Why comparative effort prediction studies may be invalid

  • Authors:

  • Barbara Kitchenham;Emilia Mendes

  • Affiliations:

  • Keele University, Keele, UK;University of Auckland, Auckland, New Zealand

  • Venue:
  • PROMISE '09 Proceedings of the 5th International Conference on Predictor Models in Software Engineering
  • Year:
  • 2009

Quantified Score

Hi-index 0.01

Visualization

Abstract

Background: Many cost estimation papers are based on finding a "new" estimation method, trying out the method on one or two past datasets and "proving" that the new method is better than linear regression. Aim: This paper aims to explain why this approach to model comparison is often invalid and to suggest that the PROMISE repository may be making things worse. Method: We identify some of the theoretical problems with studies that compare different estimation models. We review some of the commonly used datasets from the viewpoint of the reliability of the data and the validity of the proposed linear regression models. Discussion points: It is invalid to select one or two datasets to "prove" the validity of a new technique because we cannot be sure that, of the many published datasets, those chosen are the only ones that favour the new technique. When new models are compared with regression models, researchers need to understand how to use regression analysis appropriately. The use of linear regression presupposes: a linear relationship between dependent and independent variables, no significant outliers, no significant skewness, no relationship between the variance of the dependent variable and the magnitude of the variable. If all these conditions are not true, standard statistical practice is to use a robust regression or transform the data. The logarithmic transformation is appropriate in many cases, and for the Desharnais dataset gives better results than the regression model presented in the PROMISE repository. Conclusions: Simplistic studies comparing data intensive methods with linear regression will be scientifically valueless, if the regression techniques are applied incorrectly. They are also suspect if only a small number of datasets are used and the selection of those datasets is not scientifically justified.