On a measure of information gain for regression models in survival analysis