Nielsen is widely regarded as providing exception analysis of consumer data.  In fact, many of the best analysts I work with spent many years at Nielsen and they have been very interested in learning data visualization best practices because they now understand the benefits. 

I agree that Nielsen’s insights are often fantastic, yet I don’t understand why they can’t present their analysis more appropriately.  I guess my larger concern is that when a company with the influence on analysts like Nielsen has presents data visualization, even simple ones, so poorly, it make the abuse more and more pervasive.  I’m seriously considering send a link to this post to the author of the presentations I’m about to review.

I received two annual reports the other day, both authored by a high-level employee who was supplemented by many other resources, so I can’t lay the blame on any one person, but more on Nielsen as a whole.  Here is a summary of the charts they presented:

Presentation #1
Zero-based axis = 6
Missing or non-zero-based axis = 88
Pie chart with no color to differentiate the slices = 2
Charts well done = 6/96 = 6.3%

Presentation #2
Zero-based axis = 37
Missing or non-zero-based axis = 36
Smoothed line charts = 7
Charts well done = 37/80 = 46.3%

Hopefully this means someone told them presentation #1 had a lot of chart junk and they made an effort to improve presentation #2, but I doubt that’s actually the case.  It’s easy to see that most of the charts were created in Excel, which will automatically set the axis to start somewhere other than zero if the numbers in the chart are large.  I don’t know the specific business rules that Excel uses, but they should be changed.

Here is a representative example of the charts they created which did not have a zero-based axis. I’m holding out hope that this wasn’t done to intentionally deceive the reader, but to emphasize the subtle differences between the data points.

I recreated the chart as a dual-axis chart with the primary axis starting at zero and the secondary axis set to Excel’s default.  Also note that I created a line chart since this is time-based series data, which typically means you’re wanting to see the overall pattern. 

Clearly these imply a very different story.

 

Ok, we see what’s wrong, but how could Nielsen have presented the data more effectively? You have two primary options.

Dot plot

• You can replace bar charts with dot plots so that the sequence over time is de-emphasized
• Dot plots don’t require a zero-based scale
• Dot plots force the reader to refer to the scale before comparing two values
• Sizing and coloring the bubbles by change over prior year would speed up the reader’s analysis of variances between variables

Non-zero-based line chart with special alerts

This is the most effective method if you insist on NOT using a zero-based scale.  Stephen Few sums it up best in Show Me the Numbers (page 169):

You should generally avoid starting your graph with a value greater than zero, but when you need to provide a close look at small differences between large variables, it is appropriate to do so.  Make sure you alert your readers that the graph does not give an accurate visual representation of the values so that your readers can adjust their interpretation of the data accordingly.

Nielsen can learn a few lesson by reading some of the great data visualization books by Few and Tufte, or they could hire resources that know what they’re doing and allow those resources the freedom to make the best practices viral.  The alternative isn’t good for any of us.

My daughter had a soccer game last weekend across town early in the morning and the weather was predicted to be quite cold.  Naturally I went to weather.com to check the hourly forecast, but this time something struck me.

Notice the vertical scale.  It’s not zero-based.  Sure, it’s simply showing the changes in temperature, but as I scrolled through the pages, the axis values changed, that is, the range did not stay consistent.  I also noticed that 12am is repeated, that’s kind of odd.  Fusion Charts is their tool of choice.

I would have used Tableau to create a simpler chart.  Unfortunately I lose the nice pictures across the top of each hour, which I really like, and the gentle shading (though why use gold for night hours…doesn’t gold mean sunny?), but I gain a zero-based scale and a line that I can color based on temperature, with the mid-point at 32 degrees.  Below 32 = red, above 32 = green.

In this view the variances in the temperatures are even easier to see.  You can see the huge change from 6am to 3pm and then the dramatic drop as sunset approaches.  Which view works best for you?

I never knew such a chart existed, but alas I found one, and I hope it becomes extinct! Occasionally I scan through Many Eyes visualizations for ideas and/or blog inspiration. Let's review this simple line chart of Estimated Median Age at First Marriage. Click on the image below to get started.


When I first saw this I thought "Wow! What a huge variance over the years!" But then I looked a bit closer and saw that:

1. The years are backwards. A time-series line chart should nearly always start with the oldest time period on the left. I can't even think of a way to interpret time backwards. Maybe the DeLorean from Back to the Future could help.


2. The Y axis does not start at zero. This creates a misleading variance. It appears there has been a 700% variance from highest to lowest, but really it's only 35%.


3. The Y axis should be rounded to a whole number; this is unnecessary precision.


4. I find myself having to refer back to the legend to remind myself which sex is represented by which color. They are way too close in hue. Why not use blue for men and pink for women?


5. The Years on the X axis are at an angle and squished together. If you must show all of the years, the turn them a full 90 degrees. In the end though, I believe the purpose of the chart is to show a trend, so I don't need to see all of the years, just enough so that I know it's a regular interval.


6. One more thing. It's very subtle. This is NOT a regular interval after all. Between 1890 and 1940, there is only one measure per decade. Only beginning in 1947 is there data for every year. I would only display 1947-2003.

To address all of these problems, the chart could have been created like this.

I finished re-reading Show Me the Numbers by Stephen Few last week and in Chapter 10 there is a discussion about Scale Lines, a subsection of which is devoted to logarithmic scales. Stephen made one particular recommendation that caught my eye (p. 199):

"Whenever you want to compare differences in values as a ratio or percentage, logarithmic scales will do the job nicely. They are especially useful in time-series relationships if you wish to compare ratios of change across time."

On p. 200, Few goes on to say:

"When you use lines to encode time-series along a logarithmic scale, angles that are the same tell you that the rate of change is the same."

Personally, I try, whenever possible, to avoid using a logarithmic scale. They tend to be very difficult to interpret and, for me, it's much easier to understand a graph that shows % change on a zero-based axis. Let's look at an example of both starting with a bar chart.



On this bar chart, it's challenging, but you could make the assumption that both bars are changing at the same rate. This same chart is much easier to read as a line chart like the following:



On this line chart, it is much easier to see how both line have roughly the same angle of ascent leading you to believe that they have approximately equal rates of change.

However, if you look at the same line chart, but with a zero-based axis, you get a much different message:



You can now much more clearly see that the percentage change year to year for the two measures is very different. (NOTE: I have hidden 2001 from this chart because it is the reference point from which the measures begin.) In fact, the logarithmic scale would lead you to believe that there is a positive percent change, however, it is clear from this zero-based line chart that it's quite to opposite.

In the end, I urge caution when using logarithmic scale and my suggestion would be to use a zero-based scale to represent percentage change whenever you can.