Customer Reviews

Exploratory Data Analysis

5 star:		(5)
4 star:		(0)
3 star:		(1)
2 star:		(1)
1 star:		(0)

 

 

Tukey's EDA is a ground-breaking text, one that is as rich in extraordinary ideas and approaches to data analysis in 1998 as it was in 1977.

An earlier reviewer on this web page dismissed the EDA book as a pre-PC contribution, a dinosaur of the slide rule era. The comment recalled for me the experience of a colleague who, years ago, had the opportunity to visit the UIUC office of John Bardeen, who took Nobel Prizes in Physics in 1956 and again in 1972. He said that he went to the interview expecting to be impressed by the computing technology in this guy's office, but was dismayed to find that Bardeen's main tools were pencils, good quality graph paper, and a slide rule. Good enough for him to co-invent the transistor....

As my friend's experience reminds us, it's the ideas, not the tools, that really count. Tukey and his EDA book are gold mines of fresh ideas and approaches to data analysis. I recommend it strongly and without any reservation to every researcher who is interested in data analysis. It is a marvelous read.

...and lest anyone who is unfamiliar with Tukey's contributions feels that he is a light-weight on the computer side of things, he _did_ give computerese such fundamentals as the word "bit".

In the preface, Tukey writes, "this book ... exists to expose its readers and users to a considerable variety of techniques for looking more effectively at one's data." It succeeds remarkably well and is still relevant: after over 40 years, many of the techniques still have not been incorporated in commercial software and some of those that have (such as box-and-whisker plots and robust smoothing) are often emasculated.

This book has served me well for decades: I have used most of its techniques in my statistical consulting practice and, more recently, have used it as a foundation for courses in data analysis that range from a few hours to an entire semester. Students always appreciate the practical experience and set of tools they acquire. The more experienced ones comment on the insight: "I never fully understood what the box-and-whisker plot really did until now" is a recent example from a mid-career engineering professional.

Nevertheless, it is true that some of the material is outmoded due to its focus on manual calculation and some of the rest may be too idiosyncratic for most. What remains--which is plenty--can be studied on its own, because this book is designed for self-study: most of the chapter groups are independent of all but the introductory material, they provide detailed examples, ask many thought-provoking questions, and supply many datasets for practice. Tukey's methods speak for themselves through the gains in insight they provide, so he is content to show *how* to do them and to provide copious examples. What he does not do is supply the mathematical theory. If you like, you can read about that in Hoaglin, Mosteller, and Tukey's "Understanding Robust and Exploratory Data Analysis".

The highlights of this book, in terms of techniques, are:

* Chapters 1-4 on graphing data and on basic, useful data summaries: stem-and-leaf plots and n-letter summaries. Most statistical software now provides these. They are of primary interest as building blocks in more advanced analyses.

* "Straightening out plots:" simple, effective techniques to re-express the independent and dependent variables in a bivariate scatterplot so that the relationship becomes approximately linear. (Chapters 5 and 6.) I am not aware of any software that does this, but the techniques are so simple and elegant you can still carry them out with pencil and paper (or a spreadsheet) even with huge datasets.

* Smoothing sequences (chapters 7, 8, and 16). The "3rssh" method has largely been displaced by Lowess smooths in software, but the principles and ideas of smoothing, "roughing," and "re-roughing" are timeless.

* Analyses of two-way tables using median polish. This technique has recently been exploited in various fields, including spatial statistics, but only in the most elementary way. Think of this as a robust version of Analysis of Variance, with a focus on finding the effects, but without any of the mathematical apparatus. (Chapters 10 and 11.) One of the best tools for performing median polish is a spreadsheet.

* Advanced fitting of two-way tables: adding an interaction term; transforming the dependent variable; plotting the fits. (Chapters 12 and 13.)

* Techniques for re-expressing and analyzing counts and fractions. (Chapters 15, 17, and 18.)

Additional material covers "delineations" of scatterplots (chapters 8, 9, and 14) and assessing distributions (chapters 19 and 20). The latter suffers in retrospect from not using probability plotting methods. The former is worth learning but there does not seem to be any simple way to use modern statistical software to create these delineations.

In brief, this book requires no more mathematical prerequisite than facility with arithmetic, but after working through it, the diligent reader will come away with a body of techniques for understanding almost any kind of data set, including methods of time series analysis, regression, analysis of variance, and contingency table analysis.

Wow. I had assumed this was out of print, but here it is. This is simply a remarkable book about the basics of thinking numerically. Even if you do your stats on a computer I recommend this. I use it to show Arts students that quantitative analysis can be fun, powerful and simple.

I reviewed this book when it first appeared in 1977 -- having been reading preliminary versions for about a decade before that.
At the time I offered the not-very-prescient opinion that it would quickly become a classic. It has.

The various reviews on this site are all correct. Yes there are still wonderful ideas to be mined (easy univariate transformations to symmetry, transforms to linearity to aid in curve fitting, the crucial importance of robustness, of looking at outliers and fringeliers, and the dominant role that graphics plays in forcing us to see what we never expected).
And yes, there is some material that is out-dated (e.g. using break points to estimate logs in your head).

But these are beside the point -- Scholars still study Talmud and Newton's Principia. Why? Obviously there are many reasons varying with the work and the reader.
For me a principal reason for reading (and rereading) EDA is to get a close look at how a first class mind works, with the hope that when faced with a similar problem we can then try to emulate him. To this day as I read EDA I can see Tukey's smiling face patiently explaining to me how to look at data and exhorting me not to miss subtle hints.

This was a book to be treasured when it first became available, and I see no reason for that judgment to change now, or in the future.

This was a one of my recommended Text Books for my Undergraduate degree in Statistics. I never realized at that time how important a book this is. Year later I had to buy a second hand copy (I wonder what happened to mine) as the book was out of print. The fact is, now with computers we can just push the button and the machine grinds the data through the various parametric analyses, it will fit lines, whether it is appropriate to do so or not. That is where we have missed the data analysis plot. Tukey exposes us to the art of data exploration. How much we can learn from the data just visually without making (probably invalid) assumptions. If you are the sort of person who just drinks from the milk carton without sniffing it first, don't buy this book, and good luck to you. If you like me sniff, these are techniques you will need to sniff out the qualities of the data, before you drink it! Its back in Print - Buy it - it's a bargain

I was not surprised to see that one of the issues raised by previous reviewers is whether or not this book is still relevant, given the enormous improvements in computing since this book was written. While I recognize that there are a lot of good ideas in this book, I must say that far too much of the text is devoted to techniques that are only relevant if one is doing data analysis on paper. This is certainly not to slander Tukey's own extensive knowledge of computing -- simply an admission of the fact that when EDA was written, few readers would have had access to any computing resources, let alone anything comparable to those that are available today. Today, the books suffers for this. A new version of this book, updated to take advantage of today's computing resources, would be quite valuable.

From everything I've read and heard, this was the authoritative book on the subject at one time. But, having been written in 1976, before PCs, it discusses data analysis using slide rules for computation and rulers for graph analysis. I need something to help me understand predictive modeling on SPSS. Is there something better?