Frequently Asked Questions¶
This page addresses some frequently asked questions from users. Most of the question that applies to HESS also applies to \(\rho\)-HESS.
What is the sample size requirement for HESS?¶
We recommend to apply HESS on GWAS with sample size greater than 50,000. HESS tends to be downwardly biased at smaller GWAS since fewer eigenvectors are necessary in the truncated-SVD regularization to obtain a stable estimate.
On the other hand, if genome-wide SNP-heritability is available for the GWAS (e.g. estimated from individual-level data), then it may still be possible to obtain the local estimates (see here).
Why is it necessary to re-inflate genomic control factor?¶
Most GWAS studies apply genomic control factor (\(\lambda_{GC}\)) on the association statistics. This could result in a downward bias in the estimated heritability since the second step of HESS involves an estimation of the environmental effect variance (\(\sigma_e^2\)) -- with \(\lambda_{GC}\) corrected GWAS summary stats, HESS tends to overestimate \(\sigma_e^2\), resulting in downward bias in local SNP-heritability.
See here for how to specify \(\lambda_{GC}\) in estimating local SNP-heritability, and here for genetic covariance.
Why do I get negative local SNP-heritability estimates?¶
Local SNP-heritability are not constrained to be greater than 0, so that if a locus truly has zero SNP-heritability, then HESS can give a zero estimation in the expectation.
If a locus has negative SNP-heritability estimate, then this is likely because the local SNP-heritability of the trait is close to zero.
Why do I get NaN for genome-wide SNP-heritability variance estimates?¶
This is because the estimated variance of the genome-wide SNP-heritability is negative, usually caused by relatively small sample size of the GWAS.
Why do I get negative variance estimates for local SNP-heritability / genetic covariance?¶
The variance estimates for local SNP-heritbility (genetic covariance) is a random variable that is not constrained to be positive. For some loci, the local SNP-heritability variance estimates may be negative.
Usually, negative variance estimates are caused by relatively small sample size of the GWAS.
Why does HESS give different estimate than LDSC?¶
Overall, HESS and LDSC should give similar estimates. However, HESS and LDSC have different definitions of SNP-heritability and genetic covariance. Also, unlike LDSC, HESS does not account for population stratification in the GWAS summary stats. This could also cause the difference.
What does the following error message mean?¶
Some users experience the following error message in step 2 of the analysis.
[ERROR] Rank of A less than the number of loci. There might be loci with no SNP.
The second step of HESS and \(\rho\)-HESS involves inverting a matrix. If the matrix is not invertible, HESS and \(\rho\)-HESS will not attempt to estimate SNP-heritability or genetic covariance.
Usually, this error is caused by empty locus (SNP with no SNP in it).
If there are empty loci in the data, it is recommended to remove these loci or combine these loci with nearby loci.
What does the following error message mean?¶
Some users experience the following error message in step 2 of the analysis.
IOError: CRC check failed ...
This is usually caused by prematurely finished step 1.
Note that step 1 usually involves inverting large matrices and may take quite bit of memory.