Package {mditools}

mditools: Microdata Infrastructure Tools for Firm-Level Microdata Research

Description

Supports the full analysis pipeline for researchers working with firm-level microdata. Provides data tools for panel preparation (import, outlier detection, classification harmonization), analytical methods (production function estimation, capital stock measurement, markups, intensity measures, distributions, regression, clustering), and disclosure tools for tagging outputs with dominance and observation counts before aggregation and publication. Production function estimation implements methods by Ackerberg, Caves and Frazer (2015) doi:10.3982/ECTA13408, Levinsohn and Petrin (2003) doi:10.1111/1467-937X.00246, Wooldridge (2009) doi:10.1016/j.econlet.2009.04.026, Petrin, Poi and Levinsohn (2004) doi:10.1177/1536867X0400400202, and Arellano and Bond (1991) doi:10.2307/2297968 with the "too many instruments" correction by Roodman (2009) doi:10.1111/j.1468-0084.2008.00542.x. Markup estimation follows De Loecker and Warzynski (2012) doi:10.1257/aer.102.6.2437. Cost-share production function estimation follows Basu and Fernald (1997) doi:10.1086/262073. Capital stock estimation via the Perpetual Inventory Method follows OECD (2009) doi:10.1787/9789264068476-en.

Author(s)

Maintainer: Johanna Weiss johanna.weiss@iwhesl.onmicrosoft.com

Authors:

• Johanna Weiss johanna.weiss@iwhesl.onmicrosoft.com

• Daniele Aglio

• Eric Bartelsman

• Mirja H<U+00E4>lbig

• Marco Miorandi

• Alessandro Zona Mattioli

Other contributors:

• Juli<U+00E1>n D<U+00ED>az-Acosta [contributor]

• Alberto Ferreira [contributor]

• Javier Miranda [contributor]

• Marcelo Piemonte Ribeiro [contributor]

• Reetuparna Vishwanath [contributor]

• Chengzi Yi [contributor]

See Also

Useful links:

• https://github.com/Secretariat-CompNet/mditools

• Report bugs at https://github.com/Secretariat-CompNet/mditools/issues

ACF Production Function Estimation

Description

Estimates a production function using the Ackerberg-Caves-Frazer (ACF) method (Ackerberg, Caves & Frazer 2015, Econometrica). Runs a two-stage GMM procedure: a first-stage OLS (with optional polynomial and time fixed effects) to recover the productivity proxy Phi, followed by GMM minimization to identify input elasticities.

tfp = (Phi or y) - X*beta depending on TFP_minuend; with TFP_minuend = "y" this matches the y - X*beta convention of the other estimators.

Usage

mdi_acf_prodest(
  DT,
  y,
  endog,
  exog,
  instr,
  id,
  time,
  spec = "cd",
  degree = 3,
  lower_bound_theta = 0,
  upper_bound_theta = 1,
  TFP_demeaned = TRUE,
  TFP_minuend = c("Phi", "y"),
  Omega_estimates = TRUE,
  time_FE = FALSE,
  extended_instr = FALSE
)

Arguments

Value

A data.table with one row per observation in the GMM sample, containing: - id, time columns (using the names supplied) - tfp: total factor productivity - el_<input>: estimated input elasticity for each input - TFP_demeaned (if TFP_demeaned = TRUE): period-demeaned TFP - g_b_slopes, g_b_intercept (if Omega_estimates = TRUE) - NumObs: number of observations used in the GMM stage - convergence: optimiser convergence code (0 = converged)

Examples

library(data.table)
set.seed(1)
n <- 200
DT <- data.table(
  id   = rep(1:50, each = 4),
  year = rep(2000:2003, times = 50),
  y    = rnorm(n, 5, 1),
  l    = rnorm(n, 3, 0.5),
  k    = rnorm(n, 4, 0.5),
  m    = rnorm(n, 2, 0.5)
)
result <- mdi_acf_prodest(
  DT, y = "y", endog = "l", exog = "k", instr = "m",
  id = "id", time = "year", degree = 2, TFP_demeaned = FALSE
)

Generic Aggregation Function

Description

Aggregates variables from a sub-aggregate level to a higher aggregate level within a 'data.table'. Supports multiple aggregation types, including sum, standard deviation, mean, quantiles (25 missing values, number of non-missing values, number of empty strings, number of zeros, number of positive values, and the Herfindahl-Hirschman Index (HHI).

The function can optionally: - apply a weight column to variables before aggregation, - merge aggregated statistics back into the original dataset, - compute the number of unique firms in the input data, - and apply disclosure control criteria.

Usage

mdi_aggregate(
  DT,
  var_list,
  bygroups,
  agg_type = c("sum"),
  weight_col = NULL,
  mrg = FALSE,
  disclosure = TRUE,
  count_firms = FALSE,
  dom_formula = c("top_share", "residual"),
  minNumObs = 5L
)

Arguments

Value

- If 'mrg = FALSE': An aggregated 'data.table' containing the requested statistics, optionally with disclosure variables and number of firms. - If 'mrg = TRUE': The input 'data.table' with new columns containing the aggregated statistics.

