Experiments

Feature Flags

7 Statsig alternatives for A/B testing and experimentation

Aug 27, 2026

min read

A graphic of a bar chart with an arrow pointing upward.

A Statsig alternative must preserve the chain from assignment to decision. Perfect assignment still fails when exposures, units, metrics, or stopping rules are wrong.

Statsig combines flags, experiments, product analytics, and related product-data workflows. Some alternatives replace that full surface; others specialize in warehouse analysis, release testing, visual optimization, or product analytics.

OpenAI acquired Statsig in 2025, while Amplitude said it would develop the Statsig platform across cloud and warehouse deployments and support existing customers. Portability, roadmap clarity, and contract terms therefore belong in the evaluation.

This guide compares GrowthBook, Datadog Experiments, Optimizely, Kameleoon, LaunchDarkly, PostHog, and VWO. GrowthBook is the strongest overall recommendation because it combines feature delivery, rigorous experimentation, warehouse flexibility, open-source deployment, and predictable pricing.

Statsig experimentation alternatives at a glance

Alternative	Best fit	Data and analysis model	Main watchout
GrowthBook	Technical product, data, and engineering teams	Warehouse-native or managed warehouse; Bayesian, frequentist, sequential, and variance reduction options	Advanced governance and methods span paid tiers
Datadog Experiments	Data-mature teams that want Eppo methods plus Datadog context	Warehouse-native metrics with Eppo's statistical engine and observability integration	Product transition and public pricing require direct evaluation
Optimizely	Large programs spanning web and feature experimentation	Vendor event pipeline plus warehouse-native analytics options; several statistical engines	Individually packaged pricing and product boundaries can add complexity
Kameleoon	Hybrid web, product, and feature experimentation programs	Client, server, and warehouse-connected workflows with flexible statistics	Full-stack capabilities and advanced methods depend on packaging
LaunchDarkly	Experiments coupled tightly to feature releases	Flag-based assignment, event or warehouse metrics, Bayesian and frequentist analysis	Multiple usage meters and a release-first product center of gravity
PostHog	Small technical teams wanting analytics, replay, flags, and tests together	PostHog event model or data warehouse sources; Bayesian and frequentist analysis	Broad usage-based platform creates several cost and governance surfaces
VWO	Conversion teams running visual web and server-side tests	Website-oriented testing with feature experimentation and Bayesian reporting	Technical depth and package scope vary across Testing and Feature Experimentation

What to compare before choosing a Statsig replacement

Use the OpenFeature specification as an independent reference for evaluation APIs.

Randomization unit, assignment, and exposure

Choose the experimental unit before the tool. Consumer experiences often randomize a person or device. B2B products may need an account, workspace, or organization because members share the treatment. Marketplaces may require switchback or cluster designs because participants affect one another.

The platform needs a stable assignment key and consistent allocation across web, mobile, server, and edge runtimes. Test null values, identity transitions, retries, and SDK-version differences. Sticky assignment prevents variant changes but creates state that must be migrated.

Exposure is not the same as eligibility or assignment. Log a participant when the treatment can actually affect them, not merely when a page loads or a flag definition exists. Triggered analysis can improve sensitivity, but only if the trigger is not changed by treatment in a way that introduces selection bias. The broad research literature on statistical challenges in online controlled experiments treats exposure, interference, sensitivity, long-term effects, and optional stopping as connected design problems rather than dashboard settings.

Metric source and semantic governance

A good alternative makes the metric definition inspectable. Record source tables or events, joins, aggregation unit, attribution window, exclusions, outlier policy, late-data behavior, and version. A metric named revenue is not governed if one tool uses checkout events while finance uses settled payments net of refunds.

Warehouse-native analysis can reuse modeled data and run calculations near the source. An event-native platform can be faster to start and may pair naturally with real-time product analytics. Neither architecture guarantees correctness. Compare query reproducibility, freshness, compute cost, backfill behavior, row-level security, personal-data movement, and how a changed metric definition affects completed experiments.

For cross-product effects, keep assignment rows even when a subject produces no downstream event. Microsoft's guidance on A/B testing across products explains why joining only participants who appear in a second product can bias the analysis. The proof of concept should reproduce at least one metric outside the vendor UI from raw assignments, exposures, and outcomes.

Statistical method and stopping policy

Bayesian, fixed-horizon frequentist, and sequential methods answer related questions with different assumptions and decision language. The tool should document what it calculates, how priors or significance thresholds are chosen, how repeated looks are handled, and whether confidence or credible intervals remain valid under the team's actual stopping behavior.

Start from the decision risk: false launches, missed improvements, practical effect size, duration, and the cost of continuing a harmful treatment. Predefine the primary metric, guardrails, minimum detectable effect, analysis method, minimum runtime, and decision thresholds.

Multiple metrics and many variants increase the chance of a chance finding. Ask whether the platform applies a family-wise or false-discovery correction, whether it distinguishes a designated primary metric from exploratory metrics, and how segment exploration is labeled. Microsoft's paper on metric interpretation pitfalls is a useful reminder that trustworthy counts do not rescue a poorly defined decision metric.

Variance reduction, power, and data quality

CUPED and related covariate adjustment methods can improve precision by using information collected before treatment. The original Microsoft research on improving sensitivity with pre-experiment data is the conceptual baseline. Implementations differ in eligible metric types, lookback windows, covariates, outlier handling, missing history, segmented results, and how often models refresh. Ask for method documentation and compare adjusted and unadjusted outputs on known experiments.

Every finalist should detect sample ratio mismatch. SRM means the observed assignment or exposure counts differ more than expected from the configured split. It is a symptom, not an inconvenience to dismiss. Microsoft's SRM taxonomy and diagnostic guidance covers causes such as logging loss, redirects, filters, inconsistent identifiers, and variant-dependent execution. A platform should block or clearly quarantine conclusions until the root cause is understood.

Power calculators should use the same unit, variance, and metric construction as the experiment. Validate them against historical traffic. A low-traffic team often gets more value from narrowing the population, choosing a more sensitive proximal metric, reducing measurement noise, or testing a larger product change than from changing the color of the statistical badge.

Program operations and governance

Compare hypothesis templates, review checklists, mutually exclusive layers, holdouts, approvals, reports, audit logs, APIs, and decision history. Use risk tiers: a copy test may need peer review, while billing, ranking, or permission experiments may need data-science, legal, and operational sign-off.

1. GrowthBook: Best overall Statsig alternative

Best for

GrowthBook is the best fit for technical product organizations that want to keep experiment metrics in an existing warehouse, or start with a managed warehouse and retain a path to bring their own data later. It combines experimentation, feature flags, and product analytics without requiring one analytics architecture for every team.

It is also the clearest choice for buyers who want an inspectable open-source core and a supported hosted product. That deployment flexibility makes it easier to treat the current Statsig transition as an opportunity to reduce platform lock-in rather than merely swap one closed SaaS endpoint for another.

Key strengths

GrowthBook can query experiment assignments and outcomes from a warehouse, so data teams can reuse governed facts, dimensions, and SQL. Teams without that infrastructure can use the managed ClickHouse-backed path and migrate later. The practical advantage is not the phrase “warehouse native”; it is having one auditable definition for revenue, retention, activation, or cost across experiments and business reporting.

The statistical surface covers Bayesian and frequentist analysis, multiple-testing corrections, SRM detection, guardrails, and shareable results on the free tier. Pro adds sequential testing, CUPED variance reduction, sticky bucketing, power calculations, decision frameworks, and multi-armed bandits. Enterprise adds program-level controls such as holdouts, launch checklists, custom metadata, stronger access control, and audit capabilities. The public GrowthBook pricing page makes those boundaries unusually visible.

Assignment runs in SDKs using deterministic hashing, while experiment analysis can stay separate from delivery. That separation is useful during migration: a team can analyze a historical Statsig test in GrowthBook before replacing assignment, or use a different flag provider while validating the analysis layer. It also supports gradual adoption across stacks rather than a flag-day migration.

Watchouts

Warehouse-native analysis needs disciplined data modeling. Bad joins, identity drift, expensive scans, or inconsistent exposure tables remain bad even when the UI makes them easier to query. Assign an owner to each data source and metric, set cost controls, and test late-arriving events and backfills.

