We tried a new venue — the Winking Lizard on Bethel Road — and were pretty pleased with the accommodations (private room, private bar, very reasonable prices), so I expect we’ll be back.

Relatively new dad Bryan Cristina had a child care conflict with his wife…so he brought along Isabella (who was phenomenally calm and well-behaved, and is cute as a button!):

I presented on a topic I’m fairly passionate about — data visualization. The presentation was well-received (Monish Datta really did tweet that it was “stellar”)  and generated a lot of good discussion. I had several requests for copies of the presentation, so I’ve modified it slightly to make it more Slideshare-friendly and posted it. If you click through on the embedded version below, you can see the notes for each slide by clicking on the “Notes on Slide X” tab underneath the slideshow, or you can download the file itself (PowerPoint 2007), which includes notes with each slide (I think you might have to create/login to a Slideshare account, which it looks like you can do quickly using Facebook Connect).

I had fun putting the presentation together, as this is definitely a topic that I’m passionate about!

* The “Monish Datta” reference in the title of this post, while accurate, is driven by my never-ending quest to dominate search rankings for searches for Monish. I’m doing okay, but not exactly dominating.

Similar Posts:

• Dashboard Design Part 3 of 3: An Iterative Tale
• An Excel Dashboard Widget
• Dashboard Design Part 2 of 3: An Iterative Tale
• Dashboard Design Part 1 of 3: An Iterative Tale
• The “Action Dashboard” — Avinash Mounts My Favorite Soapbox

© Tim for Gilligan on Data by Tim Wilson, 2010. | Permalink | No comment | Add to del.icio.us
Post tags: color blind, dashboards, infographics, sparklines, tips

Feed enhanced by Better Feed from Ozh

I’ve been put on the spot since then a couple of times to do one of two things:

• Explain why pie charts are evil, or
• Agree that one or another particular usage of a pie chart is appropriate

After catching up on some blog reading yesterday morning and seeing an excellent example of pie chart alternatives from Jon Peltier, and then watching seven presentations yesterday, six of which used the same basic presentation template, and five of which stuck with a pie chart for the sole non-text slide in the presentation, how could I not write another post?! Let’s see how succinct I can make it (don’t hold your breath that you could read the whole thing before exhaling!).

Yes, There is ONE Thing That a Pie Chart Does Well

This kills me, because there’s one way, in a a very narrow set of circumstances, that pie charts do marginally better than alternatives. All THREE of the following criteria have to be met for this to be the case:

• Exactly 2 or 3 categories that make up the “whole”
• A fairly significant difference in % makeup for each of the categories
• Plenty of space available to present the information

99 times out of 100 when pie charts get used, all of these criteria are not met. But, there, I’ve admitted that there is a situation where pie charts are appropriate.

Of course, mullets are an appropriate hairstyle if you are prone to both warm ears and spontaneous hair donations…but that doesn’t mean I’m going to sport one!

Of Course, We Must Start with a Before/After Example

With only the category names changed, below is one of the pie charts I saw yesterday:

In my experience, a simple horizontal bar chart is a better option (among a variety of better options):

Why is this a better option? Oh, let me count the ways…

1. Rainbows Are Good in Princess Tales — Not in Data Visualization

When it comes to data visualization, a chart that doesn’t rely on multiple colors always trumps a chart that does. Four reasons:

• If you use subtle/muted colors, you can’t get past 4 or 5 categories before you are asking the person reading the chart to work hard to distinguish between subtle shading differences
• If you use bright/high-contrast colors, you’re asking your user to put on sunglasses to keep from wincing at the visual overkill
• Roughly 10% of men suffer from some form of color-blindness — it’s darn tricky to nail a palette with more than a small handful of colors that works across the various types of the condition (of course, if you’ve got a secret agenda to have women take over the world, this is one way to contribute, as color blindness is exceedingly rare in women)
• Maybe you’re presenting your chart in glorious, projected color…but are you sure no one is going to try to print it in black-and-white?

These are all issues with any pie chart that has more than 3 categories. None of these are an issue with a horizontal bar chart.

2. Labels, Labels, Labels

If you’ve every constructed a pie chart in Excel, you’ve run into the challenge of trying to get all of the wedges labeled right there on the chart. Excel continues to make odd choices as to where to wrap text in pie charts, and the circular nature of the whole layout means some wedges have plenty of horizontal labeling room, while others have almost none. You’ve tried some (or all) of the following:

• Using leader lines for some of the wedges so you can label the most troubling wedges somewhere more spacious
• Abbreviating the category names
• Strategically rotating the chart so that the labeling all happens to work (it never does)
• Rearranging the underlying data so that the pie wedges occur in a different order (which also never works)

After fiddling with the above, you finally break down and yank the labels from the chart and just use a legend. This is bad, bad, BAD! Scroll back up to the pie chart example above and pretend you’re actually trying to interpret the data, but pay attention to how many times you look back and forth between the legend and the pie. This is putting a totally unnecessary strain on your brain! Take a look at the horizontal bar chart — no jumping back and forth needed!

With a horizontal bar chart, the label sits right next to the data, and it doesn’t need to be abbreviated to do so (this is one reason that I find horizontal bar charts to be better than vertical column charts in many cases — with a horizontal orientation, the labels have more width with which to work).

