# V91 follow-up: PSID 2021–2023 timing feasibility audit
29 October 2026. Documentation audit only; no participant records acquired, sample counted, outcome estimated, calibration changed or website deployed.

## Decision
The inspected PSID food measures do not implement the v91 monthly -3:-1 / +1:+3 outcome. Keep that primary-design gate CLOSED. A between-interview spending association is a possible separate study, not an implementation or validation of the monthly design.

## Verified field anchors (partial crosswalk)
Sources are the official family codebooks linked below. Page numbers are printed PDF pages.

| Purpose | 2021 | 2023 | Scope |
|---|---|---|---|
| Family interview ID | ER78002 | ER82002 | Wave-specific; not a person linkage key |
| Interview month/day/year | ER78012/13/14 | ER82012/13/14 | Actual dates, not wave labels |
| Checking/savings/money-market amount | ER79970 (p660) | ER83940 (p647) | Current balance |
| Food-at-home amount / period, SNAP branch | ER78868/69 (p320) | ER82861/62 (p313) | Usual-week question; multiple reported units |
| Food-at-home non-SNAP branch | Not mapped | ER82871/72 (p313 note) | Branches must be reconciled |
| Main-job end month, reference person | Not mapped | ER82167 (p54) | Year and other jobs still needed |
| Main-job separation reason, reference person | Not mapped | ER82234 (p81) | Broad categories |
| Main-job separation reason, partner | Not mapped | ER82553 (p202) | Separate work history |
| Family longitudinal weight | Not mapped | ER85812 (p1357) | Variance design still unresolved |
| Total expenditure | ER81914 (p1414) | Not mapped | Constructed basket, not monthly diary |

2021 release 3 corrects annualized water expenditure and affected totals. Pin that release. Negative account balances are valid; zero amount may require routing information. Food unknown/refused codes are not zero. Branch-specific receipt, delivery, eating-out and imputation fields remain unmapped.

## Temporal-order implications and decisions
These are analysis decisions arising from the documentation, not provider-certified feasibility findings.

1. Do not use an endpoint current balance as a pre-event resource measure. A baseline balance is eligible for temporal ordering only when its actual measurement date precedes the earliest possible separation date. A job-end month overlapping baseline interview month cannot establish within-month order without additional information.
2. A 2021-to-2023 comparison could support only a between-interview usual-spending estimand, conditional on completing the crosswalk. It cannot recover missing monthly observations by dividing annualized expenses by twelve. Nor does a monthly employment history create monthly expenditure observations.
3. Baseline balances can be remote from the event. Any redesigned historical specification must define and justify permitted lags before inspecting outcomes; never choose a lag to improve significance.
4. Usual-week spending has an incompletely dated reference window. Ordering an interview after an event is insufficient to certify that all spending refers to the post-event period.
5. Do not automatically label a broad company-change category a plant closure or involuntary displacement. Audit original reason codes, job slots, employment status and temporary-layoff handling before defining exposure.
6. A constructed total-expenditure denominator is not automatically mean spending over the three baseline months required by v91. Inspect component reference periods, benefits treatment and annualization before proposing a different buffer ratio. Standalone cash is not established by the mapped account variable.
7. Family interview identifiers must be linked through the official cross-year individual file, including index-person identity changes, splitoffs and partner continuity. A household is not certified stable because a family label matches.
8. A longitudinal weight alone does not establish survey variance validity. Audit design variables, genealogy/clustering and guidance for the chosen domain before any precision claim.

## Gate record
| Gate | Result |
|---|---|
| Primary dated monthly pre/post food outcome | Not supported by inspected measures; closed |
| Liquid assets provably before event | Conditional; record-level dates and routing untested |
| Exact involuntary-event classification | Incomplete |
| Complete food basket including benefit-funded purchases | Incomplete |
| Monthly baseline total-expenditure denominator | Not established |
| Cross-wave index-person/household linkage | Not audited |
| Design-based variance and achieved precision | Not audited |
| New empirical association or macro calibration | Not authorized by this evidence |

This crosswalk is intentionally partial; it is not an extraction specification. No raw-value decoder is supplied because branching and missingness must first be completed. No participant data are published.

## Access and next research priority
The official PSID data page states that first-time downloads require registration. Public documentation was accessible; no authenticated download or registration was attempted.

Next priority: screen CE official documentation for exact separation timing/reason, dated expenditures and a strictly pre-event resource reference date. Continue without acquiring outcomes. If CE also fails the monthly-design gates, retain the prospective monthly design and prepare its practical recruitment/data-access feasibility assessment. A historical PSID association remains a separately specified alternative, not the confirmatory fallback selected by favorable results.

## Provenance and validation
- 2021 family codebook, release 3 described on p2: https://psidonline.isr.umich.edu/documents/psid/codebook/FAM2021ER_codebook.pdf
- 2023 family codebook, release 1: https://psidonline.isr.umich.edu/documents/psid/codebook/FAM2023ER_codebook.pdf
- Provider download/access guidance: https://simba.isr.umich.edu/data/data.aspx
- Parent design: [DIRECT_CONSUMPTION_DESIGN_V91.md](DIRECT_CONSUMPTION_DESIGN_V91.md)

Each mapped identifier was checked against the indicated official entry. This is a live-document audit, not a checksum-pinned local extraction; retain exact source files and hashes before implementation. No new software was changed, so the existing calculator's three passing tests do not validate PSID measurement or linkage.