Some mature-program capabilities are paid. Teams that need holdouts, approval workflows, advanced access control, SSO, SCIM, or formal support should compare the relevant GrowthBook tier with the actual Statsig contract, not compare an open-source screenshot with an enterprise deployment.

Pricing and implementation notes

The Cloud Starter plan is free for up to three users and includes unlimited flags, experiments, and traffic. Pro is currently $40 per seat per month and adds the advanced experimentation methods most scaling teams expect. Enterprise is custom. The self-hosted open-source plan is free with unlimited users, flags, experiments, and traffic; supported enterprise self-hosting is custom.

For a proof of concept, connect one representative data source, import or recreate a completed experiment, define one binary metric, one revenue or ratio metric, and one guardrail, then compare counts and intervals with the old result. Run an A/A test through the production SDK path before moving decision-critical traffic. The GrowthBook versus Statsig comparison provides a product-level starting point, but the organization's raw-data reconciliation should decide the migration.

2. Datadog Experiments: Best for warehouse metrics plus observability

Best for

Datadog Experiments is the successor path for teams attracted to Eppo's warehouse-native experimentation and advanced statistics, especially when Datadog already owns application observability and product analytics. Datadog acquired Eppo in 2025 and launched Datadog Experiments in 2026, combining warehouse business metrics, product events, and operational guardrails.

That combination is differentiated for backend, AI, performance, and reliability experiments. A model or ranking change may affect conversion, latency, error rate, infrastructure cost, and token consumption at once. Keeping technical guardrails close to traces and application telemetry can shorten diagnosis when a statistically promising treatment harms service quality.

Key strengths

Eppo's warehouse-native model remains the statistical foundation. Its documentation describes frequentist fixed-sample, sequential, hybrid-sequential, and Bayesian choices, along with SRM checks, winsorization, sample-size planning, and CUPED++. CUPED++ can use historical values from multiple eligible metrics and assignment properties as covariates, rather than limiting adjustment to the pre-period value of the same outcome.

The metric layer can be defined through a UI or a code-based semantic framework. That is useful for data teams that want reviewed, versioned definitions and for product teams that need a readable catalog. Eppo also supports layers, mutual exclusion, global holdouts, entry points, and analyses over assignments produced by another system.

The current Datadog Experiments launch announcement positions the product around warehouse business metrics, product analytics events, and real-time observability guardrails. This can make it a strong replacement when Statsig experiments were already paired with Datadog dashboards and incident workflows.

Watchouts

Treat the Eppo-to-Datadog transition as a product migration to evaluate, not merely a new logo. Ask which functions live in the Datadog Experiments application today, which remain in Eppo interfaces, how existing customers are being migrated, and whether SDKs, APIs, metric definitions, links, and data-retention terms will change. Datadog has publicly described an active customer migration and roadmap, so current-state evidence matters more than older comparison pages.

Warehouse analysis can also create significant compute. Measure scanned bytes, cache reuse, refresh duration, concurrency, and the behavior of wide metric sets. CUPED++ is powerful, but buyers should understand where it applies and note that Eppo's documentation says adjusted results are not available for every filtered or exploratory view.

Pricing and implementation notes

Public self-serve pricing for Datadog Experiments is not clearly posted, so request a written quote and meter definition. Include experiment subjects, assignments, warehouse compute, product-analytics events, observability data, feature-flag traffic, seats, retention, support, and migration services. If the organization already has a Datadog commitment, ask how Experiments changes the committed-spend model.

The proof of concept should use a change with both a business metric and a technical guardrail. Compare a warehouse result, a product event, and an observability signal on the same experimental unit. Verify identity joins and timestamps, then reproduce the effect estimate from exported assignments and outcomes. This tests the integration value that distinguishes Datadog Experiments from a standalone statistical dashboard.

Compare GrowthBook and Statsig

See how warehouse-native experiments, open-source deployment, feature flags, and predictable pricing compare with Statsig.

Explore the comparison

3. Optimizely: Best for a broad enterprise experimentation program

Best for

Optimizely fits organizations that want a mature program spanning visual web tests, feature and server-side experiments, personalization, collaboration, and warehouse-native analysis. It is strongest when experimentation is a cross-functional enterprise capability rather than a tool used by one product squad.

The product family can support marketers editing a webpage, engineers testing a backend algorithm, analysts exploring segments, and program leaders managing ideas and results. That breadth is valuable when consolidation is the goal and the organization has owners for platform architecture, implementation, and statistical policy.

Key strengths

Optimizely Feature Experimentation uses SDKs and a CDN-delivered datafile to run flags and tests across applications. The current Feature Experimentation documentation describes unlimited flags and controlled rollouts on the free Rollouts tier, with one concurrent experiment; paid plans add the full Stats Engine, collaboration, governance, and greater capacity.

Optimizely's newer results experience supports frequentist fixed-horizon, sequential, and Bayesian engines, plus CUPED, outlier management, exploration, and data-health checks. The established Stats Engine includes confidence intervals and false-discovery-rate control. The broader experimentation portfolio adds visual editing, multivariate tests, bandits, audience targeting, and personalization. Warehouse-native analytics can centralize analysis over governed data while web and feature products handle delivery.

For a large organization, program-management features and services can matter as much as calculations. Optimizely can provide structured collaboration, permissions, support, and training across many teams. It is a credible choice when the buyer wants one strategic vendor for marketing and product experimentation and is prepared to manage product boundaries.

Watchouts

Breadth creates architecture and procurement work. Establish which product performs assignment, which stores events, which computes results, and which owns the experiment record. “Optimizely Experimentation” can refer to different combinations of Web Experimentation, Feature Experimentation, Analytics, Personalization, and collaboration capabilities. A capability in one package should not be assumed to exist in another.

The new A/B results page has its own eligibility limits; Optimizely's documentation says, for example, that it applies to manual A/B experiments and not every adaptive or multivariate design. Confirm the exact statistical engine and result interface for each planned experiment type. If teams mix methods, encode the choice in templates so experimenters cannot select a friendlier-looking answer after seeing data.

Pricing and implementation notes

Optimizely says paid experimentation deployments are individually packaged based on traffic, products, and implementation complexity. Rollouts offers a free entry point with one experiment at a time, but an enterprise comparison requires a quote. Ask for separate and bundled pricing across web testing, feature experimentation, warehouse analytics, personalization, collaborators, monthly active users, raw-event export, support, and professional services.

Use two proof-of-concept tracks if both marketer and product use cases matter: a visual web test that checks flicker, performance, consent, and editor workflow; and a server-side flag experiment that checks deterministic assignment, event export, metric reconciliation, and rollback. Do not let a strong visual-editor demo substitute for the feature-experiment evidence, or vice versa.

4. Kameleoon: Best for hybrid web and feature experimentation

Best for

Kameleoon is a strong option for organizations that need one experimentation program across websites, applications, server-side features, and product teams. It sits between web-optimization suites and developer-first feature platforms: non-engineers can create visual tests, while engineers can run SDK-based experiments and controlled releases.

It is especially relevant when the statistical team wants a choice among methods rather than one mandatory engine. Kameleoon publishes detailed methodology covering assignment, frequentist tests, multiple-test correction, CUPED, sequential tests, Bayesian decisions, continuous metrics, outlier handling, and bandits.

Key strengths

Kameleoon supports visual and code-based experimentation, feature flags, audience targeting, holdouts, mutually exclusive groups, raw-data access, and web performance controls. Its current plans page lists sequential testing, multiple-testing correction, SRM detection, unlimited metrics and audiences in the PBX Starter package, with feature management, feature experimentation, CUPED, personalization, mobile testing, and bandits available in the broader enterprise configuration.

The statistics are more transparent than a generic “AI-powered significance” claim. Kameleoon documents fixed frequentist and Bayesian approaches, sequential confidence sequences, multiple-comparison controls, power planning, and CUPED using pre-experiment outcomes. It also warns users not to toggle between adjusted and unadjusted results after observing which looks more favorable, an important operational detail.

