Estimands

An estimand is the causal quantity you want to identify and estimate. CausalRL makes the estimand explicit and pairs it with assumptions.

Policy value

PolicyValueEstimand represents the expected return of a target policy.

from crl.estimands.policy_value import PolicyValueEstimand
from crl.assumptions import AssumptionSet
from crl.assumptions_catalog import BEHAVIOR_POLICY_KNOWN, OVERLAP, SEQUENTIAL_IGNORABILITY

estimand = PolicyValueEstimand(
    policy=target_policy,
    discount=0.99,
    horizon=10,
    assumptions=AssumptionSet([SEQUENTIAL_IGNORABILITY, OVERLAP, BEHAVIOR_POLICY_KNOWN]),
)

Policy contrast

PolicyContrastEstimand captures differences between two policies.

from crl.estimands.policy_value import PolicyContrastEstimand

contrast = PolicyContrastEstimand(
    treatment=policy_a,
    control=policy_b,
    discount=0.99,
    horizon=10,
)

Sensitivity estimand

SensitivityPolicyValueEstimand encodes a gamma-bounded confounding model and produces partial identification bounds.

Proximal estimand

ProximalPolicyValueEstimand encodes proximal assumptions about proxy variables and bridge functions.

Why estimands matter

• They define what you can claim.
• They connect assumptions to estimators.
• They force clarity when data are incomplete.