The tutorial in question is entitled Charts III: Create a professional-looking chart, and it is the third in a series of four tutorials. The others are Charts I: How to create a chart, Charts II: Choose the right chart type, and Charts IV: Charts for the scientist. I will probably review these other tutorials at some point, but this one is really a job for Chart Busters.

Before I begin, I’d like to thank Jorge Camoes, whose Spiffy Charts: Want to know how to create bad charts? Visit Microsoft Office Online Training blog post pointed me to this horrid tutorial.

I’d also like to acknowledge all of the bloggers and commenters who have bemoaned the fact that “making a professional chart” is really a nice sounding phrase that really means “use as much gratuitous formatting and effects to make your chart pretty, without regard for its effectiveness.”

Tutorial Example Chart 1

Microsoft kicks off their tutorial with a simple column chart that uses the defaults of Excel 2003.

The first thing the Microsoft tutorial suggests is removing the gridlines. This is perfectly reasonable as far as it goes.

But it really doesn’t go far enough. Let’s really clean up the chart (below left).

• Remove the border around the entire chart.
• Move the chart and axis titles close to the edges of the chart area, and stretch the plot area into the freed up space.
• Take away the plot area fill and border.
• Remove the tickmarks from the X axis.
• Move “First Quarter Sales” to below the title, so it’s more of a subtotal. If the X axis has dates, you probably don’t need an axis title.
• Remove the border around the legend, arrange the legend entries horizontally, and move the legend beneath the chart.

And the gridlines wouldn’t be bad if they were lighter. This allows the Y axis line to be removed (below right).

Next the Microsoft tutorial suggests putting all of the data into data labels in the chart. This renders the Y axis redundant, so it can be removed.

It may make sense to add data labels to some points, but putting a label on every point clutters up the chart. It makes one ask, why not just use a table, if the individual numbers themselves are so critical? I like my chart with labeled gridlines and no axis.

The third Microsoft enhancement to the chart is to start adding gratuitous formatting. Add a gradient background to the plot area. Add a different gradient fill to the data bars.

Well, if we add a gradient, we will no longer have a nice uncluttered chart background. So let’s ignore that advice. As far as the pseudo-columnar appearance of the columns, that’s distracting, because folks are going to spend more time thinking about how we accomplished that cool formatting. Se let’s skip that one, too. But there’s nothing wrong with using nicer colors. Here’s my updated chart with axis (below left) or with gridlines.

The fourth enhancement to this simple chart is to remove the border from the legend. Hmm, I think we’ve heard that somewhere before. Also, the tutorial suggests adding a shadow to the chart. However, that means adding back the border, which with the shadow would make the chart even more cluttered.

The best piece of advice in this step is (emphasis mine):

You could add shadows around the chart and axis titles as well, but make sure the chart doesn’t look cluttered.

Good idea.

Chart Axis Sin

The final enhancement Microsoft wants to apply to their chart, after changing the product name, is to change the axis scale. In some cases this is acceptable, even desirable. However, bar and column charts encode their values by their length. When you shorten the axis, you effectively cut the bottoms off of the bars. This makes the longer bars seem much longer in comparison to the shorter bars than they really are.

Why start the axis at 50? Why not at 100? This makes the increases from January to February all the more striking.

This is a a major chart fail. The value axis on a column or bar chart should always include zero. Always. If you want to expand the scale to help resolve the values, then a column chart is not the right chart type.

Update

Scott Ziolko commented below that a further improvement to these charts would be to rotate the Y axis title to make it easier to read. In the chart below left, the axis title has been rotated and moved above the axis, and the plot area was widened (alternatively the chart could have been made narrower with no loss in resolution). In the chart below right, I have incorporated the axis title into the chart subtitle, where it is just as effective.

Chart Type Selection

Speaking of the right chart type, it is generally more effective to show time-based data in a line chart, i.e., in a time series (below left). When the axis is rescaled (below right), there is no loss of meaning or distortion of data.

Tutorial Example Chart 2

The next chart is “a combination chart that needs help.” And it does need help, but I think the “after” chart needs as much help as the “before” chart.

My first thought is about which series was assigned to which chart type. To me it has always made more sense to use columns for the Projected or Target values, and show the Actual values with a line. Alternatively, an area series could show the Projected or Target values.