Examples

library(data.table)
DT <- data.table(
  firmid = rep(1:5, each = 2),
  year   = rep(2020:2021, 5),
  nace   = rep(c("A", "B"), 5),
  emp    = c(10, 12, 5, 6, 20, 22, 8, 9, 15, 16),
  rev    = c(100, 110, 50, 55, 200, 210, 80, 85, 150, 155)
)

# Sum by nace
mdi_aggregate(DT, "emp", "nace", "sum", disclosure = FALSE)

# Multiple agg types
mdi_aggregate(DT, "emp", "nace", c("sum", "mean"), disclosure = FALSE)

# Merge back into original DT
mdi_aggregate(DT, "emp", "nace", "sum", mrg = TRUE, disclosure = FALSE)

# Count unique firms
mdi_aggregate(DT, "emp", "nace", "sum",
  count_firms = TRUE, disclosure = FALSE)

Clustering

Description

This tool clusters observations using one of the following methods: - k-means; - hierarchical clustering with Ward linkage; - hierarchical clustering with complete linkage; - hierarchical clustering with average linkage; - hierarchical clustering with single linkage; - PAM; - Gaussian mixture; - DBSCAN. The tool can optionally compute: - total within-cluster sum of squares (WSS); - average silhouette width; - bootstrap-style ARI stability; - the WSS or silhouette selection plot used to choose the number of clusters.

Usage

mdi_clustering(
  DT,
  id_vars,
  cluster_vars,
  method = c("hc_ward", "hc_complete", "hc_average", "hc_single", "kmeans", "pam",
    "mclust", "dbscan"),
  k_selection = c("fixed", "automatic"),
  k_fixed = NULL,
  automatic_by_wss = FALSE,
  automatic_by_silhouette = FALSE,
  compute_wss = TRUE,
  compute_silhouette = TRUE,
  compute_stability = FALSE,
  plot_selection = FALSE,
  k_grid = 2:25,
  exclude_noise = TRUE,
  B_boot = 200,
  nstart = 100,
  seed = 123,
  minPts = 4,
  q = 0.95,
  eps = NULL,
  G = 1:10,
  standardize = TRUE,
  na_action = c("stop", "omit"),
  cluster_col = "cluster",
  overwrite_cluster_col = FALSE,
  bootstrap_reselect_parameters = FALSE,
  verbose = TRUE
)

Arguments

Value

a list with: - data: the original input data.table with an additional clustering column, by default called "cluster". - chosen_k: final selected number of clusters, when relevant. - wss: final WSS result, if requested. - silhouette: final silhouette result, if requested. - stability: bootstrap-style ARI stability result, if requested. - selection_plot: recorded plot object, if plot_selection = TRUE.

Examples

library(data.table)
set.seed(1)
DT <- data.table(
  firmid = 1:30,
  x1     = c(rnorm(15, 0, 0.5), rnorm(15, 5, 0.5)),
  x2     = c(rnorm(15, 0, 0.5), rnorm(15, 5, 0.5))
)

result <- mdi_clustering(DT, id_vars = "firmid",
                     cluster_vars = c("x1", "x2"),
                     method = "kmeans", k_selection = "fixed", k_fixed = 2,
                     compute_wss = TRUE, compute_silhouette = TRUE,
                     compute_stability = FALSE, verbose = FALSE)

Cost-Shares Production Function Estimator (Cobb-Douglas)

Description

Implements a cost-shares approach to production function estimation: - Build firm-time input shares from expenditures in levels. - Average shares by (bygroup, time). - Compute TFP as a log-index residual using averaged shares. - Optionally demean TFP by (bygroup, time).

Usage

mdi_cs_prodest(
  DT,
  y,
  endog,
  exog,
  id,
  time,
  bygroup,
  log_values = TRUE,
  TFP_demeaned = TRUE
)

Arguments

Value

A data.table with one row per observation (after cost-share filtering), containing: - id, time, bygroup columns - el_<endog>, el_<exog>: averaged input elasticities - tfp: total factor productivity (log-index residual) - TFP_demeaned (if TFP_demeaned = TRUE) - NumObs: total number of observations used (matches the convention of the other estimators in this suite)

Returns NULL if no valid observations remain after filtering.

Examples

library(data.table)
set.seed(1)
n <- 120
DT <- data.table(
  id      = rep(1:30, each = 4),
  year    = rep(2000:2003, times = 30),
  nace    = rep(c("A", "B"), each = 60),
  y       = log(runif(n, 5, 50)),
  labour  = log(runif(n, 1, 10)),
  capital = log(runif(n, 2, 20))
)
mdi_cs_prodest(DT, y = "y", endog = "labour", exog = "capital",
               id = "id", time = "year", bygroup = "nace")

Add Disclosure Criteria to Aggregated Data

Description

Computes and attaches disclosure-control variables to an aggregated dataset. The function calculates a dominance measure and the number of non-missing observations per group.