The hybrid delivery surface is useful for gradual organizational adoption. A conversion team can start with browser tests, while engineering adopts server-side experiments for pricing logic, recommendations, or application flows. A common result and governance layer can reduce the split between “marketing tests” and “product experiments.”

Watchouts

Package mapping is essential. The public plans page distinguishes the prompt-based starter from enterprise add-ons, so feature experimentation and advanced statistical capabilities should be confirmed in the quote. Ask whether each method applies to client-side, server-side, and warehouse-connected experiments, and whether filtered or segmented results use the same adjustment.

Browser testing introduces familiar risks: flicker, content security policy, consent interactions, tag-manager timing, single-page application navigation, and page-performance impact. Server-side testing avoids several browser constraints but requires engineers to implement assignment and exposure correctly. Test both paths rather than assuming a “unified” platform makes them operationally identical.

Pricing and implementation notes

Kameleoon currently advertises a 30-day free PBX trial for up to three experiments and a PBX Starter package beginning at $495 per month for up to ten experiments and 50,000 tested visitors per month. Enterprise is custom. The starter price does not represent every feature-experimentation requirement, so obtain a line-item proposal for tested traffic, projects, environments, web domains, feature flags, advanced methods, raw export, support, and service levels.

For the proof of concept, run the same business hypothesis through one visual implementation and one server-side implementation. Compare assignment stability, exposure counts, SRM checks, page performance, raw export, and statistical output. This reveals whether the cross-functional promise survives real implementation boundaries.

5. LaunchDarkly: Best for release-centric experimentation

Best for

LaunchDarkly is best when experiments are extensions of software delivery. Engineers already wrapping changes in flags can attach metrics, run a test, inspect uncertainty, and roll the winning variation forward without moving the release workflow to a separate system.

It can replace Statsig for teams using flags and experiments more heavily than Statsig's analytics and replay products. It is less obviously a one-for-one replacement when the organization expects the experiment vendor to be its primary product-data platform.

Key strengths

LaunchDarkly supports A/B/n tests, funnel metrics, A/A validation, mutually exclusive experiments, holdouts, multi-armed bandits, and experiments over flags across client, server, mobile, edge, and infrastructure contexts. Its experimentation documentation describes both Bayesian and frequentist analysis, uncertainty intervals, targeting, performance and cost metrics, and warehouse-native metrics.

The release connection is the core advantage. A treatment can be guarded by the same targeting and kill-switch infrastructure used for normal delivery. Product managers can test a feature while engineering retains control over rollout, and operations teams can include latency, error, or cost outcomes. A platform already standardized on LaunchDarkly flags may avoid a second assignment SDK.

The documentation is unusually explicit about experiment design. It distinguishes assignment keys, sample-size planning, metric types, mutually exclusive layers, and holdouts. The migration guidance also recognizes Statsig as a source platform, which can reduce mapping work for flags and experiment definitions.

Watchouts

Release-centric experimentation still needs a governed metric layer. Inventory whether each result uses LaunchDarkly events, warehouse-native metrics, or Data Export, and reconcile the result with business reporting. Flag exposure telemetry is not automatically the correct analysis population; test evaluation timing and repeated calls in the application's actual code path.

Pricing has several meters. Experimentation usage can be measured separately from service connections, client-side users, observability data, session replay, and other platform capabilities. Model a year of growth and concurrent experiments rather than pricing one small test. Also confirm which statistical, governance, and warehouse functions belong to the selected plan.

Pricing and implementation notes

The current free Developer plan includes unlimited flags and seats, five service connections, 1,000 client-side monthly active users, and 100,000 experimentation MAU per month. Paid plans scale through usage-based and custom packaging. Use the LaunchDarkly pricing page for the current meters, then request a quote based on server environments, client users, experiment keys, and expected growth.

The proof of concept should instrument a production-shaped flag across at least two SDK runtimes. Verify deterministic assignment, session consistency, exposure deduplication, offline behavior, warehouse or event-metric counts, and the path from a stopped experiment to a normal rollout. Run an A/A test before trusting a feature-change result.

6. PostHog: Best all-in-one choice for a small technical team

Best for

PostHog fits small engineering-led product teams that want product analytics, session replay, feature flags, experiments, error context, and warehouse data in one interface. It is fast to evaluate, transparent about usage pricing, and broad enough to replace several Statsig workloads without an immediate procurement process.

The product is most compelling when the team is comfortable using PostHog's event and identity model as a central product-data layer. A team that already has deeply governed warehouse metrics should test data-warehouse sources and reconciliation carefully rather than recreating every metric in the event pipeline.

Key strengths

Experiments attach to feature flags and can use funnel, mean, ratio, and other saved metrics. PostHog supports Bayesian analysis by default and a frequentist option. Its public Bayesian methodology describes user-level aggregation, effect estimation, uncertainty, and quality checks instead of hiding the calculation entirely.

The surrounding context is the differentiator. A result can be explored alongside funnels, cohorts, recordings, user behavior, and errors. Engineers can move from a metric change to sessions or diagnostic events without stitching together several vendors. Flags and analytics also share identity and event infrastructure, which can simplify early implementation.

PostHog exposes APIs and maintains an open-source core, which can suit teams that prefer inspectable infrastructure and agent-driven workflows. Its generous free allowances let a startup run a realistic trial before negotiating an annual contract.

Watchouts

The all-in-one model creates coupling. Identity mistakes, ingestion loss, or event-schema drift can affect analytics and experiment decisions at once. Define canonical subject IDs, event ownership, schema checks, and metric reviews before many teams self-serve. Verify how anonymous identity merges, group analytics, late events, and feature-flag calls influence exposure.

PostHog's supported production path is cloud. Although the core repository and a Docker Compose deployment are public, its self-hosting documentation says self-hosted deployments are officially unsupported, have no paid support or guarantees, and cannot use paid-plan features. Do not choose it as a supported on-premises Statsig replacement without accepting that constraint.

Usage-based pricing also spans multiple products. An experiment can consume feature-flag requests and product-analytics events, while replay, warehouse, error tracking, and other products have their own meters. The per-unit price can look small while the combined data surface grows quickly.

Pricing and implementation notes

PostHog currently includes free monthly allowances of one million product-analytics events, one million feature-flag requests, and 5,000 session recordings. Paid product analytics starts at $0.00005 per event and feature flags at $0.0001 per request, with volume pricing. Confirm current allowances and regional options on the PostHog pricing surface before committing.

For the proof of concept, instrument one feature experiment plus one existing analytics funnel. Compare raw event volume, unique experimental subjects, exposures, conversions, identity merges, and replay availability. Then model a full month across all enabled products. This catches both statistical and commercial surprises before a broad rollout.

7. VWO: Best for conversion-focused web experimentation

Best for

VWO is strongest for conversion-rate optimization teams that need visual editing, A/B and split-URL tests, multivariate testing, targeting, reporting, and a workflow accessible to marketers. VWO Feature Experimentation extends the platform to server-side and application changes, making it a plausible Statsig alternative when web optimization is the center of gravity.

It is a better candidate for landing pages, merchandising, signup flows, and campaign experiences than for a purely backend experimentation program. Engineering-heavy buyers should evaluate its SDK, raw data, flags, and metric model as deeply as the visual editor.

Key strengths

The VWO plans page lists A/B, split-URL, and multivariate tests, targeting and reporting options, custom events, mutual exclusion, API access, and feature-experimentation capabilities across its packages. The visual workflow is designed for creating variants without waiting for a full application release, while rollout features can send a winner to more traffic.

VWO's SmartStats reporting uses Bayesian analysis. The broader platform includes heatmaps, recordings, funnels, personalization, planning, and program-management tools depending on the package. That combination can help an optimization group move from research to hypothesis, test creation, analysis, and follow-up in one suite.

The product has a long track record in web experimentation and supports client-side, server-side, and mobile use cases. For an organization replacing Statsig because marketers could not independently run website tests, that mature web surface can outweigh deeper warehouse-native methods.

Watchouts

Visual changes still execute in the browser. Measure cumulative layout shift, script loading, flicker, consent behavior, CSP compatibility, bot filtering, SPA route changes, and the failure mode when VWO is unavailable. A visually correct test is not production-safe if it changes performance or appears before consent.

