violin plots should not exist 

If I ever write a paper, I'm going to not use a violin plot, and I'm going to cite you for why I didn't use a violin plot.

Hello! I am the creator of the Turnip Plot, which is similar to a violin plot but rotated around the vertical axis and rendered in 3D. Please cite my paper. Thanks!

Violin plots are overused but they have a use case for comparisons of a large number of samples that have complex distributions. We use them for this when comparing gene expression in cell populations. We can quickly see the 'shape' and get the vibe of the multimodel gene expression for large sample numbers.

each time she points to the void and a violin plot appears I absolutely and completely lose my shit they get uglier each time without missing a beat oh my god

the mad scientist plotting to take over the world cackles madly as he circles “violins?” on his list of evil plots.

I really liked the comparison you showed @33:57 that compares violin plot to ridgeline plot. That drove the message home.

So, I like all your videos for different reasons, but I have to say this one is my favorite. Fantastic points. Also, I had a (sorta) similar experience at work. Our boss usually brought snacks to meetings. She brought ice cream once. I was the only overweight person on a large team of fit, young sports nuts. So what does she do upon wrapping up her PowerPoint? She says, “there’s lots of ice cream left if anyone wants extra. Stephanie?” Everyone turns to me because of COURSE the fat girl will want more ice cream, right?! I shouldn’t have to give an excuse to why I don’t want any ice cream.. but I still had to explain I’m lactose intolerant. So I hate meetings where she brought snacks. Which is a shame.

Most typewriters allowed you to apply whiteout without removing the paper. And some even had a white ribbon, to strike over incorrect letters. :D

Thank you for yet again making things so crystal clear! Agree 99.9% -- Beginner question: Why would I ever need units on a distribution? If I know that the area is always equal to 100%, then the shape and length (in this case the vertical axis), and smoothing, are the only three things that matter for a distribution? Admittedly, I do not have ANY math OR science background, but that tiny corner of the argument seems trite to me. I will absolutely never, ever use violin plots, over my dead body. I do not wish to be perceived as one of those guys, and I don't wish to feed that side of the room, ever. As a non-english, male speaker with no math background, I apologize for any language errors of mine.

I’m old enough to remember wet copying. And making stencils. You bloody think before you write. On the really old typewriter, you would backspace, find the error and scratch it out. You may or may not use a paper filler to reinforce the integity of the paper. That’s the reason old time paper was a heavier grade eg bond paper, was for precisely this reason. You’re welcome to the good old days.

I love your videos. It's like if science was a city and you're a tour guide taking us to all the places where people get stabbed.

Have you noticed that most graph paper uses a relatively faint blue colour? That is because photocopiers (before they became scanners) couldn't see blue well. If you were careful with the contrast, you could draw a diagram or a form on graph or squared paper and the photocopier would come out with only your lines on it. It was a great way to create forms, character sheets for RPG games or, indeed, the sorts of diagrams in an academic paper. Regarding the typewriter thing, you could get whiteout on paper strips. You backspaced over the error, poked the whiteout paper behind the ink-ribbon and hit the erroneous letter again. That deleted it and you could over-type again with the correction. The line spacing on typewriters is quantised so you can go up and down by exact lines easily. It's only when you noticed the error after you had taken the paper out that you would have to resort to liquid whiteout and fiddle to get the positioning correct.

i spent 27 minutes going "okay is she going to say they look like vulvas" and then felt very validated

I used to work at a tight tolerance thin film deposition optics manufacturer. One day my current supervisor visited my workstation unexpectedly and asked, "you haven't been using violin graphs in any of your report generators, have you?"; "Never heard of such a thing. Why?"; "Good. I knew you were a smart one. Just don't. They're useless anyway." ... Later she quietly explained their inappropriateness. Turns out she wasn't just trying to prevent her team from using them either. The owner got top management together and instructed them to purge any current work of their existence. HE was smart.

The technology exists for us to put 3d objects into .pdfs. For this reason, I propose the vase plot. The diameter of the vase is the probability density. or maybe the cross-sectional area is the probability density. It is intentionally ambiguous. Please cite my comment whenever you use my case plot in your paper.

Librarian here: Physical Journals/books/etc. are often preferred when it comes to preservation and access, since when we buy a physical copy, we won't lose access to it when the publisher decides that we need to pay another $400 per user to access their system, or if a ransomware attack takes out the archive. Obviously, storing all these is impractical for all but the largest library systems, but the number of times someone needs to cite an article in the 2003 summer edition of Phrenology Today means that you can get away with only one copy in an offsite, climate controlled storage space. The number of times I've had to do dumb piracy shit because a publisher literally pulled access to a non-downloadable article I was using for a paper *mid semester* was infuriating. Not to mention the time my school literally stopped paying our publisher liaison because they upped their rates by a factor of 10. We had a lot of 3rd/4th generation photocopies floating around that year, lmao

The kooky patterns make it so that colorblind people can read the histograms. I'm red-green colorblind, and I get really passionate about data visualization, partially because I have trouble with certain color visuals that other people don't struggle with, and partly because data visualization is a very efficient way to communicate info when done correctly, and I also have adhd, so I appreciate their communicative strengths. But yeah, all that to say, I really appreciate when kooky patterns are included. It's immediately tells me that the person who made that visual is thinking about how their data will be perceived and wants to communicate it effectively to as many people as possible.