Two dominance formulas are available via 'dom_formula': - '"top_share"' (default): share of the top 'domNr' firms in the group total. - '"residual"': (Total - x_1 - x_2) / x_1, where x_1 and x_2 are the two largest values. Used when the dominance criterion is defined as the residual relative to the largest firm.

This function is normally called internally by ['mdi_aggregate()'] when 'disclosure = TRUE', but can also be used standalone.

Usage

mdi_disclose_crit(
  DT,
  domVar = "var",
  domNr = 2,
  bygroups,
  var_list = NULL,
  dom_formula = c("top_share", "residual"),
  count_firms = FALSE,
  firm_col = "firmid",
  ent_col = "entid"
)

Arguments

Details

- For 'domVar != "var"', the named column must already be present in 'DT'. One dominance column ('domPerc') is returned and 'dom_formula' is ignored. - For 'domVar = "var"', separate dominance columns are created for each variable in 'var_list' ('domPerc_<var>'), using the formula specified by 'dom_formula'. - When 'count_firms = TRUE', 'firm_col' and 'ent_col' must be present in 'DT'; the function stops with an error if either is missing.

Value

A 'data.table' with the same grouping structure as the input, plus: - One or more 'domPerc_*' columns: dominance share per group. - A column 'NumObs': number of non-missing observations per group. - 'NumFirms' and 'NumEnt' (only when 'count_firms = TRUE').

Examples

library(data.table)
DT <- data.table(
  nace   = rep(c("A", "B"), each = 5),
  year   = rep(2020L, 10),
  emp    = c(10, 20, 5, 15, 8, 12, 25, 6, 14, 9),
  firmid = 1:10,
  entid  = c(1,1,2,2,3,4,4,5,5,6)
)

# Standard top-share formula
mdi_disclose_crit(DT, domVar = "var", domNr = 2L,
            bygroups = c("nace", "year"), var_list = "emp")

# Residual formula with firm counts
mdi_disclose_crit(DT, domVar = "var", domNr = 2L,
            bygroups = c("nace", "year"), var_list = "emp",
            dom_formula = "residual", count_firms = TRUE)

# Single-column dominance (column must be present in DT)
mdi_disclose_crit(DT, domVar = "emp", domNr = 2L,
            bygroups = c("nace", "year"))

Disclosure Check and Masking for Regression Tables

Description

Applies disclosure control rules to regression output tables. Rows that do not meet minimum thresholds for degrees of freedom, number of observations, or (for Germany) number of firms are flagged and sensitive regression statistics are masked.

Usage

mdi_disclose_reg_tab(
  DT,
  min_obs = 5,
  show_disclosed = FALSE,
  disc_method = c("obs_df", "firm_count")
)

Arguments

Details

- **'disc_method = "obs_df"':** disclosure is triggered when either 'df < min_obs' or 'NumObs < min_obs'. - **'disc_method = "firm_count"':** disclosure is based on 'NumFirms < min_obs' or 'NumEnt < min_obs'. Used for Germany. - The following regression statistics may be masked if disclosure applies: '"Estimate"', '"Std. Error"', '"z value"', '"Pr(>|z|)"', '"ci.lower"', '"ci.upper"', '"R2"', '"AdjR2"', '"AIC"', '"BIC"', '"LogLik"'.

Value

A list with three elements:

Examples

library(data.table)
DT <- data.table(
  coef         = c("(Intercept)", "x1"),
  Estimate     = c(1.2, 0.5),
  `Std. Error` = c(0.1, 0.05),
  df           = c(20L, 20L),
  NumObs       = c(25L, 25L)
)
result <- mdi_disclose_reg_tab(DT, min_obs = 3L)
result$DT
result$redacted_n

Arellano-Bond Difference-GMM Production Function Estimator

Description

Estimates the dynamic Cobb-Douglas production function

y_{it} = \rho\,y_{i,t-1} + \beta_l\,l_{it} + \beta_k\,k_{it} + \alpha_i + \varepsilon_{it}

by first-differencing to eliminate the firm fixed effect \alpha_i, then applying one-step GMM with lagged levels of (y, l, k) as instruments for the differenced regressors. Reference: Arellano & Bond (1991, ReStud).

Capital is treated as predetermined (chosen at t-1), so k_{i,t-s} for s >= 1 are valid instruments for \Delta k_{it}. Labour is treated as endogenous (chosen at t with knowledge of \varepsilon_{it}), so l_{i,t-s} for s >= 2 are valid. Lagged output y_{i,t-s} for s >= 2 instruments \Delta y_{i,t-1}. The lag depth is capped by max_lag_Z to avoid the "too many instruments" problem (Roodman 2009).

Estimation is one-step GMM with weight matrix W = (Z'HZ)^{-1}, where H is block-diagonal by firm with tridiagonal blocks (2 on the diagonal, -1 on the first off-diagonals) reflecting the MA(1) structure of \Delta\varepsilon_{it} under iid \varepsilon. Standard errors use the cluster-robust sandwich formula (clusters = firms).

tfp is reported in levels as tfp = y - beta_l*l - beta_k*k, matching the convention of the other estimators (ACF/LP/OLS/WDRG). It absorbs the firm fixed effect and the lagged-y persistence; it is NOT the productivity innovation \varepsilon_{it}.

