Hi Judith,

You're right, tfci was implemented on top of tensorflow probability, not pymc.

Pycausalimpact was implemented on top of statsmodels. Very unfortunately the company that it was developed for deleted the repository so we're left with only tfci for now (I do have plans to also integrate it in tfci but couldn't find available time for doing so lately).

As for differences, tfci follows the same bayesian approach than the original R package does. pyci OTOH uses frequentist stats instead and therefore results at times may vary (you may find that in one package results are statistically significant whereas the other it's not).

Another main difference is that tensorflow probability proved to be painfully slow and poorly implemented to the point where pymc4 aborted its adoption. Hopefully we'll have better tools soon in the community to start using new and better tools for this analyses.

Hope that helps :)