Interesting evidence indicating that R0 might actually be lower than Delta but with some immune evasion that allows it to spread in areas previously hit by Delta:
I'm not really sure R0 is actually a meaningful thing to run a prediction market on. The problem is, when you dig into how R0 is calculated, it's not really a precise or well defined value. In theory its definition is simple enough but in reality, it has no biological basis - claims that R0 is this or that are always coming from models. And when you look at what the models do, well, R0 turns out to just be some kind of fudge factor or arbitrary coefficient used to make the model roughly match the data reported so far.
As a consequence I've seen models that yield totally different values of "R0" for the same virus on the same dates, simply because the model was given case curves from different cities. As such it cannot be genuinely said that this represents anything genuinely true about the virus itself. It's not like it's derived from RNA or careful lab experiments or anything like that.
There are a lot of things in the world worth making predictions about which we can't nail down with the kind of scientific precision we would like. The solution is, rather than not trying at all, to just pre-specify as precisely as possible the method by which we will make the assessment in the end.
The fine print of the Metaculus prediction says:
"This will resolve on the basis of the mean R0 that is estimated for the Omicron variant according to the first credible systematic review that estimates this value. Such a systematic review should be peer-reviewed and should incorporate R0 estimates from at least 3 studies."
Metaculus users aren't being asked to predict "what is the R0 of Omicron", but rather, "what will the estimate of the first credible systematic review of R0 studies be?" And this is close enough to being the same question that it's still useful to us.
Yes, the prediction market can be successful on its own terms if it's just trying to predict what academics think R0 is. That isn't quite what I meant by being a meaningful thing to run a prediction market on though. I'm just really skeptical the resulting value is actually useful in any way.
It might be more meaningful to ask what the Rt of Omicron is, rather than the R0 of Omicron. R0 is the number of people each person infects *assuming no public health intervention.* But there are public health interventions!
Rt also has the advantage that you can estimate it with fairly minimal modelling assumptions. You need to know the interval between infections, or at least a distribution describing the serial interval. But aside from that, you can estimate it using only case counts.
I don't see how that is more meaningful, though. Whether you're measuring number of people infected without interventions or with interventions, the value you get is going to be highly dependent on how much time the people in the area are spending breathing each other's exhales. Do a thought experiment: Town of 10,000 people, very small and congested, large families squashed together in small dwellings. Let's say there are no public health interventions. One person with the virus enters the town and we track transmission for a while, and using that info we calculate how many people the average infected person infects. Now imagine the same 10,000 people, still no public health interventions, but now the people are living all spread out over the prairie. Some people live 5 miles from the next building. No real community gathering places except a coupla general stores. No schools, even. Introduce one infected person. Virus is going to spread much slower, right? R0 is going to be way smaller. And if instead there are public health interventions -- like say 50 % of the population is vaccinated, 75% wear masks around people they don't live with -- the Rt is still going to be way higher in the congested town than in the prairie community. Neither R0 nor Rt is a measure that is independent of how much face time the members of the population you are studying have with each other. The value you get is going to vary GREATLY depending on, basically, how many cubic feet of their neighbor's exhales an average citizen of the community is breathing per day.
That is why it does not make sense to talk about the R0 of a virus variant, as though it is a property of the variant that is the same at all times in all places. A virus with a R0 (or Rt) of 6 in the congested town setting might have an R0 (or Rt) of 2 on the prairie. In a setting where everyone never get closer than a mile away from another person both R0 and Rt would be zero..
As you observe, Rt boils down to a fancy derivative of a case curve, which means such a market would have to specify the exact time period and time series (e.g. city, country) over which the prediction will be made.
The problem here (and with almost all prediction markets) is you aren't predicting what R0 is, you are predicting what academic journals say R0 is. If you believe there is absolutely no bias in academic publishing then "what academic journals say" may be a reasonable proxy for "what is reality". If you believe there is a massive amount of bias in academic publishing, then "what academic journals say" may be totally unrelated to what is reality. For things like sports reporting, the reporting source is almost always a very strong proxy for actual results.
This knowledge of who is reporting tends to present itself in the prediction market quite well. A hypothetical market participant could believe that the *real* R0 of Omicron is negligible, but they may simultaneously believe that the academic journals will strongly prefer papers that have models that insight fear (more clicks/buys) and thus they may bet that R0 will be very high for this market.
Unfortunately, prediction markets are absolutely terrible at finding truth for this exact reason. They can do a very good job of predicting how some oracle will report in the future, but that is not necessarily correlated with what is true/real.
Insider trading alone would be a massive problem if markets that bet on "what academics say" ever became worth significant real money. After all lots of people know what a meta-analysis is going to say before it gets published.
R naught is a property of the virus in a given setting. It's dependent on the virus's intrinsic transmissibility and also on how much time people in that setting log hanging out breathing each other's exhales. It depends on patterns of living, working, and socializing, and also on precautions such as masking and ventilation. No matter how transmissible a virus is there would be no transmission in a setting where each person never came within a mile of anyone else. R nought would be zero.
So yeah, talking about R naught in general is like talking about the risk of somebody getting run over. It's highly dependent on how much time the person logs crossing streets and how many cars are in the streets he's crossing..
Still, if one specifies a locale, say Brooklyn, it would be meaningful to talk about the R naught of virus X in that locale. (It would vary some depending on weather, part of Brooklyn and some other factors, but probably not all that much.)
R0 in theory is an intrinsic property of the virus, before adjustment for other factors like population density, weather, etc.
In practice epidemiologists have no way to actually measure this, so they end up making claims about R0 that are derived from pet models that are being curve fitted to the published PCR testing numbers. As such claims about R0 can easily conflict in major ways: it seems to bother nobody.
Can it be said R0 is a real thing at all, given the total inability to measure it or precisely define what it means? After all, as you astutely point out, it doesn't really make sense to talk about transmissibility in the absence of an actual context as if nobody was around and you were sealed in a vacuum jar, transmissibility of your viruses would be zero regardless of what the viral RNA was.
As such I am very dubious about all claims around transmissibility of viruses emanating from professors of epidemiology. It's just not a sound field. Values for R0 are usually on close inspection actually values for Rt that have been fiddled with a bit, where Rt is more or less some derivative of the case curve. Some models even take R0 as both an input and yield it as an output! Have fun figuring that one out ....
I've always been told that R0 is the average number of people an individual infected person infects, before recovering, in a hypothetical population where no one has any acquired immunity. That doesn't sound like an intrinsic property of the virus, unless you have a background assumption that fixes the behavioral and environmental features of the human population you're talking about.
The fact that they use the names "R0" and "Rt" is suggestive that the classical model *does* assume fixed human behavior and environment, with acquired immunity being the only relevant parameter that varies with time to change the reproductive number.
I mean, it's not dumb for scientists to use models that involve known false idealizations. Sometimes we can know that the idealizations are false in ways that don't make much difference (as when we assume that the sun and the planets are the only gravitational bodies in the solar system and there is zero friction as they move, or when we treat the ocean as homogeneous and infinitely deep in modeling the generation of waves by wind storms) and sometimes even when we know the model is going to be bad at making predictions, it can still be useful for helping us to understand the types of impact different factors can have (for instance, we treat a pool table as frictionless to understand how hitting the cue at different angles will send the other balls in different directions, and we treat an economy as homogeneous and well-mixing to understand how changes in supply, demand, taxes, and interest will affect inflation).
Yes, I understand about the need to simplify things to make useful models. But the simplification baked into the definition of R0 seems so radical that it interferes with the utility of the measure. R0 purports to be a measure of how many people an infected person will infect, but in fact number of infections passed on by an infected person depends on the environment and people's behavior in it, and there is a LOT of variability in those things across different settings. It's like treating pool tables as frictionless when some have smooth glossy surfaces and others are rough and gravelly and others are sticky with maple syrup.
Maybe it would be more useful to have a contagiousness measure where the setting is part of the definition: Number of people an infected person infects in a city of a certain population density. We could then accumulate a body of info about how variations in population density and other factors affect contagiousness in other settings. How dense is my setting compared to the setting on which the contagiousness measure is standardized? Use that as a multiplier . . .
Where nobody has any acquired immunity, where prior/adapted immunity doesn't exist, where everyone recovers at the same rate and has an immune system of identical strength, etc.
Basically, how many spherical frictionless cows are infected by another spherical frictionless cow in unit time. It's an interesting thought experiment but not a real number that can actually be calculated and have any meaning.
Well, IQ is a pretty abstract scale. IQ itself is a real thing on its own terms because it has high test-retest consistency, the methodologies for determining it are standardized, it seems to correlate well with other "intelligency" things etc. At least I believe those things are true - I'm no expert in IQ though. The problems here start when you begin claiming IQ is the same thing as generalized intelligence, which of course is a common but much more controversial claim.
R0 is on much less firm footing than IQ. There's no standardized methodology to determine it, nobody even seems to agree on the precise definition (see the multiple similar-but-different attempts in this thread), it's not even clear there can be such a definition, and different academics routinely calculate totally different values via totally different methodologies for what is supposedly the same thing about the same virus.
Given that savant abilities tend to correspond with other "deficits", the general intelligence thing is tricky indeed.
