Introduction to estimation of the causal effects in an observational study and g-methods

Graduate Programme in Public Health

Estimation of the causal effects in an observational study is a risky business. In an observational study, an exposure/treatment can be classified as either fixed or time-varying. An exposure/treatment is classified as fixed if the exposure does not change over time (for example a one-dose vaccine or a surgical intervention). Any exposure/treatment that is not fixed is classified as time-varying (for example a drug whose dose is readjusted according to the patient’s clinical response). In the absence of confounding or model misspecification, conventional methods to adjust for confounding variables (such as stratification, matching, and/or regression) deliver consistent estimators of the causal effect of a fixed exposure. In contrast, when interest focuses on the causal effect of a time-varying exposure, even when confounding and model misspecification are both absent, conventional analytic methods may be biased and result in estimates of effect that may fail to have a causal interpretation.

The g-computation algorithm formula (the “g-formula”), inverse probability of treatment weighting (IPTW) of marginal structural models (MSMs), and g-estimation of structural nested models (SNMs), is collectively referred as Robins’ generalized methods (g-methods). The g-methods provide consistent estimates of contrasts (e.g. differences, ratios) of potential outcomes under a less restrictive set of identification conditions than do standard regression methods (e.g. linear, logistic, Cox regression). In the real world, majority of the exposures/treatments are time-varying, which can potentially be handled by the g-methods. However, uptake of g-methods has been hampered by both complexity of analytic methods and technical details.
Day 1: A crash course on probability theory that required for the g-methods
Day 2: A crash course on Stata programming language that required for the g-methods
Day 3: Introduction to potential outcome frame
Day 4: An introduction to inverse probability of treatment weighting (IPTW) of marginal structural models (MSMs)
Day 5: An introduction to the g-computation algorithm formula (the “g-formula”)


After the participation in the course, the participants should be able to:
Comprehending potential outcome framework and average causal effect in a observational study
Comprehending concept of at least two types of the g-methods
Implementing the two types of the g-methods for fixed exposures in Stata
Facilitating the further studies on g- methods for time-varying exposures

Knowledge on directed acyclic graph (DAG) is a requited.
Basic knowledge on epidemiological concept of selection bias, confounding bias, and information bias
Knowledge on logistic regression is a requited.
Stata will be used during the whole process.

Course credit: 2,5 ECTS

Course leader: Associate Professor Chunsen Wu

Max number of participants: 20 PhD students

Course fee:
The course is free of charge for PhD students enrolled in  the Faculty of  Health Sciences at the University of Southern Denmark.  for other PhD students enrollede in universities that have joined the "Open market agreement", the fee is DKK 800,-.
For other participants there is a course fee of DKK 6800,-