Effects of data transformation on multivariate analyses in intracerebral hemorrhage

Abstract

Multivariate statistical approaches have been increasingly applied in hemorrhagic stroke data analysis. Nevertheless, several aspects regarding their relevance and validity in respect of the application of data transformations have not been studied in details. This paper examines the effects of different data transformations in the standard statistical methods of the multivariate analysis of the intracerebral hemorrhage (ICH) parameters in small group samples. Two different methods for data transformations (log transformation (log(Xi )), square root transformation (√Xi ))have been carried out. The initial volume of the ICH have been studied using several test for skewness, kurtosis, histogram distribution method and different quartile-quartile (Q-Q) and probability-probability (P-P) plots as criteria for normal distribution. Multivariate analyses for the prediction of the perifocal edema was performed using raw and transformed data. Our results indicate that the data transformation operations should be performed very carefully because different analytical outputs lead to different scientific conclusions.

Inexpensive spot sampling provides unexpectedly effective indicators of watershed nitrogen status

“…Logarithmic transformation is a simpler and more common solution for analyzing heteroscedastic variables for which the variance increases with the mean (Snedecor and Cochran 1989, Draper and Smith 1998), and log–log regression relating one such variable to another is widely applied in water quality analyses (Helsel and Hirsch 2002). However, statisticians have documented general problems with log transformation, including failure to eliminate heteroscedasticity and difficulty applying parameter estimates or hypothesis tests back to the untransformed variables (Feng et al 2013, 2014, Choi 2016, Greenacre 2016, Rendevski et al 2016, Curran‐Everett 2018, Ekwaru and Veugelers 2018). We chose the bootstrapping approach instead of log–log regression because it handled our heteroscedastic variables without causing these problems and because bootstrapping gave other benefits, such as quantifying sampling uncertainty.…”