Maximizing Insights from Longitudinal Epigenetic Age Data: Simulations, Applications, and Practical Guidance

TL;DR

This paper explores how to best analyze changes in epigenetic age over time using simulations and real data, offering guidance for future studies.

Contribution

The paper provides a methodological comparison of models for analyzing longitudinal epigenetic age data and validates findings with real-world examples.

Findings

Linear mixed models and generalized estimating equations using chronological age as a time variable yield the most accurate effect sizes.

Epigenetic age accelerates in males and decelerates in children with higher birthweight when using optimal modeling approaches.

Abstract

Epigenetic Age (EA) is an age estimate, developed using DNA methylation (DNAm) states of selected CpG sites across the genome. Although EA and chronological age are highly correlated, EA may not increase uniformly with time. Departures, known as epigenetic age acceleration (EAA), are common and have been linked to various traits and future disease risk. Limited by available data, most studies investigating these relationships have been cross-sectional - using a single EA measurement. However, the recent growth in longitudinal DNAm studies has led to analyses of associations with EA over time. These studies differ in (i) their choice of model; (ii) the primary outcome (EA vs. EAA); and (iii) in their use of chronological age or age-independent time variables to account for the temporal dynamic. We evaluated the robustness of each approach using simulations and tested our results in two real-world examples, using biological sex and birthweight as predictors of longitudinal EA.

Our simulations showed most accurate effect sizes in a linear mixed model or generalized estimating equation, using chronological age as the time variable. The use of EA versus EAA as an outcome did not strongly impact estimates. Applying the optimal model in real-world data uncovered an accelerated EA rate in males and an advanced EA that decelerates over time in children with higher birthweight.

Our results can serve as a guide for forthcoming longitudinal EA studies, aiding in methodological decisions that may determine whether an association is accurately estimated, overestimated, or potentially overlooked.

Figures