The connected dot plot (aka, gap chart, dumbbell chart, and barbell chart) has become my “go to” for showing the differences within demographic groups for virtually any survey data question type (Percent top two boxes, Check-all-that-apply, Median hours worked, etc.)

But what do you do when the values within different groups overlap or are the same?

In this blog post we’ll see how you can offset those values so they don’t occlude each other.

By the way, this technique works for any type of connected dot plot, not just those related to survey data.

Some examples in the wild

“How do organizations that are really good at this stuff handle situations like this?” That’s the question I ask when I’m tasked with visualizing data. I very much like how Pew Research Center handles gaps between different members within a demographic group.

Here’s an example that I featured in The Big Picture that shows the difference among Black, Hispanic, and White survey respondents to a survey Pew Research Center conducted in 2018.

Figure 1 — A connected dot plot from Pew Research Center.

Here’s how I applied the same technique to visualize Likert scale data broken down by different demographics.

Figure 2 — A connected dot plot I created to handle multiple Likert-scale questions broken down into multiple generational segments.

Yes, but how do you handle identical values?

Look closely at the last row in the preceding graphic. Do you notice there are only three dots that are visible? The gray dot for “Generation X” is behind one of the other dots.

So… how do the pros address this problem?

Let’s look at another example from Pew Research where two values are identical.

Figure 3 — Identical values are offset from each other so you can see both dots.

Ah, they offset the values from the dotted line.

How do you do something like this in Tableau?

Creating a connected dot plot with offset values in Tableau

Here’s a data set I put together to illustrate the challenges. Notice the identical values in Cat 1 and Cat 2 and the near identical values in Cat 2.

Figure 4 — The sample data.

Building a simple dot plot

Here’s a simple dot plot. We’ll hold off connecting the dots until we figure out how to do the offset.

Figure 5 — A simple dot plot with dots that occlude other dots.

In the marks card we specify that we want to display circles (1) and that we want to have a different colored circle for each Item (2). The MAX(0) on Rows (3) is what allows us to display the dotted line. By default Tableau shows a dotted line for zero values for rows and columns.

Let’s see how to offset the dots so that we can see all five dots for both categories.

Creating a simple offset

Let’s deal with items that have equal values. Instead of placing MAX(0) on Rows we’ll fashion a calculated field called [Offset 1] that will move one of the items up a little higher on the Y-axis (1).

Figure 6 — A dot plot where any dot that would occlude another dot gets offset by .2 vertically.

The field [Offset 1] is defined as follows:

IF SUM([Value])-LOOKUP(SUM([Value]),-1)=0  then .2 else 0 END

This translates as

If the current dot is equal to the previous dot (meaning when you subtract them you get zero), place the current dot at .2; otherwise, place the current dot at 0. The “.2” is arbitrary, it just needs to be any value greater than 0.

Since we’re using a table calculation (LOOKUP) we’ll need to define the scope of that table calculation so that is does the comparison across [Item] and then starts over when it gets to a new [Category].

Figure 7 — Defining the scope of [Offset 1].

Figure 8 — Fixing the axis’ start and end values.

The fixed axis creates more headroom which in turn forces the .2 dot to be closer to the dotted zero line. If we hadn’t adjusted the axis the offset dots would look like this.

Figure 9 — Tableau’s automatic axis places the offset dot too high.

We’ve come up with a solution that addresses if there are a pair of dots that have the exact same value, but what about the two Items in Cat 2 that have values that are very close (25.0 and 25.1)?

Let’s see how we can address this as well as craft a symmetrical offset so we move one of the dots up and the other dot down.

Creating a symmetrical offset that handles close values

Here’s a more sophisticated approach to the Offset field.

IF ABS(SUM([Value])-LOOKUP(SUM([Value]),-1))<=.3  then .2

ELSEIF ABS(SUM([Value])-LOOKUP(SUM([Value]),1))<=.3 then -.2

ELSE 0

END

This translates as

If the current dot is within .3 of the previous dot, then move the dot to .2, else, if the current dot is within .3 of the next dot, then move the dot to -.2; otherwise, place the current dot at 0.

Note that we need to adjust the y-axis so that it now accommodates both the dots that have been placed higher and the dots that have been placed lower. Here’s what things look like with the new [Offset 2] field.

Figure 10 — Symmetrical offset that handles values that are close, in this case within .3.

Adding the connecting line

Let’s follow the standard approach to creating the connecting line. We’ll duplicate the SUM(Value) pill (1) and change the mark type for this second pill to be a line instead of a circle (2). Here’s what we get.

Figure 11 – Running into a speedbump, in this case a zig-zaggy line, in trying to construct the connected dot plot.

The line indeed connects the dots, but because the dots are offset we get a zig-zaggy line (3). We need a way to flatten the line, as it were (we’ll deal with it being multi-colored in a bit).

