UPDATE 7/21/2021: As you doubtless know at this point, it was not over. Given the visibility of this post, I’m going to note here at the top that the prediction of a potential large wave of infections between March and May did not happen, no matter what ultimately happens with Delta (and the prediction was not made with Delta in mind anyway, only Alpha). Some more reflections on that at the bottom of this post here.
A year ago, there were reports coming out of China about a new coronavirus. Various people were saying things about exponential growth and the inevitability of a new pandemic, and urging action be taken. The media told us it was nothing to worry about, right up until hospitals got overwhelmed and enough people started dying.
This past week, it likely happened again.
A new strain of Covid-19 has emerged from southern England, along with a similar one in South Africa. The new strain has rapidly taken over the region, and all signs point to it being about 65% more infectious than the old one, albeit with large uncertainty and error bars around that.
I give it a 70% chance that these reports are largely correct.
There is no plausible way that a Western country can sustain restrictions that can overcome that via anything other than widespread immunity. This would be the level required to previously cut new infections in half every week. And all that would do is stabilize the rate of new infections.
Like last time, the media is mostly assuring us that there is nothing to worry about, and not extrapolating exponential growth into the future.
Like last time, there are attempts to slow down travel, that are both not tight enough to plausibly work even if they were implemented soon enough, and also clearly not implemented soon enough.
Like last time, no one is responding with a rush to get us prepared for what is about to happen. There are no additional pushes to improve our ability to test, or our supplies of equipment, or to speed our vaccine efforts or distribute the vaccine more efficiently (in any sense), or to lift restrictions on useful private action.
Like last time, the actions urged upon us to contain spread clearly have little or no chance of actually doing that.
The first time, I made the mistake of not thinking hard enough early enough, or taking enough action. I also didn’t think through the implications, and didn’t do things like buying put options, even though it was obvious. This time, I want to not make those same mistakes. Let’s figure out what actually happens, then act upon it.
We can’t be sure yet. I only give the new strain a 70% chance of being sufficiently more infectious than the old one that the scenario fully plays out here in America before we have a chance to vaccinate enough people. I am very willing to revise that probability as new data comes in, or based on changes in methods of projection, including projections of what people will decide to do in various scenarios.
What I do know is we can’t hide our heads in the sand again. Never again. When we have strong Bayesian evidence that something is happening, we need to work through that and act accordingly. Not say “there’s no proof” or “we don’t know anything yet.” This isn’t about proof via experiment, or ruling out all possible alternative explanations. This is about likelihood ratios and probabilities. And on that front, as far as I can tell, it doesn’t look good. Change my mind.
The short term outlook in America has clearly stabilized, with R0 close to 1, as the control system once again sets in. Cases and deaths (and test counts) aren’t moving much. We have a double whammy of holidays about to hit us in Christmas and New Year’s, but after that I expect the tide to turn until such time as we get whammied by a new more infectious strain.
Instead of that being the final peak and things only improving after that, we now face a potential fourth wave, likely cresting between March and May, that could be sufficiently powerful to substantially overshoot herd immunity.
Let’s run the numbers.
The Numbers
Predictions
Last week’s prediction: 13.1% positive rate on 11.5 million tests, and an average of 2,850 deaths per day.
Results: 13.7% positive rate on 10.7 million tests, with an average of 2,677 deaths.
We didn’t test substantially more people than last week, and the positive test percentage didn’t fall much, and the death rate didn’t rise much. Everything in a holding pattern. Some of that could be pending Christmas issues. Fool me twice, and all that.
This next week is Christmas plus New Year’s, so reporting issues are inevitable. Another week for wide error bars.
Prediction: 13.6% positive rate on 10.1 million tests, and an average of 2,500 deaths per day.
Note that I expect the deaths decline, at least, to be about reporting rather than about actual numbers, which I don’t think will start to decline much for a bit longer.
Deaths
Date | WEST | MIDWEST | SOUTH | NORTHEAST |
Oct 22-Oct 28 | 895 | 1701 | 2208 | 612 |
Oct 29-Nov 4 | 956 | 1977 | 2309 | 613 |
Nov 5-Nov 11 | 1089 | 2712 | 2535 | 870 |
Nov 12-Nov 18 | 1255 | 2934 | 2818 | 1127 |
Nov 19-Nov 25 | 1761 | 4169 | 3396 | 1714 |
Nov 26-Dec 2 | 1628 | 3814 | 2742 | 1939 |
Dec 3-Dec 9 | 2437 | 5508 | 4286 | 2744 |
Dec 10-Dec 16 | 3278 | 5324 | 4376 | 3541 |
Dec 17-Dec 23 | 3826 | 5158 | 5131 | 3772 |
Death rates didn’t rise all that much due partly to the decline in the Midwest, but they are still up in the other three regions and substantially in the West and South. It seems clearly a few weeks too early to hit peak deaths based on when we had peak infections.
