Junk Charts

Elegant way to present a pair of charts

The Bloomberg team has come up with a few goodies lately. I was captivated by the following graphic about the ebb and flow of U.S. presidential candidates across recent campaigns. Link to the full presentation here.

The highlight is at the bottom of the page. This is an excerpt of the chart:

From top to bottom are the sequential presidential races. The far right vertical axis is the finish line. Going right to left is the time before the finish line. In 2008, for example, there are candidates who entered the race much earlier than typical.

This chart presents an aggregate view of the data. We get a sense of when most of the candidates enter the race, when most of them are knocked out, and also a glimpse of outliers. The general pattern across multiple elections is also clear. The design is a stacked area chart with the baseline in the middle, rather than the bottom, of the chart.

Sure, the chart can disappoint those readers who want details and precise numbers. It's not immediately apparent how many candidates were in the race at the height of 2008, nor who the candidates were.

The designer added a nice touch. By clicking on any of the stacks, it transforms into a bar chart, showing the extent of each candidate's participation in the race.

I wish this was a way to collapse the bar chart back to the stack. You can reload the page to start afresh.

***

This elegant design touch makes the user experience playful. It's also an elegant way to present what is essentially a panel of plots. Imagine the more traditional presentation of placing the stack and the bar chart side by side.

This design does not escape the trade-off between entertainment value and data integrity. Looking at the 2004 campaign, one should expect to see the blue stack halve in size around day 100 when Kerry became the last man standing. That moment is not marked in the stack. The stack can be interpreted as a smoothed version of the count of active candidates.

I suppose some may complain the stack misrepresents the data somewhat. I find it an attractive way of presenting the big-picture message to an audience that mostly spend less than a minute looking at the graphic.

Seeking simplicity in complex data: Bloomberg's dataviz on UK gender pay gap

Bloomberg featured a thought-provoking dataviz that illustrates the pay gap by gender in the U.K. The dataset underlying this effort is complex, and the designers did a good job simplifying the data for ease of comprehension.

U.K. companies are required to submit data on salaries and bonuses by gender, and by pay quartiles. The dataset is incomplete, since some companies are slow to report, and the analyst decided not to merge companies that changed names.

Companies are classified into industry groups. Readers who read Chapter 3 of Numbers Rule Your World (link) should ask whether these group differences are meaningful by themselves, without controlling for seniority, job titles, etc. The chapter features one method used by the educational testing industry to take a more nuanced analysis of group differences.

***

The Bloomberg visualization has two sections. In the top section, each company is represented by the percent difference between average female pay and average male pay. Then the companies within a given industry is shown in a histogram. The histograms provide a view of the disparity between companies within a given industry. The black line represents the relative proportion of companies in a given industry that have no gender pay gap but it’s the weight of the histogram on either side of the black line that carries the graphic’s message.

This is the histogram for arts, entertainment and recreation.

The spread within this industry is very wide, especially on the left side of the black line. A large proportion of these companies pay women less on average than men, and how much less is highly variable. There is one extreme positive value: Chelsea FC Foundation that pays the average female about 40% more than the average male.

This is the histogram for the public sector.

It is a much tighter distribution, meaning that the pay gaps vary less from organization to organization (this statement ignores the possibility that there are outliers not visible on this graphic). Again, the vast majority of entities in this sector pay women less than men on average.

***

The second part of the visualization look at the quartile data. The employees of each company are divided into four equal-sized groups, based on their wages. Think of these groups as the Top 25% Earners, the Second 25%, etc. Within each group, the analyst looks at the proportion of women. If gender is independent of pay, then we should expect the proportions of women to be about the same for all four quartiles. (This analysis considers gender to be the only explainer for pay gaps. This is a problem I've called xyopia, that frames a complex multivariate issue as a bivariate problem involving one outcome and one explanatory variable. Chapter 3 of Numbers Rule Your World (link) discusses how statisticians approach this issue.)

On the right is the chart for the public sector. This is a pie chart used as a container. Every pie has four equal-sized slices representing the four quartiles of pay.