3. Those Pesky Near-Zero Values

Pie charts suck at the small percentages. Small percentage categories wreak havoc on the labeling issue, for sure, but they’re also nearly impossible to compare to each other. In the example above, the smallest percentage is 3%, and that’s almost manageable. But, heaven forbid you have a couple of pesky sub-one-percent categories, and you’re looking at wedges that look suspiciously like the lines between wedges.

4. Seeing Small Differences

Fundoogles & Flibbers came in at 3%, while Dracula’s Mickety Micks came in at 5%. Do the wedge sizes really look different? That’s a fundamental challenge with pie charts — we don’t do a very good job of comparing the areas of these odd sorta-triangular-but-with-one-curved-side shapes. In the case of the bar chart, all you have to compare is lengths — much easier.

5. Economy (of Space) Is a Virtue

Check out the overall size of the charts. While they have the same font size, the same text displayed, and the same width, the bar chart is 20% shorter…and it could have been shorter still! Bar charts are more efficient space-wise. With pie charts, and largely because of the other issues listed above, it’s often necessary to make the chart larger and larger to make it readable.

Of Course, This Exampel Was At Least Flat

This post would be twice as long if I went into the additional issues of using the “3D effect” version of the pie chart.

[Update] Always Room for Improvement

Of course, the danger of posting a “here’s a better way” is that you leave yourself open for suggestions as to how the better way can be improved! See Naomi’s comment below. She raises a good point — basically, that I didn’t do a great job of heeding the data-pixel ratio with my bar chart! So, below is a revised version.

In a subsequent email exchange, Naomi made the case for keeping the x-axis and the numbers, but simply removing the “%” signs entirely and putting the word “Percent” in the axis label:

Her main point is that numbers can be read more easily if they are not cluttered with symbols like dollar signs and percent signs. And, her case for keeping the gridlines and labeled axis is that it helps show that the bars are drawn to scale — there hasn’t been any incorrect or misleading scaling (intentional or not — in the same spate of presentations that spurred this post, there was a bar chart with an accompanying table of data…and one of the bars was clearly not accurate).

I’m partial to the version with all of the lines removed, but, at this point, the debate is at a much healthier level than “pie vs. bar,” so I’m happy!

© Tim for Gilligan on Data by Tim Wilson, 2009. | Permalink | 9 comments | Add to del.icio.us
Post tags: bar charts, pie charts

Feed enhanced by Better Feed from Ozh

I’ll walk through some of the components here in detail, but, first, a handful of key points:

• There is no redundant information — it’s not uncommon to see dashboards (or reports in general) where there is a table of data, and that table of data gets charted, and the values for each point on the chart then get included as data labels. This is wasteful and unnecessary.
• Hopefully, your eyes are drawn to the bold red elements (and these highlights should still pop out for users with the most common forms of colorblindness — I haven’t formally tested that yet, though) — this is really the practical application of the vision I laid out in my Perfect Dashboard post.
• I have yet to produce a dashboard solely comprised of these widgets — there are always a few KPIs that needs to be given more prominent treatment, and there are other metrics that don’t make sense in this sparkline/trend/current format
• I do mix up the specific measures on a dashboard-by-dashboard basis. In the example above, showing the past two years of trends by month, and then providing quarterly totals and comparisons, makes the most sense based on the planning cycle for the client. But, that certainly is not a structure that makes sense in all situations.

And now onto the explanation of the what and why of each element, working our way from left to right.

Metric Name

This one hardly warrants an explanation, but I’ll point out that I didn’t label that column. That was a conscious decision — the fact that these are the names of the metric is totally obvious, and Edward Tufte’s data-ink ratio dictates that, if it doesn’t add value, don’t include it!

Past 12 Months Sparkline

The sparkline is another Tufte invention, and it’s one that has really taken off in the data visualization space. That’s good, because sparklines are darn handy, and the more people get used to seeing them, the less there will need to be any “training” of dashboard users to interpret them. Google Analytics has been using sparklines for a while, even, so we’re well on our way to mass adoption!

One tweak on the sparkline front that I came up with (although I’m sure others have done something similar): I add a second, gray sparkline for either the target or the prior reporting period. I like that this gives a quick, easily interpretable view of the metric’s history over a longer period — has it been tracking to target consistently, consistently above or below the target, or bouncing back and forth? Is there inherent seasonality in the metric (signified by both the black and gray sparklines having similar spike/dip periods)?

One limitation of sparklines is that they don’t represent magnitude very well. If, for instance, a particular metric is barely fluctuating over time, then, depending on how the y-axis is set up, the sparkline can still show what looks like a wildly varying value. It’s a minor limitation, though, so I’ll live with it.

4-Month Trend Arrow

The 4-month trend is the single icon that results from a conceptually simple (but a little hairy to calculate) assessment of the most recent four data points. That was the punchline of an earlier post on calculating trend indicators. Whether the basis of the trend is months, weeks, or days can vary (not within one dashboard, generally, but as a standard for the dashboard overall), as well as whether it’s 4, 5, 6, or more data points. It’s a judgment call for both driven by the underlying business need that the dashboard supports.

