Add occupation/industry as predictors in tip income imputation model

[MaxGhenis]

Summary

The QRF model for tip income imputation uses only 4 features: employment_income, age, count_under_18, and count_under_6. Occupation and industry are the strongest predictors of who receives tips, but are not used despite being available in SIPP.

Available SIPP variables (already loaded but unused)

• TJB*_OCC — Occupation codes per job (up to 7 jobs)
• TJB*_IND — Industry codes per job (up to 7 jobs)

Why this matters

Tip income is highly concentrated in specific occupations (food servers, bartenders, hairdressers, etc.) and industries (NAICS 72: Accommodation and Food Services). Without occupation/industry, the model spreads tip income more diffusely across the income distribution, which:

1. Understates tip concentration among low-wage service workers
2. Reduces accuracy of the distributional impact of "no tax on tips"
3. May assign tips to workers in non-tipped occupations

Suggested approach

1. Map SIPP occupation/industry codes to CPS occupation/industry codes (or use broad categories like 2-digit NAICS)
2. Add these as categorical features to the QRF model in sipp.py
3. Ensure the CPS recipient dataset has matching occupation/industry variables for prediction

Mapping challenge

CPS and SIPP use different occupation/industry classification systems, so a crosswalk may be needed. At minimum, a broad industry indicator (e.g., food services vs. other) would capture most of the signal.

Context

This is the highest-impact improvement for closing the gap between PolicyEngine's tip deduction estimate ($4.7B) and JCT's score ($10.0B for FY2026).