Bayesian reporting does not eliminate design obligations. Predefine the primary metric, practical effect threshold, guardrails, audience, and minimum duration. Ask how SmartStats handles multiple goals, repeated observation, segments, missing data, outliers, SRM, and ratio metrics. If warehouse metrics are required, verify the exact import and analysis path in the purchased package.

VWO has several products and tier boundaries. Confirm whether Testing, Feature Experimentation, Insights, Personalize, and Plan are separate charges, and whether raw export, APIs, SSO, mutual exclusion, advanced targeting, and support require higher tiers.

Pricing and implementation notes

VWO offers an explore-for-free route, but its public pricing page directs buyers to schedule a demo for paid Growth, Pro, and Enterprise testing packages. Obtain a quote based on monthly tested visitors, domains, workspaces, concurrent tests, server-side use, seats, raw export, integrations, support, and additional VWO products.

Run a production-shaped landing-page test and a server-side application test if both are in scope. Measure page performance, assignment consistency, exposure and conversion counts, visual-editor maintainability, raw-data access, and result reproducibility. A web-focused product should win on the complete optimization workflow, not solely on how quickly a button color can be changed.

How the seven alternatives differ in practice

Decision priority	Strongest candidates	What to verify
Governed warehouse metrics	GrowthBook, Datadog Experiments	SQL transparency, metric versioning, warehouse cost, backfills, access controls
Release flags plus experiments	GrowthBook, LaunchDarkly	SDK parity, exposure timing, kill switches, rollout transition, offline behavior
Visual website optimization	Kameleoon, Optimizely, VWO	Editor reliability, page speed, consent, SPA support, raw export
Broad product-data suite	GrowthBook, PostHog, Datadog	Identity model, event quality, analytics depth, total metered cost
Advanced statistical choice	GrowthBook, Datadog Experiments, Kameleoon, Optimizely	Method applicability, documentation, stopping policy, CUPED scope, corrections
Open-source deployment	GrowthBook	Supported architecture, enterprise license boundary, operating cost
Enterprise cross-functional program	Optimizely, Kameleoon, GrowthBook	Governance, templates, holdouts, permissions, training, support

Evaluate the complete causal chain. A warehouse engine with unreliable exposures, an SDK paired with duplicated business metrics, or a visual editor that bypasses consent can all produce a failed experiment program.

How to migrate experiments from Statsig

Inventory the actual Statsig surface

List every gate, dynamic configuration, experiment, layer, holdout, metric, dashboard, product-analytics report, replay workflow, data export, warehouse connection, SDK, proxy, role, approval, and integration in use. Mark the owner and last activity. This prevents a team from migrating experiments while quietly breaking a feature gate or executive report.

Old results may need to remain searchable after SDK calls stop. Decide whether to migrate each record, export an immutable report, or retain time-bounded read-only access.

Freeze semantics before moving data

Export or document the randomization unit, assignment key, variant weights, targeting, namespace or layer, start and stop times, exposure definition, metric SQL or event filters, attribution windows, exclusions, capping, statistical method, correction policy, decision, and owner for each relevant experiment.

Create versioned metric specifications outside either vendor. For each metric, include an executable query or event definition plus fixtures that cover zero activity, repeat activity, late events, refunds, identity merges, and missing data. A migration is successful when both systems calculate the intended business quantity, not merely when their labels match.

Do not switch a live experiment's assignment algorithm

Different hashing functions, salts, namespaces, or bucketing ranges can reassign participants. That creates contamination: the same subject may see control under Statsig and treatment under the replacement. Even if the visible split remains 50/50, the groups are no longer the original randomized comparison.

Finish or stop active tests before changing assignment. If a critical long-running holdout cannot stop, preserve its original assignment service until the decision window closes, or migrate a durable assignment table and prove exact parity for every subject. Never concatenate results from two assignment regimes without a design reviewed by a qualified experimenter.

Shadow analysis before shadow assignment

Reanalyze a completed Statsig experiment in the candidate using the same assignment and outcome data. Compare subject counts, variant counts, exposure timestamps, exclusions, metric values, point estimates, intervals, corrections, SRM status, and conclusions. Differences are useful: they reveal semantic or methodological choices that need to be made explicit.

Next, send production assignment or exposure data to both analysis systems without changing what users receive. Only after analysis matches should the new SDK calculate decisions in shadow mode. Compare millions of representative contexts across languages and runtimes, including null attributes, semantic versions, Unicode, dates, prerequisites, and partial outages.

Run an A/A test and a bounded A/B test

An A/A test serves identical behavior under two labels. It should show the configured traffic split, low SRM incidence, stable metric baselines, and the expected rate of chance alerts under the selected method. Microsoft's SRM guidance notes that an imbalance can invalidate the experiment even when the result looks plausible; diagnose any mismatch before proceeding.

Then run a low-risk A/B test with a known hypothesis, reversible treatment, one primary metric, and operational guardrails. Pre-register the decision rule. Reproduce the result outside the UI. This is the final evidence that assignment, exposure, metric, statistics, and workflow function together.

A proof-of-concept scorecard

Score evidence, not sales claims. A useful proof of concept includes these gates:

Assignment parity: The same subject receives the same variation across supported runtimes, retries, sessions, and deployments.
Exposure accuracy: Only affected subjects enter the analysis, repeated evaluations do not inflate the population, and treatment cannot change inclusion improperly.
Metric reconciliation: Counts and values match an independently executed query within a documented tolerance.
Statistical reproducibility: Point estimates, uncertainty intervals, corrections, and quality warnings can be explained from documented methods.
Data-quality response: A seeded SRM, missing event stream, delayed partition, and identity fault are detected and prevent a casual ship decision.
Operational resilience: Applications behave safely when the control plane or network is unavailable, and operators can stop a harmful treatment.
Governance fit: Templates, permissions, audit history, APIs, and reports support the organization's risk tiers without blocking routine work.
Cost at scale: The model includes subjects, events, requests, seats, warehouse compute, retention, support, and every adjacent product activated by the workflow.
Exit test: Assignments, configurations, metrics, and results can be exported in useful formats without a professional-services project.

Use one dataset and rubric for every finalist. A common test reveals which system fits the organization's actual workload.

The best Statsig alternative for experimentation

GrowthBook is the best overall alternative for product and engineering teams that want rigorous experimentation without surrendering control of their data model or deployment path. It supports a managed start, warehouse-native growth, an open-source option, flags and experiments in one workflow, and a pricing model that does not charge by experimental subject or experiment count. Teams can start with GrowthBook for free and add sequential testing, CUPED, stronger governance, or enterprise support as the program matures.

Datadog Experiments deserves a close look for data-mature Datadog customers and teams that value Eppo's statistical depth plus operational telemetry. Optimizely is the broad enterprise-suite choice. Kameleoon is compelling for a combined web and feature program. LaunchDarkly is the release-centric option. PostHog is the fast all-in-one path for small technical teams. VWO remains strong for conversion-focused web experimentation.

Decide with a replayed experiment, an A/A test, and a production-shaped proof of concept. For help with data architecture, SDK rollout, metric governance, or migration risk, book a GrowthBook demo and bring the scorecard.

Build a trustworthy test program

Start free with feature flags, A/B testing, product analytics, and a path from managed data to your own warehouse.

Start with GrowthBook

Example H2

See All Articles

Experiments

A/B testing for healthcare: Examples and best practices

Sep 23, 2026

min read

In healthcare, “Can we randomize it?” is the wrong first question. Start with “Could either experience change care, rights, privacy, or access?”

A/B testing can improve digital intake, appointment access, patient education, clinician workflows, and administrative operations. It can also create unacceptable risk when teams treat a clinical or consent decision like an ordinary conversion funnel.

The difference is not the label on the method. A/B tests are randomized experiments. What matters is the treatment, purpose, affected population, data flow, and oversight required in the organization and jurisdiction. This guide provides a practical product framework, not a substitute for legal, clinical, privacy, security, or institutional review.

