Analyzing data sets with missing data: an empirical evaluation ofimputation methods and likelihood-based methods

摘要

Missing data are often encountered in data sets used to constructnsoftware effort prediction models. Thus far, the common practice hasnbeen to ignore observations with missing data. This may result in biasednprediction models. The authors evaluate four missing data techniquesn(MDTs) in the context of software cost modeling: listwise deletion (LD),nmean imputation (MI), similar response pattern imputation (SRPI), andnfull information maximum likelihood (FIML). We apply the MDTs to an ERPndata set, and thereafter construct regression-based prediction modelsnusing the resulting data sets. The evaluation suggests that only FIML isnappropriate when the data are not missing completely at random (MCAR).nUnlike FIML, prediction models constructed on LD, MI and SRPI data setsnwill be biased unless the data are MCAR. Furthermore, compared to LD, MInand SRPI seem appropriate only if the resulting LD data set is too smallnto enable the construction of a meaningful regression-based predictionnmodel