Dec 182011
 

Overview

Full Disclosure – my former company, WexTech Systems, developed and marketed the first commercial online help authoring tool, Doc-To-Help.  I use online assistance a lot and expect it to be, well, helpful.

Is it me, or has the quality of online help / user assistance declined in the past few years?  For example, when was the last time you saw a “what’s this?” feature inside a dialog box?  When was the last time you saw a dialog box with a help button?  Indeed, when was the last time you pressed Function Key F1 and something besides the main help page appeared, saving you the trouble of having to find the right help topic?

I spend a lot of time working with two programs, Tableau for data visualization and Sibelius for music notation.  I love both of these programs, but they both make me work a lot harder than I would like when I’m trying to understand a concept or complete a task.

The problem is not that the online help is poorly written.  On the contrary, the writing and examples are first class.  The problem is that I have to work a lot harder than I would like to actually find the topic that answers the questions I have.

In all fairness, Tableau’s online help / documentation is significantly better than Sibelius’, but Sibelius shipped with a 650+-page manual that has a really good index.

As for Tableau, under ideal circumstances I’d like it so that any time I’m flummoxed I’d be able to click someplace in the UI and the UI would take me to exactly the right topic in the online help.  Short of that, I’d at least like to see the return of a feature that disappeared with the release of Tableau 6.0:

The Index Button

While full text search – especially one with good word stemming, synonymy and disambiguation – is great to have, I maintain that it’s harder to find topics with Tableau versions 6.0 and later than it was with versions 5.2 and earlier.  Is this added difficulty solely the result of no Index feature?  Almost certainly not, as the product itself is much richer and the attendant online user documentation is more complex.

The beloved Index feature in the online help for Tableau 5.2

But putting it back wouldn’t hurt, would it?

Note: I realize that the ultimate goal is to fashion a product so easy and intuitive that a user never needs additional assistance via online help.  I think Tableau has made incredible strides with creating user affordances.

But you still need a good help system and you need to get users to the answers they need, when they need them.

 Posted by on December 18, 2011 1) General Discussions, Blog No Responses »
Oct 122011
 

Overview

I received a number of comments from my last post (see http://www.datarevelations.com/likert-scale-nirvana.html) including what I think is a better approach for staggered Likert scale visualizations from Joe Mako.

Joe’s approach with various “kitchen sink” added functionality is shown in the interactive dashboard below.  Notice that you can control whether neutral values are displayed, sorting, Likert scores, and so on.

There’s some great stuff going on here but for this blog post I will focus on a simpler version of this visualization that just presents the staggered Likert scale bars without the various bells and whistles.

How the Viz Works

Let’s look under the hood of this simpler visualization (it’s the second tab in the workbook.)

Here we have Gantt bars (1) where the start position of the bars is determined by the table calculation Gantt Percent (2) and the thickness (or size) of the bars is determined by the table calculation Percentage (3).

Indeed, if you drag the Percentage pill off the Size shelf you’ll see more clearly how Gantt Percent dictates the position of the bars.

So, all we need to do is figure out how the two table calculations work.

To do this, let’s look at the third tab in the workbook where we present a cross tab view of the data with several intermediary calculations.  Both Tableau (and Joe) recommend building this type of view when first building table calculations.

Let’s look at all the Measure Values.

Number of Records

This SUM() function gives us the number of responses for each answer category (e.g.,  for the first question 9 people responded “Poor”, 17 responded “Fair”, 73 responded “Average”, and so on.)  We could also have used CNTD(ID) to determine these values.

Count Negative

We had to address this in the bar graph approach we used in the previous blog post and we’ll run into the same problem here as we need a way to signify that all the “Poor” and “Fair” responses should go to the left of the zero, along with half of the “Average” responses.  Our formula for determining this is

IF [Score]<3 THEN 1
 ELSEIF [Score]=3 THEN .5
 ELSE 0 END

Total Count Negative

Let’s look at the table calculation for this formula.

This tells Tableau to take the TOTAL of the SUM of all the Count Negative values addressing along the Answer field.

Note that in this case addressing across Answer is the same as Pane Down; however, it is much safer to use Answer as you can never be sure how the table construct may be changed as you explore different visualizations.

Total Count

This gives us the total number of responses for each question.

Note – from here on all Table Calculations in this example are computed along Answer.

Gantt Start

We will use this calculation to determine the left / right offset for the block of Gantt bars.  That is, for each question we have a bunch of bars that are stacked together and spread out horizontally.  This calculation will determine how far to the left or right of the center (0) the stack should start.

The formula we use is

-[Total Count Negative]/[Total Count]

Very simply, this is the percentage of responses that are negative.  Questions where most of the responses are positive (e.g., “Excellent:, Good”, etc.) will be close to the center (0); questions where most of the responses are negative will start much further to the left.

Now that we know where the stack of bars will start, we need to know where to position each of the Gantt bars (the bar for “Poor”, “Fair” etc.).  The Percentage table calculation will help is figure this out.

Percentage

This formula is very straightforward

SUM([Number of Records])/[Total Count]

Remember, we are addressing across Answer so this tells Tableau to add up the responses for a particular Answer and divide by the total number of responses for all possible answers; that is,

For Poor

9 / 417 = 2.2%

For Fair

17 / 417 = 4.1%

etc.

Incidentally, we will use this table calculation to determine the size (thickness) of the Gantt bars once we’ve determined exactly where to place them.  We determine the exact placement using Gantt Percent.

Gantt Percent

Here’s the formula

PREVIOUS_VALUE([Gantt Start])+ZN(LOOKUP([Percentage],-1))

This tells Tableau to look grab the previous Gantt Percent value (if there is no previous, for the first record in the partition, use Gantt Start value as the previous) then “lookup” the previous row’s Percentage value and add that to what you have.  If there is no “previous row” the ZN() function converts the NULL value to a zero.

Let’s see how this works with an example.

For the first row there is no previous Gantt Percent value, so we start with -15.0% (Gantt Start value).  We then “lookup” the previous Percentage value (which is null) so we get a zero.  This yields a result of -15% for Gantt Percent for the first record.

