In this work, we provide a procedure for the systematic evaluation of the quantitative impact of the selection of the basic intonation unit for data-driven intonation modeling. Taking advantage of the corpus based modeling technique previously developed, we show how the number of prosodic features selected and the kind of basic unit determines the final prediction RMSE of the synthesized F0 profiles. This provides a means to fully characterize a given corpus and, also, a procedure to test how 'a priori' linguistic knowledge corresponds with real samples in that corpus. Although all the results shown are for Spanish, the methodology can be readily applied to any other similar language for which a similar annotated corpus is available.