Going forward, the vaccine will start to be effective for those who get infected next week, so to the extent we are protecting residents of nursing homes, we’ll see that effect in the death rate start to be noticable in late January. I expect deaths to be in decline by then.
Positive Test Percentages
Percentages | Northeast | Midwest | South | West |
10⁄22 to 10⁄28 | 3.68% | 9.87% | 8.58% | 6.46% |
10⁄29 to 11⁄4 | 4.28% | 12.79% | 8.86% | 7.04% |
11⁄5 to 11⁄11 | 5.56% | 17.51% | 9.89% | 8.31% |
11⁄12 to 11⁄18 | 6.99% | 18.90% | 11.64% | 10.66% |
11⁄19 to 11⁄25 | 7.00% | 16.62% | 10.41% | 11.75% |
11⁄26 to 12⁄2 | 8.38% | 17.90% | 12.45% | 12.79% |
12⁄3 to 12⁄9 | 10.47% | 17.94% | 13.70% | 12.76% |
12⁄10 to 12⁄16 | 10.15% | 15.63% | 15.91% | 13.65% |
12⁄17 to 12⁄23 | 9.88% | 14.65% | 15.78% | 13.82% |
Quiet on all fronts. Midwest clearly in slow decline, looks like other regions also ready to follow, as soon as we get past the holidays. There’s also an overall one-time bump coming of unknown size, but after that we should be clear until the new English or South African strain becomes a problem.
Positive Tests
Date | WEST | MIDWEST | SOUTH | NORTHEAST |
Oct 22-Oct 28 | 94983 | 181881 | 158123 | 57420 |
Oct 29-Nov 4 | 112684 | 252917 | 167098 | 70166 |
Nov 5-Nov 11 | 157495 | 387071 | 206380 | 108581 |
Nov 12-Nov 18 | 211222 | 452265 | 255637 | 150724 |
Nov 19-Nov 25 | 269230 | 435688 | 294230 | 170595 |
Nov 26-Dec 2 | 256629 | 357102 | 294734 | 185087 |
Dec 3-Dec 9 | 354397 | 379823 | 368596 | 263886 |
Dec 10-Dec 16 | 415220 | 315304 | 406353 | 260863 |
Also, this seems like a nice graphic:
But it’s also potentially misleading, because the increased cases in the deep South are mostly increased testing. The California increase is largely real. It would make sense that California would be the place that has stalled its crisis for the longest, between not having cold weather and imposing draconian restrictions all year, and so perhaps it’s finally time for them to face the music.
Test Counts
Date | USA tests | Positive % | NY tests | Positive % | Cumulative Positives |
Oct 15-Oct 21 | 6,461,028 | 6.4% | 865,890 | 1.2% | 2.52% |
Oct 22-Oct 28 | 6,943,470 | 7.5% | 890,185 | 1.4% | 2.67% |
Oct 29-Nov 4 | 7,349,648 | 9.5% | 973,777 | 1.6% | 2.89% |
Nov 5-Nov 11 | 8,285,878 | 10.7% | 1,059,559 | 2.4% | 3.16% |
Nov 12-Nov 18 | 9,033,621 | 12.4% | 1,155,670 | 2.9% | 3.50% |
Nov 19-Nov 25 | 10,415,393 | 11.8% | 1,373,751 | 2.9% | 3.87% |
Nov 26-Dec 2 | 9,741,057 | 11.7% | 1,287,010 | 4.0% | 4.22% |
Dec 3-Dec 9 | 10,458,644 | 13.9% | 1,411,142 | 4.9% | 4.66% |
Dec 10-Dec 16 | 10,694,845 | 13.8% | 1,444,725 | 4.9% | 5.11% |
Dec 17-Dec 23 | 10,710,356 | 13.7% | 1,440,770 | 5.1% | 5.56% |
This is the silent scandal no one is talking about. Why are we no longer expanding testing? It seems clear now that our capacity hasn’t been expanding in December. It’s clear that demand greatly exceeds supply, and that more testing would be a huge help. When things were improving slowly, at least they were improving. Now it looks like we are stalled out, well short of where we need to be. Vaccinations are important, but until we get a lot farther along on those, so are tests.
