whether i could use MTAG when there is WARNING: mean chi^2 may be too small

[lynling14]

Hi,

Before i used MTAG, i used LDSC and tried to calculate the genetic correlation of two traits. However, at the first step of LSDC, in the log of one trait showed that "WARNING: mean chi^2 may be too small". And the LDSC suggests " If mean chi-square is below 1.02, munge_sumstats.py will warn you that the data probably are not suitable for LD Score regression."

So i was wondering if i could still use MTAG when the mean chi-square is below 1.02 or if i could used MTAG when the heritability is not large.

Many thanks! Ling

[paturley]

Hello Ling,

I am nervous when I hear about applications of MTAG with mean chi2 values below 1.02. MTAG relies on having reliable estimates of the genetic correlation between traits, and with such a low mean chi2, it's possible that the estimated rg that is used will be much too high or too low, which can inflate the type I error rate. Sorry I can't be more helpful.

Best, Patrick

On Fri, Sep 16, 2022 at 3:58 AM lynling14 @.***> wrote:

Hi,

Before i used MTAG, i used LDSC and tried to calculate the genetic correlation of two traits. However, at the first step of LSDC, in the log of one trait showed that "WARNING: mean chi^2 may be too small". And the LDSC suggests " If mean chi-square is below 1.02, munge_sumstats.py will warn you that the data probably are not suitable for LD Score regression."

So i was wondering if i could still use MTAG when the mean chi-square is below 1.02 or if i could used MTAG when the heritability is not large.

Many thanks! Ling

— Reply to this email directly, view it on GitHub https://github.com/JonJala/mtag/issues/168, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5KYQQ3URY4L2RQZE2DV6QSDBANCNFSM6AAAAAAQOCW2BU . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[lynling14]

Hello Patric,

Thanks for your reply, that helps a lot.

Best, Ling

[lynling14]

Dear Patrick,

I am curious about the statistic mean chi^2 (χ2). How to understand this statistic, how to calculate mean chi^2, and why we should be careful when the mean chi^2 below 1.02 no matter in MTAG or LDSC, why is not below 1.1 or 1.05. Is there any literature support 1.02 or is there any literature could help me to understand this statistic better? Did you ever simulate what happened when the mean chi^2 of traits are below 1.02.

Many thanks! Ling

[paturley]

Hello Ling,

There is nothing magical about the 1.02 threshold. We tested a variety of cases, and the lower the mean chi2, the less precise the estimates of h2 and rg were, and therefore the less stable MTAG was. We picked a number that seemed sensible to us, but the right number would depend on how nervous you are about the increased Type I error rate or the losses in power. You can look at the MTAG paper for our simulation results at different mean chi2 levels and make a judgment for yourself if you'd like. (I presume LDSC also conducted such simulations, but I wasn't on that paper, so I'm not sure.)

Patrick

On Mon, Sep 26, 2022 at 12:01 PM lynling14 @.***> wrote:

Dear Patrick,

I am curious about the statistic mean chi^2 (χ2). How to understand this statistic, how to calculate mean chi^2, and why we should be careful when the mean chi^2 below 1.02 no matter in MTAG or LDSC, why is not below 1.1 or 1.05. Is there any literature support 1.02 or is there any literature could help me to understand this statistic better? Did you ever simulate what happened when the mean chi^2 of traits are below 1.02.

Many thanks! Ling

— Reply to this email directly, view it on GitHub https://github.com/JonJala/mtag/issues/168#issuecomment-1258270498, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFBUB5NP2NFORDUS4KFH6EDWAHCGDANCNFSM6AAAAAAQOCW2BU . You are receiving this because you commented.Message ID: @.***>