מסמך טכני מיומן המחקר · באנגלית, כפי שנכתב בשלב העבודה

MacroVision: external-source SHED panel mechanism check (v89)

Follow-up: the disjoint-window SHED extension (v90) adds the compatible 2023→2024 window, reserves one window per respondent and reports both chronological assignment priorities. It reuses observations and is not an independent confirmation of this earlier result.

Finding and decision

A separately fielded Federal Reserve SHED panel provides a new, proxy-based check of the liquidity-buffer mechanism. The primary reported-core sample contains 1,807 respondents, including 132 reporting their own job loss. The adjusted interaction is +3.83 percentage points, with a model-robust normal 95% interval [-6.87, +14.53] pp. Its half-width is 10.70 pp, exceeding the fixed 5 pp target.

The point estimate is in the hypothesized positive direction: job loss is associated with a larger change in food insufficiency among those without a prior-wave emergency fund. The interval includes zero and a reversed association. This is not clear confirmation of the mechanism. Including valid imputed core responses gives a similar estimate but does not resolve uncertainty.

This is an external data source, not an exact replication of the SIPP estimand, proof of a causal liquidity effect, or demonstrated improvement to a common precision standard. No individual-level non-overlap with SIPP has been established. Neither records nor coefficients are pooled with SIPP. Model calibration remains unchanged.

New-source feasibility and why this source was used

The current Federal Reserve release supplies 2024 and 2025 public CSVs, a persistent respondent identifier and a dedicated pair weight. That permits a prior-wave resource measure and a later-wave job-loss report rather than measuring both resources and hardship at one survey date. It does not guarantee that resources precede every job-loss event: the retrospective job-loss window can overlap the earlier interview near its boundary.

PSID remains excluded under the previously recorded AI-use restriction. Existing SIPP observations are not relabeled as an independent sample. SHED was selected on documented variables and public access before inspecting its new interaction estimates; its different measures make this a complementary mechanism check, not a replacement for a better-powered direct replication.

Fixed sequence

  • Model and source checkpoint: years, resource/outcome proxies, baseline domain, model, uncertainty method and availability gates fixed before linked-sample and outcome contrasts.
  • CSV decoding checkpoint: the official CSV contains value labels; these are mapped to codebook codes, with unknown labels rejected. This corrects ingestion, without changing the sample criteria or model, before new outcome contrasts.
  • Membership and weight audit checkpoint: sample flow, cell sizes and weights recorded before estimation.

Questionnaires, official marginal tabulations and earlier SIPP results were already known. This repository sequence is not external preregistration or a blinded confirmatory trial. The stage-audit reads outcome codes only for validity/reporting checks; it emits no outcome rate or contrast.

Sample and measurement

Respondents are linked one-to-one by shedid. Only a positive **2025 panel_weight** is used; cross-sectional weights are never substituted. The baseline domain is age 25–60, paid/profit work and no own job loss in the preceding year. Paid/profit work can include self-employment.

The resource proxy is **2024 EF1=0, meaning no designated fund sufficient for three months of expenses. The event is 2025 D44_f=1**, the respondent’s own reported layoff/job loss over the preceding 12 months, including temporary layoff. Controls require no such job loss and no voluntary quit (D44_e=0). Exposed respondents may also have quit a different job. A partner’s or other member’s job loss is not fully observed for the exposure definition; “control” does not certify an unshocked household.

The outcome is the change in past-month food insufficiency, I(FD3_2025>=3) - I(FD3_2024>=3), with valid codes 1–4 in both waves. It is not the six-item annual food-security scale. Exact event dates and event duration are not modeled; the one-year retrospective event window may briefly overlap baseline fieldwork. No household-roster stability filter is applied, so food reports can refer to changed households.

The primary domain requires the necessary SHED core-item imputation flags to equal zero. Those core items are baseline work/job loss/emergency fund/food and follow-up job loss/food, plus voluntary-quit reporting for controls. The edited sensitivity accepts valid imputed core responses under the same rules. Core reporting does not certify that demographic profile covariates are unimputed or that response selection is ignorable.

