Measuring the Impacts of Teachers: Comment

North Carolina statewide administrative panel

North Carolina public-school students and teachers.

Replicates the Chetty-Friedman-Rockoff design in a different state and adds a placebo test: value-added should not predict students' PRIOR-year scores. It does, which implies assignment is not random conditional on the controls.

The most serious challenge to value-added: a placebo test that value-added fails in North Carolina data, bounding potential bias at 10-35% of VA variance. The exchange does not overturn VA, but it should stop anyone using single-year VA as a high-stakes number.

This IS the confound argument: students are sorted to teachers on characteristics (partly heritable, partly parental) that prior-score controls do not fully absorb.

Answered by Chetty, Friedman & Rockoff's reply (doi:10.1257/aer.20170108) and by Bacher-Hicks et al., who show the placebo test rejects designs that are valid by construction in simulation. The honest unresolved band is 'bias modest, not reversal'.