# MacroVision: first SIPP household-resilience pilot (v82)

## Follow-up (v83)

The [2022-2024 replication and temporal extension](EMPIRICAL_REPLICATION_V83.md) are now complete. The person-disjoint cohort is below its prespecified event-count threshold; the separate temporal extension changes the point-estimate direction and remains imprecise. H1 is not established and the liquidity mechanism remains uncalibrated from these estimates. The results below remain the original v82 pilot, not the combined evidence.

## Later measurement-quality review (v84)

The [v84 audit](SIPP_HISTORICAL_PLAN_V84.md) identified an official Census note on food-security hot-deck pathing/flag errors affecting 2018-2022 releases, including this pilot's 2022 baseline release. A flag0 conditional item cannot alone certify an un-imputed structural skip. The table's flag-based sensitivity is therefore not a certified imputation-free sample without an answer-conditioned universe audit. Numerical results below are unchanged; they remain inconclusive and are not calibration inputs.

Official note: https://www.census.gov/programs-surveys/sipp/tech-documentation/user-notes/2020-usernotes/2020-hot-deck-alloc-error.html

## Research conclusion

**The first empirical pilot is inconclusive. It does not establish H1 and does not authorize calibration of the full model.** Lower pre-event deposits are associated with a larger unadjusted rise in food insecurity following an economic job separation. After baseline adjustment, the estimated additional increase is 2.97 percentage points, with a survey-design 95% interval of -15.92 to +21.87 points. The data are compatible with protection, no difference, or a reversed association.

This is an exploratory analysis of official US public-use microdata, not a causal study, external preregistration, consumption study, or test of automation-induced unemployment. H2 concerns contractual debt service and remains untested; controlling for a debt **balance** does not test H2. H3 and the fiscal and price channels remain untested.

## Periods, population and definitions

- Baseline: reference year **2021**, in the 2022 SIPP release. Deposits and wealth are measured at year-end, before the event year.
- Event: reference year **2022**, in the 2023 release. An employer job ends with reason code 1, 2, 3 or 6: closure/relocation, slack business conditions, abolished position/shift, or other involuntary reason. December endings require the job-continuation flag to be zero. All seven job slots are checked; repeated monthly copies do not create extra people or events.
- Outcome: annual **2023** food insecurity, in the 2024 release. Official `RFOODS=2 or 3` is the binary outcome; `RFOODS=1` is the comparison category. We use the official recode, not an improvised score. The outcome is neither monthly nor consumption expenditure.
- Population: people aged 25-60 working all of December 2021 (`RMESR=1 or 2`). This includes people with a job or business, so differences in baseline work arrangements remain a limitation. Earlier economic separations and unknown reasons for observed employer endings are excluded. Controls have no observed employer separation of any reason in 2022. Economic separation does not necessarily mean subsequent unemployment, permanent displacement, or automation exposure.
- Annual hardship is household information attached to people. We require all 12 months of each outcome-measurement year to be at the interview residence. This reduces residence-assignment ambiguity and narrows generalizability to stable residences; it does not rule out household membership changes between years.
- Deposits are six individually attributed checking/savings shares, summed once per member of the December baseline household (`SSUID+ERESIDENCEID`). Spousal joint shares are already divided between spouses; they are not doubled. Repeated household wealth/debt values are **not** summed over members. Structural out-of-universe amounts may be zero; unexplained missing owned balances remain missing.
- Lower deposits means less than one month of mean **baseline income**, not expenditure. Annual income requires 12 nonmissing monthly household-income observations and a positive annual total. The measure excludes money-market assets, CDs, housing and risky investments; it is not total liquid safe wealth.
- All tables describe **people's household hardship**, weighted by 2024 `FINYR3` for reference years 2021-2023. Multiple eligible people can share a household. These are not counts or population estimates of distinct households.

