> Unfortunately, nobody does this. Rather, people collect data and only afterwards come here to ask how to analyze it. As you write, this directly leads us into the garden of forking paths by looking at data before deciding how to analyze it.

Legit lament (courtesy @StephanKolassa)

That is the sad truth of it. In industry you can end up having access to huge amounts of historical data. Some of it is good enough to be useful, but not all of it.

True.

@Galen Sometimes you have this access. Often you don't. The usefulness question only comes after. I still remember times when I was told to develop a functionality with zero data to test it on, because the deal depended on having that functionality. (The functionality didn't work well, IMO because the signal was too weak anyway, but the customer was deliriously happy with the rest of what we do, so mostly a good outcome... except that we now had to maintain that functionality forever.)

> Nowadays a lot of science is stuck into large scale projects costing a lot of money, time and power (and rich people decide to make spaceX instead of meaningful science). The childish and naive approaches are gone and it is a full-time job.

- courtesy SextusEmpiricus

"For forty years the Russian author Vladimir Sorokin has been detonating his own parameters – physiological, sexual, scatological, psychotic, cannibalistic, bestial, surreal, absurd, hallucinatory." Times Literary Supplement 28 June 2024. It's bad enough that politicians, pundits and literary types hijacked the word parameter and by some confusion with perimeter turned into a pretentious word for limit. Any way, which parameters in statistics or machine learning should be detonated?

I am aware of detonator plot as yet another term for a plunger plot or dynamite plot, much used especially by not very statistical biologists and medics, and much criticised by most people with more statistical skill.