Analyses — Cheap Signals, Costly Proof¶
This page is the directory of analyses for Cheap Signals, Costly Proof: Award-Layer Evidence Triage in Cartel Enforcement. Each AN-NNN page is anchored by a YAML frontmatter block that drives the generated index below.
This page is the directory of designs and results. For curated claims about the world — sentences that could go in the paper — see Findings. For the testable predictions and their cross-references to AN pages, see Hypotheses.
How to read an AN page¶
Each an-NNN-<slug>.md follows the same skeleton:
- Frontmatter: identifier, status, hypothesis bearing on, script and output path, design block (sample, specification, FEs, etc.).
## Question— the precise research question.## Design— sample, variation, specification, identification threats.## Results— headline number + headline table or figure, with a link to the underlying output.## Interpretation— what the result means; confidence color with justification.## Follow-ups— forward-looking only; analyses not yet run.
Status. pending (scaffolded, not yet run or not yet interpreted) →
done (script run, headline + interpretation written, confidence set) →
stale (superseded by a newer AN or by upstream-data change).
Confidence. 🟢 green (clean identification, large sample, robust),
🟡 yellow (informative with caveats), 🔴 red (unreliable, kept for the
record only). Set only when status: done.
Generated index¶
All 42 AN pages, auto-generated from the YAML frontmatter of each docs/analyses/an-NNN-*.md via scripts/gen_analysis_index.py + scripts/render_indexes.py. The machine-readable form lives at docs/reference/analysis-index.yaml.
| AN | Type | Status | Conf. | Hypothesis | Question |
|---|---|---|---|---|---|
| AN-001 | descriptive | done | 🟢 | cobidder-concentration | How is the persistent-zero-win-participation rank constructed, and what is its distribution across always-los… |
| AN-002 | robustness | done | 🟡 | cobidder-concentration | How does the cobidder AUC change as the IQR threshold is varied, and is the median + 1.5 × IQR cutoff disting… |
| AN-003 | descriptive | done | 🟢 | cobidder-concentration | How are CADE direct defendants and adjudication-anchored cobidders linked to BEC firms via CNPJ root, and wha… |
| AN-004 | descriptive | done | 🟢 | cobidder-concentration | Does the FL14 stratum contain a disproportionate share of CADE-adjudication-anchored cobidders relative to th… |
| AN-005 | placebo | done | 🟢 | exposure-discipline | Does cobidder concentration in the FL14 stratum survive a participation-volume-matched placebo, and how far i… |
| AN-006 | robustness | done | 🟡 | timing-discipline | Does cobidder concentration survive when the FL score is formed strictly before the target window? |
| AN-007 | placebo | done | 🟢 | direct-defendants-null | Does the FL score discriminate direct CADE defendants? It should not — by design. |
| AN-008 | descriptive | done | 🟡 | cobidder-profile-distinct | Within the FL14 stratum, how do cobidders differ from non-cobidder FLs along buyer breadth and operational fo… |
| AN-009 | descriptive | done | 🟡 | cobidder-profile-distinct | Do cobidders inside the FL14 stratum operate in more concentrated product portfolios than non-cobidder FLs, a… |
| AN-010 | descriptive | done | 🟢 | award-bid-complementarity | How does the seven-feature Imhof–Wallimann bid-distribution pipeline perform on the cobidder target, and what… |
| AN-011 | descriptive | done | 🟢 | award-bid-complementarity | Does the continuous log(1+tenders_count) dominate the binary FL14 on the cobidder target? |
| AN-012 | descriptive | done | 🟡 | gatekeeping-cost-of-evidence | What are the in-sample precision@k and lift metrics for the FL ranking used as a forensic gatekeeper? |
| AN-013 | robustness | done | 🟢 | gatekeeping-cost-of-evidence | What are the temporal-holdout precision@k and lift metrics, and how much does the in-sample evaluation inflat… |
| AN-014 | robustness | done | 🟢 | gatekeeping-cost-of-evidence | How much does item-level evaluation leak relative to out-of-fold and temporal-holdout retraining? |
| AN-015 | descriptive | done | 🟢 | award-bid-complementarity | D1 gate diagnostic — does the continuous score dominate FL14 on a harmonized same-sample horse race, and do t… |
| AN-016 | descriptive | done | 🟡 | price-scope-sign-reversal | D2 gate diagnostic — does the FL screen discriminate cobidders better in Convite or in Pregão environments? |
| AN-017 | robustness | done | 🟢 | cobidder-concentration | D3 gate diagnostic — does the continuous score preserve the loser-side thesis without FL14, and what is the i… |
