Cross-country Experiment

[tm-kah-alforja]

Update as of 01/25:

Able to train on diff country splits and leave 1 country out for training. Initial findings: Training the model this way is not generalizable and performance depends on the split.

Next steps:

• EDA on country split differences that impact the result
• normalize wealth index features

[tm-kah-alforja]

Distribution of wealth and features

• Plotted wealth index distribution per country (see first pic). TL, KH, and MM seem to skew more left compared to PH. PH also covers a wider range of index values

Normalizing wealth indexes

• Normalizing wealth indexes over all countries has negligible effect on performance, we get almost the same R^2 value compared to no normalization

Normalizing wealth index per country improves the results for 3 out of 4 splits. Improves results overall.

• Mean split r^2: 0.41 (0.13) Previously: 0.32 (0.3)