The Marginal Structural Model (MSM) is used for assessing the causal effects on treatments when there exist time-varying confounders. These confounders are also affected and affect previous or future assignments. Because of the complicated structure and time-varying properties, we cannot use standard statistical methods. In the previous web pages, we show the applications and rationales behind MSM. MSM can solve the problems in time-varying confounders and even provide a Cox-type model to assess the strength of treatment effect, to estimate objects’ survival time.

However, even though we know the rationales behind MSM are IPTW and IPCW, and we can also estimate all conditional expectations via saturated regression and exposures probability model, we still need to be careful of “pitfalls” (the following bullets are directly from Tomohiro and Etsuji, 2020):

• MSM should be distinguished from inverse probability weighting
• MSM shows prespecified assumptions on causal estimands, while an exposure probability model is an imposed restriction on the observed distribution
• As MSM and exposure probability model are used for different purposes, misspecification of these models would lead bias
• MSM and exposure probability model raise different challenges in real data analysis
• G-formula, which shares identifiability assumptions with inverse probability weighting, can be used to fit MSMs only when the models are saturated

Also, MSM also depends on a number of assumptions:

1. the information on time of treatments and outcomes is accurate
2. the measured covariates are sufficient to adjust both confounding and selection bias due to loss to follow-up
3. the models for initiation of treatment and censoring, given the past, are correctly specified
4. the levels of baseline covariates are correctly specified in MSM

Last but not least, I believe the MSM is useful in clinical trials and survival analysis, because it is a Cox-type model which is easy to interpret. In the future, we can extend the MSM by releasing assumptions, and we can also consider multiple treatments rather than binary treatments. Furthermore, considering relative analysis to assess the difference between MSMs is also one of the possible future topics.