Draw the boundary before designing variants

Create an intake step that classifies the proposed change before anyone builds a treatment. At minimum, ask:

Can the change alter diagnosis, treatment, triage, dosage, or clinical recommendations?
Can it delay or discourage access to care, accommodations, or urgent help?
Does it change informed consent, privacy choice, required disclosure, or patient cost?
Does it use protected or sensitive health information for assignment or measurement?
Does it include children, people in crisis, or another population requiring added protection?
Is the purpose internal quality improvement, or is it designed to contribute to generalizable knowledge?
Could the software function fall within medical-device or clinical decision-support oversight?

The HHS quality-improvement guidance says many activities limited to improving patient care and collecting operational data are not research under the cited human-subjects regulations. It also states that some quality-improvement activities can have a research purpose, in which case human-subject protections may apply. A product team should not make that determination informally; route it to the organization’s authorized office.

Likewise, software that influences clinical decisions is not automatically an ordinary product surface. The FDA’s January 2026 clinical decision-support guidance explains that some software functions are excluded from the device definition while other patient- or caregiver-facing functions can remain subject to digital-health policy. Clinical and regulatory owners need to classify the function before experimentation.

Start with lower-risk operational questions

The safest early program tests reversible changes where both variants meet the same clinical, accessibility, privacy, and disclosure requirements.

Appointment reminder timing

Compare 2 approved reminder schedules or message structures to reduce missed appointments. Keep required details, opt-out behavior, language support, and urgent-contact instructions constant.

Use completed appointments or timely rescheduling as the primary outcome. Track cancellations, patient contacts, message delivery, opt-outs, wrong-recipient risk, and differences across language, age, disability, or access groups. A higher click rate is not enough if no-show rates or trust worsen.

Patient portal navigation

Test whether a clearer information architecture helps people complete a high-value administrative task, such as finding results, updating insurance, or sending a non-urgent message. Preserve emergency guidance and clinical escalation paths in both variants.

Measure successful task completion and time to completion. Guard against repeated navigation, abandonment, accessibility failures, mistaken message routing, and increased call-center burden. Use usability testing before the A/B test to catch failures randomization should never expose.

Administrative form sequence

Compare a long form with a staged flow, or test the order of non-clinical fields. Do not omit information needed for safe care, billing transparency, consent, or legal compliance.

Measure accurate completion, not just submission. Track validation errors, correction rates, staff rework, abandonment, and time to appointment. If the treatment collects sensitive data, confirm necessity and access controls before launch.

Educational content layout

Test 2 ways to present the same clinician-approved information: summary-first versus stepwise, text plus illustration versus text alone, or a clear action checklist versus a dense paragraph. Keep the medical meaning, risks, contraindications, and escalation advice equivalent.

Use a comprehension or appropriate next-action metric when feasible. Page time and clicks can be misleading. Accessibility, language quality, and comprehension across health-literacy levels belong in the guardrail plan.

Review the design before launch

Use a trustworthy experiment-design session to pressure-test metrics, safety checks, and decision rules before exposing patients or clinicians.

Watch the Experiment Design Session

Use stronger controls for care-adjacent products

Some product changes are not clinical interventions but can still influence care. They need clinical ownership, narrower eligibility, conservative ramps, and explicit stopping criteria.

Clinician workflow support

A test might compare how a work queue prioritizes administrative follow-up, how a note template reduces documentation work, or how a non-diagnostic alert is presented. The treatment should not silently alter the clinical standard of care.

Randomize at the unit that prevents contamination. Individual clinician assignment may fail when teams share queues and handoffs; clinic- or unit-level clusters may better match the workflow. Measure task completion and time saved, with guardrails for missed work, overrides, escalations, documentation quality, and staff workload.

Preventive-care outreach

Compare approved outreach content or channels for people already eligible under the same clinical rule. Do not experiment with whether one group receives necessary care or required notice.

Use completed appropriate follow-up as the primary outcome. Track opt-outs, unreachable patients, scheduling capacity, disparities, complaints, and downstream cancellations. If the treatment drives demand beyond operational capacity, a messaging lift can make access worse.

Digital adherence support

Test the presentation or timing of an approved reminder, checklist, or educational cue. Avoid treatment changes that could be interpreted as personalized medical advice without the corresponding validation and oversight.

Measure the intended behavior with caution. Self-reported completion or app engagement is not a clinical outcome. Include adverse-event reporting, escalation pathways, disengagement, and privacy events where relevant.

Feature rollout in health software

Use feature flags to separate deployment from release, start with internal or trained cohorts, and expand only when technical and clinical guardrails remain healthy. GrowthBook’s feature flag platform supports targeted rollouts and kill switches, while the experiment layer measures impact.

The rollback plan must describe more than turning off a flag. Determine whether the old experience remains clinically and operationally safe, how queued work is reconciled, what happens to partial workflows, and who is authorized to stop exposure.

Protect data by design

Do not send a broad event stream to an experimentation vendor and decide later which fields were unnecessary. Inventory the data before implementation:

Data question	Required decision
Assignment	What is the least identifiable stable unit that works?
Eligibility	Which sensitive attributes are truly needed?
Exposure	What event proves the treatment was delivered?
Outcomes	Can metrics be computed inside the governed data environment?
Access	Which roles can view assignments, segments, and results?
Retention	When are raw records, logs, and exports removed?

The HHS minimum-necessary guidance describes limiting uses, disclosures, and requests for protected health information to what is needed for the intended purpose, with policies based on roles and recurring versus non-routine access. Apply that principle to experiment attributes, debugging logs, dashboards, and downloaded readouts.

Pseudonymous identifiers reduce exposure but do not automatically make a dataset non-sensitive or outside applicable rules. Review linkability, small cohorts, free-text fields, URLs, device metadata, and combinations that can reveal a condition. Never put clinical details or identifiers in feature names, variation labels, or URLs.

A warehouse-native experimentation approach can query approved metrics where the organization already governs them. Architecture does not create compliance on its own; teams still need contracts, access control, auditability, retention rules, security review, and configuration that matches the approved data flow.

Keep unsafe questions out of product experimentation

An experimentation policy should name prohibited or separately governed categories. Product teams should not discover the boundary only after a proposal reaches launch review.

Do not use an ordinary product A/B test to withhold a clinically indicated service, emergency direction, safety warning, accessibility accommodation, required disclosure, or legally protected choice. Do not reduce the visibility of risks to improve completion. Do not randomize a diagnostic or treatment recommendation without the clinical, regulatory, and research framework appropriate to that intervention.

Avoid treatments that exploit fear, urgency, shame, or uncertainty about health. A message can increase appointment conversion while undermining informed choice. Likewise, do not test whether patients tolerate a harder cancellation, more confusing privacy control, or hidden cost. Both variants must meet the organization’s baseline standard for respectful and comprehensible communication.

Clinical AI and decision-support changes need an evaluation program beyond a click-based A/B test. Validate the model offline, examine performance and failure modes across relevant populations, review human factors, and stage deployment with clinical monitoring. An online comparison may contribute evidence only after both treatments meet the safety threshold for exposure.

When an activity may be human-subjects research, follow the institution’s process before enrolling or exposing anyone. HHS research-oversight training states that covered non-exempt human-subjects research requires the applicable review and that informed consent requirements apply unless the IRB authorizes otherwise. The product team should preserve the determination, protocol version, approved treatment, and reporting obligations with the experiment record.

Finally, do not interpret lack of detected harm as proof of safety. Rare adverse events, small vulnerable groups, and outcomes that occur after the experiment window may be underpowered. Use prior evidence, incident monitoring, qualitative reports, and post-rollout surveillance alongside the randomized estimate.

Define patient-centered metrics and guardrails

Healthcare teams need more than a conversion scorecard. Build a measurement hierarchy:

Primary outcome: the operational or patient-facing result that answers the decision.
Process diagnostics: steps that explain why the treatment worked or failed.
Safety guardrails: outcomes that trigger a stop or clinical review.
Equity checks: predeclared groups where access or benefit could differ.
Operational guardrails: staffing, wait time, rework, cost, and downstream capacity.

Define the practical threshold before launch. A statistically detectable change may be too small to justify implementation, and a neutral aggregate can hide meaningful harm in a protected or vulnerable group. At the same time, slicing results across many small subgroups increases false-positive risk and can expose sensitive attributes. Predeclare the equity questions that matter and use appropriate privacy and multiple-testing controls.