This is where I received some help from Ken Flerlage who came up with a clever way to change the level of detail for the line (4, above). Currently, there are five Items / dots. Suppose we came up with a new dimension, based on the dimension [Item] that had only two elements, with those elements being the minimum and maximum Items?

That’s what Ken fashioned with this calculated field called [High and Low Item]:

IF {FIXED [Category], [Item]: SUM([Value])} = {FIXED [Category]: MAX({FIXED [Category], [Item]: SUM([Value])})} OR

{FIXED [Category], [Item]: SUM([Value])} = {FIXED [Category]: MIN({FIXED [Category], [Item]: SUM([Value])})} THEN

[Item]

END

This translates as

If within each Category and Item, the sum of an Item is the same at the Maximum value within that Category and Item, or equal to the Minimum value within the same, then display that Item; otherwise, don’t display anything.

We need to use {Fixed} so that Tableau gives us a dimension that we can place on Color, Detail, or Path in the Marks card.

Let’s swap out Item with this new field and see what we get.

Figure 12 — Getting closer. The lines are straight but they are not the correct length as Tableau is adding together too many values.

The calculated field High and Low Item placed on color (1, above) produces straight lines, but the lines are too long (2) as Tableau is adding too many values together. We’re also getting some extra coloring because of the Null values (3).

If we right-click Null in the color legend and select Hide we’ll have a line that is the correct length.

Figure 13 — Very close. The lines are straight and they are the correct length, but we want them to be a light gray.

Now let’s move High and Low Item from Color to Path (you can also add it to Detail instead) and we have all the elements we need to create the connected dot plot.

Figure 14 – [High and Low Item] is now on Path (1). The Null values are still hidden, and we no longer have a multi-colored line.

The hard part is done, we just need to

1. Make this a dual axis chart by right-clicking the second SUM(Value) pill and selecting Dual Axis.
2. Right-click in the top axis and select Synchronize Axis.
3. Right-click the top axis and select Move marks to back (this just swaps the two green pills at the top).
4. Increase the thickness of the line by selecting the pill associated with the line and increasing the size on the Marks card.
5. Make the line color lighter.
6. Hide the top axis by right-clicking and deselecting Show header.
7. Hide the Offset 2 header by right-clicking on the Y-axis and deselecting Show header.

Here’s the result:

Figure 15 — A connected dot plot with offset values.

Bonus: how to ditch the legend

Whenever possible I do my best to avoid color and size legends. In the example below I’ve labeled each of the items in the first row so we don’t have to refer to a legend. You do this by placing [Item] on Label in the Marks card (for the Circle pill only) but do NOT turn on Show marks label. Instead, right-click each dot, select Mark Label and then Always show.

Figure 16 — A connected dot plot with direct labeling for the dots in the first row.

Conclusion

The connected dot plot remains my “go to” for showing gaps among different demographic groups for virtually any type of survey question. The technique discussed here addresses how you can offset elements when you have identical and near identical values.

Here’s the workbook that you can download.

The post A better way to connect the dots appeared first on Data Revelations.

Salma covers a lot of ground in her post, citing an example from Scott Berinato’s book Good Charts where Berinato underscores the importance of sketching and how he and Walter Frick brainstormed on how to compare the cost of selected Apple products to median monthly household income, based on when those products were released.

I followed Salma’s link to a Harvard Business Review article and was struck by this proportional square chart, one of several in the article (Figure 1).

Figure 1—Proportional square chart from Harvard Business Review article. See https://hbr.org/2014/10/apple-luxury-brand-or-mass-marketer.

The big squares represent the median monthly income for the year the Apple product was released and the smaller squares embedded represent the costs for the associated Apple product.

“Ah,” I thought “yet another example of trying to show actuals and percentages at the same time.”

This need to show both comes up so often that I make sure to cover it in my dashboard design workshops. Here’s an example that shows Percent of Goal and Sales across four regions (Figure 2).

Figure 2–Showing Percent of Goal and Sales at the same time

It’s a combination of two charts. The bar-in-bar chart allows an easy comparison of percent of goals and the sales chart allows an easy comparison of actual sales.

So, how might we apply this approach (or something similar) to the Apple product costs data?

Actuals vs. Percentages, Take One

Before I go further, I want to make clear that I have no problems with the way the authors of the HBR article presented their data. As there isn’t a need to make accurate comparisons, the novel chart types are likely to draw readers in and the storytelling around the visualizations are entertaining and informative.

But being a professional chart-looker-atter I couldn’t help thinking “how would I show this, especially in a business setting where I may have to compare many products?” It would be difficult to make accurate comparisons across many proportional square charts as they take up a lot of space and people are better at comparing the length of bars than they are estimating the area of squares or circles.

With this in mind, let’s look at the raw data for six Apple products, the years they were introduced, the median monthly income for each year, and the associated product costs (Figure 3).

Figure 3–Table showing year, product name, product cost and monthly median income.

Below I present all this information in a compact space that I hope makes it easy for people to make accurate comparisons and better understand the data (Figure 4).

