Профиль

This is nominally about data analysis.

Normalize the expectation of minimum costs, so that people no longer feel so entitled to fast and easy options that they will aggressively dismiss alternatives.

I’m sure if people have been more disincentivized, if not outright punished, for asking questions lately, if asking questions is a skill that people must actively learn but that learning isn’t happening, or if it’s something else entirely. Definitely unsettling, regardless.

But hopefully people might have suggestions.

In my courses (and eventually updated writing) I explicitly discuss the replicability behavior (no guarantee of bitwise reproducibility of Markov chains, but assuming a CLT MCMC estimates are consistent within the MCMC-SE, etc) but I've rarely seen applied work go into anywhere near that much depth.

There was lots of folklore and folk practices back in the BUGS days given the relative inefficiency of the sampler and poor model parameterizations, but I didn't encounter too much actually in print. Same for people who disregarded Stan because all of the sudden there are more diagnostic warnings.

To be honest, I gave up trying to collate bad MCMC takes a long time ago. I would suggest every variational inference paper offering bad takes in their introductions, but it sounds like you're more looking for discussions based on using MCMC estimates when the Markov chains hadn't fully converged?

Perhaps I'm just disillusioned and bitter, but my experience has always been people thinking that it should be the first, and only, thing that they should try. Most people are conditioned to believe that data analysis should be a-nearly automatic step and refuse to consider anything else.

People have been doing this since the first public version of Stan, long before LLMs became ubiquitous. A subtle, effortful approach will always be hard to sell against the promise of a quick fix, regardless of their relative degrees of productivity.

As always, a huge thanks to my supporters over on patreon dot com, www.patreon.com/c/betanalpha, who make freely available work like this feasible. Incidentally, covector+ supporters also have access to a 2.5 hour video review of this latest piece.

Michael Betancourt — writing about modeling building and statistical inference.Get more from Michael Betancourt on Patreon. writing about modeling building and statistical inference.. Support Michael Betancourt and get exclusive access to their work.www.patreon.com

The prior modeling is careful and deliberate, demonstrating how principled prior models can be developed from available domain expertise in real analyses.

Unlike most of my public case studies, this piece builds up the analysis in Python instead of R. That said, the modeling approach is the same: build up to an adequate model iteratively, using retrodictive tension to motivate productive updates at each step.

Or, more productively, there are lots of different insights that one could communicate, each of which is best communicated through a different visualization. Consequently, other visualizations could be useful here, but that doesn't mean that the one shown is wrong in any absolute sense.

All of that said, I also think that it's fair to argue that the point of a weekly forecast is to communicate weekly dynamics, reserving the consequences to the observer. Here it's straightforward to absorb that the temperature will mostly be the same all week but hotter for two days in the middle.

Because the physiological reaction to temperature if far from linear, I think that a few extra degrees above a high baseline can be just as impactful as 10-20 degrees from a much lower baseline. Of course, physical health verses physical comfort are different scales!