2021 年 3 巻 2 号 p. 37-45
In clinical epidemiological studies, many exposures and confounders are time dependent. In the presence of time-dependent confounders affected by previous exposures, the usual analytic methods may introduce biases. Marginal structural models are used to deal with time-dependent confounders and exposures. A marginal structural model is a regression model for a pseudo-population using the concept of a potential outcome. The inverse probability of treatment and censoring weighting method is used to create a pseudo-population in which the effects of baseline confounders and time-dependent confounders can be removed when estimating the causal effect of the exposure on the outcome event. If the accuracy of the weights is high, the inverse probability of treatment and censoring weighting method is reliable and the bias of the marginal structural model is small. After the weights are created, a weighted regression model is applied to calculate the treatment effect. This seminar series paper introduces time-dependent confounders, time-dependent treatments, and marginal structural models.
