Published online in Wiley Online Library (wileyonlinelibrary.com).Wickham and colleagues have provided a nice summary of a data-driven approach to exploratory analysis of statistical models. We would summarize the central themes as:1. Plot as much of the raw data as possible and overlay the model fits and parameters.2. Display multiple models from a collection to understand structure in the data.Some of the concepts discussed are fairly common in applied data analysis-for example, it is common to visualize the boundaries of a decision rule. Other ideas are more infrequently applied, like taking "grand tours" through the high-dimensional data space through a sequence of twodimensional slices. In general, the intuition of displaying as much data as possible is natural among practicing data analysts, but the intuition is frequently based on ad hoc experience. This paper is a nice synthesis of these intuitive ideas and provides some examples of new potential visualizations that expose specific features of complex models.Our intuition agrees with Wickham and colleagues: that displaying more data is better than displaying less. In considering the paper we were motivated by the question: What is the objective of visualizing models?. In their culminating example in Section 6, Wickham and colleagues use grand tours and decision boundary plots to understand the operating characteristics of the neural network procedure. The path suggested by this example moves from data visualization, to model fitting, to model visualization, and culminates in model understanding (Fig. 1, teal path).We envision this path as a useful approach for teaching about model fitting methods on idealized data sets. For example, visualizing the hidden nodes within the neural network (Wickham and colleagues Figure 22) highlights * Correspondence to: Jeffrey T. Leek (jtleek@jhu.edu) the way that a neural network combines multiple logistic boundaries to arrive at a nonlinear classification boundary. This type of plot is an excellent teaching tool for ensembling methods-similar plots are helpful in explaining the AdaBoost boosting algorithm (e.g., Slides 17-19 of http:// bit.ly/1EveDiP, re-hosted from: http://webee.technion.ac.il/ people/rmeir/BoostingTutorial.pdf). This plot can certainly be used to improve students' understanding of the mechanics of a particular class of models and could unearth ways to improve the algorithm itself.In our experience, a more typical objective in day-to-day data analysis is to use model visualization for the purpose of checking and updating model fits before summarization (Fig. 1, orange path). It is less clear to us that the proposals of Wickham and colleagues are suitable for the objective of model correction. Human beings have a difficult time inferring correlations or statistical significance from even simple scatterplots or boxplots [1][2][3]. While we agree that showing as much of the data as possible makes intuitive sense, we wonder how people will actually interact with the visualizations suggested in the paper...