Your story about everyone turning to look at you as the only woman... it gave me a flashback to school sex-ed where the teacher (to be inclusive) said "or a**l sex" and people turned to look at me which basically outed me to the teacher. But also what reaction were they expecting from me lmaoo

During my masters, my advisor actually ENCOURAGED the use of violin plots, and I didn't really question it at the time. Rude jokes were made. I have a publication that has FIVE SEPARATE PAIRS of violin plots (and plots that cluster points into hexagons? for some reason?). And you're right! Side-by-side histograms would have been BETTER and MORE COMPACT 😭😭😭

Angela utilizing her platform to indoctrinate the public against evil plots in an intellectual crusade is what I live for

Yes to the symmetry argument. I spent the entire video trying to figure out why they were twice the size they needed to be.

I'm not a scientist. Just a humble welder. I like your videos. I had no idea where you were going with this, but you had me nodding along, thinking, 'I get it. These violin plots are stupidly complicated and are not as efficient as other plots.' I get it! And you know, not being involved with the sciences, I didn't actually care about the plots one way or the other, but I could see how it would annoy an actual scientist/researcher. Then it finally got it where it was going, and I was like, 'Damn! That was masterfully done.' Keep up the great work.

As a data analyst, you've made amazing points for not using violin plots--scientifically. But in the business world, violin plots are **pretty** and the vibe is how you get execs to make the decisions in favor of what you want 😂😂😂😂

The comment at the end about “why have two flaps” is the most upsetting part of them to me. The only time I’ve seen something like this (where they used histograms instead of smoothing, so less yonic) that seemed even a little defensible is the man/woman population age plots for different counties. It’s still probably better to arrange them differently but at least they were using the two sides for something.

This video sort of gave me the urge to come up with plots that are even more cursed than the violin plot. Like a stick figure plot where different aspects of the data set are represented by the size and orientation of the body parts.

When I was in undergrad getting my degree in economics, I saw these every once in a while when conducting research for papers. Turns out I wasn't an idiot for not being able to read these plots, I was just an idiot for getting a degree in economics.

Hello! I like violin plots a lot; Im a data science researcher. A violin plot displays not only easily comparable mean and quartile information, but also more granular information about the shape of tge distrobution. This gives violin plots a unique ability to intuitively inform certain dicisions about further analysis, especially when youre exploring new data with numerous distrobutions for the first time. Histograms have an issue of sharing the same axis, which, when trying to understand intricacies of distributions, can be difficult to read. Box plots are easy to read but can obscure information, maybe leaving readers to question if the choice of a box plot was appropriate. A violin plot allows you to render an easily interpretable plot which lays bare qualitative aapects of the underlying dustribution. This not only allows for easy analysis via the box plot, but also high level qualitative understanding. I never, when i read a violin plot, care about the scale of the distribution, but the shape, which i think they do fairly well. Of course, when publishing I may or may not use them. I find them incredibly good for visualizing data exploration, and like to use them when explaining datasets moreso than results. On the point of smoothing, totally. Thats why ive gravitated towards swarm plots for general qualitative distribution understanding. But, smoothing is an issue within itself, histograms have essentially the same exact issue in terms of bin size. Also, worth noting, im colorblind af, so overlayed color infornation may as well be jibberish to me, which might be part of the reason i hate overlayed histograms so much.

An alternative to a violin plot is a "beeswarm" plot: Instead of a smoothed density you plot each individual datapoint as a dot at its exact y-value, and the x-values of the points are chosen so the dots don't overlap, causing y-coords with a lot of dots to bulge out. I like them because you can simultaneously see the raw datapoints, and also see the broad distribution. One problem is that in naive implementations you get chains of points extending out from the center in a line, giving a christmas-tree appearance. But good implementations can avoid this.

I'm an old man, so I know something about typewriting erasure techniques. In rough order of technological advancement: 1. hand-erasing with an ink eraser 2. Liquid ink eraser solution, in a bottle. You would apply this to the typo on the page and it would break down the ink and fade it somehow. 3. An "eraser strip" as part of the ink ribbon where you would retype the offending letter using the strip and it would abrade/absorb/dissolve the ink 4. hand painting over the typo with white-out (from a bottle) 5. hand-held whiteout strips, with dried whiteout on one side. You would retype the letter with one hand, holding the strip against the page where the hammer hits with the other hand. 6. whiteout strip built into the ink ribbon. Same as the whiteout strip, above,, but you don't need to hold the strip, its part of the ink ribbon. Most solutions did not require removing the paper, because you can *never* get it aligned again. There's typically enough space where the hammers hit the page for you to get in there with whatever erasing solution you're trying to use.

This channel really feels like having a cool older PhD friend who tells you all the secret tips and tricks and cool stories in academia

Related to that very last point about "why not just make it half the graph", when they made us do violin plots in high school biology we were told to put *half* the data on each side so that the width of the *whole* thing matched the amount of data. Which is absolutely not how anyone else does their violin plots. Also we had to do them by hand, which is as excruciating as it sounds

You'd *apply whiteout w/ the page STILL inside* the typewriter, wait about 1 minute, and then use the *backspace* key to shift _back a space_ to your mistake so you can apply new ink over the dried whiteout. Later on there were typewriters that could apply the whiteout for you. Usually electric ones. There were some other more esoteric solutions too! But yeah, most people just applied something directly to the paper and shifted back a space. Calling it "backspace" on a computer keyboard is one of many holdovers from the typewriter days. So is the stubborn yet incorrect convention of double-spacing after a sentence. (Double-spacing is essentially a typewriter trick/convention that makes things easier to read because periods are so small and on some typewriters don't offset enough. Single space after a sentence has ALWAYS been typographically correct in the world of typesetting books - plus print & graphic design.) We use a lot of old terms and symbols that don't apply anymore. Like saying "rolling" when a camera starts recording comes from the early days of film when there was a step to roll the film. Same way that many save icons are still simplified shapes of floppy disks - lots of kids grow up associating that shape form with "saving" without actually knowing it's a real physical thing. Or how how we associate the power symbol with turning things on (it actually was originally a standby-reset symbol or something but that's a whole different conversation.

