Penalized PPML Regression with penppml

1 Introduction

penppml is an R package that enables users to fit penalized Poisson Pseudo Maximum Likelihood (PPML) regressions with high-dimensional fixed effects (HDFE). Supported penalties in the current version are ridge and lasso. The original application that motivated the development of penppml was the estimation of three-way gravity models of trade with a large number of PTA provision dummies (Breinlich, Corradi, Rocha, Ruta, Santos Silva and Zylkin, 2021).

To install and load penppml, you can use the following commands:

install.packages(penppml)
library(penppml)

2 Data Sets

The penppml package features the trade data set, which integrates panel data on bilateral trade flows with information about specific provisions in trade agreements for 220 exporters and 270 importers. The provisions included in the package are a subset of 16 out of 305 “essential” provisions featured in the full data set. More information about the data set and the variables is included in the corresponding help file, accessible via ?trade.

Table 1: International Trade Data Set
exp imp time export id agreement ad_prov_14 cp_prov_23 tbt_prov_07 tbt_prov_33
ABW ANT 1988 13087119 0 0 0 0 0
ABW ANT 1996 2371576 0 0 0 0 0
ABW ARG 1988 3351 0 0 0 0 0
ABW ARG 2012 50642 0 0 0 0 0
ABW ATG 1988 558 0 0 0 0 0
ABW ATG 1996 132070 0 0 0 0 0
ABW AUS 1996 1129 0 0 0 0 0
ABW BHS 1988 1117 0 0 0 0 0
ABW BHS 1996 587 0 0 0 0 0
ABW BHS 2012 830249 0 0 0 0 0

Along with the trade data set, the package includes an auxiliary data frame, countries, which contains basic information about the country ISO codes included in the main data set.

Table 2: Country Data Set
iso name region sub.region
AFG Afghanistan Asia Southern Asia
ALA Islands Europe Northern Europe
ALB Albania Europe Southern Europe
DZA Algeria Africa Northern Africa
ASM American Samoa Oceania Polynesia
AND Andorra Europe Southern Europe
AGO Angola Africa Sub-Saharan Africa
AIA Anguilla Americas Latin America and the Caribbean
ATA Antarctica
ATG Antigua and Barbuda Americas Latin America and the Caribbean

This enables users to easily filter by region or subregion. For instance, if we want to restrict our analysis to countries in the Americas, we can do the following:

selected <- countries$iso[countries$region %in% c("Americas")]
trade2 <- trade[(trade$exp %in% selected) & (trade$imp %in% selected), -(5:6)] # We remove columns 5 and 
# 6 because these variables are not needed in our regressions.

We will show the capabilities of this package using the filtered trade2 data frame.

3 Unpenalized PPML estimation with HDFE

The package enables users to run unpenalized PPML regressions with HDFE, using the hdfeppml function as follows:

reg1 <- hdfeppml(data = trade2,
                 dep = "export",
                 fixed = list(c("exp", "time"), 
                              c("imp", "time"),
                              c("exp", "imp")))
#> User did not specify independent variables. By default, the function takes all variables
#>         not included in 'dep' or 'fixed' as regressors.

As we can see, the function is designed to be data-frame-friendly: the user sets a data frame of reference in the data argument and then picks the dependent, independent and fixed effect variables either by name or by column number within the provided data frame. Also, note the following points about the syntax:

Internally, the function will do the necessary transformations of the columns into vectors and matrices, as needed by the algorithm.

Alternatively, more advanced users who prefer to handle data transformations themselves can use the internal function hdfeppml_int. For more information and examples on this issue, run ?hdfeppml or ?hdfeppml_int. Note also that this applies to all of the main functions in our package: both the data-frame-friendly wrapper and the internal function are available for use.

