AB tests
Design, monitor, and analyze randomized controlled trials
USE CASES
Built for every modern causal challenge
Quasi-Experiments and Observational Study
Estimate effects when there is no random assignment, using causal inference techniques
Robustness and Accuracy
Check robustness and accuracy to make your decision trustworthy
You don’t choose methods — you choose a scenario. Causalis selects a best-practice default.
Each scenario maps to a modern estimator with defensible inference, implemented through a consistent fit() → estimate() API.
Backed by the literature
Causal ML
Chernozhukov et al. (2018) — DML/IRM
Bach et al. (2024) — DoubleML
Experiments
Deng et al. (2013) — CUPED
Lin (2013) — Interacted ANCOVA
Robust inference
Welch (1947) — Welch t-test
Efron (1979) — Bootstrap
Guardrails included: SRM checks (Fabijan et al., 2019) and sensitivity analysis (Chernozhukov et al., 2024).
Engineering
Production-ready by design
Causalis is built like a real analytics system: typed data contracts, test coverage, deterministic estimators, and performance defaults that scale.
Typed data contracts (Pydantic)
Validate schema, types, nullability, and column roles before any estimator runs.
CausalData enforces outcome/treatment/confounders
Numpy-style docstrings
Consistent API docs with assumptions, inputs, outputs, and failure modes.
Docstring-first public API
Tested estimators
Unit + integration tests over synthetic DGPs to prevent silent regressions.
Deterministic seeds + known ground truth
Fast inference paths
Approximation options when bootstrap is too expensive at scale.
Analytic / IF-based SEs where appropriate
Battle-tested ML defaults
Strong out-of-the-box nuisance models for modern causal ML workflows.
CatBoost default for DML/IRM
Fail-fast diagnostics
Guardrails that catch common experiment and data issues early.
SRM, balance checks, sensitivity hooks
Typed contracts. Tested estimators. Fast uncertainty. Strong defaults.
Benchmarks
DGPs you can trust
Every scenario ships with synthetic generators that encode the causal effect—so you can validate estimators against ground truth, test uncertainty, and benchmark robustness.
Level 0
Sanity checks
Minimal DGP to verify pipelines and baseline estimators end-to-end.
No confounding • constant effect
Level 1
Experimental realism
Prognostic signal + optional y_pre to benchmark CUPED / Lin adjustment.
add_pre • pre_corr • prognostic_scale
Level 2
Observational
Confounding through X only: DML/IRM should recover the effect with valid inference.
beta_d ≠ 0 • beta_y ≠ 0 • u_strength_* = 0
Level 3
Modern complexity
Copula-correlated X, nonlinearities, heterogeneity, multiple outcome families.
use_copula • g_y/g_d • tau(X) • binary/poisson
Level 4
Stress & failure
Assumptions violated on purpose to show where identification breaks.
u_strength_d ≠ 0 AND u_strength_y ≠ 0 → guarded
Oracle Columns
m, g0, g1, cate
Outcome Families
continuous / binary / poisson
Positivity Control
propensity_sharpness
Run your first notebook
Result
Ground truth ATE is 0.9509353818962034
Result
| y | d | tenure_months | avg_sessions_week | spend_last_month | premium_user | urban_resident | |
|---|---|---|---|---|---|---|---|
| 0 | -1.983895 | 1.0 | 28.814654 | 1.0 | 84.100761 | 1.0 | 0.0 |
| 1 | 7.527126 | 0.0 | 7.444181 | 0.0 | 30.890847 | 0.0 | 1.0 |
| 2 | 6.696842 | 1.0 | 23.759279 | 2.0 | 93.693180 | 0.0 | 0.0 |
| 3 | 10.337161 | 0.0 | 24.969929 | 9.0 | 127.974978 | 0.0 | 1.0 |
| 4 | 6.071955 | 0.0 | 29.943261 | 2.0 | 96.998539 | 0.0 | 1.0 |
Result
| value | |
|---|---|
| field | |
| estimand | ATE |
| model | IRM |
| value | 0.9788 (ci_abs: 0.8074, 1.1503) |
| value_relative | 40.3375 (ci_rel: 33.2610, 47.4139) |
| alpha | 0.0500 |
| p_value | 0.0000 |
| is_significant | True |
| n_treated | 3454 |
| n_control | 6546 |
| treatment_mean | 3.6747 |
| control_mean | 2.3230 |
| time | 2026-02-08 |