[Forum SIS] Short course in causal inference for the health sciences

[Dottorato Statistica]

PhD Program in Statistics, Department of Statistics, University of Milan
Bicocca

 

Topics in causal inference for the health sciences



 

by Els Goetghebeur and Stijn Vansteelandt

Dept. Applied Mathematics and Computer Science, Ghent University 

 

 

Summary

Recent developments in 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
settings. The aim of this course is to review some of these developments and
to provide state-of-the-art statistical solutions for dealing with these
problems, motivated by substantive problems from medicine and public health.


 

The first half day of the course will introduce the framework of potential
outcomes as a tool that helps articulate causal questions and model data to
seek their answers. It will consider the two basic classes of assumptions
that allow to make progress based on observed data 1) the no unmeasured
(time-varying) confounders assumption and 2) the instrumental variables
assumption with reference to Mendelian randomization.  For each class
structural mean or distribution models will be developed that allow to
analyze continuous, discrete or right censored survival type outcomes. Their
application in well chosen observational and experimental clinical studies
will be presented. Due attention will be given to interpretation and
justification of the approach and its conclusions in the clinical context.
Participants will be offered a lead into the statistical software available
for implementation.

The second half day of the course will focus on causal diagrams to  express
causal background knowledge including: (i) ways for reading off such graphs
whether a given data situation suffers problems of confounding and selection
bias (ii) and whether/how this can be accommodated, for instance via inverse
probability weighting. The usefulness of such diagrams will be illustrated
in the context of complex problems of time-varying confounding, such as
arise when estimating the effect of hospital-acquired pneumonia on
mortality. It will be shown that standard approaches based on
regression-adjustment for confounding by disease severity are fallible, and
that progress can be made using inverse probability weighting under
so-called marginal structural models.

 

 

The course will be held at the Department of Statistics, University of
Milan-Bicocca building U7, 2nd floor, room 2019, on 20th and 21th December
2011.

 

The timing of the course is as follows.

20th December 2011:

Lecture 1               14:00 - 16:00

Break                    16:00 - 16:15

Lecture 2               16:15 - 18:00

 

21th December 2011:

Lecture 1                9:00 - 11:00

Break                     11:00 - 11:15

Lecture 2               11:15 - 13:00

 

-------------- parte successiva --------------
Un allegato HTML è stato rimosso...
URL: <http://www.stat.unipg.it/pipermail/sis/attachments/20111206/93cd3a4f/attachment.html>