Statistical Issues on Optimization for Software Metric Models with Missing Data

When developing new software, software metric models are often applied in predicting certain important elements such as total work effort or error rate and so on. The procedures during the regression model construction using certain historical data, such as determination of the independent metrics, imputation of missing values and combining levels for independent categorical metrics, have been discussed already. In this paper, how to choose some important parameters for the proposed procedures during the model construction is considered in depth. The selection of critical parameters in the k-nearest neighbors (k-NN) multiple imputation is specifically focused. An example is given for illustration with data from widely used database.