For the next row we use the previous Gantt Percent value (-15.0%) and then look at the previous row’s Percentage (2.2%). When we add these together we get -12.8%.

For the middle row (what would be the position for the “Average” bar) we start with the previous Gantt Percent value (-12.8%) and add the previous row’s percentage value (4.1%) yielding -8.8%.

Conclusion

Well, I hope this really is the conclusion as this is my fourth blog post about Likert scale visualizations.

I do like the staggered (or divergent) bar approach and will use it often.  I also prefer the Joe’s Gantt Bar method we explore here to the stacked bar approach I wrote about in http://www.datarevelations.com/likert-scale-nirvana.html.

But if I were to introduce other dimensions (e.g., region, age groups, time, etc.) I would probably just use a Likert Score bar and show the percentage breakdown using a tool tip like the one we discussed in http://www.datarevelations.com/a-little-more-on-likert-scale-questions.html.

 Posted by on October 12, 2011 2) Visualizing Survey Data, Blog 56 Responses »
Oct 062011
 

Overview

I’ve written a lot about this subject (see http://www.datarevelations.com/the-likert-question-question.html and http://www.datarevelations.com/a-little-more-on-likert-scale-questions.html) but some recent discussions with friends / colleagues Joe Mako and Naomi Robbins (along with some long plane rides to and from the Midwest) encouraged me to  see how difficult it would be to create a “staggered” Likert scale chart like the one shown below.

Actually, I came up with this approach several months ago but there were some “uglies” that I needed to work around.  Let’s look under the hood and see how this works and how to hide the “uglies”.

I’ve already addressed how to display the Likert score circles (which I love now more than ever) in a previous post so we’ll just focus on how to get the colored bars with emphasis on how to get the positive attributes to the right of the zero, the negative attributes to the left, and how to split the neutrals.

Measure Names

Formula for Dark Blue and Dark Orange Bars

Here’s the underlying formula for determining the number of responses that were Excellent.

This is pretty straightforward; just take the number of records where the score was 5 (Excellent) and divide by the total number of responses.

Here’s the formula for determining the number of responses that were Poor.

Same idea as before, but this time we add up all the “Poor” responses and make them negative.

Splitting the Neutrals

I’ve always preferred even-numbered Likert scale questions as you force people to take a stand as respondents cannot take the exact middle ground.

Not so with odd-numbered Likert scale questions, so we need to take all the neutrals and make half of them positive and half of them negative, as we do here:

It’s the same concept as with the Excellent / Poor responses but we divide by twice as many total responses to split the difference.

Dealing with the “Uglies”

There are at least three things that would prevent me from publishing this in its current form.

1)      Too many items in the legend.

(The “neutrals” are broken up into two groups.)

2)      Weird ordering of the legend.  To get the bars to stack properly I had to order the items in the legend as shown above.  Ugly.

3)      Color highlighting does not display amounts the way I would like.

This is the results of clicking S_Percent_Average_Neg in the color legend.  I only see half the neutrals and I see negative numbers.

The Hack (I mean “Solution”)

So, how can I display the values for each color bar and have a legend that does what I need it to do?

I “cheated” with the legend and am just using a static graphic, specifically, a screen shot of how I would like the legend to appear.  Go ahead and wrap your knuckles in it – It’s hollow.

As for allowing people to see the underlying values, I just applied “Cotgreavean” tool tip bars, as shown here.

So, I do wish I could just select all the neutrals and see their values, but I’m satisfied with this approach and like that the offsets help convey the positive / negative sentiment.

Please download and if you come up with a nicer way to handle the legend, I’d love to see it.

 Posted by on October 6, 2011 2) Visualizing Survey Data, Blog 11 Responses »
Sep 212011
 

What’s Going to Happen when the Clinics Close?

I’ve had a lot of conversations this past week regarding my data visualization tracking STDs, HIV, and AIDS in Texas from 2006 to 2010 (see http://bit.ly/r57qeR.)

One of the bigger questions that may have been overlooked in my accompanying blog post is what is going to happen over the next two years as state budget cuts take effect? Consider the chart below that shows the increase in reported STD cases over the past five years.

Granted, combining Chlamydia, Gonorrhea, Syphilis, and HIV into one lump and not taking overall population growth into account may be misleading, so consider the chart below that shows the number of STDs and HIV cases, incident rates, and percentage changes from 2006 to 2010.

Even without drilling very deep it looks like Texas is having a lot of trouble just containing the spread of STDs and HIV.  To be fair, one of the reasons the number of Chlamydia cases has increased so much is that the Texas department of State Health Services has adopted new technology and screening techniques which in turn has led to much better detection and reporting.

That said, we pretty much see all the trend lines moving upward.

So why should we be very worried?

The numbers shown above reflect a state healthcare system that had previously budgeted over $50 million per year for family planning initiatives, including education, screening, and treatment for STDs and HIV / AIDS.  These funds have been severely cut (see http://www.texastribune.org/texas-legislature/82nd-legislative-session/day-15/.)

So, what’s going to happen when clinics close and those clinics that do remain open have to cut staff and shorten office hours?

(To explore the interactive visualizations, click here.)

 Posted by on September 21, 2011 4) Health and Social Issues, Blog 1 Response »
Sep 202011
 

You all know I love Tableau, yes? We’re cool with that…

… but…

There are times when I see Tableau guiding people to make less than ideal visualizations; situations where your have to grapple with the tool to get better results.  A good example may be found at the Tableau Public blog.  Here’s a screen capture.

There’s some great information here, but a lot of the impact is lost for two reasons:

1) Alaska and Hawaii

I’ve been griping about this since the release of Tableau 4.0, but this viz underscores the problem of always being “longitudinally correct”:  There’s no room for the information! (Damn that Pacific Ocean).  Alex Kerwin offers a solution to this here but it’s not built into the product and it requires some futzing (but it is very cool).

2) Default divergent colors

I’ve already railed about this in an earlier blog post but if you need a reminder, here’s a simulation of what this type of color combination looks like to a person who is color blind:

The solution?  Make Blue / Orange the default divergent palette.