Usage

mdi_dpgmm_prodest(
  DT,
  y,
  endog,
  exog,
  id,
  time,
  max_lag_Z = 2,
  TFP_demeaned = TRUE
)

Arguments

Value

A data.table with one row per firm-year (subset where y, endog, exog are non-missing). Columns: id, time, el_<endog>, el_<exog>, se_el_<endog>, se_el_<exog>, rho, se_rho, tfp, NumObs, the diagnostics sargan_J/sargan_df/sargan_pval and ar1_z/ar1_pval/ar2_z/ar2_pval, and optionally TFP_demeaned. Under iid \varepsilon: AR(1) should reject, AR(2) should not; Sargan tests the overidentifying restrictions.

Examples

library(data.table)
set.seed(1)
n_firms <- 40; n_periods <- 6
n <- n_firms * n_periods
DT <- data.table(
  id   = rep(seq_len(n_firms), each = n_periods),
  year = rep(seq(2000L, length.out = n_periods), times = n_firms),
  y    = rnorm(n, 5, 1),
  l    = rnorm(n, 3, 0.5),
  k    = rnorm(n, 4, 0.5)
)
result <- mdi_dpgmm_prodest(DT, y = "y", endog = "l", exog = "k",
                            id = "id", time = "year", TFP_demeaned = FALSE)

Estimate Firm-Level Markup

Description

Estimates firm-level markup following De Loecker (2012): markup is the output elasticity of an input divided by its expenditure share of revenue.

Usage

mdi_estimate_markup(DT, oe = "oe_l", rev_col = "nq", input_cost = "nm")

Arguments

Value

A 'data.table' with a single column 'markup'.

Examples

library(data.table)
DT <- data.table(
  oe_l = c(0.6, 0.7, 0.5),
  nq   = c(100, 200, 150),
  nm   = c(50,  80,  60)
)
mdi_estimate_markup(DT)

Estimate production functions with multiple estimators (panel, by-group)

Description

mdi_estimate_prodfun() is a wrapper that estimates Cobb-Douglas production functions on firm-level panel data using a selectable set of estimators (methods). It is designed for *grid runs* across within-method specifications (e.g., degree choices, time fixed effects, alternative demeaning) and returns standardized outputs that can be pooled across methods and industries.

The wrapper is typically run "by industry" (or another grouping variable) using bygroup. Within each group, the function calls one estimator at a time and binds results into a common panel-style output format including:

• firm id and year (or generic id, time)

• elasticities with standardized names el_<var>

• total factor productivity proxy (tfp) and optional de-meaned version (TFP_demeaned)

• book-keeping: NumObs, plus method/spec identifiers if requested

Usage

mdi_estimate_prodfun(
  DT,
  methods,
  y,
  endog,
  exog,
  instr = NULL,
  id,
  time,
  bygroup,
  acf_args = list(),
  lp_args = list(),
  wdrg_args = list(),
  dpgmm_args = list(),
  ols_args = list(),
  cs_args = list(),
  verbose = TRUE,
  drop_empty = TRUE,
  allowed_ids = NULL
)

Arguments

Details

## Input convention The wrapper assumes all production-function variables are in logs unless a method explicitly requires levels (e.g., cost shares with log_values=TRUE will exponentiate internally).

Minimal required columns in DT are: output y; free input(s) endog; state input(s) exog; proxy/instruments instr (ACF/LP/WRDG); panel identifiers id, time; and the grouping variable bygroup.

Each estimator performs its own NA filtering on the variables it needs; hence the effective sample can differ by method/spec.

## Methods implemented Methods are selected via methods. The wrapper recognizes:

"acf"

Ackerberg-Caves-Frazer (2015) control-function estimator (Cobb-Douglas). First-stage polynomial in inputs and proxy to construct \Phi, then a GMM stage with lagged inputs as instruments. Returns firm-level elasticities and a residual-based tfp.

"lp"

Levinsohn-Petrin (2003) proxy estimator (Cobb-Douglas).

"wdrg"

Wooldridge (2009) system-GMM estimator. Stacked two-equation GMM, estimated linearly in the parameters (the polynomial coefficients on c(x_t,m_t) and c(x_{t-1},m_{t-1}) are free, separate vectors; this is more general than the random-walk-with-drift case but does not impose the structural AR(G) restriction of a degree-G nonlinear law of motion). Returns common elasticities, se_el_<var>, tfp = y - X*beta, and alpha_hat as a diagnostic column.

"dpgmm"

Arellano-Bond (1991) difference-GMM for the dynamic PF y_{it} = \rho y_{i,t-1} + \beta_l l_{it} + \beta_k k_{it} + \alpha_i + \varepsilon_{it}. First-differences to remove the firm fixed effect; lagged levels of (y, l, k) instrument the differenced regressors (K predetermined, L endogenous). One-step GMM, cluster-robust SEs, Sargan/Hansen and AR(1)/AR(2) diagnostics. Returns el_<var>, rho, tfp = y - beta_l l - beta_k k.

