{
 "version":90,
 "stage":"fixed_two_window_SHED_extension_not_independent_confirmation",
 "windows":[[2023,2024],[2024,2025]],
 "screen":"2022 primary CSV lacks FD3 and FD3_iflag; it cannot supply a comparable food baseline and is excluded before any extension contrast.",
 "assignment":"Primary: earliest linked window with positive finite endpoint panel_weight for each shedid, BEFORE age, work, prior job loss, resource/outcome validity or reporting restrictions. Never replace a selected window after later exclusions. A sensitivity uses latest eligible linked window, with the same ordering of assignment before exclusions.",
 "domain":"Same v89 baseline age25-60, paid/profit work, no own prior-year job loss; exposure own subsequent-year D44_f=1; control D44_f=0 and D44_e=0. Any job loss in another household member is not fully observed.",
 "resource_outcome":"Unchanged v89: low buffer is baseline EF1=0; outcome is change in I(FD3>=3). Valid codes and required core iflags follow v89. No household-roster stability condition.",
 "analyses":["earliest_core_reported_primary","earliest_all_valid_edited_sensitivity","latest_core_reported_assignment_sensitivity"],
 "model":"Weighted least squares with the v89 17 baseline terms plus a 2024-2025 window indicator and its interactions with job_loss and low_buffer (20 terms). One common job_loss*low_buffer coefficient; other adjustment slopes shared. Period intercepts and main exposure/resource slopes may differ. This does not establish homogeneous moderation across periods.",
 "weights":"Use endpoint panel_weight for each pair; after each analysis-domain selection, normalize weights within each of the two windows to sum to 0.5. This defines an equal-period weighted regression loss, not an equal mean of separately fitted interaction coefficients or a nationally calibrated combined longitudinal weight.",
 "variance":"Reuse verified weighted-WLS HC3 with one pair per respondent, assuming independent respondent pairs. No full survey-design variance, cross-period PSU dependence or re-estimated sampling weights. Normal approximation only.",
 "availability":{"pooled_each_raw_cell_n":30,"pooled_each_cell_kish_n":20,"within_window_each_raw_cell_n":5,"within_window_each_cell_kish_n":5,"both_windows_required":true,"full_rank_and_finite_hc3":true},
 "gate_scope":"Exploratory pooled common-interaction model gates, not a statement of adequate power. No thresholds are changed after counts or effects are viewed.",
 "precision_target_halfwidth_pp":5,
 "known_results":"v89 latest-window result and earlier SIPP results are already known. This extension reuses respondents and measurements; it is not a fresh holdout, independent confirmation, or a replication of the exact SIPP estimand.",
 "selection_limit":"Earliest assignment reserves prior participants even if they later fail domain rules; latest sensitivity can select on future recontact. Original pair weights do not correct this additional allocation/complete-response selection. Do not claim national household prevalence or causal protection.",
 "temporal_limit":"Retrospective job-loss windows may overlap baseline fieldwork at their boundary; exact event date is unavailable. No causal timing proof.",
 "sequence":"Commit specification, code and harmonization metadata; commit membership/overlap audit before estimates. Schema and public marginal tables were inspected before this plan; no externally registered preregistration.",
 "model_calibration":false,
 "causal":false
}