The results of our PPML model are:

results <- data.frame(prov = rownames(reg1$coefficients), b = reg1$coefficients, se = 0)
results$se[!is.na(reg1$coefficients)] <- reg1$se
results
Table 3: Unpenalized PPML results
Provision Coefficient SE
ad_prov_14 -0.0033 0.3334
cp_prov_23 -0.0337 0.0514
tbt_prov_07 0.4166 0.3252
tbt_prov_33 -0.0874 0.3382
tf_prov_41 0.6866 0.1831
tf_prov_45 0.1308 0.0619
ser_prov_47 -0.2780 0.1143
inv_prov_22 0.3722 0.0617
Provision Coefficient SE
et_prov_38 -0.0246 0.1274
ipr_prov_44 NA 0.0000
env_prov_18 -0.2642 0.1343
ipr_prov_15 NA 0.0000
moc_prov_21 0.1052 0.1495
ste_prov_30 0.0290 0.0636
lm_prov_10 -0.5859 0.1739
cp_prov_26 -0.1674 0.0854

(Note that the functions is automatically dropping perfectly collinear variables from the estimation algorithm and reporting NA as the coefficient.)

4 Penalized PPML estimation with HDFE

4.1 Lasso regression

The mlfitppml function is a flexible tool for computing penalized PPML regressions with HDFE. For instance, if we want to fit a PPML regression with a lasso penalty for several values of the penalty parameter (lambda) at once, we can run:

lambdas <- c(0.05, 0.025, 0.01, 0.0075, 0.005, 0.0025, 0.001, 0.00075, 0.0005, 0.00025, 0.0001, 0)

reg2 <- mlfitppml(data = trade2,
                  dep = "export",
                  fixed = list(c("exp", "time"), 
                              c("imp", "time"),
                              c("exp", "imp")),
                  penalty = "lasso",
                  lambdas = lambdas)

The function returns a list that contains two sets of coefficients for each value of lambda: the penalized coefficients (in the beta_pre element) and the post-penalty or unpenalized coefficients (in the beta element; these are calculated by estimating a post-penalty regression with just the selected variables). We can plot the regularization path of the penalized coefficients as follows:

results <- as.data.frame(reg2$beta_pre)
names(results) <- lambdas
results$provision <- row.names(results)
results <- reshape2::melt(results, id.vars = "provision", variable.name = "lambda", 
                          value.name = "coefficient")
results$lambda <- as.numeric(as.character(results$lambda))

ggplot2::ggplot(data = results, mapping = ggplot2::aes(x = lambda, y = coefficient, col = provision)) +
  ggplot2::geom_line(show.legend = FALSE) +
  ggplot2::scale_x_reverse(expand = ggplot2::expansion(add = c(0, 0.015))) +
  ggplot2::theme_classic() +
  directlabels::geom_dl(ggplot2::aes(label = provision), 
                        method = list(directlabels::dl.trans(x = x + 0.5), "last.bumpup")) +
  ggplot2::labs(x = "Penalty parameter (lambda)", y = "Coefficient", 
                title = "Figure 1: Regularization path for lasso")

If the user wishes to obtain the penalized estimates for a single value of lambda, they can either use mlfitppml as described above (just setting lambdas == x, where x is a number) or use the penhdfeppml function, upon which mlfitppml is built, directly. For instance:

reg3 <- penhdfeppml(data = trade2,
                  dep = "export",
                  fixed = list(c("exp", "time"), 
                              c("imp", "time"),
                              c("exp", "imp")),
                  penalty = "lasso",
                  lambda = 0.005)

We can easily check that the penalized coefficient estimates of penhdfeppml and mlfitppml are equal for a lambda value of 0.005 (within a numerical tolerance):

all.equal(as.vector(reg3$beta[!is.na(reg3$beta)]), as.vector(reg2$beta_pre[, 5]), tol = 1e-05)
#> [1] TRUE

For more details on penhdfeppml, run ?penhdfeppml.