"ols"

Pooled OLS baseline with flexible polynomial controls (Cobb-Douglas). tfp = y - X*beta. Use degree = 1 for the clean naive baseline.

"cs"

Cost-shares (index-number) approach for Cobb-Douglas.

## Method-specific optional arguments (xxxx_args) Each method accepts an optional list of tuning arguments passed via a dedicated parameter name: acf_args, lp_args, wdrg_args, dpgmm_args, ols_args, cs_args. Unspecified fields fall back to method defaults.

ACF arguments (acf_args): spec, degree, lower_bound_theta, upper_bound_theta, TFP_demeaned, TFP_minuend ("Phi"/"y"), Omega_estimates, time_FE, and extended_instr (logical, default FALSE; if TRUE adds \Phi_{t-1} to the instrument set per ACF eq 28 for overidentification).

LP arguments (lp_args): spec, degree, lower_bound_theta, upper_bound_theta, TFP_demeaned, TFP_minuend ("y"/"Phi"), Omega_estimates, time_FE.

Wooldridge arguments (wdrg_args): degree, tol, TFP_demeaned, TFP_minuend ("y" only).

Dynamic panel GMM arguments (dpgmm_args): max_lag_Z (integer >= 1, default 2; instrument lag depth), TFP_demeaned.

OLS arguments (ols_args): spec, degree, TFP_demeaned.

Cost shares arguments (cs_args): log_values, TFP_demeaned.

Value

A data.table with one row per firm-year (or per used observation), containing bygroup, method, id, time, el_<var>, tfp, and method-specific extras (TFP_demeaned, rho, se_el_<var>, diagnostics), plus NumObs.

Examples

library(data.table)
set.seed(42)
n <- 200
DT <- data.table(
  firmid = rep(1:50, each = 4),
  year   = rep(2000:2003, 50),
  sector = rep(c("A", "B"), each = 100),
  y      = rnorm(n, 5, 1),
  l      = rnorm(n, 3, 0.5),
  k      = rnorm(n, 4, 0.5),
  m      = rnorm(n, 2, 0.5)
)
mdi_estimate_prodfun(
  DT, methods = c("ols", "acf"),
  y = "y", endog = "l", exog = "k", instr = "m",
  id = "firmid", time = "year", bygroup = "sector"
)

Hierarchical Aggregation of Data

Description

Aggregates variables in 'var_list' to unique values of the hierarchical dimensions specified in 'hhfile' by groups. Aggregation is performed at each level specified by 'hier' using ['mdi_aggregate()'].

Usage

mdi_hier_apply(
  DT,
  hhfile,
  var_list,
  bygroups,
  hier,
  agg_type = "sum",
  weight_col = NULL,
  mrg = FALSE,
  disclosure = TRUE
)

Arguments

Value

A 'data.table' containing the aggregated variables from 'var_list' at each requested hierarchy level, combined via 'rbindlist'. A 'node' column identifies the aggregation level of each row.

Examples

library(data.table)
hhfile <- data.table(
  h_0 = c("A1", "A2", "B1", "B2"),
  h_1 = c("A",  "A",  "B",  "B")
)
DT <- data.table(
  nace = c("A1", "A2", "B1", "B2"),
  year = rep(2020L, 4),
  emp  = c(10L, 20L, 15L, 25L)
)
mdi_hier_apply(DT, hhfile, var_list = "emp",
           bygroups = c("nace", "year"), hier = "h_1",
           disclosure = FALSE)

Import Data into R data.table

Description

Reads data from various file formats into an R 'data.table'. This function is a wrapper around multiple file-reading packages such as 'fread' (from 'data.table'), 'haven' (for Stata, SAS, SPSS), and 'readxl' (for Excel). It supports CSV, Stata (.dta), Excel (.xlsx), SAS (.sas7bdat), SPSS (.sav), and tab-delimited text files (.txt). You can also specify a list of columns to import and specify which columns should be imported as characters. However, note that for some file formats (e.g., Stata, SAS, SPSS), the function cannot directly import columns as characters during the import process. In such cases, the specified columns are converted to character types **after** the data has been loaded.

Usage

mdi_import_data(
  dir,
  file,
  format,
  col_list = NULL,
  char_columns = NULL,
  encoding = NULL
)

Arguments

Value

A 'data.table' containing the imported data.

Examples

tmp_dir <- paste0(tempdir(), "/")
write.csv(data.frame(id = 1:3, emp = c(10, 20, 30)),
          paste0(tmp_dir, "data.csv"), row.names = FALSE)
mdi_import_data(tmp_dir, "data.csv", "csv", char_columns = "id")

Compute Technology Adoption Intensity via Probit Propensities

Description

Reduces a set of binary adoption indicators to a single continuous intensity score using probit regressions. For each boolean indicator a probit model is fitted with continuous firm-level predictors and optional fixed effects. The predicted propensities are then combined into a single intensity score via the geometric mean.

