search

josefin-werme / LAVA

54 stars 9 forks source link

Ref: Model does not converge #52 #54

Closed ancolococky closed 1 year ago

ancolococky commented 1 year ago

So, how was this error solved? I am also tackling the same issue. In my case, the error occurred for some loci but not for other loci. I will really appreciate it if anyone could provide a solution. Thanks!

https://github.com/josefin-werme/LAVA/issues/52#issue-1804890626

cadeleeuw commented 1 year ago

This issue can arise if there is an extreme skew in the case-control ratio, or sometimes if there is an extreme amount of genetic signal for that phenotype in that region. Do either of these apply to your input data for those loci?

ancolococky commented 1 year ago

Hi cadeleeuw,

Thank you very much for your reply.

I am measuring the genetic correlations among three groups with different traits. Those traits are genetically correlated at a global level.

The case-control ratios are skewed towards the cases in all the groups: trait1 0.0127 trait2 0.0178 trait3 0.0577 If these skewnesses matter, the model conversion issue would happen to all the loci and the traits tested, so I guess this is not the critical reason.

The numbers of signals in each locus for all the traits tested are as below. Locus | Trait 1 | Trait 2 | Trait 3 1 | 7715 | 7701 | 9288 2 | 4453 | 4386 | 6235 3 | 6175 | 6069 | 8153 4 | 7299 | 7113 | 10093 5 | 6699 | 6593 | 8603 6 | 7536 | 7581 | 9680 7 | 7657 | 7653 | 10559 8 | 6888 | 6879 | 9120 9 | 6570 | 6626 | 8569 10 | 5768 | 5548 | 7325 <- Inversion error in Trait 1 11 | 6813 | 6597 | 8522 12 | 7408 | 7223 | 9596 13 | 6121 | 6429 | 6811 14 | 10135 | 10033 | 12428

Although the numbers of signals in Trait3 are larger than the other two in all the loci, the issue seems not to come from the differences.

Thanks for your help.

cadeleeuw commented 1 year ago

Hi,

Indeed those do not suggest an immediately obvious cause. Skewness and level of signal still may be contributing (those don't guarantee inversion errors, just make those more likely), so I would have to directly inspect the input data to determine specifically what is causing it in this instance. Though even if that can be identified, that may not mean a direct solution is available.

We are also currently working on a change in the analysis model for binary phenotypes that we are hoping to implement later this year, which would not have this problem. That is still undergoing validation however, so at present I cannot say when this will become available.

Best, Christiaan