GrowthBook supports reusable fact tables and metrics so teams can keep definitions reviewable. Use a power analysis for the primary outcome and critical guardrails. If the required sample or duration is unrealistic, do not weaken the standard; use usability research, simulation, staged quality improvement, or a larger treatment contrast.

Create a healthcare experiment review packet

Before launch, the owner should provide one reviewable packet:

purpose, hypothesis, and operational decision
classification and required oversight determination
affected population and exclusion criteria
clinical, privacy, security, accessibility, and compliance approvals
treatment screenshots or workflow diagrams
assignment, exposure, and data-flow design
primary outcome, diagnostics, guardrails, and equity checks
sample plan and stopping rule
rollout stages, monitoring owner, and rollback procedure
patient or clinician communication plan, if applicable
documentation and retention plan

Use an approval matrix that names accountable people. Product approval does not replace clinical approval; a privacy review does not settle human-subjects research status; and an IRB determination does not automatically approve the production security architecture.

The WHO clinical-trial best-practices guidance emphasizes ethical standards, regulatory considerations, patient-centered research, transparency, and stakeholder collaboration. Not every healthcare product experiment is a clinical trial, but high-risk work should inherit the same respect for people and evidence.

Build trust into the experimentation program

Start with reversible operational improvements where both experiences are already acceptable. Prove that the team can classify risk, minimize data, validate assignment, monitor safety, and document decisions before expanding scope.

Publish internal rules for what teams may test, what requires added review, and what is out of bounds. Maintain an experiment registry and audit trail. Record neutral and negative results so a new team does not repeat the same risky idea.

GrowthBook can support the controlled delivery and analysis layer through experimentation, feature flags, permissions, and warehouse-defined metrics. The organization remains responsible for the clinical, ethical, legal, privacy, and operational framework around every test.

In healthcare, speed is valuable only when the learning process protects the people whose behavior creates the data.

Build a governed test workflow

Connect controlled releases to reviewable metrics and decision rules while keeping healthcare data in your approved architecture.

Get Started With GrowthBook

Experiments

When to use a z-test vs t-test vs chi-square vs ANOVA

Sep 22, 2026

min read

The right statistical test is determined by the question and data-generating process, not by which function is easiest to run. Start with the outcome, groups, and dependence structure; the test name comes later.

Z-tests, t-tests, chi-square tests, and analysis of variance (ANOVA) all compare observed data with a null model. They differ in the kind of outcome they model, the uncertainty they estimate, and the number or structure of groups they can compare.

For a simple product experiment, a useful first pass is:

continuous outcome, two independent groups: usually a Welch two-sample t-test
binary proportion, two large independent groups: a two-proportion z-test is common
categorical counts across groups: chi-square test, if expected counts are adequate
continuous outcome across three or more groups: one-way ANOVA or Welch ANOVA

Those rules are a starting point. Paired observations, clusters, ratios, repeated measures, heavy tails, covariate adjustment, or sequential monitoring require a model that reflects the design.

Choose from the outcome and hypothesis

Write the estimand before choosing a test. An estimand is the quantity the experiment is trying to estimate: a difference in mean revenue, a difference in conversion probability, or an association between two categorical variables.

Question	Outcome	Common test
Did average order value change between A and B?	Continuous	Welch two-sample t-test
Did signup probability change between A and B?	Binary	Two-proportion z-test
Is plan choice associated with variant?	Categorical, 3+ levels	Chi-square test of independence
Do mean task times differ across four variants?	Continuous	One-way ANOVA
Did the same users' scores change before and after?	Paired continuous	Paired t-test

The number of groups alone is insufficient. Conversion in four variants is still categorical data; a chi-square or binomial model may fit. Revenue in two groups is continuous; a t-test or regression is more natural.

The University of Michigan's statistical-test guide uses the same sequence: identify variable types and the relationship being tested before selecting a method.

When to use a z-test

A z-test compares a standardized estimate with the standard normal distribution. The classical one-sample z-test for a mean assumes the population standard deviation is known. That condition is unusual in product analytics, where variability is estimated from the current sample.

Z-tests remain common for proportions. In a two-arm conversion experiment, the estimate is:

difference = p_treatment - p_control

Under the null of equal proportions and with adequate counts, the standardized difference is approximately normal. This yields a two-proportion z-test.

Use it when:

the outcome is a binary count summarized as successes and failures
assignment groups are independent
sample sizes make the normal approximation credible
the hypothesis and one- or two-sided direction were set before analysis

Do not rely on a universal “n greater than 30” rule. For rare events, 30 observations can produce almost no successes; for balanced common events, approximation quality can be good. Inspect expected successes and failures and use an exact or model-based method when counts are sparse.

In high-volume online experiments, a normal approximation is also used for many sample means through the central limit theorem. The important question is whether the estimator's sampling distribution and variance calculation are valid for the metric, not whether the raw user values look perfectly normal.

When to use a t-test

A t-test is designed for inference about means when the variance is estimated from sample data. That extra variance uncertainty produces a t distribution with heavier tails than the standard normal, especially at small sample sizes.

For two independent groups, default to Welch's t-test unless equal variance is justified. Welch's version does not assume the two population variances are equal and handles unequal group sizes. NIST's two-sample t-test reference shows the unequal-variance standard error based on each group's sample variance and size.

Use an independent two-sample t-test when:

the outcome is numeric and the mean is the target
the two groups contain different experimental units
observations are independent within the model
the mean and standard error behave well enough for the sample size

Use a paired t-test when each value has a meaningful partner: the same user's before-and-after score, or deliberately matched units. The analysis reduces each pair to a difference and tests the mean of those differences. Treating paired data as independent discards information and computes the wrong standard error.

The t-test can be sensitive to extreme values because the sample mean and variance are sensitive to them. Product metrics such as revenue or session duration are often skewed. At scale, the mean may still have a usable sampling distribution, but inspect outliers, data quality, and the estimand. Robust inference, transformations, winsorization policies, or bootstrap methods may be more appropriate when a few observations dominate the result.

Reduce variance before launch

Learn how CUPED and covariate adjustment can sharpen experiment estimates without changing the randomized comparison.

Explore Variance Reduction

When to use a chi-square test

Pearson's chi-square statistic compares observed category counts with counts expected under a null hypothesis. Two common forms are:

goodness of fit: does one categorical distribution match specified probabilities?
independence or homogeneity: is a categorical outcome distributed the same way across groups?

Suppose an onboarding experiment records three outcomes: completed, skipped, and abandoned. Cross-tabulate outcome by variant. A chi-square test asks whether the outcome distribution is independent of variant.

              Completed  Skipped  Abandoned
Control             420      110         70
Treatment           455       82         63

The test statistic sums (observed - expected)^2 / expected across cells. NIST's chi-square documentation describes the same comparison of binned frequency distributions.

Use a chi-square test when observations contribute counts to mutually exclusive categories and expected cell counts are large enough for the asymptotic approximation. With sparse cells, combine categories only when substantively justified or use an exact method such as Fisher's exact test for a two-by-two table.

A chi-square result says the distributions differ somewhere. It does not provide the most decision-friendly effect estimate by itself. Report category proportions, absolute differences, uncertainty intervals, and the cells contributing to the pattern.

For a binary two-arm experiment, the Pearson chi-square test and a two-sided two-proportion z-test are closely related: under standard conditions, the chi-square statistic with one degree of freedom equals the squared z statistic. Choose the representation that matches the hypothesis and reporting needs.

When to use ANOVA

ANOVA compares variation between group means with unexplained variation within groups. A one-way ANOVA tests the null that all population means are equal across levels of one factor.

Use it for a continuous outcome across three or more independent groups when the global question is whether any mean differs. Classical ANOVA assumes independent errors, normally distributed residuals within the model, and equal variances. Welch ANOVA relaxes the equal-variance assumption; R's 0 implements that approximation.

ANOVA's F-test is an omnibus test. A significant result means at least one mean differs, but it does not identify which one. Use planned contrasts or multiplicity-aware post-hoc comparisons to answer the product question.

