Terrible graphs
Posted: Mon Apr 27, 2020 5:40 pm
This one, via The Conversation, is... something
It reminds me of a phone conversation with a colleague on my walk last week, when he was wondering what a graph of total cases vs hospital admissions vs deaths would look like.Bird on a Fire wrote: ↑Mon Apr 27, 2020 6:13 pmI think that's quite an effective way of showing the looping effect they're talking about in the article, actually. Though it's certainly not very pretty or intuitive.
That's a very good point.jaap wrote: ↑Tue Apr 28, 2020 6:44 amWouldn't any measurement with regular fluctuations show the same loops?
It's basically plotting a sine wave against its derivative, a cosine wave, which gives you a circle. And the only reason it is like a sine wave is because it's using rolling averages to smooth out the fluctuations so that they are more wave-like.
It doesn't really tell you anything particularly interesting, except may that there are weekly fluctuations due to weekends.
I think we figured out that the "cubic model" was actually a cubic fit to the log of the data.Allo V Psycho wrote: ↑Wed May 06, 2020 5:26 pmThis came from the White House's Council of economic advisers. On Twitter https://twitter.com/WhiteHouseCEA
They say "To better visualize observed data, we also continually update a curve-fitting exercise to summarize COVID-19's observed trajectory. Particularly with irregular data, curve fitting can improve data visualization. As shown, IHME's mortality curves have matched the data fairly well."
I laughed out loud - did they use 'Add Trend Line' in Excel, and choose a cubic?
The cubic is the one that predicts zero deaths in 9 days time.
And what kind of curve fitting exercise gives the green line from the black one?
Plus, was that graph really drawn by hand using coloured sharpies, or is there a programme that makes your graphs look more user friendly, by making them more crap? I did think of posting this in Pandemic Jokes, but then, it ain't that funny.
I read it as the cubic fit range is only the dashed line, and they've projected it (dotted line) using an unspecified different method because they know the cubic itself generates nonsense. Maybe.bolo wrote: ↑Wed May 06, 2020 7:52 pmLast time I checked, a cubic goes to +infinity in one direction and -infinity in the other. Even if it's cubic in the logarithm, it can't go to zero in both directions. So if this is a cubic at all, the coefficient of the cubic term must be zero, so really it's a quadratic (in the log) which is ... I'm not sure, is it stupider or just equally stupid?
Well it does look like a Gaussian, which is what you would get if you took the log of the data and fit a quadratic to it (given that the second derivative is negative) and then exponented it back.bolo wrote: ↑Wed May 06, 2020 7:52 pmLast time I checked, a cubic goes to +infinity in one direction and -infinity in the other. Even if it's cubic in the logarithm, it can't go to zero in both directions. So if this is a cubic at all, the coefficient of the cubic term must be zero, so really it's a quadratic (in the log) which is ... I'm not sure, is it stupider or just equally stupid?
Kevin Hansett talked about the cubic model. And we're discussing someone who has form in economics, with an absolutely barking Laffer curve, which I might dig out later.bolo wrote: ↑Wed May 06, 2020 7:52 pmLast time I checked, a cubic goes to +infinity in one direction and -infinity in the other. Even if it's cubic in the logarithm, it can't go to zero in both directions. So if this is a cubic at all, the coefficient of the cubic term must be zero, so really it's a quadratic (in the log) which is ... I'm not sure, is it stupider or just equally stupid?
That Laffer curve is a fib of beauty. Following through the discussions that ensued, I found someone asserting that Norway is plotted in the wrong place: he took the total tax revenues including those coming from oil for the y axis, and plotted them against general corporate tax rate on the x axis But oil revenues in particular were at that time taxed at a much higher rate than that, so apparently Norway should be have been way over at about 52% rather than 28%.jimbob wrote: ↑Wed May 06, 2020 9:14 pmKevin Hansett talked about the cubic model. And we're discussing someone who has form in economics, with an absolutely barking Laffer curve, which I might dig out later.bolo wrote: ↑Wed May 06, 2020 7:52 pmLast time I checked, a cubic goes to +infinity in one direction and -infinity in the other. Even if it's cubic in the logarithm, it can't go to zero in both directions. So if this is a cubic at all, the coefficient of the cubic term must be zero, so really it's a quadratic (in the log) which is ... I'm not sure, is it stupider or just equally stupid?
