Cross-framework ensembling of causal machine-learning estimators for ATE and pointwise CATE, with full-pipeline bootstrap inference.
MetaCausal orchestrates multiple causal-ML estimators from different libraries — EconML, DoubleML, CausalML, stochtree, or arbitrary user-supplied callables — behind a single protocol, and aggregates their treatment-effect estimates into a single ensemble estimate. Eight aggregation strategies are provided, grouped into three tiers:
- Pointwise — Median (default), Mean, Trimmed Mean.
- Agreement-based — Consensus Based Averaging, which selects a high-agreement subset of components from pairwise Kendall's τ.
- Outcome-supervised — Causal Stacking, R-Stacking, Q-Aggregation, and Select, which learn weights (or, for Select, choose a single best component) by optimising a causal loss on cross-fitted out-of-fold predictions. These follow published aggregation rules rather than inventing new ones — see the "Reference" column in the table below.
A full-pipeline bootstrap refits every component, nuisance model, and aggregation weight on each replicate, so the resulting confidence interval reflects uncertainty from all of these sources rather than just the final point estimate — shortcut bootstraps that freeze weights or nuisance fits across replicates understate variability. A side benefit: because the same procedure applies uniformly regardless of aggregation strategy or component mix, it also yields directly comparable CIs for both ATE and pointwise CATE across heterogeneous components whose native inference machinery is otherwise incomparable.
No single causal-ML estimator dominates across data-generating processes (Curth & van der Schaar, 2023), and model selection for heterogeneous treatment effects is empirically unreliable: because the true CATE is never observed, ordinary cross-validation does not apply, and Mahajan et al. (2024) benchmark 415 CATE estimators across 78 datasets to find that reliable selection requires careful nuisance-model tuning — and that ensembling can outperform selecting any single estimator outright. Individual methods can also fail catastrophically under specific violations of their own assumptions (overlap breakdown, nuisance misspecification, tree extrapolation). MetaCausal's default pointwise median aggregation gives a 50% breakdown point with no tuning — up to half the component estimators can produce arbitrarily bad estimates without corrupting the ensemble. When outcome data allow learning weights, MetaCausal also ships the three outcome-supervised stackers from the recent CATE-ensemble literature, plus Select as a feasible-selection baseline for comparison.
pip install metacausalThis installs the core package and its required dependencies (numpy, pandas, scipy). Estimator libraries are optional extras:
# Individual libraries
pip install "metacausal[econml]"
pip install "metacausal[doubleml]"
pip install "metacausal[causalml]"
pip install "metacausal[stochtree]"
# Visualisation helpers (matplotlib)
pip install "metacausal[plots]"
# Everything (frameworks + plots)
pip install "metacausal[all]"Python 3.11 or later is required.
from metacausal import CausalEnsemble
from metacausal.datasets import load_lalonde
X, T, Y = load_lalonde()
# Default ensemble: nine estimators spanning EconML, DoubleML,
# CausalML, and stochtree, aggregated by pointwise median.
ens = CausalEnsemble()
ens.fit(X, T, Y, random_state=42)
# Point estimate. summary() prints the ensemble ATE, the per-component
# table, and the spread; the raw values stay on the object (ate.ate, ...).
ate = ens.ate()
print(ate.summary())
# Full-pipeline bootstrap confidence interval. n_boot=10 keeps this
# snippet quick to run; bump to the n_boot=200 default (or higher) for
# publication-quality intervals.
boot = ens.bootstrap(n_boot=10, random_state=42, n_jobs=-1)
print(boot.summary())The three-step fit → ate / cate → bootstrap pattern is the recommended one, because it lets you inspect intermediate state and swap aggregation strategies on an already-fitted ensemble. The convenience wrapper ens.estimate(X, T, Y, n_boot=10, ...) does fit + bootstrap (or fit + ate) in a single call.