And while I’m at it…

I really need dashboard-specific filters, but I’ve already begged/complained about that.

Maybe in Tableau 7?

 Posted by on September 20, 2011 1) General Discussions, Blog 6 Responses »
Sep 162011
 

Overview

The Texas State Legislature’s recent and sweeping funding cuts to all family planning organizations – including the complete defunding of Planned Parenthood – has lead Data Revelations to examine historical data on sexually-transmitted diseases (STDs) and HIV / AIDS and suggest what the case count and incidence rates will look like in the near future.

Note: The Center for Disease Control (CDC) divides the disease we studied into two groupings: STDs (which include chlamydia, gonorrhea, and syphilis) and HIV / AIDS.  While much of the following analysis that follows looks at overall case count and incidence rates, the interactive dashboards allow exploration by both disease category and individual diseases.

Data source: Texas Department of Health Services

The interactive dashboards may be found at the end of this blog post.

Special thanks to Joe Mako for building the county polygons and providing invaluable advice.

Key Findings

  • Roughly ten percent of Texas’ 254 counties account for 80% of all cases.
  • Within these counties, the incidence rate for STDs is up 28% from 2006.
  • Incidence rate for HIV is up 5%, but for AIDS it is down 31%.
  • The two counties that can boast the largest rate decrease for all diseases tracked in the study are Hays (-13.9%) and Travis (-7.4%).
  • The two counties with the greatest incidence rate increase for all diseases tracked in the study are Jefferson (+122%) and El Paso (+65.5%).
  • There appears to be a strong correlation between the existence of Planned Parenthood locations and decreased incidence rates (though not in all locations).
  • We believe that the recent cuts in family planning funding will lead to a large increase in cases in 2012.

Understanding the Landscape

The image below shows the incidence rate (number of cases per 100,000 persons) for all diseases broken down by county.

Note that if you hover over a county you can see information about that county.

If you click a county in the top view the table at the bottom of the screen will show results for just that county:

Where to start?

Let’s change the view and focus first on the counties that have the largest number of cases.  We can do this easily by coloring the map by Cases instead of Rate, as shown below.

Now we can see that a small number of counties are responsible for a large number of cases.  Which counties should we focus on first?

Vilfredo Pareto and the 80-20 Rule

The Pareto Principle, or 80-20 rule, is named after Italian Economist Vilfredo Pareto who early in the 20th century observed that 80% of the land in Italy was owned by 20% of the population.

In Texas it’s the 80-10 Rule

The visualization below shows that when it comes to the number of cases, just over 10% of the counties are responsible for 80% of the cases.

Note that if we highlight just the first ten percent of the bars in the top visualization we will limit the number of counties displayed in the bar graph to the 28 that account for 80% of the cases.

Specifically, selecting ten percent of the bars (1), reduces the number of counties from 254 to 28 (2) and reduces the overall case count from 697,456 to 588,416 (3).

So, now that we know what counties to focus on, what can we learn about them?

Cases, Rate, and Percent Rate Change

The next dashboard offers several ways to see how counties have performed from 2006 through 2010 as we can look at Cases, Rates, Percent Change, individual diseases, and so on.

We found “% Rate of Change” the most enlightening view so we’ll focus on that.

Let’s see what happens if we just focus on STDs and exclude HIV / AIDS from the mix.

So, what’s up with Chlamydia and Syphilis?

If you look at the individual counties using the visualization in the top portion of the dashboard you will see that without exception the percent change in Chlamydia rates from 2006 is up in all 28 counties while Syphilis is down in many counties (and way up in others.)

When asked about these numbers, an epidemiologist at the Texas Department of Health Services stated that the increase in Chlamydia rates should be attributed to an increase in better testing technologies, expansion of electronic lab reporting, and increased screening.  Chlamydia is often asymptomatic and that five years ago many cases were undetected or unreported.  As both detection and attention to reporting have improved one should see a larger number of cases.  That said, some counties are much worse in this respect than others.

As for Syphilis and Jefferson County’s “off the chart” numbers for 2009 … we’ll look at this in a moment.

Looking at the Trend for All STDs

Another way to view the data is to combine a disease group into one line by selecting Show overall from the drop down list box.

In the screen below we track overall % rate change for STDs from 2006 through 2010. It’s very easy to see where the outliers are.

Now that we have some tools to determine where counties are succeeding and failing, let’s see if we can determine why this is happening.

Location, Location, and…

The dashboard below compares the percent rate change among the 28 counties.  The size of the circles indicate the number of cases and the color indicates whether the incidence rate has increased (orange) or decreased (blue) from 2006.

We’re particularly interested in the dark orange and dark blue dots (the outliers), so let’s see what happens when we click the dark orange dot that borders Louisiana.

So, what happened in Jefferson between 2006 and 2007 to cause the initial spike, and eventual peak in 2009?  We believe it has to do with clinic locations as well as a once-in-a-generation environmental event which we will explore in a moment.  (We still don’t know why, but at least we know when the problem started.)

If we clear the selection and look at the larger, albeit not as dark, dot all the way to the west we see the following:

What happened in El Paso between 2009 and 2010 that caused such a large increase?  Again, we’ll explore this in a moment but we believe clinic location and administration has a lot to do with this.

Enough of the orange dots; let’s look at the other end of the spectrum and explore Travis County where we see an impressive decrease for all STDs.

So, we have a better sense of when problems (or improvements) occur, but we still don’t know why they occur

One additional set of data that might help us figure out why some counties are succeeding while others are failing is to look at the location of family planning clinics; i.e., clinics that provide screening, counseling, education, and treatment for STDs and HIV/AIDS.

All clinic locations

The diamond shape indicates locations of family planning centers as of August 2011.  The size of the dot indicates the number of centers within a county.  You can hover over a dot to see more information about a county and the number of clinics.

At this point there doesn’t appear to be an obvious correlation between location and number of current clinics and the percentage rate change.  Let’s see what happens if we only look at clinics run by Planned Parenthood.

