H:price-scope-sign-reversal — Price evidence carries scope information, not damages

The manuscript reports a price sign reversal when the cobidder margin is introduced. The hypothesis is that the price result should be read as scope information — telling us where the loser-side ranking applies — rather than as a damages calculation. A naive damages reading would predict a stable, signed effect; the scope reading predicts that the sign and magnitude depend on the validation target.

Intuition (plain-language)

The presence of frequent losers in a tender is associated with higher negotiated prices on average (+6.4% — a damages-like reading that a careless reader might interpret as the size of the cartel overcharge). But once we restrict comparisons to comparable cells under matched ATT weighting, the sign FLIPS to negative (−9.7%). The hypothesis is that this sign reversal is informative and disciplining: the price evidence is scope information about WHERE the loser-side ranking applies, not a damages estimate. The paper deliberately treats price evidence as secondary corroboration, and the sign reversal is what justifies that disciplined reading.

Evidence strength: Partial (strongly supported). Eight lines of evidence document the sign-reversal pattern that defeats a naive damages reading — six describe the pattern, two describe the within-cell decomposition behind it (descriptive scope evidence, not mechanism identification): (i) Headline three-spec progression (AN-037): broad +0.064 (p=0.003) → overlap unweighted +0.044 (p=0.035) → overlap ATT −0.097 (p = 1.7 × 10⁻¹⁰). Clean structural sign flip. (ii) Within-overlap subgroup decomposition (AN-037): negative coefficient survives both modalities (Convite −0.099, Pregão −0.098), all 4 PBU-size quartiles, tender-value Q1–Q3, items without direct CADE (−0.097), and items without cobidders (−0.108); only tender-value Q4 stays positive (+0.041) — the scope boundary. (iii) Item-group decomposition (AN-038): most groups flip from positive baseline to negative ATT (12, 13, 14, 29); groups 37 and 39 stay negative across specs (−0.126 / −0.103 ATT); group 10 stays positive (+0.063 ATT, n.s.). Predictably structured, not uniform. (iv) Cell-by-cell negative audit (AN-038): period 2009–2013 −0.119; PBU-size Q4 −0.126; tender-value Q1 −0.123; item-group 37 −0.182. The negative concentrates in routine, low-value, large-PBU cells — the operational deployment target. (v) RDD + DiD support (AN-019, AN-020): RDD coefficients stable across bandwidths (+0.043 / +0.055 / +0.060); McCrary ratio 0.94 (no bunching); DiD around 2018 decree null. (vi) Modal asymmetry (AN-016, AN-022): Pregão 2.45× Convite in binary price spec; joint specs flip binary FL14 sign. (vii) Selection component (AN-039): among non-treated items, mean log-price rises monotonically from 1.35 (lowest FL-share quintile) to 6.93 (highest); Δ = 5.58 log-points, full-FE FL-share coefficient +3.55 (SE 0.23). Frequent losers are concentrated in structurally high-price cells — the naive positive coefficient is selection, not damages. (viii) Within-cell component (AN-040): within overlap cells FL presence is associated with +0.507 log-bidders (≈66% more) and a winner bid −0.048 closer to reference; the winner-vs-reference association vanishes once bidder count is controlled. Selection dominates where bidders are few, the within-cell association where they are many. Descriptive associations, not a mechanism. Promotion to 🟢 (Confirmed) requires non-BEC replication of the sign-reversal pattern — see the H8 page section on why within-data evidence does not satisfy the bar.

Theory

Price effects from cartel screens are typically interpreted as damages estimates \citep{harrington2008detecting,porter1993detection}. Under incomplete observability and a triage-not-classification framing, prices play a different role: they inform the scope of the screen — which markets, modalities, or sub-periods the loser-side ranking applies to — rather than the size of the cartel overcharge \citep{chassang2022robust}.

Prediction

The sign of the price coefficient associated with the FL margin should:

• flip when the validation target shifts (direct defendants vs cobidders);
• vary in magnitude across modality (Convite vs Pregão) consistently with the loser-side scope rather than with a single damages parameter.

Competing prediction

Damages reading. A single signed price coefficient that is stable across targets and modalities, supporting a conventional damages interpretation. The hypothesis rejects this reading.

Case evidence

The price-scope distinction is consistent with CADE's separation of liability findings from damages quantification in the procurement-cartel record (administrative-law standard).

