Estimands in epigenome-wide association studies.

Area of Science:

• Epigenetics
• Genomics
• Bioinformatics

Background:

• DNA methylation analysis, including epigenome-wide association studies, assesses effects at differentially methylated CpG sites.
• Beta-values and M-values are two common outcome measures for statistical analysis, each with distinct properties and interpretations.
• Beta-values offer direct biological interpretation as methylation percentages but have statistical limitations, while M-values are statistically robust but lack direct biological interpretability.

Purpose of the Study:

• To present and discuss four approaches for obtaining estimands in DNA methylation analysis.
• To evaluate the usage and dependencies of M-values and Beta-values within bioinformatics pipelines.
• To demonstrate and correct for deviations caused by confounder effects in DNA methylation data analysis.

Main Methods:

• Comparative analysis of Beta-values and M-values using two data simulations, with and without confounder effects.
• Application of the intercept method to correct for confounder effects when analyzing Beta-values.
• Validation of theoretical findings on two large human genome-wide DNA methylation datasets.

Main Results:

• M-values provide statistically sound inference but yield biologically uninterpretable effect estimates.
• Beta-value statistics, when used alone, can overlook crucial confounder effects, rendering them unreliable for reporting.
• The intercept method effectively corrects for confounder effects when Beta-value differences are the primary focus.

Conclusions:

• M-values are suitable for exploratory analysis of CpG sites but not for biological interpretation of effect sizes.
• Reporting Beta-value statistics without accounting for confounders is not recommended due to potential biases.
• The intercept method is the recommended approach for studies focusing on Beta-value differences, ensuring accurate evaluation of hyper- or hypomethylated CpG sites.