Measuring the Landscape of Civil War

This is a package, documentation, and replication repository for the paper "Measuring the Landscape of Civil War" (provisionally accepted for publication) 2017

The Paper:

Measuring the Landscape of Civil War - Read the Paper

Measuring the Landscape of Civil War - Read the Online Appendix

The Authors:

• Dr. Rex W. Douglass(University of California San Diego)

• Dr. Kristen Harkness(University of St. Andrews)

Replication Code and Analysis

Self Contained Package

All of the files necessary for reproducing our analysis are including in a self contained R package "MeasuringLandscape." You can install the package MeasuringLandscapeCivilWar from github with:

# install.packages("devtools")
devtools::install_github("rexdouglass/MeasuringLandscape")

R-Notebooks

The analysis and figures in the paper and statistical appendix are produced in a number of R Notebooks.

NOTE: Several parts of this analysis are stochastic, specific coefficient estimates and p-values will vary with each execution. Substantive results will be consistent across runs. We encourage the reader to run the replication multiple times and observe the variation.

• 00 Project Setup: Useful commands for installing necessary packages and setting up the project.

File Preparation:

• 01 Prep Events Counts: Loads and cleans a novel dataset of violent events observed during the 1950s Mau Mau Rebellion.
• 02 Prep Gazetteers: Cleans and combines a large number of gazetteer of place names for looking up locations by name and retrieving their coordinates.

Fuzzy Matcher: A supervised learning pipeline for matching two placenames to one another even when they are spelled slightly differently.

• 03 Fuzzy Matcher Stage 1 (Locality Sensitive Hashing)
• 04 Fuzzy Matcher Stage 2 (XGBoost)

Georeferencer: A supervised learning pipeline for assigning a real-world coordinate to a placename.

• 05 Georeferencer: Takes in locations of events described as text and returns all possible matches across different gazetteers.
• 06 Ensemble and Hand Rules: Ranks the returned matches from best to worst. First, using simple hand rules of what kind of match to prefer over others. Then second, with a supervised model that attempts to predict which match will be geographically closest to the true location (fewest kilometers away from the right answer).

Analysis: Main analysis of the paper.

• 07 Recall Accuracy: Rate georeferencing options in terms of recall (how many event locations they recover) and accuracy (how far away their imputed locations tend to be from the true location)
• 08 Predict Missingness DV: Rate georeferencing options in terms of how systematic they are at recovering locations for certain kinds of events but not others.
• 09 Predicted Effects: Demonstrate what kinds of events tend to systematically get excluded. Here, in terms of whether the event would have received an original military coordinate or not.
• 10 Bias: Demonstrate that the kinds of locations that are imputed are different from the true locations, in terms of things like population, distance from roads, ruggedness, etc.
• 11 So What: Demonstrate that different georeferencing decisions will produce different results in a simple linear regression model in terms of both statistical significance and substantive effects.