Skip to content

AN-001: DiDiR price effect

Intuition (plain-language)

The cleanest reduced-form evidence that the set-aside raises prices: when medical supplies were forced back into SME-only tendering in March 2018, their prices rose about 10–11% relative to groups that never changed. The estimate is stable across 6/12/18-month windows — what you want from a real policy effect rather than noise. The v8 paper uses this only to pin down timing and sign; the magnitude comes from the structural decomposition.

Reduced-form motivation layer

The numbers below are from the v1–v4 reduced-form DiDiR pipeline (scripts/02_analysis.R + companions), which the v8 manuscript carries as motivation in §1 but does not headline. The canonical v8 result is the structural counterfactual decomposition — see AN-010 (decomposition) and AN-011 (welfare arithmetic).

Question

Does removing the SME-only restriction lower negotiated procurement prices in switched group 65? The March 2018 PGE-SP reinterpretation makes group 65 identical to all other groups in being subject to SME-only tendering by default. Comparing the pre-period (group 65 open, others SME-only) to the post-period (all groups SME-only) yields the open-vs-SME-only price effect, identified by DiDiR.

Design

  • Sample: completed items (oc_item_status == 1) in the BEC parquet cache for Sep 2016–Aug 2019 (18-month window), with 6/12-month sub-windows.
  • Variation: group-65 indicator \(\times\) pre-March-2018 indicator.
  • Specification: \(y_{pigt} = \eta_i + \gamma \text{Pre}_t + \beta\,(g65 \times \text{Pre}_t) + x'\delta + \varepsilon\) with item FE \(\eta_i\), log-quantity and sealed-bid controls, item-clustered SEs.
  • Outcomes: log price (primary).
  • Identification threats: cost shock specific to group 65 around March 2018; ruled out by placebo (AN-004) and HonestDiD (AN-006).

Results

The canonical reduced-form benchmark is Table 2 of the v8 manuscript: the \(g65 \times \text{Pre}\) coefficient on \(\log p^{\mathrm{final}}\) with PBU controls, item and month fixed effects, and item-clustered SEs. It is stable across windows and implies a 10–11% price movement at the 18-month benchmark.

Window \(g65 \times Pre\) SE N
6m −0.148 (0.014) 219,535
12m −0.108 (0.013) 439,054
18m −0.113 (0.012) 649,714

Item and month FE; auction-format and log-quantity controls; PBU controls; item-clustered SEs. Table 2 in paper.pdf.

Earlier reduced-form pipeline

The v1–v4 DiDiR pipeline (scripts/02_analysis.R → tab_prices.tex) produced slightly larger magnitudes for the same windows (18m: −0.131 base / −0.133 PBU-FE, item-clustered SEs ~0.009). The v8 benchmark above is the canonical reduced-form number; the difference reflects the benchmark specification (PBU controls and the reference-price treatment used in the structural sample), and is well within the band that supports the timing-and-sign reading.

Interpretation

The coefficient implies ~10–11% lower prices under open competition relative to the SME-only regime. The stability across windows is meaningful — short windows have less power and more pre-treatment seasonality noise, but the estimate stays in a tight band. Modern DiD estimators (Callaway–Sant'Anna, BJS) attenuate the magnitude, consistent with the conservative role assigned to the reduced form: it carries timing and sign, not the structural magnitude.

Reading: the lost-discipline channel (non-SMEs removed from the price-forming pool) is the load-bearing mechanism. The reduced-form DiDiR does not decompose this — see AN-010 for the structural decomposition.

Confidence: yellow. The point estimate is robust across windows, FE structures, alternative clustering (item, group, PBU, two-way; see docs/robustness.md), and tight Lee bounds (AN-005). HonestDiD survives substantial M̄ violations (AN-006). The reading is still single-source — own-project estimate; promotion to green requires either independent replication in another procurement setting or a converging recovery from a different identification source (e.g., the GPV recovery from Convite first-price bids; partly run, not yet documented as an AN).

Follow-ups

  • Cross-modality validation: GPV recovery of cost distribution from Convite first-price bids in the same product cells, compared to the Pregão drop-out recovery (AN-013). Lives in v7-jpube-tight/scripts/38_cross_modality.R.
  • Adherence-rate-by-PBU sensitivity for the fiscal translation (scripts/35_pref_sdid_class.R). Not load-bearing for the reduced-form coefficient, but tightens the welfare arithmetic in AN-011.
  • The quantile-DiD heterogeneity in AN-007 shows the mean coefficient masks a strongly negative effect at low quantiles and a positive effect at \(\tau=0.90\). A within-item conditional-quantile decomposition would isolate where the open-competition price gain comes from.