Request: Matching methods notebook

[bcdelodx]

Would love to see a comprehensive notebook on matching methods added to KASS.

The Need

Many policy evaluations can't use diff-in-diff or RDD because:

• Treatment timing is staggered and individual-specific
• No clear discontinuity or natural experiment
• But we have rich pre-treatment covariates that could support selection-on-observables identification

Proposed Solution

A matching methods notebook covering:

Core methods:

• Propensity score matching (1:1, 1:k, kernel)
• Coarsened exact matching
• Entropy balancing
• Mahalanobis distance matching

Diagnostic toolkit:

• Balance tests (standardized differences, variance ratios)
• Common support assessment
• Sensitivity analysis (Rosenbaum bounds)
• Placebo outcome tests

Implementation decisions:

• How to choose matching algorithm
• Caliper selection
• Handling non-overlap
• Standard error calculation (accounting for matching)

Use Cases

This would be valuable for:

• Program evaluation where participants self-select (job training, education programs)
• Policy analysis where treatment roll-out was non-random but recorded
• Observational studies with rich administrative data
• Situations where we can credibly argue "selection on observables"

Alternatives Considered

Could adapt the heterogeneous treatment effects notebook (11-HTE) since it uses propensity scores, but:

• That's focused on discovering heterogeneity, not just average effects
• Doesn't cover balance diagnostics comprehensively
• Doesn't discuss matching-specific inference issues

I Can Help By

• Providing domain expertise on where matching is used in workforce policy
• Testing the implementation on real use cases
• Writing the notebook (would need guidance on KASS standards)
• Reviewing methodology

Additional Context

Matching methods are workhorses of applied program evaluation, especially in:

• Labor economics (training program evaluation)
• Education (school choice, intervention programs)
• Healthcare (treatment effectiveness studies)

Having a rigorous, well-documented implementation would be hugely valuable for practitioners who need these methods but want to ensure they're applying them correctly.

---

Thoughts from maintainers or community on priority and approach?