[ETA: https://www.discovermagazine.com/the-sc ... tting-ever ]
And it only needs to have a turning point at the future zero for a dishonest/incompetent person to then take that zero gradient as all time future.
I don't know which he fails on, but he's certainly not both capable and honest - given his work I'm going for neither.
That Twitter link led me to this:shpalman wrote: ↑Wed May 06, 2020 5:42 pmI think we figured out that the "cubic model" was actually a cubic fit to the log of the data.Allo V Psycho wrote: ↑Wed May 06, 2020 5:26 pmThis came from the White House's Council of economic advisers. On Twitter https://twitter.com/WhiteHouseCEA
They say "To better visualize observed data, we also continually update a curve-fitting exercise to summarize COVID-19's observed trajectory. Particularly with irregular data, curve fitting can improve data visualization. As shown, IHME's mortality curves have matched the data fairly well."
I laughed out loud - did they use 'Add Trend Line' in Excel, and choose a cubic?
The cubic is the one that predicts zero deaths in 9 days time.
And what kind of curve fitting exercise gives the green line from the black one?
Plus, was that graph really drawn by hand using coloured sharpies, or is there a programme that makes your graphs look more user friendly, by making them more crap? I did think of posting this in Pandemic Jokes, but then, it ain't that funny.
https://twitter.com/potatoffel/status/1 ... 09792?s=09
No scientific basis or predictive power whatsoever.
It does seem like the IHME predictions are massively sensitive to the data so might well be overfitting as compared to have a decent prior and/or maximizing the entropy.
I think you have to pick the end points of the fit and projection carefully.Allo V Psycho wrote: ↑Thu May 07, 2020 7:00 amOK, I plotted their data as best as I could estimate it (in Excel, of course). Yes, I know I'm sad, but I don't start work till 8, so it is my own personal sadness.
Both log and ln give me a curve, but it's platykurtic compared to theirs: Closest I can get is to plot a cubic on the original data then match a gaussian to part of it.
It does appear to have been drawn using the graphical equivalent of Comic Sans...Allo V Psycho wrote: ↑Wed May 06, 2020 5:26 pmPlus, was that graph really drawn by hand using coloured sharpies, or is there a programme that makes your graphs look more user friendly, by making them more crap? I did think of posting this in Pandemic Jokes, but then, it ain't that funny.
It would be straightforward to get that in Excel, but you'd need to actually put in effort to get that - by sorting the data in descending order.Holylol wrote: ↑Tue May 19, 2020 7:59 amA new one:
https://statmodeling.stat.columbia.edu/ ... his-graph/
Although, somebody might make a graph, then sort the cells to look at the data, without realising that the graph would be automatically updated to the new order, and then not check it. Excel doesn't always handle dates well either.jimbob wrote: ↑Tue May 19, 2020 5:50 pmIt would be straightforward to get that in Excel, but you'd need to actually put in effort to get that - by sorting the data in descending order.Holylol wrote: ↑Tue May 19, 2020 7:59 amA new one:
https://statmodeling.stat.columbia.edu/ ... his-graph/
It would be very difficult to do that accidentally.
Excel handles dates better than that, though. Again, it would require sorting the data into descending order and then plottingBird on a Fire wrote: ↑Tue May 19, 2020 6:25 pmAlthough, somebody might make a graph, then sort the cells to look at the data, without realising that the graph would be automatically updated to the new order, and then not check it. Excel doesn't always handle dates well either.jimbob wrote: ↑Tue May 19, 2020 5:50 pmIt would be straightforward to get that in Excel, but you'd need to actually put in effort to get that - by sorting the data in descending order.Holylol wrote: ↑Tue May 19, 2020 7:59 amA new one:
https://statmodeling.stat.columbia.edu/ ... his-graph/
It would be very difficult to do that accidentally.