https://abtest.chat/tags/msprt/Recent content in MSPRT on 展博增长实验室 | 增长决策可信度诊断https://abtest.chat/images/og-default.svghttps://abtest.chat/images/og-default.svgHugozh-CNTue, 03 Nov 2026 00:00:00 +0800https://abtest.chat/posts/%E5%AE%9E%E9%AA%8C%E8%B7%91%E4%B8%A4%E5%91%A8%E7%9C%8B%E5%88%B0%E6%98%BE%E8%91%97%E5%B0%B1%E5%85%A8%E9%87%8F%E5%81%87%E9%98%B3%E6%80%A7%E5%8F%AF%E8%83%BD%E5%B7%B2%E4%BB%8E-5-%E6%8A%AC%E5%88%B0-20/Tue, 03 Nov 2026 00:00:00 +0800https://abtest.chat/posts/%E5%AE%9E%E9%AA%8C%E8%B7%91%E4%B8%A4%E5%91%A8%E7%9C%8B%E5%88%B0%E6%98%BE%E8%91%97%E5%B0%B1%E5%85%A8%E9%87%8F%E5%81%87%E9%98%B3%E6%80%A7%E5%8F%AF%E8%83%BD%E5%B7%B2%E4%BB%8E-5-%E6%8A%AC%E5%88%B0-20/固定窗 t 检验的 5% 假阳性只在「到点看一次」时成立。每天刷看板、一显著就停属于 optional stopping，会把假阳性抬到 20% 量级。序贯检验（sequential / always-valid）把「可以反复看」写进统计保证，但代价是弱效应更慢、区间更宽、早停点估计偏乐观。