ANOVA is more than a rule for “three or more groups.” Multi-factor ANOVA can estimate main effects and interactions in multivariate or factorial experiments. Repeated-measures or clustered data need corresponding error structures rather than a basic one-way calculation.

Why several t-tests are not a substitute for ANOVA

With four variants there are six pairwise comparisons. Testing each at 0.05 creates multiple opportunities for a false positive. An omnibus ANOVA tests one global null first, and planned follow-ups can use Tukey, Holm, Bonferroni, or another procedure appropriate to the family of claims.

The Bonferroni correction is simple and conservative. The right procedure depends on whether the goal is all pairwise comparisons, treatments versus one control, or a small set of preplanned contrasts. Define that family before looking at the ranking.

ANOVA and regression are also two views of the same linear-model machinery. R's 0 documentation describes aov as a wrapper around linear models for experimental designs. Regression is often more flexible when the analysis includes covariates, interactions, or unbalanced data.

Assumptions that change the choice

Before running any of the four tests, verify:

Independence and assignment unit

If the experiment randomizes accounts but analyzes users as independent observations, standard errors will usually be too small. Analyze at the randomization unit or use cluster-aware inference. If users can appear in both groups, repair the assignment or use a model that represents the dependence.

Paired or repeated observations

The same user measured twice is not two independent users. Use a paired test or repeated-measures model. For experiments with many events per user, aggregate to the user level or use appropriate clustered methods.

Outcome distribution and metric construction

Check missingness, zero inflation, extreme tails, ratio denominators, and censoring. A test can be mathematically correct for the supplied numbers while the metric itself misrepresents the user outcome.

Variance assumptions

Prefer Welch's t-test or Welch ANOVA when group variances may differ. Equal sample sizes do not prove equal variance, and a preliminary variance test can introduce another decision layer.

Sample size and sparse cells

Approximate z and chi-square methods need enough information in the relevant cells. Low-frequency guardrails and small segments may need exact methods or longer collection.

A product experimentation decision tree

Use this sequence before opening a statistics package:

What unit was randomized: user, account, device, session, or region?
What is the primary estimand: mean, proportion, category distribution, or model coefficient?
Are groups independent, paired, repeated, or clustered?
Are there two groups, several groups, or multiple factors?
Do expected counts and sample sizes support the approximation?
Are variances, tails, or outliers likely to break the default model?
How many confirmatory hypotheses can trigger the decision?
Was the test direction and stopping rule declared before launch?

Then choose the simplest model that answers the exact question. A two-proportion z-test may be perfect for signup conversion, while a t-test handles mean revenue and a chi-square test handles plan mix in the same experiment. Different metrics can require different tests.

Report effects, not only test names

The test produces a statistic and p-value under a null model. The guide to interpreting a t-test p-value shows why that number needs the effect, interval, and degrees of freedom beside it. The product decision needs more:

the effect estimate in business units
a confidence or credible interval
sample sizes and allocation
baseline and treatment values
assumption and data-quality checks
the planned hypothesis family
practical thresholds and guardrails

GrowthBook's statistics documentation explains the frequentist and Bayesian engines available for experiment analysis. Whichever framework is used, review effect magnitude and uncertainty together. A small p-value can accompany a trivial lift in a huge sample, while a valuable estimated lift can remain uncertain in a small one.

Choose the test by tracing the data back to the experiment design. For three or more continuous-outcome variants, the deeper ANOVA guide covers the omnibus F-test, planned contrasts, and Welch alternative. When the outcome, assignment unit, dependence, and hypothesis are explicit, the difference between z, t, chi-square, and ANOVA becomes a modeling decision rather than a memorization exercise.

Analyze tests with context

Connect experiment assignments to trusted metrics, inspect uncertainty, and keep decision rules visible to the whole team.

Get Started With GrowthBook

Experiments

What is ANOVA? Comparing multiple test variants

Sep 21, 2026

min read

An experiment with control plus three variants creates more than one comparison. ANOVA gives the team one principled global test of whether the variants differ before it starts hunting for a winner.

Analysis of variance, or ANOVA, is a family of statistical models for comparing group means and decomposing sources of variation. In a one-way product experiment, the “factor” is the assigned variant and its “levels” are control, B, C, and D.

The basic ANOVA question is deliberately broad: if all variants had the same population mean, would the observed separation among their sample means be surprising relative to the noise within variants?

That question is useful, but incomplete. A significant ANOVA result does not say which variant won, whether the lift is large enough to ship, or whether assumptions and instrumentation are sound. Those conclusions require planned contrasts, uncertainty intervals, and experiment-quality checks.

How ANOVA compares means through variance

ANOVA separates total variability into components:

between-group variation: how far each group mean is from the overall mean
within-group variation: how far individual observations are from their group mean

Each sum of squares is divided by its degrees of freedom to produce a mean square. The F statistic is:

F = mean square between groups / mean square within groups

Under the null hypothesis that all group means are equal, both quantities estimate the same underlying error variance, so their ratio should often be near 1. When group means are separated relative to the residual noise, F grows.

NIST's one-way ANOVA explanation describes this as comparing the level mean square with the residual mean square. The p-value is the probability, under the null model and assumptions, of an F statistic at least as large as the observed one.

For k groups and N total observations, one-way ANOVA usually has:

between-group degrees of freedom = k - 1
within-group degrees of freedom = N - k

The numerator asks how much the k means vary. The denominator pools information about variability inside the groups.

A four-variant experiment example

Suppose a SaaS team tests four onboarding flows and measures projects created per eligible account during the first week.

Variant	Accounts	Mean projects	Standard deviation
Control	1,000	2.30	1.80
B	1,020	2.42	1.84
C	990	2.61	1.91
D	1,010	2.36	1.79

The null hypothesis is:

mean_control = mean_B = mean_C = mean_D

The alternative is that not all four means are equal. Notice what it does not say: “C is best.” The global alternative includes any pattern where at least one mean differs.

If the F-test rejects the null, the team should evaluate the comparisons it planned. It might compare every treatment with control, or test one contrast between the current flow and the average of three new concepts. The comparison plan should reflect the decision, not the visual ranking in the finished dashboard.

Make multiple tests trustworthy

See how experimentation leaders plan hypotheses, guardrails, and review practices when a result surface contains many possible claims.

Watch the Trustworthy Experiments Talk

Why not run every pairwise t-test?

Four groups create six pairs. If the team runs six independent tests at alpha 0.05 and treats any significant result as proof, the probability of at least one false positive across the family can exceed 0.05.

ANOVA gives one global test of the equality of all means. It also estimates residual variation using all groups, which can be more efficient than estimating it afresh for each pair under the classical equal-variance model.

The global test does not eliminate multiplicity in follow-up comparisons. R's Tukey HSD documentation explicitly notes that ordinary t-tests inflate the probability of a false declaration across a family. Choose the follow-up procedure for the comparisons the decision actually needs:

every pair: Tukey-style simultaneous comparisons
every treatment versus control: Dunnett-style comparisons
a few planned product questions: predeclared contrasts with a suitable adjustment
a conservative small family: a Bonferroni or Holm correction

An omnibus test can also be nonsignificant while one carefully planned contrast is persuasive, because the hypotheses and power differ. Decide before launch whether the global null or a treatment-versus-control contrast is the primary decision test.

Unequal group sizes do not automatically invalidate ANOVA, but they make the variance assumption and contrast plan more consequential. If allocation is intentionally uneven, power the smallest comparison that drives the decision and preserve the assignment probabilities. When variances and sample sizes both differ, classical pooled ANOVA can behave poorly; Welch ANOVA or a regression with suitable standard errors is usually easier to defend.

Planned contrasts can also use product structure that the global test ignores. Instead of comparing every pair, a team might compare control with the average of three related treatments, or compare two low-intensity treatments with two high-intensity treatments. A small set of predeclared contrasts often answers the business question with more power and clearer multiplicity control than an exhaustive winner search.

ANOVA assumptions in experiments

The familiar one-way fixed-effects model can be written as:

outcome = overall mean + variant effect + residual error

Classical inference depends on the residuals and design, not on a requirement that the combined raw outcome form one bell curve. NIST's model reference assumes independent, normally distributed errors with mean zero and common variance.

