A Welfare Criterion for Models with Biased Beliefs