---

name: critique-calculation description: Critically analyze a calculated parameter for academic rigor. Checks methodology, identifies weaknesses, suggests improvements for economist audiences. allowed-tools:

• Read
• Grep
• Glob
• Bash
• TodoWrite

---

/critique-calculation <PARAMETER_NAME>

Critically analyze a calculated parameter for academic defensibility. Identifies methodological weaknesses and suggests improvements that would make the calculation more persuasive to economists.

Usage

Copy/critique-calculation POLITICAL_DYSFUNCTION_TAX_TOTAL_PCT
/critique-calculation TREATY_COMPLETE_ROI_EXPECTED
/critique-calculation DFDA_NET_BENEFIT_ANNUAL

If no parameter specified, list calculated parameters and ask which to review.

---

Phase 1: Identify the Calculation

Find the parameter definition:

Copygrep -B5 -A50 "^<PARAM_NAME> = Parameter" dih_models/parameters.py

Extract:

• Formula/compute function: How is it calculated?
• Inputs: What parameters feed into it?
• Source type: calculated, external, definition?
• Confidence interval: Is uncertainty properly propagated?

---

Phase 2: Trace Input Dependencies

For each input parameter:

Copygrep -B5 -A30 "^<INPUT_NAME> = Parameter" dih_models/parameters.py

Build dependency tree:

CopyTARGET_PARAM
├── INPUT_1 (source: external, ref: study2023)
├── INPUT_2 (source: calculated)
│   ├── INPUT_2a (source: external, ref: who2024)
│   └── INPUT_2b (source: definition)
└── INPUT_3 (source: definition)

Classify each leaf input:

• Empirically grounded: Has peer-reviewed source with specific estimate
• Theoretically grounded: Based on theory but no empirical estimate
• Definition/assumption: Policy parameter or modeling choice

---

Phase 3: Methodological Critique

A. Additivity/Independence Check

If formula uses addition (A + B + C):

• Do components overlap? (Double-counting risk)
• Would multiplicative model be more appropriate?
  • Additive: τ_total = τ_1 + τ_2 + τ_3
  • Multiplicative: τ_total = 1 - (1-τ_1)(1-τ_2)(1-τ_3)
• Is there theoretical justification for additivity?

If formula uses multiplication (A × B × C):

• Are factors truly independent?
• Could there be interaction effects?

B. Empirical Backing Assessment

Rate each input:

Red flags:

• Key inputs with no empirical backing
• Large uncertainty on dominant inputs
• Sources from advocacy organizations vs peer review
• Old data (>10 years) for fast-changing metrics

C. Uncertainty Propagation Check

Copy# Check if inputs have uncertainty metadata
grep -E "confidence_interval|distribution|std_error" dih_models/parameters.py | grep -i "<input_names>"

Verify:

• All leaf inputs have uncertainty (CI, distribution, or std_error)
• Calculated parameter derives uncertainty from Monte Carlo
• CI width is plausible (not too narrow = overconfident, not too wide = useless)

D. Sensitivity Analysis

Check if tornado/sensitivity charts exist:

Copyls knowledge/figures/tornado-<param_name>*.qmd
ls _analysis/tornado_<PARAM_NAME>.json

Review sensitivity results:

• Which inputs dominate the output variance?
• Are the dominant inputs well-grounded empirically?
• Would reducing uncertainty on key inputs substantially tighten the estimate?

---

Phase 4: Identify Improvements

Categorize potential improvements by effort/impact:

Low Effort, High Impact

• Fix additivity with multiplicative model
• Add confidence intervals to inputs missing them
• Cite better sources for key inputs
• Acknowledge overlap explicitly in text

Medium Effort, Medium Impact

• Find empirical estimates for theoretically-grounded inputs
• Add sensitivity analysis if missing
• Split aggregate parameters into measurable components

High Effort (Note for Future)

• Commission original research to fill gaps
• Build more sophisticated interaction models
• Conduct meta-analysis of existing estimates

---

Phase 5: Generate Report

Copy## Calculation Critique: <PARAMETER_NAME>

### Summary
- **Central estimate:** X (95% CI: Y-Z)
- **Formula:** [describe calculation]
- **Academic defensibility:** Strong / Medium / Weak

### Dependency Tree
[ASCII tree of inputs]

### Empirical Grounding
| Component | Backing | Strength | Key Source |
|-----------|---------|----------|------------|

### Methodological Issues
1. **[Issue]:** [Description]
   - **Impact:** [How it affects credibility]
   - **Fix:** [Recommended solution]

### Recommended Improvements
**Quick wins:**
- [ ] [Improvement 1]
- [ ] [Improvement 2]

**Consider for future:**
- [ ] [Improvement 3]

### What Economists Would Ask
1. [Likely challenge and how to respond]
2. [Likely challenge and how to respond]

---

Common Patterns

Pattern: Additive Components That Overlap

Problem: Sum of parts > whole due to double-counting Fix: Use multiplicative model: 1 - Π(1-τ_i) Example: Political dysfunction tax components

Pattern: Theoretical Decomposition Without Empirical Parts

Problem: Elegant theory, but component estimates are guesses Fix: Either find empirical estimates OR present as conceptual framework, not precise calculation Example: Time-inconsistency, information, coordination taxes

Pattern: Single Study Anchor

Problem: Entire calculation depends on one study's estimate Fix: Cite meta-analysis if available, OR widen confidence interval, OR conduct sensitivity to source Example: Del Rosal (2011) rent-seeking survey

Pattern: Stale Data

Problem: Key inputs from 5+ years ago for rapidly changing metrics Fix: Update with recent data, OR note vintage and direction of likely bias Example: GDP, spending, population figures

Pattern: Missing Counterfactual

Problem: Benefit calculated vs. status quo, but status quo is changing Fix: Specify baseline scenario, consider alternative counterfactuals Example: DFDA benefits vs. current FDA (but FDA is evolving)

---

Rules

1. Be honest about weaknesses - Credibility comes from acknowledging limits
2. Propose fixes, not just critiques - Every issue should have a recommendation
3. Distinguish fatal flaws from minor issues - Not everything needs fixing
4. Consider the audience - What would a skeptical economist ask?
5. Check if fixes already exist - Maybe the issue is presentation, not calculation