I remember vividly how my advisor emphasized how he read papers: title, authors, abstract then plots. If the plots were compelling then he'd dig in. Those plots absolutely need to tell a succinct and coherent story.

"You can't actually get data, you're just getting vibes." Brilliant.

I'm convinced people only use them because they look vaguely sexual

My theory: The inventors of the violin plot were *literally trolling,* and call it internally the *pussy plot.*

One principle of peer review is that we shouldn’t just assume authors are analyzing their data correctly. I appreciate violin plots because they provide the reader reassurance that the use of box plots is appropriate. Absent the density overlay, I worry (and sometimes rightly) that the authors are using box plots in inappropriate contexts (as evidence from the fact one sometimes sees multi mode distributions in violin plots in papers)

A violin plot to overthrow the physics department !

As a non-scientist, I've always thought these plots were confusing and just obviously above my pay grade. Very validating to hear that they are indeed as uninformative as I thought they were. Much appreciated❤️

My grad school used to have professional plot makers on staff. It was long before my time but the space they worked in was still there and there were some people still working there who remembered them.

The colors on the graph at 20:37 are really really hard to tell apart for people with deuteranopia (some 6% of males). These pastel colors are hard, it's much better if they are very definitely yellow, or blue, or red, or grey, etc. Just wanted to chip in, since we're talking about it already.

I've always felt vaguely guilty as a scientist for never using the violin plot functions in any plotting tool - thank you for lifting this weight from me.

I thought, "Well they look funny, but surely there's a reason why they'd be useful!" And then at the 8-minute mark, you finish explaining how you make a violin plot and I'm like, "Okay but why would you do that though???" I think it's a terrible plot already and there's still over 30 minutes of reasons to listen to. Brilliant!

I am not a scientist. I have never heard about violin plots before today. But now I know about them and why they are mostly useless. I cannot overstate how much talent you have for making subjects like these interesting. A lot of the time, I pause science videos while alt tabbing to other things, taking in the videos in chunks. I always seem to watch yours straight through, beginning to end. The way you break up your videos with music and title-cards really helps make them digestible. Thank you!

As to the violin plot joke thing social difficult choices. A classic is to look confused and ask why the joke is funny in a very sincere way. Another way that tends to work for me is to just say "dude, c'mon", as that puts the onus squarely on them I do find it very helpful after a moment like the one you described to reflect on what I could have done differently while keeping my goal in mind. Generally speaking with stuff like this, the best approach is to deflate the other person, so to speak. I wish you all the best in your future strange and awkward social interactions

About papers 100 years ago. Based on the memoirs of the Stephen Timoshenko it seems that there were special people at universities who prepared plots. You would give them hand scatched drawings and they prepare then versions for a paper.

I think the type of smoothing they're doing is called a Kernel Density Estimate. They didn't teach us about KDE plots in physics classes because, as you say, they mostly just show vibes, but it's still better than the arbitrary-window smoothing you're suggesting they do. See the Seaborn documentation for violinplot

Finally! Someone else hates these things!

Violin plots are just Georgia O'Keeffe paintings for STEMlords

I'm convinced the original paper was an elaborate troll. It's funny because sexual body parts are inherently funny. For the same reason farts are funny. It takes a physical (physics!) aspect of ourselves which is tied to emotions that we normally try to keep private and forces it out into the open. This emotional discomfort is mostly politely expressed as humor. Now had your joking colleague just said with clear disgust, "It looks like HONK, ugh" wouldnt that have been worse? Of course you are right, best course of action would have been to say nothing at all. The obvious humorous response to the violin plot is to come up with some sort "c*ck" plot and then come up with a logical way to overlay it or point it at the violin plot. And , get it published straight up. Never acknowledge any physical resemblance. That would be funny. I know you'd like it to be but, clearly, humanity isn't better than this.

I think the reason the histogram got mirrored is because "symmetry makes it look and function better" (which isn't necessarily true in general, and certainly not true in this specific case, but it feels like a common misconception, though that might also be personal bias because my spicybrain likes symmetry). Also, the joke is that genitals are funny. Not just AFAB people's genitals, everyone's. I've seen some radio reception graphs that look like a different set of genitals, and had to stifle a giggle. Like, the sexism stuff is definitely real and valid, and there's a time and place for genital jokes, and an academic setting definitely isn't that, and the jokes absolutely age more quickly than some short-lived isotopes, but still.

This is actually kind of funny. I'm a PhD student in statistics, and I learned about violin plots for about 10 seconds in one of my first year courses. Just a few weeks ago I ran into a situation where I actually considered using violin plots to convey the distribution of sequence lengths for a system running in different states. However, I did ultimately decide to use a different plot, because the finished violin plot just looked too weird, and would have been distracting. I admit, I've never seen them used in a professional setting by other statisticians or scientists.

I've never once used a violin plot... but for some reason i still felt like i was in trouble the whole video.

