Skip to content

Replication

Software requirements

Component Version
R 4.5+
fixest 0.12+
data.table 1.15+
arrow 14+
duckdb 0.10+
ggplot2 3.5+
scales 1.3+
Rcpp / RcppArmadillo (BNE Monte Carlo)
interval (Turnbull NPMLE)
boot (cluster bootstrap)
kdensity / np (GPV kernel inversion)

Data

The structural sample is derived from the BEC (Bolsa Eletronica de Compras) administrative microdata, accessed through a research agreement with the Secretaria da Fazenda e Planejamento do Estado de Sao Paulo (Sefaz-SP). The raw extract is not publicly redistributable under the access agreement.

Sample Window Auctions Bids
Full DiD sample Jan 2016 -- Dec 2019 832,984 ~3.7M
Structural sample (Pregao, Group 65) Sep 2016 -- Aug 2019 97,993 297,967

Data access

The raw extract is not publicly redistributable. Researchers interested in replication should contact the author. Aggregated cell-level data, the structural-sample auction-level cleaned residuals, and replication code in R are available on request and at the project replication repository (URL on acceptance).

Pipeline

The full v5 structural analysis runs from a single master script:

Rscript scripts/00_master.R

Sequence of subprocesses:

Script Purpose
01_clean.R CSV → parquet conversion, variable creation, sample filters
02_did.R Reduced-form DiD against 76 never-treated product groups (Appendix B)
10_uh_decomposition.R Method-of-moments σa2/σe2 + Krasnokutskaya shrinkage
11_fc_estimation.R Type-specific Fck via all-bidders ECDF + Turnbull NPMLE
12_gpv_convite.R Cross-modality GPV inversion on Convite first-price sealed-bid auctions
13_entry_rates.R Stratum-specific Poisson arrival rate calibration
14_bne_simulation.R Bayes-Nash counterfactual Monte Carlo (B = 2,000) over scenarios S1, S2, S3, V0--V3
15_bootstrap.R Cluster bootstrap at the auction level (Bbs = 500)
16_welfare.R DWLalloc + MCPF arithmetic, λ-grid, Saez--Stantcheva weights
17_robustness.R Strict invariance, Fc regime, kernel bandwidth, sample filter, window
03_tables.R LaTeX tables (threeparttable + booktabs)
04_figures.R PDF figures (grayscale, cairo)

Output structure

v5-jpube/output/
├── tables/             # threeparttable + booktabs
│   ├── tab_v3_bne_decomp.tex
│   ├── tab_v3_apv.tex
│   ├── tab_v3_strict_invariance.tex
│   ├── tab_v3_sensitivity_fc.tex
│   ├── tab_v3_decomp_grid.tex
│   ├── tab_did_structural_bridge.tex
│   ├── tab_v3_bootstrap_ci.tex
│   ├── tab_v3_entry_cost.tex
│   ├── tab_v3_welfare.tex
│   ├── tab_v3_welfare_ci.tex
│   ├── tab_welfare_ranking_lambda.tex
│   ├── tab_v3_window_sensitivity.tex
│   ├── tab_v3_filter_sensitivity.tex
│   ├── tab_v3_bandwidth_grid.tex
│   ├── tab_v3_pregao_fc_summary.tex
│   ├── tab_v3_turnbull_fc.tex
│   ├── tab_v3_uh_variance.tex
│   ├── tab_v3_primitive_invariance.tex
│   ├── tab_v3_uh_invariance.tex
│   ├── tab_v3_cross_modality.tex
│   └── tab_preview.tex
└── figures/            # cairo_pdf, grayscale
    ├── fig_v3_decomposition.pdf
    ├── fig_v3_entry_insurance.pdf
    ├── fig_v3_welfare_forest.pdf
    ├── fig_v3_welfare_weight.pdf
    ├── fig_v3_cross_modality_uh.pdf
    ├── fig_v3_bne_prices.pdf
    ├── fig_v3_pregao_fc_by_stratum.pdf
    ├── fig_v3_pregao_ht_bands.pdf
    ├── fig_v3_optimal_preference.pdf
    ├── fig_v3_gelbach_waterfall.pdf
    └── fig_eventstudy_prices.pdf

Manuscript compilation

cd v5-jpube/manuscript
pdflatex paper_v5.tex
bibtex paper_v5
pdflatex paper_v5.tex
pdflatex paper_v5.tex
# online appendix
pdflatex online_appendix.tex
bibtex online_appendix
pdflatex online_appendix.tex
pdflatex online_appendix.tex

The manuscript uses the elsarticle document class and follows JPubE submission formatting.

Technical notes

  • Memory management. Pipeline scripts run as separate R subprocesses to prevent OOM on systems with 16 GB RAM. The Krasnokutskaya shrinkage step holds the full bid-level table only briefly; intermediate state is cached as .rds at /tmp/p2_v3_*.rds.
  • Parquet cache. First run reads the 6.4 GB CSV file and creates a parquet cache (~73 columns of the original 373). Subsequent runs load from parquet (~5 seconds vs. ~5 minutes).
  • DuckDB for parquet workflows. Read/filter/join/aggregate operations on the parquet cache use DuckDB with PRAGMA threads=12; PRAGMA memory_limit='14GB'. Polars/pandas only as hand-off to model fitting.
  • BNE Monte Carlo. B = 2,000 simulated auctions per (class × scenario) cell. The cluster bootstrap (Bbs = 500) is the dominant compute cost; runs in ~15 minutes on a 12-thread machine.
  • Thread configuration. Both fixest and data.table use 12 threads by default (deferring 2 threads for the OS / IDE). Adjust in scripts/00_setup.R if running on a machine with fewer cores.