Findings — Cheap Signals, Costly Proof¶
Intuition (plain-language)
Catching bid-rigging is expensive: the textbook tests need detailed bid-by-bid data that procurement systems rarely keep. The economic idea organizing these findings is a cost-of-evidence one — use a cheap, always-available signal (which firms bid constantly without ever winning) to decide where to spend costly forensic effort, not to decide who is guilty. The findings below separate what that cheap signal can credibly buy (a prioritized pool to investigate) from what it cannot (a verdict on any individual firm).
This page is the curated index of what we have learned from the project's
own analyses. Each finding is a claim about the world — the kind of
statement that would go in the paper — and it may rest on several
AN-NNN analyses, cited in its Sources footer. The detail, design, and
caveats behind each number live on those analysis pages; this page is the
synthesis.
This page is the directory of conclusions. Use docs/analyses/
to look up the per-result design, raw numbers, and full result tables;
use this page to scan what we believe and at what confidence.
How to read the confidence tags¶
A traffic-light convention runs across both scales: 🟢 = strongest confidence, 🟡 = middle, 🔴 = weakest. The meaning of each color depends on whether the claim is empirical or interpretive.
Empirical findings — the color reflects the source of confidence, not the size of the effect:
- 🟢 Replicated — the finding appears in multiple independent samples or studies that agree in direction and rough magnitude.
- 🟡 Single source — one solid study, or one of our own runs, with no independent replication yet.
- 🔴 Provisional — one descriptive cut that is parser-dependent, sensitive to sample definition, or flagged with a known caveat. Read the qualifier before quoting.
Interpretations — parallel scheme:
- 🟢 Strong — multiple converging lines of evidence; alternatives have been considered and rejected.
- 🟡 Plausible — consistent with the evidence but other readings remain open.
- 🔴 Speculative — suggested by the data but unverified; flagged for follow-up rather than relied on.
For Cheap Signals, Costly Proof, the natural 🟢 promotion criterion is replication on a non-BEC-SP procurement panel (e.g. ComprasNet federal, or another state's e-procurement system). Until that exists, findings here rest on the project's own BEC 2009–2019 runs and start at 🟡.
Findings overview¶
All findings below are scaffolds at 🟡 (single-source own-project estimates) or 🟢 (load-bearing for the paper's framing) and link to the AN pages that carry the design + numbers. Promotion to 🟢 (when not already there) would require independent replication on a non-BEC procurement panel.
Award-layer triage scope (empirical + interpretive)
- CADE-adjacent cobidders concentrate in the FL14 stratum — 🟡 131/193 cobidders recovered inside the FL stratum; rank construction = persistent zero-win participation.
- The FL ranking is null on direct CADE defendants — by design — 🟢 AUC ≈ 0.49–0.50; the predicted finding under the loser-side scope.
Audit discipline (empirical)
- Cobidder concentration survives exposure, leakage, and timing audits — 🟡 result holds under sham FL placebo, out-of-fold leakage audit, and strict timing.
Cobidder economic profile (empirical)
- FL cobidders are operationally distinct from other frequent losers — 🟡 broader buyer deployment, more concentrated product portfolios, closer to legal cartel anchors, bid patterns consistent with credible losing roles; distinctness survives volume matching (AN-041).
Sequential architecture / cost of evidence (empirical + interpretive)
- Award-layer gatekeeping cuts the bid-microdata pool by 83% — 🟡 131/193 cobidders recovered after the cut; the paper's operational headline.
- Joint award + bid scoring is the full-observability upper bound — 🟡 the two layers are complementary, not substitutes.
Price scope (interpretive)
- Price coefficient sign reverses — scope, not damages — 🟡 the sign reversal defeats the damages reading and supports a scope interpretation.
Open items for this page¶
- 🟢 promotion path is replication on a non-BEC dataset. None of the planned analyses provides that on its own; that is a separate data-acquisition decision (ComprasNet federal, or another state's e-procurement platform).
- Numeric headlines are bound to
values.texmacros (e.g.,\valAUCFLfirm,\valAUClogtc,\valCobidders,\valPrecTHFivehu) auto-generated byscripts/99_make_paper_values.R. Regeneratevalues.texand the analysis-index after each upstream re-run; the finding and AN pages quote the macros' current canonical values.
Generated index¶
All 7 finding pages, auto-generated from the body of each docs/findings/<slug>.md via scripts/render_indexes.py. Confidence and one-line summary extracted from the page lede.
| Conf. | Finding | Summary |
|---|---|---|
| 🟢 | CADE-adjacent cobidders concentrate in the FL14 stratum | In São Paulo's BEC procurement platform (2009–2019), always-loser firms that bid alongside direct CADE defendants in adjudicated cartel environments concentrate disproportionately inside the FL14 stratum identified by… |
| 🟡 | FL cobidders are operationally distinct from other frequent losers | Inside the FL14 stratum, cobidders look different from other always-losers along four economically meaningful dimensions, with effect sizes (Cohen's d) in the 0.3–1.0 band — large in social-science terms. |
| 🟡 | Cobidder concentration survives exposure, leakage, and timing audits | The cobidder concentration result from CADE-adjacent cobidders concentrate in the FL14 stratum survives the three audit disciplines introduced in §4.2 of the manuscript — though the operational deployment metrics are… |
| 🟢 | The FL ranking is null on direct CADE defendants — by design | The frequent-loser ranking, evaluated against the 47 direct CADE defendants in 2009–2019 procurement-cartel adjudications, returns AUC = 0.491 [0.461, 0.520] — indistinguishable from random. The null is the predicted… |
| 🟡 | Joint award + bid scoring is the full-observability upper bound | On the cobidder target, a classifier that uses both the award- layer score (FL14 or continuous log_tc) and the Imhof-style bid- distribution features achieves the highest AUC: 0.955 [0.943, 0.967] with FL14 + Imhof, and… |
| 🟡 | Award-layer gatekeeping cuts the bid-microdata pool by 83% | Used as a gatekeeper, the FL14 ranking cuts the bid-microdata pool required for the forensic stage by 83% while still recovering 131 of 193 adjudicated cobidders (68%). This is the cost-of-evidence headline of the… |
| 🟡 | Price coefficient sign reverses — scope, not damages | The price coefficient associated with the FL margin reverses sign when overlap discipline is applied, in a pattern that defeats the naive damages reading and supports a scope interpretation: prices tell us where the… |
Confidence legend. 🟢 replicated / load-bearing for paper framing; 🟡 single-source own-project estimate; 🔴 provisional, kept for the record only. Promotion to 🟢 generally requires independent replication on a non-BEC procurement panel.