(by Michael Lavine)
After seeing Andrew Gelman’s piece on Information Visualization versus Statistical Graphics I read the pieces in Statistical Computing and Graphics Newsletter and have two comments.
1. The piece on InfoVis didn’t convince me. In the first example, the author says that the value of the spiral graphic lies in discovering that more people call in sick on Mondays than on other days. But anyone examining the data with standard graphics would quickly make the same discovery. The fancy graphic adds no value.
The graphics in the second example is just a sequence of scatterplots. The clever part of the analysis is removing the data known to be good, and having that done by computer rather than by hand. There is nothing special about the graphics, unless the author has omitted part of the explanation.
2. One advantage of graphical analysis is that it facilitates looking at individual data points, rather than merely summaries. It’s not that one can’t look at individual data points in a table — one can — but it’s daunting. It’s much easier to see the points in a graph. And seeing the individual points can often lead to good things. That’s at least part of the reason for making many plots of residuals, fitted values, leverage values, and so on.
For an example, consider the graphic I posted yesterday about how individual points contribute to the likelihood ratio or AIC, BIC, and DIC. LR, AIC, BIC, DIC are one-number summaries of goodness-of-fit, and users would never think to read a table of the contributions from individual points. But the graphic shows immediately that the one-number-summary is dominated by just one of those points.
To repeat, one of values of graphics is not the graphics per se, but that graphics encourage users to look at smaller units of analysis — often individual points — rather than just summaries