Closed Qinqiword closed 4 years ago
I thought I had fixed this, but apparently it seems like it could be caused by a a bug (or weird feature) in plinkio. I committed a fix on the master branch (v. 1.0.9). Would you mind verifying that it works?
Best, Bjarni
I'll close this now, as I believe this is fixed.
Hi Bjarni,
I am unfortunately still getting the same inverse of results (higher PRS with lower disease prevalence). Has this been fixed?
With these apparently inverted results, is it appropriate to take the inverse of the PRS (e.g. make the PRS negative to reverse the order), which strikes me as similar to switching the effect and the reference allele and consequently inversing the odds ratio?
Thanks
J
On Mon, Oct 21, 2019 at 2:43 AM Bjarni J. Vilhjalmsson < notifications@github.com> wrote:
I thought I had fixed this, but apparently it seems like it could be caused by a a bug (or weird feature) in plinkio. I committed a fix on the master branch. Would you mind verifying that it works?
Best, Bjarni
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/bvilhjal/ldpred/issues/120?email_source=notifications&email_token=AILQP6NOPQOTISUBCUD26ELQPV2VFA5CNFSM4JB3VCG2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBZXDNA#issuecomment-544436660, or unsubscribe https://github.com/notifications/unsubscribe-auth/AILQP6MHJCNC3IVS53PDJD3QPV2VFANCNFSM4JB3VCGQ .
I think it hasn't been fixed. I am using v.10 and having the same issue
Hi bvilhjal,
I am using LDpred to calculate PRS for a disease, recently. The ss_file is from one published GWAS study (containing 191764 individuals and 8441806 SNPs), and the LD-reference file and the validation file are the same genotype file, containing 2638 individuals (1542 case samples and 1096 control samples) and 18058448 SNPs, ss_file and gf have the same ancestry.
Using these files as input, I got the PRS using three models: ldpred_inf, ldpred and P+T. But when I checked the PRS of the three runs, I found that in the top 10% samples with the highest PRS, only 30% individuals have that disease, but in the bottom 10% samples with the lowest PRS, 78% individuals have that disease. The p-value of Chi-square test was significant. This is very strange; individuals with higher PRS are more likely to be healthy, but individuals with a lower PRS are more likely to have that disease.
In another try, I selected 88 SNPs from the ss_file using p-value < 5×10-8 to calculate PRS in the same population by the traditional formula: PRS = b1x1+b2x2+ ....+bkxk+bnxn. In the 88_PRS, among the top 10% samples with the highest PRS, 72% were patients, and 28% were healthy. That is expected.
It is weird that this two result were converse. I checked the input files and code but failed to find the reasons. One thing I suspected is the following lines in the log file. “PRS correlation” and “The slope for predictions with weighted effects” are both negative.
Here are my questions:
Here are my scripts and log.file:
And the log file.
Any comments or suggestions would be appreciated.
Best wishes, Jing