IQ tests certainly dont measure certain types of intelligence- social intelligence, organizational intelligence, creative intelligence, others.
The things IG tests do measure might be correlated with those things, but thats a complicated question.
I have thought about someone from a western country who was taken into a nomadic hunter gathere society. Say they had some sort of intelligence test for the sorts of intelligence that was useful in their living conditions. Would high IQ scores correlate? I don't know.
I understand there is something sort of like global intelligence, but one has to define it carefully lest it covers modal nodes of intelligence it actually does not.
Came here to say sth. similar. You need to define at least the location to make a meaningful statement about R0.
Once you do that, I think it makes sense to use this, as the R0 of different viruses (under similar conditions of population density and habit) is so different; so it allows you to compare in a given setting.
Probably also a reflection of the health system's detection capability, to identify infected people, so the 'measured' R0 in a city in Europe (such as London) could have a different 'measured' R0 than a city with similar characteristics in the USA (such as New York).
These two questions bother me. But writing good questions is hard. Prediction markets aren't supposed to address that problem, so I'm not sure I'm justified in being bothered. (Unless prediction markets are supposed to address that problem by asking more objective questions. Which they are, but that's not the only thing they're supposed to do.)
The second question on deadliness bothers me less. It seems like a real-world practical question that people want to extract from the literature, so it makes sense to take the literature as given. I don't like the particular formulation, but that's the basic problem of asking questions. Whereas, I feel like the question about R0 is more like doing science and it shouldn't be taking the literature as given, but should be pushing for predictions of more concrete things. Eg, the ratio of R of the strains in the existing population instead of the absolute R of omicron in a naive population (which is also the practically relevant question).
R0 might be impossible to calculate accurately, but the ratio between R0s of different variants is much more well defined and could be possible to calculate. So conditional on some value for R0 of the base variant, the R0 of another variant could be predicted.
Agree there's no consistent methodology to measure R0. Here's an idea for measuring the relative R0 of 2 variants: How rapidly does one take over? Once variant Herbie starts showing up in a population where all the infected are hosting variant Batilda, how long does it take for Herbie to be the variant found in 80% of tests? Of course in most cases things will be more complicated -- there will be more than 2 variants circulating. But the math would still be doable. Maybe you could do it sort of the way chess rankings are done --everybody's ranking is a measure of how hard to beat they people they beat are, and how consistently they beat them.
I think that sort of metric might make sense, although of course tremendously confounded by how much sequencing a country does. I believe the UK is responsible for half of all GISAID SARS-CoV-2 uploads for some reason.
Well, I'm not sure the differences among countries in sequencing would be a confound -- if by confound you mean a variable that distorts results. A country that doesn't do much sequencing would give us a skimpy data set, but not a distorted one -- right? Or is there something I'm not thinking of?
Perhaps confounded was the wrong word. The problem is that most countries do very little or even no sequencing, and there are gazillions of variants. It's not true that there's Alpha, Delta, Omicron etc. That's more or less a press fiction at this point. Look at GISAID to see the true family tree. So you get a lot of .... variants, strains, mutations, whatever you want to call them ... that grow rapidly but then top out at a few percent and don't go further for some reason.
So they'd never reach 80% and thus this figure could only be calculated retroactively. But the only reason people care about R0 is because fraudulent pseudo-scientists like epidemiologists claim it's very important and a high R0 number means "exponential growth" and thus P.A.N.I.C. (mathematics needs to be put in a care home for abused children after all this). As a consequence you'd end up calculating a lot of very dubious growth values for tons of variants that look initially scary and then two weeks later their growth tapped out for some reason, and you'd be doing it on like 10 data points in the best cases.
Overall I find myself thinking R0 is a concept that should just be humanely put to sleep, like the rest of the junk field that spawned it. Even if it had some actual precise meaning and could actually be calculated, so what? What's the actual actionable outcome of knowing such a value? Well, there isn't one. It's not true that high R0 values = PANIC because you'd still be left in a situation where you can't model how long the virus will grow for, or how serious it is, and even basic questions like "what is the actual capacity of our hospitals" turns out to be mired in nonsense and mis-understandings (by epidemiologists!). So it can't tell you anything and it ends up with stuff like what we see today, where the doctor who discovered Omicron points out that the person who had it didn't even realize they had "COVID" at all (because they didn't, the symptoms aren't even close to a match) yet because of years of misinformation about exponents and R0, scientists and politicians immediately panicked.
R0 is in the end highly reflective of the profession that spawned it: fuzzy, worthless, used only to create unjustified panic.
Agree that we are all walking around with a sense of the relative contagiousness of variants, but how do we arrive at that? My sense of variants' contagiousness is based on the accumulated headlines I've read. It's sort of an average -- well, a weighted average, because I trust some headline-writers more than others. And I'm not proud of that. Do you have a better calculation to propose?
The test you proposed in the other comment is roughly what I had in mind, with adjustments for level of immune escape, and assuming an equal serial interval between variants, but these can be measured independently.
Many seem to believe there's a case for omicron being *less* lethal than delta, based on near-zero-evidence throwaway quotes being shared widely on social media, best I can tell, out of wishful thinking. A metaculus question opened on this here:
I wish I was smart enough to know wether or not this article means I should be more scared or not. (I'm assuming....yes? Be more scared?). I'm boosted. Does that help or we don't know? #YourDumbestSubscriber
You should be proportionally more scared. Existing vaccines likely work poorly against the Omicron variant, but "poorly" is not the same as "not at all". All that vaccines (and other preventative measures) can do is mitigate the risk of infection. So, if you were a hermit living in some remote cave in the desert, vaccines would be irrelevant to you, since your chance of infection is zero anyway. If you have a habit of licking strangers all day, then no vaccine would help you for long. But, assuming you're a normal person, you can alter your behaviour within reason to reduce your chances of infection. With the Omicron variant, it means that pushing your behaviour a little closer to "hermit" might be a good idea.
You sound surer than many sources I'm reading that existing vaccines will work poorly against Omicron. Also, before you tell somebody how scared to be, you need to have some info about other factors: How contagious is the sucker? How virulent?
Well, it has three known antibody-evading mutations (from experimental work on exposing mutated spike proteins to vaccinated/convalescent serum) but nowhere near as many antibody-evading mutations as that work needed to fully evade antibody response. (And of course even fully evading antibody response still leaves T-cell response which is much harder to test.)
So we should have high confidence (>90%) that Omicron erodes antibody effectiveness somewhat, and that it does not come close to eliminating it. (>99%)
I don't know what Bugmaster meant by 'poorly' exactly; it's one of those things that needs numerical precision to evaluate. I'd estimate top vaccine effectiveness will remain over 70%, but that's not really based on much data.
Some of this seems quite wrong. Most obviously, the idea that a vaccine wouldn't help you if you went around licking strangers all day. If anything, the vaccine would help you most in that scenario!
More debatably, but still seemingly wrong to me, is the idea that vaccines "likely work poorly" against Omicron. No one really knows yet, but the evidence we have suggests that they may well work about as well as they ever did in protecting against serious illness. Which is the most important thing. So that seems like an at best overconfident, at worst very inaccurate characterisation.
Data from Israel last night: Vaccines protect against Omicron about as well as they do against Delta. The unvaccinated are about twice as vulnerable to Omicron as to Delta.
My gut feeling was that vaccines would protect at most 50% as well against Omicron as against the vanilla variant (and more likely around 25%); looks like I was in the ballpark ?
There is an interesting article arguing that the models being used for Covid are wrong, mostly because they ignore the structure of a population, the fact that A is much more likely to interact with B and C than with X, Y, and Z who are interacting with each other. He argues that the actual pattern of what has been happening cannot be explained with the current models. One of his conclusions is that the newer variants may not actually be much more infectious than the older.
Except it exactly matches the real David Friedman's writing style, posting style, subject interest, economic positions and beliefs. It would be an amazing impersonater if true..... What would be the point? To save the real David from getting out of bed?
Maybe he hid it somewhere in a logo like with that N in the slatestarcodex blog (because it was the gap for making it a perfect anagram for Scott Alexander)
This is really good. I've seen some people make claims that population structure would affect things, but they've done so at the length of a tweet or a Facebook comment thread, and not with any of this relevant detail. I would like to see a bit more about how sensitive this model is to certain structures of connectivity across the graph, and how things like actual transmissibility advantage of one variant over another and actual mild effectiveness of NPIs would appear in this model, but it makes a lot of sense.
I'm a bit surprised to hear him say that people have been denying the importance of this kind of structure. But this is partly because as far as I know, nothing in the US has ever been done on the basis of calculations of R or transmissibility advantages - perhaps in Europe people were trusting the simple models a lot to calculate those things.
For what it's worth, I did some work about a decade ago on damage spreading in networks which is *somewhat* similar if you squint hard enough. The key finding is that the largest eigenvalue of an appropriately scaled adjacency matrix determines if you get percolation (i.e., a pandemic) or not.
Of course, the details are vastly different between what I worked on and disease spread, but I think some of the broad conclusions are probably applicable, which are that you can get significant spread when it's unexpected by naive models--that only look at average numbers of links--just by changing the connectivity patterns. In particular, assortative networks in which highly connected nodes are more likely to connect to other highly connected nodes will have damage spread more easily.