*me, struggling through learning data analysis* time to get my opinions validated about this dumbass plot

I want to defend violin plots against some of the criticisms from the first segment of this video. But, the criticism in the second part is why I've never and will never make a violin plot. So maybe let's imagine a world where all violin plots were replaced with their half-violin equivalents, which as Dr. Collier concedes are definitely better. Here goes: - On smoothing: smoothed density plots look different depending on the value of the smoothing parameter. And histograms look different depending on the bin width. I guess histograms have an edge because you can see the bins and thus infer the amount of "smoothing" created by binning. But the basic issue is present in both: you can make two histograms or two density plots of the same data that look very different. - On violin plots vs. stacked/overlapping histograms: stacked histograms are also ugly and get really hard to read the more groups there are. Overlapping histograms are an abomination. They cannot be read. The plot at 20:17 is indecipherable. - Joyplots (which we're supposed to call ridgeline plots but I like joy more than ridgelines) are great. Yay for these. But, if we're gonna say "yay for these", how can we also say "boo for the half-violin plot"? They're almost the same thing. The joyplot puts the groups behind each other and does a cool little perspective thing. The half-violin plot puts them next to each other sideways. That's it, right? - On density plots and sample size: YES!!!! This is a huge issue. If we want to plot the distributions of two samples on the same plot, and the samples are of different sizes, we have a choice to make. And the correct choice is usually the one that conveys information about sample size. But, there are exceptions. Either way, this needs to be made clear. I just don't see why it's a violin plot problem. Isn't it also a stacked density plot or stacked histogram problem? - The plot at 26:15 is probably missing a caption, which points to a problem with these kinds of plots. Sometimes those lines are confidence intervals, meant to capture some hypothetical "true" or "population" mean. Sometimes they're "prediction interval" plots, meant to capture some % (like 95%) of data. Sometimes they're like the boxplot whiskers, extending from maximum to minimum. Sometimes they're the mean +/- a standard error, as opposed to +/- two standard errors like with most confidence intervals. Those lines could be anything, and you gotta read a caption in order to interpret the plot. Which, I think was a criticism of the violin plots. Ok, those are the defenses. But they're all pretty much moot because violin plots unfortunately look kinda like gentalia and it doesn't matter what else they might have going for them, that's enough reason to not use them. Also, I wish I worked in a field where "too many plots" was a problem. Try reading education research papers. All anyone talks about is directionality: "we made this change to the class and then there was a statistically significant increase in scores on a concept inventory, The End." Most of the time when I read a paper I'm just praying and begging for some plots. Any plots! Show me a picture of the data!!!! But no, it's all endless prose about what was or wasn't statistically significant.

And just how many histograms can you overlay, before 12:33 happens? Because showing histograms side by side certainly does not allows for easy visual comparisons of one vs another, especially for: Number of histograms > 2. Also, per 13:35 , I don't think you know what a histogram is! Those are vertical bar charts.

actually there IS a type writer that can erase stuff and uses some weird ass material science I don't know about to kinda suck the ink out after you write it, my good ol' technology connections has a whole video about corrections in type writers

While I also find violin plots a bit hard to read, for something like the paper at 13:30 where they apparently want to compare 7 different probability distributions side-by-side, I'm not sure any other option would be much more readable. It's probably too many to overlay the probability densities on top of each other (although I agree that's a good option for comparing 2 or 3 distributions). I guess they could do 7 side-by-side histograms or pdfs. 🤷🏻‍♂️ (By the way, Fig. 3 and 4 you point to aren't actually a histogram of the same thing, they're a bar chart of something else. Note the x-axis isn't numeric, unlike the y-axis of the violin plot. Sorry to nitpick!)

The animation on 18:20 could've just been a line diagram. Perfectly conveys the same data with just a single image. Super easy to plot in Excel too. No need to make it a complicated animation that’s impossible to understand. I feel like that's often the case with dataisbeautiful. It's almost a competition in presenting the most basic data in the most convoluted way possible. Like those "make the worst volume bar" UI challenges, but serious.

9:09 Histograms are pretty difficult to use for comparison of single-cell RNA sequencing data. You can also make a variation on the violin plot by making each half of the violin represent two different conditions, eg. a control condition vs. some treatment (eg. how does this drug that inhibits this pathway influence expression levels of this other thing, compared to no drug). The ridgeline plot is one alternative, but it's pretty space-inefficient unless you overlap the data and risk harming readability.

When you say that a man brought up that the plots resemble a vagina and combining that incident with the flip across the y axis being superfluous, there's some part of me that can't accept that not being on purpose. That they exist solely as some sort of joke to the original author that somehow managed to get published.

A picture is worth a thousand words. Dr. Collier's beleaguered sigh is worth a thousand data points.

When you made a mistake when typing on a typewriter you had to use a backspace then you would use a special white chalk covered piece of paper - put it between the typing tape and the paper then type the wrong character again - that would "erase" the wrong character, so you could use backspace again and press the correct key.

I enjoy that this turned into a histogram appreciation video because histograms are really great.

I worked as a statistician in the 90s and into the early 00s, and never heard of a violin plot. Knowing what they are now, I see they are entirely useless.

People love "chart art." I used to work in finance. There was pressure to replace data tables with charts when possible, even if the charts ultimately distort the data. But the chart makes the report look pretty. A very common chart that I absolutely hate is the 3D pie chart. The pie chart is already a bad chart, but someone has sabotaged what meaning a pie chart has as the areas have become distored and are no longer direct representations of the weights.

