The simple stacked % bar charts (with color coding) are scaled to cumulative %, which tends to be very useful if the categories are ordinal but worthless if the categories are nominal. You can’t read cumulative % directly from your favorite clustered column chart’s scale. Depends what you want to show.

….Now feel free to continue where you left off on the SPC discussion…. :^)

Colin

Looking back at those one-sided half-tornado charts, I can see how they would be confusing. I find that as I work on different alternatives, I become so familiar that I know what the chart means without thinking about it. Someone looking at it cold has a lot of work to do to come up to speed.

The tornado-like chart that balances on zero does seem to be a good approach. I saw that on someone’s web site the other day, I don’t recall whose. I hope they see this and speak up, so I can say thanks.

Nice response. I agree about the line chart. It shows the data, but it also shows nonexistent trends, which change drastically as the categories go from one arbitrary order to another.

If the data classes were related, like different categories of costs, the clustered columns would show if an employee group accounted for the same percentage of each type of cost. In this case, the relationships are false.

Re small multiples, I’ve been showing examples of panel charts for a few years now, since even before I heard Kelly O’Day call them “panel charts”. I think he originated the term, and I like it, so I’m sticking to it. Anyway, here’s a panel chart of this data, crafted in Excel:

I think my next big non-work project will be a utility that takes a data range, arranges it appropriately, and constructs a panel chart.

Thanks for the comment about the blog. I’m enjoying it a great deal. I’m afraid that posts like this use up more of my good material in the comments! Look at all these alternative charts I’ve put into the follow-ups.

Actually, my preferred method to represent the example data is using “small multiples”, where Hours, Cost, Profit and Sales are shown as separate bar charts. This approach addresses the following potential issues with the clustered column:

1) Eliminates the need for a legend as each chart has only one series (as a aside – when a legend is required on a column chart, it’s easier to decode the legend if it’s placed above the chart because the eyes don’t have to work as hard).

2) Since the measures are independent, they’re probably better off on separate charts anyway. When everything is on one chart, people have a tendency to compare things that they shouldn’t (the bar heights for Engr1 is an obvious example of this problem).

The clustered column alternative for the same data have similar issues, in addition to those that you mention.

It’s somewhat unfortunate that Excel only knows to plot everything on a single chart and provides no option to plot data on separate charts when you select a data range covering multiple series. However, one could adapt the VBA code in your excellent post “Creating Charts in a Grid” for thiis purpose. Hmmm…now that I think about it, that code could be generalized so that a dialog box is used to specify the grid parameters *and* the chart type.

By the way, this is a great blog with excellent discussion topics. Keep up the good work!

Colin

You asked about the Google charting API. It looks pretty simple to use, and offers some useful chart types (and like every other charting package, some junk). I’m planning to look at it when I get some free time, which at this rate will be a week from next summer.

Tony – If you need to know what the leaning percentages are, you’ll have to go back to the clustered style charts. What this style is good at is showing the strong feelings (No, Yes) and the general feelings (No + Leaning No, Yes + Leaning Yes). It’s not great, but it gives the impression.

I saw a version of this where the undecided was split across Y=0, with the negative responses added to the Y<0 side, and the positive responses to the Y>0 side.

The advantage of this style is that the more negative bars are pushed left and the more positive bars right, so you can judge relative approval by where each bar floats. The disadvantages are [1] the reduced resolution: the scale goes from -100 (full disapproval) to +100 (full approval); and [2] the same problem as with tornados: you’re trying to compare how far to the left one side goes with how far to the right the other side goes.

You can undo the tornado effect by plotting the positive and negative stacks side by side.

I wonder if it makes more sense to put the stronger feelings next to the axis.

Or change to a columnar orientation.

I agree with your favorite, which is the simple clustered column chart.