[mjohn]

It's worth bearing in mind that this is not a study on educational outcomes, but is assessing how useful value-added models are for quantifying teacher performance. While the biggest correlative factor may well be socioeconomic status, that does not preclude teacher's having an impact on within-group differences.

From the introduction:

> In this paper, we provide a stark illustration of the limitations to using value-added models to identify high-and low-performing teachers. We do this by applying commonly estimated models to an outcome that teachers cannot plausibly affect: student height. Aside from the implausibility of teacher effects on height, student height is an attractive measure for this exercise since it is symmetrically distributed, interval measured, and arguably less prone to measurement error than achievement. We find that the estimated teacher “effects”on height are nearly as large as the variation in teacher effects on math and reading achievement.

It's also an interesting approach: take a model that is apparently predictive and see if it's also predictive of something that we know is unrelated. By showing that value-added models are displaying spurious correlation maybe policymakers will take note.

Could maybe take a similar approach to validating our ML models?