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。

相關引導路徑順序測試和窺視問題
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。

  1. 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
  2. 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
  3. 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
  4. 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
  5. 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
  6. 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
  7. 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
  8. 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