Experiment Sensitivity Optimization

Experiment Sensitivity Optimization

Use this skill to redesign an experiment so it can detect meaningful effects with fewer users, less time, or clearer metrics. It focuses on minimum detectable effect, metric sensitivity, capping, variant reduction, CUPED, and variance reduction.

Source Traceability

Primary source: Next-Level A/B Testing by Leemay Nassery. Guidance is transformed and paraphrased from Chapter 3 on experiment design, sensitive metrics, minimum detectable effect, capping, reducing variants, and CUPED; and Chapter 6 on stratified random sampling and covariate adjustments.

Related skills:

• ab-test-design-brief for baseline experiment specs.
• trustworthy-experiment-insights for judging whether a result is believable.
• experimentation-throughput-strategy for capacity and test scheduling.

Reference Routing

Workflow

1. State the decision and the smallest practically meaningful effect.
2. Check whether the current primary metric is close enough to the feature's mechanism.
3. Reduce unnecessary variants and separate learning tests from launch tests.
4. Consider metric capping, CUPED, stratification, or other variance reduction.
5. Record data prerequisites, risks, and interpretation limits.
6. Update the experiment brief with the revised measurement plan.

Output Format

# Experiment Sensitivity Plan

## Decision
[What the experiment must decide.]

## Current Constraint
[Traffic | Duration | Noisy metric | Too many variants | Weak proxy | Other]

## Recommended Changes
| Change | Why It Helps | Requirement | Risk |
|--------|--------------|-------------|------|

## Metric Plan
- Primary metric:
- More sensitive alternative:
- Guardrails:
- Minimum detectable effect:

## Variance Reduction
- Technique:
- Data needed:
- Validation:

## Interpretation Notes
- What this design can conclude:
- What it cannot conclude:

Quality Bar

• Do not optimize sensitivity by switching to a metric that no longer answers the product decision.
• Do not add CUPED, stratification, or capping unless the data requirements and interpretation risks are named.
• Do not keep extra variants when they are not needed for the decision.
• Do not treat a smaller detectable effect as useful unless it is practically meaningful.