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

## Question
I extracted specific point estimates, standard errors (SEs), confidence-interval (CI) bounds and counts exactly as printed in two source extracts, each with a locator. The sources are Armendariz (2026), arxiv-2609.27732v1, a staggered DiD study of drug consumption rooms (DCRs), Tables 6–7; and Li, Pearson-Merkowitz & Yokum (2026), arxiv-2606.14621v1, the Rhode Island RESEA RCT, Table 2. I also derived one 95% CI from an SE and one SE from a 95% CI, and flagged quantities the sources do not print. Nothing is re-estimated.

## Data and exclusions
The inputs are `data/sources/arxiv-2609.27732v1.txt` (134 lines), `data/sources/arxiv-2606.14621v1.txt` (58 lines) and `data/README.md` (10 lines) (`n_input_rows`). The SHA-256 hashes match the README (`out/state/data_audit.md`). No items were excluded: all 25 requested items were assessed (`n_analysed` = 25; `out/exclusions.csv`). These are 19 printed values, 3 derived values, 1 p-value (null) and 2 booleans. Values are read from the table that each key names: Table 6 rather than Table A19, and Table 2 rather than Table 3 or the running text.

## Method
`out/analysis.py` parses the tables as text and prints the matched source lines. Stars, `$` signs and thousands separators are removed before conversion (for example "17,250" → 17250). The script stores the printed strings verbatim in `printed_tokens`, because JSON numbers drop trailing zeros: 0.210 is serialised as 0.21 and −0.340 as −0.34, while the printed values are "(0.210)" and "-0.340". Two quantities are derived:
- **DCR 5 km CI** (`derivation_notes.dcr_att_5km_ci`): b ± 1.96·SE, using the Table 6 values.
- **RESEA wage SE** (`derivation_notes.ri_wage_ate_se`): (upper − lower)/(2·1.96), using the Table 2 CI.

No randomness is involved (`random_seed` = null). The "baseline" for this task is the direct reading of each printed cell; no further method is layered on top.

## Results
**DCR study (log points; arxiv-2609.27732v1)**
- Table 6, 5 km: ATT −0.431** (`dcr_att_5km`), SE (0.210) (`dcr_att_5km_se`), 17,250 observations (`dcr_obs_5km`).
- Table 6, 3 km: ATT −0.283 (`dcr_att_3km`), SE (0.279) (`dcr_att_3km_se`).
- Table 7, Narcotics, 3 km: 0.549*** (`dcr_narcotics_att_3km`), SE (0.187) (`dcr_narcotics_att_3km_se`).
- Table 7, Drug-related (fatal), 1 km: −1.207*** (`dcr_fatal_att_1km`), SE (0.137) (`dcr_fatal_att_1km_se`).
- Table 7, HIV, 5 km: −0.340 (`dcr_hiv_att_5km`), SE (0.294) (`dcr_hiv_att_5km_se`). The cell has no star, so the source does not mark it significant at 10% (`dcr_hiv_att_5km_significant_10pct` = false). **The exact p-value is not reported** (`dcr_hiv_att_5km_p_value` = null): the source prints only stars.
- 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 RCT (arxiv-2606.14621v1, Table 2)**
- Wage ATE: $1,153 per year (`ri_wage_ate`), 95% CI [166, 2,141] (`ri_wage_ate_ci_low`, `ri_wage_ate_ci_high`).
- **No SE is printed for the wage ATE** (`ri_wage_ate_se_reported` = false). The SE derived from the CI is 503.83 USD per year (`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 weeks (`ri_weeks_ui_ate`), 95% CI [−2.24, −1.75] (`ri_weeks_ui_ate_ci_low`, `ri_weeks_ui_ate_ci_high`).

Every value has a locator in `locators`. The task is complete (`feasibility`).

## Uncertainty
- **Standard errors.** The DCR uncertainties (Tables 6, 7, A19) are SEs clustered at the Medstat region level, printed in parentheses. Neither table prints a CI or an exact p-value.
- **Confidence intervals.** The RESEA uncertainties (Table 2) are 95% CIs with two-sided p-values. The ATE intervals are "based on the methodology described in the main text", which is not part of the extract. No SEs are printed.
- **What the derived DCR CI assumes.** It assumes the estimator is asymptotically normal and uses the normal critical value 1.96. A t critical value, or a different number of effective clusters, would widen it slightly. It also depends on using the Table 6 SE (0.210). Table A19 prints 0.218 for what is described as the same specification (see `data_issues`).
- **What the derived RESEA SE assumes.** It assumes a symmetric, normal-based Wald interval. The half-widths are 987 and 988, so the interval is symmetric up to rounding. As a check, the implied z of 2.288 gives a two-sided p of 0.0221, compared with the printed p=0.022 (`derivation_notes.ri_wage_ate_se`). The printed CI bounds are rounded to whole dollars, so the derived SE is uncertain at about the ±0.3 USD level.

## Limitations and open issues
All of these are recorded in `data_issues`.
1. **Different SEs for the same estimate.** In the DCR source, Table 6 gives SE 0.210 and Table A19 gives 0.218 for the same 5 km CS-DR ATT (−0.431). I used Table 6, as the key definitions require.
2. **Two different control-region counts.** The tables say 608 never-treated control regions, but the A19 note says 640. The printed Observations (16,175 / 16,875 / 17,250) equal exactly 25 × (640 + treated regions), not 608. I report Observations as printed and leave the discrepancy unresolved.
3. **Abstract does not match Table 7.** The DCR abstract says fatal outcomes fall by 35%. The Table 7 5 km coefficient (−0.354) implies a fall of about 29.8%, and Section 6.1 says about 30%.
4. **Text overstates how small the HIV estimates are.** The text calls them "close to zero", but at 5 km the estimate is −0.340: not significant, but not small.
5. **Small rounding gaps.** Two "% change" entries in Table A19 differ by about 0.1 percentage points from exp(b)−1 computed from the printed coefficients.
6. **RESEA text issues.** The text contains a rendering defect ("p ¡ 0.001"). It also states the weeks-on-UI CI as magnitudes of a reduction, which flips the sign relative to Table 2; I used Table 2.
7. **Re-set tables.** RESEA Tables 2–3 were re-set by the extract editor. The Table 3 "Difference in means" column is verified identical to the Table 2 ATE column.
8. **Verbatim strings.** Trailing zeros of printed values cannot be represented as JSON numbers; the verbatim strings are kept in `printed_tokens`.