What also looks better is a chart without all those distracting borders. Removing the fill helps too. I found it a bit hard to look at the green and gray chart above: the clashing gradients made the bars hard to pick out.

I also found it unusual that the Actual values showed a steady progression from month to month, while the Projected jumped around. Maybe they did get the chart types correct, but the labels wrong. Of course, it is just data made up for the benefit of this illustration.

Tutorial Example Chart 3

The final chart in the tutorial is one of my favorite types (cough cough): a 3D column chart.

The tutorial correctly points out that the tall bars in the first chart below obscure the short bars. They spoil this observation by simply moving the tall bars behind the short ones. Does this make it any easier to compare their values?

The strange perspective eliminates any chance of understanding the relative values in this chart. Especially since the vertical axis tick labels in the second chart have been deleted.

The tutorial blindly follows the mantra: If 2D are good, 3D must be better. But for clearly displaying data, 2D rules. In both charts below, the 2D column chart and the line chart (hey, it’s a time series!), make comparisons easy.

Summary

I guess the takeaway here is that no matter how much effort is spent generating pretty tutorials, if the rules and best practices are not followed, the tutorial will not succeed. It illustrates how the producer of a software package may not be the best to advise users how to apply the software.

Keep your wits about you when seeking help, look for multiple sources of advice, and use good judgment when following that advice.

Peltier Technical Services, Inc., Copyright © 2009.
 
Licensed under a Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported License.
 

It’s eye-catching. It also is difficult to read, because of the use of irregularly-shaped polygonal areas to encode values. In fact, the graphic sticks several square pie charts together to make cubical square pies. I addressed square pie charts in Ineffective Chart – Partition Chart Revisited (Defragged), where I noted:

The defragged partition chart reminds me of the “square pie” we first read about in Eager Eyes (Women in IT – Squaring the Pie?), and which was taken up by Juice Analytics (Squaring the Pie), Dashboards by Example (The Square Pie Chart), Anil Dash (Pixels Are The New Pies), Info Clarity (A square pie), Bella Consults (Square pies taste bad), and others.

The square pie chart was all the rage for a while. Consensus seemed to be split between “they’re wicked cool” and they’re even less effective than round pie charts. My own opinion is that they are ineffective: the awkward shapes imposed by rows and columns of filled squares make comparison of areas even more difficult than comparing areas of rectangles with unequal lengths and widths or comparing areas of rotationally misaligned wedges in a pie chart.

As with so many ill-advised charts, a more effective replacement is a bar chart. Here is the Chart Busters version:

I suppose if you felt you had to, you could apply similar colors to those of the cubical square pie chart.

The colors may add to the chart’s ability to get attention, but they may actually detract from its ability to convey information.

Peltier Technical Services, Inc., Copyright © 2009.
 
Licensed under a Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported License.
 

Last week, Excel guru John Walkenbach conducted a poll on his wildly popular J-Walk Blog asking people to Post Your Pet. The poll results were tabulated in Tabulating The Pets.

I know John’s an expert and all, and he literally wrote the book on Excel Charts. Make that, the books. And these are good books, starting with the basics and progressing into some intermediate topics.

I know the books are good, because I helped out with technical editing. I made sure the protocols and the tricks all worked, and I checked out all the sample files. I even encouraged discussions of proper chart type selection.

Disclosure: these book covers link to Amazon.com, and if you buy one after following my links, I get a commission of a few percent.

Something-of-Pie Charts

So how did John decide to present the results?

Besides the usual yuck of a pie chart, the stacked bars in the bar-of-pie chart provide even more distortion. See the bar for Guinea Pig? If the labels weren’t there, I would have thought there were more guinea pigs than hamsters, rats, and other, combined. This is a problem with encoding values in areas, particularly when the scale of area to unit changes so drastically.

By default, Excel makes the bar 75% as tall as the pie’s diameter, or 37% of the pie’s area. The label on the “other” slice of the pie tells me that the bars comprise 5 units, or 5.4% of the total, but my visual cortex is overwhelmed by the hugeness of the bars.

How about a pie-of-pie chart?