This was before the second wave so the argument about heterogenous connectivity is mixed in with claims about much lower herd immunity thresholds but the gist of it is there - the "cutting edge" academic models at the time assumed everyone was identical in every way. When people with actual standards and scientific ability started looking at them, they not only immediately noticed this problem (that somehow the field of epidemiology itself had spent decades not noticing), but also built alternative models that fixed it and realized it meant the original predictions were wildly off.
This doesn't seem like an even vaguely accurate summary of the state of epidemiology to me. Google "network epidemiology" to see papers, journals and textbooks going back at least two decades that attempt to build models accounting for network structure.
The Imperial model which was the focus of UK policy at the beginning of the pandemic was an agent-based simulation based on multiple factors (age, location, transport links, etc), and simulating contacts on a non-homogeneous network. Obviously some relevant information was left out (that's the nature of modelling), but the idea that epidemiologists had somehow just not noticed that homogeneous SIR models don't perfectly predict reality seems very clearly false.
I don't think that model had non-homogenous contact networks, which is what makes the big difference here, but I might be mis-remembering. It did do some basic age stratification and had different "R0" (or some sort of R) values for different "places".
I've never seen any admission in epidemiology papers that their models fail to predict reality. Although they produce new models all the time, it's not like some fields where there are open competitions, robust accuracy metrics etc. Actually I rarely if ever see any comparisons between modelled predictions and what really happens, which is why as a group they've developed a reputation for being constantly wrong and yet acting as if they weren't.
"Although more detailed model validation and parameter estimation using data from past pandemics should be a priority for future research, it will be impossible to predict the exact characteristics of any future pandemic virus."
As for the details - the model itself is agent-based. It predicts individual contacts of the simulated agents, assuming the contacts take place in different types of location (workplaces, schools, households, and 'other' if I'm reading it correctly). This necessarily results in a non-homogenous contact network (kids associate in schools, adults don't).
The idea that epidemiologists weren't aware of the limitations of models assuming homogeneous contacts is, as you imply in your first post, ludicrous and as such you should be sceptical of it. My general impression is that they have become significantly more aware of it over the past decade, and models have started to try to take it into account more and more (basically, as the computation power and data to make such models useful has become available)
Here's a paper from 2019 which deals with the question of non-homogeneous mixing in equine populations:
To quote the introduction of that second paper, from 2018: 'a full consideration of the often profound effect of heterogeneous contact structure on infectious disease dynamics has really only become widespread more recently, and especially since the adoption of methods often originally conceived in the context of social network analysis'
Sometimes epidemiologists choose to use the simple models anyway, because they're faster to compute, and accurate enough, but they (or at least some of them) very definitely are aware of the limitations.
I don't doubt that some epidemiologists overstate the case for some simple models, and I don't doubt that sometimes the models are used badly, but the generalisations that you're making seem extremely uncharitable and honestly, I don't think they're very accurate.
I don't mean a generic disclaimer of the form "of course, models aren't perfect". Every paper has those but who cares? What I mean is the actual scientific process:
1. Hypothesis.
2. Experiment or observations.
3. Refinement, abandonment or acceptance as theory.
Epidemiology seems to consist near-exclusively of (1). Where are these papers by Imperial College that dissect and analyze their previous failures, which have been numerous and enormous over a period of decades? There are none. Where is the admission that the model they put in front of governments for COVID was totally unvalidated, and unfit for purpose? There has never been such an admission.
Perhaps you think it's unfair to harshly judge the whole field on the basis of Imperial College. In fact I've seen lots of papers that are like their work, but, OK, other than John Ioannidis (who has been treated with incredible disdain for doing so), where are the epidemiologists condemning Ferguson's team and furiously demonstrating how they aren't all like that? Where are the demands that "scientists" who promote ridiculous, totally unvalidated models that produce incorrect predictions be fired or otherwise penalized? There's nothing because they're all at it and don't see anything wrong with it.
As previously mentioned, Ferguson's model is very concerned with R and R0. It took such a value as input. Most of this discussion is about the fact that no such value can be determined and even what it means is totally up in the air. These guys have been working on their models, publicly funded, for more than 20 years - the world can and should expect better than this especially given the horrifying impact these incorrect model predictions had (there was never going to be a single massive wave and they knew it because epidemics have never worked like that).
Yes, the Imperial model was complicated, but it was worse than a raw SIR model. Pretending that epidemiological study of networks doesn't exist is overly charitable.
I think that's a different argument, and one I was making in a less detailed form. If some people are more likely to catch the virus than others, for either biological or behavioral reasons, then the more vulnerable are being selectively infected, so the average vulnerability of the population is falling over time, so her immunity is earlier.
I agree that it would be interesting to study more systematically how the properties of the network affect the dynamic of the epidemic. This is a natural development of the work I did in this post. However, that's a lot of work and, in that post, I just wanted to present the basic idea and some results showing that what you would expect intuitively can in fact happen in simulations, as well as explain why I think it matters.
On your second point, this is something I already thought before and some of the reactions to this post have only strengthened this view, but I do suspect that the difference between how things went down in the US vs. in Europe explains in large part this difference of perception. I have been in Paris since the beginning of the pandemic and, in France, the kind of simplistic models I criticize have played a very prominent role in the justification of the government's policy, if not in the elaboration of that policy itself although it clearly had some impact on that as well.
I think modeling also played a major role to justify policy in the UK. From what I can tell, it was very different in the US, where apart from the cubic model fiasco — which didn't really matter anyway because the power to order restrictions to contain the pandemic belong to the states — I don't think modeling played a major role in political discourse. But in France and many other European countries it was very different.
I imagine this must be a lot of work! Thanks for posting some of the things on Github. I unfortunately don't have much experience with this kind of thing, but this post was so interesting that I'm wondering if it will be able to get me to try to teach myself how to use some of it over the winter break to see if I can get a bit of understanding.
Some questions that suggested themselves to me while reading (just in case one or more of these turn out to be easy to address, or if anyone other than you is reading who can work on these models), in addition to the ones about what happens if there is some slight effect of NPIs, or slightly different transmissibility of the variants:
How sensitive is this to the seeding mechanism? If seeding happens uniformly in all sub-populations and sub-networks, then you automatically get a transmission advantage for new variants. (New variants will be equally spread across all sub-populations, while old variants will be disproportionately concentrated in sub-populations that have a high current infection rate, and thus are lower transmission for all variants.) But seeding by new variants that have become common overseas shouldn't occur uniformly, but instead should have some patterns of their own. If we treat your model as a model of the *world*, with no seeding, rather than a model of a country with external seeding, then this effect might go away, since new variants will be more likely to appear in sub-populations that have more cases. However, if we ignore new variants until they account for 100 cases, this will select for paying attention to variants that, by chance, are present in sub-populations with high current transmission rates. Does this restore the effect?
In your model, the seeding switches from one variant to another, and the second variant ends up taking over much faster than one might naively expect. If both variants continue seeding equally, how often does it occur that the new variant ends up becoming more common than the old variant? Or in a model with no seeding, but just occasional random mutations into a new variant (without any transmissibility advantage) how often does the new variant end up taking over?
How small of a transmissibility advantage yields transmission advantages that seem qualitatively similar to the data we have for alpha and delta?
I really like the idea you have of explaining geographic correlations by having networks of sub-populations that don't line up with geographic networks. It fits with a lot of things I've been thinking for years (including many topics Scott writes about) - city center networks in different cities are often going to be more closely connected to each other than to the suburban networks immediately surrounding them, and blue tribe/red tribe/grey tribe networks will be connected to each other in different ways than they are connected across these lines within a geographic area. And anecdotally, I noticed that I had several friends (around the world) who were infected with covid in March/April 2020, but then none (as far as I know) until April 2021, when a friend in Michigan, a friend in Toronto, and a friend in Paris all got infected within a few days of each other. It would make a lot of sense if these people are all multiply connected through social networks that are geographically dispersed.
But it does seem to me that even if this sort of tree-like hierarchy of network and sub-population interconnections has some geographic dispersal, that there should also be a cross-cutting set of geographically local connections. It might not be too hard to add extra connections across sub-populations that are geographically local, but maybe it is. It would be interesting to see how sensitive some of the big results are to the tree-like structure where sub-populations are grouped within a network, if instead we have sub-populations in a bit more complex of a network (like a small-world graph, for instance).
This also raises a deeper question. The idea of a population as a well-mixed and homogeneous network is obviously very unrealistic. But conversely, the idea of sub-populations as being mostly self-connected with only occasional connections within a larger network, and only very rare connections outside that network, also seems quite unrealistic. The latter model might well be a very good model of what happens during a period of very tight "lockdown" (whatever that means to people). I think you did one investigation that shows that with enough connections across these sub-populations, you end up with something that looks very much like the homogeneous model, but it would be interesting (and presumably extremely difficult) to see what happens between. And it also makes me wonder what happens if we model "lockdown" with a switch between the two types of dynamic.
