Export Market Penetration Dynamics: Replication Package

Joseph Steinberg, University of Toronto

This repository contains the code required to reproduce the results in my paper "Export Market Penetration Dynamics" in the Journal of International Economics.

0. Notes and requirements

0.1 Note about the Brazilian microdata

The main data from Brazil used in this paper cannot legally be distributed, so the raw data files are not included in this repository. I have, however, included all of the key intermediate datasets required to reproduce the empirical results in Section 2. The processed firm-level data are in the file bra_microdata_processed.pik in the programs/scripts/output/pik folder, and the files bra_microdata_agg_by_d.pik and bra_microdata_agg_by_d2.pik in the same folder contain data aggregated to the destination-year and destination levels (respectively). Starting with these files, one can reproduce all empirical results in Section 2 without the underlying raw data.

0.2. Note about other large files

There are also some other files that are too large to store here on Github. In some cases (for example, the raw data from the World Bank Exporter Dynamics database used in Appendix C), compressed versions of the files are sufficiently small to be stored here, and you will see zip files. Please use "unzip here" when decompressing these files. In other cases (for example, the raw and processed model simulation output), even compressed files are too large, so I had to omit them from the repository. Any folder with omitted files contains a text file called missing files.txt listing them. I am happy to produce these files on request via SFTP or SCP. Please note that whenever possible, intermediate files are contained here so that this is in general not necessary. Note also that the model simulation output can be generated by compiling and running the C program described in section 2 below. Other than the Brazilian microdata that I cannot legally distribute, no external datasets were omitted.

0.3. Requirements

The scripts in the programs/python folder require Python 3 and the NumPy, Pandas, Matplotlib, Statsmodels, Patsy, os, and sys libraries, as well as Stata MP with the reghdfe package. The C code in the programs/c folder requires OpenMPI, GSL, and NLOpt. I ran all programs using Ubuntu Linux 19.10 on a 56-core AMD Threadripper workstation with 192GB of RAM; it may take a very long time (or may not be feasible at all) on a computer with fewer cores or less RAM. All commands listed below assume you are working in a Linux bash terminal. Please contact me if you need help running these programs on a different operating system.

1. Python scripts

The python scripts used to process data, run regressions, and create the tables and figures in the paper are contained the folder programs/python. Next to each file name, I have provided a brief description of what this script does and highlighted in bold the output files it produces. Intermediate datasets are stored in Python pickle format in the programs/python/output/pik folder. These scripts should be run in order if one is starting from scratch, but otherwise one can generally rely on the intermediate outputs from previous scripts.

1.1 Scripts for downloading and processing data

get_wbdata.py: Pulls data from World Bank on GDP per capita and population for export destinations.

gravdata.py: Merges data from three sources: gravity variables from CEPII Gravity database; bilateral tariff data from TRAINS; and bilateral trade data from DOT. Computes trade barriers by running a gravity regression.

bra_microdata_prep.py: Processses the raw firm-level data from Brazil and merges on the World Bank and gravity variables. Note that you cannot run this script without access to the SECEX Brazilian data, which I cannot legally provide.

1.2 Scripts for processing model output

model_microdata_prep.py: Processes the simulated firm-level data from the C program. Running this script without a command-line argument will process the data from the baseline model. Running it with the optional arguments smp, sunkcost, or acr processes the simulated data from the alternative models (static market penetration, sunk cost, and exogenous new exporter dynmamics models, respectively). Running it with the optional arguments abn1, abn2, abo1, abo2, or a0 processes the simulated data from the sensitivity analyses discussed in Appendix A.2 and A.3.

model_mechanics_plots.py: Produces Figure 3 (fig3_policy_function_example.pdf). Output is stored in programs/python/output/model_mech_figs.

life_cycle_prep.py: Conducts further processing on the files created by bra_microdata_prep.py and model_microdata_prep.py. Computes variables required to estimate equations (2)-(5) and stores in Stata format. Calls the following Stata do files to estimate these equations using the reghdf command: life_cycle_data.do (estimates for Brazilian data); life_cycle_model.do (estimates for baseline model); and life_cycle_alt_models.do (for alternative models and sensitivity analyses). All output (both intermediate datasets and estimation output) is stored in the folder programs/python/output/stata.

