Add capital income predictors to asset imputation QRF

[MaxGhenis]

Summary

• Adds interest_income, dividend_income, and rental_income as predictors to the SIPP→CPS liquid asset imputation model
• Capital income is strongly correlated with asset holdings and was the major missing signal
• Updates all three imputation pathways consistently: sipp.py (model training), cps.py (CPS extraction), source_impute.py (calibration-time imputation)

Details

The QRF model imputing bank_account_assets, stock_assets, and bond_assets from SIPP 2023 to CPS previously used only:

• employment_income, age, is_female, is_married, count_under_18

Now also uses:

• interest_income (SIPP: TINC_BANK + TINC_BOND, annualized)
• dividend_income (SIPP: TINC_STMF, annualized)
• rental_income (SIPP: TINC_RENT, annualized)

These map directly to CPS variables already available in the microdata.

Motivation

With the companion PR (PolicyEngine/policyengine-us#connect-snap-assets-to-microdata) connecting snap_assets to imputed liquid assets, the SNAP asset test now binds in microsimulation. Better asset imputation is needed to avoid over/under-counting households failing the asset test when modeling BBCE elimination.

Test plan

• Verify liquid_assets.pkl retrains with new predictors
• Compare asset distribution before/after (expect closer match to SCF/SIPP actuals)
• Re-run BBCE elimination analysis with improved assets
• Existing tests pass (predictor list assertions still hold)

🤖 Generated with Claude Code