Figure 4–Product cost, monthly income, and percent of monthly income.

The bar-in-bar shows the actual product costs and actual median monthly incomes for each year / product while the pie charts (yes, pie charts!) shows the percentage, or “bite,” each product takes out of the median monthly income for that year.

Not So Fast! Actuals vs. Percentages, Take Two.

I conveniently left out products that had costs that were greater than the monthly median incomes for the year the product was released. Consider the table below that adds the Apple II and Macintosh to the list of products (Figure 5).

Figure 5–Highlighting two years when the selected Appl product costs more than the monthly median income.

How do we show more than 100% with a pie chart? Yes, I’m sure that Apple Watch enthusiasts are thinking about closing the rings on their watches and what the watch shows when you exceed goals (Figure 6).

Figure 6–How the Apple Watch shows when you have exceeded your fitness goals in two of the three categories. I’m not a big fan of the concentric circles.

In this case I would go with two bar-in-bar charts, like the one shown in Figure 7.

Figure 7–Bar-in-bar for actuals and percentages.

I think that many things that were buried in the table now become clear. For example, we can see the median monthly income almost doubled between 1984 and 1998. We can also see that those early Apple computers were very expensive!

Where to Learn More

This certainly isn’t the only way to show actuals and percentages at the same time. My fellow authors and I address challenges like these, that importance of sketching, and many other data visualization trials in the forthcoming book Dashboards That Deliver.

My thanks to Salma Sultana for both writing the original post and to Salma and Amanda Makulec for providing feedback on the article.

The post Actuals? Percentages? Why not show BOTH? appeared first on Data Revelations.

I think the number one infraction in data visualization is the misuse of color.

If I were pressed for time and had to distill my recommendations on color into a Haiku, it would be this:

Make everything gray

Except the few things you think

Should be highlighted

Yes, this is simple, bordering on the simplistic, but it works in so many situations.

Consider the kaleidoscopic mess below on the left. What do you want the audience to make of this? It’s an overdose of categorical colors that provide no value. Now look at the chart on the right. The purposeful use of a highlight color makes it clear that you want people to compare Vietnam to the other countries.

Having a different color for each country makes the chart on the left hard to read but with the chart on the right, it’s easy for your audience to know where to focus (from the book The Big Picture by Steve Wexler).

There are only five ways to use color in data visualization. Once you understand how to apply each of these approaches, your visualizations will be much more effective.

From the upcoming book Dashboards That Deliver by Steve Wexler, Jeffrey Shaffer, Andy Cotgreave, and Amanda Makulec. These “five ways” are also discussed in The Big Book of Dashboards and The Big Picture.

I make sure to cover the “five ways” in all my workshops. People who come in with headache-inducing visualizations leave with the ability to make clean, aesthetically pleasing visualizations that help your audience see what you want them to see.

Dashboards That Deliver: How to Design, Develop, and Deploy Dashboards That Work will be available September 2025.

The post A Haiku about color in data visualization appeared first on Data Revelations.

Out of work, I put a web site together and started blogging about visualizing survey data using Tableau. My hope was that these posts would help me land a consulting gig.

As I waited for the phone to ring, I realized I couldn’t keep producing content in a vacuum.

I needed validation. I needed someone to tell me what I was doing was helpful.

Shortly thereafter, two “someones” emailed me to tell me that they thought I had written something useful.

This was the fuel I needed to keep writing.

There’s a lot of noise on social media platforms. People starting out in a field have a hard time being seen and heard.

If you come across a newcomer who is producing something good, let them know. This might be the fuel they need to continue.

…

“Newcomer” doesn’t mean young. It means someone who is attempting something new or is entering a new field.

And if you are curious about the two people who gave me what I needed when I needed it, it was Joe Mako and Michael Cristiani.

The post Let people know that you see them and hear them appeared first on Data Revelations.

Last week I saw a reprehensible and delusional post from Elon Musk on his social media platform, X.

Shortly thereafter I read Robert Reich’s Guardian opinion piece on how Elon Musk has gained a concerning level of power over US national security.

For certain, Musk is a genius. He is our era’s Thomas Edison and Henry Ford combined. But he is also a man who has unfettered control of virtually all the world’s satellite internet service through Starlink. On a whim he can shut down a country’s ability to communicate and perhaps its ability to defend itself.

I cannot in good conscience be part of a platform run by such a powerful and egomaniacal person.

Musk is on pace to be the world’s first trillionaire by 2027 and my absence from X will have no impact on him whatsoever. But with my dropping out he will have one fewer active users to tout when he courts advertisers for his social media platform.

I do want to continue to have engaging data visualization discussions. I just won’t be having them on X. You can find me here:

LinkedIn: https://www.linkedin.com/in/swexler/

Bluesky: https://bsky.app/profile/datarevelations.com

The post Why I Quit X (Twitter) appeared first on Data Revelations.