Yes, modelling was the single key factor that converted the UK from a relatively Sweden-style relaxed approach to harsh China-style approach. Once Ferguson made his graph of a single giant curve totally swamping all hospitals, it was over because nobody in government seemed to be aware of his team's numerous prior failures, and suddenly all sorts of academics popped up claiming ICL were the "best epidemiologists in the world".
Lots of people in the general public caught on within months that the man was a ribald liar - despite claiming some sort of apocalypse was inevitable even if everyone did lock down, he was immediately caught breaking lockdown rules to go sleep with his lover, who not surprisingly was a far left extremist (extinction rebellion). Clearly he didn't actually believe his own predictions and the nation found out why later on, much to its own cost.
My understanding of the rationale behind lockdowns was to reduce the homogeneity of social structure, to give us contagion dynamics that look more like the waves and less like the big bell curve. I don't want to take responsibility for anyone else's messaging on that point, but that's how I understood from the start: reduce the number of edges on the graph, and reduce their strength, as much as possible, to reduce the degree to which infection was GEOGRAPHICALLY correlated (since that's also how medical resources are distributed).
Journalists may have made a hash of the point but who cares?
We saw really quite decent evidence for at least some of this population-structure based spreading in the last UK wave: it produced a very significant peak in the teenage population, and a corresponding smaller peak in the age group corresponding to their parents. The younger and older adults saw much smaller growth.
Spin-off question: Would it be theoretically possible for a variant to be, for all practical purposes, a covid vaccination? So it would be a virus that kills about 1 person in a million, but mostly causes no symptoms or very mild ones -- and confers immunity to covid for the next few months to all who have been infected. And if that's possible, mightn't it be possible to tweak the variant so as to make it so hugely transmissible that everybody gets it, even the animals who can be infected by covid? If that happened, could we truly and permanently get rid of all variants except the benign one?
I'm sure it's very unlikely that such a virus will pop up, but still would like to hear from knowledgeable people whether such a sequence of events could happen if such a virus did appear.
My guess is that it is way beyond existing biotechnology to create a variant that (1) is mostly asymptomatic or mildly symptomatic and (2) confers immunity to existing variants of SARS-CoV-2 and (3) is at least as transmissible as Delta. Being transmissible requires hijacking cellular machinery to replicate itself, and that's the kind of thing that triggers the immune system and usually provokes symptoms.
Not a virologist, but it seems that you'd need to restrict infectivity to the upper respiratory tract (as opposed to systemic infectivity that COVID has) while keeping the spike protein structure largely intact, and the mechanism that you'd use for that need to be robust enough so that a few random mutations here and there would not cause it to become systemic again. Even if it were possible, it seems very easy to make matters worse while developing such a mechanism and having an accidental lab leak.
At that point why not just have vaccines? I guess the answer is (i) vaccines you have to go out and physically jab people, and (ii) some people don't like vaccines. On (i), fair enough but at this point the infrastructure required to physically jab people isn't the main bottleneck, and on (ii) I don't think the people who dislike the vaccine will be happier with "it's like a vaccine but it spreads through the air so now you CAN'T avoid it!"
I'd also be worried that it mutates into something deadlier, apparently that can happen a lot!
I think you'd still have to go physically jab people - as contagious as COVID is, it's nowhere near contagious enough to give the high % coverage you want by natural spread.
It's sort of amazing that even at the low points of the vaccination program, there have been over 700,000 people getting vaccines per day in the United States, while the highest recorded number of infections was about 300,000 per day in the middle of last winter's wave. There are surely more actual infections than the recorded number, but probably since March or April, there has likely not been a single week where more people were infected than got vaccinated.
I think that at least some consequences of the virus (sneezing being a prime example) serve to enhance infectivity. So you wouldn't be able to both perfectly maximize infection and minimize damage. My guess is that the more harmful parts of COVID would somehow also enhance infectivity (since for example asymptomatic cases seem to have lower infectivity rates from what I last read, but that may be out of date), to the point that it would be difficult to make a low-damage competing strain.
This is sort of how the first vaccinations worked: using cowpox as a vaccine for smallpox. So like, it's *possible* for such a thing to happen by chance.
But we had to manually infect people with cowpox, it didn't outcompete smallpox on its own. And cowpox only had to be safer than smallpox, which is a much lower bar to clear.
Yes, it'll hopefully evolve into just another nasty cold within a number of years. no, deliberately exposing the population to infectious diseases is a bad idea, even if you think they're pretty innocuous.
A couple of people have mentioned the original small pox vaccination. But a better match for this is the oral polio vaccine which is an attenuated virus that is capable of (at least in limited circumstances) of spreading from person to person. This helped increase vaccination coverage in difficult to reach areas. But it can also mutate to more severe forms (see vaccine derived polio, though note is still typically milder than wild type polio) and at one point such forms were the dominate circulating strains in certain areas. As polio has gotten closer to eradication this vaccine gets phased out in favor of the injected killed virus form.
Virulence is generally a side-effect of infectiousness. Cow pox and traditional attenuated/live vaccines are mild viruses. They injection causes enough of an infection to produce an immune response in the individual, but it doesn't reproduce enough to harm the individual or spread to other people. Often live vaccines evolve to spread faster, but when they do, they are harmful like the original form.
I have read that live polio vaccines rely on the target infecting other people who were not directly vaccinated. But I don't think it spread very far; rather it has R<1, which means it spread, but burns out.
It is plausible that omicron is less infectious and less virulent than delta, but with more immune escape, so that in a naive population it would be outcompeted, but in existing highly exposed populations it is spreading. This would be mildly bad for the people who were already exposed and get a second mild illness, but potentially good for the people who have so far gotten nothing to get omicron rather than delta.
There's evidence that something like this has happened naturally in the past. Despite its name, the 1889-1891 "Russian flu" pandemic may have been caused by a coronavirus, not an influenza virus. One candidate for its identity is a coronavirus that today is implicated in up to 30% of common colds. See: https://sfamjournals.onlinelibrary.wiley.com/doi/full/10.1111/1751-7915.13889 for more details.
Do we *know* the virus itself has become virulent? Or might it be that almost everybody catches it umpteen times while young but if someone somehow managed to catch it for the first time when they're 70 it'd be as bad for then as it would have been in 1890?
Sure, it could definitely happen. It's definitely occurred to me that the 'common cold' is actually the equivalent of [Carcinisation](https://en.wikipedia.org/wiki/Carcinisation) in viral evolution. There are over a hundred endemic viral infections with symptoms and contagiousness so similar that we lump them all into the same disease; it's clearly a pretty stable evolutionary equilibrium. It's entirely possible that COVID will also eventually create a 'common cold' version of itself, with all the advantages that brings.
However, it's worth noting that we'd still probably be quite a bit more worried about it than a normal cold. The mutational distance from the hypothetical 'COVID cold' variant to something legitimately deadly with systemic infection would be much smaller than for typical colds. We'd probably keep a high alert on monitoring it, and push vaccination for it into the mandatory vaccine schedule to prevent its inevitable return as a deadly disease.
(Of course, actually tweaking such a disease and deliberately increasing its transmissibility would be folly of the highest order, even if we could do it. You don't WANT to make a disease that's a short mutational distance from deadly, highly transmissible. Its eventual deadly offspring would likely retain that transmissibility and be horrifying.)
> Metaculus didn’t want to wade in to precise lethality statistics, so they just asked for a yes-or-no answer on whether it would be deadlier than Delta. Forecasters say there’s a 34% chance it will be.
Is that "deadlier on a per case basis", or "deadlier en masse"?
That is, if I catch Delta will I breathing a sigh of relief that at least I didn't get schwacked with something really horrible like Omicron, or is it that way more people will catch Omicron than Delta so Omicron will rack up a higher body count over all by increasing the sheer number of infections?
Because... I'm given to understand that that's how endemic diseases work. It starts off lethal, then millions of millions of generations later the virus finds a sweet spot of being super infectious but relatively chill about killing the host. If Omicron spreads like Gangnam Style but is, say, half as likely to kill you... isn't that a good thing, overall?
Somebody, anybody, feel free to correct me harshly if I'm way off base.
>Is that "deadlier on a per case basis", or "deadlier en masse"?
I'm pretty sure it means "deadlier on a per case basis", i.e., "probability of death, conditional on getting infected". But looking at the wording of the question, it's not very explicit.
Sure, I mostly agree, but I think there is still a selection force toward a less severe virus (subject to the actual constraints of chemistry and biology of course.) There are certainly symptoms that COVID would benefit from losing. Lost of smell/taste is an obvious one; if people have a harder time distinguishing COVID from the flu, it could spread more easily. Similarly, the general fatigue from COVID reduces the sociability of its host; a bad trait, evolutionarily. Same with a fever; easy identification of illness without an associated upside.
There's a reason that there are over a hundred cold viruses. 90% of the things that increase infectiveness are respiratory. Almost all the symptoms from a systemic infection like COVID don't contribute to contagiousness of the host. A variant of COVID that only infected the nose and throat (rather than the whole body) would probably be much more contagious on a host-behavioral level.
Can I ask when you pulled those graphs? Because the numbers on Metaculus look substantially different now. The median prediction on mortality is now 25%.
