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

# Question

Extract, exactly as printed, selected treatment effects, standard errors (SEs), confidence-interval (CI) bounds and counts from (i) Armendariz (2026), arXiv:2609.27732v1, a staggered difference-in-differences study of drug consumption room (DCR) openings in Switzerland (Tables 6 and 7), and (ii) Li, Pearson-Merkowitz & Yokum (2026), arXiv:2606.14621v1, a randomised evaluation of Rhode Island's RESEA programme (Table 2). Give each value a locator, derive one 95% CI from an SE and one SE from a 95% CI, and identify which requested quantity the source does not report. Nothing is re-estimated.

# Data and exclusions

The inputs are the two plain-text extracts under data/sources/ and data/README.md, which was read only for provenance. Line counts are n_input_rows: README 10, DCR extract 134, RESEA extract 58. All SHA-256 hashes match the README and the bundle (out/state/data_audit.md). No records were excluded. out/exclusions.csv holds a single row that removes nothing: 22 requested source items were assessed and 0 removed, so n_analysed = 22. Those 22 are 11 printed DCR values, the HIV p-value lookup, the HIV significance flag, 8 printed RESEA values and the RESEA SE-reported flag. The three derived values are computed from these items. The problems found in the sources are listed under data_issues and summarised under Limitations below.

# Method

out/analysis.py parses the table blocks from the text using the " | " cell separators. In the DCR tables, the line after each coefficient row holds the SEs in parentheses. Each value is taken from the table the key names, and the script prints every source line it matched. Significance stars are counted to set dcr_hiv_att_5km_significant_10pct. The script searches the whole DCR extract for an exact printed p-value ("p = 0.x") and finds none. Derived quantities (conventions in derivation_notes):

• 95% CI = ATT ± 1.96 × SE, using the Table 6 printed ATT and SE, rounded to 3 decimals.
• SE = (upper − lower) / (2 × 1.96), using the Table 2 printed wage-ATE CI, rounded to 2 decimals.

The code uses no randomness (random_seed = null).

# Results

DCR study (arXiv:2609.27732v1), log points:

• Table 6, 5 km: ATT dcr_att_5km = -0.431 (**), SE dcr_att_5km_se = 0.210. Observations dcr_obs_5km = 17250.
• Table 6, 3 km: ATT dcr_att_3km = -0.283, SE dcr_att_3km_se = 0.279.
• Table 7, Narcotics, 3 km: dcr_narcotics_att_3km = 0.549 (***), SE dcr_narcotics_att_3km_se = 0.187.
• Table 7, Drug-related (fatal), 1 km: dcr_fatal_att_1km = -1.207 (***), SE dcr_fatal_att_1km_se = 0.137.
• Table 7, HIV, 5 km: dcr_hiv_att_5km = -0.340, SE dcr_hiv_att_5km_se = 0.294.
  • The source does not print an exact p-value, so dcr_hiv_att_5km_p_value = null. The paper reports significance only through stars.
  • The estimate has no star, so dcr_hiv_att_5km_significant_10pct = false.
• Derived 95% CI for the 5 km ATT: [dcr_att_5km_ci_low = -0.843, dcr_att_5km_ci_high = -0.019].

RESEA study (arXiv:2606.14621v1), Table 2:

• Wage ATE ri_wage_ate = 1153 USD/yr, 95% CI [ri_wage_ate_ci_low = 166, ri_wage_ate_ci_high = 2141].
  • The table prints no SE (ri_wage_ate_se_reported = false).
  • Derived SE ri_wage_ate_se_derived = 503.83.
• Control-group mean wage ri_control_mean_wage = 33043 USD/yr.
• Reemployment ATE ri_reemployment_ate_pp = 1.47 pp.
• Weeks on UI ATE ri_weeks_ui_ate = -1.99 weeks, 95% CI [ri_weeks_ui_ate_ci_low = -2.24, ri_weeks_ui_ate_ci_high = -1.75].

feasibility = "complete". Every requested quantity was either extracted, derived, or correctly identified as not reported (the HIV p-value).

Note on format: a JSON number cannot keep trailing zeros, so results.json stores 0.210 as 0.21 and -0.340 as -0.34. The numeric values are identical, and the printed forms (0.210, -0.340) are shown above.

# Uncertainty

• Standard errors. In the DCR paper (Tables 6 and 7) the uncertainty is given as SEs clustered at the MedStat region level, printed in parentheses, plus significance stars. It prints no CIs or exact p-values for these cells.
• Confidence intervals. In the RESEA paper (Table 2) the uncertainty for each ATE is a two-sided p-value and a 95% CI "based on the methodology described in the main text". That methodology is not in the extract, and no SE is printed.
• What the derived CI assumes. The 5 km CI assumes the CS-DR estimator is approximately normal and that the clustered SE is valid, using a normal critical value of 1.96. It is not a source value. Using the Table A19 SE (0.218) for the same estimate would widen it. A t critical value with few clusters would also widen it, but the number of clusters behind the SE is not stated in the extract.
• What the derived SE assumes. The wage SE assumes the printed CI is a symmetric Wald interval with z = 1.96. Two things support this: the interval is symmetric (1153 − 166 = 987 and 2141 − 1153 = 988), and the SE reproduces the printed p = 0.022 (checked in out/state/checks.md). With N = 23,549, a t critical value would make no material difference. The CI bounds are rounded to whole dollars, which makes the derived SE uncertain by about ±0.3 USD.

# Limitations and open issues

• Inconsistencies in the sources. These are recorded in data_issues and none of them changed the extraction rule.
  1. Table 6 prints an SE of 0.210 for the 5 km ATT, while Table A19 prints 0.218 for the same CS-DR estimate. I used Table 6, as the key requires.
  2. Tables 6 and 7 say there are 608 control regions, while the Table A19 note says 640. The Table 6 Observations equal (640 + treated regions) × 25 in every column. The 25-period panel length is my inference; the extract does not state it.
  3. The prose sometimes differs from the tables: the abstract's 35% fall in fatal outcomes against about 30% implied by -0.354, and HIV described as "close to zero" against -0.340. In the RESEA text, the weeks-on-UI CI is given as positive magnitudes.
  4. There is a rendering defect ("p ¡ 0.001"), and RESEA Tables 2–3 were re-set by the extract editor.
• Limits on verification. "As printed" can only be checked against the extracts, which may contain transcription errors relative to the original papers. I did not consult the originals, as the task rules require.
• Derived values. These are conventions, not values from the sources (see Uncertainty).