Covid Machine Learning Projections
Machine learning projections say infections have been static since about November 25, which mostly matches the testing data. We can assume that the projections will keep saying similar things for the next two weeks.
Their predicted total infected is up to 19.2% on December 9, stabilized at 623k new infections per day. The total is up from 17.9% on December 2. As a reminder, I consider these lower bound estimates.
The immunity effects here compound fast. Even if you assume people get infected completely at random, going from 17.9% to 19.2% immune reduces R0 by 1.6%, which reduces infection levels by that much every five days or so, or 9% per month, and we’re introducing that effect permanently each week. After a month of this level of effect, you’ll see a 16% decline from newly immune people alone. After two months, infection levels would be cut in half.
Selection of who gets infected makes the effect bigger, and also we get to add vaccinations.
Of course, the control systems ensure it does not work that way, as people will notice things improving and take more risks, but it’s worth noting that things will start rapidly getting better if we can only hold onto our current levels of prevention, and let immunity from all sources do the job from there.
Europe
Going to use short-dated graphs to improve readability. If you want the longer view you can get it at OurWorldInData, or previous weekly Covid posts.
Sh
The positive test percentages chart is so incomplete and all over the place that I’m going to stop posting it, but you can go to the source if you still want it.
Deaths in Europe continue to run close to those in the United States, suggesting the Europeans are finding cases less often than we are, or have worse medical care or are worse at protecting vulnerable populations.
Then there’s that United Kingdom graph going rapidly vertical in infections. Turns out, there’s a reason, and it’s not that they lifted their restrictions…
The English Strain
The big news this week is that England has identified a new strain of Covid-19 that is ‘up to 70% more infectious’. The new strain dominates in southern England, including London, and the graphs tell a rather clear story.
Oh no.
You’re probably wondering the same question I was when I read that, which is that we know that ‘up to’ means we’re not willing to commit to anything at all (did you know I am up to 15 feet tall? It’s true!), but what the hell does ‘70% more infectious’ mean?
It could mean a lot of different things.
To me, there are two natural hypotheses for what it means.
One sensible definition of this is that 70% more people get infected each day, so it raises R0 by that percentage. If previously things were stable, 70% more infectious would cause infections to rise 70% each serial interval, which I’ve been approximating at about five days. So if it was this, things would about double each week.
Alternatively, it could mean that any given physical interaction was 70% more likely to infect you. This seems unlikely to be it, because how would anyone know what this value was, but it still makes at least some intuitive sense and has some practical value. So if before, if you went home for Christmas and someone was infected, you’d have a 10% chance of getting Covid-19, now that number is 17%.
The difference between those two is that if you get exposed multiple times, you can only get infected once, so the first scenario is a bigger jump in cases than the second one. Depending on how much ‘overkill’ you think takes place when people get infected, the difference could be big or it could be small.
As a third option, it could mean any given exposure is effectively as if you were exposed to 70% more virus. Chance of catching the virus is non-linear with viral load, so in some ways this is a more than 70% increased risk (if previously load was below threshold to get infected, and now it isn’t), and in other ways it could be less once you get to the other end of the curve. This also changes the distribution of initial viral loads, in ways that might be good or bad for outcomes and death rates. If you are reliably getting higher initial loads, that’s bad. If you are getting infected despite low loads then, given that we know you’re infected, that’s good, perhaps quite good.
What I definitely didn’t consider was the possibility that this was measuring the length of the doubling time because that’s not remotely a fixed number and using this doesn’t make any sense and arrrrrgh and then I saw this on this post:
What is still driving me crazy is that CellBioGuy presumed that this was what they meant. I mean, I certainly hope it is what they meant in terms of ‘that physical property of the world would kill less people’ but I can’t help but notice it would be completely insane in a way that even my model of predicted general insanity isn’t handling well. The model isn’t even handling the presumption by CellBioGuy very well here.
So it turns out that CellBioGuy was wrong here, and this refers to the sensible thing of “percent rise in infections each cycle” R0 thing:
(Do check out the rest of that comment by CellBioGuy anyway – even though the presumption in question turned out to be wrong, the rest of the comment has a lot of good gears-level details on various issues, but it’s too long to put here in full. I’d like to know how well the rest holds up.)
So on the plus side, statistics are being reported in a way that is relative sane.
On the minus side, this seems rather like it can be summed up as: We’re fucked, it’s over.