Background

I recently had to address some serious health issues that required a lengthy hospital stay followed by a long in-patient rehab. I won’t go into the details, but over the course of my stays people would ask me what I did for a living.

I would answer I was involved in data visualization.

Not one person who asked knew what that was, so I would try to explain what I did.

Sometimes it was easy, like when I had a follow up appointment where I was tethered to an EEG monitor for five days and it produced a real-time chart like the one shown here.

Sparklines, sparklines, sparklines! If you are into sparklines, it doesn’t get better than this!

I would then point out that if you know how to read this constantly changing chart (I certainly don’t… same with MRIs), it was a lot easier than trying to read a constantly changing table full of numbers.

But for the people who were not involved with the real-time EEG monitor… how would I explain what data visualization is and how it could be so useful for the people trying to care for me?

Trying to explain without being able to show

It was hard not to lean into Tableau’s mission statement and say that “I help people see and understand data.” I would provide examples of where a really good chart–or a combination of charts–would reduce time to insight. Indeed, time to insight became the driving force of getting people to see the value of data visualization. In this setting we weren’t interested in getting people to “feel” something about the data.

We discussed everything from the number of patients on the floor, to how long they stayed, to time to respond to patient call bells, to the most pernicious of all the things I saw during my stay: the 1970s-styled lab reports.

You know them, the reports that show albumin levels, cholesterol, white blood cell counts, etc. You see your numbers and the “acceptable” range.

What happens when you fall out of the acceptable range? How far out of the range are you? When should you be mildly concerned and when should you panic? Is it easy to see this? I wrote about this in my book The Big Picture and shared an example that Amanda Makulec presented at a visualization conference several years ago. Here’s a snippet from the book.

Reimagining Lab Tests

Amanda Makulec is a consultant and Executive Director at the Data Visualization Society. After her father had a second battle against kidney cancer, Makulec decided that she should submit to a genetic test and discovered that she, like her father, had Lynch syndrome that put her at greater risk for several different types of cancers. The lab results she received were presented in a table like the one shown in Figure 9.44

Lots of body parts, lots of numbers, and lots of red. And a lot of frustration with the way the data was presented.

It’s hard not to be particularly alarmed by the first set of numbers (55%–82%) and just focus on the 82%, suggesting a four out of five risk of that body part developing cancer.

Armed with both a master’s degree in public health and data visualization expertise, Makulec could interpret the results but wondered about those not as well steeped in reading and understanding health-related numbers. She wished the testing company had presented the results this way instead.

If you are wondering why there is a dot for the general population and a very wide range for Makulec, it’s because of the amount of data the testing organization maintains. They have a great deal of data about the general population but sparse data from people with Lynch syndrome, so the margin of error is very wide, hence the wide range bars.

These institutions can do better

I’m scratching my head wondering why these institutions do such a poor job with both lab reports and operations data (I probably shouldn’t be scratching my head as there are still some scarring). I’m sure if the data visualization community got the chance, we could help make both the internal operations data and the patient-facing data so much easier to interpret.

What do you think about this and again, how do you explain what you do and how data visualization can help? Join the discussion here.

Note: the in-patient rehab where I recovered for two weeks was so helpful I plan to volunteer my services. I am very grateful to them.

The post How do you explain what data visualization is? appeared first on Data Revelations.

Thank you to Alli Torban, Anna Foard, and Ben Jones for their feedback.

My Delight

A few weeks ago, I saw this terrific interactive graphic in one of my social media feeds. It was part of a recent Makeover Monday challenge.

What a great way to show rank and magnitude at the same time, as opposed to a bump chart, which only shows ranking. There are lots of nice touches in this, including the text along the top of the bar, the thin reference lines that show where 100% is, and the right-aligned numbers.

And it got even better. Here’s what happens when you see the same graphic rendered on a phone.

Two great approaches to showing the data. Just who is this shivaraj c practitioner who makes such good stuff?

I saw some other renderings that I liked a great deal, including this connected dot plot from Jhoie Victor, below. Notice the color of the connecting line to highlight where Young exceeds Old (blue) and where Old exceeds Young (orange.)

And just look at the huge gaps between older and younger people, especially with respect to saving more money, pursuing a career ambition, taking up a hobby, spending less time on social media, and improving diet.

Those are really large gaps.

Which got me curious about the data.

Which led to my dismay.

My Dismay

Let’s have a look at the Makeover Monday site to see what was available to participants.

We have a link to the data (1), an article that queues up the discussion (2), and the data source (3).

Here’s the data set.

This seems straightforward… but did anybody pause to vet the quality of the data? That is, did anybody stop and think “I’m about to amplify the findings in this data… am I sure the data is correct?”

My guess is that few did, as a visit to the source data would show that the number of respondents who participated in this part of the survey was quite small, which means the findings are going to be fuzzy, at best.

Let’s have a look.

Everything above the line (1) is a response count, and everything below the line is a percentage. The survey sample is robust, with over 2,000 participants (2).

