Hacker News Reader

Comment by mcnamaratw | Hacker News Reader

It's not clear what technical content this post has. A cubic polynomial vs time would be a terrible idea, but the dashed and/or dotted red line (that people seem upset about) is obviously not a cubic polynomial vs time.

At that point I run out of guesses about exactly what's being said by anybody.

View on HN

I think there's actually two things going on here. There's what this post is talking about, which is a general lack of understanding of statistics, which is likely true and correct, and likely for both reasons that are explored here.

Then there's the reason to actually even talk about this specific instance, which is that there is a complete lack of operating on the assumption of good faith and clarifying intent before making assertions as to other peoples intentions which is rampant currently. Whether it's rampant on twitter and between political parties or has spilled into other areas such that it's harder to have coherent discussions in general now than it was in years past I'm not sure.

There are a lot of assumptions in that reply to the tweet in question that's shown. It's not even a novel set of assumptions, it's the standard twitter fare of "I assume he means X and this thing doesn't explicitly show X therefore he must not understand what he's talking about." Any nuance such as using a secondary aspect of something to outline a potion of what you mean is immediately ignored, and if pointed out later assumed to be covering up after the fact.

Well, if you want to express nuance, Twitter might not be the medium for you...

Lol, I believe that. The problem is that these days it seems like the message either has to be only visible by friendlies, or reduced to the simplest and rote version that's impossible to misinterpret if you want to avoid people trying to tear it apart, and even then they'll sometimes pull it into some wider framework of society and what it means when views through the lens of X, Y and Z and why you should be mad. That's a high bar to hit just to tweet something.

It's crazy, and twitter's where it's easiest to see, but you can also use a news aggregator like news.google.com and get a good dose of it just from the headlines about what is ostensibly the same story from different news agencies. Wild times.

> ...there is a complete lack of operating on the assumption of good faith and clarifying intent before making assertions as to other peoples intentions which is rampant currently.

This is basically what all politics is. People projecting thoughts (or lack thereof) in bad faith on each other. Though it does seem that the fraction of society sucked into this toxic discourse has grown to cover practically everything lately (in the US anyway).

But also the outrage machine of Twitter etc is obviously biased toward the extreme views and outliers. Where 99/100 people might say 'meh, another failure to predict the future like most', the remaining 1/100 blows his top and gets the attention.

For all the negativity out there, there is a lot more clarity in challenging bad interpretations. I havent learnt as much online recently as I have in 15 yrs on the net.

> It's not clear what technical content this post has.

Indeed. The post is a big exercise in strawmanning (and of course ad hominem but I am willing to forgive it because it adds entertainment value). Andrew Gelman (a well-known statistics expert) dunks on some guy because he dared to defend a shitty graphic and "made a statistical error". Which is kind of unfair, because he didn't do any statistics. Of course when our brains see some curves they can't resist the temptation to do some statistics on their own, but bad dataviz is a different sin than bad statistics and should be treated as such.

Um, yes.

The statistical community has not been doing well with the coronavirus epidemic. Nobody's models seem to be predicting well. Nor is the source data for anything but deaths very good.

This matters, because the current plan US plan seems to be "open things up and wait for herd immunity". How much time, and how many deaths, lie between now and that point? I dunno.

> The statistical community has not been doing well with the coronavirus epidemic.

I'm curious how you concluded this. Can you (or anyone) recommend a reasoned evaluation of the existing models? We've got a few months of hindsight now and I'm curious hear what any reputable data scientist might have to say.

I see criticism of the models online (in op-eds and social media), but the criticisms are usually agenda driven and opaque on technical details, which isn't particularly helpful.

The actuals from now and the predictions from a few weeks ago aren't aligning quantitatively. The shapes of the curves, though, do make sense.

Here's the Financial Times graph of actuals, country by country.[1] No predictions. The pattern that shows up in some of the actuals is "huge spike, tight lockdown, big drop". See Italy, Belgium and the Isle of Man. The Isle of Man puts people in jail for weeks for violating quarantine rules. The US didn't have a huge spike, but isn't seeing a big drop, either. US deaths are at about 2/3 of peak.

[1] https://ig.ft.com/coronavirus-chart/?areas=usa&areas=gbr&are...

> This matters, because the current plan US plan seems to be "open things up and wait for herd immunity". How much time, and how many deaths, lie between now and that point? I dunno.

This seems pretty straightforward, no? You re-open cautiously, wait a few weeks, make sure your hospital resources aren't being overwhelmed, then open up a bit more, rinse and repeat. I wasn't aware there was another way.

The trace model: identify each infected person, trace all their contacts, test all of them, expand outwards. Full quarantine the infected. Intensive but it works; see South Korea and New Zealand.