After getting over your impression of a wagon with mismatched wheels, you are misled even worse by this graph. The default “other” pie has 75% the diameter and 56% the area of the main pie. It’s obvious that guinea pigs is as large a wedge as cats, right?

Can you say “cognitive dissonance”? Sure you can. As Professor Tufte has said, the only thing worse than a pie chart is more than one of them.

Improved Chart Selection

So how do you show all elements in a single chart with a single encoding scale? Class?

That’s right, our old boring friend, the bar chart. Look, guinea pigs are nowhere near as popular as cats after all.

What if you want to show the actual counts in the chart?

Easy. Data labels add the information, without being critical to your understanding of the chart.

This is the wrong chart type for a handful of reasons.

First, we all are aware of the evils of 3D charts. The first green bar in this chart has a data label proclaiming a value of 11, but it sure looks much closer to 12 than to 10. Let’s try to improve this with a 2D clustered column chart:

Okay, we can better judge the values of the bars without the 3D-induced parallax.

But what was the analyst trying to show? There was scant description in the post that accompanied the graph.

Is there any relationship between the two measures? A column chart will never show such a relationship.

Here is that same information on an XY chart. There’s a cluster of points in the middle and an outlier (Lidge) in the upper right.

There isn’t a strong dependence of Blown Saves on FIP, as we see from the blue symbols, regression line, and regression equation below. If we throw out the outlier, there’s no relationship at all, illustrated by the red features in the chart.

What was the analyst trying to show? Was he making some comment in the title of the chart?

“When you need them the least”: Does he think these are poor pitchers? What do you think?

It’s a striking graphic – The colors capture your attention, and you can decipher quickly what the chart is showing. A weakness is that the use of bold fill colors precludes direct comparisons of players on the same graph. You could use side-by-side diamonds, but that’s not as clear a comparison as overlaying the data.

What if we used an unfilled diamond for each player? Carlos looks no less impressive in this less colorful chart:

This type of chart is called a radar chart, because it has its origin in the center and the data sweeps around the periphery, like a radar screen. It is also called a spider chart, because with more spokes and with circumferential gridlines between the min and max, it looks like a spider’s web.

It is easy in this case to make comparisons between a couple players. Adam Dunn does not have two Power ratings, of course. His fielding is so bad, the corresponding rating is negative. Apparently something about his ratings means he’s worse than the worst fielder at 0%. Not unusual, I remember teammates like that in Babe Ruth ball.

Suppose we have a whole set of players to compare:

Conceivably you could compare many players in this diamond chart.

There are a couple of problems with this kind of chart, as discussed in Radar Charts are Ineffective. The resolution is worse at the lower numbers, and any zeros make interpretation more difficult. The negative value for Dunn’s fielding is difficult to pick out, and we see a similar calculation glitch with Ben Zobrist’s fielding score, which is greater than 100%. Maybe he could give Adam Dunn some lessons?

Where have we seen this style of chart before? Perhaps when we discussed Spider Chart Alternatives. Among the alternatives we’ve evaluated is the Parallel Coordinates chart, developed by Alfred Inselberg and described in Interactive Parallel Coordinates Chart and in Sample Parallel Coordinate Chart. Stephen Few discusses parallel coordinates charts in Multivariate Analysis Using Parallel Coordinates (html) and Multivariate Analysis Using Parallel Coordinates (pdf).

Justin has already done the hard work for us: the ratings are compiled and scaled to 0%-100% (with a couple computational irregularities). All we have to do is unwind the chart, that is, convert it to a line chart. Here are Carlos Pena’s ratings again.

Because the coordinates are parallel, we can compare the scores in different categories more easily than in the radar plot. We can readily compare a couple players, and any off-scale values can be represented by enlarging the scale.

We can add more players, and the resulting chart is arguably less cluttered than the spider chart equivalent.

That’s still not easy to read, if you want to compare two players. In Interactive Parallel Coordinates Chart we embellished the chart above in order to let the user select any two quarterbacks from a list. We’ll apply the same techniques here: Color all series light gray for background, Add two more colorful series, Incorporate a couple of dropdown controls, and link the pieces together.

In the following chart, we can compare Pena and Dunn to each other, but also see how they fall within the larger population.

To change our comparison, we need only select a different player, Albert Pujols for example, from a dropdown list.

