Diagnostics

Diagnostics are the evidence you present alongside any policy value estimate. They answer: should you trust this number?

What diagnostics measure

• Overlap: support violations or near-zero behavior probabilities.
• ESS: how much effective data remains after weighting.
• Weight tails: risk of estimates dominated by a few samples.

How to read them

• Overlap violations mean identification is broken.
• Low ESS signals high variance and fragile estimates.
• Heavy tails mean that even unbiased estimators can be unstable.

Where diagnostics appear

• EstimatorReport.diagnostics
• EstimatorReport.warnings
• HTML reports generated by OpeReport.to_html()