Peer Effects Overcontrolling in Value-Added Models
RCTClinical Trial
Including classroom-level baseline mean scores as covariates in value-added models strips away variance that actually represents true teacher effectiveness rather than non-random student composition.
Picture this
Imagine evaluating a gardener's skill by measuring plant growth, but subtracting points if the plants are placed in a sunny greenhouse. If the gardener deliberately created or earned that ideal environment, penalizing them for the setting hides their actual gardening capability.
What the evidence says
The component removed by peer effect controls predicted student achievement after random assignment with a coefficient of 1.150 (p < 0.01, SE = 0.336), proving it contains causal teacher effectiveness. In contrast, the component removed by individual baseline score controls yielded a coefficient of 0.047 (SE = 0.042), showing baseline score controls successfully remove non-teacher sorting noise.
- Who was studied
- N = 27,255 randomized students in grades 4–8 across 619 randomization blocks in 6 urban school districts.
- How
- LIML IV regression decomposing value-added model variance to test whether components subtracted by peer effect controls predict student achievement following random assignment.
What to do
Omit classroom-level baseline mean achievement controls (peer effect covariates) from within-school teacher value-added specifications to prevent subtracting legitimate teacher impacts.
From the source
"In other words, the component removed by controls for peer effects does seem to reflect causal teacher effects-not just factors outside a teacher's control."
Have We Identified Effective Teachers? Validating Measures of Effective Teaching Using Random Assignment