A better coefficient of determination for genetic profile analysis

Area of Science:

• Genetics
• Biostatistics
• Epidemiology

Background:

• Genome-wide association studies (GWAS) enable disease risk prediction using Single Nucleotide Polymorphism (SNP) markers.
• Quantifying the predictive accuracy of genetic profiles typically involves calculating coefficients of determination (R-squared).
• Existing pseudo-R-squared measures for binary outcomes lack standardization due to difficulties in defining residual variance on the probability scale.

Purpose of the Study:

• To address the limitations of current R-squared measures in genetic profiling.
• To propose a novel, globally standard R-squared measure interpretable on the liability scale.
• To enable direct comparison of genetic profile performance with heritability estimates.

Main Methods:

• Utilized a liability threshold model to link observed probabilities with underlying liability.
• Evaluated existing R-squared measures for interpretability, bias, and comparability to heritability.
• Developed and validated a new R-squared measure on the liability scale.

Main Results:

• Current R-squared measures for binary responses are often difficult to interpret and may be biased by ascertainment.
• These measures are not directly comparable to heritability, hindering benchmarking.
• The proposed novel R-squared measure is interpretable on the liability scale and comparable to heritability.

Conclusions:

• A standardized R-squared measure on the liability scale is crucial for accurate genetic profiling.
• The new measure overcomes limitations of existing methods, offering better interpretability and comparability.
• This approach facilitates robust evaluation of genetic risk predictors, even in ascertained case-control studies.