The "herd immunity" plan can only work with more than 50% of the population infected, of which about 1% will die and some slightly higher percentage suffer lingering ill health, which in the US means at least a million people.

That's current US policy. That's what the "Get and Keep America Open" plan does.[1] Current death rate for the US is around 1,400 per day.

[1] https://www.cdc.gov/coronavirus/2019-ncov/php/open-america/i...

This could be a naive question, but everything I've heard is that coronaviruses don't disappear, they just sorta integrate and mutate into our normal "known set" that we deal with year-to-year.

If that is true, doesn't the strategy taken by SK and NZ put them at continued risk for a another outbreak if the virus sneaks back in? Without a vaccine, and then significant uptake by the population yearly, doesn't the risk of covid-19, and its mutations, come back every year?

We do eventually require a vaccine or similar.

SK and NZ can "end" the outbreak. The "herd immunity" strategy will simply continue it straight through the whole year, with a lot more deaths.

The models are mostly easy to understand (from simple differential equations for the simplest models to agent-based monte carlos modeling everyone ina region), but the parameters in the models depend on people's behavior, which is not something that can easily be predicted.

Argonne National Lab has simulations where they simulate every person Chicago and where they might go and who they might run into, which, given the correct set of behaviors would probably work very well. But... how do you predict how people will act in a week? That will depend on the weather, court decisions, number of deaths in random countries, how rousing a speech a politician might give, etc. If we could control how people acted, I bet the models would work very well.

> for anything but deaths very good.

Even that's iffy when compared to "excess deaths" (in lots of places, excluding all covid deaths still results in more people dying this year compared to previous years).

> A cubic polynomial vs time would be a terrible idea, but the dashed and/or dotted red line (that people seem upset about) is obviously not a cubic polynomial vs time.

What? The chart's creator said it was a cubic polynomial. Mr. Hassett said he had employed ‘just a canned function in Excel, a cubic polynomial.’”

https://www.vox.com/2020/5/8/21250641/kevin-hassett-cubic-mo...

on the other hand, for reals calculus (namely taylor polynomial estimation) also says, "why not a cubic polynomial?"

Because Taylor approximations can lose accuracy very quickly as you move away from the point they are centered around.

There are available epidemiological models that are actually grounded in a scientific understanding of the problem. "Why not a cubic polynomial?" is a stupid question.

"Why not a cubic polynomial?" taken literally is not a stupid question at all. No, it's not a contribution to understanding the problem at the professional level, but it is a fundamental and important question.

Somebody persists after hearing the answer (which you gave), that might be stupid.

Yes, and until you do the correct analysis and estimate just how bad the estimate is going to be (off by 10%? Off by 50%? An order of magnitude?) You can't just dismiss it out of hand. So how far off will it be?

The cubic is going to go to either +infinity or -infinity, so it will eventually be of by infinity%. Or infinity orders of magnitude.

I feel like I can dismiss it now.

There's such thing as a convergence interval and your flippant comment betrays your ignorance.

Convergence interval doesn't help at all when you're working with a fixed finite polynomial (cubic, in this case).

You don't need any flippancy to betray your ignorance, it's already clearly on display.

Sorry to shock you, but not every piece of text on the internet holds water mathematically.

Plot a cubic polynomial in time, at^3 + bt^2 + c*t + d. Use Excel. The curve either diverges to + or - infinity at each end, or it's a constant. That's not the behavior of the dashed red curve in the OP.

What it could be is some kind of cubic spline. But it is not a cubic polynomial in time.

It’s apparently a cubic polynomial fit to the logged data — see http://bactra.org/weblog/1176.html

A cubic polynomial goes to +infinity on one side and -infinity on the other. So taking exp, log, etc won't make it go to 0 on both sides. I can think of two possible explanations: 1) the leading coefficient happens to be 0, so it's actually quadratic fit 2) the leading coefficient is nonzero, so somewhere to the left or right the graph starts rising again, but it's not visible due to cropping. Anyway by now we've spent more time thinking about this than any of the participants, so we should call it a day.

The article may not have technical content, but the title certainly is true: http://pseudoexpertise.com/

Well the flame is saying it's a misunderstanding of the difference between data smoothing and model-based forecasting.

To which I'd say that just begs the question of if there's a difference to begin with, and what criteria you would use to distinguish them if they are.

Of course there's a difference between smoothing and forecasting. It's the difference between interpolation and extrapolation.

But I can't figure out what anyone in the original discussion is really saying, if anything.

To begin, if the curve doesn't extend into the future, it can't forecast anything.

I think it resembles a cubic spline kernel, FWIW.

Yeah that could be it. But in that case what's the big deal? The data is a messy bump on a log, the fit is a smooth bump on a log. That kind of fit is useless for prediction, but they say so explicitly.

Reply on news.ycombinator.com