When I wrote my bachelor's thesis in uni my supervisor insisted that I would use violin plots to show some data. The problem was, the outliers in my dataset where not many but they were far out, like really really far out. So in the plots the often just weren't visible at all but I had to include them to represent the data accurately, at least that's what I was told. In the end the captions for every figure featuring these plots ended up absurdly long because I had to explain what the hell was going on there lest I forget it myself. So I 100% agree with you that these plots are just bad in every regard. The bit about these plots looking like genitalia is also true in every regard: because of these outliers some of my medians ended up at the bottom of the plot so that's of course where the belly was located, this however had the fun little side effect of making these particular plots look like a cock and balls. So my supervisor basically insisted that I would draw a bunch of dicks in my thesis. These things are truly terrible, they just always look like genitalia

In 1956, Bette Nesmith Graham (mother of future Monkees guitarist Michael Nesmith) invented the first correction fluid in her kitchen. Working as a typist, she used to make many mistakes and always strove for a way to correct them. Starting on a basis of tempera paint she mixed with a common kitchen blender, she called the fluid "Mistake Out" and started to provide her co-workers with small bottles on which the brand's name was displayed.

There's a strong link between embarrassment and humour, and a lot of embarrassment over taboo topics like anatomy that roughly half of people have (particularly among teenagers and people who haven't got over having been teenagers). As for the use of references as a form of comedy, the basic idea is "we all laughed at {thing} then; remembering it will bring you to a similar state of mind and probably make you laugh now". If you didn't laugh as a teenager when someone broke taboo, then you're not going to find it funny when people try to evoke the experience as an adult. On the other hand, if you're someone who found Monty Python hilarious, then someone saying "this parrot - " (the pause is essential) is going to remind you of John Cleese screaming at Michael Palin and very likely get at least a smile, if not a chuckle, out of you. There's also a whole in-grouping thing going on - "you and I share an understanding of this reference, therefore I am a member of the in-group and popular and successful"

A few counterpoints: 1) Violin plots are hard to understand. I don't see how violin plot would be any more difficult to understand than a KDE density cure or a histogram. If you know it's a density estimate - you know it's a density estimate. I think it's pretty obvious what violin plots show an how to read them. Then again, it seems there are multiple people commenting here that they have had trouble with this so maybe you are right, I don't know. 2) Why not just use a histogram? Well, couldn't the same point could be made against KDE (or any other smoothed density estimate): why not use histograms? You don't suggest this, but histograms aren't a panacea for density estimation and have their own problems. (Andrew Gelman has many times criticized R's bad default settings for histograms and I agree). 3) You suggest that multiple violin plots could be replaced with KDE curves that have a z-axis. I STRONGLY disagree with this one. Adding a z-axis creates more problems than it solves, since now you have to adjust in your brain for perspective. I honestly don't understand how you can sincerely suggest that this would make comparisons easier. You also have another example in which you replace violin plot with KDE curves (around 10:23). I think you made the graph significantly worse. I don't think it's easier to compare the distributions in your version, quite the contrary: the original violin plot version clearly shows the different shapes of the distributions, and you get an idea about their location and spread with a single galce. Your version, in my opinion, is more cluttered and difficult to read: it takes more time to disentangle the distributions from each other. There's also a greater probability of mixing up the curves (which curve is from what parameter/dataset or whatever is being plotted). --- I think violin plots are useful when you have to display many density estimates at once, for example, when visualizing marginals of a Bayesian posterior distribution. I don't think your idea of overlaying KDE curves is a single plot, or trying to use z axis to de-clutter them is any better or clearer; I think they might actually be worse. As a kind of "worst case scenario" consider that you want to compare marginal distributions from two 10-dimensional posterior distributions. Would you just shove all 20 density estiamates on top of each other in a single plot? Would you make 10 graphs, taking up huge swathes of space? Or just a single violin plot, in which the relevant density estimates are next to each other? You might be right that violin plots are overused in contexts in which information should be conveyed in some simpler way to people who might not be that familiar with density estimation.

Ha, only a minute in and I already adore the video. Out curiousity I had a professor give me old dissertations to see how they used to do data visualisation back in their days. And the solution is glue. Glued in graphs hand drawn on graph paper. Glued in photographs of the setups. And then it clicked in my head why we learned to do all that glueing stuff on paper in elementary school.

I'm only 8 minutes into the video, so you may well change my mind before the end, but I have made good use of violin plots in my work. When I've been comparing posterior distributions of multiple parameters from multiple different MCMC chains, the violin plots have been an excellent way for me to tell at a glance what the data is doing, and if there are any severe problems. Boxplots do not tell you if your posteriors are multimodal, violin plots do, and a histogram with 30 variables is going to be completely unreadable. I don't really care about the precise values of the interquartile ranges, I want to see if chains are converging to the same unimodal distributions. None of this information is for presenting in a paper (I'll give sensible posterior distribution plots there), it's for me (and my collaborators) to understand how well my MCMC is converging, and where the problems are. For that they work pretty well. Okay, now I'm going to shut up and continue watching to hear what you have to say! EDIT: all my violin plots were horizontal. It never actually occurred to me what they resembled when viewed vertically.... Also I work in veterinary epidemiology (I'm a mathematical modeller), where the majority are women, and my supervisor is a German woman (also did maths at undergraduate), who has no issues with speaking her mind, so I don't think anyone would be as daft as to joke about it!

I have used violin plots and liked them. You have convinced me that I was wrong.

