Small Multiple Area Charts vs. Line Charts

I posted this on twitter a while ago …

Enrico Bertini @FILWD

What explains that area charts are easier to read than line charts in a configuration like the one below? Or maybe it's just my impression? (cc @SteveFranconeri @sharoz)

I can’t totally recall when I came up with this idea first, but I am pretty sure it’s one of the myriad effects I noticed while grading student assignments and figuring out I did not have a good explanation for questions my students had.

Before delving into the sweet set of replies I received, I think it would be fair to share with you what I used to think before writing that tweet. My rough explanation for it went something like that:

1. Areas form a filled shape;

2. Humans eyes are much better at perceiving shape from filled areas than from lines;

3. Hence, area charts are better than line charts.

Of course, this is not even an explanation because, why is it easier to compare filled rather than empty shapes? Admitting this is even true.

Let’s see what we got from the tweets …

Squinting

The first to reply was Steve. I knew he would come up with something:

Steve Haroz 📊👁️🧠 @sharoz

@FILWD @SteveFranconeri The first thing to think about in anything related to vision is spatial frequency or scale. Squint your eyes, so the image is blurry. The area plots are easily differentiable. But the line charts disappear.

Steve Haroz 📊👁️🧠 @sharoz

@FILWD @SteveFranconeri Here are the images with a 40 pixel Gaussian blur. You can still segment the area chart, but the lines in the line chart almost disappear.

This is interesting as a general rule. I see how it can be useful. But it does not help me too much to figure out what is going on. I guess, the real point here is that better segmentation makes it easier to distinguish between charts and because of that it’s easier to perceive their patterns? I am not sure.

Frame of reference

A more convincing case is the idea that area charts have a frame of reference that does not depends on the location of the axes. In fact, the very presence of the axes can be distracting for the line charts. See this image Steve shared in the thread:

Steve Haroz 📊👁️🧠 @sharoz

@FILWD @SteveFranconeri Let's compare two data points. For position (1), no idea what the reference frame is. Different reference frames (2, 3) carry very different meanings. You have to segment the facets to subtract facet baseline position from line position. For area (4), no facet segmentation needed

If you change the frame of reference, lines can represent completely different values, whereas areas do not even need a frame of reference; it’s implicit in their design.

Nick had a similar take with this illuminating saliency analysis:

Nick Holliman @binocularity

@FILWD @SteveFranconeri @sharoz If you look at a saliency analysis the line graphs have no strong baseline - you can't see what to judge them against. Remove the bounding boxes and grid lines from both plots and put in a strong base line might work better? Also area plots suggests false zeros at the ends.

Alignment and shared axis

David had another interesting comment:

David Gotz @DavidGotz

@FILWD @SteveFranconeri @sharoz I suspect this effect would be reduced if charts were arranged horizontally to share a common baseline? Of course that would create a long thin series of charts which is not practical.

To which I replied with an image showing an example of what he was proposing:

Enrico Bertini @FILWD

@DavidGotz @SteveFranconeri @sharoz @DavidGotz What do you think? I agree that you may be onto something here. I was skeptical at first but the images look pretty good.

What I liked of this part of the conversation is that it made me explore a slightly different solution in which the charts are aligned horizontally rather than vertically. When I replied I was a bit more optimistic than I am now while I am writing. It does look like the problem for lines is mitigated with this solution. But I can’t help it. I still perceive area charts as somewhat easier to read.

Processing multiple (similar) objects

This one below from Glenn is probably my favorite explanation, in addition to the baseline effect discussed above.

glenn mcdonald @glenn_mcdonald

@FILWD @SteveFranconeri @sharoz I think it's that the area-chart version reads as N objects, where the line-chart version requires you to visually parse the lines and spaces (and baselines) separately, which is sort of 2.5*N objects.

This is brilliant and it’s the closest thing to an actual explanation. A different way to describe this is that where in area charts there is a unique object that forms a “whole”, in line charts there are many separate objects that interfere with the perception of a whole. This is actually a very common problem that I have noticed in many other visualizations.

This also reminds me of a classic effect I read first about in one of Ed Tufte’s books years ago (which he took from visual artist Josef Albers): the idea that in graphic design 1+1 = 3.

That’s the relevant section from the book:

So, maybe equally in the line charts version we see a lot more objects and it gets more difficult for our eyes to remove the “visual noise” from the actual signal we need to perceive.

Gestalt: Figure and Ground

Some others mentioned the figure-ground effect described in Gestalt Psychology. Here is one relevant tweet from Renata:

Renata Steffen @resteffen

@FILWD @SteveFranconeri @sharoz Figure & ground (gestalt principles)

I think this is probably relevant even though it’s also maybe another way to describe the same thing above. A filled area against a background with a strong difference in contrast produces grouping, which in turns leads to better segmentation and all the things we mentioned. Steve here seems to express a similar concept:

Steve Franconeri @SteveFranconeri

@FILWD @resteffen @sharoz These are all the same. Clear Figure ~= easy segmentation ~= good ref frame ~= low spatial freq friendly area encoding. Wiggles are because any of these explanations will devolve to hand waving even when pushing a perception expert. Would be delighted to be proven wrong.

Reflections

This thread helped me enormously to think about this problem at a deeper level. If anything, it demonstrated how hard it is to explain how data visualization works. I wish we had some sort of mega theory to guide us in visualization. Or at least a number of piecemeal theories that cover the gaps. I don’t think we are there yet, but we do have amazing researchers around. Unfortunately, all we have in visualization is mostly a super coarse theory of channels with their extremely limited rankings.

One thing I like about this exercise is that some of these things we discussed here can be reused as tools for thought in other situations. In particular, the idea of areas providing a natural frame of reference seems to be very powerful. Similarly, the idea that some configurations may unnecessarily lead us to perceive too many objects to parse seems also very powerful. And even squinting and saliency analysis seem like a first step to always take when dealing with visual perception.

I hope you found this useful. Add your comments below if you think you have anything to add!