Biases in twin estimates of the return to schooling