What makes a good graph?

I want to begin by talking about some general issues when presenting data. SPSS and other packages make it very easy to produce snazzy-looking graphs (see Section 4.9), and you may find yourself losing consciousness at the excitement of colouring your graph bright pink (really, it's amazing how excited my psychology students get at the prospect of bright pink graphs - personally I'm not a fan of pink). Much as pink graphs might send a twinge of delight down your spine, I want to urge you to remember why you're drawing the graph - it's not to make yourself (or others) purr with delight at the pinkness of your graph; it's to present information (dull, perhaps, but true).

Tufte (2001) wrote an excellent book about how data should be presented. He points out that graphs should do the following, among other things:

Show the data.

Induce the reader to think about the data being presented (rather than some other aspect of the graph, like how pink it is).

Avoid distorting the data.

Present many numbers with minimum ink.

Make large data sets (assuming you have one) coherent.

Encourage the reader to compare different pieces of data.

Reveal the underlying message of the data.

However, graphs often don't do these things (see Wainer, 1984, for some examples). Let's look at an example of a bad graph. When searching around for the worst example of a graph that I have ever seen, it turned out that I didn't need to look any further than myself - it's in the first edition of this book (Field, 2000). Overexcited by SPSS's ability to add pointless fluff to graphs (like 3-D effects, fill effects and so on - Tufte calls these chartjunk), I literally went into some weird orgasmic state and produced an absolute abomination (I'm surprised Tufte didn't kill himself just so he could turn in his grave at the sight of it). The only consolation was that because the book was published in black and white, it's not bloody pink! The graph is reproduced in Figure 4.2. What's wrong with this graph?

The bars have a 3-D effect: Never use 3-D plots for a graph plotting two variables because it obscures the data.¹ In particular, 3-D effects make it hard to see the values of the bars: in Figure 4.2, for example, the 3-D effect makes the error bars almost impossible to read.

Patterns: The bars also have patterns, which, although very pretty, distract the eye from what matters (namely the data). These are completely unnecessary.

Cylindrical bars: Were my data so sewage-like that I wanted to put them in silos? The cylinder effect muddies the data and distracts the eye from what is important.

Badly labelled y-axis: 'Number' of what? Delusions? Fish? Cabbage-eating sea lizards from the eighth dimension? Idiots who don't know how to draw graphs?

Now, take a look at the alternative version of this graph (Figure 4.3). Can you see what improvements have been made?

A 2-D plot: The completely unnecessary third dimension is gone, making it much easier to compare the values across therapies and thoughts/behaviours.

The y-axis has a more informative label: We now know that it was the number of obsessive thoughts or actions per day that was being measured.

Distractions: There are fewer distractions like patterns, cylindrical bars and the like.

¹If you do 3-D plots when you're plotting only two variables then a bearded statistician will come to your house, lock you in a room and make you write I 75,172 times on the blackboard. Really, they will.

Minimum ink: I've got rid of superfluous ink by getting rid of the axis lines and by using lines on the bars rather than grid lines to indicate values on the y-axis. Tufte would be pleased.