Hospital business analyst here... the reason I use violin plots is to specifically highlight the fact that the data is wonky and it cannot accurately be represented by an average... despite it being reported out that way on ten prior occasions

This went from ha ha, to not ha ha real fast I want to say that you do an amazing job explaining the female side of these interactions The fact that you articulate why it was not ok is a great source of information for people who want to do good but don't yet see how certain things are problematic

It is funny, but my master thesis was actually one of first theses in my university that were typeset using LaTeX. Probably the first thesis, actually. And my diagrams were drawn using Postscript. Yes, I wrote the programs to draw the diagrams. Of course, Knuth created the whole digital typesetting thing because the expert typesetters (actual people) were retiring. And the new generation could not do his "The Art of Computing" well enough because of all the diagrams and mathematics. Yeah, we came a long way. I was just one of the people right in the middle of the old and the new. Later, I was using gplot to generate Postscript graphs. It still exists, I believe.

My wife in astrophysics -- she has a very similar experience to what you described with the Violin plots, where she'll go out to dinner with a bunch of male physicists and the waiter will come up and say "Well, ladies first!" And she can't explain how frustrating it is to have the entire table be alerted/reminded of the fact that she's the only woman there. It's tough in that situation because she can't comment on how it makes her uncomfortable because the waiter's not being a bad person about it and it'll make her look bad if she brings it up at the table. So she just kind of has to deal with it. It sucks!

i just got through an 'intro to data science in python' course for my masters of analytics (it was a bloodbath- ~1100 people enrolled in the course. 110 people made a straight up ZERO on the final. got NO points. and 1/3 of the class got a failing grade on the final. which was worth 25% of our grade. but anyways..) there was a violin plot on one of the homeworks- i remember it because i was like "hey, what the hell is this thing?" it was supposed to be visualizing a group of horseshoe crabs, the length of their carapaces, and the average time of survival of the different lengths. i saw the violin plot and thought it was a plot of 'this is what the scientifically ideal horseshoe crab body looks like.' i read it like 5 times before i realized the violin wasn't supposed to be the shape of a horseshoe crab.. smh. shitass plot, zero out of 10.

There are so many data visualization that just should be wiped from existence. Violin plots are at the top of my personal list, they are outlassed in every way in modern data visualization

Do be careful with plots that heavily rely on color. Color vision impaired folks can't always read them well.

I totally agree about the "feminism" point. The level of defensiveness, especially in STEM is insane. Re: the plot: I think they doubled it assuming we'd get a more intuitive sense of the "area" corresponding to an increased histogram value. Still a shitty plot

I’m confused. How do box plots “show the average of a data set”? I don’t think averages are part of box plots by default. Box plots show the quantiles. Some plotting software I use has options to add means to box plots, but it is not traditionally part of them, right?

I love your videos. Like why would I care why a certain plot is horrendous? 40 plus minutes later I'm super invested and ready to go on the war path about violin plots.

I don't know if anyone has mentioned yet, but a population pyramid comes to mind as a use case where it makes sense to have both the mean or median and quartiles, but the overall shape of the distribution is still important and useful. Mind you, I'm typing this at the 9:20 mark, so maybe that does come up in the video at some point and I'm just jumping the gun.

I sometimes have a lot of possible plots I could do, and generating all of them as violin plots is useful because I don't know in advance if my data is bimodal or whatever. And I can put lots of violins next to each other and compare them, unlike histograms. But sure, I won't put them in my presentations because by then I know the best way to show the data. Fair enough. Also excellent point about the smoothing! People not understanding their statistical models bothers me a lot.

I've never seen these before. I do think they would make sense in One particular situation. You give an example of temperature data distribution per month (12 separate plots) but suppose instead of months, you want to plot the annual temperature distribution of all the world's capital cities, and you want to put a continuous variable (latitude) on the X axis, with temperature on the Y axis. In that situation, the symmetry of the violin helps centre the data correctly on the X axis. I completely agree that this format should only be used for overview information (but I note that it can work without colour, while certain other types of plots can be difficult to read without colour.)

I loved this video. I don't get the whole inappropriate stuff from what looks like gross bugs / fish / sting rays. I'm not a woman though, so who cares about my opinion? XD Seriously though, these plots are horrible, simply for being unintuitive. I can't tell what's going on, and it looks like someone is comparing leafs, or blobs of slime, or rocks, or weird sting rays or something, instead of doing actual science. If I saw this in a paper, I would be like "Yo, what is this supposed to be? Where's the data?" would be my actual comment while presenting. Is that a somewhat unprofessional comment? Well, yes, but it's better than calling them something irrelevant, like a sting ray or part of a human body or something. Counselor: What do you see? Me: I'm setting the foot print of a dinosaur after it rained for a few minutes and the foot print got smoothened out. Counselor: Oh. Well, you were supposed to see a 2-headed butterfly... I guess that means your a psychopath now. Me: What does this ridiculous and non-sensical image have to do with my mental well being anyways?

I've been watching this video thinking "Mmm, I don't think I'm seeing where Angela is coming from here. They aren't that bad." ...until the bean plot at 23:07. An absolutely chaotic dumpster fire of borderline illegible meaninglessness. I get it now. I'm revolted that it took me this long.

