3 ms·
Generalized Sequential Probability Ratio Test for Families of Hypotheses [pdf]
- measurablefunc 8mo agoGreat. How do I use this in my life to make things better?
- glial 8mo agoThe SPRT is probably already making your life better: it's used to decrease the cost of medical trails, optimize classifications in high-stakes examinations (i.e. for medical certifications), detect defective manufacturing processes, etc. It sounds like this paper extends the method to groups of hypotheses, whereas the basic version is limited to a null hypothesis and an alternative hypothesis.
- data-ottawa 8mo agoThis helps with determining when have you observed enough data to make a decision. A/B tests, monitoring metrics, health, quality control all use this. If you use LLMs, you might use this to determine if a model update or prompt change impacts results using fewer tokens.
- wdkrnls 8mo agoImplement a statistical software suite that ubiquitously uses this framework instead of the usual hierarchical mixed modeling tools whose assumptions often don't match what experiments were actually done.
- srean 8mo agoYou can search for the "peeking" problem in A/B testing. SPRT also very likely helped win a major war that involved many nations.
- noelwelsh 8mo agoFrom 2014, with 43 cites in Google Scholar: https://scholar.google.com/scholar?cites=4354840623154369813&as_sdt=2005&sciodt=0,5&hl=en https://scholar.google.com/scholar?cites=4354840623154369813... Might want to check the cites for more recent work ("Efficiency in sequential testing" looks relevant) and also the literature on "bandit best arm identification", which seems to be distinct from this line of work but tackles broadly the same problem.