## Sample flow

| Stage | People |
|---|---:|
| December baseline employed, age 25-60 | 13,105 |
| Exclude prior economic separation/unknown ending reasons | 12,883 |
| Linked in all three release years | 3,845 |
| Positive longitudinal weight, full event year, known ending reasons | 3,836 |
| Economic separation or no employer separation | 3,328 |
| Stable baseline and outcome residence | 2,781 |
| Complete descriptive inputs | **2,766** |
| Economic-separation subgroup | **83** |
| Closure **or relocation** subgroup within those 83 | **15** |

The large linkage reduction includes panels that cannot span the entire window, not just nonresponse. It must not be labeled a pure attrition rate. Fifteen closure/relocation cases do not support a separate public effect estimate under the pilot's minimum-cell rule (30).

## Descriptive results

Percentages are weighted with `FINYR3`; sample sizes are unweighted. Changes are percentage points. These small-cell estimates are unstable; no precision is implied by the two-decimal presentation.

| 2022 status | Baseline deposits | n | 2021 insecurity | 2023 insecurity | Change |
|---|---|---:|---:|---:|---:|
| No employer separation | Below one month of income | 1,156 | 13.75% | 14.21% | +0.46 pp |
| No employer separation | At least one month | 1,527 | 2.79% | 3.90% | +1.11 pp |
| Economic separation | Below one month | 38 | 16.26% | 32.12% | +15.86 pp |
| Economic separation | At least one month | 45 | 13.48% | 21.00% | +7.52 pp |

The unadjusted difference between the two separated groups' changes is 8.34 points. Accounting additionally for the corresponding no-separation changes gives a descriptive contrast of 8.99 points. Neither is a causal effect; the groups differ before separation.

## Exploratory adjustment and uncertainty

We fit a person-weighted linear model of the 2023-minus-2021 hardship indicator. Regressors are separation, lower deposits, their interaction, log baseline annual household income, inverse-hyperbolic-sine baseline net wealth and debt balance, age, age squared, and baseline hardship. All resources and controls precede the event year. No post-separation income or wealth controls are used.

The interaction estimates the excess hardship change associated with lower deposits among separated people relative to the no-separation contrast, conditional on these covariates. It is not the causal effect of giving a household cash. Education, industry, occupation, health, anticipation, local shocks and other confounding remain. This simple model also does not replace the proposed event-study validation.

We use the matching **240 longitudinal replicate weights**, check their reference period and that `REPWGT0` matches `FINYR3`, and re-estimate the model for every replicate. Fay BRR uses factor 0.5: variance = sum of squared replicate deviations / (240 x 0.5²). Reported intervals use the normal 1.96 multiplier and remain exploratory. Imputation uncertainty is not fully captured by these survey-design intervals.

| Analysis | n | Economic separations | Interaction | Design SE | 95% interval |
|---|---:|---:|---:|---:|---:|
| All edited observations | 2,766 | 83 | +2.97 pp | 9.64 pp | -15.92 to +21.87 pp |
| Flag-based food sensitivity (see v84 quality review) | 2,569 | 75 | +9.81 pp | 9.88 pp | -9.54 to +29.17 pp |

For the sensitivity, items 1-3 must be reported (`AFOOD1..3=1`); conditional items 4-6 may be reported or outside their universe (`0 or 1`). Requiring six reported answers would wrongly select on food hardship. A recode flag of 9 alone is not sufficient evidence of reported components.

Census documents food-security allocation changes affecting comparisons with earlier releases. The sensitivity is therefore necessary, but it changes the sample and does not eliminate selection or all measurement problems. Both intervals span zero and substantial effects in either direction. **We neither confirm nor reject H1 from this pilot.**

## Specification history and reproducibility

The 2021/2022/2023 timing, economic-separation codes, official food-security outcome and one-month-income deposit threshold were selected before viewing the first outcome rates. During implementation we corrected weight-column capitalization, preserved identifiers as strings, restricted residence assignment, and respected conditional food-question universes. The adjusted model was specified **after** the first descriptive rates were viewed; it is explicitly exploratory. There was no external registration or untouched confirmatory sample in this pilot.

