Transforming Data to Make Better Predictions

Presenter: Tom Donnelly

Understanding and Comparing Data Transformation

Transforming data can be useful to stabilize the variance and make the error more uniform across the design region. It serves to remedy lack of fit and plot predictions in a way that does not violate physical limits, display negative counts or erroneously report yields as greater than 100%.

See how to:

  • Transform data on the fly using Graph Builder and change scales to improve graph readability and interpretability
  • Use square root transformation to eliminate negative values and examine how using Box-Cox power transformation on response might change fit
  • Compare no transformation, log transformation with prediction in raw units and log transformation with prediction in log units
  • Use square root transformation to construct model effects using Polynomial to Degree 2 macro, identify best transformation, use selected transformation, save residuals to data table and evaluate results using distributions
  • See how square root transformation might display a more uniform count spread and more linear relationship

Back to Top