Calibrate SSI disability and take-up for eligibility-expanding reforms

[MaxGhenis]

SSI disability status is currently based on SSI receipt. As a result, we will underestimate the impact of reforms that expand SSI eligibility, e.g. by income or assets. We could impute SSI disability status from SSI eligibility and general disability flags.

[MaxGhenis]

Discussing with @nikhilwoodruff we'll actually say that anyone with a CPS reported disability is SSI disabled, and then add a SSI take-up flag which will be true for anyone claiming SSI in the CPS, otherwise a fraction.

[MaxGhenis]

This is now blocking credible SSI asset-limit reform estimates. The issue is not just imputing a boolean disability flag; the observed outcome is censored by income/assets, application, take-up, and SSA medical determination.

Recent local no-asset-limit SSI reform check:

• PolicyEngine: +1.399M SSI recipients (+24%).
• CBPP no-limit study Kathleen Romig shared: +462k participants (+6%).
• PolicyEngine adults 18-64: +1.006M vs CBPP +298k.
• PolicyEngine seniors 65+: +393k vs CBPP +115k.
• PolicyEngine children: +0 vs CBPP +49k.
• PolicyEngine cash SSI spending: +$12.84B/year; Medicaid enrollment +584k and Medicaid cost +$20.1B/year.

Why current model overstates:

• takes_up_ssi_if_eligible defaults to true.
• is_ssi_disabled is currently broad survey disability minus SGA, not probability of passing SSA disability determination and applying.
• Current SSI receipt cannot directly identify latent SSI disability for people failing income/resources; eligibility censoring means non-receipt combines no disability, denial, non-application, and financial ineligibility.

Proposed modeling direction:

1. Separate latent SSA medical eligibility / successful application from broad survey disability.
2. For baseline recipients, force probability/take-up to 1 where SSI is observed.
3. For non-recipients, estimate or parameterize a probability using survey disability, age, SSDI/OASDI disability receipt, Medicare disability indicators if available, work limitation, earnings/SGA, family structure, state, and income/assets.
4. Calibrate baseline SSI recipients/spending to SSA admin totals and calibrate SSI asset-limit counterfactuals to external evidence, starting with CBPP's +462k no-limit estimate.
5. Keep separate knobs for medical allowance and take-up so reform uncertainty is explicit rather than hidden in is_disabled.

Acceptance criteria for first PR:

• Add a parameterized probability/take-up mechanism that leaves observed baseline SSI recipients intact.
• Asset-limit no-limit reform should no longer mechanically add all resource-ineligible broad-disabled people.
• Add tests showing non-recipient broad disability does not automatically imply SSI receipt when resources are relaxed.
• Add a calibration note/test target comparing no-limit participation to CBPP as an external benchmark.

[MaxGhenis]

Implementation should move primarily to policyengine-us-data; the policy model consumes is_disabled / take-up inputs, but the latent SSA medical allowance and application/take-up problem is data-generation/calibration. Filed the us-data implementation issue here: PolicyEngine/policyengine-us-data#1094.

Leaving this issue open as the consumer-side contract: policyengine-us should not have to infer true SSA disability from broad CPS disability alone, and tests should protect against asset-limit reforms mechanically granting SSI to every broad-disabled resource-ineligible person.

[MaxGhenis]

Opened draft PR #8364 to add the PolicyEngine-US contract variable for this work: #8364

Implementation detail: meets_ssi_disability_criteria is data-overridable and falls back to the current is_disabled & ~ssi_engaged_in_sga behavior. is_ssi_disabled now delegates to that variable, so policyengine-us-data can later publish a calibrated/imputed SSI disability criteria flag without changing downstream SSI formulas.