Quick definition: the ability for another analyst or the original team to recreate an experiment’s result and decision from documented inputs, code, definitions, and data rules.
What is Reproducibility?
Reproducibility is a core concept in disciplined experimentation because it connects a proposed change to a credible decision. Reproducibility differs from replication. Reproducibility reruns the same analysis on the same governed data; replication tests whether a result appears again in a new population or time period. Both strengthen trust, but they answer different questions. In a useful experiment, the term must be defined in relation to a population, a version of the product, a measurement window, and a decision owner. Those details keep a familiar label from becoming a vague dashboard claim.
Online experiments are valuable because they create a concurrent comparison between experiences. The comparison can support a causal conclusion only when eligibility, assignment, delivery, and measurement are understood. Read Reproducibility as part of that system: it is not a substitute for a clear hypothesis, trustworthy event data, or a decision criterion. For an overview of the full method, see A/B testing.
How reproducibility works in an experiment
A reproducible experiment stores the hypothesis, eligibility query, assignment version, exposure rule, metric definitions, analysis population, exclusions, statistical method, code version, and data snapshot or lineage. The readout should show counts and transformations sufficiently clearly that a reviewer can rebuild the estimate. Versioned definitions prevent a later dashboard change from silently rewriting a past conclusion.
The experiment record should also distinguish pre-treatment facts from outcomes created after assignment. A person may be eligible but never exposed, may be exposed but not engage, or may engage without completing the primary outcome. Conflating those states changes the question being answered. Stable definitions make the result interpretable by product, engineering, analytics, and finance teams months later.
Measurement needs an explicit numerator, denominator, attribution window, and treatment version. For example, a conversion rate can mean orders per assigned user, orders per exposed user, or orders per session; each can be defensible for a different question, but they are not interchangeable. Use a single prespecified primary metric for the central decision and retain supporting metrics to diagnose mechanisms and unintended effects.
Practical scenario
A growth team revisits a year-old referral test after revenue dashboards are updated. Its stored analysis notes reveal that the original primary metric excluded refunds within a 30-day window and used account-level assignment. Re-running the versioned query matches the published estimate. A new query using the current dashboard differs, but the team can explain the difference instead of questioning whether the original winner was fabricated.
This example illustrates why an observed number is only one input. The team should report group counts, outcome rates or means, absolute and relative differences where helpful, an uncertainty interval, data maturity, and prespecified guardrails. It should also say what the result does not establish. A conclusion tied to a defined population and window is more useful than an inflated claim that obscures conditions.
Decision workflow
Create an experiment record at planning time and link it to code, configuration, and metric versions. Lock the final analysis dataset or retain durable lineage and access controls. Have an independent reviewer rerun the primary estimate before a high-impact decision. Document deviations, late data, and manual corrections. Use replications or holdouts when generalizability, not merely calculation, is uncertain.
- Frame the decision: state what action could change and what evidence would support it.
- Lock core definitions: document eligibility, assignment, exposure, outcomes, and the primary comparison before launch.
- Verify execution: inspect allocation, event completeness, version delivery, and data latency while the test runs.
- Interpret the whole result: combine effect size, uncertainty, practical value, guardrails, and operational context.
- Record the learning: preserve the configuration, analysis, decision, and unresolved questions for later reuse.
When evidence is insufficient, “not yet decided” is often the most accurate outcome. An interval that spans a meaningful gain and a meaningful loss does not prove the intervention has no effect; it says the current information cannot separate those possibilities. Conversely, a narrow interval around a negligible effect can be a valuable reason to stop investing. Guidance on planning a viable comparison is available in the experiment hypothesis and baseline glossary entries.
Limitations and common mistakes
Exact reproduction may be impossible when identifiers expire, source data are legally deleted, vendor systems change, or real-time data are not snapshotted. Reproducible code can still implement a biased design. Avoid treating a polished notebook as evidence if it lacks assignment validation, provenance, or a clear definition of the decision population.
- Changing rules after seeing results: moving success metrics, denominators, or stopping conditions toward a favored answer weakens credibility.
- Ignoring quality checks: sample-ratio mismatch, duplicate identities, missing exposure events, and late outcomes can invalidate a polished analysis.
- Over-reading segments: unplanned cuts are useful for generating hypotheses, but many comparisons produce chance patterns.
- Forgetting guardrails: a local gain can conceal harm to reliability, customer experience, costs, or long-term value.
- Generalizing too far: a result applies first to the tested population, implementation, and time period.
These risks are not arguments against testing. They are reasons to make decisions in a way that is reviewable. An A/A test can help validate a new assignment and reporting pipeline, while a sample-ratio mismatch check can surface allocation problems before teams call a result a win.
Frequently asked questions
Why does reproducibility matter in A/B testing?
It clarifies a part of the causal and operational chain between a product change and a decision. Without that clarity, teams can compare incompatible populations, interpret noisy metrics as proof, or ship changes that do not serve the original objective.
Can a statistically significant result settle the decision?
No. Statistical evidence addresses uncertainty under a specified model. The decision also requires a meaningful effect, valid data, guardrail review, costs, risk, and a clear match between the tested and deployable experience.
What should be documented before launch?
At minimum, record the hypothesis, eligible population, assignment and exposure rules, control and treatment versions, primary metric, guardrails, planned sample or monitoring rule, attribution window, and intended decision.
How do we know whether the data are trustworthy?
Check expected versus observed allocation, assignment persistence, event completeness, metric calculations, data freshness, and unexplained differences in pre-treatment characteristics. Investigate anomalies before interpreting outcome differences.
When should the experiment be repeated?
Repeat or extend it when the estimate is too imprecise for the decision, delivery or measurement was compromised, the production implementation changed materially, or the decision must generalize to a new population, season, or risk level.
Summary
Reproducibility the ability for another analyst or the original team to recreate an experiment’s result and decision from documented inputs, code, definitions, and data rules. Used well, it supports decisions that are explicit about population, evidence, uncertainty, and trade-offs. Used casually, it can turn a convenient metric into an overconfident claim. Plan the comparison, verify execution, assess practical consequences alongside statistical evidence, and retain a reproducible decision record.
Sources
- Kohavi, Tang, and Xu, Trustworthy Online Controlled Experiments.
- NIST/SEMATECH e-Handbook of Statistical Methods.
- GOV.UK Service Manual: A/B testing.