Skip to content

The under-the-gun gap is selection-bounded at 15.9–21.1%

🟡 Within the administrative urgent channel — the closest feasible urgent-procurement comparison to litigation-driven orders — items procured "under the gun" show a naive price gap of 29.5% (coef −0.259), but once the gap is bounded for selection into the urgent channel using Lee (2009) bounds, the price difference compresses to a 15.9%–21.1% range (coefs −0.148/−0.192), with a mean trimming rate of 26.9% (AN-002).

Economic intuition

The honest question is not "how much dearer are litigated orders?" but "how much survives once we admit the comparison group was screened?" Administrative requests are filtered, so the items in that channel are not a clean counterfactual. Lee bounds answer the harder question: under a monotonicity restriction, the gap lies in a range that stays well above zero. Bounding — not the raw 29.5% — is what makes the claim defensible.

The bounded effect is statistically robust under inference that accounts for the small number of clusters: a wild cluster bootstrap returns p = 0.0080 in the preferred specification and p = 0.0390 under the more demanding item-by-year-month clustering (AN-007). The Lee-bounds exercise is the central correction in this finding: the naive gap overstates the difference because the administrative urgent channel is selected and larger, so a substantial part of the raw 29.5% reflects which items flow through that channel rather than the price consequence of urgency alone.

Caveat. The administrative urgent channel is the closest feasible urgent-procurement comparison, not a random or clean assignment. The Lee bounds discipline selection on observed differential channel composition, but they rest on a monotonicity assumption and do not convert the contrast into a structural counterfactual. The bounded 15.9%–21.1% range is a selection-bounded gap, not a point estimate of a causal treatment effect. The reading is 🟡 because it is a single-source own-project estimate in São Paulo BEC alone; promotion to 🟢 would require independent replication.

Sources.

  • Own analysis: AN-002 (naive gap, Lee bounds, mean trimming), AN-007 (wild cluster bootstrap inference).
  • Cross-refs: H:utg-gap-selection-bounded.
  • Validation: backing scripts 40_utg_lee_bounds.R, 44_wild_bootstrap.R, 50_v9_outputs.py.