A sub-field of statistics concerned with estimation and uncertainty quantification which are sequentially valid, i.e., remain valid at all stopping times. This enables procedures that are (i) immune to (many forms of) p-hacking (see issues with p-values), (ii) allow stopping rules to be data-dependent, and (iii) allow one to continuously monitor incoming data.
Though valid measure-theoretically, SAVI results typically rely on game-theoretic statistics and game-theoretic probability, both intuitively and mathematically.
Common tools in SAVI include:
- confidence sequences
- e-values and e-processes
- test-martingales
- non-negative supermartingales and submartingales (indeed, in some sense, SAVI must rely on these: https://arxiv.org/abs/2009.03167)