Τομέας Κοινωνικής Ιατρικής - Ψυχιατρικής και Νευρολογίας
Library of the School of Health Sciences

2015-06-02

2015

Βουρλή Γεωργία

Παναγιώτα Τουλούμη, Φώτιος Σιάννης, Miguel Hernan

Causal models and their application to estimate the effect of Chronic Hepatitis B treatment

Greek

Classic statistical analysis methods examine whether observed relationships are
due to chance and provide us with inference concerning non-circumstantial
associations between variables, that may however be non-causally interpreted.
Unfortunately, it has been shown that in the context of a longitudinal
observational study, when a covariate affected by past exposure is both a
predictor of the future exposure and the outcome, i.e there exists
time-dependent confounding, standard analysis approaches for the estimation of
the exposure’s effect, may produce biased estimates. The g-methods are a class
of methods introduced to estimate causal effects. The most recent of them is
the Inverse Probability of Treatment Weighting (IPTW), which is applied to
estimate the parameters of the Marginal Structural Models (MSMs).
The aim of this thesis was to assess the performance of the MSMs in situations
often met in longitudinal observational studies with survival endpoints.
Following an exact simulation method, we explored two classes of scenarios: a)
missed visits, which resembles clinical cohort studies; b) missing confounder’s
values that corresponds to interval cohort studies. In the first class, data
were analyzed initially without any correction and subsequently using either
weights’ truncation or normalization. In the second class, data were analyzed
either filling-in the missing values with the Last Observation Carried Forward
(LOCF), or after imputing them by Multiple Imputation (MI) or accounting for
missingness through additional weighting with the Inverse Probability of
Missingness (IPMW).
Furthermore, we analysed data from the HEPNET.Greece study for viral hepatitis
B, in order to evaluate the effect of treatment and the treatment type
(Interferon (IFN) +/- Nucleos(t)ide (NA) vs. NA) to the occurrence of a
clinical event in CHB patients.
Results of this thesis suggest that data from observational studies can provide
us with useful inference, given that they are analyzed appropriately. Even in
the presence of problems related to missing confounding data or irregular
observational plans that are often met in observational studies, sophisticated
approaches to account for potential bias can be incorporated into the MSM.

Causal models, Marginal structural models, Longitudinal data, Survival analysis

Yes

ix-xv

Yes

113

xv,192

File access is restricted only to the intranet of UoA.

https://pergamos.lib.uoa.gr/uoa/dl/object/1308436