Recent developments on causal inference within the statistical and
artificial intelligence literature have led to important new insights on how
to address problems of confounding and selection bias in a wide variety of
The aim of this Course is to review these developments and
to provide state-of-the-art statistical solutions for genetic path
identification, and for dealing with problems of confounding due to
population admixture and of selection bias resulting from ascertainment
conditions in genetic association studies.
The first day of the Course will
focus on probabilistic graphical models, which have their origins in genetic
path analysis and which provide a natural general framework for expressing
and manipulating many important concepts in statistical genetics. Local
computational algorithms can be described in this way, but also complex
issues of identification in forensic settings, for example, together with
genetic mapping and pedigree uncertainty can be handled in this context, as
can issues of causal inference and identification of regulatory networks.
This part of the course will introduce the basic ideas and illustrate how
graphical models can be used in a variety of settings.
The second day of the
Course will focus on statistical techniques to adjust for measured
confounding. Specifically, we will discuss limitations of ordinary
regression adjustment and focus on successful alternatives, such as inverse
probability weighting estimators in marginal structural models and
G-estimators. These methods will be applied to detect gene-environment
interactions in family-based association studies and to identify genetic
The third day of the Course will focus on the use of Mendelian
randomization for examining the causal effect of a modifiable exposure on
disease by making use of measured variation in genes of known function.
Emphasis will be on the assumptions required for using these methods and on
state-of-the-art analysis techniques. We will end with an introduction to
graphical search algorithms for unravelling causal genetic networks.
Course will include hands-on computer practical sessions during the
afternoons. It is intended for an audience with some previous familiarity
with statistics, in particular with linear and logistic regression.