1) The original covid strain is deadlier than most pathogens (and most coronaviruses)
2) Omicron has more mutations than other variants and is therefore the least like the original strain
3) We should expect reversion to the (less deadly) mean for Omicron which may be more like an average pathogen and less like the original covid variant
4) Therefore, all else being equal, we might expect Omicron to be less deadly?
Of course, all else is probably not equal. Notably, there's reason to believe it would have better vaccine escape, and if it's better at spreading that may be a sign that it has qualities that would make it deadlier. (Better at evading immune response? Better at replicating quickly?) But perhaps there's some reason to think these scary features may be offset by less deadliness otherwise? Maybe that's just wishful thinking, though.
I think you are wrong at least about 1). Look at SARS on wikipedia, the 2002-2004 outbreak of CoV-1 had some 8000 cases, with >700 deaths. I think this ratio is way over twice the lethality of COVID-19, and we were lucky that it did not spread as widely. (Note that I may be wrong about their relation, and also perhaps that initial rates of death might be higher due to the novelty, or perhaps we are doing better now due to advances in medicine. However I think we can agree that we did not get an uniquely lethal variant in ‘19. (MERS seems to have a cfr around 34.4?!))
Covid-19 is almost certainly the third or fourth most deadly coronavirus to humans that has affected people in the last century or so. It's far more deadly and dangerous than the majority of them. (We've known about four of them that cause "common colds" for a while, but two of them were only detected in the 1990s, and there just hasn't been a huge amount of virus sampling to determine whether or not there are many others.)
For the intrinsic deadliness that bortrand asks about, covid-19 is much less deadly than the common cold coronaviruses. The difference is that we get them as children and build immunity, whereas people are getting their first exposure to covid-19 as adults.
That actually does sound plausible, though I don't know how much evidence we have for that. The best evidence for this is the hypothesis that OC43 might have been the cause of the 1889 pandemic, but this is definitely far from conclusive.
I'm skeptical of the claim that we *all* get OC43 as children; there are too many varieties of the common cold for everyone to get them all, and even if OC43 is one of the more common strains, I'd expect a significant number of people to avoid it by luck (or maybe geography). In which case, there would be a significant number of people first contracting OC43 as adults, and enough of them becoming really sick or dead that this would be noticed.
But that needs more careful analysis than I expect anyone is likely to give it any time soon.
There are maybe 200 cold viruses, but the 4 coronaviruses are the common ones. 10-40% of colds are said to be from coronaviruses. In a sample of 105 people, no one tested negative for antibodies for OC43, 1 person for 229E, 2 for NL63, and 8 for HKU1:
Metaculus says it will resolve the claim based on "the first credible systematic review that estimates this value," citing Liu and Rocklöv (2021) as an example of the sort of thing they consider a “credible review.” But the methodology of Liu and Rocklöv seems atrocious: they write that across five studies they have identified, "the basic reproductive number for Delta ranged from 3.2 to 8, with a mean of 5.08." This has the following problems:
(1) The five studies are one article in English about Guangdong, one article in Chinese about Guangdong, two white papers about the UK, and one MedRxiv preprint about “China”. The five studies all used different methodologies for estimating R0. Liu and Rocklov say nothing about the strengths and weaknesses of these methodologies. Only one of the five studies indicated an explicit confidence interval (95% CI: 2.0–4.8).
(2) The reported “mean of 5.08” is simply the arithmetic mean of the midpoints of the R0 estimates of the five studies. Given the disparity of methodologies, calculating such a mean in the first place seems meaningless, and reporting it to three significant figures is ludicrous. The fact that this went through to publication does not speak well of the peer review process at the “Journal of Travel Medicine”.
(3) It doesn’t even make sense in the first place to talk about “the” R0 for a virus or variant — R0 is a parameter used in certain models for describing the dynamics of a pathogen *in a particular epidemiological situation*. Liu and Roclov half admit this, writing: “Given that the reproductive number in the studies identified here was estimated at a time when most countries still enforce a variable extent of lockdown measures, there is a risk that the real reproductive number may be even higher than the estimated 5.08.” But they fail to realize that the same logic implies that there is no such thing as the “real reproductive number” of a virus: there is only a range of estimated R0 parameters used in various epidemiological models to fit the observed case data in particular times and places.
(Epistemic status: I am not an epidemiologist, and I will be happy if anyone who knows what they are talking about can tell me why I’m wrong.)
> It doesn’t even make sense in the first place to talk about “the” R0 for a virus or variant — R0 is a parameter used in certain models for describing the dynamics of a pathogen *in a particular epidemiological situation*. Liu and Roclov half admit this, writing: “Given that the reproductive number in the studies identified here was estimated at a time when most countries still enforce a variable extent of lockdown measures, there is a risk that the real reproductive number may be even higher than the estimated 5.08.” But they fail to realize that the same logic implies that there is no such thing as the “real reproductive number” of a virus: there is only a range of estimated R0 parameters used in various epidemiological models to fit the observed case data in particular times and places.
I want to address only this point out of the good points you made. There is (obviously, IMO) some 'contagiousness' trait of a given disease. For instance, we know that measles spreads extremely rapidly through unvaccinated communities with little existing immunity, while a new flu (say, H1N1 when that came out a few years ago) can take weeks or months to cover the same community. So (to use Eliezer's map/territory metaphor) there's a real trait in the territory that is a property of a specific variant of disease which affects how quickly it would spread in a given community. The fact that the speed of spread will vary between *communities* for all viruses (due to population densities, precautions, behavioral differences, etc.) doesn't eradicate the fact that the speed of spread will vary between *viruses* for a given community. They both contribute.
R0 when used as a property of the virus is a crude estimate in our map of this real trait in the actual territory. It is definitely a meaningful thing to try to measure in spite of the fact that it's horribly confounded.
Right, I have no quibble with the idea that some variants of COVID are more contagious than others, and that the differences in contagiousness can be reasonably expressed in (ranges of) R0 values. My problem is with the framing by the Metaculus people: the idea that it is a sensible thing to bet on whether the R0 of omicron is 4.2 or 5.1 or 8.3. They get around this by saying they will resolve on the "first credible systematic review", but they seem to be unable to identify whether a published review is in fact credible or systematic. (Of course, Metaculus being unable to identify a credible review is somewhat less of a problem than the "Journal of Travel Medicine" being unable to identify a credible review...)
Interesting evidence indicating that R0 might actually be lower than Delta but with some immune evasion that allows it to spread in areas previously hit by Delta:
https://twitter.com/trvrb/status/1465364300936085506?s=20
Came here to post a different tweet in the same thread.
Paula came here to post the same tweet in all threads.
I'm not really sure R0 is actually a meaningful thing to run a prediction market on. The problem is, when you dig into how R0 is calculated, it's not really a precise or well defined value. In theory its definition is simple enough but in reality, it has no biological basis - claims that R0 is this or that are always coming from models. And when you look at what the models do, well, R0 turns out to just be some kind of fudge factor or arbitrary coefficient used to make the model roughly match the data reported so far.
As a consequence I've seen models that yield totally different values of "R0" for the same virus on the same dates, simply because the model was given case curves from different cities. As such it cannot be genuinely said that this represents anything genuinely true about the virus itself. It's not like it's derived from RNA or careful lab experiments or anything like that.
There are a lot of things in the world worth making predictions about which we can't nail down with the kind of scientific precision we would like. The solution is, rather than not trying at all, to just pre-specify as precisely as possible the method by which we will make the assessment in the end.
The fine print of the Metaculus prediction says:
"This will resolve on the basis of the mean R0 that is estimated for the Omicron variant according to the first credible systematic review that estimates this value. Such a systematic review should be peer-reviewed and should incorporate R0 estimates from at least 3 studies."
Metaculus users aren't being asked to predict "what is the R0 of Omicron", but rather, "what will the estimate of the first credible systematic review of R0 studies be?" And this is close enough to being the same question that it's still useful to us.
Yes, the prediction market can be successful on its own terms if it's just trying to predict what academics think R0 is. That isn't quite what I meant by being a meaningful thing to run a prediction market on though. I'm just really skeptical the resulting value is actually useful in any way.
It might be more meaningful to ask what the Rt of Omicron is, rather than the R0 of Omicron. R0 is the number of people each person infects *assuming no public health intervention.* But there are public health interventions!
Rt also has the advantage that you can estimate it with fairly minimal modelling assumptions. You need to know the interval between infections, or at least a distribution describing the serial interval. But aside from that, you can estimate it using only case counts.
I don't see how that is more meaningful, though. Whether you're measuring number of people infected without interventions or with interventions, the value you get is going to be highly dependent on how much time the people in the area are spending breathing each other's exhales. Do a thought experiment: Town of 10,000 people, very small and congested, large families squashed together in small dwellings. Let's say there are no public health interventions. One person with the virus enters the town and we track transmission for a while, and using that info we calculate how many people the average infected person infects. Now imagine the same 10,000 people, still no public health interventions, but now the people are living all spread out over the prairie. Some people live 5 miles from the next building. No real community gathering places except a coupla general stores. No schools, even. Introduce one infected person. Virus is going to spread much slower, right? R0 is going to be way smaller. And if instead there are public health interventions -- like say 50 % of the population is vaccinated, 75% wear masks around people they don't live with -- the Rt is still going to be way higher in the congested town than in the prairie community. Neither R0 nor Rt is a measure that is independent of how much face time the members of the population you are studying have with each other. The value you get is going to vary GREATLY depending on, basically, how many cubic feet of their neighbor's exhales an average citizen of the community is breathing per day.
