Manuscript¶

Download¶

Submission-ready version, April 2026.

Contribution¶

The paper makes three contributions to the literature on procurement-preference rules.

Methodology — within-auction structural decomposition from a clean legal trigger. The paper combines five conditions rare in the procurement-preference literature: a total-exclusion legal trigger (clean separation of regimes), a Pregao drop-out auction format (point-identification of cost distributions following Haile and Tamer 2003), an auction-level unobserved-heterogeneity correction (Krasnokutskaya 2011), an observed-equilibrium-entry treatment in the spirit of Athey, Coey, and Levin (2013) and Athey and Seira (2011), and a Saez and Stantcheva (2016) welfare-weight identity. Each component is established; the combination delivers a within-auction decomposition of an SME procurement preference where the partial-preference structural literature identifies the two margins jointly with cross-equation constraints rather than from a single legal trigger.
Endogenous SME entry as a welfare object, not a side issue. A fixed-pool counterfactual — the standard reduced-form shortcut — misstates the realized policy cost by roughly 50--60%. The participation response is the government's partial offset to a larger latent within-auction shock, not the source of the markup.
A market-design statement, not a universal ranking. A 10% price preference welfare-dominates the set-aside in thick, standardized markets across both the main and the strict-invariance specifications. In thin, heterogeneous markets the comparison depends on how entry selection is modeled. The bifurcation identifies precisely where equilibrium-selection vs. primitive-invariance treatment of the protected pool is first-order, and converts a sensitivity check into a substantive claim about where bidder exclusion delivers proportional returns and where it does not.

Institutional Background¶

Brazil's federal SME statute (Complementary Law 123/2006, amended 2014) defines micro- and small enterprises by annual gross revenue (R$360,000 and R$4,800,000 ceilings, respectively) and mandates exclusive SME tenders for items valued at R$80,000 or less. Public buyer units (PBUs) can avoid the SME-only default only by formally justifying the departure to the audit courts — a costly and scrutinized process. State-level adherence to SME-only tendering for eligible items rose from roughly 13% to nearly 70% in the years following the 2014 amendment.

Since 2005, all public procurement in Sao Paulo passes through Bolsa Eletronica de Compras (BEC), a centralized electronic reverse-auction platform that handled approximately R$13 billion (US$3.7 billion) in 2019 alone and has cumulatively moved more than R$105 billion (US$30 billion) since implementation, spanning 860,000 purchase offers and 4.8 million item-level transactions.

Group 65 (medical, dental, and hospital supplies) was a singular exception to the 2014 rule. Between August 2014 and February 2018, the state government and the Tribunal de Contas (TCE-SP) operated under a joint interpretation that medical supplies were strategic goods requiring open competition. In November 2017, the Procuradoria-Geral do Estado (PGE-SP) issued a legal opinion reversing the exception on isonomy grounds, arguing that no substantive feature of medical supplies justified procedural distinctions from other product groups. Effective March 2018, opt-out costs extended to Group 65.

Groups	Before March 2018	After March 2018
Group 65 (switched)	Opt-out costs = 0	Opt-out costs > 0
Others (always treated)	Opt-out costs > 0	Opt-out costs > 0

The change was triggered by a legalistic reinterpretation of the isonomy principle inside PGE-SP, not by anything happening in the Group-65 procurement market. Public documentation turns up no sign that the decision was a response to price levels, competition patterns, or supplier complaints specific to Group 65 in 2017.

Data and Sample¶

Feature	Detail
Source	BEC administrative records (Sefaz-SP)
Coverage	All standardized goods procurement in Sao Paulo state
Window (full)	January 2016 -- December 2019
Window (structural)	September 2016 -- August 2019 (18 months on each side of the March 2018 cutoff)
Treated group	Group 65 (medical/hospital supplies, ~27% of transactions)
Control set (DiD)	76 never-treated product groups
Auction format (structural)	Pregao only (descending-clock electronic English-reverse)
Structural sample	297,967 firm-auction observations, 97,993 distinct auctions
Pre vs Post	48,740 vs 49,253 auctions (balanced)
Bid filter	c_ε = b/p^ref ∈ (0, 3]; n ≥ 2 firms per auction
Type flag	SME / non-SME from BEC registry, validated against historical bidding
Pharma flag	CADMAT classes 6531, 6532, 6536, 6581 (CMED-regulated pharmaceuticals and biologicals)

The structural narrowing to Pregao is methodological: the descending-clock format admits point identification of cost distributions from drop-out bids (Haile and Tamer 2003), which the smaller-value Convite first-price sealed-bid format does not afford in the same form. Convite is reserved for a cross-modality consistency check via Guerre, Perrigne, and Vuong (2000) inversion.

