Learning from Failure: Validating and Refining Structural Models with a School Voucher Experiment

This paper leverages a unique experiment that randomized access to school vouchers across markets and students to study what rigorous out-of-sample validation of structural econometric models delivers in practice. We hold out data from treatment markets while developing and estimating school choice models on control markets alone, and commit to predictions before analyzing the treatment data. Across a range of modeling assumptions and specifications, we find that all models underpredict demand for the experimental voucher offer. Rather than treating this failure as a verdict, we use it to diagnose why the models failed: the program induced household search and led private schools to use program surplus to attract students. A refined model reconciles the choice patterns induced by the experiment with those in control markets, providing a credible basis for evaluating alternative voucher policies. Its estimates imply large welfare gains from vouchers targeted to households with limited ability to pay.