I promise, promise, promise to make a simplified example of this arrow calculation and post it in a future post — check the Comments section for this post to see if a linkback exists (I’ll come back and update this entry as well once it’s done)

Current

Typically, when sparklines are used, the exact value of the last point in the sparkline is included. In the example above, I’ve done something a little different, in that I actually provide the sum of the last three data points. This is a quarterly dashboard, but the sparkline has a monthly basis to it to show intra-quarter trends. If the current value is sufficiently below the target threshold, then the value is automatically displayed as bold and red.

There are certainly situations where “Current” would actually be the last point on the sparkline. Like the trend arrow calculations, it’s a judgment call based on the business need that the dashboard supports.

YOY

In the example above, there is a comparison to the prior year. But, this could be a comparison to the target instead. Target-based comparison is even better — straight period-over-period comparisons tend to feel like something of a cop out, as prior periods really are more “benchmarks” than true “targets.” Now, setting a target as something like “15% growth over the prior year” has some validity! That would then impact both the gray sparkline, the “when does Current go bold red,” and this %-based calculation.

28 Data Points

In the version of the widget above, there are 28 unique pieces of data presented for each metric: the metric name (1), the black sparkline (12), the gray sparkline (12), the trend indicator (1), the current value (1), and the year-over-year growth percentage (1). And that’s not counting the conditional formatting that highlights values as bold and red when certain criteria are met. That’s a key aspect of the widget design. 28 sounds like a lot of data to represent for a single metric. Yet, they seem pretty digestible in this format, don’t they?

Let me know what you think. Does this work? What doesn’t work?Similar Posts:

• Calculating Trend Indicators
• Data Visualization Tips and Concepts (Monish Datta calls it “stellar”)*

© Tim for Gilligan on Data by Tim Wilson, 2009. | Permalink | 4 comments | Add to del.icio.us
Post tags: sparklines, trend indicators

Feed enhanced by Better Feed from Ozh

With the well-warranted increasing use of sparklines, which are inherently trend-y representations of data, I like to be able to put a meaningful trend indicator that complements the sparkline. Throughout this post, I will illustrate trendlines, but I’m really focussed on trend indicators, which are a symbol that indicates whether the trend in the data is upward, downward, or flat. Although there are a few minor tweaks I’d love to make once Excel 2010 is released and allows the customization of icon sets, I’m reasonably happy with their 5-arrow set of trend indicators:

They’re clean and clear, and they work in both color and in black and white. And, with conditional formatting, they can be automatically updated as new data gets added to a dashboard or report. While I won’t show these indicators again in this post, the trendlines I do show are the behind-the-scenes constructs that would manifest themselves as the appropriate indicator next to a sparkline or numerically reported measure.

I’ll use a simple 12-period data set throughout this post to illustrate some thoughts (not as a sparkline, but the principles all still apply):

Trends are slippery beasts for several reasons:

• Noise, noise, noise — all data is noisy, which means it’s easy to over-read into the data and spot a trend that is not really there
• The aircraft carrier vs. the speedboat conundrum — the more data points you use, the more stable your trend, but the longer it takes to collect enough data to identify a trend, or, worse, to determine if you’ve truly impacted the trend going forward

Let’s start this exploration by walking through some of the common ways that “trend” judgments get made and point out why they’re troubling. I will then show an alternative that, while only marginally more complex to implement, works better when it comes to specifying trend-age.

Trending Approaches of which I’m Leery

Trending Based on the Change Over the Previous Period

The most common way I see trends reported is on a “change since the previous period” basis.

In this example, the trend would be an “up” because the data went up from the prior period to the current period. The problem with this is that, if you look at the longer pattern of data, you see that the data is pretty noisy, and it’s entirely possible that this “trend” is entirely a case of noise masking the true signal.

Trending Over an Extended Period

Another way to trend your data, which Excel makes very simple, is to add a trendline using Excel’s built-in trending capabilities (converting this trendline to an indicator would require some use of a couple of Excel functions that I’ll go into a bit in my recommended approach later in the post).

With this method, the trend would be indicated as “slightly up.” While this may be a valid representation of the overall trend…it seldom seems quite right to use it. The trend gets impacted heavily by any sort of big spikes (or dips) in the data. These keep the same upward or downward trend for a very long period of time. I had a blog post during March Madness one year that wound up driving a big spike in traffic to my site. While it was legitimate for that spike to show an upward trend when I looked at my traffic that week or month, that spike has now wreaked havoc on the macro trend indicator that Google Analytics has shown ever since — for several months that spike kept my overall trend up, and, then, once that spike passed the fulcrum of the tool’s trend calculation, it caused the reporting of a downward trend for severals subsequent months. Through the whole period, I had to mentally discount what the trend indicator showed.

Year-Over-Year Trending

Because seasonality wreaks havoc with trendlines, it’s not uncommon to see trend indicators based on year-over-year results — if the current reporting period is a higher number than the same period a year ago, then the trend is up. For trending purposes, this combines the worst of the two prior examples — it takes a very small number of data points (subjecting the assessment to noise) and it uses ancient history data in the equation.

This isn’t to say that comparisons to the same period in the prior year (or even the same period in the prior quarter, since many companies see an intra-quarter pattern) are bad. But, the question those comparisons answer differs from a trend: a trend should be an indication of “where we are heading of late such that, if we continue on the current course, we can estimate whether we will  be doing better or worse next week/next month,” while a year-over-year comparison is more a measure of “did we move positively from where we were last year at this time?”