Usage

mdi_intensity(DT, uniqdim, boollist, contlist, fe)

Arguments

Value

A data.table keyed on 'uniqdim' with one additional column 'intens_probit': the geometric mean of all predicted propensities.

Examples

library(data.table)
set.seed(1)
n <- 100
DT <- data.table(
  firmid = seq_len(n),
  year   = sample(2010:2012, n, replace = TRUE),
  bool1  = sample(0L:1L, n, replace = TRUE),
  cont1  = rnorm(n),
  cont2  = rnorm(n)
)
mdi_intensity(DT, uniqdim = "firmid", boollist = "bool1",
          contlist = c("cont1", "cont2"), fe = "year")

Calculate Joint Distributions

Description

Computes joint distributions for specified variables within a data table. Calculates distributional moments (deciles, quintiles, or quartiles) for one or more variables, aggregated by specified groups and potentially hierarchical structures.

Usage

mdi_jointdist(
  DT,
  hhfile,
  qnames,
  var_names,
  moment = c("decile", "quintile", "quartile"),
  bygroups,
  hier,
  agg_type,
  prefix = agg_type,
  weight_col = NULL,
  mrg = FALSE,
  disclosure = TRUE
)

Arguments

Value

A 'data.table' with computed joint distributions including the distributional moments for specified variables aggregated by the given criteria.

Examples

library(data.table)
DT <- data.table(
  nace = rep(c("A", "B"), each = 5),
  year = rep(2020L, 10),
  emp  = c(10, 20, 5, 15, 8, 12, 25, 6, 14, 9)
)
hhfile <- data.table(h_0 = c("A", "B"), h_1 = c("X", "X"))
mdi_jointdist(DT, hhfile,
  qnames = "emp", var_names = "emp", moment = "quartile",
  bygroups = c("nace", "year"), hier = "h_1",
  agg_type = "sum", disclosure = FALSE)

Levinsohn-Petrin (2003) Production Function Estimator

Description

Two-stage proxy-variable estimator using an intermediate input (typically materials) to control for unobserved productivity. First stage: regress y on a flexible polynomial in (state, proxy) plus the free input, recovering Phi. Second stage: GMM on the law-of-motion residual identifying the state elasticities. Reference: Levinsohn & Petrin (2003, ReStud).

tfp = (y or Phi) - X*beta, depending on TFP_minuend. Defaults to "y" to match the convention shared by the other estimators.

Usage

mdi_lp_prodest(
  DT,
  y,
  endog,
  exog,
  instr,
  id,
  time,
  spec = "cd",
  degree = 3,
  lower_bound_theta = 0,
  upper_bound_theta = 1,
  TFP_demeaned = TRUE,
  TFP_minuend = c("y", "Phi"),
  Omega_estimates = TRUE,
  time_FE = FALSE
)

Arguments

Value

A data.table with one row per observation in the GMM sample, containing: - id, time columns (using the names supplied) - tfp: total factor productivity - el_<input>: estimated input elasticities - TFP_demeaned (if TFP_demeaned = TRUE): period-demeaned TFP - g_b_slopes, g_b_intercept (if Omega_estimates = TRUE) - NumObs: number of observations used in the GMM stage - convergence: optimiser convergence code (0 = converged)

Examples

library(data.table)
set.seed(1)
n <- 200
DT <- data.table(
  id   = rep(1:50, each = 4),
  year = rep(2000:2003, times = 50),
  y    = rnorm(n, 5, 1),
  l    = rnorm(n, 3, 0.5),
  k    = rnorm(n, 4, 0.5),
  m    = rnorm(n, 2, 0.5)
)
result <- mdi_lp_prodest(DT, y = "y", endog = "l", exog = "k", instr = "m",
                         id = "id", time = "year", degree = 2,
                         TFP_demeaned = FALSE)

Harmonize a Classification Over Time

Description

Creates a harmonized concordance for a classification of interest over a time period, starting from year-by-year concordance tables. Each row in 'conc_table' maps a code in year 't-1' (column 'left') to a code in year 't' (second column), along with a 'year' column indicating which transition the row belongs to.

The function handles 1:1, m:1, 1:m, and m:m code changes, and appends '"D"' to harmonized codes that disappear before the last year.

Usage

mdi_make_conc(conc_table, year_list, code_name)

Arguments

Value

A 'data.table' in long format with columns:

• 'year' — the year of the observation.

• '<code_name>' — the original code for that year.

• '<code_name>_harmonized' — the harmonized code mapped to the most recent non-missing code in the group.

Examples

library(data.table)
conc <- data.table(
  year  = c(2011L, 2011L, 2012L, 2012L),
  left  = c("A",   "B",   "A",   "C"),
  right = c("A",   "B",   "A2",  "C")
)
mdi_make_conc(conc, 2011:2012, "pcc")

OLS Production Function Estimator

Description

Fits a Cobb-Douglas production function by OLS on the free and state inputs together with a degree-G polynomial in the same inputs, following the prodest-package convention. Returns TFP as the residual between output and the fitted input index, with optional period demeaning.

