Abstract:
|
Data transformation is a commonly used technique to make data more normal and to allow inference from linear regression. Transforming data has benefits: it enables data to meet assumptions needed for inference, higher power can be achieved, and certain transformations are less sensitive to outliers. Transforming also has disadvantages: interpretation can be difficult, significant results on one scale do not mean significant results on another scale, and there can be arbitrariness in the choice of transformation. In new-onset Type 1 diabetes clinical trials, an LN(mAUC+1) transformation is routinely applied to the primary endpoint, C-peptide mean area under the curve (mAUC). C-peptide mAUC is best summarized and understood as measured, and so we argue that the best way to analyze these data may not involve transformation. This conclusion could also apply to other continuous measures that do not have a natural interpretation on a transformed scale. The following alternative ways to evaluate the data are presented: loess regression/splines, linear regression with robust standard errors, bootstrap methods, generalized linear models, and quantile regression.
|