That is why it does not make sense to talk about the R0 of a virus variant, as though it is a property of the variant that is the same at all times in all places. A virus with a R0 (or Rt) of 6 in the congested town setting might have an R0 (or Rt) of 2 on the prairie. In a setting where everyone never get closer than a mile away from another person both R0 and Rt would be zero..
That's the point - Rt is explicitly about a specific population at a specific time, while R0 claims to be about a population in a timeless sense.
As you observe, Rt boils down to a fancy derivative of a case curve, which means such a market would have to specify the exact time period and time series (e.g. city, country) over which the prediction will be made.
The problem here (and with almost all prediction markets) is you aren't predicting what R0 is, you are predicting what academic journals say R0 is. If you believe there is absolutely no bias in academic publishing then "what academic journals say" may be a reasonable proxy for "what is reality". If you believe there is a massive amount of bias in academic publishing, then "what academic journals say" may be totally unrelated to what is reality. For things like sports reporting, the reporting source is almost always a very strong proxy for actual results.
This knowledge of who is reporting tends to present itself in the prediction market quite well. A hypothetical market participant could believe that the *real* R0 of Omicron is negligible, but they may simultaneously believe that the academic journals will strongly prefer papers that have models that insight fear (more clicks/buys) and thus they may bet that R0 will be very high for this market.
Unfortunately, prediction markets are absolutely terrible at finding truth for this exact reason. They can do a very good job of predicting how some oracle will report in the future, but that is not necessarily correlated with what is true/real.
So what you're saying is that 4 teams of epidemologists could collude to subtly fudge the numbers and make a killing on the prediction markets?
Is there a prediction market with this prediction where you play with actual money?
You realize you would essentially be betting on how many people were going to die/get sick.
I think a lot of people feel intuitively that such betting would be distasteful.
Insider trading alone would be a massive problem if markets that bet on "what academics say" ever became worth significant real money. After all lots of people know what a meta-analysis is going to say before it gets published.
Yeah, I agree, mostly.
R naught is a property of the virus in a given setting. It's dependent on the virus's intrinsic transmissibility and also on how much time people in that setting log hanging out breathing each other's exhales. It depends on patterns of living, working, and socializing, and also on precautions such as masking and ventilation. No matter how transmissible a virus is there would be no transmission in a setting where each person never came within a mile of anyone else. R nought would be zero.
So yeah, talking about R naught in general is like talking about the risk of somebody getting run over. It's highly dependent on how much time the person logs crossing streets and how many cars are in the streets he's crossing..
Still, if one specifies a locale, say Brooklyn, it would be meaningful to talk about the R naught of virus X in that locale. (It would vary some depending on weather, part of Brooklyn and some other factors, but probably not all that much.)
R0 in theory is an intrinsic property of the virus, before adjustment for other factors like population density, weather, etc.
In practice epidemiologists have no way to actually measure this, so they end up making claims about R0 that are derived from pet models that are being curve fitted to the published PCR testing numbers. As such claims about R0 can easily conflict in major ways: it seems to bother nobody.
Can it be said R0 is a real thing at all, given the total inability to measure it or precisely define what it means? After all, as you astutely point out, it doesn't really make sense to talk about transmissibility in the absence of an actual context as if nobody was around and you were sealed in a vacuum jar, transmissibility of your viruses would be zero regardless of what the viral RNA was.
As such I am very dubious about all claims around transmissibility of viruses emanating from professors of epidemiology. It's just not a sound field. Values for R0 are usually on close inspection actually values for Rt that have been fiddled with a bit, where Rt is more or less some derivative of the case curve. Some models even take R0 as both an input and yield it as an output! Have fun figuring that one out ....
I've always been told that R0 is the average number of people an individual infected person infects, before recovering, in a hypothetical population where no one has any acquired immunity. That doesn't sound like an intrinsic property of the virus, unless you have a background assumption that fixes the behavioral and environmental features of the human population you're talking about.
The fact that they use the names "R0" and "Rt" is suggestive that the classical model *does* assume fixed human behavior and environment, with acquired immunity being the only relevant parameter that varies with time to change the reproductive number.
Yeah, and it's just dumb for the classical model to assume human behavior and environment are the same all over the world, right? Because they're not.
I mean, it's not dumb for scientists to use models that involve known false idealizations. Sometimes we can know that the idealizations are false in ways that don't make much difference (as when we assume that the sun and the planets are the only gravitational bodies in the solar system and there is zero friction as they move, or when we treat the ocean as homogeneous and infinitely deep in modeling the generation of waves by wind storms) and sometimes even when we know the model is going to be bad at making predictions, it can still be useful for helping us to understand the types of impact different factors can have (for instance, we treat a pool table as frictionless to understand how hitting the cue at different angles will send the other balls in different directions, and we treat an economy as homogeneous and well-mixing to understand how changes in supply, demand, taxes, and interest will affect inflation).
Yes, I understand about the need to simplify things to make useful models. But the simplification baked into the definition of R0 seems so radical that it interferes with the utility of the measure. R0 purports to be a measure of how many people an infected person will infect, but in fact number of infections passed on by an infected person depends on the environment and people's behavior in it, and there is a LOT of variability in those things across different settings. It's like treating pool tables as frictionless when some have smooth glossy surfaces and others are rough and gravelly and others are sticky with maple syrup.
Maybe it would be more useful to have a contagiousness measure where the setting is part of the definition: Number of people an infected person infects in a city of a certain population density. We could then accumulate a body of info about how variations in population density and other factors affect contagiousness in other settings. How dense is my setting compared to the setting on which the contagiousness measure is standardized? Use that as a multiplier . . .
Where nobody has any acquired immunity, where prior/adapted immunity doesn't exist, where everyone recovers at the same rate and has an immune system of identical strength, etc.
Basically, how many spherical frictionless cows are infected by another spherical frictionless cow in unit time. It's an interesting thought experiment but not a real number that can actually be calculated and have any meaning.
Have you no heart? I'm sure the number means quite a lot to those spherical frictionless cows!
For those of us who are neither spherical nor frictionless, do we still need to wear masks?
Can the glasgow coma scale be said to be a real thing at all?
Can IQs?
https://slatestarcodex.com/2014/08/11/does-the-glasgow-coma-scale-exist-do-comas/
Feels like this complaint leans in the same direction
Well, IQ is a pretty abstract scale. IQ itself is a real thing on its own terms because it has high test-retest consistency, the methodologies for determining it are standardized, it seems to correlate well with other "intelligency" things etc. At least I believe those things are true - I'm no expert in IQ though. The problems here start when you begin claiming IQ is the same thing as generalized intelligence, which of course is a common but much more controversial claim.
R0 is on much less firm footing than IQ. There's no standardized methodology to determine it, nobody even seems to agree on the precise definition (see the multiple similar-but-different attempts in this thread), it's not even clear there can be such a definition, and different academics routinely calculate totally different values via totally different methodologies for what is supposedly the same thing about the same virus.
Given that savant abilities tend to correspond with other "deficits", the general intelligence thing is tricky indeed.
IQ tests certainly dont measure certain types of intelligence- social intelligence, organizational intelligence, creative intelligence, others.
The things IG tests do measure might be correlated with those things, but thats a complicated question.
I have thought about someone from a western country who was taken into a nomadic hunter gathere society. Say they had some sort of intelligence test for the sorts of intelligence that was useful in their living conditions. Would high IQ scores correlate? I don't know.
I understand there is something sort of like global intelligence, but one has to define it carefully lest it covers modal nodes of intelligence it actually does not.
Came here to say sth. similar. You need to define at least the location to make a meaningful statement about R0.
Once you do that, I think it makes sense to use this, as the R0 of different viruses (under similar conditions of population density and habit) is so different; so it allows you to compare in a given setting.
Probably also a reflection of the health system's detection capability, to identify infected people, so the 'measured' R0 in a city in Europe (such as London) could have a different 'measured' R0 than a city with similar characteristics in the USA (such as New York).
Wastewater monitoring can be used to control for that.
The prediction market question should probably be "what R0 will (sensible authority) settle on" so you know which methodology you're guessing against.
These two questions bother me. But writing good questions is hard. Prediction markets aren't supposed to address that problem, so I'm not sure I'm justified in being bothered. (Unless prediction markets are supposed to address that problem by asking more objective questions. Which they are, but that's not the only thing they're supposed to do.)
The second question on deadliness bothers me less. It seems like a real-world practical question that people want to extract from the literature, so it makes sense to take the literature as given. I don't like the particular formulation, but that's the basic problem of asking questions. Whereas, I feel like the question about R0 is more like doing science and it shouldn't be taking the literature as given, but should be pushing for predictions of more concrete things. Eg, the ratio of R of the strains in the existing population instead of the absolute R of omicron in a naive population (which is also the practically relevant question).