Independent observations

The analysis unit must respect randomization. If accounts are assigned but every user within an account is treated as independent, the standard error ignores clustering. Aggregate at the account level or use cluster-robust or hierarchical methods.

Repeated events from one user create the same problem. Ten sessions from one user do not carry the same independent information as ten users.

Appropriate residual behavior

ANOVA is often robust to moderate non-normality with balanced, sufficiently large groups, but severe skew, outliers, censoring, or zero inflation can make the mean unstable or the F approximation unreliable. Diagnose residuals and assess whether the mean is still the business estimand.

Equal variance for classical one-way ANOVA

Classical ANOVA assumes a common population variance. This can fail when a treatment changes both the mean and spread, or when groups serve different traffic mixes. Unequal group sizes make the problem more consequential.

SciPy's 0 supports Welch ANOVA when equal_var=False. Welch's method relaxes equal population variances and adjusts the degrees of freedom.

Correct outcome model

ANOVA targets a continuous mean. Conversion is binary; event counts are discrete; time-to-churn can be censored. Large-sample mean inference can sometimes work, but logistic, Poisson or negative-binomial, survival, or other generalized models may better represent the outcome and produce interpretable effects.

One-way, two-way, and repeated-measures ANOVA

“ANOVA” names a family rather than one calculation.

One-way ANOVA

One categorical factor with multiple levels, such as four assigned onboarding variants. This is the usual A/B/n example.

Two-way or factorial ANOVA

Two controlled factors, such as headline and layout. The model estimates each main effect plus their interaction. The interaction asks whether one factor's effect changes with the other. This is central to a properly designed multivariate test.

Repeated-measures ANOVA

The same units are observed under multiple conditions or times. Dependence is part of the design and must be modeled. A basic independent one-way ANOVA is invalid for repeated measurements.

ANCOVA

Analysis of covariance adds continuous covariates to the group comparison. In randomized experiments, pre-experiment covariates can improve precision when they are chosen and measured without post-treatment contamination. GrowthBook's guide to variance reduction explains the same motivation in online experimentation.

Run one-way ANOVA in Python

At the action boundary, keep one numeric observation per independent analysis unit in each group. In SciPy:

from scipy.stats import f_oneway

control = [2, 1, 4, 3, 2, 2, 5]
variant_b = [3, 2, 4, 4, 3, 2, 5]
variant_c = [4, 3, 5, 4, 4, 3, 6]

# Classical one-way ANOVA: assumes equal population variances.
result = f_oneway(control, variant_b, variant_c, equal_var=True)
print(result.statistic, result.pvalue)

# Welch ANOVA: does not assume equal population variances.
welch = f_oneway(control, variant_b, variant_c, equal_var=False)
print(welch.statistic, welch.pvalue)

Before running it, confirm that rows match the randomization unit and missing values have a documented policy. Afterward, inspect group summaries and residual behavior. The p-value alone cannot reveal a broken exposure join or a few enormous outliers.

In R, aov(outcome ~ variant, data = experiment) fits the classical model. R documents 1 as a linear-model interface, which helps explain why ANOVA, regression, and contrasts are closely connected.

Interpret the ANOVA table

A standard output contains:

degrees of freedom
sum of squares
mean square
F statistic
p-value

Suppose the output reports F(3, 4016) = 6.8, p < 0.001. Under the model, the observed ratio of between-variant to within-variant variation is unlikely if all four population means are equal. It does not mean every treatment beats control or that any effect is commercially important.

Add the quantities the product decision needs:

each mean and sample size
differences from control in original units
simultaneous or comparison-specific intervals
an effect-size measure when useful
guardrail and data-quality results
the follow-up comparison method

Avoid ranking noisy means without uncertainty. The highest observed variant has benefited from both its true effect and sampling variation, especially when many variants were screened.

Common ANOVA mistakes

Treating events as independent users

Repeated events make the nominal sample size huge and uncertainty too narrow. Preserve the assignment unit.

Using ANOVA for every metric shape

The word “variant” does not imply ANOVA. Match the outcome distribution and estimand to a model.

Checking assumptions after selecting a winner

Write the model, outlier policy, transformation, and variance choice before the ranking is visible. Result-driven switching creates hidden researcher degrees of freedom.

Treating a significant F-test as a winner declaration

Follow with the planned contrasts. The omnibus test only rejects equality of all means.

Ignoring practical significance

A very large experiment can detect a tiny difference. Compare intervals with a minimum practical effect and account for implementation cost and guardrails.

Use ANOVA as part of an experiment plan

Before launch, specify the factor and levels, independent unit, primary continuous outcome, minimum effect, sample-size plan, variance assumption, global or contrast hypothesis, comparison family, and stopping rule.

Then verify assignment and exposure before interpreting the model. A sample ratio mismatch can signal that observed group counts no longer reflect the planned randomization. No F-test can repair biased exposure data.

ANOVA is valuable because it turns a field of variant means into a structured model of signal and noise. The broader z-test, t-test, chi-square, and ANOVA guide shows when the outcome and hypothesis call for another member of that family. Use the omnibus test for the global question, planned contrasts for the decision, and effect estimates for practical judgment. That sequence makes a multiple-variant test easier to defend than a dashboard full of uncoordinated p-values.

Compare variants with discipline

Run controlled experiments, connect trusted metrics, and review treatment effects and uncertainty in one shared workflow.

Start With GrowthBook

Ready to ship faster?

No credit card required. Start with feature flags, experimentation, and product analytics—free.

Get Started

Book a Demo

Simplified white illustration of a right angle ruler or carpenter's square tool.

White checkmark symbol with a scattered pixelated effect around its edges on a transparent background.

A Statsig alternative must preserve the chain from assignment to decision. Perfect assignment still fails when exposures, units, metrics, or stopping rules are wrong.

Statsig experimentation alternatives at a glance

What to compare before choosing a Statsig replacement

Randomization unit, assignment, and exposure

Metric source and semantic governance

Statistical method and stopping policy

Variance reduction, power, and data quality

Program operations and governance

1. GrowthBook: Best overall Statsig alternative

Best for

Key strengths

Watchouts

Pricing and implementation notes

2. Datadog Experiments: Best for warehouse metrics plus observability

Best for

Key strengths

Watchouts

Pricing and implementation notes

Compare GrowthBook and Statsig

3. Optimizely: Best for a broad enterprise experimentation program

Best for

Key strengths

Watchouts

Pricing and implementation notes

4. Kameleoon: Best for hybrid web and feature experimentation

Best for

Key strengths

Watchouts

Pricing and implementation notes

5. LaunchDarkly: Best for release-centric experimentation

Best for

Key strengths

Watchouts

Pricing and implementation notes

6. PostHog: Best all-in-one choice for a small technical team

Best for

Key strengths

Watchouts

Pricing and implementation notes

7. VWO: Best for conversion-focused web experimentation

Best for

Key strengths

Watchouts

Pricing and implementation notes

How the seven alternatives differ in practice

How to migrate experiments from Statsig

Inventory the actual Statsig surface

Freeze semantics before moving data

Do not switch a live experiment's assignment algorithm

Shadow analysis before shadow assignment

Run an A/A test and a bounded A/B test

A proof-of-concept scorecard

The best Statsig alternative for experimentation

Build a trustworthy test program

Table of Contents

Related Articles

A/B testing for healthcare: Examples and best practices

In healthcare, “Can we randomize it?” is the wrong first question. Start with “Could either experience change care, rights, privacy, or access?”

Draw the boundary before designing variants

Start with lower-risk operational questions

Appointment reminder timing

Patient portal navigation

Administrative form sequence

Educational content layout

Review the design before launch

Use stronger controls for care-adjacent products

Clinician workflow support

Preventive-care outreach

Digital adherence support

Feature rollout in health software

Protect data by design

Keep unsafe questions out of product experimentation

Define patient-centered metrics and guardrails

Create a healthcare experiment review packet

Build trust into the experimentation program

Build a governed test workflow

When to use a z-test vs t-test vs chi-square vs ANOVA

The right statistical test is determined by the question and data-generating process, not by which function is easiest to run. Start with the outcome, groups, and dependence structure; the test name comes later.

Choose from the outcome and hypothesis

When to use a z-test