| Tier | Strategy | String alias / class | Data used | Reference |
|---|---|---|---|---|
| Pointwise | Median (default) | "median" / Median |
Component predictions only | — |
| Pointwise | Mean | "mean" / Mean |
Component predictions only | — |
| Pointwise | Trimmed Mean | "trimmed_mean" / TrimmedMean |
Component predictions only | — |
| Agreement | Consensus Based Averaging | "cba" / CBA |
Component CATE predictions on training data | Machluf et al., 2026 |
| Supervised | Causal Stacking | CausalStacking |
Cross-fitted OOF predictions + nuisance | Han & Wu, 2022 |
| Supervised | R-Stacking | RStacking |
Cross-fitted OOF predictions + nuisance | Nie & Wager, 2021 |
| Supervised | Q-Aggregation | QAggregation |
Cross-fitted OOF predictions + nuisance | Lan & Syrgkanis, 2024 |
| Supervised | Select (best-by-risk) | Select |
Cross-fitted OOF predictions + nuisance | — (MetaCausal baseline) |
# By string alias (default configuration)
ens = CausalEnsemble(aggregation="trimmed_mean")
# By object (lets you configure hyperparameters)
from metacausal.aggregation import QAggregation
ens = CausalEnsemble(aggregation=QAggregation(nu=0.5, greedy=True))See the accompanying paper (forthcoming) for the mathematical details of each strategy.
MetaCausal supports two outcome types:
- Continuous (default) — any numeric Y not detected as binary. Quietly absorbs counts, bounded continuous, and ordinal-as-numeric. The base learner choice is the user's responsibility (an
HistGradientBoostingRegressorby default; user-supplied components can pass a Poisson booster if appropriate). - Binary — numeric Y with values ⊆ {0, 1} or boolean dtype. The estimand is the risk difference ATE/CATE (mean difference of probabilities).
Detection happens at fit time: CausalEnsemble().fit(X, T, Y) inspects Y, picks the right pool from default_methods, and routes nuisance estimation through predict_proba for binary outcomes. To force an interpretation, pass outcome_type="continuous" or outcome_type="binary" at construction. Multi-class / nominal and survival outcomes are out of scope; encoding-as-multiple-binary or a dedicated survival library is the recommended path.
The default binary pool (8 components) drops DoubleMLPLR, EconML S/T/X-Learners, the BaseRRegressor, and stochtree BCF — all of which either lack a binary-capable code path in their upstream library or would silently fit a linear-probability model — and substitutes the CausalML classifier siblings (BaseSClassifier, BaseTClassifier, BaseXClassifier, BaseRClassifier). DoubleMLIRM, CausalForestDML, DRLearner, and TMLELearner remain.
Snippets below that use X, T, Y without redefining them assume the
load_lalonde() call from Quick start; snippets that need
different data (e.g. a binary outcome) load it explicitly.
Estimators from EconML and CausalML are auto-detected by module prefix; DoubleML, stochtree, and arbitrary callables go through explicit adapters.
from metacausal import CausalEnsemble, GenericATEAdapter
from metacausal.adapters import DoubleMLAdapter, CausalMLAdapter
from econml.dml import CausalForestDML
from econml.metalearners import TLearner, XLearner
from doubleml import DoubleMLIRM
from causalml.inference.meta import BaseDRRegressor
from sklearn.ensemble import (
HistGradientBoostingRegressor as HGBR,
HistGradientBoostingClassifier as HGBC,
)
def naive_diff(X, T, Y):
return float(Y[T == 1].mean() - Y[T == 0].mean())
ens = CausalEnsemble(
methods=[
CausalForestDML(discrete_treatment=True), # auto-wrapped (EconML)
TLearner(models=HGBR()), # auto-wrapped (EconML)
XLearner(models=HGBR(), propensity_model=HGBC()), # auto-wrapped (EconML)
DoubleMLAdapter(DoubleMLIRM, ml_g=HGBR(), ml_m=HGBC()),
CausalMLAdapter(BaseDRRegressor(learner=HGBR())),
GenericATEAdapter(naive_diff, name="naive_diff"),
],
aggregation="median",
)
ens.fit(X, T, Y, random_state=42)
print(ens.ate().ate)To configure analytical upstream inference, wrap the estimator explicitly instead of relying on auto-detection:
from causalml.inference.meta import BaseTRegressor
from doubleml import DoubleMLIRM
from econml.dml import CausalForestDML
from metacausal.adapters import DoubleMLAdapter, EconMLAdapter, CausalMLAdapter, StochtreeAdapter
from sklearn.ensemble import (
HistGradientBoostingRegressor as HGBR,
HistGradientBoostingClassifier as HGBC,
)
dml = DoubleMLAdapter(DoubleMLIRM, ml_g=HGBR(), ml_m=HGBC(), alpha=0.10)
econ = EconMLAdapter(CausalForestDML(model_y=HGBR(), model_t=HGBC(), discrete_treatment=True), alpha=0.10, inference="statsmodels")
cml = CausalMLAdapter(BaseTRegressor(learner=HGBR(), ate_alpha=0.10)) # upstream CausalML control
st = StochtreeAdapter(alpha=0.10)For CausalML, analytical ATE CI settings stay on the wrapped upstream estimator (ate_alpha on meta-learners / TMLE, alpha on CausalTreeRegressor); the adapter does not override them.