Empirical test

• Sample: BEC items with FL participation.
• Outcomes: log negotiated price; price ratio to reference; markup proxy.
• Specifications:
• target = direct defendants: price coefficient with direct-defendant presence;
• target = cobidders: price coefficient with cobidder presence;
• modality split: same specifications by Convite vs Pregão.
• Identification: cross-target and cross-modality sign comparison; the test rejects the damages reading if signs/magnitudes differ systematically.

Data requirements and limitations

Requires reference prices and the FL-presence indicator at the item level. Limitation: the sign-reversal interpretation depends on the assumption that the loser-side scope is correctly identified; the price reading is descriptive and supportive, not causal.

Evidence

Analysis	Bearing	Status	Key takeaway
AN-019 (RDD price + overlap)	Direct	done	Baseline +6.36% → overlap −9.72% → PS −30.67%
AN-020 (DiD around decree)	Supports	done	Callaway-Sant'Anna ATT +0.014 (SE 0.039); stacked DiD −0.006 — null
AN-016 (modal AUC asymmetry)	Supports	done	Pregão 0.952 vs Convite 0.816 — opposite direction to inst. theory
AN-022 (modal falsification)	Supports	done	Pregão/Convite price ratio 2.45×; joint specs flip binary sign
AN-037 (sign-reversal headline + subgroups)	Direct	done	Broad +0.064 → overlap ATT −0.097 (p = 1.7e-10); both modalities, all PBU-Q, tender-value Q1-Q3 all negative
AN-038 (item-group + cell audit)	Direct	done	Item-group 37 −0.126 (stays negative); group 10 +0.063 (stays positive); period 2009-2013 −0.119 — predictably structured
AN-039 (selection component — Test 1)	Direct	done	Non-treated log-price rises Q1 1.35 → Q5 6.93 (Δ 5.58); full-FE FL-share coef +3.55 (SE 0.23) — frequent losers concentrate in high-price cells, so the naive positive coefficient is selection, not damages
AN-040 (within-cell component — Test 2)	Direct	done	Within overlap cells FL presence → +0.507 log-bidders (≈66%) and winner −0.048 closer to reference; association runs through bidder count. Selection dominates sparse tenders, within-cell association dominates dense. Descriptive, not mechanism identification

Open tests

• Triangulation of price scope with bid-level dispersion patterns (AN-031 bid- level moments).
• Robustness to reference-price construction (administrative vs market-implied).
• Item-group 37 deep-dive: which product category produces the consistently negative coefficient (AN-038)?
• Tender-value Q4 deep-dive: why does the positive coefficient persist in high-value tenders?
• Cross-modality + cross-quartile × period interaction (full decomposition).

Why not 🟢 Confirmed?

The within-data sign-reversal evidence is strong. The headline progression (broad +0.064 → overlap ATT −0.097, p = 1.7 × 10⁻¹⁰) plus the within-overlap subgroup decomposition (negative in 14 of 15 relevant cells) plus the item-group structure (predictable heterogeneity, not noise) is the cleanest H8 evidence the paper has.

Three obstacles to Confirmed:

1. Specification-dependent identification. The sign reversal is produced by overlap discipline + ATT weighting on top of item + year + PBU fixed effects. A reviewer could argue that the ATT reweighting is itself a specification choice that drives the sign — and that an alternative weighting scheme would give a different result. The data do not adjudicate which weighting is the "correct" causal specification. The scope reading is the interpretive answer to this objection: the price coefficient is not a damages parameter for ANY weighting; it is scope information that varies with weighting because the underlying price-formation process varies across cells.

2. No causal identification of the FL margin. The headline coefficient is a within-cell association under FE + ATT weights, not an estimate identified by an instrument or a sharp design. A clean instrument would help determine whether the sign reversal is structural or specification-driven. The cap-RDD (AN-019) and the 2018- decree DiD (AN-020) come close, but the cap-RDD is positive (4–6%) and the DiD is null — neither delivers a causal sign-reversal estimate.

3. Same-data limitation. All evidence comes from BEC × CADE; the sign-reversal pattern could be a feature of BEC's price-recording convention combined with CADE's adjudication selection. Non-BEC replication of the three-spec progression would settle whether the sign-reversal generalizes.

The mr-frequent reading: H8 is scope information, not a structural causal claim. The within-data evidence is strong enough to defeat the naive damages reading, which is what the paper claims. Promotion to 🟢 would require treating H8 as a causal claim, which the paper explicitly avoids. Stays at Partial (strongly supported): the descriptive sign-reversal is robust; the causal interpretation is deliberately limited to scope, not damages.