This is an unfortunate reality that I have encountered in many (dare I say most?) fields, including in physics where my path in statistics first began a long, long time ago. A short thread.
Good applied statisticians who can work with scientists and industrial practitioners to design experiments and tease apart subtitles in how those experiments are actually realized to make robust and generalizable inferences from data and domain expertise are rare.
Those practitioners are then left to fend for themselves, learning what they can from whatever resources they happen upon and (re)deriving the rest. Unfortunately this often leads to foundational misinterpretations that poisons everything build up from there.
Those self-taught and self-derived methods then become the orthodoxy in those fields so that even when applied statisticians are available and interested in collaborating their criticisms conflict with the orthodoxy and are ignored if not outright attacked.
It's a dangerous feedback cycle that allows dangerous misconceptions about modeling, computation, and inference to propagate under the guide of empirical heuristics and folklore in those fields.
But wait, it gets worse! With applied fields fending for themselves there were less opportunities for applied statisticians. As the stats academy largely retreated into clean theory the funding for applied stats work dried up, and even fewer applied statisticians were trained.
So then when there is coherent interested in more formal analysis methods there aren't enough applied statisticians around to teach good foundations and robust methods. Instead it falls to other fields like computer scientists.
Which then leads to analysis "innovations" like machine learning and data science which piece together some relevant material while missing critical aspects that applied statisticians have been yelling about _since the beginning of the 20th century_.
Applied statisticians continue on in smaller communities where they can sustain themselves, the fields lucky to be integrated early enough or industries that were build on proper experimental designs. Their wisdom hidden in plain sight.
I do not like Fisher's approach to inference but hot damned did he understand the limitations of experiment and lay out a wealth of knowledge on basic experimental design that is relevant to anyone trying to learn from data today.
In grad school I was that young, overconfident physicists who had just learned a little bit about statistics and ranted to anyone who would listen about how poorly the field was doing things.
While I was correct about a few things I was completely ignorant about so, so, so much more. It was only after having the extreme luck to learn prob and stats foundations properly in my post doc that I realized just how difficult the problem is.
I'm going to keep working to provide the pedagogy and tools to help propagate better methods into applied fields, but practitioners have to do their share, too. They can't ignore criticisms from applied statisticians by rationalizing the sufficiency of their heuristics.
Listen to applied statisticians and place their critiques in the context of your own work to understand the relevance. Read and collaborate as much as you can outside of your field. While the system is broken _everyone_ can have a positive impact towards fixing it.
Sorry for the decidedly not short thread. I should have known better.