load_lalonde(binarize_y=...) returns the 1978-earnings outcome as a binary indicator — "median" for the (~50/50) above-median split, "positive" for the (~69/31) "any 1978 earnings" indicator. Useful as a real-data fixture without leaving the package.
from metacausal import CausalEnsemble
from metacausal.datasets import load_lalonde
X, T, Y = load_lalonde(binarize_y="median")
# outcome_type="auto" detects binary Y, materialises the binary
# default pool (8 components targeting the risk difference), and
# fits. ATE is on the risk-difference scale, in [-1, 1].
ens = CausalEnsemble()
ens.fit(X, T, Y, random_state=42)
print(ens.ate().ate)from metacausal import CausalEnsemble
from metacausal.aggregation import CausalStacking
from metacausal.datasets import load_lalonde
X, T, Y = load_lalonde() # continuous outcome; reload since the previous
# recipe rebound X, T, Y to a binary outcome
ens = CausalEnsemble(aggregation=CausalStacking())
ens.fit(X, T, Y, random_state=42)
# Pointwise CATE CIs at X_eval (any covariate matrix; omit to
# evaluate at the training X instead). n_boot=10 keeps this snippet
# quick to run; bump to 200+ for publication-quality intervals.
X_eval = X
boot = ens.bootstrap(X_eval, n_boot=10, random_state=42, n_jobs=-1)
print(boot.cate) # ensemble CATE at X_eval, shape (n_eval,)
print(boot.cate_ci_lower) # pointwise 95% lower bound
print(boot.cate_ci_upper) # pointwise 95% upper bound
# Inspect the learned ensemble weights
print(boot.ensemble_weights)An aggregation=... argument to ate() or cate() re-aggregates from cached predictions without refitting components — useful for quick comparisons.
ens = CausalEnsemble(aggregation="median")
ens.fit(X, T, Y, random_state=42)
for agg in ["median", "mean", "trimmed_mean", "cba"]:
ate = ens.ate(aggregation=agg)
print(f"{agg:<15} ATE = {ate.ate:.1f}")The optional metacausal.plots submodule (installed via the [plots] extra) provides four matplotlib helpers that consume the result types above:
forest(boot)— component and ensemble ATEs with bootstrap CIs.weights(ens)— aggregation weight bars (agreement-based and supervised strategies).cate_profile(source, x, xlabel=...)— ensemble CATE along one covariate, with optional bootstrap band and per-component overlay.disagreement(ens, X)— pairwise component-CATE rank-correlation heatmap.
Each function is also available as a thin method on the class it plots — e.g. boot.forest() is equivalent to forest(boot), and ens.weights() to weights(ens).
from metacausal.plots import forest, cate_profile
forest(boot)
grid = X[:, 6] # `re74`; see load_lalonde()'s docstring for column order
cate_profile(boot, x=grid, xlabel="re74 (1974 earnings, USD)")MetaCausal exposes five injection points that let researchers extend the package without forking it: custom component adapters, custom aggregation strategies, replacement nuisance pipelines (fit_nuisance_fn), replacement pseudo-outcome functions (pseudo_outcome_fn), and custom cross-fitting splitters. The accompanying paper (forthcoming) covers each injection point in detail.
The lowest-effort path for adding a new estimator is GenericCATEAdapter, which wraps a fit function, a CATE prediction function, and (optionally) an ATE prediction function into a component without implementing the full protocol:
from metacausal import CausalEnsemble, GenericCATEAdapter
def fit_fn(X, T, Y, **kwargs):
# Train your model and return any state you need.
...
