Investigate ACA take-up and plan-choice inputs driving high PTC estimates

Context

PR #68 fixes several healthcare target-mapping bugs in the fiscal target registry. This issue tracks the remaining ACA data/calibration concerns that are intentionally out of scope for that PR.

What I observed

Using the published 2024 Populace dataset in a PolicyEngine-US microsimulation:

• takes_up_aca_if_eligible is true for every tax unit in the released HDF5.
• selected_marketplace_plan_benchmark_ratio is 1.0 for every tax unit in the released HDF5.
• Because of those inputs, assigned_aca_ptc, used_aca_ptc, and raw aca_ptc are identical in the current dataset.
• Total PTC is about $92.10B, and positive PTC returns are about 12.95M tax units.

Compared with the SOI TY2022 all-income PTC targets:

• PTC amount target: $53.91B; current sim: $92.10B (+70.8%).
• PTC returns target: 7.84M; current sim: 12.95M (+65.2%).
• Average PTC per positive return is not wildly off: target is about $6.9k, current sim is about $7.1k. So the main issue appears to be too many positive PTC returns, not credits that are too large conditional on receipt.

The excess is concentrated in middle/high AGI bins:

• $75k-$100k: 1.94M positive returns vs 0.47M target.
• $100k-$200k: 2.16M positive returns vs 0.36M target.
• $200k-$500k: 0.31M positive returns vs 0 target.

For contrast, the CMS APTC recipient benchmark looks much better after the PR #68 mapping fix: eligible people with positive assigned PTC are about 20.24M vs the CMS target of 19.74M (+2.5%), with state-level correlation around 0.996.

Marketplace enrollment is still a separate concern. The reported Marketplace coverage input is about 8.77M people vs the CMS target of 21.45M. A direct-purchase union gets closer nationally but looks worse state-by-state, so this should be handled as a data/calibration design question rather than a simple variable-map change.

Why PR #68 does not cover this

PR #68 fixes cases where targets were wired to variables with the wrong semantics. The remaining ACA concern is different: the current dataset appears to have degenerate ACA take-up and selected-plan inputs. Even after mapping targets to assigned_aca_ptc, the model still overstates SOI PTC returns/spending because everyone eligible is effectively assumed to take up ACA coverage and select a benchmark-like plan.

Suggested next steps

• Locate where takes_up_aca_if_eligible and selected_marketplace_plan_benchmark_ratio are generated or defaulted; this may be outside PolicyEngine/populace.
• Add a dataset diagnostic or release check that flags degenerate health input columns, especially all-true take-up and all-1.0 plan ratio.
• Reproduce the SOI PTC amount/return diagnostics after PR #68 merges, so target mappings and source data issues are separated cleanly.
• Decide what the CMS marketplace enrollment benchmark should calibrate: reported Marketplace coverage, modeled ACA take-up, direct-purchase coverage, or some combined construct.
• Design an ACA take-up/plan-choice imputation or calibration strategy that jointly considers CMS enrollment, CMS APTC recipients, and SOI PTC amount/return targets.

Start in PolicyEngine/populace by locating where takes_up_aca_if_eligible and selected_marketplace_plan_benchmark_ratio are generated or defaulted. Reproduce the SOI PTC amount and return diagnostics after PR #68, then define a diagnostic or release check for degenerate inputs and document the agreed ACA take-up and plan-choice calibration strategy.