Understanding the potential bias of variance components estimators when using genomic models

Area of Science:

• Genetics
• Statistical Genetics
• Bioinformatics

Background:

• Genomic models link phenotypes to genotypes for variance parameter inference.
• Restricted maximum likelihood (REML) is commonly used but can yield biased estimates if the model's covariance structure is misspecified.
• Misspecification occurs when the genomic model's covariance structure poorly reflects the true data structure, leading to incorrect inferences.

Purpose of the Study:

• To theoretically analyze genomic models with misspecified likelihoods.
• To develop a method for quantifying the bias in variance component estimates.
• To provide insights into the accuracy of heritability estimates derived from genomic models.

Main Methods:

• Theoretical analysis of genomic models by decomposing misspecified likelihood equations.
• Introduction of a quantitative measure to compare covariance structures between genomic and true models.
• Analysis of bias in variance component estimation under misspecification.

Main Results:

• A method is presented to isolate components of misspecified likelihood equations that cause incorrect inferences.
• An informative measure is defined to compare the covariance structure specified by a genomic model against the true data structure.
• The analysis determines the expected occurrence and direction of bias in variance component estimates.

Conclusions:

• The theoretical framework explains the effectiveness of methods aimed at reducing bias in heritability estimates.
• In simple genomic models for heritability (single component, i.i.d. SNP effects), existing bias tends to be downward.
• The study provides a robust theoretical foundation for assessing and correcting bias in genomic variance component estimation.