Intended as a naive baseline for the simultaneity-bias comparison against the proxy-variable estimators. NOTE: at degree > 1 the reported linear-input coefficients are partial regression coefficients conditional on the higher-order polynomial terms, not structural elasticities. Use degree = 1 for the clean naive-OLS baseline.

tfp = y - X*beta (same convention as ACF/LP/WDRG/dpGMM).

Usage

mdi_ols_prodest(
  DT,
  y,
  endog,
  exog,
  id,
  time,
  spec = "cd",
  degree = 3,
  TFP_demeaned = TRUE
)

Arguments

Value

A data.table with one row per observation in the estimation sample, containing: - id, time columns (using the names supplied) - tfp: total factor productivity (OLS residual) - el_<endog>, el_<exog>: estimated input elasticities - TFP_demeaned (if TFP_demeaned = TRUE): period-demeaned TFP - NumObs: number of observations used

Examples

library(data.table)
set.seed(3)
n <- 200
DT <- data.table(
  id   = rep(1:50, each = 4),
  year = rep(2000:2003, times = 50),
  y    = rnorm(n, 5, 1),
  l    = rnorm(n, 3, 0.5),
  k    = rnorm(n, 4, 0.5)
)
result <- mdi_ols_prodest(DT, y = "y", endog = "l", exog = "k",
                          id = "id", time = "year", degree = 2,
                          TFP_demeaned = FALSE)

Outlier Routine for Data Cleaning

Description

Executes a specified outlier handling routine (trimming, winsorizing, or flagging) on a dataset for selected continuous variables. This function can trim or winsorize the data at specified quantiles or flag observations as outliers based on the fraction provided. Trimming replaces outliers with NA, winsorizing replaces outliers with the closest value within the non-outlier range, and flagging marks outliers with a flag variable.

Usage

mdi_outlier(
  DT,
  var_list,
  routine = c("trim", "winsorize", "flag"),
  fraction,
  both_tails = FALSE,
  group = NULL
)

Arguments

Value

A modified copy of 'DT' with the outlier routine applied to the specified variables. If 'routine = "flag"', new columns named 'flag_<var>' are added (value '1' for flagged observations, 'NA' otherwise).

Examples

library(data.table)
DT <- data.table(id = 1:10, income = c(50, 55, 45, 60, 200, 40, 45, 55, 65, 1000))
mdi_outlier(DT, "income", "trim", 0.1, both_tails = TRUE)
mdi_outlier(DT, "income", "winsorize", 0.1)
mdi_outlier(DT, "income", "flag", 0.1)

Estimate Capital Stock Using Perpetual Inventory Method (PIM)

Description

Estimates capital stock based on the Perpetual Inventory Method (PIM), as outlined in Halle & Mairesse (1995). This function is tailored for data structured as panel data, with indexing based on a firm identifier ('firmid') and a year variable. It involves the following steps: - Reading the specified variables from the input data.table. - Depending on whether the depreciation rate is provided directly or inferred from the asset type, the function calculates the capital stock. - The output is the input data.table augmented with the new capital stock variable.

The function supports different depreciation formats and asset types, offering flexibility in estimating capital stock across various contexts.

Usage

mdi_pim_capital(
  DT,
  id = "firmid",
  K0 = "K0",
  I = "ni_tan",
  delta = "d_GFCF",
  output_name = "k_new",
  time = "year"
)

Arguments

Value

A modified 'DT' with the new capital stock variable appended.

References

Halle, P., & Mairesse, J. (1995). "Estimation of the Perpetual Inventory Method for Capital Stock".

Examples

library(data.table)
DT <- data.table(
  firmid = rep(c("F1", "F2"), each = 4),
  year   = rep(2010:2013, 2),
  K0     = c(100, NA, NA, NA, 200, NA, NA, NA),
  ni_tan = c(10, 12, 11, 13, 20, 22, 21, 23),
  d_GFCF = rep(0.08, 8)
)

mdi_pim_capital(DT)

Weighted Regression with Automatic Disclosure Check

Description

Runs regression analysis on a 'data.table' using models from the **fixest** package ('feols', 'feglm', etc.), with automatic disclosure control applied.

Disclosure rules ensure that regressions are only reported if they meet minimum thresholds for sample size: both the number of observations and the residual degrees of freedom must be at least 'minNumObs'.

When 'count_firms = TRUE', the number of unique firms and enterprises in the regression sample are also computed and added to the output ('NumFirms', 'NumEnt'). This is controlled by the caller – typically set based on the country-specific disclosure requirements of the NSI running the code.

The function supports clustered standard errors, regression weights, instrumental variables, and optional LaTeX export of results.

