References

One of the main objectives in clinical epidemiology is to detect a relation between a factor, e.g. treatment dose, and outcome. We address data where treatment is applied repeatedly in time and the respective dose rate is set according to actual measurements. If such mea-surements are subsequently affected by given treatment, they might act as time-dependent confounders. Marginal structural models (MSMs) proposed by Robins [1] adequately address such confounders which allows a causal interpretation of the estimated treatment effect. In this talk, different aspects of Robins ’ approach are illustrated by analysing data of the GEPARDUO study [2]. This is a randomised clinical trial in breast cancer which compares two chemotherapies that are applied preoperatively. Outcome is pathohistological response measured at subsequent surgery. The chemotherapies are given in repeated cycles. Clinical values such as palpation result measured before each cycle may lead to stop chemotherapy prior to the last planned cycle. To provide insight into Robins ’ approach, we address two different questions. In a first step, we investigate the difference in outcome between the two chemotherapy regimen. We account for confounding due to early stopping by inverse-probability-of-censoring weighting. In a secon