Pitfalls in predictions of rock properties using multivariate analysis and regression methods

Statistical methods are commonly used for prediction of geoscience and engineering properties. This commonly involves selection of a small number of variables among a large number of available geological, geophysical, petrophysical and engineering variables. The conventional view is to select the variables that have highest correlations with the variable of concern. In this article, we show that this may not always be a wise approach because it ignores a critical aspect of the variable interaction — suppression. We review the suppression phenomenon, and discuss three types of suppression in multiple linear regression of geoscience and reservoir properties. We present examples using wireline logs, seismic attributes, and other engineering parameters. We show that understanding the suppression phenomenon is important for selecting appropriate variables for optimal prediction of geoscience and reservoir properties.

► We discuss multivariate analysis and modeling of geoscience phenomena. ► Pitfalls in selecting variables for multivariate analysis are highlighted. ► Examples from wireline logs, seismic attributes, and reservoir variables are given. ► Two interaction phenomena include redundancy and suppression. ► We demonstrate the importance of understanding these variable interactions.