Mendelian randomization with repeated measures of a time-varying exposure: an application of structural mean models
Author: Joy Shi
Last updated: May 19, 2021
R code is provided to replicate the simulations presented in Appendix 3 of Mendelian randomization with repeated measures of a time-varying exposure: an application of structural mean models.
The simulations assess the use of structural mean models (SMMs) when conducting Mendelian randomization analysis of time-varying exposures. In all simulations, we consider data-generating models with three relevant exposure time points and assess under which conditions we can identify the causal effect of interest. The following table provides a summary of the causal estimand of interest and the assumptions made in the data-generating models for each simulation:
| Simulation | Causal estimand of interest | # of exposure measurements considered in the model | Instrument-exposure relationship changes over time? | Effect of exposure modified by previous exposure? | Presence of time-varying outcome-exposure confounding? |
|---|---|---|---|---|---|
| A.3.1 | Point effect | One | Yes | No | No |
| A.3.2 | Period effect | All (three) | Yes | No | No |
| A.3.3 | Period effect | All (three) | Yes | No | Yes |
| A.3.4 | Period effect | All (three) | Yes | Yes | No |
| A.3.5 | Period effect | One | Yes | No | No |
| A.3.6 | Period effect | One | No | No | No |
| A.3.7 | Period effect | Subset (two) | Yes | No | No |
| A.3.8 | Period effect | Subset (two) | No | No | No |
| A.3.9 | Period effect | Subset (two) | Yes a | No | No |
a Instrument-exposure relationship changes over certain (but not all) time intervals