# F4-conf-01 — Treatment effects and uncertainty extracted from two published tables

## Question
Extract, exactly as printed, specified treatment-effect estimates, standard errors (SEs), confidence-interval (CI) bounds and counts from (i) Armendariz (2026), arXiv:2609.27732v1, a staggered DiD study of drug consumption room (DCR) openings in Switzerland (Tables 6 and 7), and (ii) Li, Pearson-Merkowitz & Yokum (2026), arXiv:2606.14621v1, an RCT of Rhode Island's RESEA programme (Table 2). Attach a locator to each value, derive a 95% CI from a printed SE and an SE from a printed CI, and flag the requested quantity that the source does not print. Nothing is re-estimated.

## Data and exclusions
Inputs: the two source extracts and `data/README.md`, all SHA-256-verified against the README/bundle (`out/state/data_audit.md`). Line counts: `n_input_rows` (README.md 10, arxiv-2609.27732v1.txt 134, arxiv-2606.14621v1.txt 58). I assessed 25 requested items (`n_analysed`) and excluded none (`out/exclusions.csv`). I logged eight input problems in `data_issues`. The ones that matter most:
- The 5 km CS-DR ATT has SE 0.210 in Table 6 but 0.218 in Table A19. I used Table 6 because the key names it.
- Table A19's note says there are 640 never-treated regions; Tables 6 and 7 say 608.
- For the fatal 5 km effect, the abstract says 35% and Section 6.1 says about 30%.
- The prose gives the weeks-on-UI CI as positive magnitudes, while Table 2 prints negative bounds.
- One paragraph has a rendering defect: "p ¡ 0.001".

## Method
`out/analysis.py` parses the " | "-delimited table rows. In the DCR tables, the SE is the parenthesised value on the row after each coefficient. The script removes "$", thousands separators, units and significance stars, keeps the printed sign, and prints every matched source line. Locators follow `<source> | Table n | row "…" | column "…"`. Significance at 10% is read from the stars (`*` means p<0.10). The script searches the extract for an exact p-value and finds none. Derivations use the normal critical value 1.96 (`derivation_notes`). There is no randomness (`random_seed` = null). The baseline is simply the direct extraction, since the estimands are the printed numbers.

## Results
Note that JSON drops trailing zeros (0.210 is stored as 0.21). The verbatim strings are kept in `printed_strings`.

**DCR (arxiv-2609.27732v1)**
- Table 6, overall ATT:
  - 5 km: −0.431** with SE 0.210 (`dcr_att_5km`, `dcr_att_5km_se`).
  - 3 km: −0.283 with SE 0.279 (`dcr_att_3km`, `dcr_att_3km_se`).
  - Observations at 5 km: 17,250 (`dcr_obs_5km`).
- Table 7, by outcome:
  - Narcotics at 3 km: 0.549*** with SE 0.187 (`dcr_narcotics_att_3km`, `dcr_narcotics_att_3km_se`).
  - Drug-related (fatal) at 1 km: −1.207*** with SE 0.137 (`dcr_fatal_att_1km`, `dcr_fatal_att_1km_se`).
  - HIV at 5 km: −0.340 with SE 0.294 and no star (`dcr_hiv_att_5km`, `dcr_hiv_att_5km_se`). It is therefore not marked significant at 10% (`dcr_hiv_att_5km_significant_10pct` = false). The source prints no exact p-value, so `dcr_hiv_att_5km_p_value` = null.
- Derived 95% CI for the Table 6 5 km ATT: [−0.8426, −0.0194] (`dcr_att_5km_ci_low`, `dcr_att_5km_ci_high`).

**RESEA (arxiv-2606.14621v1, Table 2)**
- Wage ATE: $1,153 per year, 95% CI [166, 2,141] (`ri_wage_ate`, `ri_wage_ate_ci_low`, `ri_wage_ate_ci_high`).
  - No SE is printed (`ri_wage_ate_se_reported` = false).
  - The SE implied by the CI is 503.83 (`ri_wage_ate_se_derived`).
- Control-group mean wage: $33,043 (`ri_control_mean_wage`).
- Reemployment ATE: 1.47 pp (`ri_reemployment_ate_pp`).
- Weeks on UI ATE: −1.99, 95% CI [−2.24, −1.75] (`ri_weeks_ui_ate`, `ri_weeks_ui_ate_ci_low`, `ri_weeks_ui_ate_ci_high`).

## Uncertainty
- **SEs:** all DCR uncertainties (Tables 6 and 7) are SEs clustered at the MedStat region level, printed in parentheses (`dcr_*_se`). The stars use the source's thresholds: *** p<0.01, ** p<0.05, * p<0.10.
- **CIs:** the RESEA uncertainties are two-sided 95% CIs, printed with p-values. They come from the "methodology described in the main text", which is not part of the extract.
- **Derived DCR CI:** `dcr_att_5km_ci_low`/`_high` = ATT ± 1.96·SE. This assumes the estimator is approximately normal with the printed clustered SE. With a t critical value or 2.0 the interval would be slightly wider. The interval excludes 0, which agrees with the ** marking (|t| ≈ 2.05; `out/state/checks.md`).
- **Derived RESEA SE:** `ri_wage_ate_se_derived` = (2,141 − 166)/(2·1.96). This assumes a symmetric normal-approximation CI. The printed CI is symmetric to within 1 USD of rounding. Plugging the derived SE back in gives a two-sided p of about 0.022, matching the printed p=0.022 (checks.md).
- Inputs are printed to only 3 decimals (DCR) or whole dollars (RESEA), so the derived values carry that rounding error.

## Limitations and open issues
- The source does not print the exact p-value of the HIV 5 km ATT. It cannot be obtained without a distributional assumption, so I left it null. Its absence is the expected "not reported" item.
- The task is otherwise complete (`feasibility` = "complete").
- Using Table A19's SE (0.218) instead of Table 6's would give a different derived CI. The source does not explain the discrepancy, and I did not resolve it.
- The method behind the RESEA CI is not in the extract, so the derived SE depends on the normality and symmetry assumption.
- I found the internal inconsistencies listed under `data_issues` but did not adjudicate them. None of them changes a requested key.