Update cross country rollouts

[tm-jace-peralta]

Problem:

• Current model with StandardScaler predicts little to no negative wealth indices for rollout on certain countries (Cambodia, Philippines, Indonesia)
• The split-quintile approach would fail in the above cases since the split would only contain positive values
• The rest of the rollout countries do have substantial negative wealth indices, but the overall distribution is still overly left-skewed
• Mean of key features had considerably changed from training to rollout. Rollout has much lower mean due to inclusion of remote areas.

Proposed solution:

1. Use MinMaxScaler to simplify interpretation

• Min is stable but similar to StandardScaler, the effect will instead pull down predictions due to scaling using Max
• Comparable R2 values to StandardScaler when tested using leave-one-out for DHS countries
• Interpretation can be improved during income category quintile postprocessing

2. Use normal quantiles for wealth category classification
   • Split quantiles would not work when MinMaxScaler is used since prediction range is [0,1]

PR Contents

1. Retrain cross country model using MinMaxScaler
2. Rerun and edit rollout notebooks using new .pkl file
3. Update expected output map images
4. Add code cell to quickly save features to a file for easier model diagnosis

[review-notebook-app[bot]]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB