Tangetial Wasserstein Projections

Gunsilius, Hsieh & Lee (2022)

Overview

The code in this repository constructs the results, plots, and tables found in the accompanying paper, available at https://arxiv.org/abs/2207.14727.

Data Availability and Provenance Statements

Summary of Availability

• All data used herein are publicly available.

Details on Each Data Source

• For Lego image replication, the data can be downloaded from Kaggle: https://www.kaggle.com/datasets/joosthazelzet/lego-brick-images. In our repository, select Lego_bricks in the folder Data, and the images used are contained therein. We kept the same file names for the files downloaded from the link above.

Datafiles: Data/Lego_bricks

• For Medicaid expansion application, the data can be downloaded from IPUMS: https://usa.ipums.org/usa/. We downloaded the variables stated in the main text: HINSCAID, EMPSTAT, UHRSWORK, INCWAGE. We selected additional technical person-level ID variables to allow us to select household head and spouse (if any). We applied further sample selection criteria mentioned in Appendix B.2 of the accompanying paper.

Datafiles available here: https://www.dropbox.com/sh/y43s568l44ny8pz/AADmFeUdanq9PzKHYhESaPBaa?dl=0

Dataset list

Data file	Source	Notes	Provided
Data/Lego_bricks/0001.png	Listed above	Target image in Section 4.2 of main text	Yes
Data/Lego_bricks/0040.png	Listed above	Control image in Section 4.2 of main text	Yes
Data/Lego_bricks/0080.png	Listed above	Control image in Section 4.2 of main text	Yes
Data/Lego_bricks/0120.png	Listed above	Control image in Section 4.2 of main text	Yes
Data/Lego_bricks/0160.png	Listed above	Control image in Section 4.2 of main text	Yes
Data/Lego_bricks/0200.png	Listed above	Control image in Section 4.2 of main text	Yes
Data/Lego_bricks/0240.png	Listed above	Control image in Section 4.2 of main text	Yes
Data/Lego_bricks/0280.png	Listed above	Control image in Section 4.2 of main text	Yes
Data/Lego_bricks/0320.png	Listed above	Control image in Section 4.2 of main text	Yes
Data/Lego_bricks/0360.png	Listed above	Control image in Section 4.2 of main text	Yes
Data/Lego_bricks/0400.png	Listed above	Control image in Section 4.2 of main text	Yes
Data/Medicaid_Data/	Listed above	ACS data used to obtain optimal weights lambda star	Yes; in Dropbox folder above
Data/Medicaid_Data/Counterfactual	Listed above	ACS data used to obtain counterfactual distributions in Section 4.3	Yes; in Dropbox folder above

Computational requirements

Software Requirements

The simulations and applications were ran using Python/3.8.1.

• ot (pip install pot)
• cvxpy
• numpy
• matplotlib
• pandas
• seaborn

Memory and Runtime Requirements

Summary

Approximate time needed to reproduce the analyses on a standard (CURRENT YEAR) desktop machine:

• 1-3 days

Details

The code was last run on a 8-core Apple-based laptop with MacOS version 12.4.

Portions of the code were last run on a 36-core Intel server with 180 GB of RAM. Computation took 2 hours.

Description of programs/code

• Programs in Python Code will perform all simulations and estimations described in the main text. These programs will also generate all tables and plots found therein.

Instructions to Replicators

• Download the repository to your local computer.
• Download Medicare data from the Medicaid Data/Pre-Intervention Data subfolder on Dropbox (linked above) to the directory: Data/Medicaid_Data.
• Download Medicare data from the Medicaid Data/Counterfactual Data subfolder on Dropbox (linked above) to the directory: Data/Medicaid_Data/Counterfactual.
• Move the Data folder to the Python Code subdirectory.
• Run Python Code/Mixed Gaussian Simulation.ipynb for the Gaussian simulations.
• Run Python Code/Lego Block Simulation.py for the Lego Brick image replication.
• Run Python Code/Medicaid.ipynb for the Medicaid expansion application.

The .png files will be stored in Python Code subdirectory by default. The tables are generated within each .ipynb program. If you wish to save the .png files in alternative directories, please change the corresponding plt.savefig() commands in the programs.

Details

• Python Code/twp_utils.py: defines barycentric projection and projection method described in the main text.
• Python Code/Mixed Gaussian Simulation.ipynb: contains the Gaussian simulations described in the main text.
• Python Code/Lego Block Simulation.py: contains the Lego Brick image replication described in the main text.
• Python Code/Medicaid.ipynb: contains the Medicaid expansion application described in the main text.

Apart from Python Code/Lego Block Simulation.py, it takes less than 5 minutes to finish running all the programs on a laptop with Apple M1 processor. Running Python Code/Lego Block Simulation.py takes 2 hours from start to finish on a cluster computer with 36 cores.

List of tables and programs

The provided code reproduces:

• All tables in the paper
• Figures 2 through 7 in the paper

Figure/Table #	Program	Line (Block) Number
Table 1	Python Code/Mixed Gaussian Simulation.ipynb	8
Table 2	Python Code/Medicaid.ipynb	9
Table 3	Python Code/Mixed Gaussian Simulation.ipynb	12
Figure 2 (Right)	Python Code/Lego Block Simulation.py	152
Figure 3 (Weights)	Python Code/Lego Block Simulation.py	144
Figure 4	Python Code/Medicaid.ipynb	19, 20
Figure 5	Python Code/Medicaid.ipynb	21, 22
Figure 6	Python Code/Medicaid.ipynb	10, 12
Figure 7	Python Code/Medicaid.ipynb	14, 15

References

Steven Ruggles, Sarah Flood, Ronald Goeken, Megan Schouweiler and Matthew Sobek. IPUMS USA: Version 12.0 [dataset]. Minneapolis, MN: IPUMS, 2022. https://doi.org/10.18128/D010.V12.0