# MacroVision: empirical household-resilience protocol (v80)

## Historical protocol status (v80)

**Update, v82:** The first public-use SIPP pilot has now been run. See [EMPIRICAL_PILOT_V82.md](EMPIRICAL_PILOT_V82.md) for the actual sample, exploratory estimates and wide uncertainty intervals. H1 remains inconclusive and model calibration stays closed. The no-download/no-estimation statements below describe the original v80 protocol stage, not the current research status.

## Original research decision

This is a protocol prepared before estimation, not an empirical result or an externally registered study. No household microdata have been downloaded, no coefficients have been estimated, and no causal or fiscal calibration is approved.

The next test examines whether resources measured **before involuntary job loss** predict subsequent material hardship, conditional on initial income and net wealth. This tests a household-resilience mechanism. It cannot establish that automation causes 20% global unemployment, lowers prices, or finances adequate transfers.

## Source selection and an important exclusion

**SIPP is selected for a feasibility pilot on material hardship**, subject to a verified year-specific variable crosswalk, panel linkage, timing, weights and public-use conditions. Census describes longitudinal employment, income, assets, liabilities and adult well-being, including food security and housing expenses. Food insecurity and payment difficulties are candidate outcomes, not measures of real consumption expenditure. A SIPP pilot does not replace a full consumption test.

**PSID is excluded from this AI-assisted data workflow.** Its official Conditions of Use explicitly prohibit AI programs and LLMs in conjunction with PSID data and prohibit transferring records to third parties. Only public documentation was consulted. No PSID records or derived extracts were obtained. The initial candidate choice was rejected after the terms review; it must not be used by the generic preflight tool in this workflow.

The full consumption study remains **source not selected**. HRS/CAMS and Understanding Society are documentation-only candidates until access conditions, relevant ages, expenditure coverage, job-loss timing and compatibility have been verified. No registration or agreement has been submitted.

Official source documentation reviewed:

- Census SIPP content: https://www.census.gov/programs-surveys/sipp/about/sipp-content-information.html
- Census SIPP datasets and release-specific documentation: https://www.census.gov/programs-surveys/sipp/data/datasets.html
- Census SIPP 2024 variable metadata (documentation, not records): https://api.census.gov/data/2024/sipp/variables.json
- PSID Conditions of Use: https://simba.isr.umich.edu/U/CondUse.aspx

## Falsifiable hypotheses

**H1, liquidity:** at similar initial income and net wealth, larger pre-loss liquid safe assets predict a smaller increase in material hardship after involuntary job loss. In the later consumption study, the corresponding outcome is a smaller decline in real nondurable consumption.

**H2, debt burden:** conditional on initial liquidity, net wealth and income, higher pre-loss contractual debt service predicts a larger increase in hardship. Debt balances are not interchangeable with payment obligations. If cash debt service cannot be verified, a separately labeled debt-to-income sensitivity analysis may proceed, but H2 as formulated remains untested.

**H3, wealth composition:** debt-free ownership does not ensure preservation of real net wealth when housing or financial asset prices fall. Wealth and hardship are separate outcomes. A hardship result cannot confirm wealth protection.

Opposite-direction effects would challenge the corresponding mechanism. Wide intervals spanning meaningful protection and harm are inconclusive. Practical-effect thresholds and a power calculation must be frozen before inspecting outcome differences, after eligible sample sizes are audited.

## Pilot sample and measures

Proposed working-age population: adults aged 25–60, employed before a verified involuntary separation. Distinguish plant closure from individual layoff. Do not infer involuntary job loss from unemployment alone, or automation exposure from any job loss. Voluntary departures, retirement and disability exits require separate definitions.

Track stable individuals and their household membership. Verify identifiers against the selected release; never merge independent samples merely because household numbers coincide. Require pretreatment assets and outcomes, plus a post-event observation, with actual reference periods aligned. Monthly employment information does not make annually measured hardship a monthly outcome. No invented or interpolated observations.

For food hardship, select the official item or validated scale, response universe, reference period, missing/imputation codes and scoring before extracting records. Do not select a favorable item after testing many outcomes. Housing-payment difficulty is a secondary outcome only if comparable pre/post items exist. Measure liquid safe assets separately from risky holdings and housing; negative net wealth and zero debt are valid values. All predictors must precede the event and outcome window.

## Estimation and validation plan

Start with weighted descriptive trajectories, missingness, attrition and overlap. Compare displaced people with not-yet-displaced people using pretreatment age, education, industry, income, net wealth, liquidity and debt. Do not match on post-loss income or assets; they may be mediators.

A stacked event study with cohort-specific comparisons, individual and calendar effects is a candidate once timing and sample support are verified. Avoid already-displaced controls and an unexamined pooled staggered-treatment coefficient. Account for household dependence and the documented survey design. Plant closures do not automatically remove local shocks, selection or anticipation. Associations do not become causal merely by including controls.

Freeze chronological training and holdout windows only after release compatibility, event counts and measurement periods are audited, and before comparing outcomes. Report pretrends, common support, repeated separations, composition changes, imputation sensitivity and confidence intervals. A US sample does not automatically identify Israeli or global parameters.

## Data gates and tooling

`data/metadata/household_resilience_contract_v80.json` separates the selected hardship pilot from the still-unselected consumption study. Raw variable mappings remain unverified. Metadata inventories are not completed crosswalks.

`scripts/household_resilience_preflight_v80.py` is a **generic structural validator for a future permitted consumption panel**, not a SIPP extractor, hardship scorer, estimator or permission check. It checks duplicate person-wave keys, reference years, finite inputs, zero debt, negative wealth and dated/reasoned involuntary-loss flags. Tests use synthetic fixtures. Passing this tool does not open any research gate or authorize PSID data use.

Next gate: verify the SIPP release-specific employment-event, pretreatment asset and hardship crosswalk; confirm the necessary pre/post reference periods and public-use terms; then define a separate hardship input schema. If these do not align, stop that pilot and select another source. No household records are published or served by the website. Only approved aggregates can enter the site; model calibration stays closed until held-out performance and interpretation are reviewed.