This is the repository for "Turning the Lights on to Keep Them in the Fold: How Governments Preempt Secession Attempts."

Replication

All analyses reported in the paper and supplemental information can be replicated using the following data files found in the Input Data directory:

• groups_nightlights.RDS
• gadm data.RDS
• grid data.RDS
• myers_data.RDS

The primary analysis dataset groups_nightlights.RDS is created by nl_prep.R, which merges variables from many additional data sources into the primary geospatial data output, group data.RDS. Datasets are also included in .csv format for software-independent use.

Results replication

This section assumes that you are starting from the datasets in the Input Data directory and not creating all datasets from scratch. All R scripts are located in the Code directory. To replicate the results:

1. Run Analysis nl_prep.R via Analysis nightlights prep.sl
2. Run all remaining Analysis_nl_*.R files via their respective SLURM scripts
3. Run results_plots.R
4. Run acd_trend.R, grid_plots.R, Map Creation.R, nl_sudan.R, and separatism.R to generate results

To create group data.RDS, gadm data.RDS, and grid data.RDS from original spatial data sources, see the section below.

Geospatial data replication

To recreate all geospatial data from scratch, the workflow is as follows:

1. Linearly interpolate population rasters with Raster Prep/Count Interpolation.R
2. Perform intercalibration between satellites on nightlights data with Raster Prep/Nightlights Intercalibration.R
3. Create overlap-corrected rasters with Raster Prep/Overlap Correction hi-res.R and Raster Prep/Overlap Correction lo-res.R a. Note! These R scripts were originally run on an older cluster, and the resources requested by the SLURM submission scripts are a best estimate of the resources needed to complete the job since re-running them on Longleaf would be computationally wasteful. These jobs may fail due to out of memory errors and require adjustment of the requested resources in their SLURM submission scripts.
4. Fix topology errors in GeoEPR with GeoEPR Cleaning.R
5. Generate geospatial datasets with Group Variable Creation.R, GADM Variable Creation.R, and Grid Variable Creation.R

Once this is done, use Analysis nl_prep.R as described above to generate the analysis dataset and proceed with analyses.

Computing environment

Replicating these scripts requires access to a high performance computing (HPC) environment, such as a university cluster. Most R scripts are written to be run via job scheduler software in an HPC environment. These job submission scripts are written for the SLURM Workload Manage and have a '.sl' file extension. They may require adapting to work with a different HPC environment even if that environment uses SLURM. Of special note are the following scripts:

• Code/Group Variable Creation.sl
• Code/GADM Variable Creation.sl
• Code/Raster Prep/Overlap Correction hi-res.sl
• Code/Raster Prep/Overlap Correction lo-res.sl
• Code/Raster Prep/Nightlights Intercalibration.sl
  • See Code/Raster Prep/README for information on this script

They are configured to use the 'bigmem' partition on UNC's Longleaf cluster and will likely require editing to use the appropriate large memory partition in other HPC environments. Additionally, the code to load R and any required geospatial libraries uses the "module" command and may need to be modified depending on the HPC environment. To submit a job to the SLURM scheduler, use the following command: "sbatch <script>.sl".

The following scripts are also of special note:

• Code/Analysis nightlights grid border country year.sl
• Code/Analysis nightlights grid pop country year.sl

They are set to request 11 day time limits, the maximum on UNC's Longleaf cluster. However, both may require more runtime than that, and thus may require extending the job time limit by contacting HPC system administrators if your system sets a similar maximum time limit.

The majority of results are the product of Markov chain Monte Carlo estimation of Bayesian models. As such, even with attempts at exact reproducibility via the setting of random number generator seeds, it is possible that results will fail to replicate exactly. This can be the results of operating system differences and differences in external library as well as R package versions.

All analyses were carried out using R 3.5.0.

Analysis scripts

The following R scripts can be run in a standard desktop computing environemnt and thus do not have corresponding SLURM scripts:

• Code/acd_trend.R
• Code/grid_plots.R
• Code/Map Creation.R
• Code/nl_sudan.R
• Code/results_plots.R
• Code/separatism.R

The following R scripts are auxiliary files called from other scripts and should not be run directly:

• Code/cshapes Recode.R
• Code/sfFunctions.R
• Code/Raster Prep/lintemp.R

All other R scripts should be run via their respective SLURM submission scripts.

Datasets

The following datasets are needed to fully replicate the data creation and analyses:

• cshapes 0.6
• GeoEPR 2014
• GPW v3
• GPW v4
• NOAA DMSP OLS v4
• PRIO PETRODATA VERSIONXXX
• GADM 3.4
• UCDP/PRIO Armed Conflict Data 18.1
• UCDP/PRIO Dyadic Data 18.1
• EPR 2014
• ACD2EPR 2018
• FORGE 1.0
• Strategies of Resistance Data Project 2019
• de Jaun and Pierskalla 2015 replication data