TMLE update: include option of place clever covariate as the weights

[miaow27]

We have used zepid:TMLE for a while and notice the estimate as well as estimate's variance under some setting are majorly different from R:tmle. (Due to company policy, I cannot share data with you).

We think we are able to identify the root-cause of this discrepancy: R's implementation of TMLE use the "clever covariet" by weighting (target_gwt = True) in the latest release. After changing zepid to have the same setting, we observe a more consistent alliance between R and zepid. Here we want to share you our slight changes of codes and hope this can bring the function to be beneficial to a broader audience.

Based on our empirical experience, target_gwt = True might be a more stable setting for general user.

• under extreme cases, target_gwt = False tends to give extreme result that is majorly different from IPW or AIPW. While target_gwt = True gives more closer and stable result.
• for most cases, target_gwt = True will gives similar ATE as target_gwt = False with larger p-value. So the weighting approach might be more conservative and robust to weird data or model specification.

Attached with some additional resources about this topics:

• R TMLE source code snippet: https://github.com/cran/tmle/blob/master/R/tmle.R#L1583-L1602
• IBM's TMLE implementation also incorporate this feature (they named it importance_sampling):
  • https://github.com/BiomedSciAI/causallib/blob/master/causallib/estimation/tmle.py
  • https://github.com/BiomedSciAI/causallib/blob/master/examples/TMLE.ipynb

[pzivich]

Hi @miaow27

I haven't looked at the latest release of R's tmle. What you say and have linked to covariate vs. weighting in the target model is consistent with my knowledge of TMLE. One of my other libraries for TMLE with network-dependent data (mossspider) uses the weighting approach.

Thanks for the suggestions, I will try to integrate in the next few weeks. I may change the syntax a bit (target_gwt is not the clearest term to me, even though it is what they use in tmle). I know there are some tests that will need to be updated. I will also need to make.

This is also a reminder for me, but I also need to check and update the cross-fit TMLE's as well.