SDID time weights: study inventory

2026-10-19

Under development — computational feasibility inventory.

Shared spreadsheet: readiness summary and paper-level inventory.

Findings

The collection contains 262 distinct papers: Peter Hull’s original 22-paper collection and 240 extension papers. Existing analysis inputs permit direct inspection of 281 specifications across 65 papers, yielding 1,728 cohort–outcome blocks. This inventory reads the actual panels, including 86 reconstructed from existing preparation code with sample counts checked against archived receipts. The remaining 195 specifications use saved matrix inputs.

Thirteen papers have at least one balanced block with two or more controls, two or more pre-treatment periods, a post-treatment period, and no outstanding controls/fixed-effects adapter requirement. These are computationally usable inputs for a time-weight investigation; successful new optimization and specification review remain separate checks.

Seven papers offer at least five pre-treatment periods. Six are observational candidates in the current audit. The seventh is the randomized government-audit study preserved in the original collection and can be analyzed as a separate experimental stratum. Six of the seven have a block with at least ten pre-periods. Five and ten periods are planning thresholds, not SDID identification requirements.

Longer-history candidate Pre-periods Post-periods Donors Existing SDID experience
Medical implants / product liability 5 6 2,462 Archived fit
Rebel on the Canal 176 86 455 Archived fit
Reconciliation Narratives / Birth of a Nation 25–65 1–43 494–560 Previous solver failed
The Gift of a Lifetime / hospitals 5–18 3–16 49–52 Archived fits
Government audits and corruption (randomized) 5–12 1–8 3,523 Archived fits
Copyrights and Creativity / Italian opera 31–58 12–39 4 Some fits succeeded; others failed
Credit Markets, Property Rights, and the Commons 7–50 3–46 194 Archived fit

Ranges refer to qualifying blocks, with at least five pre-periods, within the existing prepared specifications. Counts are observed periods, not uniformly years: retain original calendar labels and frequency before comparing distances. Staggered studies are assessed cohort by cohort, using the existing control pool.

The six shorter-history cases are Curriculum and Ideology (2–3 pre-periods), school construction in Indonesia (2–3), collectivization/draft-animal deaths (3), ARRA road spending and voting (3), cap-and-trade (4), and Religion, Division of Labor, and Conflict (2). The school-construction and religion papers retain pending design/specification labels; school construction’s earlier fit timed out.

Full inventory

Categories are mutually exclusive at the paper level, assigned by the most usable inspected target. They do not classify every analysis in a paper.

Readiness category Papers
Balanced inputs with at least 5 pre-periods 7
Balanced inputs with only 2–4 pre-periods 6
Balanced subset or earlier-cohort/not-yet-treated window possible 14
Current controls, fixed-effects or weighting adapter needs work 21
Inspected targets have only one pre-period or one donor 14
Other missingness/window/preparation work needed 3
Contributed absorbing-treatment candidates without inspected panels here 48
Other contributed binary temporal contrasts needing treatment-path review 26
No prepared target inspected and no such contributed candidate flag 123
Total 262

The 14 sample/window candidates have an explicit mechanical route to a nontrivial balanced block: either retaining complete units over the existing window, or using later cohorts as controls strictly before their adoption. The exports give donor/treated retention and alternative post horizons. These changes were not applied. They can change the estimand and selection of units, so they need a declared sampling/window rule before fitting.

The 21 adapter cases are limitations of the current integration, rather than general claims about SDID. The authors’ package accepts covariates. Matching the original specification, aggregation and weights still requires work. The 48 additional absorbing candidates come from the contributed Deep DiD audit; their panel completeness and pre-period lengths remain unmeasured here. The final 123 are also not whole-paper exclusions. Earlier inventory notes and proposed authored targets remain visible in the paper-level export.

A single pre-period forces its time weight to one. With one donor and a free time-weight intercept, the balance residual can be absorbed by the intercept; regularization and algorithmic conventions then determine the weights. These cases provide little evidence about data-driven temporal concentration. For the no-covariate time-weight problem, only donor outcomes are required; checking complete donors alone yields the same 13-paper set in these inputs.

Weight availability and estimation protocol

There are 142 archived SDID estimates across 26 papers. The saved summaries and effect matrices do not retain the pre-treatment time-weight vectors. None of these inventory counts describes an observed weight distribution. No models, standard errors or bootstrap runs were recomputed for this inventory.

The archived backend is crabbymetrics.SyntheticDID. Its cohort-specific implementation estimates time weights from donor pre-period outcomes to match donor post-period means, with a free intercept. The existing report writer discards the time weights after using them to adjust pre-treatment gaps.

For a study of non-zero weights, use a pinned version of the authors’ synthdid package, save the actual weights$lambda, and record penalty, convergence settings and sparsification. The inspected upstream version is 70c1ce3: its default time penalty multiplier is eta.lambda = 1e-6, it uses a time-weight intercept, and its default optimization includes a sparsification step. The archived backend’s support pattern should not be substituted for the authors’ implementation without a direct cross-check.

Recommended first pass:

  1. Use the five observational longer-history papers with at least one existing SDID fit (medical implants, canal, hospitals, opera and credit markets). Add Birth of a Nation after resolving optimizer behavior. Keep randomized audits as a separately reported stratum. Existing estimates are retained.
  2. Fit each cohort as a separate block. Save time weights, calendar dates, outcome, treatment definition, donor identifiers, pre/post windows, package commit and optimization diagnostics. Keep the outcome/window selection rule explicit and avoid treating many outcomes from one paper as independent studies.
  3. Report exact support and support above 1e-8, 1e-6 and 1e-4, together with effective pre-periods (1 / sum(lambda^2)), mass on the last 20% of pre-periods, weighted lag from adoption, and the shortest interval containing 90% of mass. For clustering, add the number of contiguous support runs and gaps between active periods. Normalize location/span by pre-period length, while retaining frequency-specific calendar summaries.
  4. Compare default sparsification with unsparsified optimization and tighter convergence tolerances. Vary the post horizon: it changes the donor target being matched and can change time weights even with the same pre-period data.
  5. Summarize by paper and show cohort/outcome variation within paper. The 247 qualifying long-history blocks are heavily concentrated in a few studies; they are neither 247 independent studies nor necessarily 247 distinct time-weight problems. This available-input sample is not representative of the entire DiD literature.

Files and reproducibility

Run .venv/bin/python scripts/inventory_sdid_time_weights.py from the repository with the existing local analysis inputs available. The script is read-only with respect to models and inputs; only these inventory exports are written. It checks binary absorbing prepared treatment, cached row/unit counts and recovered sample counts. Source paths and SHA256 hashes are recorded at specification and cohort level. Temporary reconstruction folders are removed automatically. Input availability is required; exported source paths are provenance rather than a redistribution of restricted microdata.

All original 22-paper membership, public pages, fitted estimates, calibration plots and eligibility decisions remain unchanged. These inventory artifacts belong to the community’s Under development work.