Much as I love Tufte’s work, he does base much of what he says on aesthetics and his (considerable) intuition. Another approach, of which Cleveland has been an important pioneer, is to draw conclusions based on empirical testing of which forms of data representation are most effective. In their 1991 paper Displaying Proportions and Percentages, Ian Spence and Stephan Lewandowsky compared the efficacy of bar charts, divided bar charts, pie charts and tables. Their focus was on asking three different types of questions about sets of data containing between four and seven data points (A, B, C,…). These were to determine which was greater A or B, A or B+C and A+B or C+D (note that the data was not necessarily presented in that order). They concluded that subjects found it harder to make all three judgements with a table than with either bar charts or pie charts.

One important point here though is that the subjects of the experiment were deliberately selected from groups that were not scientists, statisticians or experts at reading charts and data. So, these conclusions are only really applicable for presenting information to a general audience. Still, that is an important domain to consider.

Table: Cases of Marmalade Sold

         US       UK
Jan    4,000     9,800
Feb    6,000    11,000
Mar    6,400    11,500

Love the karma/dogma line…I plan to steal it!

Sean.

Thank you for sharing. I would feel more sense of closure about Cleveland and bar charts at this moment myself . . . except I’ve written myself into a corner.

Ironically, I noticed today that an Excel app that I’d set up years ago is still being used to report results to my organization using bar charts with, well, partial bars. Clearly, I am guilty of an engineering compromise, made long before I’d stumbled onto this website and learned about how to make dot plots using Excel, or read Cleveland to understand why they are better. My coworkers seem quite content with these bar charts — and it’s not like the abstract index they represent is something one can actually count or measure with a ruler, anyway. No harm, no foul?

Suddenly I’m appreciating your shades of gray, Sean!

What to do when one’s karma runs over one’s dogma? I guess that my best way of atoning is to try to update my less than ideal bar charts with a dot plot. If that radical idea isn’t accepted, my fall-back recommendation would be to show the whole bars, even though the underlying data is about 90% locational and 10% length (or retinal position), and showing the whole bars will hurt the resolution somewhat.

I just write what I write, and I am always happy and surprised when a post generates this much discussion.

Don’t miss Derek’s comments above. He no longer tells people to include zero, he changes emphasis and tells them to include the whole bar. So for a floating column chart you need not include zero, just include the minimum and maximum extent of the bars.

I decided to bite the bullet and buy the relevant Cleveland article from JSTOR. Here is the most relevant quote (which does indeed support the “bar charts are different” line):

“The bar of a bar chart has two aspects that can be used to visually decode quantitative information–size (length and area) and the relative position of the end of the bar along the common scale. The changing sizes of the bars is an important and imposing visual factor; thus it is important that size encode something meaningful. The sizes of bars encode the magnitudes of deviations from the baseline. If the deviations have no important interpretation, the changing sizes are wasted energy and even have the potential to mislead (Schmid 1983)”.

The reference to Schmid is to a book called “Statistical Graphics”.

For me this brings a bit of closure: I have always been a fanatic for zero for bar charts too, but was thrown by Cleveland’s comments about not getting too hung up on zero. It is clear now that he only applies this comment to charts that encode solely by position (e.g. dot charts, x-y plots, line charts).

So Jon, thanks again for the post…the discussion here has led me to enlightenment!

But in order for the bar’s retinal properties to be an accurate depiction of the quantity, you have to show the whole bar. To hide part of the bar beneath the bottom of the chart is to mislead the eye. I’ve stopped trying to convince people to include zero on their bar charts; I just say “fine, leave the zero off, but you still have to show me the whole bar.” It’s just a coincidence that this forces them to include the zero :-)

Note that my “show the whole bar” rule not only short circuits the objections, but also allows the circumstances where you genuinely can have a bar chart with no zero on the scale: when it’s a floating bar chart. Then you only have to go down as far as the bottom of the floating bar, not all the way down to zero.

I still suspect that Cleveland’s earlier 1983 manuscript / 1984 paper “Graphical Methods for Data Presentation: Full Scale Breaks, Dot Charts, and Multibased Logging” would reveal what he really thought about bar charts that don’t start at zero. His other 1984 paper that you found does seem to suggest he’d concluded bar length distracted from the judgment of endpoint position relative to scale.

Yes, a general audience is more likely to be mislead by bar charts that don’t start from zero. A technical audience — and especially those who read Chart Busters — is more likely to notice the exception and perhaps wonder if the presenter is trying to mislead them visually or simply doesn’t know how to use Excel and/or chart selection that well.

The pattern of tick marks and powers of ten labeling seems a dead giveaway for a proper log axis. One usually doesn’t have to investigate the numbers (if any!) to notice that a log scale is not a simple linear scale starting from zero. Also, a log scale has clear redeeming social or scientific value, and I’m not sure you can make that case for a bar chart whose bars don’t start from zero.