Extending a Time-Varying Multivariable Mendelian Randomisation Model to Accommodate Two Outcome Measurements

Abstract

The application of multivariable Mendelian randomisation (MVMR) to analyse time-varying data with multiple measurements of both an exposure and an outcome is unclear. The purpose of this thesis is to develop and examine the properties of a potential model to extend MVMR to handle two measurements of both an outcome and an exposure. The exposure effect at Time 1 is estimated using univariable Mendelian randomisation (MR), while the exposure effects at Time 2 are estimated using MVMR by using a set of single nucleotide polymorphisms (SNPs) exclusive to the first outcome measurement. Simulations examining the properties of the causal effect estimates in the model under different scenarios were undertaken. The scenarios included different sampling schemes (1, 2, or 4 samples) for summary statistics. Confidence intervals were too wide, over-coverage was present when following the one-sample scheme, while slight under-coverage in both the two-sample and four-sample schemes was observed. Parameter estimators appeared to be mainly unaffected by increasing instrument strength. Increasing the number of SNPs pertaining to each exposure led to increased biases for the parameters affecting the second outcome measurement. Lastly, parameter estimates maintained acceptable coverage and small biases for different scenarios of overlapping SNPs. The inclusion of SNPs pertaining to the first outcome measurement in a time-varying MVMR model with two exposure and two outcome measurements allows for the estimation of exposure effects at both time points. However, the apparent drop in performance when the number of SNPs increases is of concern.