Skip to contents

Reproducing the SUBE Paper

Source: vignettes/paper-replication.Rmd
paper-replication.Rmd

1. What this vignette replicates

The 2024 SUBE paper computes Leontief multipliers, elasticities, and regression coefficients from WIOD international supply-use tables. This vignette walks through the end-to-end reproduction of the paper’s raw supply, use, and net-supply matrices using only the functions exported by sube. All code chunks are shown eval = FALSE so the vignette builds cleanly on CRAN and in CI without requiring the ~4 GB WIOD data archive.

This vignette is specific to WIOD data because the paper itself is WIOD-based. The pipeline from build_matrices() onward is identical for any SUT source in canonical format - the WIOD-specific steps are the import (section 3) and the correspondence tables (section 4).

The gated test at tests/testthat/test-replication.R asserts bit-level numerical equivalence for a representative subset (AUS, DEU, USA, JPN for 2005). Running it locally is covered in section 9.

Tip. For a one-call equivalent of the full chain - import through multipliers in a single function - see run_sube_pipeline() and the Pipeline Helpers vignette.

2. Obtaining the WIOD data

Download the WIOD international supply-use tables and companion files from Eurostat / the WIOD website. The expected layout under $SUBE_WIOD_DIR:

$SUBE_WIOD_DIR/
  International SUTs domestic/
    Int_SUTs_domestic_SUP_2005_May18.csv
    Int_SUTs_domestic_USE_2005_May18.csv
    ...
  Correspondences/
    CorrespondenceCPA56.dta
    CorrespondenceInd56.dta
  GOVAcur/      GO_{country}_{year}.dta
  EMP/          EMP_{country}_{year}.dta
  CO2/          CO2_{country}_{year}.dta
  Regression/data/   {country}_{year}.csv  (paper ground-truth W matrices)

Set the environment variable before running the gated test:

export SUBE_WIOD_DIR=/path/to/wiod

3. Importing the domestic block 

sut <- import_suts(file.path(Sys.getenv("SUBE_WIOD_DIR"),
                             "International SUTs domestic"))
domestic <- extract_domestic_block(sut)
domestic[REP == "AUS" & YEAR == 2005][1:3]
#> Returns a sube_domestic_suts long table with REP == PAR rows only.

4. Aggregation via correspondence tables 

root <- Sys.getenv("SUBE_WIOD_DIR")
cpa_map <- data.table::data.table(haven::read_dta(
  file.path(root, "Correspondences", "CorrespondenceCPA56.dta")))
ind_map <- data.table::data.table(haven::read_dta(
  file.path(root, "Correspondences", "CorrespondenceInd56.dta")))
data.table::setnames(cpa_map, "CPAagg", "CPA_AGG")
data.table::setnames(ind_map, "Indagg", "IND_AGG")

The CPA56 and Ind56 maps project the raw 56-sector NACE codes onto the 22-sector aggregation used throughout the paper.

5. Computing Leontief multipliers and elasticities 

results <- compute_sube(domestic, cpa_map, ind_map)
results$tidy[COUNTRY == "AUS" & YEAR == 2005 & measure == "multiplier"][1:5]

compute_sube() builds the aggregated S and U matrices, inverts (I - A), and emits tidy multiplier and elasticity tables.

6. Applying the paper’s outlier treatment 

models <- estimate_elasticities(results)
comp <- prepare_sube_comparison(results, models)
comp_filtered <- filter_paper_outliers(comp)
#> Applies the six exclusion layers from the 2024 paper's outlier
#> treatment - see ?filter_paper_outliers for the full rule list.

Pass apply_bounds = FALSE to keep multiplier outliers in the output, or pass variables = c("GO", "VA", "EMP") to skip the CO2-specific rules when CO2 data is unavailable.

7. Running the regressions 

models <- estimate_elasticities(results)
models$OLS[term == "P01" & variable == "GO"]
#> The SUBE paper's OLS, pooled, and between estimators are all produced
#> in one call. See ?estimate_elasticities for the full return shape.

8. Comparing with the 2024 paper output

Running the full pipeline on the 2018 WIOD data yields:

#> Raw supply and use matrices: exact match (bit-identical)
#> Net-supply model matrix W = t(SUP_agg - USE_agg): exact match
#> Leontief multipliers after outlier treatment: ~2.7% mean abs. diff.
#> OLS coefficients (significant terms): match to 4+ decimal places

The matrix-level differences are zero; the multiplier- and coefficient-level differences are methodological (averaging order, filter interaction) and are documented in full in inst/scripts/replicate_paper.R.

9. Running the gated test locally

With SUBE_WIOD_DIR pointing at a populated WIOD tree:

SUBE_WIOD_DIR=/path/to/wiod Rscript -e 'devtools::test(filter = "replication")'

Without the env var (or on CRAN / CI), the test auto-skips:

Rscript -e 'devtools::test(filter = "replication")'
#> SKIP (SUBE_WIOD_DIR not set - paper replication test skipped)

Beyond this vignette

  • IHS regression variants. The 2024 paper additionally fits inverse-hyperbolic-sine transformed variants (linear, IHS, and mixed lin-IHS / IHS-lin combinations). Those are not part of sube v1.1; contact the maintainer if you need the extension.
  • Paper figures. plot_paper_comparison(), plot_paper_regression(), and plot_paper_interval_ranges() reproduce the published figures. Each ships with its own man page.
  • Full comparison runbook. inst/scripts/replicate_paper.R is the reference runbook; this vignette is the narrated subset.