The female proportion is encoded in both the size and color of the pie slices. The size encoding is more precise while the color encoding has only 4 levels so it provides a “binned” summary view of the same data.

For the public sector, the lighter-colored slice shows the top 25% earners, and its light color means the proportion of women in the top 25% earners group is between 30 and 50 percent. As we move clockwise around the pie, the slices represent the 2nd, 3rd and bottom 25% earners, and women form 50 to 70 percent of each of those three quartiles.

To read this chart properly, the reader must first do one calculation. Women represent about 60% of the top 25% earners in the public sector. Is that good or bad? This depends on the overall representation of women in the public sector. If the sector employs 75 percent women overall, then the 60 percent does not look good but if it employs 40 percent women, then the same value of 60% tells us that the female employees are disproportionately found in the top 25% earners.

That means the reader must compare each value in the pie chart against the overall proportion of women, which is learned from the average of the four quartiles.

***

In the chart below, I make this relative comparison explicit. The overall proportion of women in each industry is shown using an open dot. Then the graphic displays two bars, one for the Top 25% earners, and one for the Bottom 25% earners. The bars show the gap between those quartiles and the overall female proportion. For the top earners, the size of the red bars shows the degree of under-representation of women while for the bottom earners, the size of the gray bars shows the degree of over-representation of women.

The net sum of the bar lengths is a plausible measure of gender inequality.

The industries are sorted from the ones employing fewer women (at the top) to the ones employing the most women (at the bottom). An alternative is to sort by total bar lengths. In the original Bloomberg chart - the small multiples of pie charts, the industries are sorted by the proportion of women in the bottom 25% pay quartile, from smallest to largest.

In making this dataviz, I elected to ignore the middle 50%. This is not a problem since any quartile above the average must be compensated by a different quartile below the average.

***

The challenge of complex datasets is discovering simple ways to convey the underlying message. This usually requires quite a bit of upfront analytics, data transformation, and lots of sketching.

 

 

Watching a valiant effort to rescue the pie chart

Today we return to the basics. In a twitter exchange with Dean E., I found the following pie chart in an Atlantic article about who's buying San Francisco real estate:

The pie chart is great at one thing, showing how workers in the software industry accounted for half of the real estate purchases. (Dean and I both want to see more details of the analysis as we have many questions about the underlying data. In this post, I ignore these questions.)

After that, if we want to learn anything else from the pie chart, we have to read the data labels. This calls for one of my key recommendations: make your charts sufficient. The principle of self-sufficiency is that the visual elements of the data graphic should by themselves say something about the data. The test of self-sufficiency is executed by removing the data printed on the chart so that one can assess how much work the visual elements are performing. If the visual elements require data labels to work, then the data graphic is effectively a lookup table.

This is the same pie chart, minus the data:

Almost all pie charts with a large number of slices are packed with data labels. Think of the labeling as a corrective action to fix the shortcoming of the form.

Here is a bar chart showing the same data:

***

Let's look at all the efforts made to overcome the lack of self-sufficiency.

Here is a zoom-in on the left side of the chart:

Data labels are necessary to help readers perceive the sizes of the slices. But as the slices are getting smaller, the labels are getting too dense, so the guiding lines are being stretched.

Eventually, the designer gave up on labeling every slice. You can see that some slices are missing labels:

The designer also had to give up on sequencing the slices by the data. For example, hardware with a value of 2.4% should be placed between Education and Law. It is shifted to the top left side to make the labeling easier.

Fitting all the data labels to the slices becomes the singular task at hand.

 

Form and function: when academia takes on weed

I have a longer article on the sister blog about the research design of a study claiming 420 "cannabis" Day caused more road accident fatalities (link). The blog also has a discussion of the graphics used to present the analysis, which I'm excerpting here for dataviz fans.

The original chart looks like this:

The question being asked is whether April 20 is a special day when viewed against the backdrop of every day of the year. The answer is pretty clear. From this chart, the reader can see:

• that April 20 is part of the background "noise". It's not standing out from the pack;
• that there are other days like July 4, Labor Day, Christmas, etc. that stand out more than April 20

It doesn't even matter what the vertical axis is measuring. The visual elements did their job. 