And the chart shows our new comparison.

I’ve posted the zipped Excel 2003 workbook on my web site at Composite_Player_Evaluation.zip.

In this analysis, Justin has related attendance to the home team’s won-lost record. The season isn’t over yet, so the data covers games through the past weekend. Justin covers the National and American Leagues in two separate posts as part of his Graph of the Day series:

MLB 2009 Attendance Comparison I – National League – Graph of the Day
MLB 2009 Attendance Comparison II – American League – Graph of the Day

Justin’s analyses use some awkward charts and he introduces an awkward statistic.

Beyond the Box Score Charts

Here is Justin’s analysis of the American League’s Eastern Division, where my Sox play:

There are a total of six charts, since each league has three divisions. I’ve reduced them here:

The charts are dark, and the colors and gradients bold, so the charts weigh heavy on the eyes. There are only twelve data points per chart, six on each axis, and part of the analysis involves comparing the relative heights of the bars on primary and secondary axes. I’ve discussed the problems of comparing series on primary and secondary axes in Secondary Axes in Charts. Any conclusions you may reach are affected by the relative arbitrary scales on the primary and secondary axes.

It’s not easy to make comparisons among these charts, and you miss the opportunity to see how attendance and number of wins are correlated. In fact, you don’t even see attendance in these charts: Justin has introduced a statistic called Attendance Per Win, which he is comparing to the team’s wins.

Chart Busters Charts

Rather than plot the two variables of interest on primary and secondary columns, Chart Busters have made an XY chart (or as statisticians and Microsoft call it, a scatter chart). All three divisions of both leagues fit onto the chart, which shows wins on the X axis and attendance along the Y axis. The teams are denoted by the data labels, and we only have one case of overlapping, the red squares for the Minnesota Twins and the Chicago White Sox, in the middle of the chart.

There is an obvious positive correlation between wins and attendance: makes sense, because teams love to cheer for a winner.

Justin’s use of the derived Attendance Per Win statistic initially makes some sense. But when Chart Busters replace the Y axis of the XY chart with Attendance Per Win, we lose the clear correlation. The National League looks purely random, while the American League has at best a slight positive trend.

Justin’s intent was to show which teams gain more attendance than expected for the number of wins. Without inventing a new statistic, Chart Busters show this by simply drawing the lines of best fit. There are separate lines for each league, since there’s some difference in average attendance between leagues.

Linear regression supports our eyeballs: there is a positive correlation for each league. R² for the American League is 74%, and for the National League is 39%. Not a bad correlation, since it ignores such factors as weather, stadium size, and scheduling variations. The points Justin makes are visible in this chart: fans in Florida, Oakland, Texas, and Tampa do not support their teams, as shown by the distance of these points below the fitted lines. In fact, Florida has abysmal fan support.

The Mets, in contrast, have the strongest fan support. This in conjunction with their sub-mediocre win total shows that my previous statement, that fans love to see a winner, is not strictly true. The Brewers, Cubs, Phillies, and Dodgers also have stronger than expected fan support, but then, the Dodgers and Phillies are leading their respective divisions.

Let’s see how Justin’s Attendance Per Win stands up to best fits.

The American League shows a positive correlation, not very steep, and the National League has only a very slight positive slope. The R² values are 23% for the AL and 2% for the NL. We see the same teams with strong and weak fan support as in the previous chart, without clouding the analysis with an unnecessary derived statistic.

Yeah, NPR is great and all, I even listen occasionally. But this is a Chart Busters column, so I’m going to talk about the visualizations that accompanied the FastCompany article. Specifically I want to draw attention to the following two graphs.

Look at the first graph: it’s a pie chart, in 3D, not only exploded, but with the pieces rearranged. At least they’re in order, but how are we supposed to interpret it?

You could put the pie back together, and take a bird’s eye view.

Yeah, not much better, but at least we’re not distracted by the thickness. The pie isn’t really much different than a set of labels. Here is the misaligned non-tabular table that results when we hide the pie pieces.

Time to revert to an old standby. This bar chart clearly shows the relationships between the five revenue sources, and if we look very carefully at the Y axis labels, we see that they are formatted as percentages, so it’s no huge jump to realize they all add to 100%.

