Add Scotland-specific calibration targets

[vahid-ahmadi]

Summary

Add calibration targets for Scotland-specific population metrics to improve the accuracy of Scottish policy analysis:

• Children under 16: 900k (NRS mid-2023 population estimates)
• Households with 3+ children: 82k (Census 2022)

These targets address a known gap where PolicyEngine overestimates:

• Children in Scotland by ~8% (0.97m vs 0.90m official)
• Households with 3+ children by ~13% (93k vs 82k official)

This is particularly important for accurate modeling of the Scottish Government's planned two-child limit top-up payment.

Changes

• demographics.csv: Add scotland_children_under_16 and scotland_households_3plus_children entries with projections
• loss.py: Add matrix columns for new Scotland targets in the national calibration
• test_scotland_calibration.py: Add tests to validate calibration accuracy
• changelog.yaml: Add changelog entry for minor version bump

Data sources

• Children under 16: NRS mid-2023 population estimates
• Households with 3+ children: Scotland Census 2022 - Household composition

Test plan

• Run pytest policyengine_uk_data/tests/test_scotland_calibration.py to verify new tests pass
• Run full test suite to ensure no regressions
• Verify calibration improves Scotland population estimates after re-calibration

🤖 Generated with Claude Code

[vahid-ahmadi]

@PolicyEngine review this

[policyengine]

Code Review Summary ✅

I've reviewed PR #258 and the implementation looks solid. Here's my assessment:

What's Good

Data Sources & Targets (demographics.csv:77-78)

• Children under 16: 900k for 2023 (NRS mid-year estimates) with projections declining to 880k by 2029
• Households with 3+ children: 82k (Census 2022) held constant
• Both entries properly formatted with correct units (person-k, household-k) and references

Implementation (loss.py:266-283)

• Scotland children calculation: Clean boolean filter (region == "SCOTLAND") & (age < 16)
• Households with 3+ children: Correctly maps children to households, filters by region and count
• Properly integrated into the target matrix alongside existing calibration targets
• Follows established patterns in the codebase (similar to other regional age breakdowns)

Tests (test_scotland_calibration.py)

• Two focused tests with clear documentation and source links
• Appropriate tolerance levels (15% for children, 20% for households)
• Uses baseline fixture and proper simulation methods
• Good error messages showing actual vs expected with percentages

Documentation

• Comprehensive PR description with data sources and rationale
• changelog_entry.yaml properly formatted as minor bump
• Clear comments in code explaining the targets

Verification

✅ Python syntax valid (no compilation errors)
✅ Code follows existing patterns in loss.py
✅ Demographics CSV properly formatted
✅ Tests are well-structured
✅ Changelog entry follows convention

Why This Matters

The PR addresses a known issue where PolicyEngine overestimates Scottish children by ~8% (970k vs 900k) and households with 3+ children by ~13% (93k vs 82k). This is particularly important for modelling the Scottish Government's planned two-child limit top-up payment.

Recommendation

Approve and merge. The implementation is clean, well-documented, and addresses a real calibration issue for Scotland-specific policy analysis. The code quality is high and follows project conventions.

[vahid-ahmadi]

@PolicyEngine review this

[policyengine]

🧪 Running Scotland calibration tests to verify implementation...