I come out and say that I used the Violin Plots in my PHD thesis. I had grain size distributions to display. Histograms have the big problem that you can not well put a lot of them over each other. I had I think ten different samples I wanted to display next to each other (to make them comparable). Furthermore, I wanted to use the median to simplify the further discussion, but I also wanted to show the actual distribution. Of the Particles, as it was important for the behavior I was looking at. The Violin plot was a good combination of: 1. The median is visually represented. 2. I do show the actual distribution, so I can discuss the skew, if there is one. 3. I can pack a lot of them next to each other, the reader gets a good visual representation of the different distributions. 4. Yes, I find them visually pleasing, if done right. I did plot them horizontal, so. 5. I think the Violin plot is symmetrical, for the same reason as the Boxplot is symmetrical. And when I think about a metal grain, which I worked with). I thought like it represents the form of the actual grains. I did not add another histogram. I did give the smoothing value and normalized the width to one. As, I saw the second part of your video, I am sorry to hear about the unfortunate situation and that this plot makes you and other women uncomfortable, this is unfortunate. I did not know about this, and it was absolutely not obvious to me. I am sorry for that.

I would argue that in the era of information, violin plots have an important purpose. I get that many dislike them, and I appreciate why. It's annoying to take a ruler and run it across a list of plots to figure out the values. But 99.9% of people nowadays are not actually doing knee-deep research that requires precise fact-checking for these plots. Also, usually whoever made the plot has also published their data so you can just make your own plots or use the data directly, which is actually way more accurate than eyeballing some numbers if you're actually doing math for something in your paper. And violin plots have a very useful feature, skimmability. They reduce a plot down to its basic shapes and make it so the human brain, upon skimming, can quickly identify "hey, this visualized 2D object has three humps, whereas this one has two. And this one looks just like a ball. These are fundamentally different samples!" Instead of just showing you the curve of the shape (like a normal histogram), the violin plot also gives you the ability to "feel" the volume of the displayed "object" at a specific point. This is useful because our monkey brains are better at identifying and classifying objects than curves. As an example, what might look like a slightly more sudden downturn (but no big deal) in a histogram could give you a "wow, this sample is tapering out like crazy at the bottom!" effect. Also you can often just throw violin plots on top of your box plots and it will just give readers a better understanding of the data than pure box plots. However, I do agree that smoothing diagrams so they look pretty can be quite obnoxious and sometimes even obstructive to science. And it happens most often with violin plots, since they give more of a "wow" effect if you smooth them like crazy. Additionally, I think violin plots really only "work" with certain types of small data (i.e. just look at these three humps). If there are too many metrics, it becomes a nightmare to read and interpret violin plots (as you have stated) and it should just be thrown in a different diagram. Basically, if your violin plot looks more like a weird snake than a violin, you should rethink how you're presenting your data. And honestly, I do just want more research on these plots and their effects on scientific research. If you can get a paper together that actually proves violin plots are trash and obstruct science, I would love that. It would help me accept that everything in this comment (so far) is just a subjective personal experience from me and my professors, and I'd be more inclined to never use one ever again. My counterarguments to the video are subjective experiences, after all. Also, then I could cite you, which would "appreciably increase your market value" as they say on LinkedIn, I think. Oh, and I'd love to see the feedback of the people I've worked with who have conducted extensive research primarily relying on violin plots. I'm cool with sharing videos like this with my friends, but I'd rather give my professors something more academic, you know. Addendum: Wow.. I don't have a vagina and I leave thoughts about genitals at home when doing research, I can't believe I never realized they look like . Yeah, that will actually make me think twice about using them... Thanks to the creator for the story and transparency, it must suck to put themselves out there to this extent. And I'm glad they decided to ignore the idiots who go "huhuhuh va" at work. Especially since many of those people will likely see this video, potentially even including that f-ing guy. The "joke" is so thoroughly unfunny that it made me almost laugh at how badly the guy conducted himself. I think it's best to just pretend those people don't exist, in most cases. I really hope some of his coworkers or friends gave him a serious talking-to after the presentation... If any reader experiences this kind of behaviour from someone they know, please just sit them down in private and tell them it's not ok. That said, as many commenters have mentioned before, something looking kinda suggestive does not mean it is wrong to use scientifically. It's just really nice to have a box plot with extra data about a specific metric sometimes. Sometimes you just wanna be able to directly compare like 10-20 different f-ing plots without having them overlap and look like a horrible unreadable cluster-f. Especially if you expect the reader to find something you didn't expect. What if someone finds a cool association that would be hidden behind the back layers of the fancy 3D layered histogram? You cannot seriously tell me you can clearly see the beginning of each distribution at 21:00... Also, accessibility! Not everyone can see colours, colourblind people exist. That's why the one histogram in the beginning had kooky patterns, colourblind people wanna understand data distributions too. Also, I know it's the digital age, but many researchers like printing out their important references and notes. As a starving student, it costs soooo much more to print something in color rather than greyscale. I implore you to take your "useful diagram" at 20:40, print it in greyscale and explain to someone what it means. Or better yet, just ask your red-green colourblind friends to read it. In this case, not even kooky patterns can save you, because you want them all to overlap! Not every design decision is good, many of them have problems. But I would say violin plots have a clear reason to exist, even if that reason is a bit nuanced and overlooked. And yeah, I don't like making women uncomfortable, but anything is suggestive if you look at it too much. I think most of us can agree that the awkward situation was created by that one weirdo, not by the plot. However, I hold hope that perhaps one day some psychology student will set out to prove me wrong on that. Until then, I will probably keep using violin plots every now and then

