What a regrettably large head you have. I would very much like to hat it
January 14, 2019 10:00 AM   Subscribe

“Visualizing data through graphs can be an effective way to communicate one’s results. A ubiquitous graph and common technique to communicate behavioral data is the bar graph. The bar graph was first invented in 1786 and little has changed in its format. Here, a replacement for the bar graph is proposed. The new format, called a hat graph [PDF], maintains some of the critical features of the bar graph such as its discrete elements, but eliminates redundancies that are problematic when the baseline is not at 0.”
posted by not_the_water (21 comments total) 19 users marked this as a favorite
 
It looks kind of a step chart with just two steps.
posted by paper chromatographologist at 10:09 AM on January 14, 2019


I could totally see using hat-plots in some of my work.

Obligatory posting of R::CatterPlots.
posted by Made of Star Stuff at 10:19 AM on January 14, 2019 [4 favorites]


Whether you like this idea or not, this is the important part:
The effectiveness of the hat graph was tested in two empirical studies for which participants had to find and identify the condition that lead to the biggest improvement from baseline to final test. Performance with hat graphs was 30 - 40% faster than with bar graphs.
posted by straight at 10:22 AM on January 14, 2019 [7 favorites]


Also, this fits with some of the low-ink chart design principles one of my committee co-chairs pushes. Minimalism in plot design should help people interpret things more quickly, as long as there are enough elements to communicate every important piece of information. Which the hat-plot has, since it's stripped down to just the relative differences. I also like this recommendation:

graph design indicates that for behavioral studies for which effect size is measured in terms of standard deviations, the range of the y-axis should be 1.5 standard deviations (Witt, in press). Setting the y-axis in this way helps to maximize compatibility between the visual impression of the effect and the size of the effect. Small effects will look small and big effects will look big. In cases for which the size of the effect is too big to fit within 1.5 SDs (as was the case in the current studies), the y-axis should be expanded and its range noted in the figure caption.

mutable y-axes are a pox on this land, and graphing practices that make effect sizes intuitive are good practices.

Yeah, I'm good with this.

straight, sure, it works in practice, but how's the theory?
posted by Made of Star Stuff at 10:28 AM on January 14, 2019 [6 favorites]


i don't see the need for a brim; just having this floating bar tells you the scale, the difference, and - with an x-axis - change over time. It's pretty effective, and far tidier than clustered columns which really need to die a fast death.
posted by entropone at 10:43 AM on January 14, 2019


> It is unclear how to expand hat graphs to allow comparison across 3 or more conditions.

Not so useful for engineered system monitoring or large data sets.

Seems to me just a 'center aligned' bargraph with a reference point. Often instead of the brim of the 'hat', the X-axis is just moved up to the reference value.
posted by anthill at 10:44 AM on January 14, 2019


It's like a parade of tiny Abe Lincolns.
posted by The Underpants Monster at 10:45 AM on January 14, 2019 [1 favorite]


I came to read about people with large heads. Maybe even a few paragraphs about Stetson hats. Disappointed.
posted by AugustWest at 11:12 AM on January 14, 2019 [2 favorites]


It's like a parade of tiny Abe Lincolns.
...or a parade of French legionnaires.
posted by elgilito at 11:16 AM on January 14, 2019 [1 favorite]


These seem like they could easily be misleading, since there's nothing to indicate whether the change-from-baseline is large or small, or whether that's important or not.

If the zero point really isn't interesting and comparing change-from-baseline is, well, why not a bar chart of changes-from-baseline?
posted by zompist at 11:20 AM on January 14, 2019 [7 favorites]


Who really wants a webpage to load instantly anyway? I really enjoyed staring at a blank, white page for several seconds.
posted by humboldt32 at 11:47 AM on January 14, 2019 [5 favorites]


Strongly seconding zompist. This is a nice visualization solution to the wrong problem. Plot the thing you want to communicate. If only change-from-baseline is important, just plot that. If both change-from-baseline and value-at-baseline are important, it seems like this hat plot de-emphasizes the latter at least as badly as traditional bar plots de-emphasize the former.

