State of the Art RD Packages For Python

Installation

Please note that installation should be done with a mixture of pip and direct installations in R so that the correct versions of packages are installed.

With pip for python: pip install py_rdpackages

Important: you need to have the original rdrobust installed in R: install.packages('rdrobust')

Notes:

• rdrobust requires the newest version of ggplot2 so please: install.packages('ggplot2')
• The rdrobust and ggplot2 versions in Anaconda lag behind the current release for R. Please install in R directly as opposed to through conda.

Introduction and Use

These packages are a work in progress, but are an attempt to create a wrapper to implement the wonderful RD packages found here (https://sites.google.com/site/rdpackages/rdrobust) which utilize R or Stata, so that they can be used in Python directly.

I utilize the graphics package of Seaborn to make the plots look nice and to enable additional features, like binning points by size.

There are three packages in py_rdpackages:

1. rdplot creates plots of the regression discontinuity with a variety of options.
2. rdrobust does the RD and reports the regression results.
3. rdbwselect selects the optimal bandwidth size.

A Example of the Output

The code in the testing folder produces the following example.

Requirements and Stability

Use of the programs requires all of the following packages in Python:

1. rpy2 for running R in Python
2. matplotlib and seaborn for producing high quality graphics
3. pandas, numpy and for data manipulation and dataframe reading

Currently tested and stable for:

• rdrobust version 0.99.9
• rpy2 version 2.9.4 (last version updated through conda -- may work for later versions)
• ggplot2 version 3.3.0

Limitations

1. ryp2 produced slow pandas DF to R DF conversions, so I use pd.df.to_csv('temp_file_for_rd.csv') as a solution and then delete that same file after doing the analysis. This should be fixed in future versions.
2. Of course, Python calling R and then converting back to python is not ideal. Some future version should code this from scratch.