Estimation¶

Methods for estimating causal effects after identification.

Backdoor Estimators¶

Estimator	Type	Best For	Key Parameters
`LINEAR_REGRESSION`	Parametric	Simple, interpretable	None
`PROPENSITY_SCORE_MATCHING`	Matching	Few covariates, balance needed	`n_neighbors`, `caliper`
`PROPENSITY_SCORE_WEIGHTING`	Weighting	Many covariates	`stabilized`, `trim_quantiles`
`DOUBLY_ROBUST`	Augmented IPW	Default recommendation	`propensity_model`, `outcome_model`
`TARGETED_MAXIMUM_LIKELIHOOD`	TMLE	Efficient, asymptotic	`initial_estimator`
`CAUSAL_FOREST`	ML / CATE	Heterogeneous effects	`n_estimators`, `min_samples_leaf`
`DOUBLE_ML`	Orthogonal ML	High-dim confounders	`ml_g`, `ml_m`

Choosing an Estimator¶

Is treatment effect homogeneous?
├── Yes → linear_regression, doubly_robust (ATE)
└── No (need CATE) → causal_forest, double_ml, or EconML metalearners

Are covariates high-dimensional?
├── Yes → double_ml, causal_forest
└── No → any

Need inference (CI, p-values)?
├── Yes → doubly_robust, tmle, causal_forest (with honest forest)
└── Must be simple → linear_regression

EconML CATE Estimators¶

Metalearner	Description	When to Use
`T_LEARNER`	Separate models for T=0, T=1	Simple, sparse data
`S_LEARNER`	Single model with T as feature	Strong overlap, interactions
`X_LEARNER`	Learns treatment/control effects separately	Many control, few treated
`R_LEARNER`	Residual-on-residual (Robinson)	General purpose, orthogonality
`DR_LEARNER`	Doubly robust residual	Best overall properties

from causal_toolkit.wrappers import EconMLWrapper
from causal_toolkit.core.base import EstimatorType

econml = EconMLWrapper(
    data=df,
    treatment="treatment",
    outcome="outcome",
    covariates=["X1", "X2", "X3"],
    effect_modifiers=["X1", "X2"]  # CATE varies by these
)

# Get CATE
cate = econml.estimate_cate(EstimatorType.CAUSAL_FOREST_CATE)

# Get ATE (average of CATE)
ate = econml.estimate_ate(EstimatorType.CAUSAL_FOREST_CATE)

Instrumental Variables¶

Estimator	Method
`TWO_STAGE_LS`	2SLS, linear first/second stage
`DEEP_IV`	Deep learning for complex IV
`ORTHO_IV`	Orthogonal IV (Chernozhukov)

Estimator Parameters¶

estimation:
  method: "doubly_robust"
  params:
    propensity_model: "logistic"        # or sklearn estimator
    outcome_model: "linear_regression"  # or sklearn estimator
    cv: 5                               # cross-validation folds

For causal_forest:

estimation:
  method: "causal_forest"
  params:
    n_estimators: 1000
    min_samples_leaf: 10
    max_depth: 10
    honest: true        # Use honest forest for valid inference
    random_state: 42

Diagnostics¶

estimate = dowhy.estimate(EstimatorType.DOUBLY_ROBUST)
print(estimate.diagnostics)
# {
#   "propensity_model_score": 0.72,
#   "outcome_model_score": 0.65,
#   "effective_sample_size": 847,
#   "weight_summary": {"mean": 1.0, "std": 0.3, "max": 5.2}
# }