GeoRDD

This julia package implements the methods introduced in the paper “A Bayesian Nonparametric Approach to Geographic Regression Discontinuity Designs: Do School Districts Affect NYC House Prices?”

What is a GeoRDD?

A GeoRDD is a natural experiment where units on one side of a geographical border are given a treatment while units on the other side are not. It's the spatial analog of univariate RDDs, which are an increasingly popular tool in econometrics and other social sciences.

Installation

The package is currently not registered, so it has to be installed with a direct link to github. This can be accomplished by running import Pkg; Pkg.add("https://github.com/maximerischard/GeoRDD.jl.git") (or add https://github.com/maximerischard/GeoRDD.jl.git from the Pkg REPL).

Quick Start

You first need to prepare a dataframe, for example with columns:

• X1 and X2 for the spatial covariates of the units,
• outcome for the outcome of interest
• region giving the treatment indicator (for example containing strings "treatment" and "control")
• covarA and covarB for two additional non-spatial real-valued covariates.
• categ for an additional categorical covariate

You also need the border coordinates as a LibGEOS LineString or MultiLineString object. You can either create this yourself or read it from a shapefile or geojson file. If what you have is polygons of the treatment and control regions, you can use the provided get_border function to obtain the border between the two polygons.

With this in hand, you can use this package to obtain

1. Estimates of the treatment effect everywhere along the border (the “cliff face” estimate),
2. Estimates of the local average treatment effect (LATE),
3. A (p)-value for the significance of the treatment effect.

See the notebooks/SimulatedExample.ipynb jupyter notebook for a step-by-step example of how to perform the GeoRDD analysis. If you run into any trouble, please open a GitHub issue.

For a more complete example using real data, see notebooks/NYC_analysis-2015.ipynb, which reproduces the example in the paper of estimating the difference in house prices between school districts in NYC.

Reproducibility

The figures in the paper and supplementary materials are generated in notebooks, so they are fully reproducible. In each notebook, the first cell downloads and installs the versions of the julia package dependencies that were used for the paper's analysis. The table below indicates which notebook each figure was generated in.

#	File Name	Notebook	Short Caption
1	NYC_pairwise_mean_se	NYC_analysis-2015	Map of NYC showing the property sales and estimates of the inverse-variance weighted local average treatment effect between pairs of school districts.
2	NYC_cliff_face	NYC_analysis-2015	Estimate of the difference in log prices per square foot at the border between school districts 19 and 27.
S-1	confounding	(drawn in a vector graphics editor)	Illustration of the confounding due to spatial variation in the projected 1D RDD method.
S-2	mississippi_projection_methods	Mississippi_projection_illustration	Illustration of the projected finite-population and projected-land local average treatment effect estimators.
S-3	wiggly_boundaries_setup	Wiggly Boundaries	Setup of the wiggly boundaries simulations.
S-4	wiggly_boundaries_posteriors	Wiggly Boundaries	Results of the wiggly border simulations.
S-5	weight_functions	Wiggly Boundaries	Illustration showing the behavior of the border and unit weight functions for each local average treatment estimator.
S-6	mississippi_sim	Mississippi_sharp_null_sims	Set-up of an imaginary experiment at the Louisiana-Mississippi border
S-7	NYC_placebos	NYC analysis	Placebo tests for additional significance tests.