4.2 Ridge regression (in development)

mlfitppml also allows user to use the ridge penalty in their PPML regressions. Simply run:

lambdas <- seq(0.0001, 0, length.out = 10) 

reg4 <- mlfitppml(data = trade2,
                  dep = "export",
                  fixed = list(c("exp", "time"), 
                              c("imp", "time"),
                              c("exp", "imp")),
                  penalty = "ridge",
                  lambdas = lambdas)

Note that this feature is still in development and may contain bugs.

4.3 Penalty selection

4.3.1 Cross-validation

The mlfitppml function enables users to carry out cross-validation of their models via the xval and IDs arguments. When xval is set to TRUE, the function performs cross-validation using a user-provided vector of IDs. For instance, if we want to do k-fold cross validation with k = 20, splitting the data set by agreement (not by observation), we can do the following:

id <- unique(trade[(trade$exp %in% selected) & (trade$imp %in% selected), 5])
nfolds <- 20
unique_ids <- data.frame(id = id, fold = sample(1:nfolds, size = length(id), replace = TRUE))

cross_ids <- merge(trade[(trade$exp %in% selected) & (trade$imp %in% selected), 5, drop = FALSE], 
                   unique_ids, by = "id", all.x = TRUE)

Now we activate the cross-validation option in mlfitppml and input the ID vector:

reg5 <- mlfitppml(data = trade2,
                  dep = "export",
                  fixed = list(c("exp", "time"), 
                              c("imp", "time"),
                              c("exp", "imp")),
                  penalty = "lasso",
                  lambdas = c(seq(0.5, 0.1, by = -0.1), 0.05, 0.01, 0.005, 0.001, 0.0005, 0.0001, 0),
                  xval = TRUE,
                  IDs =  cross_ids$fold)

Now the function also returns a rmse element that includes the cross-validation results (the average RMSE or root mean squared error for each value of lambda). Users can employ this tool to choose the value of the penalty parameter that minimizes the RMSE:

reg5$rmse
Table 4: Cross-validation results
Penalty (lambda) RMSE
5e-01 108.5813
4e-01 108.5813
3e-01 108.5813
2e-01 108.5813
1e-01 108.5813
5e-02 108.5813
1e-02 108.5750
5e-03 108.5743
1e-03 108.5740
5e-04 108.5739
1e-04 108.5739
0e+00 108.5739

In this case, the RMSE criterion suggest that the penalty should be set at 0. In other words, the results in Table 3 are the ones that minimize the mean squared error of the regression, according to 20-fold cross-validation.

4.3.2 Plugin lasso

This package also enables the use of the plugin lasso method, incorporating coefficient-specific penalty weights calculated automatically by the package. The most convenient way to do this is to set method = "plugin" in mlfitppml. Note that the plugin algorithm requires a clustering variable - we are using the interaction of the exporter and importer variables in the trade data set:

reg6 <- mlfitppml(data = trade2,
                  dep = "export",
                  fixed = list(c("exp", "time"), 
                              c("imp", "time"),
                              c("exp", "imp")),
                  penalty = "lasso",
                  method = "plugin",
                  cluster = c("exp", "imp"))
results <- data.frame(prov = rownames(reg6$beta), b_pre = reg6$beta_pre, b = reg6$beta, se = 0)
results$se[!is.na(reg6$beta)] <- reg6$ses
results
Table 5: Plugin Lasso results
Provision Lasso Coefficient Post-Lasso Coefficient SE
ad_prov_14 0.0000 0.0000 NA
cp_prov_23 0.0000 0.0000 NA
tbt_prov_07 0.0000 0.0000 NA
tbt_prov_33 0.0000 0.0000 NA
tf_prov_41 0.0000 0.0000 NA
tf_prov_45 0.1353 0.4033 0.0848
ser_prov_47 0.0000 0.0000 NA
Provision Lasso Coefficient Post-Lasso Coefficient SE
inv_prov_22 0.0122 0.1676 0.0619
et_prov_38 0.0000 0.0000 NA
env_prov_18 0.0000 0.0000 NA
moc_prov_21 0.0000 0.0000 NA
ste_prov_30 0.0000 0.0000 NA
lm_prov_10 0.0000 0.0000 NA
cp_prov_26 0.0000 0.0000 NA

We can see that the plugin lasso is very strict: only two provisions have non-zero coefficients when the penalties are set at the level suggested by the plugin algorithm. Compare this to the cross-validation results, which didn’t remove any of the provisions.