Of these participants we see that 29% of younger people but only 6% of older people will make new year’s resolutions for 2024 (3). THAT is a very interesting finding with a solid response count, but not the focus of the Makeover Monday exercise.

Now here’s where things get dicey. Look at the count of people who in fact plan to make resolutions (4). There are only 330 respondents. How do you get that number?  If only six percent of 541 people who are 65 or older plan to make a resolution, that means the total number of respondents in that demographic category is 32. And 29% of the 154 people who are between 18-24 years old planning to make resolutions yields a response count of 45. Do the same multiplication across the other two age groups and you get 330 responses.

Yeah, Yeah, I know 29% of 154 is 44.66. Clearly, they rounded off the percentages.

So, we have 32 respondents who are 65+ and 45 respondents who are between 18 and 24.

When you have a response count this low, the margin of error is extremely high, but in the ten Makeover Monday submissions I reviewed nobody mentioned this.

Let’s Explore Showing Margin of Error

Sure, there’s a very big gap between the first and fourth questions (“saving more money” and “pursuing a career ambitions”) but there also looks to be a big gap for the sixth question, “taking up a new hobby” (32% vs 14%).

But… there’s also a massive margin of error.

Let’s see how to compute the margin of error for the 32% of the respondents aged 18 – 24 who stated they plan to take up a new hobby.

The formula to compute the margin of error is

MoE = z * sqrt(p̂ ⋅* (1-p̂)) / sqrt(n)

Where z is the z-score associated with the level of confidence. A z-score of 1.645, indicating that we want to be within 1.645 standard deviations above or below the population mean, is what we use for a 90% confidence level. If we want a 95% confidence level we would increase the Z-score to 1.96.

The p̂ represents the percentage of respondents who said yes to the question, in our case 32%, and n is the number of respondents (45). Using this formula, we get

MoE = 1.645 * sqrt(0.32 * (1-0.32)) / sqrt(45)

and that yields a margin of error of a little over 11%.

It turns out that the 32% for the younger cohort is 32%… + / – 11 points! This means that if we were to survey a sample of the population repeatedly, we can state with 90% confidence that the true value would be between 21% and 43%.

That’s a very wide range.

Here’s the same data, but with showing the margin of error.

There are only three questions where you can say, with 90% confidence, that there is a gap between older and younger participants (“saving more money”, “pursuing a career ambition”, and “spending less time on social media”). All the other questions have an overlap.

If you are thinking “the range of values is so large… how can anyone report results with authority?” then you have an idea why I’m dismayed.

What should be done, and who should be doing it?

Before going any further, I don’t think disseminating questionable conclusions from this data is going to do any harm, except for maybe denting the reputations of folks who didn’t bother to vet the data and at least display a footnote noting the margin of error.

As for who is responsible for vetting the data, when I first brought this to the attention for the team that’s now running Makeover Monday (Irene Diomi, Chimdi Nwosu, and Harry Beardon), here’s how one of the leaders responded:

“It is on me… I picked the dataset. These are good insights on the data quality, I have to admit I didn’t spend enough time looking at it but it is important to take into account sample size and weights.

I’ll keep this in mind for future datasets.”

I love it when people own up to a mistake, but I don’t think this is just the leaders’ responsibility. I think ALL OF US have a responsibility to interrogate the data thoroughly before publishing visualizations that make assertions.

Realize this was not Real World, Fake Data (one of my favorite data visualization social media projects). With this Makeover Monday challenge, every time a person publishes a visualization asserting the source conclusions, that person is amplifying the idea that “these are facts!”

As I said before, in this case I think there was little harm in asserting questionable differences in New Year’s Resolutions proclivities. Who knows, a larger survey sample may reveal the findings from this survey are spot on.

But there have been cases where people amplified false assertions that may have done some real harm. I encourage folks to read this blog post that I wrote with my friend and colleague, Jeff Shaffer.

I don’t want to come off as “holier than thou” as I have screwed up on countless occasions and I have to actively remind myself that vetting the data is not somebody else’s job. It’s a critical part of my job.

And it’s a critical part of your job, too, so please… do your job well.

Recommended reading

If you don’t already own it, purchase a copy of Ben Jones’ gem of a book Avoiding Data Pitfalls.

Good Calculators has a very good Margin of Error calculator that allows you to input values.

I’ve written several blog posts about how to visualize margin of error in survey data. Here are two that you may find valuable.

Showing uncertainty in survey results – Data Revelations

and

More thoughts on showing Margin of Error in survey data with Tableau – Data Revelations.

The post Just how carefully did you look at the data? appeared first on Data Revelations.

Overview

There are a lot of ways to show measures over time for multiple categories. I want to explore what works, when it works, and make sure that a particular technique catches your interest.

But first… a big shout out to Nick Desbarats whose book, Practical Charts, provides both observations and approaches that I had not considered. Strongly recommended.

Background

I enjoyed a great discussion during one of my data visualization and storytelling workshops. The catalyst was a combination of this hot mess “before” picture…

… combined with this uncluttered, curated “after” picture.