But, again, I'm conflicted.
R0 might be impossible to calculate accurately, but the ratio between R0s of different variants is much more well defined and could be possible to calculate. So conditional on some value for R0 of the base variant, the R0 of another variant could be predicted.
Calculate how? There is no consistent methodology.
Agree there's no consistent methodology to measure R0. Here's an idea for measuring the relative R0 of 2 variants: How rapidly does one take over? Once variant Herbie starts showing up in a population where all the infected are hosting variant Batilda, how long does it take for Herbie to be the variant found in 80% of tests? Of course in most cases things will be more complicated -- there will be more than 2 variants circulating. But the math would still be doable. Maybe you could do it sort of the way chess rankings are done --everybody's ranking is a measure of how hard to beat they people they beat are, and how consistently they beat them.
I think that sort of metric might make sense, although of course tremendously confounded by how much sequencing a country does. I believe the UK is responsible for half of all GISAID SARS-CoV-2 uploads for some reason.
However, it wouldn't be R0.
Well, I'm not sure the differences among countries in sequencing would be a confound -- if by confound you mean a variable that distorts results. A country that doesn't do much sequencing would give us a skimpy data set, but not a distorted one -- right? Or is there something I'm not thinking of?
Perhaps confounded was the wrong word. The problem is that most countries do very little or even no sequencing, and there are gazillions of variants. It's not true that there's Alpha, Delta, Omicron etc. That's more or less a press fiction at this point. Look at GISAID to see the true family tree. So you get a lot of .... variants, strains, mutations, whatever you want to call them ... that grow rapidly but then top out at a few percent and don't go further for some reason.
So they'd never reach 80% and thus this figure could only be calculated retroactively. But the only reason people care about R0 is because fraudulent pseudo-scientists like epidemiologists claim it's very important and a high R0 number means "exponential growth" and thus P.A.N.I.C. (mathematics needs to be put in a care home for abused children after all this). As a consequence you'd end up calculating a lot of very dubious growth values for tons of variants that look initially scary and then two weeks later their growth tapped out for some reason, and you'd be doing it on like 10 data points in the best cases.
Overall I find myself thinking R0 is a concept that should just be humanely put to sleep, like the rest of the junk field that spawned it. Even if it had some actual precise meaning and could actually be calculated, so what? What's the actual actionable outcome of knowing such a value? Well, there isn't one. It's not true that high R0 values = PANIC because you'd still be left in a situation where you can't model how long the virus will grow for, or how serious it is, and even basic questions like "what is the actual capacity of our hospitals" turns out to be mired in nonsense and mis-understandings (by epidemiologists!). So it can't tell you anything and it ends up with stuff like what we see today, where the doctor who discovered Omicron points out that the person who had it didn't even realize they had "COVID" at all (because they didn't, the symptoms aren't even close to a match) yet because of years of misinformation about exponents and R0, scientists and politicians immediately panicked.
R0 is in the end highly reflective of the profession that spawned it: fuzzy, worthless, used only to create unjustified panic.
Agree that we are all walking around with a sense of the relative contagiousness of variants, but how do we arrive at that? My sense of variants' contagiousness is based on the accumulated headlines I've read. It's sort of an average -- well, a weighted average, because I trust some headline-writers more than others. And I'm not proud of that. Do you have a better calculation to propose?
The test you proposed in the other comment is roughly what I had in mind, with adjustments for level of immune escape, and assuming an equal serial interval between variants, but these can be measured independently.
Many seem to believe there's a case for omicron being *less* lethal than delta, based on near-zero-evidence throwaway quotes being shared widely on social media, best I can tell, out of wishful thinking. A metaculus question opened on this here:
https://www.metaculus.com/questions/8766/omicron-variant-less-deadly-than-delta/
I wish I was smart enough to know wether or not this article means I should be more scared or not. (I'm assuming....yes? Be more scared?). I'm boosted. Does that help or we don't know? #YourDumbestSubscriber
I'd say yes, be more scared of lockdowns, but no, be less scared of dying from it. And we don't know about the vaccines yet, but they won't hurt
It's all at the level of "maybe could be bad" at this point, so far as I can tell.
It's not the thought that counts, it's the data.
You should be proportionally more scared. Existing vaccines likely work poorly against the Omicron variant, but "poorly" is not the same as "not at all". All that vaccines (and other preventative measures) can do is mitigate the risk of infection. So, if you were a hermit living in some remote cave in the desert, vaccines would be irrelevant to you, since your chance of infection is zero anyway. If you have a habit of licking strangers all day, then no vaccine would help you for long. But, assuming you're a normal person, you can alter your behaviour within reason to reduce your chances of infection. With the Omicron variant, it means that pushing your behaviour a little closer to "hermit" might be a good idea.
You sound surer than many sources I'm reading that existing vaccines will work poorly against Omicron. Also, before you tell somebody how scared to be, you need to have some info about other factors: How contagious is the sucker? How virulent?
Well, it has three known antibody-evading mutations (from experimental work on exposing mutated spike proteins to vaccinated/convalescent serum) but nowhere near as many antibody-evading mutations as that work needed to fully evade antibody response. (And of course even fully evading antibody response still leaves T-cell response which is much harder to test.)
So we should have high confidence (>90%) that Omicron erodes antibody effectiveness somewhat, and that it does not come close to eliminating it. (>99%)
I don't know what Bugmaster meant by 'poorly' exactly; it's one of those things that needs numerical precision to evaluate. I'd estimate top vaccine effectiveness will remain over 70%, but that's not really based on much data.
Some of this seems quite wrong. Most obviously, the idea that a vaccine wouldn't help you if you went around licking strangers all day. If anything, the vaccine would help you most in that scenario!
More debatably, but still seemingly wrong to me, is the idea that vaccines "likely work poorly" against Omicron. No one really knows yet, but the evidence we have suggests that they may well work about as well as they ever did in protecting against serious illness. Which is the most important thing. So that seems like an at best overconfident, at worst very inaccurate characterisation.
Data from Israel last night: Vaccines protect against Omicron about as well as they do against Delta. The unvaccinated are about twice as vulnerable to Omicron as to Delta.
https://www.mako.co.il/news-lifestyle/2021_q4/Article-0e660b77fe17d71027.htm
> twice as vulnerable
Is that "twice as likely to get infected" or "twice as likely to get sick/die"?
My tentative understanding of Omicron, definitely subject to update, is that it is less deadly but spreads much faster.
I wonder about that too. We need someone who can read Hebrew to go thru whole article and summarize.
My gut feeling was that vaccines would protect at most 50% as well against Omicron as against the vanilla variant (and more likely around 25%); looks like I was in the ballpark ?
Wait, no FDA approval?! That's outrageous!
There is an interesting article arguing that the models being used for Covid are wrong, mostly because they ignore the structure of a population, the fact that A is much more likely to interact with B and C than with X, Y, and Z who are interacting with each other. He argues that the actual pattern of what has been happening cannot be explained with the current models. One of his conclusions is that the newer variants may not actually be much more infectious than the older.
https://cspicenter.org/blog/waronscience/have-we-been-thinking-about-the-pandemic-wrong-the-effect-of-population-structure-on-transmission/
Why did you drop a letter from your name in your last few comments?
It really is giving some of us fits. [no not really. but we are curious]
I mean if you were Grampa Simpson you could just say you can't use 'n' cause the Germans stole it.
Not the real David F. would be my guess.
That’s a good guess. Like the spam I get from Amazonn.con
Except it exactly matches the real David Friedman's writing style, posting style, subject interest, economic positions and beliefs. It would be an amazing impersonater if true..... What would be the point? To save the real David from getting out of bed?
And he's been here long enough that the real DF could've shown up.
It's a profile setting for substack, though.
GPT-DDF
I have no idea why the letter got dropped; I've been blaming Scott. It's probably somewhere on his floor or maybe in one of his pockets.
Maybe he hid it somewhere in a logo like with that N in the slatestarcodex blog (because it was the gap for making it a perfect anagram for Scott Alexander)
That does sound like him.
This is really good. I've seen some people make claims that population structure would affect things, but they've done so at the length of a tweet or a Facebook comment thread, and not with any of this relevant detail. I would like to see a bit more about how sensitive this model is to certain structures of connectivity across the graph, and how things like actual transmissibility advantage of one variant over another and actual mild effectiveness of NPIs would appear in this model, but it makes a lot of sense.
I'm a bit surprised to hear him say that people have been denying the importance of this kind of structure. But this is partly because as far as I know, nothing in the US has ever been done on the basis of calculations of R or transmissibility advantages - perhaps in Europe people were trusting the simple models a lot to calculate those things.
For what it's worth, I did some work about a decade ago on damage spreading in networks which is *somewhat* similar if you squint hard enough. The key finding is that the largest eigenvalue of an appropriately scaled adjacency matrix determines if you get percolation (i.e., a pandemic) or not.
Of course, the details are vastly different between what I worked on and disease spread, but I think some of the broad conclusions are probably applicable, which are that you can get significant spread when it's unexpected by naive models--that only look at average numbers of links--just by changing the connectivity patterns. In particular, assortative networks in which highly connected nodes are more likely to connect to other highly connected nodes will have damage spread more easily.
