Always getting 'pve estimate =0.99xxx se(pve) =-nan' about gemma

[pcarbo]

@xiekunwhy I was able to reproduce this error using the simpler command:

gemma -bfile mydata -k mydata_kinship.sXX.txt -lmm

Further, I did not get an error using the linear regression instead of the LMM:

gemma  -bfile mydata -k mydata_kinship.sXX.txt -lm

The eigenvalue decomposition seemed to work okay (using the -eigen flag).

So my hunch is that GEMMA is behaving strangely with your phenotype data. To test this, I created a "dummy" phenotype (see the attached file), and re-fit the LMM to the dummy data:

$ gemma -bfile mydata -k mydata_kinship.sXX.txt -lmm -n 4
Reading Files ... 
## number of total individuals = 120
## number of analyzed individuals = 120
## number of covariates = 1
## number of phenotypes = 1
## number of total SNPs = 361505
## number of analyzed SNPs = 131480
Start Eigen-Decomposition...
pve estimate =0.0863376
se(pve) =0.148312

This time, it worked, and produced reasonable estimates.

I noticed that all your phenotype values are either 1 or 2.

@xiangzhou, do you have any thoughts as to why the LMM might not work with a binary-valued phenotype that is either 1 or 2, and only 120 samples?

mydata.fam.gz

[pcarbo]

@xiekunwhy One possible explanation is that the rows of the phenotype and genotype files do not match up. I have seen people make this error before.

[pjotrp]

We should check whether this is a duplicate of #58

[pjotrp]

With gsl1 in both cases we get results in the assoc file . With gsl2 we get nan in the assoc file. In all cases we get se(pve) =-nan for the first fam file. So I think it is mostly a GSL issue.

[pcarbo]

@xiekunwhy In case this is still useful for you, the new GEMMA v0.96 binaries should solve your problem. See also Issue #58.

[pcarbo]

@pjotrp Is it okay if I close this issue? Based on your test it looks like we solved the issue.

[xiekunwhy]

OK，thanks you two.

[pjotrp]

Please reopen if you still see a problem.

[pjotrp]

Running the test with my latest checks

    $gemma -bfile data/issue26/mydata -k data/issue26/mydata_kinship.sXX.txt \
           -miss 1 -maf 0.01 -r2 1 -lmm \
           -debug -issue 26 \
           -o $outn

we now get

testIssue26
Reading Files ... 
## number of total individuals = 120
## number of analyzed individuals = 120
## number of covariates = 1
## number of phenotypes = 1
## number of total SNPs    = 361505
## number of analyzed SNPs = 361505
**** DEBUG: Validating K in src/debug.cpp at line 23 in void do_validate_K(const gsl_matrix*, bool, const char*, int)
**** WARNING: K has small or negative eigenvalues! in src/gemma.cpp at line 2720
**** DEBUG: Eigenvalues Min 1.72861e-15 Max 11.9122 Ratio 6.89122e+15
**** WARNING: K is ill conditioned! in src/gemma.cpp at line 2720
**** FAIL: K is not positive definite! in src/gemma.cpp at line 2720

So, it looks like the Kinship matrix is a problem.

[pcarbo]

Based on followup discussion, this may not necessarily be an issue because the centered relatedness matrix is expected to have one zero eigenvalue. More to come...

[pjotrp]

Yes, the centered relatedness has one zero eigen value. Reporting is fixed now. Please reopen if you encounter issues again.