Trending Approaches I Feel Better About

I’ve spent an embarrassing amount of time thinking about trending over the past four or five years, but I’ve finally settled on an approach that meets all of these criteria:

• It balances the number of data points available for the trend with the sluggishness/timeliness of the results
• It’s reasonably intuitive to explain
• It passes the “sniff test” — while a trend indicator may initially be a little surprising, on closer inspection, the user will realize it’s legit

The last bullet point is really a combination/result of the first two.

My Failed Exploration: Single Point Moving Range (mR)

Because of criteria above, I’ve discarded what I thought was my most promising approach — using the single point moving range (mR). A light bulb went off last spring when I took an intermediate stats class, and, although the professor glossed over the moving range formulas, I thought it was going to be the answer that would allow me to solve my trendline quandary — it would look at the “change over previous period” and determine if that change was sufficiently large to warrant reporting a measurable trend. After noodling with it quite a bit… I don’t think that it works for the purposes of trend indicators. For chuckles, a moving range chart for the example in this post looks like the following:

If you want to read more about moving ranges, the best explanation I found was on the Quality Magazine web site. I’ll just stop there, though. We’ve already lost on the “reasonably intuitive” front, and I haven’t even calculated the control limits yet!

And Another Failed Exploration: the Moving Average

There’s also the “moving average” approach, which smooths things out quite a bit:

I always feel like the moving average is some sort of narcotic applied to the data — it makes things fuzzy by having a single data point factored into multiple points represented on the chart. But, I’ll grudgingly admit that it does have its merits in some cases.

My Approach to Trending (At Last!!!)

There are two key elements to my trending approach, and neither is particularly earth-shattering:

1. Break the data into smaller components than the reporting cycle
2. Trend only over recent data, rather than over the entire reported timeframe

Going back to the original example here, let’s say that I update a dashboard once a month, and that the dashboard primarily looks at data for the prior 3 months. In that case, the 12 data points each represent (roughly) one week. IF I simply reported the data on a monthly basis, then the chart would look like this:

That shows a clear upward trend, regardless of whether I look at the last month or the last two months of data. It would be hard not to put an upward trend indicator on this plot. But, we’re relying on all of three data points, and we’re going back three full reporting periods to draw that conclusion. Both of these are a bit concerning. Invariably, we’d want to go back farther in time to get more data points to see if this trend was real…and then we’re falling into the aircraft carrier dilemma.

Instead, though, I can keep the granularity of the reporting at a week, but only trend over the last four periods:

I don’t actually plot the trendline shown in the chart above. Rather, I calculate the formula for the line using the SLOPE and INTERCEPT  Excel functions. I then calculate the value of the 4-weeks-ago endpoint of the line and the most-recent-week endpoint of the line and look at the percentage change from one to the other. I actually set some named cells in my workbook to specify how many periods I report over (so I can vary from 4 to 6 or something else universally) as well as what the different thresholds are for a strong up, weak up, no change, weak down, or strong down trend.

In the example in this post, the change is a 16% drop, which usually would garner a “strong down” trend — very different from all the upward trends in the early examples! And, even somewhat counter-intuitive, as the most recent change was actually an “up.” If the entire range has been trending upward, as shown by the 3-point plot as well as by a close inspection of the raw basic data (think of it as a sparkline), then you already have that information available as the longer term trend, but, of late, the trend seems to be somewhat downward.

A Note of Caution

This post has gone through what works for me as a general rule. As I read back over it, I realize I’m setting myself up for a case of, “Yeah, you CAN make the data say whatever you want.”

I’m less concerned about prescribing a universally-effective approach to trend calculation as I am about putting out a cautionary tone on the various “obvious” ways to calculate a trend. The sniff test is important — does the trend work for your specific situation when you actually apply it? Or, have you adopted a simplistic, formulaic approach that can actually provide a very clear misrepresentation of the data?

And…a Nod to Efficiency and Automation

The prospect of introducing SLOPE and INTERCEPT functions may seem a little intimidating from a maintenance and updating perspective, but it really doesn’t need to be. By using built-in Excel functionality, these can be set up once and then dynamically updated as new data comes in. I like to build spreadsheets with a data selector so that the dashboard is a poor man’s BI tool that allows exploring how the data has changed over time. The key is to use some of Excel’s most powerful, yet under-adopted, features:

• Conditional formatting — especially in Excel 2007 where conditional formatting can make use of customized icon sets
• Named cells and named ranges — these are handy for establishing constants used throughout the workbook (thresholds, for instance) that you may want to adjust
• Data validation — using a cell as your “date range selector” that references a named range of the column that lists the dates for which you record the data
• VLOOKUP — because you used data validation, you can then use VLOOKUP to find the current data based on what is selected by the user
• Dynamic charts — these actually aren’t a “feature” of Excel so much as the clever combination of several different features; Jon Peltier has an excellent write-up of how to do this

If set up properly, a little investment up front can make for an easily updated report delivery tool…with meaningful trend indicators!Similar Posts:

• An Excel Dashboard Widget

© Tim for Gilligan on Data by Tim Wilson, 2009. | Permalink | 6 comments | Add to del.icio.us
Post tags: dynamic ranges, INTERCEPT, moving average, moving range, SLOPE, trend indicators, trendlines

Feed enhanced by Better Feed from Ozh

The third point in Goodman’s take is worth quoting almost in its entirety, as it is both spot-on and eloquent:

Everyone has a creative side, but someone who has never picked up a design book with an emphasis on data visualization should not implement dashboards for their own company and certainly not as a consultant. Dashboard development is not the forum to unleash creative juices when the intent is to monitor business performance. Working with clients who have educated themselves have[sic] definitely facilitated more productive engagements. Reading a book does not make you an expert, but it does allow for more constructive discussions and a smoother delivery of a dashboard.

