Verteilungsbasierte kausale Inferenzmodelle zur Schätzung von Therapieffekten in randomisierten kontrollierten klinischen Studien

In prospective randomized trials differences in population based
means can be considered as estimates of mean causal exposure or
treatment effects. Nevertheless, occurence of intermediate events
in the course of a trial may lead to biased estimates of treatment
effects. Particularly this will be the case, if the probability of
such an event is not independent of initial treatment allocation
and the intermediate event is both a risk factor for the main
outcome parameter (e.g. survival) and a predictor of subsequent
treatment. This situation was referred as 'treatment by indication
problem' by Robins (1992) and is common in epidemiological trials.
Robins demonstrated that the usual approach of an adjusted
estimation of treatment effect using a time-dependent proportional
hazards model may be biased in this situation, whether or not one
further adjusts for past confounder history in the analysis. In
this thesis a novel inference procedure for randomized trials is
introduced which is based on the idea of Robin's G-Estimation
principle but particularly consideres specifics of randomized
trials. The suggested procedure allows for an unbiased and
consistent estimation of a treatment effect parameter (or paremeter
vector) which is connected by a real-valued function to the
parameters of an underlying distribution of survival times. In
fulfilment of the requirements of a causal individual-level based
model, a link of a subject's observed and counterfactual survival
time is directly achieved by the inverse distribution function of
survival times in a reference treatment arm. In this term, causal
inference is feasible based on the likelihood function and its
corresponding test statistics, even under appropriate consideration
of time-dependent confounders.