Comparison of Marginal Structural Models to a missing data approach illustrated by data on breast cancer chemotherapies

One of the main objectives in clinical epidemiology is to detect a relation between treatment and outcome. We address data where treatment is applied repeatedly in time and the dose given at a specific time-point may be modified due to actual measurements on disease parameters. If such measurements are subsequently affected by the treatment, they might act as time-dependent confounders. Standard statistical methods cannot adequately address such confounders, but Marginal Structural Models (MSMs) proposed by Robins cope with them. However, these models are still controversely discussed because they are defined within the counterfactual framework. We illustrate Robins' approach as an extension of a common approach developed for the handling of missing outcomes which does not explicitely use counterfactuals. We address two questions on breast cancer chemotherapy schemes given in repeated cycles. First, we examine the therapy effect and compare two different chemotherapy schemes by the outcome after the fully applied chemotherapy regimen. We account for confounding due to early stopping by Inverse-Probability-of-Censoring-Weighting. Secondly, we investigate the dose effect of one chemotherapy, i.e. the influence of the number of given cycles on the outcome which is modeled by a MSM. Now, the effect is defined by coun-terfactual variables and time-dependent confounders are accounted for by estimating the parameters of the MSM via Inverse-Probability-of-Treatment-Weighting. We illustrate the concepts of MSMs by showing parallels to the first analysis and pointing out the differences.