Statistical Analyst

Mode 2 — Size an Experiment (Pre-Launch)

Use before launching a test to ensure it will be conclusive.

1. Define — Baseline rate, minimum detectable effect (MDE), significance level (α), power (1−β)
2. Calculate — Run sample_size_calculator.py to get required N per variant
3. Sanity-check — Confirm traffic volume can deliver N within acceptable time window
4. Document — Lock the stopping rule before launch to prevent p-hacking

Mode 3 — Interpret Existing Numbers

Use when someone shares a result and asks "is this significant?" or "what does this mean?"

1. Ask for: sample sizes, observed values, baseline, and what decision depends on the result
2. Run the appropriate test
3. Report using the Bottom Line → What → Why → How to Act structure
4. Flag any validity threats (peeking, multiple comparisons, SUTVA violations)

Tools

scripts/hypothesis_tester.py

Run Z-test (proportions), two-sample t-test (means), or Chi-square test (categorical). Returns p-value, confidence interval, effect size, and a plain-English verdict.

# Z-test for two proportions (A/B conversion rates)
python3 scripts/hypothesis_tester.py --test ztest \
  --control-n 5000 --control-x 250 \
  --treatment-n 5000 --treatment-x 310

# Two-sample t-test (comparing means, e.g. revenue per user)
python3 scripts/hypothesis_tester.py --test ttest \
  --control-mean 42.3 --control-std 18.1 --control-n 800 \
  --treatment-mean 46.1 --treatment-std 19.4 --treatment-n 820

# Chi-square test (multi-category outcomes)
python3 scripts/hypothesis_tester.py --test chi2 \
  --observed "120,80,50" --expected "100,100,50"

# Output JSON for downstream use
python3 scripts/hypothesis_tester.py --test ztest \
  --control-n 5000 --control-x 250 \
  --treatment-n 5000 --treatment-x 310 \
  --format json

scripts/sample_size_calculator.py

Calculate required sample size per variant before launching an experiment.

# Proportion test (conversion rate experiment)
python3 scripts/sample_size_calculator.py --test proportion \
  --baseline 0.05 --mde 0.20 --alpha 0.05 --power 0.80

# Mean test (continuous metric experiment)
python3 scripts/sample_size_calculator.py --test mean \
  --baseline-mean 42.3 --baseline-std 18.1 --mde 0.10 \
  --alpha 0.05 --power 0.80

# Show tradeoff table across power levels
python3 scripts/sample_size_calculator.py --test proportion \
  --baseline 0.05 --mde 0.20 --table

# Output JSON
python3 scripts/sample_size_calculator.py --test proportion \
  --baseline 0.05 --mde 0.20 --format json

scripts/confidence_interval.py

Compute confidence intervals for a proportion or mean. Use for reporting observed metrics with uncertainty bounds.

# CI for a proportion
python3 scripts/confidence_interval.py --type proportion \
  --n 1200 --x 96

# CI for a mean
python3 scripts/confidence_interval.py --type mean \
  --n 800 --mean 42.3 --std 18.1

# Custom confidence level
python3 scripts/confidence_interval.py --type proportion \
  --n 1200 --x 96 --confidence 0.99

# Output JSON
python3 scripts/confidence_interval.py --type proportion \
  --n 1200 --x 96 --format json

Test Selection Guide

When NOT to use these tools:

• n < 30 per group without checking normality
• Metrics with heavy tails (e.g. revenue with whales) — consider log transform or trimmed mean first
• Sequential / peeking scenarios — use sequential testing or SPRT instead
• Clustered data (e.g. users within countries) — standard tests assume independence

Decision Framework (Post-Experiment)

Use this after running the test:

Always ask: "If this effect were exactly as measured, would the business care?" If no — don't ship on significance alone.

Effect Size Reference

Effect sizes translate statistical results into practical language:

Cohen's d (means):

Cohen's h (proportions):

Cramér's V (chi-square):

Proactive Risk Triggers

Surface these unprompted when you spot the signals:

• Peeking / early stopping — Running a test and checking results daily inflates false positive rate. Ask: "Did you look at results before the planned end date?"
• Multiple comparisons — Testing 10 metrics at α=0.05 gives ~40% chance of at least one false positive. Flag when > 3 metrics are being evaluated.
• Underpowered test — If n is below the required sample size, a non-significant result tells you nothing. Always check power retroactively.
• SUTVA violations — If users in control and treatment can interact (e.g. social features, shared inventory), the independence assumption breaks.
• Simpson's Paradox — An aggregate result can reverse when segmented. Flag when segment-level results are available.
• Novelty effect — Significant early results in UX tests often decay. Flag for post-novelty re-measurement.

Output Artifacts

Quality Loop

Tag every finding with confidence:

• 🟢 Verified — Test assumptions met, sufficient n, no validity threats
• 🟡 Likely — Minor assumption violations; interpret directionally
• 🔴 Inconclusive — Underpowered, peeking, or data integrity issue; do not act

Communication Standard

Structure all results as:

Bottom Line — One sentence: "Treatment increased conversion by 1.2pp (95% CI: 0.4–2.0pp). Result is statistically significant (p=0.003) with a small effect (h=0.18). Recommend shipping."

What — The numbers: observed rates/means, difference, p-value, CI, effect size

Why It Matters — Business translation: what does the effect size mean in revenue, users, or decisions?

How to Act — Ship / hold / extend / kill with specific rationale

Related Skills

When NOT to use this skill:

• You need to design or instrument the experiment — use marketing-skill/ab-test-setup or product-team/experiment-designer
• You need to clean or validate the input data — use engineering/data-quality-auditor first
• You need Bayesian inference or multi-armed bandit analysis — flag that frequentist tests may not be appropriate

References

• references/statistical-testing-concepts.md — t-test, Z-test, chi-square theory; p-value interpretation; Type I/II errors; power analysis math