AN-015: Classical bid-coordination screens¶
Intuition (plain-language)
If the price rise were really the surviving SME pool quietly coordinating, the drop-out-as-cost reading would collapse. Standard collusion screens look for a post-policy jump in suspicious bid-clustering. There is baseline clustering but no post-policy intensification, so a new coordination shock is not driving the result.
Question¶
The structural decomposition rests on the maintained IPV-clock interpretation of Pregão drop-outs (H:ipv-clock-admissible). If bidder coordination intensifies across the March 2018 policy break — because the smaller eligible SME pool meets the same competitors more often — drop-outs are strategically suppressed in the post-period and the IPV reading breaks down. The identifying question is narrower than "is there coordination": baseline clustering is endemic in procurement. The question is whether clustering intensifies after non-SMEs are removed.
Design¶
- Sample: BEC Pregão bid logs spanning the policy break; non-pharma standardized and pharma cells separately.
- Variation: pre-policy vs post-policy bidder pools in switched group 65.
- Specification:
- Conley-Decarolis (2016) close-pair screen
(
v7-jpube-tight/scripts/58_collusion_screen.R,59_collusion_screen_schurter.R): share of firm pairs with abnormally close residual bids ("close" = within a calibrated bid distance after partialling out auction-level controls). The realized share is compared against a permutation null mean drawn by re-shuffling firm pairs across auctions. - Bajari-Ye (2003) T1 ratio
(
v7-jpube-tight/scripts/60_collusion_screen_bajari_ye.R): observed-to-null moment ratio on residual bid co-movement; a ratio above 1 indicates clustering, with stronger ratios signaling tighter coordination. - Pair-classification cross-check
(
v7-jpube-tight/scripts/61_collusion_screen_pair_classcond.R): persistent co-bidding pairs are flagged and their coordination signatures (within-auction bid dispersion, exit-order patterns) tested for pre-vs-post invariance. - Outcomes: realized share vs null mean (Conley), T1 ratio (Bajari-Ye), per period × class.
Results¶
Conley-Decarolis close-pair shares (%, realized vs permutation null):
| Class | Pre realized | Pre null | Post realized | Post null | Δ realized |
|---|---|---|---|---|---|
| Non-pharma | 16.9 | 10.6 | 16.8 | 10.2 | −0.1 (flat) |
| Pharma | 27.6 | 16.3 | 24.4 | 14.4 | −3.2 (falls) |
Realized shares exceed the null in every cell (baseline clustering is real). But the realized share is flat in non-pharma (16.9 → 16.8) and falls in pharma (27.6 → 24.4) post-cutoff. The pharma drop in null mean (16.3 → 14.4) tracks the realized drop — i.e., the screen's permutation distribution shifts proportionally as the auction-pool composition changes; the gap between realized and null is stable.
Bajari-Ye T1 ratios (observed / null):
| Class | Pre ratio | Post ratio | Δ ratio |
|---|---|---|---|
| Non-pharma | 2.63 | 1.83 | −0.80 |
| Pharma | 1.29 | 1.11 | −0.18 |
Both ratios fall post-cutoff. The non-pharma fall is large (31% reduction). The pharma ratio is close to 1 in both periods (consistent with a sparser, less-connected bidder pool to begin with).
Output: v7-jpube-tight/output/collusion_screens.parquet; documented in
paper_v8.tex §6.4 (Coordination).
Interpretation¶
The screens land clean for the differential question. If the post-policy price increase came from a new coordination shock among the surviving SME pool, both screens should intensify post-cutoff — Conley realized share should rise, Bajari-Ye ratio should rise. Neither does. The non-pharma Conley realized share is essentially unchanged (−0.1 pp); the pharma realized share falls by 3.2 pp; both Bajari-Ye ratios fall. The differential-coordination explanation for the exclusion-dominant decomposition is rejected.
What the screens do not say. They do not show that baseline clustering is absent. It is not: realized shares are 1.6–1.9× the null mean in non-pharma and 1.7× in pharma. The IPV-clock interpretation of drop-outs is not literally the dominant strategy played by every bidder; what the screens do is rule out the specific threat that the post-policy environment generated more coordination, which is the threat that would mechanically produce the exclusion-dominant decomposition through bid suppression rather than through the type-cost gradient.
Confidence: yellow. The differential null is clean and load-bearing for H:no-collusion-confound; the screens come back invariant or falling post-cutoff. The reading remains yellow rather than green because (i) the screens were designed for a different institutional context (US highway, Italian roads) and the BEC adaptation requires interpretation of test-statistic baselines, (ii) residual baseline clustering remains a limitation acknowledged in the paper, and (iii) the screens detect first-moment clustering, not all conduct deviations from the IPV-clock benchmark.
Follow-ups¶
- Buyer-level heterogeneity
(
v7-jpube-tight/scripts/62_buyer_heterogeneity.R): if differential coordination (if any) is concentrated in specific buyer types, the decomposition could still hold conditional on buyer characteristics. - Within-firm bid-pattern stability: a within-firm DiD on bid moments would falsify the maintained IPV-clock reading at the individual-bidder level. Not yet specified as a standalone AN.
- Pair-classification AN: the pair-classification result deserves its own AN-page summary; currently bundled in this page's design.