I've had the good fortune to have never encountered a violin plot before being a science geek (NB: I am not a scientist, nor have I ever been. I've simply had a lifelong interest in science). I agree with this young lady, in general - violin plots look less informative than histograms or box-plots. As a means of presenting data, if I came across some article or paper that interested me that used a violin plot, I suspect I'd be little the wiser for having seen the violin plot. However, I do recall a video on evolution where, for various families of species, they plotted time on the vertical axis against some measure of how common the species in that family were on teh horizontal scale, and they did this all in a single image. Now, whilst it did not give me anything in the way of numeric data, I could see that this family arose around then, whilst this other family were dominant, then it expanded over time, there seems to be rivalry with a different family a bit later, and so on. You could see how various sections of the evolutionary tree began, blossomed and died out over time quite easily. For that, to give a rough overview, it worked well. And yes, it looked like a bunch of simultaneous violin plots., but because of the way in which it was done, and bearing in mind its intended audience - the general public - it adequately conveyed what it needed to convey simply and quickly. However, if I were sufficiently interested in biology and evolution in particular to the extent that I wanted actual data, I would not be a happy bunny given something like that. A histogram would, IMO, make it much easier to extract the information I might want from it. I'm prepared to accept that there may be edge cases (when are there not edge cases?!) where a violin plot might be the best way to display something, but I find it hard to imagine such a situation. And yes, I have read the other commentors including those speaking well of violin plots, whose arguments I find unconvincing. But perhaps I just need to see cases where violin plots are appropriate rather than ones where they are not?

😂😂😂 OMG, when I first saw these violin plots I had the same thought. I feel validated that someone else had the same thought, because I felt that my profession was slowly corrupting my brain - I'm an OB/GYN. Thank you for providing content for me to watch on my post-call day. It sucks being a woman in a science based field, and unfortunately this bias has seeped its way over into medicine as well (shocker). This is why representation matters, but we must also continually actively dismantle the patriarchy

I love these kooky patterns too! I first ran into this in Trello, where the labels (categories) are colored rectangles; but that sucks for accessibility. So they have Color Blind Friendly mode where they add distinct patterns to the colors. Everything should do this! I'm not colorblind and it helps SO MUCH!

Once, I loudly expressed that a piece of lab equipment looked similar to a "toy" - I still cringe thinking about my own behavior. The key difference being that I was an 11 y/o

I'll agree with you that violin plots have some problems, but uhm... 11:03 "If I open a paper and I see a violin plot, if I want to actually get data out of it.." Who's trying to get accurate data out of plots in papers?? Is that something people do??? I always think of plots as a way to tell a story about your data. I don't necessarily think you're meant to be reading them in that much detail. Like, you said "you can't get data from them, just vibes", and that's ... the point of them? Like, you talk about ridge line plots and 3D plots and I'm sorry but I think they're pretty much just as bad. It feels like they're just your æsthetic preference at this point.

While i agree that violinplots are not very good, some of the "better" alternatives you give are equally bad: - The 3D-plot that you claim still lets you compare/measure the data does not allow you to do that. 3D quite generally becomes near unreadable if your data is not neatly aligned or interesting features end up hidden behind layers in front. - Overlaying histograms on top of each other becomes entirely unparseable very quickly with the number of conditions/data groups I will agree that raincloud plots (with or without the "rain") are better, and that the researcher (and reviewers) should decide what contrasts are of relevance and what features of the data are worthy of visualization (or can simply be mentioned in text; or given in the supplementary materials), and will also concede i have myself opted for histograms/smoothed histograms instead of violinplots. Do i find them ugly? Yes, i do, mostly -- however, that's not a valid reason. 3D barplots look pretty cool, but are one of the single worst visualizations possible. Scatterplots are kinda ugly (depending on the data), but are a great visualization.

i love watching your videos while i procrastinate on my physics 101 homework. makes me feel like i'm actually working

Never heard of these, thanks for letting me know about them. Also I do think it’s important to talk about the woman’s perspective. I know there are a lot of men like the one in the badman example. But there are also men who want to fill in their blind spots. Hearing women’s experiences helps them do that. So thanks for that to.

I was thinking "man I've basically never seen these" but then I realized - they're employed for ALMOST EVERY chart showing the age demographics of a country. And it's always weird and hard to parse and takes a second. It probably started with people using the two different sides to show men and women (which honestly you don't need to, just turn it sideways and give a second color for one of them, it would be way more legible, 'down' is not generally the direction we think of time passing in) but nowadays most of them don't even do that. They just turn it into one of these godforsaken graphs because that's how you're supposed to show age demos and it wastes like three seconds of my life every single time

While I also don't like violin plots, a lot of this commentary is pretty bad. They're not impossible to read, often they are more intuitive people are used to seeing dependent values on the Y-axis, and if you have a large number of data sets then superimposed histograms become impossible to read. At some point you "improve" a violin plot by rotating it, but you end up with an axis that goes from high to low. If people don't tell you how they made the plot and what smoothing they used, then that's an issue with underreporting the methods, not an issue with the plot itself. That Vanderwaals paper you showed does NOT show the same data in fig 3 and 4 as in the violin plots in fig 5 and 6. (Also you say fig 3 and 4 are histograms, but they're not. They're a lot of bar graphs without error bars packed in tight, that's not the same thing.) But where you really lost me is about halfway where you show the 3D plot... You cannot be serious. You can no longer interpret the X-axis, a lot of it becomes hard to see and you completely obscure part of the data on the left side of the back few data sets. How on earth is that better? Yes, violin plots are not great and I get that it's fun to have a go at them if that happens to be a pet peeve. But all of these bad takes really don't help your argument or the quality of the video.