StageRespondent pairs/records
2024 records12,295
2025 records12,934
Exact linked pairs4,420
Linked pairs with positive 2025 panel weight4,419
Baseline age/work/no-prior-job-loss domain1,972
Valid core answers1,972
Job loss or no observed own exit1,841
Valid prespecified baseline covariates1,841
Primary: reported required core responses1,807

One linked pair lacks a positive panel weight and is excluded. The link fraction among annual records is not an attrition/nonresponse rate: recontact invitations and sampling eligibility are not reconstructed here.

Primary groupNo three-month emergency fundHas three-month emergency fund
No own layoff/job loss or voluntary quit6301,045
Own layoff/job loss6666

The two exposed-cell Kish effective sample sizes are 58.34 and 57.54. They meet the predeclared raw N≥30/Kish N≥20 gates. These gates only permit the exploratory fit; they do not establish high power or adequate confounding control.

Model and uncertainty

The model is 2025-pair-weighted least squares with 17 parameters:

change_food_insufficiency ~ intercept + job_loss + low_buffer + job_loss:low_buffer + baseline_food_insufficiency + (baseline_age-40)/10 + baseline_household_size + baseline_female + baseline_income_category_dummies + baseline_education_category_dummies

All adjustments come from the 2024 file. Income is represented by six dummies for the seven categories and education by three dummies for four categories; neither is treated as an interval scale. No follow-up resources, income, employment or household-size controls are used. The interaction’s positive direction was specified before fitting. No specification search, model-selection correction, propensity adjustment or causal identification is claimed.

Prespecified domainNOwn job lossInteraction (pp)Model HC3 normal 95% interval (pp)Half-width (pp)
Primary: reported required core1,807132+3.83[-6.87, +14.53]10.70
Sensitivity: all valid edited core1,841135+3.84[-6.47, +14.15]10.31

The HC3 sandwich treats respondent pairs as independent units in a weighted regression working model. It was checked against an independent statsmodels WLS/HC3 implementation on synthetic and actual analytic data. No replicate-weight, PSU or stratum fields were identified in the inspected CSV headers. These intervals are model-robust normal approximations, not full complex-survey design confidence intervals. They cannot be directly ranked against the cross-cohort SE-envelope intervals in v88 as if both used the same estimand and variance method.

What this says about the thesis

The new point estimate has a different direction from the earlier SIPP estimate, but both uncertainty intervals span zero and the measures, population domain, timing and variance methods differ. This does not establish either a replicated benefit or a statistically demonstrated disagreement between surveys. It leaves the liquidity-protection claim unresolved.

This test addresses respondent-reported resource buffers and food insufficiency around own job loss. It does not test actual consumption spending, net-wealth preservation, debt-service H2, automation-related unemployment or a fiscal mechanism. Resources and exposure are observational; unmeasured differences, job-loss reporting, imputation and panel selection can affect the association. A nominal positive coefficient is not an empirical calibration parameter.

The next priority is a precision and identification decision, rather than further tuning of this coefficient: design a larger harmonized panel with one prespecified window per respondent and a defensible variance method, or obtain a direct-consumption sample with demonstrably prior resource measurement and dated job loss. Existing SHED waves cannot simply be appended as independent people because respondents repeat. Any extension requires codebook harmonization, a respondent-overlap audit and a new fixed analysis plan before outcomes are contrasted. The 5 pp target has not yet been achieved here.

Reproduction and artifacts

Board of Governors of the Federal Reserve System, Survey of Household Economics and Decisionmaking [dataset], https://doi.org/10.17016/datasets.002. The downloaded current public ZIPs and both codebooks are SHA256-bound in the source manifest. Microdata stay outside the checkout and deployment; only aggregate results, audit, metadata and analysis code are published.

```bash python scripts/shed_independent_v89.py --stage audit --data-dir /outside/shed --output /outside/audit.json

Compare with the committed audit before estimation.

python scripts/shed_independent_v89.py --stage estimate --data-dir /outside/shed --audit /outside/audit.json --output /outside/result.json python -m unittest discover -s calibration -p 'test_shed_independent_v89.py' -v

`

Estimation recomputes the audit and requires exact equality of membership, spec, source-manifest and analysis-code hashes. Seven synthetic tests cover one-to-one linkage, panel-weight requirements, label decoding, unknown-label failure, rank deficiency, weight-scale invariance and HC3 agreement with statsmodels.

Official sources

← חזרה ליומן המחקר · לעמוד התזה · קובץ המקור (Markdown)