***

If you look closely, you can even assess the "magnitude" of the evidence, not just the "direction." While April 20 isn't special, it nonetheless is somewhat noteworthy. The vertical line associated with April 20 sits on the positive side of the range of possibilities, and appears to sit above most other days.

The chart form shown above is better at conveying the direction of the evidence than its strength. If the strength of the evidence is required, we use a different chart form.

I produced the following histogram, using the same data:

The histogram is produced by first locating the midpoints# of the vertical lines into buckets, and then counting the number of days that fall into each bucket.  (# Strictly speaking, I use the point estimates.)

The midpoints# are estimates of the fatal crash ratio, which is defined as the excess crash fatalities reported on the "analysis day" relative to the "reference days," which are situated one week before and one week after the analysis day. So April 20 is compared to April 13 and 27. Therefore, a ratio of 1 indicates no excess fatalities on the analysis day. And the further the ratio is above 1, the more special is the analysis day. 

If we were to pick a random day from the histogram above, we will likely land somewhere in the middle, which is to say, a day of the year in which no excess car crashes fatalities could be confirmed in the data.

As shown above, the ratio for April 20 (about 1.12)  is located on the right tail, and at roughly the 94th percentile, meaning that there were 6 percent of analysis days in which the ratios would have been more extreme. 

This is in line with our reading above, that April 20 is noteworthy but not extraordinary.

 

P.S. [4/27/2019] Replaced the first chart with a newer version from Harper's site. The newer version contains the point estimates inside the vertical lines, which are used to generate the histogram.

 

 

 

 

 

How to describe really small chances

Reader Aleksander B. sent me to the following chart in the Daily Mail, with the note that "the usage of area/bubble chart in combination with bar alignment is not very useful." (link)

One can't argue with that statement. This chart fails the self-sufficiency test: anyone reading the chart is reading the data printed on the right column, and does not gain anything from the visual elements (thus, the visual representation is not self-sufficient). As a quick check, the size of the risk for "motorcycle" should be about 30 times larger than that of "car"; the size of the risk for "car" should be 100 times larger than that of "airplane". The risk of riding motorcycles then is roughly 3,000 times that of flying in an airplane. 

The chart does not appear to be sized properly as a bubble chart:

You'll notice that the visible proportion of the "car" bubble is much larger than that of the "motorcycle" bubble, which is one part of the problem.

Nor is it sized as a bar chart:

As a bar chart, both the widths and the heights of the bars vary; and the last row presents a further challenge as the bubble for the airplane does not touch the baseline.

***

Besides the Visual, the Data issues are also quite hard. This is how Aleksander describes it: "as a reader I don't want to calculate all my travel distances and then do more math to compare different ways of traveling."

The reader wants to make smarter decisions about travel based on the data provided here. Aleksandr proposes one such problem:

In terms of probability it is also easier to understand: "I am sitting in my car in strong traffic. At the end in 1 hour I will make only 10 miles so what's the probability that I will die? Is it higher or lower than 1 hour in Amtrak train?"

The underlying choice is between driving and taking Amtrak for a particular trip. This comparison is relevant because those two modes of transport are substitutes for this trip. 

One Data issue with the chart is that riding a motorcycle and flying in a plane are rarely substitutes. 

***

A way out is to do the math on behalf of your reader. The metric of deaths per 1 billion passenger-miles is not intuitive for a casual reader. A more relevant question is what's the chance of dying from the time I spend per year of driving (or riding a plane). Because the chance will be very tiny, it is easier to express the risk as the number of years of travel before I expect to see one death.

Let's assume someone drives 300 days per year, and 100 miles per day so that each year, this driver contributes 30,000 passenger-miles to the U.S. total (which is 3.2 trillion). We convert 7.3 deaths per 1 billion passenger-miles to 1 death per 137 million passenger-miles. Since this driver does 30K per year, it will take (137 million / 30K) = about 4,500 years to see one death on average. This calculation assumes that the driver drives alone. It's straightforward to adjust the estimate if the average occupancy is higher than 1. 

