Analysis of the type 2 error of the PARDelta method for different Delta values

The empirical evaluation of algorithms yields a large set of data that needs to be statistically evaluated. Commonly, the generated set has some missing inputs due to an algorithm's inability to find a solution for a given test case. These missing data strongly limit the use of statistical tests, as classical hypothesis tests cannot be directly applied in the evaluation of these datasets. The ParDelta method emerges as a pre-processing method that inputs the missing data, creating a complete dataset that can be statistically analyzed. The literature uses the PARDelta method with Delta = 10. However, there are no justifications for such a value. In this work we evaluate the power of the PARDelta method for Delta values between 2 and 50 employing a dataset generated by an integer linear programming resolver. The results showed little variation for the different Delta values.