Cross-calibration of categorical variables: An evaluation of the genetic algorithm approach

This paper highlights the importance of the cross-calibration of categorical variables, models cross-calibration as the forecast of a joint probability distribution, and proposes a non-traditional method that can be applied to any observed sample of joint data points. The sample is generally distorted due to measurement errors and differences among raters. The approach uses a genetic algorithm that predicts the true joint probability of two categorical variables. Unlike existing methods, the proposed approach does not explicitly account for any prior knowledge, does not impose any constraint, does not define a specific agreement, and does not specify the type of dependence that exists between the variables. However, the approach produces good logical estimates of the probability forecast both at a specific point in time and longitudinally across time. The computational investigation quantifies this performance using different scoring measures and provides computational evidence of its validity and superiority.