The revenue data isn’t all that fascinating, but the audience statistics tell an interesting tale.

The second FastCompany graph is another 3D wonder, a bar chart rendered in isometric view. The angles make the bar comparisons difficult, and also make the labels less readable than if they’d been printed straight. And in fact, the use of huge bars for percent change in audience takes away any analysis of audience size itself, which is an interesting variable in its own right.

Let’s compare the audience sizes. In this line chart we can tell at a glance that NPR as increased greatly, while TV and newspapers have gone down.

This stacked bar doesn’t show the relative audience sizes as clearly, but it shows that the overall audience has declined a bit, and we can still tell that NPR’s audience has nearly doubled.

We can compare share of the market, that is, share of the total audience. Again, in this line chart, the trend over the past decade is clear. NPR’s share has doubled at the expense of newspapers and especially television.

A stacked bar chart of audience share shows much the same trend.

There are obviously more than three sources for news. Newspapers and television stations have web sites, and there are numerous blogs reporting news. I would think overall that the audience is much greater than the approximate 100 million shown in my first line chart.

The decline in television news is not surprising, and I’m sure the television outlets are not too worried. They have plenty of so-called reality programming now, which is cheap to produce and generates huge advertising revenues.

From all of the gloom and doom stories, I would hve expected to see a much greater decline in newspaper audience. I no longer get a daily paper, and in many cities, the number of major papers has declined. If we re-examine these numbers in five years, we may see another 11% drop in newspaper audiences.

I was asked whether a certain type of chart could be done. The data was weekly sales of 14 products, but the user wanted to show only the top ten products each week, in a clustered column chart, with the bars within each cluster arranged in descending order. Here is a representative three-week period of data:

Well, long story short, this kind of chart is certainly possible. All it takes is a couple sets of formulas, one set to rank the weekly figures, the next to determine which product goes into which position for each week. The chart is reproduced below. No matter how much I’d like to show off how clever I am when it comes to writing formulas that define an intricate chart source data range, I won’t show how I did this. It’s really an ineffective way to display this kind of data, and I don’t want to encourage anyone by publishing the recipe.

What makes this chart so ineffective? There are too many colors (as Wayne or Garth once said, “That looks like a plate of hurl!”), and the colors are critical to understanding the information. The order of the colors (and the items they represent) changes from one cluster to the next. Because of the excessive colors and their variable positions, the user spends so much time bouncing from one cluster, to the legend, to the next cluster, back to the first cluster, then to the legend, then back to the cluster, then to the legend, ad nauseum. It’s very hard to see where a product falls in on any given week, and tracking trends from week to week is nigh impossible.

As we’ve discussed, the best way to show data over time is with line charts. Showing more than a few differently formatted lines will confuse the data, but I’ve shown how to work this out in other articles on this blog. If we treat the weekly line chart like a parallel coordinates chart, and provide a means for the user to selectively highlight a small subset of the data in the chart, it will speed understanding. For this exercise I’ve decided that three series will be highlighted.

Construction of this chart entails plotting of weekly data for all products, and formatting this data in the same dull color. This gives the user an overview of the data. Then a few more ranges are populated. Three columns are used for the weekly data of three highlighted series, then this data is plotted and given distinctive formatting. Three cells are used to identify each highlighted series so this weekly data can be easily extracted from the full data range using Excel’s INDEX function. Finally, the list of products is placed in a column so we can populate dropdown controls to enable the user to make his selections.

The resulting chart has all the data in a dull blue, with three highlighted series corresponding to the series which were best sellers in each of the weeks.

Can you even follow these three products in the column chart above? Probably not, and certainly nowhere near as easily as this line chart makes it.

I’ve described the dropdown-based chart highlighting techniques in Sample Parallel Coordinate Chart, Easier Interactive Multiple Line Chart, and Interactive Parallel Coordinates Chart, so I will not go over the construction of this chart. I have posted my workbook in a handy zip file, for any interested readers to peruse.

So what’s wrong with the chart?

Seth says, “Now I’m not a big believer in 3-D charting that simply adds depths, supposedly creating visual appeal, to a 2-D chart.” As you all know, neither are we here at Chart Busters, Inc. And of course, adding 3D effects to a pie chart is adding insult to injury.

