引導式測驗 · 中等 等級
順序測驗和偷看問題測驗
解釋為什麼計劃外的偷看會改變錯誤率,並比較固定的臨時觀察、阿爾法支出和始終有效的推理。
問題 1 的 8
Why can stopping when an ordinary p-value first falls below 0.05 inflate false positives?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- Why can stopping when an ordinary p-value first falls below 0.05 inflate false positives?
- Repeated looks create multiple chances to cross the threshold
- The p-value stops being numeric
- Randomization is removed automatically
- The effect size becomes zero
- What does a group-sequential design specify in advance?
- The exact future treatment effect
- Interim analysis times and evidence boundaries
- Every user's response
- The final metric value
- What does an alpha-spending function allocate?
- Model weights across layers
- Sample labels across treatment groups
- Type-I error across planned interim looks
- Traffic costs across regions
- How do always-valid methods differ from unadjusted repeated p-values?
- They remove the need for random assignment
- They assume the test ends before data arrive
- They guarantee every effect is real
- They are designed for valid inference under data-dependent stopping
- Why define interim looks before seeing experiment results?
- It makes the data independent
- The schedule and boundaries determine the inference procedure
- It guarantees early significance
- It removes delayed labels
- What can early stopping do to an observed effect estimate?
- Remove uncertainty
- Guarantee the estimated effect is larger
- Make an extreme early estimate less stable
- Prevent multiple testing
- Which principle remains necessary after choosing a valid sequential stopping method?
- Report the smallest ordinary p-value as if it were prespecified
- Assume early estimates are precise because the test is sequential
- Stop at any time without accounting for the chosen method
- Prespecify randomization, metrics and handling of delayed outcomes
- What should a report include after a sequential experiment?
- The stopping rule and timing and number of looks
- Only the smallest p-value observed
- 模型的訓練損失
- A statement that all future results will match