1.3 Scripts for creating tables and figures in paper

sumstats_regs.py: Conducts the empirical analyses in Section 2.1, some of the analysis comparing the models to data in Sections 4.3 and 4.4, and some of the sensitivity analyses in Appendices A.2 and A.3. Produces Table 1 (table1_sumstats_regs.tex), Table 3 (table3_model_results.tex), Table 4 (table4_sumstats_regs_costs.tex), Table A2 (tableA2_alpha_beta.tex), and Table B1 (tableB1_sumstats_regs_cross_sec.tex). All output is stored in programs/python/output/sumstats_regs.

life_cycle_figs.py: Loads intermediate output from previous script and creates Table A.1 (tableA1_life_cycle_dyn_v_data_tests.tex), Figure 1 (fig1_life_cycle_dyn_v_data.pdf), Figure 2 (fig2_life_cycle_dyn_x_data.pdf), Figure 4 (fig4_life_cycle_dyn_v_model.pdf), Figure 5 (fig5_life_cycle_dyn_x_model.pdf), Figure 6 (fig6_life_cycle_dyn_c_model.pdf), Figure A.1 (figA1_life_cycle_dyn_v_alpha_beta.pdf), and Figure A.2 (figA2_life_cycle_dyn_x_alpha_beta). All output is stored in programs/python/output/life_cycle_figs.

transition_dynamics.py: Loads aggregated transition dynamics files created by C program and creates Figure 7 (fig7_tr_dyn_perm_tau_drop.pdf). Output is stored in programs/python/output/transitions.

`by_nd.py': Conducts all additional analyses reported in Appendix B.2. Creates Table B.2 (tableB2_drank_regs.tex), Table B.3 (tableB3_exit_by_nd_drank.tex), Figure B.1 (figB1_dist_by_nd_model_vs_data.pdf), Figure B.2 (figB2_by_nd_drank_model_vs_data.pdf), and Figure B.3 (figB3_cost_by_nd_drank.pdf). All output is stored in programs/python/output/by_nd.

1.4 Scripts for analyzing World Bank Exporter Dynamics Database data

Appendix C shows that the main empirical results also obtain in firm-level data from Mexico and Peru. This data is publicly available and can be legalluy distributed, unlike the Brazilian data usedd in the main text of the paper. The following scripts produce the results found in this appendix. These scripts are structured in much the same way as the ones listed above. Please note that all intermediate datasets and output files are stored in a single folder programs/python/wbedd/output.

app_wbedd/wbedd_microdata_prep.py: Processes World Bank Exporter Dynamics database microdata from Mexico and Peru.

app_wbedd/sumstats_regs.py: Creates Table C.1 (tableC1_sumstats_regs_wbedd.tex).

app_wbedd/life_cycle.py: Computes variables required to estimate equations (2)-(5) and stores in Stata format. Calls the Stata do file app_wbedd/life_cycle_data.do to estimate these equations using the reghdf command. Creates Figure C.1 (figC1_life_cycle_dyn_v_wbedd.pdf) and Figure C.2.(figC2_life_cycle_dyn_x_wbedd.pdf).

2. C program

The program to solve the model is written in C. It is contained in the folder programs/c. It uses OpenMP to parallelize the solution of the firm's problem and simulate microdatasets.

2.1 Source code

All source code is contained in the folder programs/c/src.

dyn_mkt_pen.c: Source code for baseline model and sensitivity analyses.

static_mkt_pen.c: Source code for static market penetration alternative model.

sunk_cost.c: Source code for sunk cost alternative model.

acr.c: Source code for exogenous new expoerter dynamics alternative model.

2.2 Compiling and running the program

The source code is compiled by running make from the command line in the programs/c folder. To compile the baseline model, type make dyn_mkt_pen (or just make). To compile the alternative models, type make static_mkt_pen, make sunk_cost, or make acr respectively. To run the models type ./bin/dyn_mkt_pen, ./bin/static_mkt_pen, ./bin/sunk_cost, or ./bin/acr. The baseline model program ./bin/dyn_mkt_pen has several command line options that allow the user to perform sensitivity analyses as well as the results for the baseline calibration.

2.3 Output files

The output of the programs is contained in the folder programs/c/output. The simulated microdata data are stored in files named <x>_microdata.csv, where <x> is the name of the model. The aggregated transition dynamics are stored in files named tr_dyn_perm_tau_drop_<x>.csv where again <x> denotes the name of the model.