Add calibration targets for retirement contributions

[PavelMakarchuk]

Summary

Retirement contribution variables are significantly underestimated relative to administrative benchmarks. Most lack calibration targets entirely, and the two that exist (traditional_ira_contributions and roth_ira_contributions) have issues — the IRA target is sourced from back-of-envelope math and overshoots the actual deduction, and the Roth IRA target is structurally ineffective due to a bug in the CPS allocation logic.

This matters for any reform that expands the AGI base to include retirement contributions (e.g., the CRFB AGI surtax reform on surtax_reform branch in policyengine-us).

Current Model Values vs Benchmarks (2026 simulation)

Variable	Model (2026)	Benchmark	Gap	Calibrated?
traditional_ira_contributions	$26.8B	$13.2B	2x over	Yes ($25B — too high)
traditional_401k_contributions	$245.4B	$567.9B	-57% under	No
traditional_403b_contributions	$0.0B	(bundled in 401k)	N/A	No
self_employed_pension_contribution_ald	$5.9B	$29.5B	-80% under	No
self_employed_pension_contributions (input)	$15.4B	—	—	No
roth_ira_contributions	$0.0B	~$39B	Broken	Yes ($39B — ineffective)
roth_401k_contributions	$0.7B	—	Unknown	No

Benchmark Sources

IRS SOI Publication 1304, Table 1.4 (Tax Year 2022)

• IRA payments (deduction): $13,166,590,000 (2.4M filers)
• Payments to a Keogh plan (deduction): $29,483,344,000 (972K filers)
• Source: IRS SOI Tax Stats - Individual Information Return, Table 1.4 — file 22in14ar.xls, Row "All returns, total", columns 124 and 116 respectively.

BEA/FRED National Income Accounts

• Total DC employer + employee contributions: $815.4B — FRED series Y351RC1A027NBEA
• Employer DC contributions only: $247.5B — FRED series W351RC0A144NBEA
• Employee DC contributions (derived): $815.4B - $247.5B = $567.9B
• This covers 401(k), 403(b), 457, and TSP elective deferrals.

Proposed Calibration Changes in HARD_CODED_TOTALS (loss.py)

1. Fix traditional_ira_contributions: $25B → $13B

The current $25B target is from SOI IRA accumulation tables (total contributions including non-deductible). Since traditional_ira_contributions flows directly into the ALD with no deductibility logic in policyengine-us, the target should match the actual deduction claimed on returns: $13.2B from SOI 1304.

2. Add traditional_401k_contributions: target ~$568B

Not currently calibrated. The variable is a plain input that flows directly into pre_tax_contributions.yaml (subtracted from wages). The BEA employee DC figure ($567.9B) is the right benchmark since it represents actual elective deferrals.

3. Add self_employed_pension_contribution_ald: target ~$29.5B

Not currently calibrated. Unlike the other variables, this one has a formula: min(contributions, self_employment_income). The SOI 1304 Keogh figure ($29.5B) represents the actual deduction claimed, which is the right target for the ALD variable. Note: calibrating the ALD directly may be more effective than calibrating the input (self_employed_pension_contributions), since the SE income cap is binding for many filers.

4. Remove roth_ira_contributions: $39B → remove

The CPS allocation logic (cps.py:713-728) gives traditional_ira_contributions the full IRA limit first, then sets roth_ira_limit = limit_ira - traditional_ira_contributions. This is mathematically guaranteed to produce $0 for Roth IRA in all cases — either traditional IRA exhausts the limit (roth_ira_limit = 0) or it exhausts the remaining pool (remaining = 0). The $39B target is dead weight. Fixing the allocation logic is a separate issue.

How Variables Flow Into AGI

Variable	Mechanism	Deductibility Logic
traditional_ira_contributions	ALD (deductions.yaml)	None — raw value IS the deduction
traditional_401k_contributions	Pre-tax payroll (pre_tax_contributions.yaml)	None — raw value subtracted from wages
traditional_403b_contributions	Pre-tax payroll (same file)	None — raw value subtracted from wages
self_employed_pension_contribution_ald	ALD via formula min(contributions, SE_income)	Yes — capped at SE income
roth_ira_contributions	Does not reduce AGI	N/A (post-tax)
roth_401k_contributions	Does not reduce AGI	N/A (post-tax)

CPS Allocation Context

All retirement contributions originate from a single CPS variable: RETCB_VAL. The allocation waterfall in cps.py:620-728:

1. Self-employed pension (if person has SE income) — full amount
2. Traditional 401(k) — up to annual limit
3. Roth 401(k) — up to annual limit from remainder
4. Traditional IRA — up to IRA limit from remainder
5. Roth IRA — remainder within IRA limit (structurally $0, see above)

No 403(b) or 457 allocation (line 631 comment: "Assume no 403(b) or 457 contributions for now").