“The book” of choice (in my mind, and, I suspect, in Goodman’s) is Few’s Information Dashboard Design: The Effective Visual Communication of Data (which I’ve written about before). Data visualization is one of those areas where spending just an hour or two understanding some best practices, and, more importantly, why those are best practices, can drive a permanent and positive change in behavior, both for analytical-types with little visual design aptitude and for visual design-types with little analytical background.

Goodman goes on in his post to be somewhat ambivalent about tool vendors’ responsibility and culpability when it comes to data visualization misfires. On the one hand, he feels like Few is overly harsh when it comes to criticizing vendors whose demos illustrate worst practice visualizations (I agree with Few on this one). But, he also acknowledges that vendors need to “put their best foot forward to prove that their technology can deliver adequate dashboard execution as well as marketing sizzle.” I agree there, too.

© Tim for Gilligan on Data by Tim Wilson, 2009. | Permalink | 2 comments | Add to del.icio.us
Post tags: Edward Tufte, Stephen Few

Feed enhanced by Better Feed from Ozh

Several weeks ago, one of the presenters in a meeting showed some data as a map overlay. As soon as she projected the first map, someone in the meeting quipped, “Good luck understanding this one, Jim!” Jim, you see, is colorblind. And, apparently, most of the people in the meeting knew it. Approximately 8% of men have some form of color blindness (it’s much more rare in women — only 1 in 200). And the overlays on the map were color-coded very subtly. Jim commented that it was hopeless!

As it happened, I was exploring a fresh set of data that same week, as we’d recently rolled out some new customer data capture capabilities. As I worked through how best to present the results, I decided to grab a colorblind-friendly palette from the web and use it in the visualization of the information. I’d hoped to find a site with one or more Excel files that I could download with such a palette, but, worst case, I was prepared to snag a palette and manually update my Excel file (for future sharing on this blog, of course!).

No. Such. Luck!

What I did find was a slew of information on the different types of color blindness (which I’ll touch on briefly in a bit), as well as a bevy of almost-useful tools and palettes:

• How to make figures and presentations that are friendly to Colorblind people — ultimately, I used the palette that is ~2/3 of the way down this page for my spreadsheet (the figure labeled “Set of colors that is unambiguous both to colorblinds and non-colorblinds”).  Mr. Excel actually references this palette and provides a macro that will update a workbook’s palette with this palette. The downside of this palette is that, while it may be plenty functional, I can’t say I’m wild about it from an aesthetic viewpoint. But, I’d spent the 30 minutes I’d given myself to dig, so I ran with it.
• Colorjack Color Blindness Simulation — a view of the color spectrum as seen by people with eight different forms of color blindness. That’s informative…but doesn’t really provide a realistic way to build a functional palette for data visualization purposes.
• Colorjack — a nifty tool for finding a color palette. Unfortunately…there’s no way to test how colorblind-friendly any of the palettes are
• Colorblind Web Page Filter — there were a number of tools for sale that would simulate how content would appear to people with different forms of colorblindness, but this is the (free) online tool I wound up using for the exercise below. It couldn’t be easier to use — you just provide a URL and what form of color blindness you’re interested in, and it renders it

So, aside from the one palette that was solely focussed on functionality and not at all on aesthetics, I struck out. As I pondered this over the next few days, it occurred to me that, perhaps Excel’s default colors always seemed so gosh-awful because they were actually developed explicitly with colorblindness in mind. I could not find any documentation to support the theory…so I turned left and headed down that rathole to see if I could figure it out myself.

The exercise was pretty simple. I created a 10-color bar chart using the Excel 2007 default palette. Note: This was created purely for palette-testing — this actual chart is a great example of needlessly using more color than is needed! Here’s the chart:

Like the one colorblind-friendly palette I found online, I really don’t like the aesthetics of this palette. It’s been toned down a bit from the Excel 2003 (and earlier) versions, which is good, but it still seems rather harsh. Could that be for colorblind compatibility? I think so! I took the chart above and ran it through the Colorblind Web Page Filter mentioned above for the four most common types of color blindness (as described in a Pearson report by Betsy J. Case):

Overall, the palette seems workable in all four situations. The first three colors absolutely work. Color 4, as well as color 5, start to lose a little contrast from color 1, but they still seem manageable. Color 5 and color 7, as well as color 10, start to get a little problematic in some cases, but, if you’re going beyond four colors in a single chart, you might need to reconsider your chart type anyway. Right?

