Flipped alleles and phantom power

[soren-rand]

Hi Patrick,

I am writing to get your thoughts on a behavior I observed while running MTAG on aortic stenosis (AS) and several highly correlated quantitative traits (rg range 0.6–0.9). I inadvertently ran MTAG twice. Once with an allele error and once corrected. I observed a counter-intuitive difference in the gain in effective sample size.

In the first instance, I had been a fool and misaligned alleles:
In the first run, the REF/ALT alleles for the binary trait (AS) were flipped. Genetic correlations were negative (opposite of the expected biological direction), but MTAG reported a massive increase in power (Neff increase of 198%)

In the second instance, I corrected my foolish behavior:
I corrected the column alignment so alleles and effect sizes were properly oriented. The genetic correlation signs resolved to the correct direction, but the increase in power waaaay lower than the first run (Neff increase of 130%).

I briefly discussed this with Raymond Walters. My hypothesis is that in the first instance, the misalignment caused MTAG to misinterpret the covariance structure, leading to spurious inflation that manifested as better discovery, whereas the second run would represent the true biological signal...
Could you confirm if this is the expected under-the-hood behavior? Specifically, does an allele flip cause MTAG to overestimate the signal-to-noise ratio in this manner?

Thanks for your time and for the tool!

Best,
Soren

Log from the first run:
flipped_as_pv_ava.log
Log from the second run:
as_pv_ava.log

[paturley]

Hmm. I find this very puzzling. If all you did was swap the ref and alt allele, the MTAG method should account for the sign flip and the power increase should be the same (though all your betas will have the wrong sign for the misaligned trait). The only thing I can think of is if there is a munge step where it compares the allele frequency of your summary statistics to that of a reference panel. In such a case, misaligning the ref and alt allele would cause SNPs with allele frequencies distant from 0.5 to be dropped since those SNPs are definitely wrong. But looking at your log file, that would lead to a very different number of SNPs remaining at the end, and they look pretty similar. The fact that they aren't identical is strange to me though. If I were you, I'd look through the log file and figure out at which step the SNPs get dropped in one analysis and not the other. Then see if there is anything unusual about those SNPs.

…

On Mon, Nov 24, 2025 at 5:01 PM soren-rand ***@***.***> wrote: *soren-rand* created an issue (JonJala/mtag#245) <#245> Hi Patrick, I am writing to get your thoughts on a behavior I observed while running MTAG on aortic stenosis (AS) and several highly correlated quantitative traits (rg range 0.6–0.9). I inadvertently ran MTAG twice. Once with an allele error and once corrected. I observed a counter-intuitive difference in the gain in effective sample size. *In the first instance, I had been a fool and misaligned alleles:* In the first run, the REF/ALT alleles for the binary trait (AS) were flipped. Genetic correlations were negative (opposite of the expected biological direction), but MTAG reported a massive increase in power (Neff increase of 198%) *In the second instance, I corrected my foolish behavior:* I corrected the column alignment so alleles and effect sizes were properly oriented. The genetic correlation signs resolved to the correct direction, but the increase in power waaaay lower than the first run (Neff increase of 130%). I briefly discussed this with Raymond Walters. My hypothesis is that in the first instance, the misalignment caused MTAG to misinterpret the covariance structure, leading to spurious inflation that manifested as better discovery, whereas the second run would represent the true biological signal... Could you confirm if this is the expected under-the-hood behavior? Specifically, does an allele flip cause MTAG to overestimate the signal-to-noise ratio in this manner? Thanks for your time and for the tool! Best, Soren Log from the first run: flipped_as_pv_ava.log <https://github.com/user-attachments/files/23732815/flipped_as_pv_ava.log> Log from the second run: as_pv_ava.log <https://github.com/user-attachments/files/23732814/as_pv_ava.log> — Reply to this email directly, view it on GitHub <#245>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFBUB5OS64WMKINDOBIEKU336N5Z7AVCNFSM6AAAAACNCMEEDKVHI2DSMVQWIX3LMV43ASLTON2WKOZTGY3DANRXGEZTOOA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>