Reverse-direction scenario for the bias sensitivity (S3; exploratory, added after the results were seen, not part of the decision rule)
Background: the tipping-point analysis (ext_bias_sensitivity_output.txt) looks in one direction only: truly unrelated ordinary pairs miscoded "related" would inflate the sound difference.
Here we look in the other direction: the 31 labelled pairs that the first blind pass coded unrelated are treated as related on the strength of the label itself (as if the coders had missed a link), added to the related group, and the near shares are compared again.
This is an extreme scenario, not an estimate: it assumes that all 31 pairs really are related in meaning, so it is a stress test.

1. What the 31 pairs are
  In the frame, 634 pairs: labelled 172 (coded related 141, coded unrelated 31), ordinary 462 (coded related 79, coded unrelated 383).
  The 31 labelled pairs coded unrelated: near 8 (26%), not near 23 (74%). Middle Chinese category: other 19, identical 7, alt_other 4, alt_departing 1.
  Source of the 31 pairs: shared by Da Xu and Xiao Xu 25, Da Xu only 5, collation undecided 1; with paronomastic glosses 0; new items 23, first-sample items 8.
  For comparison: of the labelled pairs coded related, 141 pairs, 73 are near (52%). So the 31 pairs added have a much lower near share and pull the near share of the labelled group down.

2. The main test under the three choices (near ~ label among the pairs coded related; GEE clustered by phonetic, one-sided p; rule: p < .05 and upper confidence limit >= 2 supports C)
  Reference: the main test (first blind pass, pairs coded related)
      73/141 vs 23/79, OR 2.35 [1.43, 3.86], p1 0.00038; MH 2.01 [0.75, 5.39]; pooled (unweighted) OR 2.61; rule: supports C (same word / minimal derivation)
  S3: all 31 labelled pairs coded unrelated are added to the related group
      81/172 vs 23/79, OR 2.00 [1.23, 3.26], p1 0.0026; MH 2.18 [0.86, 5.48]; pooled (unweighted) OR 2.17; rule: supports C (same word / minimal derivation)
  S3, least favourable choice: only the not-near ones among the 31 are added
      73/164 vs 23/79, OR 1.76 [1.11, 2.80], p1 0.0081; MH 1.75 [0.67, 4.56]; pooled (unweighted) OR 1.95; rule: supports C (same word / minimal derivation)

3. Identity (S1 definition) under the three choices
  Reference: the main test (first blind pass, pairs coded related)
      54/141 vs 15/79, OR 2.57 [1.35, 4.87], p1 0.0019; MH 1.57 [0.57, 4.32]; pooled (unweighted) OR 2.65; rule: supports C (same word / minimal derivation)
  S3: all 31 labelled pairs coded unrelated are added to the related group
      61/172 vs 15/79, OR 2.31 [1.22, 4.38], p1 0.0051; MH 1.69 [0.66, 4.33]; pooled (unweighted) OR 2.34; rule: supports C (same word / minimal derivation)
  S3, least favourable choice: only the not-near ones among the 31 are added
      54/164 vs 15/79, OR 2.03 [1.09, 3.78], p1 0.013; MH 1.37 [0.51, 3.67]; pooled (unweighted) OR 2.09; rule: supports C (same word / minimal derivation)

4. Comparison with a hand computation from the paper's Table 8
  Hand computation from the counts of Table 8: 81/172 against 23/79, OR about 2.17. This script: 81/172 vs 23/79, OR 2.00 [1.23, 3.26], p1 0.0026; MH 2.18 [0.86, 5.48]; pooled (unweighted) OR 2.17.
  The 2.17 is the unweighted pooled odds ratio (the script gives 2.17, in agreement); the GEE clustered by phonetic gives a point estimate of 2.00, lower than the pooled value, and the interval is the GEE interval.

Notes:
  1. Exploratory, post hoc; the main test remains the first-blind-pass result of analysis plan section 4.2 and the rule has not been changed; no new coding was done and the official numbers are not changed.
  2. The 31 pairs added have a near share of only 8/31 (26%), lower than the 52% of the labelled pairs coded related, so adding all of them lowers the near share of the labelled group; the choice that lowers it most adds only the 23 not-near pairs among them (the near ones are not added). Pooled (unweighted) OR: not folded 2.61, all folded 2.17, only the not-near ones folded 1.95. The GEE estimates are in section 2; only these choices were computed, not all subsets.
  3. Result: with all folded in, GEE OR 2.00 [1.23, 3.26], against 2.35 [1.43, 3.86] for the reference and 1.76 [1.11, 2.80] for the least favourable choice. The lower confidence limit is above 1 in all three, and by the rule all three read as "supports (C)" (upper limit >= 2 and one-sided p < .05); but the point estimate falls to about 2, the smallest effect of interest.
  4. Taking the two directions together: false positives in the ordinary group (ext_bias_sensitivity_output.txt) can bring the OR down to 1 if their rate is large enough; this reverse direction, even in its most extreme form, does not bring the OR down to 1, only to about 2.
  5. The GEE and the pooled odds ratio weight the pairs differently: 2.17 is the pooled odds ratio, whereas the GEE of the main test is clustered by phonetic; the paper cites the GEE figures.