4.3.3 Iceberg lasso (in development)

This package also allows users to implement the two-step lasso described in Breinlich et al. (2021). This method consists in, first, running a plugin lasso estimation and second, running individual lasso regressions on each of the variables selected in the first stage. The iceberg function is designed precisely with this second step in mind. It takes a vector of dependent variables and returns a lasso regression for each one of them:

iceberg_results <- iceberg(data = trade2[, -(1:4)],
                           dep = results$prov[results$b != 0],
                           selectobs = (trade2$time == "2016"))

Currently, the function returns a matrix with coefficient estimates for each of the selected provisions. Support for standard errors (including clustered) is in development as of the current version of the package. The iceberg lasso coefficients are:

iceberg_results
Table 6: Iceberg Lasso coefficients
tf_prov_45 inv_prov_22
ad_prov_14 0.0000 0.0000
cp_prov_23 0.5912 0.0000
tbt_prov_07 0.0961 0.0000
tbt_prov_33 0.0000 0.0000
tf_prov_41 0.2144 0.0000
ser_prov_47 0.0000 0.5083
et_prov_38 0.0000 0.3315
ipr_prov_44 0.0000 0.0000
env_prov_18 0.0000 0.0000
ipr_prov_15 0.0000 0.0000
moc_prov_21 0.0000 0.0000
ste_prov_30 0.0580 0.0668
lm_prov_10 0.0000 0.0000
cp_prov_26 0.0334 -0.0040

Since PPML coefficients can’t be easily interpreted, you may find it useful to see the raw correlations between the variables in the iceberg lasso step:

provcorr <- cor(trade2[, results$prov])
(provcorr <- provcorr[, results$prov[results$b != 0]])
Table 7: Iceberg Lasso correlations
tf_prov_45 inv_prov_22
ad_prov_14 0.1580 0.2260
cp_prov_23 0.9691 0.1612
tbt_prov_07 0.7023 0.3039
tbt_prov_33 0.1813 0.1318
tf_prov_41 0.3306 -0.0285
tf_prov_45 1.0000 0.1469
ser_prov_47 0.0605 0.6679
inv_prov_22 0.1469 1.0000
et_prov_38 0.0771 0.5961
env_prov_18 -0.0065 0.0752
moc_prov_21 0.0784 0.6007
ste_prov_30 0.7716 0.5569
lm_prov_10 0.0231 0.0752
cp_prov_26 0.7975 0.2416

5 References

Breinlich, H., Corradi, V., Rocha, N., Ruta, M., Santos Silva, J.M.C. and T. Zylkin, T. (2021). “Machine Learning in International Trade Research: Evaluating the Impact of Trade Agreements”, Policy Research Working Paper; No. 9629. World Bank, Washington, DC.

Correia, S., P. Guimaraes and T. Zylkin (2020). “Fast Poisson estimation with high dimensional fixed effects”, STATA Journal, 20, 90-115.

Gaure, S (2013). “OLS with multiple high dimensional category variables”, Computational Statistics & Data Analysis, 66, 8-18.

Friedman, J., T. Hastie, and R. Tibshirani (2010). “Regularization paths for generalized linear models via coordinate descent”, Journal of Statistical Software, 33, 1-22.

Belloni, A., V. Chernozhukov, C. Hansen and D. Kozbur (2016). “Inference in high dimensional panel models with an application to gun control”, Journal of Business & Economic Statistics, 34, 590-605.