Not sure if mentioned below, but tree-based methods play around w/ adding layers and layers of interactions. Hyperparameters and holdout data prevents fitting a fully saturated final model. Approach does well for prediction/probability attempts, but can be problematic for stat/causal inference.

Yes, you are correct. But OP was only interested in the 'direct' effect of A on Y. Adjusting for D closes the partial effect (A->D->Y) and opens path A->D->B->Y. So you then need to adjust for B to close it. Comment, DAG should have B to the left of D to help elucidate causal structured ordering.

I keep staring at this and it definitely looks way more like a heelflap! Or a variant of a kickflip (e.g., varial or tre flap), but not a traditional kickflap.

Yes, space constraints resulted in use of acronyms. BS: bootstrap and EF was supposed
to be: efficient influence function (EIF)! Thanks.

Fun read. Comments: 1) You fed the models the correct covariates, just didn't define correct DGF. You didn't included IV's, mediators, or colliders as terms - you provided the exact vars it needed to estimate the DGF. 2) Show it recovers treatment prevalence. 3) Thought BS or EF needed to get SEs?

I read his Simulation City (sp?) book and loved it. If you haven't read that book yet, I would strongly recommend it.

Unless I can see observed data and/or observed holdout data plotted as well, I really don't know if one of these fits is actually better. Regardless, nice figure!

I was reading an old book last night and the author spelled it as "dis-ease". That may help you frame where it may have originally came from.

Once you can do high rep pull-ups, it is night and day. You look forward to them. I have found for quantity not quality, just to do as many as possible as fast as possible before your body even realizes what is going on.

This is also used when creating synthetic data (e.g., generative AI), and examining new data versus source data.

There was an episode on Quantitude about it earlier this season with references on their page for show notes.

I like this one, but also make the median estimate visible (overlaid).

Could a randomly selected human correctly predict all of the images? Is there a true gold-standard comparator or are there inherent mistakes in the original labels? If no mistakes, than ML just isn't there yet!

Productivity gains in society? One could look at changes as preprint servers increased.