Planned Parenthood locations

Three questions come to mind upon seeing this visualization:

  • Why are there no Planned Parenthood locations in El Paso (1)?  Is this related to the spike in cases from 2009 to 2010 that we saw earlier?
  • Why are there no Planned Parenthood locations in Jefferson (2)? And as we saw earlier, why the large spike in cases, particularly Syphilis, between 2006 and 2007?
  • Why is Potter County (3) succeeding where almost all the other counties are struggling?
  • Why is it that some clinics appear to be succeeding while others are failing?  Are there other issues besides location?

El Paso

Dr. Fran Hagerty, CEO of the Woman’s Health & Family Planning Association of Texas, states that there was in fact a Planned Parenthood office in El Paso but it was forced to close at the end of 2008 and the new entity that took over for it went through a fair amount of turmoil.  The epidemiology office at Texas Department of Health Services agrees that things were indeed in flux at that time.  We see that as of 2010 there are still problems.

Jefferson

While I have not yet found out why there is no Planned Parenthood office in Jefferson (or if there ever was one) there’s one thing that might explain the spike in cases (particularly syphilis):

Katrina

At the end of 2005 and through much of 2006 there was a wave of what can best be described as refugees from Hurricane Katrina that settled in Jefferson County. Many were poor and without jobs and adequate housing.  Incidence rates peaked in 2009 but the significant decrease in case count in 2010 suggests that the Texas Department of Health Services has now controlled the epidemic.

Potter

There’s no Planned Parenthood location here, but Potter County can boast only a very small increase in STDs from 2006:

What is this county doing differently from the others?

A call to Dr. Ron Barwick, CEO of the Haven Health Clinics in Amarillo indicates that the clinic he runs had been affiliated with Planned Parenthood but broke off from them in 2006 as the then-named Texas Panhandle Family Planning & Health Center was no longer going to offer abortion services.

We asked Barwick why he believed his clinic was doing so much better than most others and he stated that the main reason was that, in addition to emphasis on education, Haven had opened a male health center and this was having a significant impact on reducing STDs.  Specifically, men that would be uncomfortable sitting in a waiting room where there were women were not uncomfortable going to the male-only clinic.

Clearly, Dr. Barwick and his colleagues are doing something right.  Unfortunately, the recent budget cuts will make it difficult for them to continue, let alone share their best practices with others.

Is it All About Location?

We started to perform an analysis where we looked at the number of cases vs. the number of clinic locations and realized that we would be missing a critical data point: How well-funded and well-staffed is the clinic?

We would want to explore the relationship between clinic headcount and funding before being able to state whether a particular location is succeeding or failing.  For example, Harris counry has the greatest number of cases and certainly has many clinic locations.  Are these clinics staffed by one person or dozens?

The same question would apply to Hidalgo where there appears to be a lot of clinics given the number of cases.  How many people staff these clinics?  What is the funding for each one?

Clearly, just mapping clinic location to the number of cases is not enough.

What Now?

In addition to exploring the relationship between cases, funding, and staffing, one obvious next step would be to look at just what it is that high-performing clinics do differently from the lower  performers and have the lower performers adopt the practices of the high performers.

Unfortunately, the unprecedented state budget cuts will prevent this from happening as many clinics – including those that are getting good results – will either have to close or severely curtail their operations.

To get an idea of the impact of the budget cuts we can look back to 2006 when the Texas legislature enacted far less sweeping cuts.  Indeed, had we used 2005 as our baseline year rather than 2006 when the cuts were enacted we would see that the rate of STDs is up 39% from the baseline year vs. 28%.  But those budget cuts pale in comparison to Texas’ complete defunding of Planned Parenthood and severe defunding of other family planning entities.

We believe that a view of these dashboards in 2012 will show more, larger orange dots, indicating a much larger number of cases which will lead to an over-taxed health system, lost productivity, and increased human suffering.

 Posted by on September 16, 2011 4) Health and Social Issues, Blog 5 Responses »
Aug 222011
 

Some thoughts on Size, Color, Usability, and Engagement

I am consumed by doing what I can to make sure that my Tableau visualizations (both public and for clients) are, well, consumed.

While Tableau excels at helping people explore data and turn information into insights, it is still all too easy to create dashboards that neither please the eye nor enlighten the mind.  In fact, too many Tableau Public visualizations I’ve seen are both ugly and confusing.

I have a lot of experience in this area having created many dashboards I thought were brilliant but left people flummoxed and uninspired. In this blog post I will share some of the things I’ve learned along the way to attract viewers and engage them in my stories.

Note: Andy Cotgreave recently addressed some of these same issues (see his blog post) and will explore how to create exemplary Tableau Public dashboards at the upcoming Tableau Customer Conference in Las Vegas in October.

Overview

There are many things one must take into account when designing good dashboards (colors, fonts, layout, etc.) but for this discussion I’m going to focus on the following elements.

  • Size
  • Color
  • Usability
  • Engagement

Size Matters

Actually, it’s the width that matters and in a moment we’ll see why you will tempt fate if you create a dashboard that is wider than 650 pixels.

But first, let’s look at just why you have to worry about this in the first place.

Every Tableau Public visualization has a Share button like the one shown below.

Tableau Public's Share button

This button allows people that view your brilliant dashboards to not just create links to your work but to actually embed your work within their web sites, blogs, etc.

This is great but it means you don’t have control over how your work is displayed.  For example, you may find that your dashboard that looks terrific with a width of 900 pixels looks horrible when crammed into a blog post.

Consider the case of the recent winner in Tableau’s Sports visualization contest.  The dashboard looks great when seen as the author intended, but here’s how it looks in the Tableau blog post:

Viz designed for a wide screen and not a narrow blog post

As Bill the Cat might say, “Ack!”

Whenever I publish interactive visualizations I hope that other bloggers pick up my viz feeds and embed my work in their blogs and web pages, but I’ve had enough of my work mangled at this point that I try very hard not to create dashboards that are wider than 650 pixels (As a general rule I shoot for a width of 625 pixels).

While you might think this is a tough constraint it turns out that being forced to work with a smaller canvas often results in dashboards that are easier for people to figure out how to use than larger dashboards.  We’ll discuss this in the section on Usability below, but first let’s look at what color combinations to avoid at all costs.

