> This is a weird concern to raise -- the idea that EF's data isn't representative because it's pulled from a sample of citizens extraordinarily interested in English would strongly imply that the true proficiency level in every country is much lower than reported, and to a lesser extent that the gaps between countries are probably wider than reported.
No, if your samples aren't representative then any conclusion you draw can be flawed for thousands of different reasons and in all possible directions. Two examples.
1) Let's say that the top 1% of English speakers of countries A and B took the test. You say "the gaps then would be wider than reported". Wrong, even if in this 1% sample country A scores the highest, it might very well have lower overall proficiency. If proficiency is normally distributed, but has a much greater variance in country A than country B, the maximum of A will be higher than of B, even with a lower average.
2) Now you take the tests at time t and t+1. However, at time t the test was less popular, so only 0.5% of the top speakers took the test, while at t+1 1% of the top speakers took the test. You would conclude that a lower score a time t+1 implies that the skill got worse, while it might have very well improved by a lot.
(In fact, I'm pretty sure that's what happened. I never heard of EF before 2017, so I couldn't have taken the test before then)