Now, let's consider someone who flies once a month (one outbound trip plus one return trip). We assume that each plane takes on average 100 passengers (including our protagonist), and each trip covers on average 1,000 miles. Then each of these flights contributes 100,000 passenger-miles. In a year, the 24 trips contribute 2.4 million passenger-miles. The risk of flying is listed at 0.07 deaths per 1 billion, which we convert to 1 death per 14 billion passenger-miles. On this flight schedule, it will take (14 billion / 2.4 million) = almost 6,000 years to see one death on average.

For the average person on those travel schedules, there is nothing to worry about. 

***

Comparing driving and flying is only valid for those trips in which you have a choice. So a proper comparison requires breaking down the average risks into components (e.g. focusing on shorter trips). 

The above calculation also suggests that the risk is not evenly spread out throughout the population, despite the use of an overall average. A trucker who is on the road every work day is clearly subject to higher risk than an occasional driver who makes a few trips on rental cars each year.

There is a further important point to note about flight risk, due to MIT professor Arnold Barnett. He has long criticized the use of deaths per billion passenger-miles as a risk metric for flights. (In Chapter 5 of Numbers Rule Your World (link), I explain some of Arnie's research on flight risk.) The problem is that almost all fatal crashes involving planes happen soon after take-off or not long before landing. 

 

The unreasonable effect of chart labels

In discussing the bar-density and pie-density charts with a buddy (thanks LB!), it became obvious that the labeling is a challenge. And he's right.

Here is the pie-density chart for the Youtube views with the labels as originally conceived.

These labels are trying too hard to provide precise data to the reader.

Here are some simplified labels that get at the message rather than the data:

Here is a slightly different version:

 

 

 

The Economist on the Economist: must read now

A visual data journalist at the Economist takes a critical eye on charts published by the Economist (link). This is a must read! (Hat tip: Fernando)

Here are some of my commentary on past Economist charts.

Bar-density and pie-density plots for showing relative proportions

In my last post, I described a bar-density chart to show paired data of proportions with an 80/20-type rule. The following example illustrates that a small proportion of Youtubers generate a large proportion of views.

Other examples of this type of data include:

• the top 10% of families own 75% of U.S. household wealth (link)
• the top 1% of artists earn 77% of recorded music income (link)
• Five percent of AT&T customers consume 46% of the bandwidth (link)

In all these examples, the message of the data is the importance of a small number of people (top earners, superstars, bandwidth hogs). A good visual should call out this message.

The bar-density plot consists of two components:

• the bar chart which shows the distribution of the data (views, wealth, income, bandwidth) among segments of people;
• The embedded Voronoi diagram within each bar that encodes the relative importance of each people segment, as measured by the (inverse) density of the population among these segments - a people segment is more important if each individual accounts for more of the data, or in other words, the density of people within the group is lower.

The bar chart can adopt a more conventional horizontal layout.

Voronoi tessellation

To understand the Voronoi diagram, think of a fixed number (say, 100) of randomly placed points inside a bar. Then, for any point inside the bar area, it has a nearest neighbor among those 100 fixed points. Assign every point on the surface to its nearest neighbor. From this, one can draw a boundary around each of the 100 points to include all its nearest neighbors. The resulting tessellation is the Voronoi diagram. (The following illustration comes from this AMS column.)

 

The density of points in the respective bars encodes the relative proportions of people within those groups. For my example, I placed 6 points in the red bar, 666 points in the yellow bar, and ~2000 points in the gray bar, which precisely represents the relative proportions of creators in the three segments.

Density is represented statistically

Notice that the density is represented statistically, not empirically. According to the annotation on the original chart, the red bar represents 14,000 super-creators. Correspondingly, there are 4.5 million creators in the gray bar. Any attempt to plot those as individual pieces will result in a much less impactful graphic. If the representation is interpreted statistically, as relative densities within each people segment, the message of relative importance of the units within each group is appropriately conveyed.

A more sophisticated way of deciding how many points to place in the red bar is to be developed. Here, I just used the convenient number of 6.

The color shades are randomly applied to the tessellation pieces, and used to facilitate reading of densities.

