MRC INTEGRATIVE EPIDEMIOLOGY UNIT
SEMINAR SERIES

 Thursday, 17^th May, 2018 : 12.30 – 13.30
Room OS6, Second Floor, Oakfield House

  Rhian Daniel
Reader in Medical Statistics, Division of Population Health,
Cardiff University

  Mediation analysis with high-dimensional mediators

 

Abstract

In many modern biomedical applications, interest lies in decomposing the effect of an exposure on an outcome into its effect via a large number of mediators. For example, when trying to understand the mechanism through which a particular genetic variant affects cardiovascular disease, one might attempt to decompose the total effect of the variant into individual path-specific effects through hundreds or even thousands of blood protein and metabolite measurements. Or, suppose that we find a trend over time in mortality at a hospital ICU; it would be natural then to investigate what explains this trend – is it explained by changes over time in patient prognosis at admission, as captured by a high-dimensional set of admission variables? In the former example, it is clear that the genetic variant is the exposure, with the omics variables being mediators. In the latter example, I will argue that it is natural to think of time as the exposure, and the admission characteristics as high-dimensional mediators.

Mediation analysis in such settings poses formidable methodological challenges. First, the mediators are likely to be highly-correlated according to an unknown causal structure, including unmeasured common causes of one mediator and another. Second, the identification of natural path-specific effects in such a setting would rely on a large number of so-called "cross-world independence" assumptions, which are impossible to justify. Third, if we were to use a parametric estimation approach, as is most commonly done in mediation analysis, then, as the dimensionality of the mediators increases, so too does the extent to which our inferences rely on incorrectly-specified and arbitrarily-chosen parametric models.

We propose that the first two problems be overcome by focusing on interventional multiple mediator effects (Vansteelandt and Daniel, 2017) and the third by adopting a data-adaptive (machine learning) estimation approach.

In this talk I will outline a few possible approaches (including one based on targeted minimum loss-based estimation, TMLE), illustrating their use on data from one or both of the motivating applications.

Biography

Rhian Daniel (@statnav) is a Reader in Medical Statistics at the Division of Population Health, Cardiff University. Her research into methods for high-dimensional mediation analysis is supported by a Sir Henry Dale Fellowship funded by the Wellcome Trust and the Royal Society. Rhian moved to Cardiff (a move back “home”) in June 2017, after spending over a decade at the London School of Hygiene and Tropical Medicine. She is mainly interested in statistical methods for making causal inferences from observational data, particularly in settings, such as non-linear mediation analysis, where standard approaches have been shown to be invalid. She is interested in the development of these methods, particularly to settings with high-dimensional data, and is also passionate about their dissemination to a wider audience.

 

ALL WELCOME