Entry response (Pregao)¶

	Non-pharma Pre	Non-pharma Post	Pharma Pre	Pharma Post
SMEs per auction	0.94	1.87	0.55	1.22
Non-SMEs per auction	2.68	1.50	2.61	1.66

SMEs roughly double in both classes; non-SMEs fall sharply. The set-aside changes both the admissible pool and the equilibrium participation response.

Empirical Strategy¶

The structural model reads the BEC platform's Pregao auction as an asymmetric independent-private-values (IPV) environment. Bidders draw costs from type-specific distributions F_c^SME and F_c^¬ and compete under a descending clock; under IPV the weakly dominant strategy is to exit when the running clock price reaches one's own cost (Haile and Tamer 2003). Drop-out prices of losers therefore point-identify F_c^k, without inverting a bid function or imposing parametric structure on bidding.

Three substantive restrictions discipline the model:

Asymmetric IPV. Costs are drawn independently across bidders; conditional on type, c_i ~ F_c^k. SME and non-SME distributions are not constrained to coincide.
Multiplicative auction-level heterogeneity. Normalized log-bid decomposes as y_it = a_t + e_it, with auction-level scale a_t common to all bidders in auction t (Krasnokutskaya 2011). A best-linear-predictor shrinkage strips a_t out of every log-bid; the cleaned residuals enter all downstream structural estimation.
Set-aside with equilibrium SME selection. Under the open regime both types enter; under SME-only, only SMEs are admissible. Within a short window around the cutoff, the underlying technology of supply is stable, but the equilibrium distribution of active SMEs under the restricted regime may differ from its pre-regime counterpart (Athey and Seira 2011 reading).

A Turnbull NPMLE that treats the winner formally as left-censored at c₍₂₎ dispenses with the multiplicative-heterogeneity assumption altogether and corroborates the baseline numbers within 16% of the total effect across classes.

The decomposition¶

Three counterfactual scenarios anchor the exercise:

S₁ — open regime at the pre-period pool (baseline)
S₂ — SME-only at the pre-period SME pool (within-auction component isolated)
S₃ — SME-only at the observed post-period SME pool (full policy effect with endogenous entry)

The simulated total effect on the second-order statistic decomposes arithmetically as

\[\underbrace{p_{S_3} - p_{S_1}}_{\text{simulated total}} = \underbrace{(p_{S_2} - p_{S_1})}_{\text{within-auction (sheltered bidding)}} + \underbrace{(p_{S_3} - p_{S_2})}_{\text{participation-margin}}.\]

The within-auction component captures equilibrium price formation under reduced competition; under ascending-clock IPV, the dominant strategy is exit-at-cost regardless of pool composition, and the within-auction component is mechanically a property of which order statistic is realized at each pool composition. The decomposition is a counterfactual-simulation accounting under the modeling assumptions stated, not a separation of independently identified structural channels.

What the model does not attempt¶

Five extensions are bracketed below or in robustness rather than imposed as the central exercise: (i) entry is taken as observed Poisson arrival rates, not solved jointly with bidding under a free-entry condition; (ii) implied κ^k are zero-profit margins at the entry margin, not structurally identified fixed costs; (iii) procurement quantity is treated as inelastic; (iv) the 10% preference V3 is computed under Vickrey-equivalent translation, with the FPSB equivalent under Maskin and Riley (2000) bracketed; (v) asymmetric IPV is assumed.

Software and Estimation¶

Component	Detail
Language	R 4.5
Reduced-form FE	`fixest::feols()` with lean estimation
Cost distributions	All-bidders ECDF on UH-cleaned residuals; Turnbull NPMLE robustness
UH correction	Method-of-moments + best-linear-predictor shrinkage (Krasnokutskaya 2011)
Cross-modality check	GPV (2000) inversion on Convite first-price sealed-bid auctions
BNE simulation	Monte Carlo, B = 2,000 simulated auctions per scenario
Standard errors	Cluster bootstrap at the auction level, B_bs = 500
Welfare arithmetic	DWL_alloc + λ · Δ_gov with λ ∈
Welfare weights	Saez--Stantcheva (2016) generalized social-marginal-welfare-weight identity
Data handling	DuckDB / `data.table` / `arrow` (parquet cache)
Figures	`ggplot2` with `cairo_pdf`, grayscale theme