LSR and CG behavioral responses produce nonsensical results when combined

[MaxGhenis]

Summary

When both labor supply responses (LSR) and capital gains (CG) behavioral responses are enabled simultaneously, the simulation produces nonsensical results — e.g., $8.6 trillion in federal tax revenue impact for a single year, compared to $6.2B (LSR only) or $21.9B (CG only).

Root cause

LSR and CG behavioral responses each create simulation branches to measure marginal tax rates, but they don't neutralize each other's responses in their measurement branches. This creates a compounding feedback loop:

The branch nesting problem

1. LSR creates lsr_measurement and baseline_lsr_measurement branches, neutralizing employment_income_behavioral_response and self_employment_income_behavioral_response
2. Inside those branches, relative_income_change computes household_net_income, which depends on long_term_capital_gains
3. long_term_capital_gains = long_term_capital_gains_before_response + capital_gains_behavioral_response
4. capital_gains_behavioral_response is NOT neutralized in the LSR branches, so it fires
5. capital_gains_behavioral_response calls relative_capital_gains_mtr_change, which creates its OWN sub-branches (baseline_cgr_measurement, cgr_measurement)
6. Inside those, marginal_tax_rate_on_capital_gains creates FURTHER branches (adult_1_cg_rise, adult_2_cg_rise)

The result is deeply nested branches where each behavioral response computes MTRs in a context where the other response's variables have been partially neutralized, producing incorrect MTR measurements. These wrong MTRs feed into wrong behavioral responses, which compound into the $8.6T result.

Specifically

• LSR measurement branches see CG-adjusted household_net_income (because CG response isn't neutralized), so the income change and wage change used for LSR are wrong
• CG response inside LSR branches computes MTRs in a context where employment income responses are neutralized, getting wrong CG MTRs
• The responses amplify each other instead of being independent

Reproduction

from policyengine_core.reforms import Reform
from policyengine_us import Microsimulation

DATE_RANGE = "2020-01-01.2100-12-31"

# Any reform that changes taxes
reform = Reform.from_dict({
    "gov.contrib.congress.watca.in_effect": {DATE_RANGE: True},
    "gov.contrib.congress.watca.surtax.in_effect": {DATE_RANGE: True},
}, country_id="us")

# Enable both LSR and CG responses
lsr = Reform.from_dict({
    "gov.simulation.labor_supply_responses.elasticities.income": {DATE_RANGE: -0.05},
    "gov.simulation.labor_supply_responses.elasticities.substitution.all": {DATE_RANGE: 0.25},
}, country_id="us")

cg = Reform.from_dict({
    "gov.simulation.capital_gains_responses.elasticity": {DATE_RANGE: -0.62},
}, country_id="us")

sim_baseline = Microsimulation()
sim_reform = Microsimulation(reform=(reform, lsr, cg))

# This produces ~$8.6T instead of a reasonable ~$15-25B
tax_baseline = sim_baseline.calculate("income_tax", period=2026, map_to="household")
tax_reform = sim_reform.calculate("income_tax", period=2026, map_to="household")
print(f"Impact: ${float((tax_reform - tax_baseline).sum()) / 1e9:.1f}B")

Individual responses work correctly:

• LSR only: $6.2B (reasonable)
• CG only: $21.9B (reasonable)
• LSR + CG: $8,651.3B (broken)

Proposed fix

Each behavioral response's measurement branches should neutralize ALL other behavioral responses, not just their own. Specifically:

1. In labor_supply_behavioral_response.formula, add capital_gains_behavioral_response to the list of neutralized variables in the measurement branches
2. In relative_capital_gains_mtr_change.formula, add employment_income_behavioral_response and self_employment_income_behavioral_response to the list of neutralized variables in the measurement branches

This makes the responses additive (each measured against the static baseline) rather than compounding.

Related issues

• Applying behavioral responses to empty reform causes nonzero change #5504 — behavioral responses breaking with empty reform
• Behavioral responses breaking in web app #5654 — behavioral responses breaking in web app
• Childcare subsidy programs should use *_before_lsr income variables to avoid LSR circular dependency #7024 — childcare subsidy LSR circular dependency
• Bug — LSR flips sign when baseline earnings < 0 #6053 — LSR sign flip with negative earnings

[MaxGhenis]

Update: narrowed root cause

The neutralization fix (PR #7788) works for parameter-only reforms (tested: standard deduction change → combined LSR+CG impact is -$41.3B, bounded at 0.3x individual sum). But WATCA still produces $8.6T.

Key finding

The bug is NOT a general LSR+CG interaction issue. It's specific to structural reforms that override income_tax_before_credits (like WATCA).

