Probably better off creating an index if the goal is dimensionality reduction.

Comments welcome! We are currently working on a more econ-audience version of this paper which will replicate a few existing studies that use remotely sensed dependent variables.

We outline several other ways this type of error can occur in the paper, and propose methods for fixing it - described by @lcsanford.bsky.social in his thread above.

Another example - say you are measuring air pollution using satellite data. We know that satellite measures of pollution saturate at higher concentrations. If your treatment only has an effect in highly polluted areas, the satellite data might detect no change, even if there was an effect.

Imagine that you are measuring deforestation and your ML model sometimes confuses forests for irrigated cropland. If your treatment affects cropland, this might mistakenly look like an increase in forest to your model.

But it is really differential measurement error that matters - errors that are correlated with treatment status. How can this happen, even in an RCT?

One thing that sometimes surprises people, is that, even in an experimental setting, treatment effects can be biased if the dependent variable has measurement error. This is clearly the case if the dependent variable is a prediction from a machine learning model.

not that i know of unfortunately but you can try messaging the organizers

Awesome it looks like I am staying close to there

In 2015 Pakistan was signing 30 year contracts with coal producers. That was the year solar became cheaper than coal.