Certainly it's difficult to imagine someone motivated by accuracy rather than making political points about migrants (and education and desire to work) concluding that what industry people have worked in, the years of experience they have, the level of seniority they've obtained, how much of their previous work is in Denmark, previous salaries they've earned and whether they're willing to take a lower paid job etc are all less important than whether they're non-Western and what their highest level of education is...

Admittedly, it's also the type of model you'd expect someone working back from very limited available data to come up with, but I struggle to believe it's a particularly accurate model

They could well be the few factors that, in aggregate at a population level, give the greatest explanatory power. And which have then been roughly calibrated.

The resulting model can be well be accurate in aggregate for informing policy decision tradeoffs. Just totally irrelevant for an individual's particular circumstances...

Suspect if they're getting more explanatory power from which part of the world their family originated from than any kind of indicator of actual employment history, it's because they either don't have enough usable employment history data or they badly misspecified the employment history indicators. Agree that the macro model they have might nevertheless be relevant to certain policy decisions, but not individual ones.

If this algorythm was a fancy neural net, we wouldn't even be able to inspect it and have this discussion

Very interesting. I'm working on visualization and explainability of ML algorithms.

Thank you for posting this.

I doubt it is a real ML algorithm, there is some discussion of the algorithm in the comments, so it seems the one currently in use is very downgraded from an earlier one that had more variables but not quite sure about that part.