***

In this section, I provide R code for those who want to explore this some more. This is code used for prototyping, and you're welcome to improve them. The general strategy is as follows:

• Set the rectangular area (bar) in which the Voronoi diagram is to be embedded. The length of the bar is set to the proportion of views, appropriately scaled. The code utilizes the dirichlet function within the spatstat package to generate the fixed points; this requires setting up the owin parameter to represent a rectangle.
• Set the number of points (n) to be embedded in the bar, determined by the relative proportion of creators, appropriately scaled. Generate a data frame containing the x-y coordinates of n randomly placed points, within the rectangle defined above.
• Use the ppp function to generate the Voronoi data
• Set up a colormap for plotting the Voronoi diagram
• Plot the Voronoi diagram; assign shades at random to the pieces (in a production code, these random numbers should be set as marks in the ppp but it's easier to play around with the shades if placed here)

The code generates separate charts for each bar segment. A post-processing step is currently required to align the bars to attain equal height. I haven't figured out whether the multiplot option helps here.

library(spatstat)

# enter the scaled proportions of creators and views
# the Youtube example has three creator segments

# number of randomly generated points should be proportional to proportion of creators. Multiply nc by a scaling factor if desired

nc = c(3, 33, 965)*2

# bar widths should be proportional to proportion of views
# total width should be set based on the width of your page

wide = c(378, 276, 346)/2

# set bar height, to attain a particular aspect ratio
bar_h = 50

# define function to generate points
# defines rectangular window

makepoints = function (n, wide, height) {
    df <- data.frame(x = runif(n,0,wide),y = runif(n,0,height))
    W <- owin( c(0, wide), c(0,height) ) # rectangular window
    pp1 <- as.ppp( df, W )
    y <- dirichlet(pp1)
    # y$marks <- sample(0:wide, n, replace=T) # marks are for colors
    return (y)
}

y_red = makepoints(nc[1], wide[1], bar_h) # height of each bar fixed
y_yel = makepoints(nc[2], wide[2], bar_h)
y_gry = makepoints(nc[3], wide[3], bar_h)

# setting colors (4 shades per bar, one color per bar)

cr_red = colourmap(c("lightsalmon","lightsalmon2", "lightsalmon4", "brown"), breaks=round(seq(0, wide[1],length.out=5)))

cr_yel = colourmap(c("burlywood1", "burlywood2", "burlywood3", "burlywood4"), breaks=round(seq(0, wide[2],length.out=5)))

cr_gry = colourmap(c("gray80", "gray60", "gray40", "gray20"), breaks=round(seq(0, wide[3],length.out=5)))

# plotting

par(mar=c(0,0,0,0))

# add png to save image to png

# remove values= if colors set in ppp

plot.tess(y_red, main="", border="pink3", do.col=T, values = sample(0:wide[1], nc[1], replace=T), col=cr_red, xlim=c(0, wide[1]), ylim=c(0,bar_h), ribbon=F)

plot.tess(y_yel, main="", border="darkgoldenrod4", do.col=T, values=sample(0:wide[2], nc[2], replace=T), col=cr_yel, xlim=c(0, wide[2]), ylim=c(0,bar_h), ribbon=F)

plot.tess(y_gry, main="", border="darkgray", do.col=T, values=sample(0:wide[3], nc[3], replace=T), col=cr_gry, xlim=c(0, wide[3]), ylim=c(0,bar_h), ribbon=F)

# because of random points, the tessellation looks different each time
# post-processing: make each bar the same height when aligned side by side

***

A cousin of the bar-density plot is the pie-density plot. Since I'm using only three creator segments, which each account for about 30-40% of the total views, it is natural to use a pie chart. In this case, we embed the Voronoi diagrams into the pie sectors.

If the distribution were more even, that is to say, the creators are more or less equally important, the pie-density plot looks like this:

***

Something that is more like 80/20

The original chart shows the top 0.3 percent generating almost 40 percent of the views. A more typical insight is top X percent generates 80 percent of the data. For the YouTube data, X is 11 percent. What does the pie-density chart look like if  top 11 percent <-> 80 percent, middle 33 percent <-> 11 percent, bottom 56 percent <-> 8 percent?