Related

• CRFB AGI surtax reform (surtax_reform branch in policyengine-us) needs accurate retirement contribution data
• Roth IRA allocation bug should be tracked separately

Microdata Sources

Variable	CPS	PUF	Source Field
traditional_ira_contributions	Yes (from RETCB_VAL waterfall)	No	CPS ASEC RETCB_VAL, allocated after 401k in priority
traditional_401k_contributions	Yes (from RETCB_VAL waterfall)	No	CPS ASEC RETCB_VAL, first allocation for wage earners
traditional_403b_contributions	Not allocated	No	CPS comment: 'Assume no 403(b) or 457 contributions for now'
self_employed_pension_contributions	Yes (from RETCB_VAL waterfall)	No	CPS ASEC RETCB_VAL, allocated first if person has SE income
roth_ira_contributions	Yes (from RETCB_VAL waterfall)	No	CPS ASEC RETCB_VAL, allocated last — structurally $0
roth_401k_contributions	Yes (from RETCB_VAL waterfall)	No	CPS ASEC RETCB_VAL, allocated after traditional 401k

All retirement contributions originate from a single CPS variable RETCB_VAL (person.RETCB_VAL in cps.py:682). The PUF does not separately report retirement contributions — they are embedded in the AGI calculation.

Pre-Calibration Values (extended_cps_2024, full weights)

Variable	Pre-Cal Value	Target	Ratio
traditional_ira_contributions	$0.0B	$13.2B	0.00x
traditional_401k_contributions	$441.1B	$567.9B	0.78x
self_employed_pension_contribution_ald	$13.7B	$29.5B	0.46x

Important: traditional_ira_contributions is $0 for all records in the extended CPS because the RETCB_VAL allocation waterfall consumes contributions in 401k before reaching IRA. Calibration cannot fix a variable that is $0 for every record — the allocation logic (cps.py:694-728) needs fixing first.

CPS RETCB_VAL Documentation

From the 2024 CPS ASEC Data Dictionary (p. 47):

RETCB_VAL — Retirement contribution, amount
Values: 0 = none or NIU; 1–99999 = amount contributed
Universe: RETCB_YN = 1

RETCB_YN — Retirement contribution, y/n
Values: 0 = NIU; 1 = yes; 2 = no
Universe: All people 15 years and over

RETCB_VAL is a single bundled total with no account-type breakdown. Census asks "how much did you contribute to retirement accounts?" but not "to which type?" The distribution variables (RINT_SC1/SC2) have source codes (401k, 403b, Roth IRA, Regular IRA, Keogh, SEP) but RETCB_VAL does not.

Root Cause: Sequential Waterfall

The old allocation waterfall (cps.py:682–728) gave 401(k) first priority, consuming nearly all of RETCB_VAL before reaching IRA. Since most CPS respondents report RETCB_VAL under the $23K 401(k) limit, IRA always received $0. The Roth IRA allocation was also mathematically guaranteed to produce $0 (traditional IRA either exhausted the limit or the remaining pool first).

Fix: Proportional Split (PR #554)

Replace the waterfall with a proportional allocation using administrative shares:

Split	Traditional	Roth	Source
DC vs IRA	90.8% DC	9.2% IRA	BEA/FRED + IRS SOI Tables 5 & 6
Within DC	85% traditional	15% Roth	Vanguard How America Saves 2024, PSCA 67th Annual Survey
Within IRA	39.2% traditional	60.8% Roth	IRS SOI Tables 5 & 6 (TY 2022): trad $22.5B / roth $35.0B

[PavelMakarchuk]

Missing Roth calibration targets

After extensive research, no authoritative direct administrative source exists for aggregate Roth 401(k) contribution dollars:

• IRS SOI publishes Roth IRA data (Tables 5 & 6) but has no equivalent for Roth 401(k) — 401(k) contributions aren't reported on individual tax returns (excluded from W-2 Box 1)
• DOL Form 5500 reports total participant contributions but doesn't break out Roth vs traditional
• BEA/FRED gives total employee DC deferrals ($567.9B) but no Roth/traditional split
• Vanguard, Fidelity, T. Rowe Price, PSCA, Alight all report Roth participation rates (15–21% of participants) but not dollar shares
• JCT lumps all DC plans together in tax expenditure estimates

Best available estimates

Source	Roth participation rate	Estimated dollar share
Vanguard How America Saves 2024	18% of participants	~15% (Roth participants contribute slightly less on average)
PSCA 67th Annual Survey	21% of participants	—
Alight 2024 Universe Benchmarks	12.5% of participants (2023)	—

Current state

• Roth IRA: $35.0B direct source exists (IRS SOI Tables 5 & 6, TY 2022) — could add as calibration target
• Roth 401(k): Best estimate is derived ($567.9B × 15% ≈ $85B) — no direct government source available
• The roth_share_of_dc_contributions: 0.15 parameter in imputation_parameters.yaml already uses the best available evidence for the allocation split

Decision needed

Should we:

1. Add a Roth IRA target ($35.0B from IRS SOI) since we have a direct source?
2. Add a derived Roth 401(k) target (~$85B) even though it's not a direct administrative total?
3. Skip Roth targets for now since the proportional allocation already uses these shares?