Files:

- `data/metadata/sipp_pilot_spec_v82.json`: design and specification history.
- `data/metadata/sipp_crosswalk_v82.json`: used variable labels/value codes and official metadata/dictionary hashes for each release. It verifies the pilot's selected fields, not every future SIPP variable.
- `scripts/sipp_pilot_v82.py`: streaming extraction, timing checks, household deposit construction, sample audit, aggregate tables and replicate-weight fits.
- `calibration/test_sipp_pilot_v82.py`: synthetic regression tests for missing values, December continuations, food skips, small cells and weighted/Fay calculations.
- `data/processed/sipp_pilot_v82.json`: aggregate results with source archive and specification SHA-256 hashes. No person identifiers or microdata are committed or served.

Download the following official public-use archives into a directory outside the repository:

1. `https://www2.census.gov/programs-surveys/sipp/data/datasets/2022/pu2022_csv.zip`
2. `https://www2.census.gov/programs-surveys/sipp/data/datasets/2023/pu2023_csv.zip`
3. `https://www2.census.gov/programs-surveys/sipp/data/datasets/2024/pu2024_csv.zip`
4. `https://www2.census.gov/programs-surveys/sipp/data/datasets/2024/lgtwgt2024yr3_csv.zip`
5. `https://www2.census.gov/programs-surveys/sipp/data/datasets/2024/lgtrw2024yr3_csv.zip`

Install `requirements-research.txt`, then run:

```bash
python scripts/sipp_pilot_v82.py --data-dir /path/outside-checkout --output /path/outside-checkout/sipp_pilot_v82.json
python -m unittest discover -s calibration -p 'test_sipp_pilot_v82.py' -v
```

The reader streams the large official ZIP members and caches selected columns beside the archives, bound to archive hashes and extraction columns. Keep this directory outside the deployed checkout. Only aggregates enter MacroVision. PSID remains excluded; no PSID data were accessed. The web model's numbers, formulas and calibration parameters are unchanged.

## Next empirical gate

1. Before opening the next outcome data, specify a replication window: baseline 2022, separation 2023, outcome 2024, initially restricted to the **2023-origin SIPP panel**. This offers a person-disjoint cohort relative to the current linked pilot, subject to verifying release compatibility and the 2025 longitudinal weights. This is an analysis holdout, not a future-data forecast.
2. Audit eligible event counts without comparing hardship outcomes, freeze a power/precision plan, and add education, industry, baseline employment arrangement and household composition to the adjustment. Do not broaden definitions to obtain a favorable coefficient.
3. Verify direct contractual debt-service variables before attempting H2. Debt balances remain separate.
4. Report replication, overlap, composition, imputation and timing sensitivities, including null and opposite-direction findings. Only after adequate precision and external validity review consider bounded model calibration. US job separations do not identify a global automation shock or Israeli fiscal parameters.

## Official documentation

- [2024 release, public-use files and longitudinal weights](https://www.census.gov/programs-surveys/sipp/data/datasets/2024-data/2024.html)
- [2024 user's guide, sections 3.4-3.5, 4.3.1 and 7.2-7.3](https://www2.census.gov/programs-surveys/sipp/tech-documentation/methodology/2024_SIPP_Users_Guide.pdf)
- [2024 dictionary](https://www2.census.gov/programs-surveys/sipp/tech-documentation/data-dictionaries/2024/2024_SIPP_Data_Dictionary.pdf)
- [2023 dictionary](https://www2.census.gov/programs-surveys/sipp/tech-documentation/data-dictionaries/2023/2023_SIPP_Data_Dictionary.pdf)
- [2022 dictionary](https://www2.census.gov/programs-surveys/sipp/tech-documentation/data-dictionaries/2022/2022_SIPP_Data_Dictionary.pdf)
- [2024 longitudinal weight dictionary](https://www2.census.gov/programs-surveys/sipp/tech-documentation/data-dictionaries/2024/lgtwgt2024_dictionary.txt)
- [Food-security and well-being allocation issues](https://www.census.gov/programs-surveys/sipp/tech-documentation/user-notes/2024-usernotes/2024-food-security-well-being-matrix-issues.html)