Skip to contents

Getting Started with OECDscaleup

Source: vignettes/OECDscaleup.Rmd
OECDscaleup.Rmd

What this package does

This package builds the firm-level “high-growth firm” (HGF) / scale-up indicators used in the OECD Scale Up project, and produces the disclosure-checked aggregate tables shared back with the OECD. It is a behaviour-preserving translation of the OECD’s original Stata pipeline (MASTER_OECD_p2.do and 5 worker .do files) — see Analysis/stata.md and Pseudocode/*.md in the package source repository for the full analysis this translation was built from.

Firm-level microdata is confidential and never leaves the country whose national statistical institute holds it, so no real data ships with this package. Every example below uses small, fabricated data purely to demonstrate the interface — the actual numbers are meaningless.

The pipeline at a glance

Five pipeline functions, matching the five original .do files one-to-one:

  1. build_accounts_indicators() — cleans raw firm-level data and constructs every derived indicator (growth rates, HGF flags, categories, multi-factor productivity, …). Every other stage depends on its output.
  2. compute_characteristics() — descriptive statistics of scaler populations.
  3. compute_persistence() — growth trajectories before/after a scaling event (“transitions” and “evolutions”).
  4. compute_contribution() — job/turnover creation-vs-destruction accounting.
  5. compute_predictors() — classifies scalers by pre-period predictor quartile and regresses scaling on those predictors.

Stages 2-5 are independent of each other and only depend on stage 1’s output. All five are pure functions — data frames in, data frames out — so they can be tested and used without touching the filesystem.

run_scaleup_pipeline() is the only function that performs file I/O: it reads the raw .dta inputs, calls the five functions above in order, and writes every output as a CSV — mirroring MASTER_OECD_p2.do’s role in the original package.

Configuration

Every stage takes a scaleup_config, built with config_scaleup(). This replaces the Stata global macros ($country, $inputdir, …) set once in MASTER_OECD_p2.do:

cfg <- config_scaleup(
  country = "FR",
  input_dir = "V:/SCALE_UP/DTA/",
  output_dir = "V:/SCALE_UP/TMP/",
  dominance = FALSE # set TRUE (with dom_vars/dom_num) if your country
  # requires dominance-rule confidentiality statistics
)

Running the full pipeline from files

If you have base_financials_<CT>.dta (and optionally plants_RAW_<CT>.dta) prepared per variable_dictionary_v1.xlsx, alongside this package’s own tl3_typology/exchange_rates reference data saved as .dta files in the same input folder (exactly as the original PDF instructions describe):

cfg <- config_scaleup(
  country = "FR",
  input_dir = "path/to/input",
  output_dir = "path/to/output"
)
outputs <- run_scaleup_pipeline(cfg)

This writes n13_characteristics_FR.csv, n13_transitions_FR.csv, and so on to output_dir, plus the intermediate accounts_indicators_FR.dta back to input_dir (used for the manual confidentiality-check procedure described below). It also returns every output table as an R list, for programmatic use.

Running the pure functions directly

If you already have data loaded in R — for instance because you’re testing, or piping in data from elsewhere in an R workflow — call the five functions directly instead. Here’s a small synthetic example:

set.seed(1)
years <- 2010:2018
base_financials <- data.frame(
  idf = rep(1:20, each = length(years)),
  year = rep(years, times = 20),
  sector_1d = "C",
  sector_2d = 26,
  sector_3d = 262,
  region_2d = "FR10",
  birth_year = 2000,
  employment = as.vector(sapply(1:20, function(i) {
    pmax(10, round(stats::rlnorm(1, log(30), 0.3) * (1 + stats::rnorm(1, 0.02, 0.1))^(seq_along(years) - 1)))
  })),
  stringsAsFactors = FALSE
)
base_financials$turnover <- base_financials$employment * 100

cfg <- config_scaleup(
  country = "FR",
  input_dir = tempdir(), # unused by the pure functions themselves
  output_dir = tempdir()
)

accounts_indicators <- build_accounts_indicators(cfg, base_financials)
characteristics <- compute_characteristics(cfg, accounts_indicators)
head(characteristics)
#> # A tibble: 6 × 19
#>    year sample category group       firm_count mean_employment_growth3y
#>   <dbl> <chr>  <chr>    <chr>            <int>                    <dbl>
#> 1  2013 HGE10  ALL      full sample          2                    0.509
#> 2  2014 HGE10  ALL      full sample          2                    0.501
#> 3  2015 HGE10  ALL      full sample          2                    0.517
#> 4  2016 HGE10  ALL      full sample          2                    0.390
#> 5  2017 HGE10  ALL      full sample          3                    0.456
#> 6  2018 HGE10  ALL      full sample          3                    0.456
#> # ℹ 13 more variables: mean_turnover_growth3y <dbl>,
#> #   mean_employment_futgrowth <dbl>, mean_turnover_futgrowth <dbl>,
#> #   share_acquisition <dbl>, L3share_young <dbl>, median_L3employment <dbl>,
#> #   median_L3turnover <dbl>, median_employment <dbl>, median_turnover <dbl>,
#> #   median_employment_growth3y <dbl>, median_turnover_growth3y <dbl>,
#> #   median_employment_futgrowth <dbl>, median_turnover_futgrowth <dbl>

compute_persistence() returns a named list (transitions, evolutions_p1, evolutions_p2) since a country’s panel may be too short to produce all three; compute_predictors() similarly returns list() entirely if added_value, tangible_fixed_assets, or labour_costs aren’t present in the input data.

persistence <- compute_persistence(cfg, accounts_indicators)
names(persistence)
#> NULL

persistence has no elements here — this random 9-year panel doesn’t happen to contain a firm that scales specifically within the 2014-2016 window (transitions) or a panel reaching 2020 (evolutions), and neither is an error. Real country panels spanning more years, with genuine scaling firms, will populate one or more of transitions, evolutions_p1, evolutions_p2.

Disclosure control

Every output table can include dom_<var> dominance-concentration columns — the share of a group’s total held by its largest dom_num firms — when config$dominance is TRUE. This package’s own dominance_share() function implements that statistic and is directly usable to independently verify a value, the same way the original PDF’s confidentiality-check worked example describes doing by hand in Stata:

cfg_dom <- config_scaleup(
  country = "FR", input_dir = tempdir(), output_dir = tempdir(),
  dominance = TRUE, dom_vars = c("employment", "turnover"), dom_num = 2
)
characteristics_dom <- compute_characteristics(cfg_dom, accounts_indicators)
characteristics_dom[1:3, c("category", "group", "sample", "dom_employment")]
#> # A tibble: 3 × 4
#>   category group       sample dom_employment
#>   <chr>    <chr>       <chr>           <dbl>
#> 1 ALL      full sample HGE10               1
#> 2 ALL      full sample HGE10               1
#> 3 ALL      full sample HGE10               1

Cells are not automatically blanked. As with the original Stata pipeline, the decision on which cells to suppress before sharing results externally rests with the analyst / national statistical institute — this package reports the statistics needed to make that decision, it does not make it for you.

Where to look next