| AN-018 | descriptive | done | 🟢 | direct-defendants-null | D4 gate diagnostic — what share of direct CADE defendants are always-losers, and what is their win-rate distr… |
| AN-019 | descriptive | done | 🟡 | price-scope-sign-reversal | Does the negotiated-price coefficient at the procurement-cap threshold reverse sign when FL14 presence is int… |
| AN-020 | descriptive | done | 🟡 | price-scope-sign-reversal | Does the 2018 procurement decree shift price dynamics differently across modalities, consistent with the scop… |
| AN-021 | robustness | done | 🟡 | cobidder-profile-distinct | Does the "first-time FL" effect on cobidder concentration survive propensity-score matching? |
| AN-022 | placebo | done | 🟡 | price-scope-sign-reversal | Do FL-margin price effects differ by procurement modality, and does the Pregão-only subsample replicate the f… |
| AN-023 | robustness | done | 🟡 | cobidder-concentration | Does the operational mapping from theory (loser-side concentration) to implementation (FL14) survive an expli… |
| AN-024 | descriptive | done | 🟡 | cobidder-profile-distinct | How does the unified mechanism profile (HHI × pairs × heterogeneity quadrants) characterize FL cobidders rela… |
| AN-025 | robustness | done | 🟢 | cobidder-concentration | How does cobidder AUC vary as the FL cutoff sweeps from FL2 through FL100, and is FL14 picking up an arbitrar… |
| AN-026 | robustness | done | 🟢 | cobidder-concentration | Does the cobidder concentration result survive across always-loser sub-populations defined by bid-microdata a… |
| AN-027 | descriptive | done | 🟢 | exposure-discipline | How does AUC behave when the universe and the positive class are systematically varied — does the loser-side… |
| AN-028 | descriptive | done | 🟢 | exposure-discipline | Within the always-loser stratum, are cobidders distinguishable from non-cobidder FLs along dimensions other t… |
| AN-029 | robustness | done | 🟢 | timing-discipline | Does the FL screen preserve discrimination under three progressively-earlier train windows, evaluated against… |
| AN-030 | descriptive | done | 🟢 | timing-discipline | How much do the firms, markets, and procuring buyers in 2017–2019 overlap with those in 2009–2016? Is the out… |
| AN-031 | descriptive | done | 🟡 | cobidder-profile-distinct | Do cobidders display bid-level behavior distinct from non-cobidder FLs, independent of participation volume? |
| AN-032 | robustness | done | 🟡 | cobidder-profile-distinct | Does the quadrant-level heterogeneity (HHI × pairs) of the cobidder profile survive propensity-score matching… |
| AN-033 | descriptive | done | 🟢 | award-bid-complementarity | How significant is the incremental value of the award-layer score added to the Imhof bid-distribution pipelin… |
| AN-034 | descriptive | done | 🟢 | award-bid-complementarity | When deployed sequentially (FL gatekeeper → Imhof forensic stage) vs jointly, how does the cost-of-evidence t… |
| AN-035 | descriptive | done | 🟢 | gatekeeping-cost-of-evidence | Across the full architecture × k × regime grid, what are the recall, precision, and bid-microdata cost trade-… |
| AN-036 | robustness | done | 🟡 | gatekeeping-cost-of-evidence | Are the precision@k metrics stable across cross-validation folds, or do they depend on a specific random spli… |
| AN-037 | descriptive | done | 🟢 | price-scope-sign-reversal | How does the FL-margin price coefficient transform across baseline → overlap-cell → ATT specifications, and d… |
| AN-038 | descriptive | done | 🟡 | price-scope-sign-reversal | At the item-group and operating-cell level, where does the negative FL-price coefficient hold and where does… |
| AN-039 | descriptive | done | 🟢 | price-scope-sign-reversal | Do cartels with cover bidders endogenously select into cells where the underlying (non-treated) price level i… |
| AN-040 | descriptive | done | 🟢 | price-scope-sign-reversal | Within overlap cells, does FL presence depress the observed winner bid relative to the reference price? Does… |
| AN-041 | descriptive | done | 🟢 | cobidder-profile-distinct | Does the within-FL distinctness of cobidders (AN-028 participation dimensions, AN-031 bid-level gap-to-winner… |
| AN-042 | descriptive | done | 🟡 | cobidder-profile-distinct | Are cobidders distinct from non-cobidder FLs on bid TIMING (revision intensity, inter-bid interval, last-bid… |
Status legend. done = analysis run and interpretation written; pending = scaffolded only; stale = superseded.
Confidence legend. 🟢 green (clean identification, robust); 🟡 yellow (informative with caveats); 🔴 red (kept for the record, not load-bearing).