32- vs 64- Discrepancy

[Tim Olson]

We are running a --fast-epistasis analysis and noticed that the results differ whether we use the 32-bit version or 64-bit version of PLINK 1.9.  The 64-bit version seems to give reasonable results, but the 32-bit version, running on the same data, gives much lower p-values.  We have seen p-values from the 32-bit version as low as 1e-100 which is just crazy.  Note that we are also using the set x set feature, specifying a .set file with two sets of SNPs.  The exe versions in question are:

PLINK v1.90b3c 64-bit (2 Feb 2015)

PLINK v1.90b3c 32-bit (2 Feb 2015)

Unfortunately, we cannot share the data in question, but we have been able to repeat this problem with multiple datasets.

Tim Olson

[Christopher Chang]

Okay, I'll try to reproduce this.  Can you post a sample log file from a bugged 32-bit run, so I know the dataset and set sizes in question, etc.?

[Christopher Chang]

The March 2 development build includes a set-processing bugfix, which may or may not be related to the problem you're observing.

If it does not fix the problem, I'll need some more information to reproduce it.  I tried generating several synthetic datasets large enough to cause 32-bit integer overflows on various dimensions, and I wasn't able to get my 32-bit --fast-epistasis set-by-set results to diverge from my 64-bit results.

On Sunday, March 1, 2015 at 9:09:36 PM UTC-8, Tim Olson wrote:

[Tim Olson]

The dev build 1.90p exhibits the same problem.  I should note that we are running the Windows build not Linux.  I have included the command-line we are using as well as the logs from 32- and 64- bit runs for both the stable and dev builds.  I'm also sharing our SET file since it's not sensitive.  If you still can't reproduce the issue, maybe we can synthesize a dataset which has the problem.

plink.exe --fast-epistasis --bfile 0004b6a4-4557-445f-b72f-2da6200c07a0 --fam 1d186cd0-9512-4a03-82de-d764c36e4d99.fam --set 0004b6a4-4557-445f-b72f-2da6200c07a0.set --out 0004b6a4-4557-445f-b72f-2da6200c07a0

--

You received this message because you are subscribed to the Google Groups "plink2-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to plink2-dev+...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Christopher Chang]

Still haven't been able to reproduce the problem, so yes, a synthetic dataset which exhibits the discrepancy on your end would be helpful.

To unsubscribe from this group and stop receiving emails from it, send an email to plink2-dev+unsubscribe@googlegroups.com.

[Christopher Chang]

Bug should be fixed in the March 5 builds; thanks for the help.

[Tim Olson]

Thank YOU for the help!

To unsubscribe from this group and stop receiving emails from it, send an email to plink2-dev+...@googlegroups.com.