Never Use Red and Green as Contrasting Colors

Once you visit this web site I guarantee you will never use red and green as contrasting / diverging colors.

This web site simulates color blindness.  I’ll focus on red-green colorblindness as 7% of American men either cannot distinguish red from green or they see red and green differently from most people (Note that only 0.4% of women are so affected.  See http://www.hhmi.org/senses/b130.html for more information.)

So, what’s the big deal? Here is a snippet of a red-green heat map from a popular Tableau Public visualization.

Heat map using Tableau's popular Christmas Poinsettia color palette

And here’s a simulation of what the image looks like to a person with red-green color blindness.

Same image showing how it might look to somebody with red-green colorblindness

So, what should you do? I really like Tableau’s colorblind-friendly orange-blue diverging pallet. It looks good and is readable by virtually everyone.

Same image using blue-orange diverging palette

You are now officially warned – anyone using red-green contrasting colors can expect a serious whupping from the viz police.

Usability

Here we’ll address the ease with which people both understand the story you are trying to tell as well as how easily people can figure out how to manipulate the filters, actions, and tabs in your dashboards.

Before you publish anything I strongly encourage you to find a friend or colleague who is not as enamored of your data / visualization as you are.  Remember, at this point you have probably fallen in love with your data and attendant visualizations, so it important (albeit, sobering) to show your work to others to see if they “get” it.

If you are like me, you will probably go through the following three stages upon observing the reaction to your work.

Anger and Disgust – This may be accompanied by thoughts of “HOW CAN YOU NOT SEE HOW THIS WORKS?  ARE YOU AN IDIOT?  THIS IS SO SIMPLE AND CLEAR!  ALL YOU HAVE TO DO IS THIS, THIS, AND THAT!”

Depression and Torpor – These are the feelings you may have once you realize that your friend / colleague is not in fact an idiot and that many people may not see the beauty and utility of your work.  Get over it!

Assimilate and Improve – It turns out that you probably don’t have to ditch all your work as very often there are small, easy things you can do to help people “get” it.  So, if folks don’t see that there are multiple tabs, filters, action controls, etc., there are things that you can do (besides yelling and gesticulating) to help them “get” your dashboard.  Here are some of the things that I’ve tried that work.

Hover Help

Screen real estate is at a premium, especially if you adhere to my recommendation of keeping your dashboards narrow.

So, how can you display useful instructions without crowding your dashboard?

Create a tool tip that contains your help / instructions.  Consider the screen below.

Is the UI truly intuitive or do you at least need some on-screen instructions?

I think the UI for this is quite friendly but when I sent early cuts of this out to people to try, very few knew instinctively what to do, so I added a little help screen that appears when you hover over the dot.

Adding Hover Help to the dashboard

It’s extremely simple to create this type of “Hover Help” tool tip.

1)      Create a calculated field called “Instructions” and define this field as follows.

2)      Place this field on the Rows shelf and change the chart type to Shape, as shown here.

3)      Create a tool tip that contains your instructions.

4)      Add this worksheet to your dashboard but hide the title so you just see the instructions and the circle.

Navigation – Dealing with Multiple Tabs

Many people not familiar with Tableau visualizations are going to miss the tabs at the top of your workbook.

You can help them discover your workbook’s other views either by pointing out the tabs with some on-screen instructions (or Hover Help) or by adding navigation elements to your dashboards.

Here’s an example from a workbook I prepared earlier this year.

Make it as easy as possible for people to figure out what they should do, and what they should do first

While the tabs at the top may get lost, my audience didn’t.

Here’s another example of embedding easily-discoverable navigation on a dashboard.

Embedded navigation controls

This technique uses the same calculated field approach we saw in Hover Help in that each navigation element is a separate worksheet in the workbook.  The difference is that on the dashboard we define an action so that when a user clicks a mark a different tab in the workbook gets activated.

Note: This feature did not work with versions of Tableau Public prior to the July 2011 release.

Wording

The way something is worded in a filter or legend can either clarify or obfuscate.  Consider the example below that shows an early cut of the filters I created for the Batting category in the Personal Baseball Entertainment Index.

Filters with ambiguous wording

Here we ask people to indicate the importance of certain offensive categories.  If you like teams that score lots of runs you should move that slider to the right.  If you don’t care that a team walks a lot you should move that slider to the left.

There was a lot of confusion regarding Strikeouts.  Some people correctly gleaned that pushing the slider to the right meant that you wanted to find teams that didn’t strike out a lot.  Others moved the slider to the left thinking that this meant find teams with few strikeouts.

This ambiguity was corrected simply by changing the wording on the filter.

Size and Complexity

I love intricate dashboards with multiple visualizations that have intelligently-defined actions so clicking on one chart affects results in another chart.

But folks not steeped in Tableau dashboard usage may be confused by all the different charts competing for attention.  In addition, without prior instruction very few people will understand that clicking on one chart can impact results in another.

This is why I like the challenge of having a smaller canvas.  If forces me to consider the less sophisticated audience and to craft dashboards that are simple and clear.

Engagement

Tableau is much more about telling the right story than assembling and displaying shiny objects, but you do in fact need to add some degree of visual bling to get people to stop and interact with your dashboards.

I’ve started adding a splash screen to both attract viewers and set the table for what will be found within the workbook.  Here’s an example.

Tableau workbook splash screen

Likewise, I try to add fun visual elements within the meat of the workbook as well.  Here we see two graphics that spice up the dashboard without being distracting.

Dashboard with a little graphical spice added

One needs to be careful not to be too cutesy with this stuff.  Here’s an early cut of the same visualization where I went overboard with the baseball motif.

Going overboard with cutesy graphics

This USA Today approach did nothing to help tell the story and in fact made it harder to distinguish between leaders and laggards.

Summary

Before publishing your next opus using Tableau Public please consider incorporating the following recommendations:

Size – Keep the dashboard width under 650 pixels.

Color – Do not use red-green for contrasting / diverging colors and check out your viz using a color-blind simulator (I use this one).

Usability – Have a friend (or friends) try your dashboard and observe them as they try to figure things out.

