Signing bias in fixed effects models.