Personally, I think there are two modes of data presentation in graphs that should inform your design. (I mean, there are obviously more than two, but this is the distinction I want to emphasize.) If what you want to communicate is a simplified, summary statistic, then just plot that, using bar graphs, connected line graphs, whatever. If what you want to communicate is the full complexity of the dataset to allow others to draw their own conclusions about it that may not agree with yours (which I think is what scientists should usually do), then plot that using techniques like violin plots, jittered dot plots, etc. Basically, I think you've got two choices: communicate only a summary statistic, in which case go full Tufte and simplify as much as possible, or communicate a distribution so you're not hiding anything. This hat plot (and the paired bar graphs it attempts to improve) falls in a middle ground between these strategies that I'm not sure is all that valuable, since it obscures both the simple summary and the real-world complexity.

I'm also not sure that response time for extracting change-from-baseline is necessarily the best measure to use for graph comprehension. Communicating key information quickly is certainly valuable, but not the only valuable thing in graph design.

Anyway, it's a neat idea and a good paper, and despite my misgivings, I like that Witt looked for a way to assess the communication technique experimentally. Maybe after seeing some of these "in the wild" I'll change my mind.
posted by biogeo at 12:11 PM on January 14, 2019 [7 favorites]


I like these! The emphasis on the delta and the integration of error bars are useful features for graphs that are about comparisons.
posted by pulposus at 12:12 PM on January 14, 2019 [1 favorite]


I can't get past the fact that the vertical axes don't go to zero in their first few figures. They go to, like, 0.7 or so.

These people are monsters.
posted by gurple at 12:24 PM on January 14, 2019 [3 favorites]


i don't see the need for a brim; just having this floating bar tells you the scale, the difference, and - with an x-axis - change over time.

The brim lets you distinguish increases and decreases from baseline - without it, you wouldn't know which way the change went. Funny that all the example plots in the paper show increases from baseline, with the "baseline" moved to the left depending on what increased.

With a floating bar, I suppose you could just have an arrow?
posted by bbuda at 1:24 PM on January 14, 2019


A floating bar would look too much like a box-plot with whiskers, anyway. I like the hat plot. But I agree that there's an issue with multiple-category comparisons.

zoompist, if the y-axes are always 1.5 standard deviations in range, then not only are changes from baseline pretty clear, they're also comparable across multiple graphs, no?
posted by Made of Star Stuff at 1:55 PM on January 14, 2019


This is really clear, I like it. You could easily expand the idea to include a baseline, and two increases, although it wouldnt look like a hat anymore.
posted by subdee at 5:37 PM on January 14, 2019


These seem like they could easily be misleading, since there's nothing to indicate whether the change-from-baseline is large or small, or whether that's important or not.

That's what the y-axis should be scaled for. People might not use it correctly but that's an issue with any plot; this makes it neither better nor worse. (Actually I think it makes it slightly more likely people will make a considered choice.)

If the zero point really isn't interesting and comparing change-from-baseline is, well, why not a bar chart of changes-from-baseline?

You lose the absolute numbers in that case. Knowing that one of your control groups is had a lower or higher response is often relevant secondary information when you think about data critically. So better to include it. Besides, you still need to think about scaling even if you decide "change from baseline" is the measure.

Whether you like this idea or not, this is the important part:

Nah, that quote is only important if you like the idea. The important quote if you don't like it is the one admitting the quite limited applications to 2 x N data sets. :)

I like the idea FWIW.
posted by mark k at 9:44 PM on January 14, 2019 [1 favorite]


*cough* To my knowledge this paper is still under active peer review, FYI.

It's an approachable graph, not sure I'd use it and we'll see what the reviewers say, but I like the idea. Jessica Witt typically does research on embodied cognition, I'll be curious to see if she continues forays into data visualization.
posted by nicodine at 9:59 AM on January 15, 2019


Minimalism in plot design should help people interpret things more quickly, as long as there are enough elements to communicate every important piece of information.
However, familiarity is important too. A note is much longer than a hobo code, but after several attempts my wife has found that the note to me is much more reliable.

Similar with bar graphs vs. J. Random WhippyChart.
posted by Gilgamesh's Chauffeur at 4:27 PM on January 16, 2019


> Funny that all the example plots in the paper show increases from baseline

Right? I scrolled through the whole thing waiting to see an upside down hat and now I will spend my day disappointed.
posted by lucidium at 5:03 AM on January 17, 2019


« Older Bharat Bandh - “General Strike”   |   Representation Roundup Newer »


This thread has been archived and is closed to new comments