Engagement – Look at what you can do to draw viewers to your work without distracting people from the story you’re trying to tell.

Jul 292011
 

Overview

Perhaps it’s a Sunday afternoon and the baseball team you root for isn’t playing because the game is rained out.  You’ve got a satellite season ticket and can watch any game you want.

Or maybe you’re fed up with the team that has broken your heart for one too many seasons and want to find a new team that will at the very least entertain you.

Note: The fully working dashboards described here may be found at the end of this blog post.

Which team should you watch?

The Personal Baseball Entertainment Index (PBEI) looks at statistics for major league teams from 2009 through the All-Star Game break in 2011 and ranks each team using 16 different metrics.  You can specify which of these metrics are more important to you than others (e.g., “Triples” vs. “Walks”) and then the system will calculate a customized ranking based on your preferences.

What the Index Tracks

There are literally hundreds of different facets of baseball that people track with mind-numbing precision and affection.  Don’t believe me? Check out www.fangraphs.com.  For PBEI we look at the following 16:

Batting:

Home Runs
Triples
Stolen Bases
Runs
Hits
Walks
Strikeouts (actually, NOT striking out)

Pitching:

Strikeouts
Not Allowing Hits
Earned Run Average (ERA)
Not Allowing Home Runs
Not Allowing Walks
Time Between Pitches

Fielding:

Errors
Preventing Stolen Bases
Turning Double Plays

Ranking Each Statistic and Each Category

There are three major categories (Batting, Pitching, and Fielding) and within each category there are three to seven different metrics.

Each team has a ranking of between 1 and 30 inclusive for each metric.  For example, over the past two and one-half seasons the Seattle Mariners are ranked 26th for Home Runs while the Minnesota Twins are ranked 4th for Hits.

If you weigh each of the metrics equally, the top teams in each category would be…

Batting:

Pitching:

Fielding:

And if you weigh all three categories equally and combine the results into a master ranking…

Your Personal Index:

I’m a lifelong Mets fan.  Do you have any idea how painful it is to see The Phillies and Yankees at the top of this list?

Fortunately, all these metrics do NOT have to be weighed equally.  Indeed, I don’t think anyone would argue that a walk is as entertaining as a triple, so let’s explore how you can apply your own weights to the rankings.

Personalized Batting Index

If you click the Step 1 of the interactive dashboard you will see a screen that looks like this:

Notice that all the slider controls are set in the center.

Move the sliders to the right for measures that are important to you and to the left for those that are less important.  If you want to exclude a measure completely, just move the slider to 0 (all the way to the left).

In the example below we rank Triples and Home Runs as most important and ignore Walks.

Note: If you hover over a bar for a particular team you can see the underlying measures and rankings:

Personalized Pitching and Fielding Indices

The Step 2 and Step 3 tabs allow you to specify your Pitching and Fielding preferences using similar slider controls we saw in the Step 1 tab.

I do want to draw your attention to one statistic that is a bit different from the others and that is Time Between Pitches.

Nothing sucks the joy out of watching a baseball game quite as much as a pitcher that takes a long time between pitches so I’ve included this somewhat atypical metric in the Pitching index.  Moving this slider to the right will move teams that have pitchers that work fast up the list.

Note: Special thanks to Lucas Apostoleris for aggregating this metric by team for the 2010 season (I do not have the statistics for 2009 and 2011 yet.)

Putting it all Together – Your Personal Index

Once you’ve specified the relative importance of each of the 16 metrics you should click the Step 4 tab.

Here you specify the relative importance of the three main categories.  In the sample below the settings indicate that Pitching is most important, followed closing by Fielding.

Raw Data and Rankings

The final tab allows you to look at the raw statistics and rankings for each team in the three categories.

You can look at a different category by clicking the Category drop down filter.

A Quick Peek Under The Hood

I plan to write a separate blog post that explains some of the inner workings of the underlying Tableau workbook as well as some of the design decisions.  In the meantime, here’s a quick look at the logic behind the indices.

For each category I wanted to have a maximum possible score of 100.  That is, if a team ranked first in every metric and the user indicated that every one of those metrics should get a “10” then the top score would be 100.

Here’s the algorithm for Fielding.

((31-[Strikeouts Rank])* User Weight /3
+ (31-[Avg Time Between Pitches Rank])* User Weight /3
+ (31-[ERA Rank]) * User Weight /3
+ (31-[HR (Pitching) Rank])* User Weight /3
+ (31-[Hits (Pitching) Rank])* User Weight /3
+ (31-[Walk (Pitching) Rank])* User Weight /3)
/(6-[Number_Pitching_Ignores])

Let’s just look at the first piece.

((31-[Strikeouts (Pitching) Rank])* User Weight Metric /3

There are 30 teams so the top ranked team would 1.  31 minus 1 = 30.

If the user indicates that this measure is of highest importance, the 30 will be multiplied by 10, giving us 300.  Dividing by 3 gives us 100.

We repeat this for the other five metrics and get a total possible high score of 600.

Finally, we divide everything by 6 (the number of metrics in this category) and get a top possible score of 100.

(Actually, we divide by 6 minus the number of filters that may have been set to zero indicating that the user doesn’t care about the metric.)

The Batting and Fielding indices are computed in similar fashion.  The PBEI (the custom master index) is computed as follows.

(([BattingIndex]* User Weight)+
([PitchingIndex]* User Weight
([FieldingIndex]* User Weight))/(30-(10*[Number_Composite_Ignores]))

Remember that the top possible Batting Index is 100, so if the user indicates this should be weighted as most important (a 10) this would give us 100 x 10, or 1,000.

Do this three times and you get a maximum score of 3,000 so we divide by 30 (okay, 30 minus 10 times the number of ignores) to get a maximum custom index score of 100.

What Team Should You Watch?
The Personal Baseball Entertainment Index

Perhaps it’s a Sunday afternoon and the baseball team you root for is rained out. You’ve got a satellite season ticket and can watch any game you want.

Or maybe you’re fed up with the team that has broken your heart for one too many seasons and want to find a new team that will at the very least entertain you.

