Dear GEMMA developers,
I hope this email finds you well.
I am writing to seek your advice on a GWAS design issue related to the assumptions of linear mixed models implemented in GEMMA.
I am analyzing a binary trait (heat tolerance: survival vs. death) using whole-genome sequencing data. The dataset includes 1,000 individuals (500 survivors and 500 non-survivors). However, due to experimental constraints, the samples were sequenced in two separate batches/platforms:
Batch A: 500 individuals, all survivors
Batch B: 500 individuals, all non-survivors
Thus, the sequencing batch is completely confounded with the phenotype (batch = phenotype).
After standard variant calling and quality control (including strict filtering on SNP missingness, e.g. retaining only SNPs genotyped in ≥95–99% of individuals in both groups), I used GEMMA to estimate SNP-based heritability with a linear mixed model and obtained an estimate of approximately h² ≈ 0.6 for heat tolerance. A subsequent GWAS using GEMMA identified several thousand genome-wide significant SNPs.
From a variant-calling perspective, sequencing experts suggested that differences between platforms primarily affect missingness or callable regions, rather than systematically biasing allele frequency estimates at well-genotyped SNPs, and that stringent missingness filtering should therefore mitigate batch effects.
However, I am concerned about the statistical validity of GWAS and heritability estimation under a design where batch is perfectly confounded with phenotype. Specifically, I would greatly appreciate your guidance on the following points:
In a linear mixed model framework using a GRM (as in GEMMA), can strict filtering on SNP and sample missingness fully mitigate bias when batch and phenotype are completely confounded?
Even if allele frequency estimates at retained SNPs are unbiased, could batch-specific missingness or callable-region differences still induce spurious genetic similarity correlated with phenotype, leading to inflated heritability estimates or false-positive associations?
In such a scenario, would the heritability estimate (e.g. h² ≈ 0.6) and the GWAS signals be considered unreliable due to lack of within-batch phenotypic contrast?
From your perspective, is it fundamentally necessary that each technical batch contains phenotypic variation (cases and controls) for GEMMA-based GWAS to be statistically identifiable?
My goal is to understand whether the observed results could reflect true genetic signals, or whether the complete confounding between batch and phenotype violates the assumptions of the model in a way that cannot be corrected by QC alone.
Thank you very much for your time and for developing GEMMA. Any insight or reference you could provide would be greatly appreciated.
Dear GEMMA developers,
I hope this email finds you well.
I am writing to seek your advice on a GWAS design issue related to the assumptions of linear mixed models implemented in GEMMA.
I am analyzing a binary trait (heat tolerance: survival vs. death) using whole-genome sequencing data. The dataset includes 1,000 individuals (500 survivors and 500 non-survivors). However, due to experimental constraints, the samples were sequenced in two separate batches/platforms:
Batch A: 500 individuals, all survivors
Batch B: 500 individuals, all non-survivors
Thus, the sequencing batch is completely confounded with the phenotype (batch = phenotype).
After standard variant calling and quality control (including strict filtering on SNP missingness, e.g. retaining only SNPs genotyped in ≥95–99% of individuals in both groups), I used GEMMA to estimate SNP-based heritability with a linear mixed model and obtained an estimate of approximately h² ≈ 0.6 for heat tolerance. A subsequent GWAS using GEMMA identified several thousand genome-wide significant SNPs.
From a variant-calling perspective, sequencing experts suggested that differences between platforms primarily affect missingness or callable regions, rather than systematically biasing allele frequency estimates at well-genotyped SNPs, and that stringent missingness filtering should therefore mitigate batch effects.
However, I am concerned about the statistical validity of GWAS and heritability estimation under a design where batch is perfectly confounded with phenotype. Specifically, I would greatly appreciate your guidance on the following points:
In a linear mixed model framework using a GRM (as in GEMMA), can strict filtering on SNP and sample missingness fully mitigate bias when batch and phenotype are completely confounded?
Even if allele frequency estimates at retained SNPs are unbiased, could batch-specific missingness or callable-region differences still induce spurious genetic similarity correlated with phenotype, leading to inflated heritability estimates or false-positive associations?
In such a scenario, would the heritability estimate (e.g. h² ≈ 0.6) and the GWAS signals be considered unreliable due to lack of within-batch phenotypic contrast?
From your perspective, is it fundamentally necessary that each technical batch contains phenotypic variation (cases and controls) for GEMMA-based GWAS to be statistically identifiable?
My goal is to understand whether the observed results could reflect true genetic signals, or whether the complete confounding between batch and phenotype violates the assumptions of the model in a way that cannot be corrected by QC alone.
Thank you very much for your time and for developing GEMMA. Any insight or reference you could provide would be greatly appreciated.