Now, one final test: for achromatopsia. On the one hand, this is extremely rare. On the other hand…it’s common when your office has a lot of black-and-white printers:

Apparently, any palette that works in grayscale is a quick way to check for compatibility with all forms of colorblindness. It’s also…a best practice. Interestingly, the Excel 2007 palette really lays an egg here, in that colors 1, 2, and 4 are all barely distinguishable!

Clearly, there is an opportunity here to test a variety of functional, attractive palettes for grayscale printability and the top four forms of colorblindness and develop something better than the Excel defaults. But, that’s an exercise for another time. I think I’ll aim for the first four colors of the palette being “highly distinguishable” in all scenarios and the next four being “functionally distinguishable.” What do you think? Would this be useful? What else should I take into consideration?

© Tim for Gilligan on Data by Tim Wilson, 2009. | Permalink | 9 comments | Add to del.icio.us
Post tags: colorblind, palette

Feed enhanced by Better Feed from Ozh

It’s hard to imagine anyone taking issue with the proclaimed intent of recovery.gov, though — make the process as transparent as possible, including how much money is going where, when it’s going, and what ultimately comes of it. It was a day or two before I found myself at a computer with time to check out the site…and I was disappointed. In the NPR interview, the interviewer commented how the site was slick and clean. Reality is “not so much.”

Now, I did once take a run at downloading the federal budget to try to scratch a curiousity itch regarding, at a macro level, where the federal government allocates its funds. On the one hand, I was pleased that I was able to find a .csv file with a sea of data that I could easily download and open with Excel. On the other hand, the budget is incredibly complex, and it takes someone with a deeper understanding of our government to really translate that sea of data into the answers I was looking for. Really, though, that wasn’t a surprise:

The data is ALWAYS more complex than you would like…when you’re trying to answer a specific question.

To the credit of recovery.gov, they clearly intended to show some high-level charts that would answer some of the more common questions citizens are asking. Unfortunately, it looks like they turned over the exercise to a web designer who had no experience in data visualization.

Examples from the featured area on the home page:

The overall dark/inverse style itself I won’t knock too much (althought it bothers me). And, the fact that the gridlines are kept to a minimum is definitely a good thing. My main beef is admittedly a bit ticky-tack. There was an earlier version where there was a $30 B gridline, and that has since been removed — that gridline clearly showed the “30.5 B point” being below the midway point between 20 B and 40 B. Clearly, someone would have to really be scrutinizing the graph to identify this hiccup, but someone will.

When presenting data to an audience, the data as it stands alone needs to be rock solid. If it contradicts itself, even in a minor way, it risks having its overall credibility questioned.

So, moving on to some more egregious examples:

We get a triple-whammy with this one:

• Pie charts are inherently difficult for the human brain to interpret accurately
• Pie charts are even worse when they are “tilted” to give a 3D effect — the wedges on the right and left get “shrunk” while wedges on the top or bottom get “stretched”
• Exploding a pie chart and then providing a pie chart of just the wedge…just ain’t good

Two questions this visualization might have been trying to answer:

• How much of the stimulus plan is devoted to tax benefits?
• How much of the stimulus plan is going to the “Making Work Pay” tax credit?

Without doing any math, can you estimate either one of these? For the first question, you’re estimating the size of the small wedge on the left pie chart. It looks like it’s ~ 1/4 of the pie, doesn’t it? In reality, it’s 37%! For the second question, you have to combine your first estimate with an estimate of the lavender wedge in the right pie chart…and that’s way more work than it’s worth. If you do the math, you’ll get that the lavender wedge works out to ~7% of the entire left pie. A simple table or a bar graph would be more effective.

And, finally, the estimated distribution of Highway Infrastructure Funds:

Well, that’s just silly. There is NO value of making these bars come flying out of the graph. Really.

Now, to the site’s credit, it takes all of 3 clicks to get from the home page to downloading .csv files with department-specific data and weekly updates (which includes human-entered context as to major activities during the prior week). That’s good (assuming it’s not unduly cumbersome to maintain)! And, I’m sure the site will continue to evolve. But, I’d love to see them bring in some data visualization expertise. The model for the visualization should be pretty simple:

1. Identify the questions that citizens are asking about the stimulus money
2. Present the data in the way that answers those questions most effectively
3. Link to the underlying data — the aggregate and the detail — directly from each visualization

As it turns out, Edward Tufte has already been engaged (thanks to Peter Couvares for that tip via Twitter), and is doing some pro bono work. But, it’s not clear that he’s focussing on the high-level stuff. I would love to see Stephen Few get involved as well — pro bono or not! Or, hell, I’d offer my services…but might as well get the Top Dog for something like this.

Starting today, the site is hosting a weeklong online dialogue to engage the public, potential recipients, solution providers, and state, local and tribal partners about how to make Recovery.gov better. I’ve submitted a couple of ideas already!

