The effect of log transformation in multivariate liner regression models for software effort prediction

Multivariate linear regression models have been commonly used as software efort prediction models. To improve the prediction accuracy, it is a common practice to transform (especially, log-transform) the data before building a model, although its theoretical basis is not necessarily clear. This paper reveals that the log-transformed linear regression model (log-log regression model) is equal to the exponential model, which is suitable to characterize various relationships among software related metrics. However, when using a log-log regression model, the result of inverse transformation tends to under-estimate the efort. This paper also introduces a method to correct such bias.