Infodemic-replication

Replication files for "Battling the Coronavirus 'Infodemic' Among Social Media Users in Kenya and Nigeria"

• arXived here
• Pre-registration here.

Note: The analysis reported in the paper is implemented using macOS, platform: aarch64-apple-darwin20. Due to differences in compilers across platforms, if this code is run on a Windows machine, the random forests used for policy learning and evaluation may have slightly different splits, even using the same random seed. This behavior is noted in the grf package documentation here. As a consequence, point estimates may vary slightly from those reported, and the restricted policy learned may have slightly different assignment rules; our replications suggest that these variations do not change the substantive findings reported in the paper.

Reproducing results

Python is required to generate the contextual probabilities, but a copy of the probabilities is saved in this replication repository. R is required for primary analysis

1. Save most recent data in the data/ folder.
2. Generate contextual probabilities for adaptive weights for inference:
   • You may wish to create a conda environment if you are re-generating the contextual probabilities:

   conda create --name contextual_probs python=3.7
   conda activate contextual_probs
   source code/contextual_probabilities/install.sh    
   

   • Then, re-generate contextual probabilities for adaptive weights for inference:

   cd code/contextual_probabilities
   Python gen_probabilities.py
   Rscript convert_contextual_probs.R
   

   • File code/contextual_probabilities/gen_probabilities.py creates an object that records treatment assignment probabilities for every observation at every point in time, conditional on an algorithm and observed data.
     • Depends on:
       • code/contextual_probabilities/utils.py
       • data/cleaned-data_*.csv
     • Generates data/contextual_probabilities.npy
   • File code/contextual_probabilities/convert_contextual_probs.R converts numpy probability object to rds
     • Depends on data/contextual_probabilities.npy
     • Generates code/objects/contextual_probabilities.RDS
3. If you are NOT using a conda environment, install R packages outside of conda:
   • Start in the infodemic_replication directory
   • Run Rscript --verbose code/r_packages.R
4. To replicate analysis, compile
   • code/misinformation_replication.Rmd (primary analysis in paper)
     • Depends on:
       • code/utils.R
       • data/cleaned-data_*.RDS
     • Generates:
       • Resulting html file (misinformation_replication.html) will be saved in the same folder.
       • Figures and latex tables will be saved in figures/ and tables/ folders respectively.