LeoYeAI/agent-master-skills / skills/ab-testing
ab-testing
VerifiedWhen the user wants to plan, design, or implement an A/B test or experiment, or build a growth experimentation program. Also use when the user mentions "A/B test," "split test," "experiment," "test this change," "variant copy," "multivariate test," "hypothesis," "should I test this," "which version is better," "test two versions," "statistical significance," "how long should I run this test," "growth experiments," "experiment velocity," "experiment backlog," "ICE score," "experimentation program," or "experiment playbook." Use this whenever someone is comparing two approaches and wants to measure which performs better, or when they want to build a systematic experimentation practice. For tracking implementation, see analytics. For page-level conversion optimization, see cro.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 13.0 KB |
| evals/evals.json | config | 8.8 KB |
| references/sample-size-guide.md | markdown | 9.0 KB |
| references/test-templates.md | markdown | 6.3 KB |
references/sample-size-guide.md
View on GitHub# Sample Size Guide
Reference for calculating sample sizes and test duration.
## Contents
- Sample Size Fundamentals (required inputs, what these mean)
- Sample Size Quick Reference Tables
- Duration Calculator (formula, examples, minimum duration rules, maximum duration guidelines)
- Online Calculators
- Adjusting for Multiple Variants
- Common Sample Size Mistakes
- When Sample Size Requirements Are Too High
- Sequential Testing
- Quick Decision Framework
## Sample Size Fundamentals
### Required Inputs
1. **Baseline conversion rate**: Your current rate
2. **Minimum detectable effect (MDE)**: Smallest change worth detecting
3. **Statistical significance level**: Usually α = 0.05 (5%)
4. **Statistical power**: Usually 80% (β = 0.20)
### What These Mean
**Baseline conversion rate**: If your page converts at 5%, that's your baseline.
**MDE (Minimum Detectable Effect)**: The smallest improvement you care about detecting. Set this based on:
- Business impact (is a 5% lift meaningful?)
- Implementation cost (worth the effort?)
- Realistic expectations (what have past tests shown?)
**Significance level (α = 0.05)**: For the planned test and its assumptions, limits the long-run probability of rejecting a true null hypothesis to 5%. It is not the probability that an observed difference is “due to chance.” Repeated unadjusted looks or multiple comparisons do not retain that error rate. See the [ASA's interpretation guidance](https://www.amstat.org/asa/files/pdfs/p-valuestatement.pdf).
**Statistical power (80%)**: Means if there's a real effect of size MDE, you have 80% chance of detecting it.
---
## Sample Size Quick Reference Tables
All tables assume α = 0.05 (two-sided) and 80% power, two-proportion test, rounded to two significant figures. Use a calculator for your exact numbers.
### Conversion Rate: 1%
| Lift to Detect | Sample per Variant | Total Sample |
|----------------|-------------------|--------------|
| 5% (1% → 1.05%) | 640,000 | 1,280,000 |
| 10% (1% → 1.1%) | 160,000 | 320,000 |
| 20% (1% → 1.2%) | 43,000 | 86,000 |
| 50% (1% → 1.5%) | 7,700 | 15,400 |
| 100% (1% → 2%) | 2,300 | 4,600 |
### Conversion Rate: 3%
| Lift to Detect | Sample per Variant | Total Sample |
|----------------|-------------------|--------------|
| 5% (3% → 3.15%) | 210,000 | 420,000 |
| 10% (3% → 3.3%) | 53,000 | 106,000 |
| 20% (3% → 3.6%) | 14,000 | 28,000 |
| 50% (3% → 4.5%) | 2,500 | 5,000 |
| 100% (3% → 6%) | 750 | 1,500 |
### Conversion Rate: 5%
| Lift to Detect | Sample per Variant | Total Sample |
|----------------|-------------------|--------------|
| 5% (5% → 5.25%) | 120,000 | 240,000 |
| 10% (5% → 5.5%) | 31,000 | 62,000 |
| 20% (5% → 6%) | 8,200 | 16,400 |
| 50% (5% → 7.5%) | 1,500 | 3,000 |
| 100% (5% → 10%) | 430 | 860 |
### Conversion Rate: 10%
| Lift to Detect | Sample per Variant | Total Sample |
|----------------|-------------------|--------------|
| 5% (10% → 10.5%) | 58,000 | 116,000 |
| 10% (10% → 11%) | 15,000 | 30,000 |
| 20% (10% → 12%) | 3,800 | 7,600 |
| 50% (10% → 15%) | 690 | 1,380 |
| 100% (10% → 20%) | 200 | 400 |
### Conversion Rate: 20%
| Lift to Detect | Sample per Variant | Total Sample |
|----------------|-------------------|--------------|
| 5% (20% → 21%) | 26,000 | 52,000 |
| 10% (20% → 22%) | 6,500 | 13,000 |
| 20% (20% → 24%) | 1,700 | 3,400 |
| 50% (20% → 30%) | 290 | 580 |
| 100% (20% → 40%) | 81 | 162 |
---
## Duration Calculator
### Formula
```
Duration (days) = (Sample per variant × Number of variants) / (Daily traffic × % exposed)
For unequal allocation, calculate each arm separately:
Days per arm = Required sample for that arm / (Daily eligible traffic × Exposure fraction × Arm allocation fraction)
Total collection time = longest arm duration
```
Use fractions for exposure and allocation (50% = 0.5). Count eligible randomized units consistently with the sample-size calculation; repeat pageviews are not additional participants. The balanced formula assumes equal sample requirements and equal allocation.