This is estimated as a 65% increase in infectiousness. If we want to stabilize infections in an area that was previously stable we’d need what would previously have been an R0 of about 0.6. If you have an R0 of 0.6 that means you would have previously been cutting infections in half each week or so.
Does that sound like something any Western country could possibly accomplish from here? What would even trying to do that even look like? Is there any chance people would stand for what was necessary to do that?
And that’s only what it takes to get a holding pattern.
Under such dire circumstances, a phase 4 lockdown has been invoked. What does that mean? Glad you asked, here are the guidelines.
The missing restrictions that stick out are not shutting down houses of worship, and allowing people to move house. Also funerals can go up to 30 people, whereas weddings are capped at 6. Also ‘support groups’ can meet up to 15 people and I don’t see anything saying it needs to be outdoors.
Whereas essentially any social contact of any kind is forbidden (e.g.: “You cannot meet people in a private garden, unless you live with them or have formed a support bubble with them”), your ‘bubble’ is highly restricted in its sizing, and you’re not allowed to be outside without a ‘reasonable excuse’ although those include groceries, going to the bank or exercising.
So basically, if you’re outside where it’s safe, they’ll harass you and maybe worse. Whereas if you stay inside, technically it’s not allowed but in practice it’s a lot less likely anything happens to you, unless the anything in question is ‘you catch Covid-19.’ The rules are porous enough that they aren’t enforceable against the things that are risky but enforceable enough to shut down the relatively safe actions that keep people sane. And with weird exceptions for remarkably large indoor gatherings for certain events that are textbook superspreaders.
All of which is what our model expects to see, and none of which seems likely to be remotely sufficient if the new strain is as infectious as they estimate.
The strain has already been seen in several other countries. Flights between the United States and the United Kingdom have not been shut down. Many European countries are shutting down some travel, which will slow things down a bit, but headlines like this one…
…illustrate that slowing things down is all that’s being aimed at. Which is good, because it’s too late anyway. There would not be any drivers to test if this was a real attempt at containment.
If the estimate of 65% more infectious is correct: The strain doubles every week under conditions where other strains are stable.
My father sent me this video (24 min) that makes the case for all of this being mostly a nothingburger. Or, to be more precise, he says he has only low confidence instead of moderate confidence that the new strain is substantially more infectious, which therefore means don’t be concerned. Which is odd, since even low confidence in something this impactful should be a big deal! It points to the whole ‘nothing’s real until it is proven or at least until it is the default outcome’ philosophy that many people effectively use.
Note that he also suggests the new strain is likely to be less virulent, and make us less sick, which could also be part of why it’s more infectious. If so, that’s great news (I can think of a scenario where it is actually bad news, but it’s an unlikely corner case).
He also points out correctly that a lot of nations don’t do much sequencing, so we should assume the new variant can’t be contained to England at this point. Doesn’t mean we shouldn’t try in order to slow it down, but such efforts will still fail.
The video seems strong on the scientific details, and the speaker strikes me as an excellent explainer/teacher, which is why I’m willing to link to it.
Alas, as is often the case with academics that are good at learning and explaining scientific things, the epistemology is bonkers. His core argument is: “You cannot use epidemiological data to prove a biological property.” With a side of Covid-19 being spread mostly by super-spreaders (true) and thus the new variant could be winning at random.
All of which is not how knowledge or Bayes’ rule works. It’s not how any of this works.
There is a valid point here, of course. Relying solely on the numerical growth of the strain or of infections in England generally, without looking at the context, doesn’t provide that much evidence. There are often other explanations. And his points about mutation in and of itself being commonplace and mostly harmless are well taken.
That doesn’t change that evidence is evidence, and a likelihood ratio a likelihood ratio. Experiments are not some special class of thing that are the only way one can make predictions or assign probabilities, and it’s weird that people can be so good at academic scientific thinking while not understanding this, and in fact it seems that when we train people to do academic science we also train them to not think about other information and to be careful not to use Bayes’ rule and to ridicule anyone who tries to use non-academic information in order to know things.
With that in mind, we look at the evidence and think about possible explanations, and mostly I find that there aren’t other plausible ones worth assigning much weight to.
Let’s start with those charts above. They aren’t quite this:
But then again, when you consider the context of England being under lockdown conditions that had previously turned the tide, and that have stabilized the situation elsewhere in Europe…
And combine it with the share of infections from the new strain from the earlier chart, and work out what those combine to imply…
This definitely does qualify under “hot damn, look at this chart.” This is a huge, dramatic increase in infections happening very quickly. A doubling in one week.
