WeightItMSM only "ATE"?

[mkrasmus]

Hi Noah many thanks for maintaining this package, very grateful it exists. I have a question regarding MSMs and ATE/ATT and the current functionality of WeightItMSM. The reference material on the website states that for the estimand: ""ATE", currently the only estimand for MSMs with binary or multi-category treatments." I was wondering if this is true for the MSM approach in general, or if it just a reflection of the WeightItMSM package in its current version? I was hoping to find an R package for MSM that yields ATT estimand and I think I might need to use the twang package's iptw function for this, but a little unsure on what is possible with MSM and ATT given that statement.

[ngreifer]

If you can find any description of MSMs for the ATT, I will happily look into implementing it :) But as far as I am concerned, the estimand isn't well defined for MSMs. So the fact that it would be possible to do with twang::iptw() is an error, not an advantage. You should not attempt this method without a theoretical understanding of whether it is possible. Again, if you find a reference for the ATT in this context, please let me know, as I would love to implement this in WeightIt. It could be as simple as multiplying ATT weights at each time point (in which case you can already do this by calling weightit() once per time point and multiplying the weights yourself, which is all weightitMSM() does), but it might be a lot more complicated than that. I'm working on a massive update to WeightIt so if you can find a reference for the method it would be a cool thing to add.

[mkrasmus]

Sorry for delayed response. I was aware of just this one example "Doubly Robust Estimation of Average Treatment Effects on the Treated through Marginal Structural Models" https://doi.org/10.1353/obs.2023.0025 . But this appears to be more how-to and I'm not sure if the theoretical consideration is addressed. But if there is something here then at least there is sufficient code and a generous repo: https://github.com/PFMB/causalchall

[ngreifer]

Yes, I did come across that paper, but it doesn't describe a methodology for estimating the ATT using weighting for MSMs. TMLE is a fundamentally different method since no weights are estimated. In the same way weighting for MSMs cannot be used to assess effect modification by time-varying confounders, I have doubts that it can be used to estimate the ATT because doing so involves conditioning on a time-varying variable (the treatment). In contrast, G-computation (and TMLE, which is just G-computation with a correction) can be used to assess effect modification by time-varying confounders because it doesn't directly condition on the confounders. So although the paper you linked to describes the estimand and one method of estimating it, I am still not convinced that there is a method that relies on weighting alone to target the ATT. Again, if you do find such a method, please send it along as this would be a useful addition to WeightIt.