| paper | attempted | returned | zero_noise | min_pre | max_pre | min_donors | max_donors | |
|---|---|---|---|---|---|---|---|---|
| aeraej-105 | Right to counsel | 2 | 2 | 0 | 147 | 156 | 2 | 4 |
| aeraej-120 | Women’s political voice | 8 | 8 | 0 | 2 | 16 | 2 | 16 |
| aeraej-159 | Village officials | 5 | 5 | 0 | 4 | 11 | 28 | 67 |
| aeraej-388 | Market integration | 7 | 7 | 0 | 4 | 4 | 34 | 34 |
| aeraej-59 | Procurement auctions | 1 | 1 | 0 | 3 | 3 | 12 | 12 |
| aeraej-6 | Retail pharmacies | 136 | 136 | 0 | 6 | 83 | 36855 | 52563 |
| aeraej-97 | E-governance rollout | 1 | 1 | 0 | 4 | 4 | 277 | 277 |
| study-02 | Powerful politicians | 23 | 23 | 0 | 16 | 25 | 6 | 201 |
| study-03 | Powerful politicians: reply | 36 | 36 | 0 | 18 | 29 | 249 | 725 |
| study-09 | Vertical integration | 6 | 6 | 0 | 2 | 13 | 239 | 251 |
| study-13 | Ideology and public organizations | 3 | 1 | 2 | 16 | 16 | 39870 | 39870 |
| study-17 | Trading nontradables | 3 | 3 | 0 | 3 | 10 | 6 | 10 |
| study-18 | Fiscal capacity (mixed-design) | 140 | 84 | 56 | 121 | 323 | 88 | 190 |
| study-22 | Nigerian oil | 47 | 47 | 0 | 3 | 10 | 155 | 246 |
SDID time-weight patterns: second tranche
14 additional papers · changed samples and windows · Under development
Coverage
This tranche examines the 14 papers previously identified as possible balanced subsets or earlier-cohort windows. The source outcome and treatment definitions are retained. All 418 candidate cohort–outcome blocks were attempted: 360 return time weights across all 14 papers, and 58 have zero default noise scales (56 Fiscal Capacity blocks and two Ideology and Public Organizations blocks). No substitute weights are assigned to those failures.
Together with the preceding 13-paper tranche, this gives returned examples from 27 distinct papers. It does not imply 27 source replications or 754 independent samples: outcomes and cohorts often share donors, and the new samples/horizons are explicitly changed.
The computation targets the separable no-covariate SDID time-weight problem. It retains the authors’ default penalty, intercept, sparsification and stopping rule. An independently checked optimized reference uses the same penalty. No new unit weights, treatment-effect estimates or bootstrap results are produced.
Sample and window changes
The frozen rule first keeps all complete units across the full prepared window when at least two donors and one treated unit remain. Otherwise the window ends immediately before the next cohort; later adopters can be donors while untreated, and complete units are selected within that window. The start of the pre-period is unchanged. This yields 315 full-window and 103 earlier-window blocks. No selection depends on the values or fitted quality of the outcomes, but complete-case selection can still change the population and induce selection bias.
All retained treatment paths were checked against the prepared source indicators, including observations removed by the earlier reversal-handling adapter. Donors remain untreated in the selected window; retained treated units do not reverse within it. Sampling weights, additional covariates and nonstandard fixed-effect adapters are outside this tranche’s selected specifications.
Retention is an essential part of these examples. Some corporate-downsizing windows retain less than 1% of eligible donor or treated units. They are useful for inspecting a well-defined time-weight calculation, but their behavior should not be generalized to the original study’s full sample. Several long-history examples also have only two or four donors, leaving many pre-period combinations with similar fit. The interactive report displays these dimensions and retention for every block.
| paper | full_window_blocks | earlier_window_blocks | min_control_retention | min_treated_retention | design_status | |
|---|---|---|---|---|---|---|
| aeraej-105 | Right to counsel | 0 | 2 | 1.0000 | 1.0000 | included |
| aeraej-120 | Women’s political voice | 0 | 8 | 1.0000 | 1.0000 | included |
| aeraej-159 | Village officials | 4 | 1 | 0.1677 | 0.1818 | included |
| aeraej-388 | Market integration | 7 | 0 | 0.9189 | 0.3958 | included |
| aeraej-59 | Procurement auctions | 1 | 0 | 0.8571 | 1.0000 | included |
| aeraej-6 | Retail pharmacies | 128 | 8 | 0.4463 | 0.2500 | included |
| aeraej-97 | E-governance rollout | 0 | 1 | 1.0000 | 1.0000 | pending |
| study-02 | Powerful politicians | 0 | 23 | 0.0007 | 0.0053 | included |
| study-03 | Powerful politicians: reply | 0 | 36 | 0.0183 | 0.0711 | included |
| study-09 | Vertical integration | 5 | 1 | 0.3092 | 0.5000 | included |
| study-13 | Ideology and public organizations | 0 | 3 | 0.0125 | 0.2007 | included |
| study-17 | Trading nontradables | 3 | 0 | 0.1250 | 0.2500 | included |
| study-18 | Fiscal capacity (mixed-design) | 120 | 20 | 0.1497 | 0.5000 | excluded |
| study-22 | Nigerian oil | 47 | 0 | 0.9044 | 0.5000 | included |
Numerical sensitivity
All 360 optimized references meet the normalized objective-gap target of 10^-8. Only 36 defaults meet that stricter check; 43 default time-weight runs reach their iteration limit. This is a numerical comparison at the same statistical penalty, rather than tuning selection.
Across the returned blocks, median L1 weight change is 9.94^{-4}; 95 change support above 10^-6, and 12 change mass in the final fifth of pre-periods by more than one percentage point. These block-level counts are dominated by papers with many outcomes or cohorts; the 14-paper profiles are provided to avoid treating them as independent evidence.
Six QP blocks were sensitive to initialization. Those six and all blocks with at most 20 donors—20 distinct blocks—were refined at 80-digit precision. Both initializations agree within 10^-45 in L1, and high-precision objective gaps are below 10^-50. The remaining references retain double precision; these checks do not universally certify the accuracy or statistical stability of weights under such weak regularization.
For large panels, an orthogonal QR reduction preserves the author time-weight objective without subsampling donors. All 248 reduced blocks pass a direct 20-step check; eight preselected examples also pass a full direct-author check. The largest full-fit maximum weight difference is below 2 × 10^-14.
One preselected example per paper
Examples were frozen before estimation: first eligible specification, median cohort, preferring at least five pre-periods. Support uses weight >10^-6. The underlying exports retain all zeros and small entries. The x-axis counts observed periods before adoption; source-period labels are supplied separately. Long histories with few donors and low-retention samples remain visibly identified.
Downloads
- Interactive explorer
- 14-paper profile PDF
- Paper summary / sample manifest
- Default and double-QP vectors / optimized reference vectors
- Paired comparisons / unavailable fits
- Base analysis / precision refinement
- Independent checks
The original 22-paper landing page, previously delivered results, classification logs and Pedro’s contributed audit remain unchanged. This tranche is an exploratory extension under development, and does not admit or exclude whole papers based on whether these time-weight calculations run.