Differences in mean chi2 of the trait of interest and included traits

[pjordab]

Dear MTAG developers,

We are performing an MTAG analysis in which the trait of interest has a mean chi2 of 1.093.

When performing MTAG with traits that have the following parameters, I get a lower mean chi2 than in the original trait of interest and a FDR of 15.3%.

	SNPs	FDR	mean chi2	GWAS equiv. max. N
Trait_of_interest	6,970,600	-	1.093	8799
MTAG result for trait of interest	5,882,923	0.153	1.080	9318
	mean chi2	LDSC rg with trait 1	LDSC p-val with trait 1	LDSC Z-score with trait 1	Pairwise MTAG mean chi2	Pairwise MTAG FDR
Trait 1	1.093
Trait 2	1.400	0.468	1.07E-12	7.122	1.078	0.141
Trait 3	1.058	-0.427	1.00E-04	-3.830	1.076	0.132
Trait 4	1.586	-0.323	9.85E-06	-4.421	1.077	0.143

I assume that this is presumably due to the low sample size of my trait of interest and differences in the mean chi2 with the included traits. 

Is it possible during the MTAG analysis to shrink the mean chi2 of the included traits in some way? In order to avoid the increase of the maxFDR?

If not, do you think in this scenario with high FDR and absence of replication cohort, is it appropriate to take the MTAG values as valid but assume a more restrictive mtag_pval value, e.g. 1e-10? Or, might it be possible to identify the subset of SNPs that violate MTAG assumptions and lead to this high maxFDR?

Thanks, any advice would be greatly appreciated!

[paturley]

I don't know that there is a right answer to your question. The decreased mean chi2 and the high maxFDR is likely a consequence of having a moderate genetic correlation and a low-powered GWAS for your trait of interest. I imagine that using a stricter p-value threshold would tame the maxFDR, but then it would also cause you to lose any power you gained from using MTAG in the first place. (Also, I don't think the software currently allows you to set a significance threshold for the maxFDR calculation, so you'd need to customize the code yourself.)

You could use genomicSEM to fit an MTAG-like model. That software has functionality to try to identify SNPs that violate the model's assumptions, but my experience with those tests is that they are low powered, so it's difficult to reduce the FDR by testing for violations and removing SNPs that don't pass that test.

Sorry I can't be more helpful.

On Thu, Jun 9, 2022 at 1:40 PM pjordab @.***> wrote:

Dear MTAG developers,

We are performing an MTAG analysis in which the trait of interest has a mean chi2 of 1.093.

When performing MTAG with traits that have the following parameters, I get a lower mean chi2 than in the original trait of interest and a FDR of 15.3%.

SNPs FDR mean chi2 GWAS equiv. max. N Trait_of_interest 6,970,600 - 1.093 8799 MTAG result for trait of interest 5,882,923 0.153 1.080 9318

mean chi2 LDSC rg with trait 1 LDSC p-val with trait 1 LDSC Z-score with trait 1 Pairwise MTAG mean chi2 Pairwise MTAG FDR Trait 1 1.093 Trait 2 1.400 0.468 1.07E-12 7.122 1.078 0.141 Trait 3 1.058 -0.427 1.00E-04 -3.830 1.076 0.132 Trait 4 1.586 -0.323 9.85E-06 -4.421 1.077 0.143

I assume that this is presumably due to the low sample size of my trait of interest and differences in the mean chi2 with the included traits.

Is it possible during the MTAG analysis to shrink the mean chi2 of the included traits in some way? In order to avoid the increase of the maxFDR?

If not, do you think in this scenario with high FDR and absence of replication cohort, is it appropriate to take the MTAG values as valid but assume a more restrictive mtag_pval value, e.g. 1e-10? Or, might it be possible to identify the subset of SNPs that violate MTAG assumptions and lead to this high maxFDR?

Thanks, any advice would be greatly appreciated!

— Reply to this email directly, view it on GitHub https://github.com/JonJala/mtag/issues/160, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5NP7DJSDDNMXFKTODDVOIUBTANCNFSM5YK3YVEA . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[pjordab]

Thank you very much for your answer!! I'll try to use GenomicSEM. I'll keep you updated.