AN-042: Volume-matched timing audit — is there a second bid-conduct channel?¶
Intuition (plain-language)
AN-041 left the bid-conduct case resting on a single channel (closeness to the winner). A one-channel story is thin, so this page hunts for a second: do cobidders bid on a different clock — revising less, spacing bids out, finishing early? Holding volume fixed, the answer is no — every timing dimension is statistically indistinguishable, the two largest effects merely tail-driven. A documented null: the cobidder conduct signature is genuinely single-channel, and the paper says so rather than dressing a noisy timing difference up as a second one.
Question¶
H5's bid-conduct distinctness rests on a single surviving channel (median
gap-to-winner). A second, orthogonal behavioral channel would lift the
bid-conduct claim above one residual. Bid timing is the natural
candidate: in pregão's real-time auction, a cover bidder placing token
bids might revise less, disengage earlier, or bid on a characteristic
rhythm. Do cobidders differ on timing, and does it survive matching on
tenders_count?
Design¶
- Sample. Within-FL stratum, cobidder vs FL non-cobidder, direct CADE removed; firms with \(\geq 5\) pregão tender-items (cob 187, ctrl 1,946). Pregão only — convite is sealed-bid, so revision/timing dynamics are degenerate there.
- Metrics (per firm, from
Data Hr Proposta): revision intensity (mean bids/item), median inter-bid interval (log), last-bid position in the tender's bidding window \([0,1]\), engagement-span fraction \([0,1]\). - Matching. Identical to AN-041: cobidder vs FL non-cobidder on
tenders_count; M1 PS nearest-neighbour caliper 0.2, M2 CEM deciles.
Results¶
Balance: tenders_count SMD 0.398 → 0.002 (M1), 183 matched pairs.
| Timing dimension | d raw | d M1 (PS) | d M2 (CEM) | Wilcoxon p (M1) |
|---|---|---|---|---|
| Revision intensity (bids/item) | −0.002 | −0.202 | −0.177 | 0.86 |
| Inter-bid interval (log s) | −0.081 | +0.054 | −0.032 | 0.96 |
| Last-bid position in window | +0.167 | +0.105 | +0.058 | 0.64 |
| Engagement-span fraction | −0.197 | −0.310 | −0.295 | 0.23 |
No timing dimension clears |d| ≥ 0.2 and p < 0.05 after matching.
Interpretation¶
The second channel does not exist. The two largest matched effects --- cobidders revise less (revision intensity d ≈ −0.20) and disengage earlier (engagement-span d ≈ −0.31) --- point in the direction a token-cover-bidding story would predict, but neither rank-sum test rejects the null (p = 0.86 and 0.23). The Cohen's \(d\) is driven by the tails of a non-normal, bounded distribution rather than by a central shift, so the distribution-free test --- the standard used throughout this project --- returns no significant difference. We report the null straight.
Consequence for the bid-conduct claim. Cobidders' bid-conduct distinctness is single-channel: the median gap-to-winner (AN-031, AN-041) survives volume matching, but both the within-firm dispersion sub-signal (AN-041) and the timing channel (this page) are documented nulls. The distinctness of cobidders is therefore structural --- product specialization, winner spread, proximity to the winning bid --- not a behavioral timing signature.
Bearing on hypotheses¶
- H:cobidder-profile-distinct (H5). Bounds the claim: the profile/structural distinctness is robust to volume (AN-041); the bid-conduct distinctness is one channel, not two. The promotion case for H5 rests on the profile leg, not on a multi-channel behavioral signature. Reported as a scope limit, not as support.
Follow-ups¶
- Within-tender bid-sequence shape (monotone undercut vs erratic) under the same match — a finer timing object than the four moments here; lower priority given the null.
- Cross-jurisdiction replication (ComprasNet): whether the structural distinctness (not timing) reproduces on a non-BEC panel is the path to Confirmed.