Average imputed earnings percentiles with national values

[vahid-ahmadi]

Summary

• Missing ASHE percentiles (91–98, 100) were imputed using ratio scaling only — this assumes every area has the same income tail shape as the national distribution, causing instability for areas with unusual tails (e.g. wealthy London boroughs get inflated upper percentiles, poorer areas get theirs deflated)
• Now blends the ratio-scaled estimate with the national reference value (50/50 by default), pulling extreme local estimates toward the national centre (shrinkage)
• Extracts the duplicated fill_missing_percentiles() and reference_values from both LA and constituency scripts into a shared earnings_percentiles.py module
• The national_weight parameter controls blending: 0.0 = old behaviour (pure ratio), 0.5 = equal blend (default), 1.0 = pure national

Test plan

• 8 unit tests covering: known values preserved, missing values filled, all-NaN unchanged, blending pulls toward national, weight=0 matches old behaviour, weight=1 gives national values, monotonicity, downward extrapolation
• CI passes (ruff, pytest)

Closes #68

🤖 Generated with Claude Code