A comparison of analytical approaches to obtain Mendelian randomization estimates with longitudinal exposures

Joy Shi¹, Sonja A. Swanson^1,2, Peter Kraft^1,3, Bernard Rosner³, Immaculata De Vivo¹, Miguel A. Hernán^1,3,4

¹ Department of Epidemiology, Harvard T.H. Chan School of Public Health, Boston, Massachusetts, USA
² Department of Epidemiology, Erasmus Medical Center, Rotterdam, The Netherlands
³ Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, Massachusetts, USA
⁴ Harvard-MIT Division of Health Sciences and Technology, Cambridge, USA

Background. Mendelian randomization (MR) is often used to estimate the effects of time-varying exposures on health outcomes using observational data. However, conventional methods for MR were designed to handle a single measure of exposure. As such, conventional MR effect estimates for time varying exposures are often difficult to interpret. We compared three analytical approaches to incorporate longitudinal measures of time-varying exposures in MR studies.

Methods. We used MR to estimate the effects of alcohol, high-density lipoprotein (HDL) cholesterol and low-density lipoprotein (LDL) cholesterol interventions on C reactive protein (CRP) levels, gamma-glutamyl transferase (GGT) levels and ankle-brachial index (ABI) using data from the Framingham Heart Study. We proposed weighted allele scores—based on genetic variants of alcohol consumption, HDL cholesterol and LDL cholesterol— as instruments. Multiple scores were generated for each exposure using different selection algorithms for genetic variants. We considered two exposure time points and used g-estimation to estimate parameters of structural mean models that (1) included each exposure time point separately; (2) included the average exposure across time points; or (3) included both exposure time points jointly.

Results. We observed increases in GGT levels and ABI as a result of increases in alcohol intake and LDL cholesterol across both time points. Effect estimates were similar regardless of the approach used, but the precision decreased with increasing number of time points included in the structural mean model and increasing correlation between weighted allele scores used for each time point.

Conclusions. This study demonstrates the first application of structural mean models to incorporate repeated exposure measurements to estimate the effect of a time-varying exposure on an outcome using MR. These methods provide a basis for considering time-varying exposures in future MR studies.