# MacroVision v91: direct-consumption feasibility and prospective design
Date: 29 October 2026. Status: documentation-based source screen and hypothetical sample planning; no new empirical estimate, extraction, calibration or deployment.

## Decision
Prioritize a PSID variable-and-timing feasibility audit. Retain CE Interview as a secondary source. Neither is certified analysis-ready. A separate prospective monthly panel is the fallback if historical reference periods or baseline liquid assets cannot support temporal ordering. This is a repository planning checkpoint, not external preregistration.

V90 has exhausted the two harmonized SHED windows; do not alter its model, select favorable allocation, pool its coefficient with SIPP/PSID, or claim independent confirmation. Its food-insufficiency interaction and the expenditure interaction below have different units and estimands. V90 remains inconclusive.

## Documentation-based source screen
| Candidate | Verified documentation | Unresolved gate / decision |
|---|---|---|
| PSID | Official consumption inventory documents food at home, delivery and eating out, with wave-specific coverage; official FAQ documents monthly job histories; main study uses biennial interviews since 1997. | Priority for audit, not estimation. Need exact released wave files, job-end reasons, liquid-asset reference dates, expenditure reference periods, family continuity and longitudinal design/weight variables. A usual-week report is not a dated monthly spending diary. |
| CE Interview | BLS guide documents four quarterly interviews, previous-three-month reporting, monthly MTBI expenses and annual assets/income; questionnaire includes LIQUDYRX one-year-ago balances. | Secondary. Prove baseline balance date predates separation; do not treat retrospectively reported balances as prospectively measured. Check household replacement, repeated release duplicates, job-loss reason/month and suitable longitudinal weights. |
| Prospective monthly panel | Design proposal only; no participant recruitment or data provider selected. | Could supply dated spending and separation, but cost, access, consent, coverage and achieved event counts are unknown. |

The PSID consumption inventory's coverage through 2025 is not evidence that every required 2025 microdata file is available. Select no wave set based on outcome estimates. Freeze a source release list and variable crosswalk before membership counts; record unresolved fields as unavailable, not inferred.

## Target and prospective measurement specification
Question: is involuntary job loss associated with a larger fall in directly measured real food expenditure for households with low pre-event liquid buffers?

Primary unit: household with one index paid employee aged 25–60, selected at enrollment as the highest baseline labor-income employee (stable person-ID tie break). Track partner employment; simultaneous partner job loss is a prespecified secondary stratum, not a hidden exclusion. Historical data must support the same scope or require a separately dated amendment before outcomes. Restrict interpretation to the supported household members.

Exposure: first documented involuntary separation of the index employee after baseline, including layoff or employer closure; distinguish quits, retirement, health exits and temporary layoff. Use only verified reason codes. Controls have no separation in the matched window and were employed at its start. Never select controls by future continuous employment after the outcome window. Assign risk-set controls the same calendar index month; one analysis window per household, allocated chronologically before outcome validity.

Prospective baseline: monthly observations -3,-2,-1; exposure month 0; follow-up +1,+2,+3. Freeze assets at end of month -1, and count anticipation as an unresolved identification threat. Historical PSID/CE cannot inherit this timing label unless actual reference periods support it. A biennial change is a distinct longer-horizon design requiring a separate checkpoint, not a three-month proxy.

Primary spending basket: food at home, delivered food and eating out, with no duplicate components; include benefit-funded purchases if measured consistently in both periods. This is food expenditure, not total consumption, calories or welfare. Primary outcome: log(mean real monthly post food spending / mean real monthly pre food spending), per baseline square-root household size in both periods. Positive spending in both periods is required for this log estimand; report zero-spending exclusions, and report a secondary real level change including zeros. Never replace zeros by an arbitrary constant. Fixed baseline equivalence size cancels in the log ratio and must not be portrayed as a demographic adjustment.

Use a frozen official monthly food price index and reference month; specify series IDs and treatment of home/away components before estimation. Low buffer: baseline checking+savings+cash / mean baseline monthly total expenditure <1; document whether CDs are accessible and whether total expenditure exists. If denominator or dates unavailable, this classification fails; do not substitute SHED's emergency-fund question or sample median. Secondary continuous buffer and baseline debt-service ratio require their own documented fields.

## Estimator and interpretation
Primary weighted household change regression:
change_log_food = intercept + loss + low_buffer + loss*low_buffer + baseline log real food + baseline age + baseline household size + baseline partner-work indicator + baseline income-category and education-category indicators + calendar-index-month effects.

