Nonparametric Inference on State Dependence in Unemployment

This repository contains code for reproducing the empirical estimates and Monte Carlo simulations in Torgovitsky, A., "Nonparametric Inference on State Dependence in Unemployment," Econometrica, Vol. 87, No. 5 (September, 2019)

Important

The code included in the supplemental material for Econometrica is from May 3, 2019.
Please download the most recent version of the code from the GitHub repository.

Software Requirements

• MATLAB. No special toolboxes are required.

• A Mathematical Programming Language (AMPL). The trial/student version of AMPL is size restricted. Some limited versions of the code might run with the trial version, but a full license is required to reproduce the results in the paper.

• A linear programming solver for AMPL. The default is CPLEX. The default can be changed by passing e.g. Settings.Solver = 'gurobi' when calling ./src/DPO.m.

• The AMPL-MATLAB API

• Linux (or perhaps OSX). I coded this on a Linux system and made no attempt to be platform-independent. However, the code is primarily in MATLAB, so should be mostly platform-independent. Some file operations are used for recording the results. These would be likely sources of issues for other operating systems, but should be easy enough to fix.

Reproducing the Results

• Important first step: Open ./cfg/Config.m.

  • Change ResultsPath to a directory where you want results to be saved.
  • Change AMPLAPISetupPath to the location where you installed the AMPL-MATLAB API.

• The primary code for the DPO model is contained is ./src/DPO.m and the routines called from within. It contains many options, which are given default values in the structure called Settings that is defined at the top of that file. The code for the comparison parametric dynamic binary response (PDBR) model is contained in ./src/PDBR.m.

• The directory ./bin/ contains a file called RunSIPP.m that generates the empirical results in the paper.

  • Selecting SimSet = main and SimNum = n will produce column n (for n between 1 and 12) in Table 2 of the paper.

  • Selecting SimSet = sigma or SimSet = sigma-young and SimNum = n will produce results for the nth gridpoint in Figures 2 and 3 (n between 1 and 9).

  • Selecting SimSet = extra will produce results for Appendix Table S4 in the supplemental appendix.

• Running all of the empirical results in the paper will take a long time, primarily due to the procedure for constructing confidence regions.

  • Confidence regions can be turned off by changing Settings.BuildConfidenceRegions = 1 to Settings.BuildConfidenceRegions = 0 in the function LoadSpec in ./bin/RunSIPP.m.

  • The number of bootstrap replications can be reduced by lowering Settings.B = 250 to some smaller number, also in the function LoadSpec in ./bin/RunSIPP.m.

  • Multiple results for each SimSet can be produced simultaneously by using the file ./bin/BatchRunSIPP.m This is basically a poor-man's parallel that opens up multiple MATLAB threads. (Unfortunately, the AMPL-MATLAB API is not easy to parallelize, which is why I am using this crude workaround.) Using this function, all of the results of (e.g.) Table 1 can be produced with the command BatchRunSIPP('your-save-dir', 'main'). Note that this will open 12 MATLAB and AMPL instances at one time, which will strain a typical system.

• The directory ./bin/ also contains a file called RunMonteCarlo.m that generates the simulation results for the Monte Carlos reported in the supplemental Appendix.

  • SimNumber = 1 produces the results for Table S1 and Figure S1.
  • SimNumber = 2 produces the results for Table S2.
  • SimNumber = 3 produces the results for Table S3.

• The directory ./bin/ contains a batching file for the Monte Carlos called BatchRunMonteCarlo.m. This opens three MATLAB threads that produce results for three different sample sizes for SimNumber = 1 or 3.

Reproducing the Data

• The cleaned data used for both the empirical results and simulations is contained in ./data/sipp08.tsv and ./data/sipp08-young.tsv. These files are included with the repository. There is also a wide-form ./data/sipp08-wide.tsv that I use for producing the table of summary statistics (Table 1).

• I have included a Bash script ./data/DownloadAndCleanSIPP.sh that downloads the raw 2008 SIPP data from the NBER page, converts it to Stata format and then creates my extract:

  • The script uses the Stata dictionary and do files provided by the NBER, except that I have edited the beginning of the do files to remove their annoying directory hardcoding. These files are included with the repository in ./data.

  • By default, the script deletes the ~8--10GB of SIPP data once my extract is created. This can be changed by running ./data/DownloadAndCleanSIPP.sh nocleanup instead.

  • Part of the script involves running ./data/CleanSIPP.do in Stata, which implements the sample selection rules discussed in the paper.

Reproducing the Tables and Plots

The directory ./post contains some Python scripts and LaTeX templates used to turn the empirical and simulation results into tables and figures.

For the empirical results:

• Table 1 (summary statistics) is generated by running ./post/BuildSumStatsTable.py ./data/sipp08-wide.tsv destdir where destdir is the output location.

• Table 2 (main empirical results) is generated by ./post/BuildResultsTable.py simdir/results/main where simdir is the location of a simulation directory and main is the directory name that is created when the SimSet variable (see above) is set to main to produce the main results. In order to fully reproduce Table 2, all SimNum's (1 through 12) should have been called so that there are directories simdir/results/main/001 through simdir/results/main/012. The completed table will be located in /simdirs/results/main/.

• Figure 2 (sensitivity analysis) is generated by ./post/BuildResultsTable.py simdir/results/sigma where simdir is the location of a simulation directory and sigma is the directory name that is created when the SimSet variable (see above) is set to sigma to produce the sensitivity analysis.

• Figure 3 (sensitivity analysis with young sample) is generated like Figure 2 but with sigma-young in place of sigma.

• Figure S4 (extra results) is generated like Table 2 but with extra in place of main.

For the Monte Carlo simulations:

• Figure S1 and Table S1 are generated by running ./post/BuildMCEstTable.py simdir/results/ where simdir is the location of a simulation directory containing the results from BatchRunMonteCarlo.m with SimNumber = 1.

• Similarly, figure S3 is generated by running ./post/BuildMCEstTable.py simdir/results/ after BatchRunMonteCarlo.m with SimNumber = 3.

• Figure S2 is generated by running ./post/BuildMCTestTable.py simdir/results/ when after running RunMonteCarlo.m with SimNumber = 2.

My Software Versions

The results in the published paper were run with:

• MATLAB version 9.3.0.713579 (R2017b)
• AMPL version 20180927
• Gurobi version 8.1.0
• AMPL-MATLAB API version 1.4.0
• Stata 13.1 (for data cleaning only)

Software Acknowledgments

The code uses two user-written MATLAB functions both located in ./ext:

• csvimport.m written by Ashish Sadanandan

• GaussQuad written by Geert Van Damme

Problems or Bugs?

Please use GitHub to open an issue and I will be happy to look into it.