We explored how the after picture only highlights one line, shows only the starting and ending values, has a clear title takeaway, and has interactivity so the audience can focus on a different category.

Does this work?

This led to one attendee sharing a visualization he had been working on which had five lines, each a different color. It looked something like this.

The attendee’s biggest concern was having a different color for each line. Indeed, anytime I see more than four categorical colors in a visualization I get nervous. Will this be too much for my audience to process?

We discussed the visualization for a bit and realized that because there was little or no overlap among the lines, we could even ditch color completely and it was still easy to understand the chart.

But what if there’s overlap?

The example we had wasn’t particularly noisy, so I wanted to explore what happens when you have five lines that bounce around a bit.

For me, this is on the threshold of non-readability. I know that is very subjective, but as a professional chart looker-atter, if I find something difficult, I can almost guarantee that my non-professional chart looker-atter audience will find it difficult, too.

Also, the “noisy” chart fails the monochrome test, at least in my opinion, as shown below.

Answering the “It Depends” questions

My friend and co-author, Andy Cotgreave, has a t-shirt with the words “It Depends” emblazoned on the front. It’s how he (and I, for that matter) address all of the “which chart is best / how many colors should I have / is it okay to use a pie chart” questions we get during our workshops and presentations.

This is one of the reasons I like Nick Desbarats book so much. It does an effective job of moving the discussion from “It Depends” to “under these circumstances, do this.”

For example, I was ready to say, unequivocally, that ten lines, each a different color, will never work.

But in his book, Nick points out where a 10+ line chart could work.

Source: Practical Charts by Nick Desbarats

I agree that the chart on the left is readable, but I still think it’s risky to try something like this in a “living” dashboard because when the data changes, the readability may change, too.

So, what would be a safer approach?

THIS is the one you should have in your arsenal

I’m going to jump to the punch line and share my preferred method for when you either have an interactive dashboard, or when you know for sure which lines are the “important” lines:

Highlight the few lines that are important and use gray for everything else.

It works when there is little overlap.

And it works when there is a lot of overlap / noise.

Indeed, this is the technique I used to turn the hot mess “before” picture into a clear and easy-to-understand “after” picture at the beginning of this article.

Yes, but what if I need to show all the values for all the categories at the same time?

If for whatever reason you cannot just focus on a few lines. and need to show all the categories at the same time, a trellis chart (also called a “small multiples” chart) may solve your problem.

Here’s a trellis line chart.

And here’s a trellis area chart.

The big drawback with the trellis arrangement is that it’s difficult to make comparisons when things are in different rows or columns.

Consider the example below. While comparing (1) and (2) is reasonably easy, comparing (1) and (3) is difficult as they don’t share a common baseline.

You can get around this by adding some interactivity so that when you highlight a particular point for one segment (e.g., Chairs) it will highlight the applicable point in all segments. Below we compare measures for the month of May across the ten different segments.

But … if you are offering your audience an interactive dashboard, why not just do the line highlighting that we saw earlier?

One more approach worth considering

If stacking ten lines is too much, but a full trellis is too distributed, Nick Desbarats suggests trying an intermediate small multiples / trellis chart, like the one shown below.

Give it a try

I’ve embedded a downloadable Tableau workbook at the end of this article that has all of the approaches we’ve discussed here, plus a few more. Use the tabs along the top to explore the different views.

The post How many lines are too many? appeared first on Data Revelations.

I recently participated in a LinkedIn Live discussion about how to visualize uncertainty in survey data with Bob Walker, Anna Foard, and Jon Cohen. I find myself contemplating whether we have a type of moral obligation to try to make sure our audience understands that there may be a big difference between the survey results we report and the true results.

A question I posed to my colleagues was why show the result as a dot with error bars behind the dot? Why not just show the error bars? The same goes for showing a gradient—are we in fact misleading people to think that the true value is more likely to be in the center?

A snapshot of our LinkedIn Live session.

We’re not the only people grappling with this question. Financial Times graphics guru John Burn-Murdoch discussed this during his Outlier Conference keynote presentation and asserts humans hate uncertainty and may ignore “best practice” techniques to help them see the range of possible outcomes. He then showed other approaches to get people to see and understand uncertainty.

John Burn-Murdoch shows different ways 538 has shown confidence intervals.

Here’s a link to Burn-Murdoch’s presentation, queued to where he starts discussing the challenge. Great stuff.

And here again is a link to a recording of our LinkedIn Live session. Note that our discussion about whether we should just show error bars and not over-emphasize the point in the middle starts at around 29 minutes in.

Finally, in our get-together we had planned to show a survey sample simulator (say *that* three times fast!) to show that if you sampled, say, 75 people 100 times, you’d see that the surveyed results are sometimes outside of your expected error bars. You can try it for yourself here.

Survey Sample Simulator from ArtOfStat.com.