return state
def cate_fn(state, X):
# Return per-observation CATE estimates, shape (n,).
return state.predict_cate(X)
def ate_fn(state, X): # optional; defaults to mean of cate_fn(state, X)
return float(cate_fn(state, X).mean())
my_method = GenericCATEAdapter(
fit_fn, cate_fn, fn_ate=ate_fn, name="my_method",
)
ens = CausalEnsemble(methods=[my_method, ...])Parallelism cooperation contract. MetaCausal automatically suppresses the internal n_jobs/cv_n_jobs of the estimators it wraps for EconML, CausalML, DoubleML, and stochtree whenever they run inside one of MetaCausal's own workers (fit/bootstrap, or supervised cross-fitting), so those defaults can't nest a second worker pool and oversubscribe cores. That automatic suppression can't reach into an opaque fn_fit/fn_cate/fn_ate callable — if your model parallelizes internally, either construct it with its own parallelism knob fixed at 1 up front, or check the same signal MetaCausal's built-in adapters use:
import os
from metacausal.adapters import INNER_WORKER_ENV
def fit_fn(X, T, Y, **kwargs):
model = MyParallelModel(n_jobs=-1)
if os.environ.get(INNER_WORKER_ENV):
# Already inside a MetaCausal worker; nesting our own pool here
# would oversubscribe cores, so fall back to serial.
model.n_jobs = 1
model.fit(X, T, Y)
return modelA single random_state seed deterministically propagates to every stochastic sub-step — component models, their sub-estimators, cross-fitting folds, nuisance fits, and bootstrap replicates — so reruns are bit-identical.
A single n_jobs knob on fit, bootstrap, and estimate routes parallelism to the outermost applicable level (bootstrap replicates when n_boot > 0; otherwise supervised cross-fitting or component fits) and pins BLAS/OpenMP threads inside each worker to prevent oversubscription. The accompanying paper (forthcoming) explains the rationale.
This exclusive-outer-parallelism design also covers the estimators MetaCausal wraps, not just its own workers: for EconML, CausalML, DoubleML, and stochtree components, MetaCausal suppresses the wrapped estimator's own internal parallelism knob (EconML/DoubleML's n_jobs, CausalML's cv_n_jobs) whenever it runs inside one of MetaCausal's workers, as an ongoing guarantee rather than a one-off fix. Custom components (GenericCATEAdapter, arbitrary callables) are not covered automatically — see "Extending MetaCausal" above for the cooperation contract if your own model parallelizes internally.
The outer process (your main script) keeps the platform-default BLAS thread count, which is fine on macOS and Windows. On Linux, where joblib's loky backend can occasionally deadlock at fork time when the parent's BLAS pool is already running threads, defensive users may want to set the standard thread env vars (OMP_NUM_THREADS=1, OPENBLAS_NUM_THREADS=1, MKL_NUM_THREADS=1, NUMEXPR_NUM_THREADS=1, VECLIB_MAXIMUM_THREADS=1) before invoking Python. This package's own test suite (tests/conftest.py) sets these automatically, so contributors running pytest locally do not need the shell prefix.
Known issue —
CausalForestDMLunder heavy bootstrap parallelism. EconML's generalized-random-forest tree builder (econml.tree) intermittently segfaults (SIGSEGV/SIGBUS) under the repeated component refits thatbootstrap()/estimate(..., n_boot=...)perform on the default pool — a latent out-of-bounds bug in the upstream library, triggered probabilistically and more often at highern_jobs/n_boot(py-why/EconML#470, open and unfixed as of EconML 0.16.0). The other eight default components are unaffected. If you hit it, dropCausalForestDMLfrom the pool:from metacausal import CausalEnsemble from metacausal.defaults import default_methods methods = [m for m in default_methods() if type(m).__name__ != "CausalForestDML"] ens = CausalEnsemble(methods=methods)
from metacausal import CausalEnsemble
from metacausal.aggregation import CausalStacking
# Parallelise supervised cross-fitting, deterministic:
ens = CausalEnsemble(aggregation=CausalStacking())
ens.fit(X, T, Y, random_state=42, n_jobs=-1)
# Or: full fit + bootstrap pipeline with bootstrap-level parallelism.
# n_boot=10 keeps this snippet quick to run; bump to 500+ for
# publication-quality intervals.
boot = ens.estimate(X, T, Y, n_boot=10, random_state=42, n_jobs=-1)A BibTeX entry will be added here when the arXiv preprint of the accompanying manuscript is posted. For interim references to the software itself, see the PyPI listing.
- Paper: a preprint covering the methodology, architecture, and extensibility hooks is in preparation. An arXiv link will be added here once it is posted.
- Replication material: will be included as ancillary files with the forthcoming arXiv submission.
See CHANGELOG.md for the full version history.
MetaCausal is distributed under the MIT License. See LICENSE.