1. Definition
- A stopping rule is the predefined condition under which an experiment or test is stopped.
- It tells you when to stop collecting data and make a decision.
- Without clear stopping rules, you risk inflating Type I error (false positives) by “peeking” until results look significant.
2. Types of Stopping Rules
(a) Fixed-Horizon Stopping Rule (Traditional A/B Test)
- Decide sample size beforehand (e.g., 10,000 users per group).
- Stop once sample is reached.
- One analysis at the end.
- Simple, protects α, but inflexible.
(b) Group Sequential Stopping Rules
- Pre-plan interim looks at data (e.g., every 25% of sample).
- Use α-spending rules (like O’Brien–Fleming, Pocock).
- Can stop early if results are overwhelming (efficacy) or hopeless (futility).
- More efficient, widely used in clinical trials.
(c) Sequential Probability Ratio Test (SPRT)
- Evaluate likelihood ratio after each data point.
- If ratio exceeds thresholds → stop for H₁ or H₀.
- Otherwise → keep sampling.
- Often requires fewer samples than fixed designs.
(d) Bayesian Stopping Rules
- Stop when posterior probability of H₁ (or H₀) exceeds a threshold.
- Example: Stop when $P(H_1 \mid \text{data}) > 0.95$.
- More intuitive (gives probability statements), no α-spending needed.
(e) Ethical / Practical Stopping Rules
- In clinical trials:
- Stop if treatment shows harm (safety).
- Stop if treatment shows clear benefit (unethical to withhold).
- Stop if treatment shows futility (unlikely to ever show benefit).
3. Examples
A/B Test Example
- Fixed rule: Run until 50,000 visitors per group.
- Sequential rule: Check every 5,000 visitors. Stop early if p < 0.001 (OBF rule).
- Bayesian rule: Stop when posterior probability > 0.95 that B > A.
Clinical Trial Example
- Planned for 1,000 patients.
- Interim checks after 250, 500, 750.
- Stopping rules:
- If new drug reduces mortality significantly early → stop.
- If harm observed → stop immediately.
- If no chance of benefit → stop for futility.
4. Why Stopping Rules Matter
- Prevent p-hacking (stopping as soon as results look “good”).
- Ensure correct Type I error control.
- Save time, cost, and resources (stop early if results are clear).
- Protect patients/participants (clinical trials).
5. Key Takeaways
- Stopping rules = when and how you end an experiment.
- They can be fixed-horizon (traditional) or sequential/adaptive (group sequential, SPRT, Bayesian).
- Good stopping rules protect against false positives while improving efficiency and ethics.
In short:
Stopping rules are pre-specified conditions for when to stop an experiment. Traditional rules use a fixed sample size, while modern sequential and Bayesian rules allow interim looks or continuous monitoring without inflating false positives.
