Improve calibration weight initialization with country-aware divisors (#262)

* Improve calibration weight initialization for devolved nations

Use country-aware initialization for calibration weights. Previously,
each household's weight was divided by total area count (e.g., 650 for
constituencies), causing Scottish households to start at ~9% of their
target (59/650). Now weights are divided by areas in that household's
country, so all countries start at ~100% of their targets.

This should improve convergence speed and final accuracy for Scotland,
Wales, and Northern Ireland specific targets like Scottish Child Payment.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Format with black

---------

Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>

Author: MaxGhenis

changelog_entry.yaml

@@ -0,0 +1,4 @@
+- bump: patch
+  changes:
+    changed:
+      - Improved calibration weight initialization to use country-aware divisors. Households now start with weights divided by the number of areas in their country rather than total areas, improving convergence for devolved nation targets.

policyengine_uk_data/utils/calibrate.py

@@ -46,8 +46,18 @@ def calibrate_local_areas(
     m_n, y_n = m_national.copy(), y_national.copy()
 
     # Weights - area_count x num_households
+    # Use country-aware initialization: divide each household's weight by the
+    # number of areas in its country, not the total area count. This ensures
+    # households start at approximately correct weight for their country's targets.
+    # The country_mask r[i,j]=1 iff household j is in same country as area i.
+    areas_per_household = r.sum(
+        axis=0
+    )  # number of areas each household can contribute to
+    areas_per_household = np.maximum(
+        areas_per_household, 1
+    )  # avoid division by zero
     original_weights = np.log(
-        dataset.household.household_weight.values / area_count
+        dataset.household.household_weight.values / areas_per_household
         + np.random.random(len(dataset.household.household_weight.values))
         * 0.01
     )