Note: The fully working dashboards described here may be found at the end of this blog post.

Which team should you watch?

The Personal Baseball Entertainment Index (PBEI) looks at statistics for major league teams from 2009 through the All-Star Game break in 2011 and ranks each team on 16 different metrics. You can specify which of these metrics are more important to you than others (e.g., “Triples” vs. “Walks”) and then the system will calculate a customized ranking based on your preferences.

What the Index Tracks

There are literally hundreds of different facets of baseball that people track with mind-numbing precision and affection. Don’t believe me? Check out www.fangraphs.com. For PBEI we look at the following 16:

Batting:

Home Runs
Triples
Stolen Bases
Runs
Hits
Walks
Strikeouts (actually, NOT striking out)

Pitching:

Strikeouts
Not Allowing Hits
Earned Run Average (ERA)
Not Allowing Home Runs
Not Allowing Walks
Time Between Pitches

Fielding:

Errors
Preventing Stolen Bases
Turning Double Plays

Ranking Each Statistic and Each Category

There are three major categories (Batting, Pitching, and Fielding) and within each category there are three to seven different metrics.

Each team has a ranking of between 1 and 30 inclusive for each metric. For example, over the past two and one-half seasons the Seattle Mariners are ranked 26th for Home Runs while the Minnesota Twins are ranked 4th for Hits.

If you weigh each of the metrics equally, the top teams in each category would be…

Batting:

Description: C:\Users\STEVEM~1\AppData\Local\Temp\SNAGHTMLd1174b.PNG

Pitching:

Description: C:\Users\STEVEM~1\AppData\Local\Temp\SNAGHTMLd3a43e.PNG

Fielding:

Description: C:\Users\STEVEM~1\AppData\Local\Temp\SNAGHTMLd776e8.PNG

And if you weigh all three categories equally and combine the results into a master ranking…

Your Personal Index

Description: C:\Users\STEVEM~1\AppData\Local\Temp\SNAGHTMLd71162.PNG

I’m a lifelong Mets fan. Do you have any idea how painful it is to see The Phillies and Yankees at the top of this list?

Fortunately, all these metrics do NOT have to be weighed equally. Indeed, I don’t think anyone would argue that a walk is as entertaining as a triple, so let’s explore how you can apply your own weights to the rankings.

Personalized Batting Index

If you click the Step 1 of the interactive dashboard you will see a screen that looks like this:

Description: C:\Users\STEVEM~1\AppData\Local\Temp\SNAGHTMLddbbd3.PNG

Notice that all the slider controls are set in the center.

Move the sliders to the right for measures that are important to you and to the left for those that are less important. If you want to exclude a measure completely, just move the slider to 0 (all the way to the left).

In the example below we’ve ranked Triples and Home Runs as most important and ignore Walks.

Description: C:\Users\STEVEM~1\AppData\Local\Temp\SNAGHTMLe68653.PNG

Note: If you hover over a bar for a particular team you can see the underlying measures and rankings:

Description: C:\Users\STEVEM~1\AppData\Local\Temp\SNAGHTMLf6fc3d.PNG

Personalized Pitching and Fielding Indices

The Step 2 and Step 3 tabs allow you to specify your Pitching and Fielding preferences using similar slider controls we saw in the Step 1 tab.

I do want to draw your attention to one statistic that is a bit different than the others and that is Time Between Pitches.

Description: C:\Users\STEVEM~1\AppData\Local\Temp\SNAGHTMLeaea62.PNG

Nothing sucks the joy of watching a baseball game quite as much as a pitcher that takes a long time between pitches so I’ve included this somewhat atypical metric in the Pitching index. Moving this slider to the right will move teams that have pitchers that work fast up the list.

Note: Special thanks to Lucas Apostoleris for aggregating this metric by team for the 2010 season (I do not have the statistics for 2009 and 2011 yet.)

Putting it all Together – Your Personal Index

Once you’ve specified the relative importance of each of the 16 metrics you should click the Step 4 tab.

Here you specify the relative importance of the three main categories. In the sample below the settings indicate that Pitching is most important, followed closing by Fielding.

Description: C:\Users\STEVEM~1\AppData\Local\Temp\SNAGHTMLf31f09.PNG

Raw Data and Rankings

The final tab allows you to look at the raw statistics and rankings for each team in the three categories.

Description: C:\Users\STEVEM~1\AppData\Local\Temp\SNAGHTMLfd0909.PNG

You can look at a different category by clicking the Category drop down filter.

A Quick Peek Under The Hood

I plan to write a separate blog post that goes explains some of the inner workings of the underlying Tableau workbook as well as some of the design decisions. In the meantime here’s a quick look at the logic behind the indices.

For each category I wanted to have a maximum possible score of 100. That is, if a team ranked first in every metric and the user indicated that every one of those metrics should get a “10” then the top score would be 100.

Here’s the algorithm for Fielding.

((31-[Strikeouts Rank])* User Weight /3

+ (31-[Avg Time Between Pitches Rank])* User Weight /3

+ (31-[ERA Rank]) * User Weight /3

+ (31-[HR (Pitching) Rank])* User Weight /3

+ (31-[Hits (Pitching) Rank])* User Weight /3

+ (31-[Walk (Pitching) Rank])* User Weight /3)

/(6-[Number_Pitching_Ignores])

Let’s just look at the first piece.

((31-[Strikeouts (Pitching) Rank])* User Weight Metric /3

There are 30 teams so the top ranked team would 1. 31 minus 1 = 30.

If the user indicates that this measure is of highest importance, the 30 will be multiplied by 10, giving us 300. Dividing by 3 gives us 100.

We repeat this for the other five metrics and get a total possible high score of 600.

Finally, we divide everything by 6 (the number of metrics in this category) and get a top possible score of 100.

(Okay, we don’t divide by 6. We divide by 6 minus the number of filters that may have been set to zero indicating that the user doesn’t care about the metric.)

The Batting and Fielding indices are computed in similar fashion. The PBEI (the custom master index) is computed as follows.

