But you can't measure that.
If you have a model, such as:score_at_end_of_year = score_at_start_of_year + a * parents_income + b * parents_education_level + c * teacher_skill + random_noise
You can figure out the values of a and b by comparing the performance of different students in the same class. If you have a large enough sample size to average out the random noise, you're left with (c * teacher_skill).
Not a perfect method, obviously - but if the best teachers at schools full of poor kids could get more incentive pay by switching to schools full of rich kids, would that be incentivising the right thing?