I got the following error.

Error: HTTP error 404.
Not Found

Did you spell the repo owner (jgscott) and repo name (FDRreg) correctly?

• If spelling is correct, check that you have the required permissions to access the repo.

Thanks!
Ling

Hi,

I am running into the error M0 > 0 are not all TRUE using nulltype = "empirical" with the additional warning Warning message: In sqrt(-1/{ : NaNs produced. Looks like this is thrown from the following line of the efron() function:

sig = as.numeric(sqrt(-1/{ 2 * b0[3] }))

I’m not sure what this line is doing, or why it expects that b0[3] is negative, so I'm not sure how to fix.

Here’s a toy reproducible example where it occurs - when the test statistics look a bit bimodal (but I’m not sure if this is the only type of scenario):

set.seed(57)
res <- FDRreg::FDRreg(c(rnorm(500,1,0.8), rnorm(500,-1,0.8)), 
               features = model.matrix( ~  splines::bs(rnorm(1000), df = 3) - 1),
               nulltype = 'empirical’)

Please let me know if this can be resolved. Thanks!

Best,
Keegan

Hi,

Just a heads up - I was just trying to install FDRreg and was unable due to the dependency on the package BayesLogit which was removed from CRAN (https://cran.r-project.org/web/packages/BayesLogit/index.html).

Best,
Keegan

I am hoping to use the results from a previous test (Z1) to re-weight the p-values from a new-related test (Z2). However, when I run FDRreg (as shown below), I end up with FDRreg-corrected results that do not seem to incorporate the information from the covariate Z1 distribution. For example, a result that was significant at Q < 0.05 had a very large Z stat in Z2 and a large but NEGATIVE Z stat in Z1. In fact, it seems like the Q-values are in the same order as the original p-values from Z2. This leads me to believe that the info from Z1 isn't being incorporated properly.

The main file is attached:

library(FDRreg)

dat=read.table("independent_aligned_effects.txt",header=T)

#convert p to z
dat$Z1 = (1 - qnorm(dat$P1)) * sign(dat$BETA1)
dat$Z2 = (1 - qnorm(dat$P2)) * sign(dat$BETA2)

dat2=dat[complete.cases(dat),]

fdr1 = FDRreg(dat2$Z2, as.matrix(dat2$Z1), nulltype="theoretical")

#get the subset of significant SNPs
Q=0.05
fdrout=fdr1$FDR
tmp=cbind(dat2,fdrout)
outmat=subset(tmp, fdrout < Q)
nrow(outmat) #8403

write.table(outmat, file="fdrreg_corrected.txt", quote=F, row.names=F)

Many of the test declared significant at Q have Z-scores in the opposite direction. For example:

subset(outmat, SNP == "rs35238445", select=c(Z1,Z2,fdrout))

          Z1       Z2      fdrout

33471 -0.3448948 5.183861 0.001698456

sessionInfo()
R version 3.2.4 (2016-03-10)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Red Hat Enterprise Linux Server release 6.3 (Santiago)

locale:
[1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C
[3] LC_TIME=en_US.UTF-8 LC_COLLATE=en_US.UTF-8
[5] LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8
[7] LC_PAPER=en_US.UTF-8 LC_NAME=C
[9] LC_ADDRESS=C LC_TELEPHONE=C
[11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C

attached base packages:
[1] splines stats graphics grDevices utils datasets methods
[8] base

other attached packages:
[1] FDRreg_0.2-1 mvtnorm_1.0-5 BayesLogit_0.5.1 fda_2.4.4
[5] Matrix_1.2-4

loaded via a namespace (and not attached):
[1] Rcpp_0.12.4 magrittr_1.5 MASS_7.3-45 munsell_0.4.3
[5] colorspace_1.2-6 lattice_0.20-33 R6_2.1.2 ggdendro_0.1-18
[9] minqa_1.2.4 mosaicData_0.13.0 stringr_1.0.0 car_2.1-2

independent_aligned_effects.txt