10x plan for deconfounding with causal EO-ML

[cjerzak]

BIG PICTURE: Our goal is to make a large-scale diff-in-diff analysis of the effects of aid in Africa using the framework of Cindy Conlin's thesis.

IMMEDIATE NEXT STEPS:

1. Mattias + Connor setup cluster (Markus will save us when we're in a tough stop)
2. Mattias finds Cindy's image data; if unable to find or access, re-download
3. Replicate(ish) Cindy's initial findings using the imagery + the hugging face data
4. Do minor changes / tweaks that Sarah V. wanted

MEDIUM-TERM NEXT STEPS:

1. Formal diff-in-diff analysis (consider the role of measurement error)
2. We might have to re-download data?
3. Consider using the new v2 version of the outcome imputations from Markus (if available)

WHAT'S THE INTUITION?

The intuition is that the deconfounding approach first outlined in Cindy Conlin's thesis used reasonable but still imperfect assumptions about how the imagery contains all the confounding information about the relationship between the treatment (AID) and outcome (LIVING CONDITIONS). The idea here is to exploit over-time variation in the treatment to "partial out" some of the static confounders not present in the imagery. The imagery will still be used as ADDITIONAL confounding information. We may also use this approach to explore heterogeneity in a way that we didn't before (i.e., WHAT KINDS OF PLACES where most receptive to the aid). Consider also using information from audits on the aid programs (?).

WHY IS THIS GOOD TO WORK ON?

Mattias learns more about ML/EO. The aid community learns more about the effects (marginally and conditionally) of development programs. The geoAI community will see how to handle large-scale Planetary Causal Inference analyses with natural experimental designs, measurement error handling, heterogeneity explorations. We'll also get more information about poverty traps (and how to detect/avoid).

[adeldaoud]

I really like this plan!