Similar Posts:

• Data Visualization — Few’s Examples
• Test Your Data Visualization IQ
• Data Visualization — March Madness Style
• Stephen Few’s Derivation of Tufte: The Data-Pixel Ratio

© Tim for Gilligan on Data by Tim Wilson, 2009. | Permalink | 3 comments | Add to del.icio.us
Post tags: Edward Tufte, recovery.gov, Stephen Few

Feed enhanced by Better Feed from Ozh

• The Intermediate Stats class I’ve been taking (taught by an Ohio State professor on-site at my company, Nationwide…as in Nationwide: Car Insurance) is wrapping up, and the last day of the class includes a group presentation
• Jon Peltier really knocked one out of the park on the PTS Blog with his re-working of a weird, radial chart showing oil prices over time
• Jim Knight wrote Will Death by PowerPoint Soon be a Thing of the Past, where he summarized the highlights of a one day Presentation // Reboot workshop put on by Garr Reynolds and Nancy Duarte
• Knight’s post introduced me to Duarte (and her blog), who, according to Knight, is “most famous for developing the design of Al Gore’s Academy Award-winning An Inconvenient Truth presentation“
• In Duarte’s latest post — a video entry while she is on vacation in Hawaii — she reviews Advanced Presentations by Design: Creating Communication that Drives Action by Andrew Abela, which sounds like it is a bit dry, but makes a strong case for putting a lot more rigor behind developing a presentation

Put all of those together, and I’ve got a mental convergence of PowerPoint usage, presenting effectively (which goes well beyond “the deck”), and data visualization. These are all components of “effective communication” — the story, the content, how the content is displayed, how the content is talked to. In one of Reynolds’s sets of sample slides, you can clearly see the convergence of data visualization and PowerPoint. And, even he admits that this is a tricky thing to post…because it removes overall context for the content and it removes the presenter. Clearly, there are lots of resources out there that lay out fundamental best practices for effectively communicating in a presentation-style format. Three interrelated challenges, though:

• The importance of learning these fundamentals is wildly undervalued — it sounds like Abela’s book tries to quantify this value through tangible examples…but it’s a niche book that, I suspect, will not get widely read by the people who would most benefit from reading it
• “I need to put together a presentation for <tomorrow>/<Friday>/<next week>” – we’re living under enormous time pressure, and it’s incredibly easy to get caught up in “delivering a substantive deliverable” rather than “effectively communicating the information.” When I think about the number of presentations that I’ve developed and delivered over the past 15 years, the percentage that were truly effective, compelling, and engaging is abysmally small. And that’s a waste.
• Culture/expectations — every company has its own culture and norms. For many companies, the norms regarding presentations are that they are linear, slide-heavy, logically compiled, and mechanically delivered affairs. For recurring meetings, there is often the “template we use every month” whereby the structure is pre-defined, and each subsequent presentation is an update to the skeleton from the prior meeting. Walk into one of those meetings and deliver a truly rich, meaningful, presentation…and your liable to be shuttled off for a mandatory drug test, followed by a dressing down about “lack of proper preparation” because the slides were not sufficiently text/fact/content-heavy. <sigh>

What’s interesting to me is that I have spent a lot of time and energy boning up on my data visualization skills over the past few years. And, even if it takes me an extra 5-10 minutes in Excel, I never send out something that doesn’t have data viz best practices applied to some extent. As you would expect, applying those best practices is getting easier and faster with repetition and practice. So, can I do the same for presentations? And, again, that’s presentations-the-whole-enchilada, rather than presentations-the-PowerPoint-deck. Can I balance that with cultural norms — gently pushing the envelope rather than making a radical break? Can you? Should you?Similar Posts:

• Harvey Balls: A Good Way to Ramp Back Up
• PowerPoint the Application vs. the Application of PowerPoint
• Am I Ever BeHIND on Posting…

© Tim for Gilligan on Data by Tim Wilson, 2009. | Permalink | 2 comments | Add to del.icio.us
Post tags: Andrew Abela, Garr Reynolds, Jon Peltier, Nancy Duarte, Nationwide, PowerPoint, Presentation Zen, presentations, presenting, PTS Blog

Feed enhanced by Better Feed from Ozh

The “game flow” graph…how in the WORLD is that telling me anything? That the score goes up as the game goes on? Really? Ya think?

My friend was referring to the diagrams that ESPN.com is providing for every game in the tournament. The concept of these graphs is pretty simple: plot the score for each team over the course of the game. For instance, the “Game Flow” graph for the Oklahoma vs. Morgan State game looks like this (you can see the actual graph on the game recap page — just scroll down a bit and it’s on the right):

This isn’t an exact replication, but it’s pretty close — best I could manage in Excel 2007 (the raw data is courtesy of the ESPN.com play-by-play page  for the game). ESPN’s graph is a Flash-based chart, so it’s got some interactivity that the image above does not (we’ll get to that in a bit).

The graph shows that the game was tight for the first 4-5 minutes, then Oklahoma pulled away, Morgan State made it really close mid-way through the first half, and then Oklahoma pulled away and never looked back. My friend had a point, though –  the dominant feature of the graph is that both lines trend up and to the right…and any chart of a basketball game is going to exhibit that pattern (actually, the play-by-play for that game has a couple of hiccups such that, when I originally pulled the data, I had a couple places where the score went down due to out-of-sequence free throw placement…but I noticed the issue and fixed it). In business, we’re pretty well conditioned to see “up and to the right” as a good thing…but it’s meaningless in the case of a basketball game.

Compare that graph to a game that was much closer – the Clemson vs. Michigan game (the graph on ESPN’s site is on the recap page, and the raw data is on the play-by-play page):

This was a tighter game all through the first half. Clemson led for the first 7-8 minutes, Michigan pulled substantially ahead early in the second half, and then things got tight in the last few minutes of the game. But, again, both lines moved up and to the right.

