Verhulst · European journal of epidemiology 2013 · longitudinal cohort re-analysis and simulation study · n=?

Do leukocyte telomere length dynamics depend on baseline telomere length? An analysis that corrects for 'regression to the mean'.

Cited 142 times in the scientific literature.

Level 3 - non-randomized controlled study

Re-analysis of longitudinal cohort data and statistical simulation

PubMed 23990212 · doi:10.1007/s10654-013-9845-4 · record verified 2026-08-30

What was done

The authors used replicate leukocyte telomere length (LTL) measurements and simulation modeling to demonstrate how measurement error drives regression to the mean (RTM), creating spurious associations between baseline LTL and attrition rate. They then re-analyzed longitudinal LTL data collected from four study populations to evaluate whether baseline LTL predicts attrition rate after statistically correcting for RTM.

What was found

Measurement error created artifactual dependencies between baseline LTL and attrition in simulated data. In the empirical re-analysis of four cohorts, correcting for RTM reduced the slope of the association between baseline LTL and attrition rate by 57% under low measurement error (coefficient of variation ~2%). A modest statistically significant association persisted after correction: baseline LTL explained 1.3% of the variation in LTL attrition overall, with significant differences between study samples, and appeared primarily attributable to an association in men (3.7%).

Why it matters

Prior reports of faster telomere attrition in individuals with longer baseline telomeres are substantially inflated by regression to the mean caused by measurement error. While a real biological association appears to persist, its effect size is much smaller than previously estimated and may vary by sex.

Limits

The abstract does not report the total participant sample size, the specific demographic characteristics of the four cohorts, or follow-up durations. It also does not state whether measurement error in all datasets was at or above the modeled 2% coefficient of variation.

Cited by