### Examples
**Scenario 1: High-traffic page**
- Need: 10,000 per variant (2 variants = 20,000 total)
- Daily traffic: 5,000 visitors
- 100% exposed to test
- Duration: 20,000 / 5,000 = **4 days**
**Scenario 2: Medium-traffic page**
- Need: 30,000 per variant (60,000 total)
- Daily traffic: 2,000 visitors
- 100% exposed
- Duration: 60,000 / 2,000 = **30 days**
**Scenario 3: Low-traffic with partial exposure**
- Need: 15,000 per variant (30,000 total)
- Daily traffic: 500 visitors
- 50% exposed to test
- Effective daily: 250
- Duration: 30,000 / 250 = **120 days** (too long!)
### Minimum Duration Rules
Even with sufficient sample size, run tests for at least:
- **1 full week**: To capture day-of-week variation
- **2 business cycles**: If B2B (weekday vs. weekend patterns)
- **Through paydays**: If e-commerce (beginning/end of month)
### Maximum Duration Guidelines
Avoid running tests longer than 4-8 weeks:
- Novelty effects wear off
- External factors intervene
- Opportunity cost of other tests
---
## Online Calculators
### Recommended Tools
**Evan Miller's Calculator**
https://www.evanmiller.org/ab-testing/sample-size.html
- Simple interface
- Bookmark-worthy
**Optimizely's Calculator**
https://www.optimizely.com/sample-size-calculator/
- Business-friendly language
- Duration estimates
**AB Test Guide Calculator**
https://www.abtestguide.com/calc/
- Includes Bayesian option
- Multiple test types
**VWO Duration Calculator**
https://vwo.com/tools/ab-test-duration-calculator/
- Duration-focused
- Good for planning
---
## Adjusting for Multiple Variants
With more than 2 variants (A/B/n tests), you need more sample:
| Variants | Multiplier |
|----------|------------|
| 2 (A/B) | 1x |
| 3 (A/B/C) | ~1.5x |
| 4 (A/B/C/D) | ~2x |
| 5+ | Consider reducing variants |
**Why?** More comparisons increase chance of false positives. You're comparing:
- A vs B
- A vs C
- B vs C (sometimes)
Apply Bonferroni correction or use tools that handle this automatically.
---
## Common Sample Size Mistakes
### 1. Underpowered tests
**Problem**: Not enough sample to detect realistic effects
**Fix**: Be realistic about MDE, get more traffic, or don't test
### 2. Overpowered tests
**Problem**: Waiting for sample size when you already have significance
**Fix**: This is actually fine—you committed to sample size, honor it
### 3. Wrong baseline rate
**Problem**: Using wrong conversion rate for calculation
**Fix**: Use the specific metric and page, not site-wide averages
### 4. Ignoring segments
**Problem**: Calculating for full traffic, then analyzing segments
**Fix**: If you plan segment analysis, calculate sample for smallest segment
### 5. Testing too many things
**Problem**: Dividing traffic too many ways
**Fix**: Prioritize ruthlessly, run fewer concurrent tests
---
## When Sample Size Requirements Are Too High
Options when you can't get enough traffic:
1. **Increase MDE**: Accept only detecting larger effects (20%+ lift)
2. **Lower confidence**: Use 90% instead of 95% (risky, document it)
3. **Reduce variants**: Test only the most promising variant
4. **Combine traffic**: Test across multiple similar pages
5. **Test upstream**: Test earlier in funnel where traffic is higher
6. **Don't test**: Make decision based on qualitative data instead
7. **Longer test**: Accept longer duration (weeks/months)
---
## Sequential Testing
If you must check results before reaching sample size:
### What is it?
Statistical method that adjusts for multiple looks at data.
### When to use
- High-risk changes
- Need to stop bad variants early
- Time-sensitive decisions
### Tools that support it
- Optimizely (Stats Accelerator)
- VWO (SmartStats)
- PostHog (Bayesian approach)
### Tradeoff
- More flexibility to stop early
- Slightly larger sample size requirement
- More complex analysis
---
## Quick Decision Framework
### Can I run this test?
```
Daily eligible traffic (randomization units): _____
Fraction exposed to the experiment: _____
Number of arms (including control): _____
Allocation fraction for each arm: _____
Baseline conversion rate: _____
MDE I care about: _____
Sample needed per variant: _____ (from tables above)
Balanced allocation: Days = Sample per arm × Number of arms / (Daily eligible traffic × Exposure fraction)
Unequal allocation: Days = max across arms of [Sample required for arm / (Daily eligible traffic × Exposure fraction × Arm allocation)]
Round collection time up, then apply the minimum-duration rules above.
If days > 60: Consider alternatives
If days > 30: Acceptable for high-impact tests
If days < 14: Likely feasible
If days < 7: Easy to run, consider running longer anyway
```
**Duration check (illustrative):** with 2,000 participants required per arm and 1,000 eligible participants/day, two arms at 50/50 with full exposure require four days, not two. At 50% exposure they require eight days. With full exposure and 90/10 allocation, the smaller arm needs 20 days to reach 2,000. These are collection estimates; honor planned business cycles and the analysis stopping rule.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the work | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixes | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation code | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |