A/B testing — design and read a defensible experiment
An experiment without a pre-committed sample size and a single primary metric is not an experiment. It is a dashboard you stare at until it tells you what you wanted to hear. The discipline lives almost entirely before traffic ships: a falsifiable hypothesis, one primary metric, a sample size derived from the smallest effect worth detecting, and a stop rule you cannot renegotiate at 2pm on day four.
Pre-test checklist — every line true before any traffic
Each one is a place experiments die silently.
- A falsifiable hypothesis — names the change, the direction, and the metric it moves.
- Exactly ONE primary metric. More than one primary = multiple comparisons = inflated false positives.
- Guardrail metrics — what you refuse to harm (latency, refunds, unsubscribes) even for a win.
- The randomization unit = the analysis unit (usually the user). Mixing them is pseudoreplication.
- An MDE — the smallest lift that would change a decision. Not "any difference."
- A computed sample size and the duration it implies at your real daily eligible traffic.
- A fixed stop rule — a date or an n you commit to before launch. No "we'll see how it looks."
Step 1 — Hypothesis and metrics
State a null you can reject. "The new checkout button changes purchase conversion" with H0: conversion equal across arms, H1: it differs. Vague aspirations ("improve the funnel") have no rejection region.
Pick one primary metric and freeze it. Why: every extra primary metric is another coin flip at α, so three "primary" metrics turn a 5% false-positive rate into roughly 14%. Demote the rest to secondary.
Randomize on the same unit you analyze on. If a user sees the variant on every visit, randomize by user, not by session — analyzing 50k sessions from 8k users treats correlated observations as independent and fabricates significance.
textBad: "We think the redesign will improve engagement and revenue and retention." (no null, 3 primaries, no number) Good: "H0: 30-day purchase conversion is equal between control and the new one-click button. H1: it differs. Primary: purchase conversion. Guardrails: refund rate, p95 checkout latency. Randomize by user_id. MDE: +1.5pp absolute on a 12% baseline."
Step 2 — Sample size from MDE, baseline, and power
Defaults: power 0.80, α 0.05 (two-sided). The MDE is yours to choose — it is the smallest effect that would actually change what you do.
Rule: required n scales with ~1/MDE². Why: halving the smallest effect you care to detect roughly quadruples the traffic and time. This is the single most expensive decision in the design, so set the MDE to a business threshold, never to "whatever is small."
For a conversion rate (proportion):
pythonfrom statsmodels.stats.power import NormalIndPower from statsmodels.stats.proportion import proportion_effectsize p1, p2 = 0.12, 0.135 # baseline, baseline + MDE (1.5pp) h = proportion_effectsize(p1, p2) # Cohen's h (arcsine transform) n = NormalIndPower().solve_power(effect_size=h, alpha=0.05, power=0.80, ratio=1.0) print(int(-(-n // 1))) # n PER ARM, rounded up
For a continuous metric (revenue per user, time on page) use Welch-style sizing:
pythonfrom statsmodels.stats.power import TTestIndPower effect = mde_in_units / pooled_std # Cohen's d n = TTestIndPower().solve_power(effect_size=effect, alpha=0.05, power=0.80, ratio=1.0)
Then convert n to a calendar plan: days = ceil((n_per_arm * num_arms) / daily_eligible_users). If that
is 9 days, run a clean two full weeks anyway — weekday/weekend mix is part of the population, and a
6-day test oversamples whoever shows up Tuesday. Full worked example (12% baseline, +1.5pp MDE, 80%
power) plus runnable sizing, n→duration, CUPED θ and SRM snippets: references/sample-size-and-cuped.md.
Step 3 — Run discipline
Fixed horizon is the default. Commit to the n/date from Step 2 and read the result once, at the end.
Do not peek and stop at first significance. Why: checking repeatedly and stopping the moment p < 0.05 inflates the Type-I error far above 5% — with enough looks, a null test crosses 0.05 most of the time. If you genuinely need to stop early, use a sequential / always-valid method (confidence sequences, e.g. Netflix's anytime-valid CIs) that holds Type-I error under continuous monitoring. Sequential is strong for killing losers early and weak for calling winners early — for a confident win, the fixed-horizon read is tighter.
Gate on SRM before you trust anything. Compute a chi-square test on the observed split versus the intended ratio. If p < 0.001 the assignment or logging is broken — a bot filter dropping one arm, a redirect, a caching bug. Fix the instrumentation and rerun; do not "adjust for it."
The peeking Type-I math, sequential/always-valid options, SRM diagnosis, novelty/primacy effects,
Simpson's paradox in segments and HARKing all live in references/pitfalls.md.
Step 4 — Analyze
Pick the test by metric type:
| Metric type | Test |
|---|---|
| Binary conversion (proportion) | Two-proportion z-test (statsmodels.stats.proportion.proportions_ztest) |
| Continuous, roughly normal / large n | Welch's t-test (scipy.stats.ttest_ind(..., equal_var=False)) |
| Continuous, heavy-tailed / skewed (revenue) | Mann-Whitney U, or t-test on a log/winsorized metric |
Report lift + confidence interval + p-value together. Never p alone. Why: p < 0.05 with a CI of [+0.1pp, +5pp] is "statistically there, practically a coin toss" — the CI tells you the size, p only tells you it is not exactly zero. Practical significance = compare the CI to your MDE: if the whole interval sits above the MDE, ship; if it straddles the MDE, you detected something too small to matter.
Multiple comparisons. Two regimes:
- Small set of pre-declared decision metrics → Bonferroni (divide α by the count). Conservative, simple.
- Large exploratory scan of many metrics/segments → Benjamini-Hochberg (FDR). It keeps far more power than Bonferroni on big scans (in a 20-effect example, ~17 detected vs ~12 under Bonferroni).
Step 5 — CUPED variance reduction
CUPED (Controlled-experiment Using Pre-Experiment Data) subtracts predictable pre-period noise so the same traffic buys more power — or the same power needs less traffic. The adjusted metric:
textY_cuped = Y − θ · (X − E[X]) where θ = Cov(Y, X) / Var(X)
Estimate θ by regressing the in-experiment metric Y on the pre-experiment covariate X (e.g. each
user's spend in the 4 weeks before the test), then analyze Y_cuped with the same test as Step 4.
When it pays: recurring users with a strong pre-period signal. Reported wins — Netflix ~40% variance reduction on engagement, Statsig 50%+ on common metrics → significance in roughly half the time/traffic.
When it does nothing — do not bother: brand-new users (no pre-period data), a covariate uncorrelated
with the outcome, or — the cardinal sin — a covariate measured after assignment, which biases the
estimate. The covariate MUST be pre-treatment and independent of which arm a user lands in. Runnable
θ-via-OLS snippet in references/sample-size-and-cuped.md.
Anti-patterns
| Bad | Why it is wrong | Do instead |
|---|---|---|
| Peek daily, stop the day p < 0.05 | Repeated looks inflate Type-I error far above α | Fix n/date up front; or a sequential method that holds α |
| No sample size set before launch | You will stop on noise and call it a win | Compute n from MDE/baseline/power in Step 2 |
| Several "primary" metrics | Each is a coin flip at α; 3 metrics ≈ 14% false-positive | One frozen primary; the rest are secondary |
| Ignore the observed split | An SRM means assignment/logging is broken; results are garbage | Chi-square SRM gate before reading anything |
| Report only the p-value | Hides effect size — p < 0.05 can be practically zero | Always lift + CI + p; compare CI to MDE |
| CUPED on a post-assignment covariate | Covariate correlated with the arm biases θ | Use only pre-treatment, assignment-independent covariates |
| Call a winner from an underpowered test | "Not significant" then ≠ "no effect"; you lacked power | Reach planned n, or report the CI and say "inconclusive, here is the range" |
| Decide the hypothesis after seeing results (HARKing) | Turns the whole analysis into a fishing expedition | Pre-register hypothesis + primary metric before launch |
| Run 6 days because it "looks significant" | Oversamples one weekday slice of the population | Run full weeks; honor the fixed horizon |
Checkable artifact
When this skill emits a Python sizing/analysis script or an experiment-design doc, run
scripts/verify.sh from your project root. It confirms the script executes under python3 and prints a
numeric sample size, and that any design doc names a primary metric, an MDE, and power/alpha. It is
read-only and soft-passes when no artifact is present (a design-only conversation).