Usage

mdi_regress(
  DT,
  formula,
  model = "feols",
  family = NULL,
  vcov = "iid",
  cluster = FALSE,
  tex = FALSE,
  output_name = NULL,
  desc_file = NULL,
  weights = NULL,
  iv = FALSE,
  num_firms = NULL,
  num_ent = NULL,
  count_firms = FALSE,
  firm_col = "firmid",
  ent_col = "entid",
  minNumObs = 5L,
  dirOUTPUT = NULL
)

Arguments

Details

- Regression output is only returned if both 'nobs >= minNumObs' and residual degrees of freedom '>= minNumObs'. - When 'count_firms = TRUE': unique firm and enterprise counts are computed from the regression sample using 'firm_col' and 'ent_col'. If a column is absent from the data, the corresponding 'num_firms' or 'num_ent' argument is used instead. If neither the column nor the manual count is provided, the function stops with an error before any regression is run. - If 'tex = TRUE', results are saved to '<dirOUTPUT>/<output_name>.tex' and an entry is appended to '<dirOUTPUT>/<desc_file>.txt'.

Value

A 'data.table' of regression results, including: - Coefficients and standard errors - Confidence intervals ('ci.lower', 'ci.upper') - Sample size ('NumObs'), residual df ('df') - 'NumFirms' and 'NumEnt' (only when 'count_firms = TRUE') - Fit statistics (R2, Adj. R2, AIC, BIC, LogLik, and IV/F tests if applicable)

If disclosure criteria are not satisfied, the regression is skipped and a message is printed.

Examples

library(data.table)
set.seed(1)
DT <- data.table(
  y      = rnorm(100),
  x1     = rnorm(100),
  x2     = rnorm(100),
  firmid = paste0("F", sample(1:20, 100, replace = TRUE)),
  entid  = paste0("E", sample(1:15, 100, replace = TRUE))
)
# Basic regression
mdi_regress(DT, formula = "y ~ x1 + x2", minNumObs = 5L)

# With firm/enterprise counts (e.g. when NSI rules require it)
mdi_regress(DT, formula = "y ~ x1 + x2", minNumObs = 5L,
            count_firms = TRUE)


# Weighted regression with clustered SEs
DT[, w := runif(.N)]
mdi_regress(DT, formula = "y ~ x1 + x2", weights = ~w,
            cluster = TRUE, vcov = "~firmid")

# With LaTeX export (writes .tex file to tempdir)
mdi_regress(DT, formula = "y ~ x1", tex = TRUE,
            output_name = "reg1", desc_file = "log",
            dirOUTPUT = paste0(tempdir(), "/"))

Count Classification Switches Between Consecutive Periods

Description

Identifies pairs of classification codes between which a unit switched in a given year, and counts how often each specific switch is observed across units. The output has one row per (old code, new code, year) triple.

Usage

mdi_transition(DT, id, time, classvar)

Arguments

Value

A data.table with four columns: 'old_code', 'new_code', 'year_shifted', and 'N' (count of units making that switch in that year).

Examples

library(data.table)
DT <- data.table(
  firmid = c("A", "A", "A", "B", "B", "B"),
  year   = c(2010L, 2011L, 2012L, 2010L, 2011L, 2012L),
  nace   = c("C10", "C10", "C20", "D30", "D40", "D40")
)
mdi_transition(DT, id = "firmid", time = "year", classvar = "nace")

Wooldridge (2009) System-GMM Production Function Estimator

Description

Implements the stacked two-equation GMM from Wooldridge (2009, Economics Letters 104, 112-114). The system is estimated linearly in the parameters: the polynomial coefficients on (x_t, m_t) (Wooldridge eq 3.3) and on (x_{t-1}, m_{t-1}) (eq 3.4) are free, separate vectors. This is more general than the random-walk-with-drift restriction (eq 3.10) which forces them equal, but it does NOT impose the structural rho_g * lambda^g form that eq (3.4) implies for a general degree-G polynomial law of motion (G > 1, which would require nonlinear GMM and is not implemented here). The implementation matches the standard empirical Wooldridge GMM used in Petrin, Poi and Levinsohn (2004).

Returns elasticities common to all firms in the group, firm-year tfp = y - X*beta (the intercept alpha_hat is reported as a separate column for diagnostics, not subtracted), and analytical SEs.

Usage

mdi_wdrg_prodest(
  DT,
  y,
  endog,
  exog,
  instr,
  id,
  time,
  degree = 2,
  tol = 1e-10,
  TFP_demeaned = TRUE,
  TFP_minuend = "y"
)

Arguments

Value

A data.table with one row per observation in the GMM sample, containing id, time, tfp, alpha_hat, el_<input>, se_el_<input>, NumObs, and optionally TFP_demeaned.

Examples

library(data.table)
set.seed(1)
n <- 200
DT <- data.table(
  id   = rep(1:50, each = 4),
  year = rep(2000:2003, times = 50),
  y    = rnorm(n, 5, 1),
  l    = rnorm(n, 3, 0.5),
  k    = rnorm(n, 4, 0.5),
  m    = rnorm(n, 2, 0.5)
)
result <- mdi_wdrg_prodest(DT, y = "y", endog = "l", exog = "k", instr = "m",
                           id = "id", time = "year", TFP_demeaned = FALSE)