Breadcrumb

Advanced Instrumental Variable Methods

Supervisor: Vanessa Didelez

Theme: Bayesian Modelling & Analysis

It is generally accepted that if we want to find out the causal effect of X (e.g. a drug) on Y (e.g. disease) we should randomly allocate the different levels of X (e.g. placebo and real drug) to a randomly selected group of people and then check for differences in Y. For many questions of interest this is not possible, e.g. when we want to assess the effect of alcohol consumption on some health outcome we cannot randomly allocate people to different levels of alcohol consumption. However, it may sometimes be possible to find situations where due to nature or some outside factor an (imperfect) random allocation has taken place - this is called an instrumental variable. For example, people with a certain genetic mutation react immediately very adversely to alcohol, and this mutation is essentially randomly distributed in the population. So we can still obtain some information on the effect of X on Y by considering such genetic instrumental variables.
The instrumental variable methodology has been developed in econometrics and its use for medical and epidemiological problems is still in its infancy. The latter pose new problems, such as non-linear effects, multiple instruments and weak (i.e. very imperfect) instruments, multiple outcomes, large sets of covariates etc. Furthermore it appears to be promising to consider a Bayesian approach especially for more complex problems, but as this involves modelling latent structures and hence underidentified models, careful attention has to be paid to informative prior assumptions and sensitivity analyses. A PhD project will address the latest challenges in this area of research, preferably from a Bayesian point of view. The project will combine interesting statistical modelling and theory of inference, with computational challenges, while being very much driven by practical applications. In collaboration with the School of Community and Social Medicine (especially CAiTE) any new methods will be applied to practical epidemiological questions.


Publications

  • On the choice of parameterisation and priors for the Bayesian analyses of Mendelian randomisation studies (2012)
    Jones, Thompson, Didelez, Sheehan
    Statistics in Medicine, vol: 31, Issue: 14, Pages: 1483 - 1501
    DOI: 10.1002/sim.4499
    URL provided by the author
  • Assumptions of IV methods for observational epidemiology (2010)
    Didelez, Meng, Sheehan
    Statistical Science, vol: 25, Issue: 1, Pages: 22 - 40
    DOI: 10.1214/09-STS316
    URL provided by the author
  • Mendelian randomisation as an instrumental variable approach to causal inference (2007)
    Vanessa Didelez and Nuala Sheehan
    Statistical Methods in Medical Research, vol: 16, Pages: 309 - 330
    URL provided by the author