G-estimation of structural mean models using instrumental variables

By Joy Shi

Description

G-estimation of structural mean models (SMMs) is one method of estimating the causal effect of a time-varying treatment on an outcome using instrumental variable analysis:

Figure 1. Direct acyclic graph with a time-fixed instrument (Z), time-varying exposure (A), time-fixed outcome (Y), and time-fixed treatment-outcome confounder (U)



Simulated Data Examples

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

Posted on:
December 6, 2021
Length:
1 minute read, 26 words
See Also: