Preprint Open Access
Preprint, Supplementary Material, associated R script and dataset.
Longitudinal studies have an important role in telomere epidemiology. In analysing effects of exposures on change in leukocyte telomere length (LTL), it is common to control for baseline LTL. However, collider bias arising from measurement error could cause overestimation of the difference in LTL attrition between groups with different exposures. We evaluated this using smoking as a test case.
We simulated LTL data to ask whether controlling for baseline LTL biases estimates of the difference in LTL attrition between smokers and non-smokers. We tested predictions from our simulation in a meta-analysis of previously-published longitudinal cohorts.
Our simulations show that if baseline LTL is shorter in smokers and LTL measurement error is non-zero, then controlling for baseline LTL overestimates the difference in LTL attrition between smokers and non-smokers. The size of this bias increases synergistically with increasing baseline difference and increasing LTL measurement error. Supporting these simulation results, the estimated difference in LTL attrition between smokers and non-smokers in empirical data is greater when models control for baseline LTL and the size of this discrepancy is positively correlated with LTL measurement error.
The false-positive error rate for reports of effects of smoking on telomere attrition is likely to exceed 5%. The bias responsible is not specific to smoking and will affect all exposures for which baseline differences in LTL exist. To avoid bias, models of LTL attrition should not control for baseline LTL. Many claims of accelerated LTL attrition in individuals exposed to adversity need to be re-assessed.
Bateson Eisenberg & Nettle - R script revised.R
Bateson Eisenberg & Nettle revised MS with tables and figs.pdf
Bateson Eisenberg & Nettle Supplementary Material.pdf