Yeah, my immediate reaction was that it looked like a practically uncorrelated scatter plot with most of the correlation signal coming from a few outliers on th

Yeah, my immediate reaction was that it looked like a practically uncorrelated scatter plot with most of the correlation signal coming from a few outliers on the high/low pay ends. At $35k, the outcomes scatter from 100% to 35%, which would be the same magnitude as raising pay from $15k to $55k according to their fit. There are so many systemic differences between those countries that could correlate with pay that interpreting it is pointless without controls. Not to mention that there's a difference between test performance and learning, there could be "optimization by proxy" effects in play.