The post More thoughts on visualizing uncertainty in survey data appeared first on Data Revelations.

A deep thanks to Anna Foard and Jonathan Drummey for their assistance, Ben Jones for the foundational work behind for my initial explorations, and Ryan Corser for asking me to look into this.

Update: an earlier version of this post was titled “Visualizing Uncertainty in Longitudinal Survey Data.” Robert Walker of Surveys & Forecasts LLC points out that a longitudinal study asks the same questions to the same people over time. In my example we survey different people. 

Background

I offer an on-demand course about visualizing survey data with Tableau and a perk that comes with the course is that I conduct monthly “what’s on your mind” sessions.

Last month Ryan Corser asked me to explore showing confidence intervals / margin-of-error around longitudinal data. That is, he wanted his stakeholders to be able to see just what plus-or-minus X points looks like with respect to Likert responses over time.

I get it. I’ve seen too many clients brush off how unreliable survey results are when you don’t have enough responses. Maybe some visual ammunition will help get across why “n=24” when you are surveying a large population is probably not good enough to make a business decision.

For this example, I look at the percentage of people who either strongly or moderately agree with various statements using a 7-point Likert scale (percent top two boxes).  You can find the data set here. Note that I broke the responses up into four different years to show changes over time. The original data set did not have a date field.

Leveraging existing work

I first delved into showing statistical significance / confidence in 2015 when a client asked me to have a visual designator indicating whether an increase or decrease from a previous period was “statistically significant.” They wanted to be able to say “The percentage of people who agree went from 40% to 48%, and we can say with 95% confidence that there was in fact an increase.”

Great. The dot means the change is, for lack of a better term, noteworthy.

But wouldn’t it be great if we could easily see why there’s a dot?

Showing error bars / confidence intervals

This desire to show people just how shaky their survey results might be led to explorations of how to show error bars / confidence intervals in Tableau.

Given the overlap of the error bars in the example above (the lines going through the dots) one cannot assert with confidence that there’s a substantive difference between survey responses from men versus those from women.

This in turn led to forays into visualizing how the number of responses directly impacts the confidence interval.

Armed with these step-by-step blog posts, let’s see how we can tackle showing the margin of error associated with Likert-scale longitudinal data.

Computing Top 2 Boxes

I copied and pasted a lot of calculated fields from the dashboards embedded in the two blog posts I cited. Not all of these fields worked “as is.” For example, the Percent Top 2 Boxes from the first example was computed like this:

SUM(

    IF [Value]>=4 THEN 1 ELSE 0 END

    )

/ SUM([Number of Records])

This translates as “If the response was a 4 or a 5, count it, then divide by everyone who answered the question.”

In our example, the questions use a 7-point Likert scale, where the “strongly agree” and “agree” were 1 and 2, respectively. I altered the calculation as follows.

SUM(IF [Values]<=2 then 1 ELSE 0 END)/COUNT([Data])

This translates as “If the response was a 1 or a 2, count it, then divide by everyone who answered the question” where COUNT([Data]) provides the same functionality as SUM([Number of Records]).

If you’re curious as to why I didn’t just create a field called [Number of Records] it’s because I used Tableau Relationship feature and didn’t do any data prep. See this blog post for a discussion.

Iterate, Iterate, Iterate

Here’s a quick walk-through of some of my visualization attempts.

I started by emulating the dots with error bars approach.

Note that the Confidence is set for 90% (P=.10). Don’t worry, the embedded dashboards at the end of this article allow you to set the Confidence to 80%, 90%, 95%, and 99%.

Yes, the margin of error for 2020 and especially 2022 are very wide (+/- 8 points for 2022). This is because there are considerably fewer responses for those years than for 2019 and 2021.

By the way, if you change the Confidence level to 99%, the margin of error for 2022 is +/- 13 points.

But I digress.

I then thought that since we’ve got longitudinal data, we should use lines instead of dots and came up with this.

The same calculations I use to get the top and bottom ranges for the error bars drives the dotted lines.

I like the dotted lines a lot but decided that I would rather shade the regions between the dotted lines to better show the range of possible values.

If you’re curious as to how to shade the region between two lines using Tableau, just perform an internet search. I went with what I thought was the easiest approach: an Area chart where I don’t stack the marks. You’ll find the “don’t stack” option under the “Analysis” menu. Color the bottom area white, the top area gray, and make sure to change the opacity to 100%.

Done.

So… what now?

What Assertions Can We Make?

Just what do we do with this lovely chart showing the responses with the error range shaded?

I find myself thinking about the things my clients would want to report. Something like “in 2019, X percent indicated that they agree or strongly agree that fluffernutters are delicious. In 2022 that percentage increased to Y percent.”

With that in mind, let’s just focus on the first year and the last year.

Can you say, with confidence, that for the second question (I feel that my agency on the whole is well managed) that there was an increase from 2019 to 2022? That is, that there is no overlap between the range of responses for 2019 and 2022?