There’s an interesting negative value, Loans, in the table below the chart. It may or may not have been added to this chart, but since it comprises less than 0.01% of the total, it isn’t visible, and they mercifully did not label it. We have seen in Pie for Dessert Again? that Excel allows negative numbers in pie charts, and merely plots their absolute values. Makes perfect sense.

Seth also asks us to ignore the labels for a moment, and judge which area is largest: the purple wedge in the front right of this ugly ovoid, the brick red wedge at the left. or the combined orange and blue wedge in the top right? To my untrained eye, they look about the same. But the labels show that these areas represent 38%, 47% and, 14% of the total. As Seth points out, this is not just caused by the magical distortion of 3D rendering, but also because the numbers don’t agree with the graphics.

We’ve also seen in Pie Chart Plotting Deficiency that even Microsoft doesn’t always care that numbers in pie charts add up. But the Microsoft chart merely left off the “Other” category, while the USAspending chart is just way off.

Seth made a 2D chart to show a more accurate depiction of the data. Note that he included the Loans wedge, but it’s not even as thick as the leader line attached to the data label.

Then Seth tries to reproduce the 3D appearance of the original chart, and he almost succeeds, except that his numbers are still faithful to the inputs.

Seth has demonstrated without doubt that the original pie was doctored. He suspects that at some point after the chart was constructed, the site design was turned over to a graphic artist who “improved” the display. Seth hasn’t demonstrated that 2D or 3D pies are a good way to illustrate this information. Okay, chart busters, what’s the best way to show this kind of data?

A boring. old. bar. chart.

I’ve made two versions, one with the minuscule Loans category:

and one without:

Seth’s planning a review of the USAspending.gov dashboard. I’m looking forward to that. I spent a few minutes on the dashboard, and found it awkward, with labels scrolling out of view before you could read them.

Polling and Election Results on FiveThirtyEight

I’ve stated in this blog that the FiveThirtyEight web site uses pie charts in a very effective way. It helps that there are only two or three segments in their pies, and the third, if present, is very small. Here is a partial view of the FiveThirtyEight analysis of the November 2008 US elections.

This isn’t a full-blown Chart Busters article, but I thought I’d discuss some alternatives.

In an unfinished post from last fall, I undertook to replace FiveThirtyEight’s pies with bars, using roughly the same amount of space. The bars may have slightly better resolution, but I didn’t include a third bar for the third tiny segment. Take your pick.

Survey Results

FiveThirtyEight used pie charts to illustrate the survey results. As every second grade student will tell you, pies are the chart of choice when showing parts of a whole. The educational system has trained our children to behave like sheep, and follow the herd.

Editorializing aside, these charts of the poll results are what motivated this post.

Is Mark Sanford more or less ethical than other politicians?

FiveThirtyEight didn’t even change the Excel defaults for their pie. Black chart border, ugly fill colors, and way more space everywhere than needed, especially on either side of the chart. It’s nice that the data points  are labeled, but they are only partially labeled with the percentages, and the reader still must journey back and forth to the legend to see which segment means what. I also think it’s useful to put the results of a poll not in numerical order (55%-18%-18%-9%) but in some value order of the categories (more-same-less).

I have redesigned the pie chart, shrinking the white space and using arguably better colors. I also reordered the points, and centered the chart differently to try to make comparisons easier.

All well and good, but without the percentages, it’s still a guess whether More or Less had a greater response.

So I pulled a bar chart out of my magic hat. We can see that most of the cynical poll responders think Sanford’s ethics are no different than the other politicians, and equal numbers thought he was more ethical and less ethical. 9% have been busy watching Real Housewives and don’t know who Mark Sanford is.

Should Mark Sanford Resign?

FiveThirtyEight used another pie to show whether respondents want Sanford to resign. The colors have been changed from the gruesome defaults, but the large wasted spaces remain, as does the legend and border. Also, FiveThirtyEight didn’t use its usual style where competing results A and B meet at the vertical line at 12 o’clock.

I’ve addressed these stylistic issues in my own version of the pie chart, and it doesn’t totally suck.

Here’s the obligatory bar chart. Apparently 15% of the respondents think he should stay at the junkyard with his son.