Major changes since the first edition include the following:

- Conversion to R with R used to run all examples in the text
- Conversion of the book source into knitr reproducible documents
- All code from the text is executable and is on the web site
- Use of color graphics and use of the ggplot2 graphics package
- New text about problems with dichotomization of continuous variables and with classification (as opposed to prediction)
- Expanded material on multiple imputation and predictive mean matching and emphasis on multiple imputation (using the Hmisc aregImpute function) instead of single imputation
- Addition of redundancy analysis
- A brief survey or new directions in predictive modeling
- Added a new section in Chapter 5 on bootstrap confidence intervals for rankings of predictors
- Replacement of the U.S. presidential election data with analyses of a new diabetes dataset from NHANES using ordinal and quantile regression
- More emphasis on semiparametric ordinal regression models for continuous Y, as direct competitors of ordinary multiple regression, with a detailed case study
- A new chapter on generalized least squares for analysis of serial response data
- The case study in imputation and data reduction was completely reworked and now focuses only on data reduction, with the addition of sparse principal components
- More information about indexes of predictive accuracy
- Augmentation of the chapter on maximum likelihood to include more flexible ways of testing contrasts as well as new methods for obtaining simultaneous confidence intervals
- Binary logistic regression case study 1 was completely re-worked, now providing examples of model selection and model approximation accuracy
- Single imputation was dropped from binary logistic case study 2
- The case study in transform-both-sides regression modeling has been reworked using simulated data where true transformations are known, and a new example of the smearing estimator was added
- Addition of 222 references, most of them published 2001-2014
- New guidance on minimum sample sizes needed by some of the models
- De-emphasis of bootstrap bumping for obtaining simultaneous confidence regions, in favor of a general multiplicity approach.