Behavior Propensities Missing

Many real datasets do not include logged behavior propensities. You have two options:

1. Estimate propensities from data.
2. Use estimators that do not require propensities (when applicable).

Option 1: estimate propensities

CausalRL includes behavior policy estimation utilities (requires scikit-learn). Install with python -m pip install -e ".[behavior]".

from crl.behavior import fit_behavior_policy

fit = fit_behavior_policy(dataset, method="logit", seed=0)
with_props = fit.apply(dataset)

This attaches estimated propensities and records metadata so reports can warn that propensities were modeled.

Option 2: choose prop-free estimators

• For MDPs, model-based estimators like FQE can run without propensities.
• For bandits, you typically need propensities to use IS/WIS. evaluate will skip propensity-based estimators when propensities are missing.

Best practices

• Always log true behavior propensities when possible.
• If you estimate them, run diagnostics and be explicit in reporting.
• Sensitivity analysis can help quantify robustness under misspecification.