Here’s another view where we use reference bands to see if the lowest possible value for 2019 might be greater than the highest value for 2022, and that the lowest for 2022 might be greater than the highest for 2019.

So, for the second and third questions you can assert, for 90% of samples, the resulting confidence intervals would imply the true values for 2019 are smaller than the true values for the second and third questions for 2022, although the gap for the second question is small. You should not make this assertion with the first and fourth questions as there is overlap with the margin of error bands.

Incidentally, if you increase the Confidence level to 95% the only thing you can assert is that there was an increase from 2019 and 2022 for the third question.

So, with the Confidence level set to 90%, we can see there are gaps for the second and third questions. Wouldn’t it be nice if we could just have a dot letting our audience know when the changes are “significant” and when they are not, rather than having bands and asking them to look for gaps?

And wouldn’t it be great if we could just leverage the work from the first blog post and use already-figured-out calculations for displaying a dot when the change between two periods is significant?

Steve Learns About Directional and Non-Directional Hypothesis

And that’s exactly what I did, and I got this confusing result.

There’s no gap for the first question, so why does my calculation indicate that the change between 2019 and 2022 is statistically significant, using the tried and true z-test calculation I wrote about back in 2015?

My head-scratching led to two calls for help, the first one being to Anna Foard. Anna explained the difference between a directional and non-directional hypothesis which eventually led me to this wonderful blog post from Analytics Vidhya.

With the error bars and error bands, the assumption that there is a degree of fuzziness for 2019 and fuzziness for 2022.

With my test for statistical significance dot, the assumption is that the results for 2019 are pin-point accurate but there is fuzziness for the 2022 values.

Now I can see why we have that extra dot for the first example. There’s nothing “wrong” with it; it’s just performing a directional vs. non-directional test.

But this got me thinking about how to create a calculation that looks for overlap in the fuzzy values for 2019 and the fuzzy values for 2022.

Making a Different Dot

I’ll confess that I struggled with this as some decisions I made early in the process forced me into having to craft a table calculation.

This led to my second call, this time to Jonathan Drummey. Jonathan conducted a clinic on how to assess the data and what capabilities we might want. For example, we discussed whether we would want to have the ability to show differences for more than two periods; that is, looking at multiple periods and not just the starting and ending period.

For the sake of simplicity, I elected to just look at the starting and ending periods. My apologies in advance that all the calcs are hard coded to look at 2019 and 2022.

Here’s the pseudocode for what we need to determine.

If the upper limit for Top 2 Boxes % for 2019 is greater than or equal to the lower limit of the Top 2 Boxes % for 2022

OR

If the lower limit for the Top 2 Boxes for 2019 is greater than or equal to the upper limit of the Top 2 Boxes % for 2022 then there is an overlap so don’t display a dot; otherwise, display a dot.

This appears simple enough. Why not just have a calculation like this one?

IF MAX(YEAR([Date]))=2019 THEN [Top 2_Lower limit] END

The problem is that our visualization needs to break things up into different years and that the calculation for 2019 is only available for 2019. When I’m in the 2022 column, Tableau can’t “see” what the value was for 2019, it just knows what the value is for 2022.

Here’s a snippet of the worksheet I crafted with Jonathan to see how the fields all work together.

In (1) we can see the lower limits for 2019 and 2022. In (2) and (3) I have separate calculations for 2019 and 2022 but because they render nulls and aren’t present across both columns, we can’t perform calculations on them.

But notice the values for (4) and (5) and how they render the same results for both columns.

Ah, now we’re talking!

Just how did we get that?

I had started to look into LOOKUP() and PREVIOUS_VALUE() calcs (and if I ever plan to compare more than two periods I’ll look into this again). Jonathan suggested something far simpler to pad the values across both columns and ensure that the lower limit for 2019 is present in both the 2019 and 2022 columns.

WINDOW_MAX(IF MAX(YEAR([Date]))=2019 then [Top 2_Lower limit] END)

By computing this by [Date] it makes Tableau apply the calculation across all columns.

We fashioned three additional table calculations and then combined them into this calculation for Non-direct dot.

IF [2019 Top 2 Boxes %  Upper Limit Padded]>=[2022 Top 2 Boxes %  Lower Limit Padded]

OR

[2019 Top 2 Boxes %  Lower Limit Padded]>=[2022 Top 2 Boxes %  Upper Limit Padded]

THEN "" else "⬤"

END

Here’s a view that has a dot for both the directional and non-directional approaches. You’ll also find an interactive version of this at the end of the blog post.

Important: I don’t think I would subject my stakeholders to two different dots. Decide which approach is most appropriate and just show one dot.

Conclusion

The main purpose of data visualization is to help people better see – and understand – their data. My hope is that with this article you have more tools to help people see when they can, and when they should not, make assertions about longitudinal survey results.

Here’s the embedded workbook with various dashboards. Feel free to tinker and download.

The post Visualizing Uncertainty in Likert Data Over Time appeared first on Data Revelations.