Roughly speaking, the second segment includes 3 times the people as the largest, and the third has 5 times as the largest.

 

P.S.

1) Check out my first Linkedin "article" on this topic. 

2) The first post on bar-density charts is here.

 

 

 

 

 

 

 

 

 

Visualizing the 80/20 rule, with the bar-density plot

Through Twitter, Danny H. submitted the following chart that shows a tiny 0.3 percent of Youtube creators generate almost 40 percent of all viewing on the platform. He asks for ideas about how to present lop-sided data that follow the "80/20" rule.

In the classic 80/20 rule, 20 percent of the units account for 80 percent of the data. The percentages vary, so long as the first number is small relative to the second. In the Youtube example, 0.3 percent is compared to 40 percent. The underlying reason for such lop-sidedness is the differential importance of the units. The top units are much more important than the bottom units, as measured by their contribution to the data.

I sense a bit of "loss aversion" on this chart (explained here). The designer color-coded the views data into blue, brown and gray but didn't have it in him/her to throw out the sub-categories, which slows down cognition and adds hardly to our understanding.

I like the chart title that explains what it is about.

Turning to the D corner of the Trifecta Checkup for a moment, I suspect that this chart only counts videos that have at least one play. (Zero-play videos do not show up in a play log.) For a site like Youtube, a large proportion of uploaded videos have no views and thus, many creators also have no views.

***

My initial reaction on Twitter is to use a mirrored bar chart, like this:

I ended up spending quite a bit of time exploring other concepts. In particular, I like to find an integrated way to present this information. Most charts, such as the mirrored bar chart, a Bumps chart (slopegraph), and Lorenz chart, keep the two series of percentages separate.

Also, the biggest bar (the gray bar showing 97% of all creators) highlights the least important Youtubers while the top creators ("super-creators") are cramped inside a slither of a bar, which is invisible in the original chart.

What I came up with is a bar-density plot, where I use density to encode the importance of creators, and bar lengths to encode the distribution of views.

Each bar is divided into pieces, with the number of pieces proportional to the number of creators in each segment. This has the happy result that the super-creators are represented by large (red) pieces while the least important creators by little (gray) pieces.

The embedded tessellation shows the structure of the data: the bottom third of the views are generated by a huge number of creators, producing a few views each - resulting in a high density. The top 38% of the views correspond to a small number of super-creators - appropriately shown by a bar of low density.

For those interested in technicalities, I embed a Voronoi diagram inside each bar, with randomly placed points. (There will be a companion post later this week with some more details, and R code.)

Here is what the bar-density plot looks like when the distribution is essentially uniform:

The density inside each bar is roughly the same, indicating that the creators are roughly equally important.

 

P.S.

1) The next post on the bar-density plot, with some experimental R code, will be available here.

2) Check out my first Linkedin "article" on this topic.

 

 

 

 

 

Not following direction or order, the dieticians complain

At first glance, this graphic's message seems clear: what proportion of Americans are exceeding or lagging guidelines for consumption of different food groups. Blue for exceeding; orange for lagging. The stacked bars are lined up at the central divider - the point of meeting recommended volumes - to make it easy to compare relative proportions.

The original chart is here, on the Health.gov website.

The little icons illustrating the food groups are cute and unintrusive.

It's when you read further that things start to get complicated. The last three rows display a flipping of the color scheme, with orange on the right, blue on the left. Up to this point, you may understand blue to mean over the recommended value, and orange is under. Suddenly, the orange is shown on the right side.

The designer was wrestling with a structural issue in the data. The last three food groups - sugars, fats and sodium - are things to eat less. So, having long bars on the right side is not good. The orange/blue colors should be interpreted as bad/good and not as under/over.

***
The problem with this design is that it draws attention to this color flip - that is to say, it draws attention to which food groups are favored and which ones are to be avoided. This insight is actually in the metadata, not what this dataset is about.

In the following chart, I enforce the bad/good color scheme while ignoring the direction of good. The text is adjusted to use words that do not suggest direction.

Dieticians are probably distressed by this chart, given that most Americans are lagging on almost all of the recommendations.

In a final edit, I re-ordered the categories.