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

# Question

The task was to copy specified treatment-effect estimates, standard errors (SEs), confidence-interval (CI) bounds and sample counts exactly as printed in two sources: Armendariz (2026, arXiv:2609.27732v1, drug consumption rooms (DCRs), Tables 6–7) and Li, Pearson-Merkowitz & Yokum (2026, arXiv:2606.14621v1, Rhode Island RESEA, Table 2). Each value gets a locator. Two quantities that the sources do not print are derived from printed values, and any requested quantity that is not reported is identified. Nothing is re-estimated.

# Data and exclusions

• data/sources/arxiv-2609.27732v1.txt (134 lines), data/sources/arxiv-2606.14621v1.txt (58 lines) and data/README.md (10 lines) were read. No other evidence was used.
• I used only the tables named in each key's definition: DCR Tables 6 and 7, and RESEA Table 2. DCR Appendix Table A19 and RESEA Table 3 were used only for consistency checks.

# Method

out/analysis.py parses the pipe-separated table rows. It finds each row by its label and each column by its header, and treats the next non-empty line after a coefficient row as that row's SE line. It prints every source line it matched. Thousands separators are removed (e.g. "17,250" → 17250). Signs and decimals are kept as printed. One limitation: JSON numbers cannot keep trailing zeros, so 0.210 is stored as 0.21 and -0.340 as -0.34. The verbatim strings are therefore also kept in results.json["printed_strings"]. Significance is read from the stars in the table, using the table note: * means p<0.10, ** p<0.05, *** p<0.01. The script also checks that the RESEA Table 2 ATE column matches the Table 3 "Difference in means" column (it does, for all three outcomes).

# Results

Values are as printed, with locators in results.json["locators"].

DCR (log points; standard errors in parentheses)

RESEA (Table 2, "Average treatment effect estimate" column)

25 reported or derived items were assessed (n_analysed).

# Uncertainty

• The DCR figures are standard errors. The values in parentheses in Tables 6 and 7 are SEs clustered at the MedStat region level, according to the table notes. The paper prints no exact p-values or CIs for these cells; it marks significance with stars only. An exact p-value for HIV at 5 km is therefore reported as null. Computing one from -0.340/0.294 would need an extra distributional assumption, and the task asks for the printed p-value.
• The RESEA figures are 95% confidence intervals (together with two-sided p-values). The paper does not print SEs.
• Derived DCR CI. Computed as ATT ± 1.96 × SE: -0.431 ± 1.96 × 0.210, which gives [-0.8426, -0.0194]. This assumes the estimator is approximately normal. The paper's own figure CIs might use another critical value, such as one from a multiplier bootstrap. If the Appendix A19 SE of 0.218 were used instead, the CI would be [-0.858, -0.004].
• Derived RESEA SE. Computed as (2,141 − 166)/(2 × 1.96) = 503.83. This assumes the CI is a symmetric normal-theory (Wald) interval. The printed interval is nearly symmetric around 1,153 (987 below, 988 above), which fits that assumption. With N = 23,549, a t critical value would give essentially the same result. Because the printed bounds are rounded to whole dollars, the SE is uncertain by about ±0.5 USD.

# Limitations and open issues

Problems found in the sources (also listed in data_issues):

1. DCR SE conflict. The 5 km CS-DR ATT of -0.431 has SE 0.210 in Table 6 but 0.218 in Appendix Table A19, for the same specification with N = 17,250. I used Table 6, as the key definition requires.
2. DCR control-region count. Tables 6/7 and the A19 rows give 608 never-treated control regions, but the A19 note says 640.
3. DCR abstract vs Table 7. The abstract says fatal hospitalisations "fall by 35%". Table 7's 5 km fatal ATT of -0.354 log points corresponds to about −29.8%, and Section 6.1 itself says about 30%. The abstract appears to read log points as percent.
4. DCR HIV wording. The text calls the HIV estimates "close to zero", but at 5 km the estimate is -0.340 (SE 0.294). It is insignificant, but not small.
5. RESEA weeks-on-UI CI. The text prints the CI as magnitudes, [1.75, 2.24]. Table 2 prints [-2.24, -1.75], and I used Table 2's version.
6. Rendering defect. The RESEA extract contains "p ¡ 0.001" in the short-term weeks-on-UI paragraph. No key uses it.
7. Re-set tables. RESEA Tables 2–3 were re-set by the extract editor. I accepted them as verbatim, as the extract header states, and they agree with each other.

The task is complete. The only null value, the HIV p-value, is null because the source does not report it.