Negative interaction is the hypothesized larger expenditure fall in low-buffer households. No post-loss income, employment, assets or household size enters adjustment. Exact categories, coding, releases and variance inputs must be frozen in the variable audit before estimation. Historical source failure does not authorize silently dropping an unavailable covariate.

Use source-approved longitudinal weights and design-based variance with published strata/PSUs or replicate procedures. Household clustering alone does not account for PSID genealogical/design dependence. If a defensible survey variance cannot be constructed, label model-only variance and keep the confirmatory gate closed. Prospective sampling must define clusters, recruitment probabilities and follow-up adjustment in advance.

Observational job loss and asset ownership remain confounded. Temporal ordering and pre-event trends are diagnostics, not causal identification. Do not interpret this interaction as the effect of granting cash, automation, a government benefit, or global 20% unemployment. US results cannot directly calibrate Israel or a global model.

## Precision and power fixed before new outcome estimates
Planning target: normal two-sided 95% interval half-width <=0.05 log units; 80% power at two-sided alpha .05 for a 0.05-log-unit interaction. These are NEW expenditure targets, not V90's five-percentage-point binary target; 0.05 log units is about 5.13% multiplicatively.

Calculator assumes independent four cells, equal within-cell outcome-change SD, exposure/buffer independence for allocation, no precision gain from covariates, design effect 1.5 and 75% complete follow-up. These are hypothetical sensitivity assumptions, not data-derived estimates. SD grid .20/.35/.50; exposure share .05/.10; low-buffer share .30/.50. Correlation between exposure and buffer can sharply change required N.

For cell shares s_j and outcome-change SD sigma:
Var(interaction) = design_effect * sigma^2 * sum(1/s_j) / N_complete.
N_precision = ceil(variance_factor * (z_.975/.05)^2).
N_power = ceil(variance_factor * ((z_.975+z_.80)/.05)^2).
Use the larger N and divide by retention for recruitment. Normal approximations do not replace a final survey-design simulation. Final power must use audited joint cell shares, plausible variance ranges, design degrees of freedom and missingness scenarios; fix these before outcome contrasts. Report every planned scenario, not just the cheapest.

The executable calculator and full 12-scenario output accompany this plan. Three synthetic unit tests verify a balanced analytic reference, design/attrition scaling and invalid-input rejection. No empirical power or achieved precision is claimed.

## Ordered gates and stop rules
1. Obtain official codebooks/release metadata and freeze exact variable IDs, reference dates, missing/imputed codes, asset units, price series, reason codes, IDs and weights. Access via provider terms; no account creation or participant data publication is authorized by this plan.
2. Audit longitudinal family/person links, duplicates, mixed households, event ordering, missingness and four joint cell counts without outputting spending contrasts. Preserve one household window; do not drop genuine household changes to improve results. Report inability to follow split households.
3. Freeze a feasibility report and design-based precision simulation; require full rank and no empty interaction cells, and judge target precision independently of computational availability. If power inadequate, report infeasibility rather than tune model/window.
4. Commit the final source-specific specification before estimates. Report primary and all prespecified sensitivities, intervals, attrition and zero-spending exclusions regardless of sign. Any altered outcome/window is explicitly exploratory.
5. Keep calibration closed unless measurement, precision, transportability and identification are separately justified. This mechanism check alone cannot validate the complete macro thesis.

Next concrete action: PSID codebook crosswalk and date-alignment audit. This step completed planning only.

## Reproduction
Run `python scripts/consumption_design_v91.py` for the full planning grid.
Run `python -m unittest discover -s scripts -p test_consumption_design_v91.py -v`.
Only standard-library dependencies; no participant data, network requests or model changes.

## Official sources inspected
- PSID consumption inventory: https://psidonline.isr.umich.edu/data/sl/Consumption-PSID.pdf
- PSID study cadence/content: https://psidonline.isr.umich.edu/Studies.aspx
- PSID work-history scope: https://psidonline.isr.umich.edu/Guide/FAQ.aspx?Type=7
- PSID release/codebook catalog: https://psidonline.isr.umich.edu/Guide/documents.aspx
- BLS CE public microdata guide: https://www.bls.gov/cex/pumd-getting-started-guide.htm
- BLS 2024–2025 questionnaire (LIQUDYRX): https://www.bls.gov/cex/research_papers/pdf/2024-2025-ceqquestionnaire.pdf