That’s almost what it is. In the original discussion people were looking for a stronger way to remind the viewer about the scale changes in such a case.

I would say you’ve made a weaker way. It’s not at all obvious that the scale is discontinuous visually—you have to read the fine print.

If you are going to graph it this way, I greatly prefer larger breaks between the categories plotted on different scales, both horizontally and vertically, also different colors used for the dot plots at different scales would also serve to alert the viewer that something is going on. It’s extremely difficult to visualize from your graph, the difference at the boundary of the scales.

There appears to be a bigger difference between recreation and airports/airways, than between conservation and sewing, where the plots are right next to each other.

That would be somewhat alleviated by more obvious visual breaks, but a simple log scale would avoid the problem completely while conveying at least as much visual information as this graph.

I prefer the log scale. or, as your last comment suggests, a table with full dollar amounts (or amounts in thousands/millions) which would make the arithmetical differences clearer without losing information.

My impression:

At first glance, the graph seems to say that there were spending categories distributed more or less evenly across the whole range of spending levels, with roads only a modest outlier. Second glance, you notice the key up top and think it’s a straight line on a log scale. Naturally, you expect a horizontal increment under “hundreds of millions” to be ten times as much as one under “tens of millions”. But no, that’s not the case either. Or is it that the items in the same panel are comparable in size, with each white space representing an order-of-magnitude break? But no, the middle break isn’t really a break at all.

As far as I can tell, the visual impression given by the graph—the smooth slope of the points—just has no relationship to the underlying data at all. So what’s the advantage of this presentation over a simple table? or over Agnew’s double graph?

But this particular graph makes visual comparison impossible without looking up the actual numbers on the scale, and the breaks and their relative importance are only noticable if you study the graph really well. If you just printed the numbers, the larger number of digits would show the breaks at least as clear as this graph.

As such, i think I like Agnew’s solution more, also because it warns more clearly that it uses a “non-standard” way of plotting. If you really want to do it in your way, I would make at most one break and make the white-space between them a lot larger. Perhaps you can also introduce horizontal whitespace?

My own solution to these kinds of situations is to use the small map/large map trick from newspapers: I make two graphs, and draw (in an external drawing program) a thick rectangle around the part of the billions-graph that goes on to be the millions-graph.

Says the person who is writing in grammatically appropriate sentences, which obey generic rules for the communicative act of writing. Or in a different context, would you have said rriiion aymt pswiht audnlascoa and expected to have gotten your point across all the more efficiently? Being a human being, communicating to other human beings who share a common language, implies enough context to start making at least rudimentary judgments about efficiency.

That having been said, the panel discontinuity of this graph is an excellent tweak. What a log scale tends to mask, this clarifies pretty well. I still get the immediate impression of what spending dwarfs what other spending, and I can compare two values relatively easily.

that means that a first glance shows an almost flat rise when, in fact, there are different scales involved

…in general, for this kind of data, we’re not trying to determine whether there’s a linear or cubic or what-have-you relationship between points. We’re comparing quantities that in some sense are not determined by one another. People who aren’t scientists and don’t work frequently with models themselves won’t make the immediate association between an approximate-linear graph and the rather sophisticated notion of slope. You look at this graph and you see roads all by itself, in a block set aside all for itself: you get the idea that spending on roads dwarfs nearly everything else. It’s not as skull-crushingly obvious “at first glance” as a basic log scale, but Kieran’s striking a balance.

It’s striking how complex even quite simple data can be to present effectively.

I’d have preferred: 1, more room between each continuous block. 2, that billions be expressed as thousands of millions along the horizontal axis (the use of 1bn, 2bn, etc was a little ill-advised). Other than those minor aesthetic tweaks, this presentation seems very effective.

first, the scales have been chose so that the blue dots form a “line” across category boundaries. that means that a first glance shows an almost flat rise when, in fact, there are different scales involved (agnew’s plot shows the actual structure much more clearly). in other words, structure in the data has been purposefully(!) hidden by the careful choice of scale boundaries (an apparently straight line across different scales does smell – misleadingly – like a power law).

second (in a sense the same point repeated), there’s no single zero point for the x axis. there’s no way of seeing that one point is worth twice another, even within the same scale, until you go look at the numbers.

i find it particularly depressing because it imitates tufte while ignoring the basic principles behind his work. the point is not that something look simple, but that it actually is simple. agnew’s plot loses (slightly) on the “style” stakes, but more than makes up by being faithful to the original data. clarity of presentation must emerge from clarity of thought; arranging things in an ad-hoc manner to make them look simple is – well, you must know, in your heart of hearts, that this sucks deeply?

Yes, I was just in the process of writing up a comment to that effect—- I should have realized it myself, as I’d tried it already yesterday and that’s exactly what happened.

For this chart you could perhaps use geom_blank to add a dummy dataset and layer that pushed out the limits of the individual scales – that would ensure that you at least had consistent scales across the different orders of magnitude.

I’ll check it out.

p.s. it’s really nice to see ggplot2 in the “wild” :)

It’s a great package—thanks for writing it. An earlier effort in the same vein as this one is here. I’m much more used to the lattice framework, so it’s taken me a bit of time to get used to the ggplot approach (which is also one of the reasons I mess around with this sort of data—to learn the syntax). The new version of the book is very helpful, btw.

For this chart you could perhaps use geom_blank to add a dummy dataset and layer that pushed out the limits of the individual scales – that would ensure that you at least had consistent scales across the different orders of magnitude.

Maybe you want facet_grid(…, shape = “free”)?

The basic plot is an object with custom breaks which is then built up first with geom_point() and then facet_grid(. ~ Order, scale=”free”), where Order is the factor that distinguishes the panels (“Millions”, “Tens of Millions” etc). If I substitute shape=”free” instead, I get an error:

Error in get(“new”, env = FacetGrid, inherits = TRUE)(FacetGrid, …) :
unused argument(s) (shape = “free”)

The full code is:

llabs < -c(“1”,”2”,”3”,”4”,”30”,”40”,”50”,”60”,
“80”,”100”,”200”,”300”,”400”,”500”,”1bn”,”2bn”,”3bn”,”4bn”, “5bn”)

p < - ggplot(data, aes(Spending, reorder(Category,Spending))) +
scale_y_discrete(name=” “) + scale_x_continuous(name=”Millions of Nominal Dollars”, breaks=c(1,2,3,4,30,40,50,60,80,100,200,300,400,500,1000,2000,3000,4000,5000),labels=llabs) + geom_point(colour=”blue”,size=2.5)

p + facet_grid(. ~ Order, scalee=”free”) + opts(axis.text.y=theme_text(colour=”#000000”,hjust=1),
axis.text.x=theme_text(colour=”#000000”,vjust=1)) + opts(title=”Scale and Allocation of Total Expenditure from WPA and Sponsors’ Funds, through March 1943n”)

Without the custom breaks (which aren’t ideal) I found some ticks being repeated on adjacent panels. Is there a a way to control the width of separate panels, or would that entail constructing completely different objects one at a time?

(For the rest of you following along, Hadley is the author of the R package that does all the heavy lifting in this graph.)

Even after noticing the discontinuity at Sewing, I had almost finished this comment and was just looking at the chart again when I noticed that writing (just below $30 million, and arbitrarily at the bottom of the tens of millions category) received more than 5 times what Household workers training received (just over $4 million and arbitrarily at the top of the millions category), even though the two dots are very close together.

The idea of adding headers naming each order of magnitude is very nice, and would work well with a true log scale as well.