Skip to content

AN-013: Pregão drop-out extraction

Intuition (plain-language)

The structural primitives come from BEC's auction event logs: each losing firm's drop-out price is read as its willingness to supply. This page extracts those drop-outs and checks the simulation is not sensitive to how the unobserved winner's cost is treated — across three censoring methods the net effect stays positive and economically large.

Question

The structural sample is built from BEC Pregão event-log data — losing-bidder exit prices and final-state bids per item. How is the extraction implemented, and how sensitive are downstream simulations to the choice of how to treat the winner (who does not drop out, so the final bid is a one-sided observation)?

Design

  • Sample: BEC Pregão event log spanning March 2018; cells by class (non-pharma standardized; pharma) × period (pre/post) × bidder type (SME / non-SME).
  • Variation: identification of which bid in the event log represents a losing-bidder exit price — the bid after which that firm submits no further bids and another firm wins.
  • Specification: parse the event log per item; for each firm, take the final observed bid as the willingness-to-supply observation; normalize by the buyer reference price.
  • Winner-censoring regimes tested downstream:
  • Losers-only: discard the winner's final bid entirely.
  • All-bidders (baseline): treat the winner's final bid as a tight upper-bound observation on the winner's cost.
  • Turnbull NPMLE (v7-jpube-tight/scripts/48_turnbull_fc.R): treat the winner's final price as left-censored at that bound; non-parametric maximum likelihood estimation of the resulting censored distribution.

Results

Per-cell drop-out coverage: documented in the extraction parquet (v7-jpube-tight/output/dropouts/dropouts_panel.parquet). The structural sample covers thousands of pregão auctions per cell across the four cells (NP/PH × pre/post) and two types (SME / non-SME); cell sizes are reported in paper_v8.tex §3.

Robustness of net \(S_3 - S_1\) to winner-censoring regime (paper §6.2):

Regime NP net \(S_3-S_1\) PH net \(S_3-S_1\) NP dev. PH dev.
Losers-only 0.275 0.347 +6% +13%
All-bidders (base) 0.259 0.308 — —
Turnbull NPMLE 0.246 0.357 +5% +16%

The Turnbull within-auction (exclusion) share remains large: 74.0% in non-pharma and 82.0% in pharma — close to or above the baseline 72.0% / 68.8%. The result is therefore not driven by treating the winner's final bid as an exact cost observation.

Interpretation

The drop-out extraction is the foundation of the structural sample. It produces empirical per-cell distributions used by all downstream AN pages (UH correction in AN-014, BNE decomposition in AN-010, welfare arithmetic in AN-011, strict invariance in AN-017).

Winner censoring is not the load-bearing assumption. Three treatments (losers-only, all-bidders upper bound, Turnbull NPMLE left-censored) give net \(S_3 - S_1\) in the range [0.246, 0.275] for non-pharma and [0.308, 0.357] for pharma. All three are positive and economically large. The Turnbull NPMLE — which is the most agnostic about the winner's true cost — gives a larger pharma exclusion share (82.0% vs 68.8% baseline), strengthening rather than weakening the qualitative result.

Confidence: yellow. The extraction is mechanical (it parses event logs); the downstream sensitivity to winner-censing regime is well-bounded; but the interpretation of drop-outs as cost observations is the maintained IPV-clock restriction (H:ipv-clock-admissible). That restriction is disciplined by the UH correction (AN-014), the cross-modality check (§6.2), the Gaussian-copula relaxation (§6.2), and the collusion screens (AN-015) — collectively yellow rather than green because none of these prove IPV; they show the decomposition is not mechanically produced by the most obvious deviations from IPV.

Follow-ups

  • Per-firm filter sensitivity: minimum drop-out count per firm, minimum auction size, etc. (v7-jpube-tight/scripts/52_filter_sensitivity.R). Currently bundled in §6.2.
  • Cross-modality contrast: the same conceptual recovery on Convite first-price bids (GPV inversion) produces a structurally different object — a strategic bid rather than a willingness-to-supply exit. v7-jpube-tight/scripts/38_cross_modality.R shows convergence in the key pharma non-SME pre-period cell after UH correction.