Note also that the warnings went out to the public on December 19. The out-of-sample data from the next few days strongly reinforce the hypothesis that we’re screwed.
The other plausible causes of such a rapid rise are not present. England didn’t suddenly relax its conditions this much. The law of large numbers is more than sufficient to make me very dismissive of ‘random chance via superspreader events’ as an explanation. How big are these superspreader events?
If my understanding of the situation is correct, there is only one conclusion:
This variant cannot be stopped short of mass vaccinations. It is not going to be stopped short of mass vaccinations.
All that is left is a holding action. Realistically, we can’t make enough of a dent to turn the tide of cases of the new strain until at least May. That’s about twenty weeks from now. That’s twenty doublings. So for every case that’s escaped to the United States so far, we can expect (does quick math) a million cases. Then another few doublings in June and July.
I do think we can do better than that, because it appears the tide is starting to turn now on the old strain, we’ll get a bunch of incremental help from increased immunity along the way, and control systems will set in.
But mostly, it seems like if you have vaccines and people who don’t want to die, you might want to hurry. I’ll get back to gaming the scenario out in the conclusion section.
Although Zvi’s overall output is fantastic, I don’t know which specific posts of his should be called timeless, and this is particularly tricky for these valuable but fast-moving weekly Covid posts. When it comes to judging intellectual progress, however, things are maybe a bit easier?
After skimming the post, a few points I noticed were: Besides the headline prediction which did not come to pass, this post also includes lots of themes which have stood the test of time or remained relevant since its publication: e.g. the FDA dragging its feet wrt allowing Covid tests, which is somehow still the case a year later; governments worldwide utterly failing at buying capacity or otherwise incentivizing more production of vaccines or tests; endorsing an analysis that Covid spreads via aerosol transmission; etc.
I do have one outsized nitpick with this essay, however. The headline “We’re F***ed, It’s Over” is alarmist, which is in principle fine given its dire prediction, and might make sense to regular readers of Zvi’s blog. But once such a post is shared more widely, sometimes the only thing a reader sees is the headline, and it seems to me like this headline cannot, and will not, be properly understood by those who do not know who Zvi is, and how he reasons. Even on LW, many people will not read the entirety of such a long essay, and might lack the context to understand the headline.
Regarding that context:
This perspective makes sense. In these terms, the original prediction suggests shifting one’s strategy towards folding or losing gracefully, rather than fighting a losing battle. But that’s because this pandemic was never an x-risk. We could afford to play this game of containing the disease (possibly unfortunately so, as argued in this comment chain), whether we had a realistic shot at victory or not. But game over, in this context, is indeed merely meant as a game loss, by someone who professionally played card games with plenty of variance, and who knows that folding is an entirely valid strategy in such a situation.
But that’s not how I expect most people to interpret that headline? Even now, part of me interprets it as claiming “we’re all going to die”. And if, for example, a sufficient number of x-risk researchers wrote the same headline, that interpretation might be entirely accurate.
In this perspective, there’s only one real game we play, and that game must be won. As one fictional character put it:
To conclude, I understand that these posts are written at a speed premium, and would not complain about a random sentence like this; but as a headline of a >270-karma post, this is rather suboptimal.
Focusing on the Alpha (here ‘English Strain’) parts only and looking back, I’m happy with my reasoning and conclusions here. While the 70% prediction did not come to pass and in hindsight my estimate of 70% was overconfident, the reasons it didn’t happen were that some of the inputs in my projection were wrong, in ways I reasoned out at the time would (if they were wrong in these ways) prevent the projection from becoming true. And at the time, people weren’t making the leap to ‘Alpha will take over, and might be a huge issue in some worlds depending on its edge spreading and how fast we vaccinate’ at all.
We also saw with Omicron how, when the variables turn out differently, we do see the thing I was pointing towards, and how people are slow to recognize that it might happen or is going to happen. I do think this had the virtue of advancing the understanding of what was plausibly going to happen. If it overshot a bit in terms of how likely it had its core predictions coming true, that’s something to improve going forward, but very much a ‘man in the arena’ situation, and much better than my ‘not be confident so say little or nothing’ approach that I shared with most others in Jan/Feb 2020.
To what extent this justifies inclusion in a timeless list is up for grabs, but I think it’s important that the next time we notice something like this, we speak up fast and loud (while also striving for good calibration on the chance it happens, and its magnitude)