These charts are not difficult to interpret:

• The line on top is the team that is leading
• The distance between the lines is the size of the lead
• The lines crossing signifies a lead change

But, could we do better? Well, my wife and kids are out-of-town for the week (spring break), I have the social life you’d expect from someone who blogs about data and data visualization, and the fridge is well-stocked with beer. Party. ON!

At best, my level of basketball fan-ness hovers right around “casual.” Still, I follow it enough to know the key factors of a game update or game upset (Think: “Hey, Joe. What’s the score?”). Basically:

• Who’s winning?
• By how much?

(If there’s time for a third data point, the actual score is an indication of whether it’s a high scoring shootout or a low scoring defense-oriented game.)

Given these two factors as the key measures of a game, take another look at the graphs above. When the game is tight, you have to look closely to assess who is winning. And, determining how much they’re winning by requires some mental exertion (try it yourself: look back at the last graph and ask yourself how much Michigan was winning by halfway through the second half).

This is just begging for a Stephen Few-style exercise to see if I can do better.

First, the Oklahoma/Morgan State game:

Rather than plotting both team’s scores, with the total score on the Y-axis, this chart plots a single line with the size of the lead — whichever side of the “0″ line the plot is on is the team that is winning. The team on the top is the higher seed, and the team on the bottom is the lower seed. I added the actual score at halftime and the end of the game, as well as each team’s seed. Compare that chart to the much closer Clemson/Michigan game:

The chart looks very different — focussing on what information fans really want and presenting it directly, rather than presenting the data in a way that requires mental exertion to derive what the fan is really interested in: who’s winning and by how much? While the graphs on ESPN’s site allow you to mouse over any point in the game and see the exact score and the exact amount of time remaining, it’s hard to imagine who would actually care to do that – better to come up with an information-rich and easy-to-interpret static chart than to get fancy with unnecessary interactivity.

A few other subtle changes to the alternative representation:

• I tried to dramatically increase the “data-pixel ratio” (Few’s principle that the ratio of actual data to decoration should be maximized) — this is a little unfair to ESPN, as their site is working with an overall style and palette for the site, but it’s still worth keeping in mind
• I used color on the Y-axis to show which team’s lead is above/below the mid-line. The numbers below the middle horizontal line are actually negative numbers, but with a little Excel trickery, I was able to remove the “-” and change the color of the labels (all done through Custom number formatting)
• By putting the top seed on the top, looking at a full page of these charts would quickly highlight the games that were upsets

I’m my own worst critic, so here are two things I don’t like about the alternate charts above:

• The overall palette still feels a little clunky — the main data plot doesn’t seem to “pop” as much as it should, even though it’s black, and the shaded heading doesn’t feel right
• While the interpretation of the data requires less mental effort once you understand what the chart is showing, it does seem like this approach requires another half-second of interpretation upr front that the original charts don’t require

What do you think? What else could I try to improve the representation?Similar Posts:

• Test Your Data Visualization IQ
• Stephen Few’s Derivation of Tufte: The Data-Pixel Ratio
• Recovery.gov Needs Some Few and Some Tufte
• Data Visualization — Few’s Examples
• The Best Little Book on Data

© Tim for Gilligan on Data by Tim Wilson, 2009. | Permalink | 7 comments | Add to del.icio.us
Post tags: ESPN, game flow, March Madness, Stephen Few

Feed enhanced by Better Feed from Ozh

I was struck by how absolutely horridly the information was presented. A note at the bottom of each chart indicated that the same staff member had compiled each chart. Yet, there was absolutely no consistency from one chart to the next: the color palette changed from chart to chart (and none of the palettes were particularly good), a 3-D effect was used on some charts and not others (3-D effects are always bad, so I suppose I’d rather inconsistency than having 3-D effects on every chart), and totally different chart types were used to present similar information. On several of the bar charts, each bar was a different color, which made for an extremely distracting visualization of the information. 

I glanced around the room and saw that most of the other committee members had furrowed brows as they studied the information. It occurred to me that there was an undue amount of mental exertion going on to understand what was being presented that would have been better spent thinking through the implications of the information.

Ineffective presentation of data can significantly mute the value of a fundamentally useful report or analysis.

Later that evening, I found myself popping around the web — ordering my own copy of Stephen Few’s Show Me the Numbers, and, later, poking around on Few’s site. Specifically, I spent some time on his Examples page, browsing through the myriad before/after examples that clearly illustrate how the same information, presented with the same amount of effort, but using some basic common principles, dramatically reduce the mental effort required to understand what is going on.

It’s a fascinating collection of examples. And Show Me the Numbers is a seminal book on the topic.Similar Posts:

• Stephen Few’s Derivation of Tufte: The Data-Pixel Ratio
• Recovery.gov Needs Some Few and Some Tufte
• Test Your Data Visualization IQ
• Data Visualization — March Madness Style
• The Best Little Book on Data

© Tim for Gilligan on Data by Tim Wilson, 2009. | Permalink | 6 comments | Add to del.icio.us
Post tags: Stephen Few

Feed enhanced by Better Feed from Ozh