• Empty reform + LSR + CG → $0 (correct)
• Parametric reform (std deduction) + LSR + CG → -$41.3B (correct)
• WATCA (structural) + LSR + CG → $8,651B (broken)

The "params only" test also produced $8.6T because create_structural_reforms_from_parameters (reforms.py L352) automatically applies the WATCA structural reform when watca.in_effect = True.

Next step

Investigate how the WATCA structural override of income_tax_before_credits (which adds watca_millionaire_surtax to the formula) interacts with behavioral response simulation branching. The surtax may be computed on wrong values inside measurement branches, or applied multiple times across nested branches.

[MaxGhenis]

Update: two-pass approach blocked by circular DAG dependencies

Attempted a Yale-style two-pass framework (two_pass_behavioral_response.py) that creates ONE set of measurement branches with ALL responses neutralized, measures all MTRs, then applies responses. This hits a recursion error:

household_net_income → household_benefits → snap → tanf → dc_tanf_eligible 
→ dc_tanf_meets_work_requirements → weekly_hours_worked 
→ income_elasticity_lsr → relative_income_change → household_net_income

The DAG has circular dependencies mediated by behavioral responses — benefits depend on hours worked, hours depend on LSR, LSR depends on net income, net income depends on benefits. The old branching code breaks this cycle via the _lsr_calculating re-entry guard and specific branch naming. A two-pass framework would need to break all these cycles simultaneously.

Summary of findings

1. Parametric reforms + LSR + CG: Fixed by neutralization (PR Fix LSR and CG behavioral responses compounding when combined #7788). Combined impact is -$41.3B (0.3x individual sum).
2. Structural reforms (WATCA) + LSR + CG: Still broken ($8.6T). The structural override of income_tax_before_credits interacts with behavioral response branches in a way that neutralization alone doesn't fix.
3. Empty reform + LSR + CG: $0 (correct). The issue is NOT in core's branching.
4. Yale-style two-pass: Blocked by circular DAG. Needs deeper refactor.

Recommended path forward

The structural reform issue likely requires changes in policyengine-core to how get_branch + clone_system interacts with structural variable overrides during behavioral response calculations. Filing a core issue to track this.

[MaxGhenis]

Technical spec for fixing combined LSR+CG with structural reforms

What works

PR #7788 fixes parametric reforms by neutralizing cross-responses in measurement branches. This is merged-ready.

What's broken

Structural reforms (like WATCA, which overrides income_tax_before_credits via update_variable) still produce $8.6T when LSR+CG are both enabled. Parametric-only reforms produce correct bounded results.

The circular dependency problem

The DAG has a cycle mediated by behavioral responses:

household_net_income → household_benefits → snap → tanf 
→ dc_tanf_meets_work_requirements → weekly_hours_worked 
→ income_elasticity_lsr → relative_income_change → household_net_income

The existing code breaks this cycle via:

1. _lsr_calculating re-entry guard on labor_supply_behavioral_response
2. Branch-specific variable neutralization
3. requires_computation_after on income_elasticity_lsr and substitution_elasticity_lsr

What a fix needs to do

Compute both LSR and CG responses in a single coordinated pass that:

1. Creates ONE set of measurement branches with ALL behavioral responses neutralized
2. Measures all MTRs (wage via marginal_tax_rate, CG via marginal_tax_rate_on_capital_gains) in those branches
3. Computes income/substitution effects and CG response from the measured MTRs
4. Returns responses WITHOUT triggering the intermediate variables (relative_income_change, relative_wage_change, income_elasticity_lsr, substitution_elasticity_lsr) that cause the circular dependency

Key files

• policyengine_us/variables/gov/simulation/labor_supply_response/labor_supply_behavioral_response.py — LSR entry point
• policyengine_us/variables/gov/simulation/capital_gains_responses.py — CG response + MTR measurement
• policyengine_us/variables/gov/simulation/labor_supply_response/relative_income_change.py — reads lsr_measurement branch
• policyengine_us/variables/gov/simulation/labor_supply_response/relative_wage_change.py — reads baseline_lsr_measurement branch
• policyengine_us/variables/household/marginal_tax_rate.py — creates sub-branches for MTR calculation
• policyengine_us/variables/household/income/person/weekly_hours_worked.py — triggers the circular dep via income_elasticity_lsr

Test

RUN_HEAVY_TESTS=1 python -m pytest policyengine_us/tests/microsimulation/test_lsr_cg_interaction.py::test_combined_lsr_cg_structural_reform -v

Should produce bounded result (< 3x sum of individual LSR and CG impacts). Currently produces $8,651B.