(([BattingIndex]* User Weight)+

([PitchingIndex]* User Weight

([FieldingIndex]* User Weight))/(30-(10*[Number_Composite_Ignores]))

Remember the top possible Batting Index is 100, so if the user indicates this should be weighted as most important (a 10) this would give us 100 x 10, or 1,000.

Do this three times and you get a maximum score of 3,000 so we divide by 30 (okay, 30 minus 10 times the number of ignores) to get a maximum custom index score of 100.

Interactive Dashboards

 Posted by on July 29, 2011 6) Sports and Entertainment, Blog 8 Responses »
Jul 252011
 

Overview

Last year, UN Global Pulse launched a large-scale mobile phone-based survey that asked people from India, Iran, Mexico, Uganda, and Ukraine how they were dealing with the effects of the global economic crisis.

The survey (conducted from May-August 2010) asked two multiple choice and three open-ended questions focusing on economic perceptions.

Note: The fully working dashboards may be found at the end of this blog post.

Key Findings

Responses from Uganda – a country that ranks in the bottom 15th percentile in the UN’s Human Development index – were consistently more optimistic than responses from other countries.

What could account for this? Is it that Ugandans are, as a group, more hopeful and optimistic than people in the other countries surveyed?

Or could it be that survey responses were somehow skewed?

Let’s explore the data to find out.

Voices of Vulnerable Populations during Times of Crisis

Clicking the second tab displays the following view.

Economic Change Index

So, why in the first graphic does Uganda warrant a positive blue bar and Mexico a negative orange bar?  By moving your mouse pointer over a bar you can see just what it is that drives the Economic Change Index.

Here are the results for Uganda…

… and here are the results for Mexico:

The index itself (1.2 for Uganda and -1.6 for Mexico) is computed by applying Likert-scale values to each of the possible question responses.  We’ll discuss the advantages of using this approach in a moment.

Fixed Responses vs. Using One’s Own Words

The first two questions in the survey gave respondents four choices from which to choose.  The remaining three questions allowed people to respond in their own words.

You can explore these responses yourself by picking a question and a country from the drop down list boxes.

So, does the sentiment shown in the first fixed-response apply the open text responses as well?

Promising vs. Uncertain

Here is how people from Uganda responded to the question “In one word, how do you feel about your future?”…

… and here is a visualization of the responses from Mexico.

This, combined with responses to other questions, left me scratching my head. What are we not seeing that would lead to responses from Uganda — a country that is arguably in worse condition than the others — being so upbeat?

If you can’t wait for the answer, click here.

A Word about Word Clouds

I’ve analyzed a lot of survey data and I hate analyzing survey results where people get to provide free-form text responses because aggregating responses based on a common sentiment can be very difficult.

In many cases Word Cloud generators can convey the overall sentiment from multiple text responses.  They are also interesting to look at and I do believe the ones shown above are a good reflection of respondent sentiment.

A problem occurs, though, when respondents use different terms that describe the same or similar sentiment.  Consider the Word Cloud shown below.

One might think that most respondents were happy, but look what happens if we “linguistically normalize” the terms that are synonyms of “sad”:

It turns out that more people are in fact sad.

Note: There are products that are capable of parsing full sentences and are able to “disambiguate” and then normalize terms under umbrella concepts. The text responses to this particular survey, however, do not warrant this type of heavy artillery.

How We Calculate the Indices

The next tab in the workbook shows some alternative ways of visualizing the fixed-response survey results.

For these questions respondents were given four choices:

Easier / Better

Same

Worse / More Difficult

Much Worse/ Very Difficult

Notice that we display the calculated index atop the Likert-scale stacked bar charts.  There are three advantages to calculating an index for Likert-scale responses:

  1. It makes it easy to weigh sentiment across many responses.
  2. It makes is possible to track sentiment changes over time.
  3. It makes it possible to compare results against various objective economic indices (e.g., GDP, UN HDI, etc.).

Note: I have no problem using even-numbered Likert scales, but I do think in this case sentiments will be skewed towards the low end as there are two levels of pessimism (e.g., “worse” and “much worse”) and only one of optimism (e.g., “better”).

I attempted to combat this by applying the following values to the responses:

Easier / Better = 3

Same = 0

Worse / More Difficult = -2

Much Worse/ Very Difficult = -4

While I think these values make sense, users of this dashboard are welcome to use the sliders and apply different values to each of the answers.  The indices will be recalculated automatically.

A Composite Index

In an earlier version of this dashboard I created a “composite index” that combined results from the two fixed-response questions:

I think this is a valuable metric and one that I would include should UN Global Pulse make this study longitudinal (see below).

Mobile Pulse Survey Results vs. Objective Economic and Human Development Indicators

In the next tab we see survey responses (first column) vs. the United Nations Human Development Index Ranking (second column).

What could account for Ugandan survey respondents being the most optimistic despite the fact that they rank 143 out of 169 countries in the UN’s HDI Ranking?

I believe that the survey’s SMS Text-based approach is skewing the results.

Consider the third column where we see the number of mobile subscribers within a country as a percentage of that country’s population.  In Uganda, at most 29% of the population has a mobile phone suggesting that those completing the survey may be better-off financially than others within their country. Survey responses may not, therefore, be a reflection of the country as a whole. (See The CIA World Factbook for mobile phone subscription information.)

This would not be the first time premature reliance on phone polls has derailed a survey (or in this case, just part of a survey).  See A Couple of Interesting Examples of Bias and Statistical Sampling.

Make the Survey Longitudinal

Despite the shortcomings, I think there is a lot of value in conducting these types of agile, real-time surveys.

One ongoing challenge will be comparing subjective data among different countries as there are so many cultural / proclivity issues that are difficult to compare.

One way to do this would be to conduct a longitudinal study and see how sentiment changes over time.  That is, instead of comparing Uganda with Mexico or India with Ukraine for a given year, track the changes over time, using an index.  This would allow you to see the percent change in sentiment between time periods without having to worry about normalizing cultural differences.

I hope that UN Global Pulse will update this survey on a regular basis as there’s much we would be able to learn from such a study.

 Posted by on July 25, 2011 4) Health and Social Issues, Blog 2 Responses »