Not quite true. There were people writing papers pointing this out in 2020.
https://nicholaslewis.org/why-herd-immunity-to-covid-19-is-reached-much-earlier-than-thought-update/
This was before the second wave so the argument about heterogenous connectivity is mixed in with claims about much lower herd immunity thresholds but the gist of it is there - the "cutting edge" academic models at the time assumed everyone was identical in every way. When people with actual standards and scientific ability started looking at them, they not only immediately noticed this problem (that somehow the field of epidemiology itself had spent decades not noticing), but also built alternative models that fixed it and realized it meant the original predictions were wildly off.
This doesn't seem like an even vaguely accurate summary of the state of epidemiology to me. Google "network epidemiology" to see papers, journals and textbooks going back at least two decades that attempt to build models accounting for network structure.
The Imperial model which was the focus of UK policy at the beginning of the pandemic was an agent-based simulation based on multiple factors (age, location, transport links, etc), and simulating contacts on a non-homogeneous network. Obviously some relevant information was left out (that's the nature of modelling), but the idea that epidemiologists had somehow just not noticed that homogeneous SIR models don't perfectly predict reality seems very clearly false.
I don't think that model had non-homogenous contact networks, which is what makes the big difference here, but I might be mis-remembering. It did do some basic age stratification and had different "R0" (or some sort of R) values for different "places".
I've never seen any admission in epidemiology papers that their models fail to predict reality. Although they produce new models all the time, it's not like some fields where there are open competitions, robust accuracy metrics etc. Actually I rarely if ever see any comparisons between modelled predictions and what really happens, which is why as a group they've developed a reputation for being constantly wrong and yet acting as if they weren't.
This is just really uncharitable. Let's take as an example the paper that describes the model I mentioned above. The full text is here: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7095311/
Here's a quote from the final paragraph:
"Although more detailed model validation and parameter estimation using data from past pandemics should be a priority for future research, it will be impossible to predict the exact characteristics of any future pandemic virus."
As for the details - the model itself is agent-based. It predicts individual contacts of the simulated agents, assuming the contacts take place in different types of location (workplaces, schools, households, and 'other' if I'm reading it correctly). This necessarily results in a non-homogenous contact network (kids associate in schools, adults don't).
The idea that epidemiologists weren't aware of the limitations of models assuming homogeneous contacts is, as you imply in your first post, ludicrous and as such you should be sceptical of it. My general impression is that they have become significantly more aware of it over the past decade, and models have started to try to take it into account more and more (basically, as the computation power and data to make such models useful has become available)
Here's a paper from 2019 which deals with the question of non-homogeneous mixing in equine populations:
https://www.nature.com/articles/s41598-019-40151-2
Here's some work considering the effect if the structure of the network changes over time:
https://www.sciencedirect.com/science/article/pii/S1755436518300173
To quote the introduction of that second paper, from 2018: 'a full consideration of the often profound effect of heterogeneous contact structure on infectious disease dynamics has really only become widespread more recently, and especially since the adoption of methods often originally conceived in the context of social network analysis'
Sometimes epidemiologists choose to use the simple models anyway, because they're faster to compute, and accurate enough, but they (or at least some of them) very definitely are aware of the limitations.
I don't doubt that some epidemiologists overstate the case for some simple models, and I don't doubt that sometimes the models are used badly, but the generalisations that you're making seem extremely uncharitable and honestly, I don't think they're very accurate.
I don't mean a generic disclaimer of the form "of course, models aren't perfect". Every paper has those but who cares? What I mean is the actual scientific process:
1. Hypothesis.
2. Experiment or observations.
3. Refinement, abandonment or acceptance as theory.
Epidemiology seems to consist near-exclusively of (1). Where are these papers by Imperial College that dissect and analyze their previous failures, which have been numerous and enormous over a period of decades? There are none. Where is the admission that the model they put in front of governments for COVID was totally unvalidated, and unfit for purpose? There has never been such an admission.
Perhaps you think it's unfair to harshly judge the whole field on the basis of Imperial College. In fact I've seen lots of papers that are like their work, but, OK, other than John Ioannidis (who has been treated with incredible disdain for doing so), where are the epidemiologists condemning Ferguson's team and furiously demonstrating how they aren't all like that? Where are the demands that "scientists" who promote ridiculous, totally unvalidated models that produce incorrect predictions be fired or otherwise penalized? There's nothing because they're all at it and don't see anything wrong with it.
As previously mentioned, Ferguson's model is very concerned with R and R0. It took such a value as input. Most of this discussion is about the fact that no such value can be determined and even what it means is totally up in the air. These guys have been working on their models, publicly funded, for more than 20 years - the world can and should expect better than this especially given the horrifying impact these incorrect model predictions had (there was never going to be a single massive wave and they knew it because epidemics have never worked like that).
Yes, the Imperial model was complicated, but it was worse than a raw SIR model. Pretending that epidemiological study of networks doesn't exist is overly charitable.
I think that's a different argument, and one I was making in a less detailed form. If some people are more likely to catch the virus than others, for either biological or behavioral reasons, then the more vulnerable are being selectively infected, so the average vulnerability of the population is falling over time, so her immunity is earlier.
http://daviddfriedman.blogspot.com/2020/08/covid-invisible-elephant-in-room.html
I'm glad you found the post interesting!
I agree that it would be interesting to study more systematically how the properties of the network affect the dynamic of the epidemic. This is a natural development of the work I did in this post. However, that's a lot of work and, in that post, I just wanted to present the basic idea and some results showing that what you would expect intuitively can in fact happen in simulations, as well as explain why I think it matters.
On your second point, this is something I already thought before and some of the reactions to this post have only strengthened this view, but I do suspect that the difference between how things went down in the US vs. in Europe explains in large part this difference of perception. I have been in Paris since the beginning of the pandemic and, in France, the kind of simplistic models I criticize have played a very prominent role in the justification of the government's policy, if not in the elaboration of that policy itself although it clearly had some impact on that as well.
For instance, the leader of the group who published this ridiculous paper (https://www.medrxiv.org/content/10.1101/2021.02.14.21251708v2) was invited to a press conference organized by the French government in February during which she pretended that it showed that curfews were very effective, which the government explicitly used to justify the continuation of that policy, even though the methods in that paper can in no way support that claim. (For more on why, see https://necpluribusimpar.net/lockdowns-science-and-voodoo-magic/ and https://cspicenter.org/blog/waronscience/the-british-variant-of-sars-cov-2-and-the-poverty-of-epidemiology/ in addition to what I explain in my post on population structure.) Throughout the pandemic, every time Macron announced new restrictions, he explicitly cited projections by this group or another at the Institut Pasteur.
I think modeling also played a major role to justify policy in the UK. From what I can tell, it was very different in the US, where apart from the cubic model fiasco — which didn't really matter anyway because the power to order restrictions to contain the pandemic belong to the states — I don't think modeling played a major role in political discourse. But in France and many other European countries it was very different.
I imagine this must be a lot of work! Thanks for posting some of the things on Github. I unfortunately don't have much experience with this kind of thing, but this post was so interesting that I'm wondering if it will be able to get me to try to teach myself how to use some of it over the winter break to see if I can get a bit of understanding.
Some questions that suggested themselves to me while reading (just in case one or more of these turn out to be easy to address, or if anyone other than you is reading who can work on these models), in addition to the ones about what happens if there is some slight effect of NPIs, or slightly different transmissibility of the variants:
How sensitive is this to the seeding mechanism? If seeding happens uniformly in all sub-populations and sub-networks, then you automatically get a transmission advantage for new variants. (New variants will be equally spread across all sub-populations, while old variants will be disproportionately concentrated in sub-populations that have a high current infection rate, and thus are lower transmission for all variants.) But seeding by new variants that have become common overseas shouldn't occur uniformly, but instead should have some patterns of their own. If we treat your model as a model of the *world*, with no seeding, rather than a model of a country with external seeding, then this effect might go away, since new variants will be more likely to appear in sub-populations that have more cases. However, if we ignore new variants until they account for 100 cases, this will select for paying attention to variants that, by chance, are present in sub-populations with high current transmission rates. Does this restore the effect?
In your model, the seeding switches from one variant to another, and the second variant ends up taking over much faster than one might naively expect. If both variants continue seeding equally, how often does it occur that the new variant ends up becoming more common than the old variant? Or in a model with no seeding, but just occasional random mutations into a new variant (without any transmissibility advantage) how often does the new variant end up taking over?
How small of a transmissibility advantage yields transmission advantages that seem qualitatively similar to the data we have for alpha and delta?
I really like the idea you have of explaining geographic correlations by having networks of sub-populations that don't line up with geographic networks. It fits with a lot of things I've been thinking for years (including many topics Scott writes about) - city center networks in different cities are often going to be more closely connected to each other than to the suburban networks immediately surrounding them, and blue tribe/red tribe/grey tribe networks will be connected to each other in different ways than they are connected across these lines within a geographic area. And anecdotally, I noticed that I had several friends (around the world) who were infected with covid in March/April 2020, but then none (as far as I know) until April 2021, when a friend in Michigan, a friend in Toronto, and a friend in Paris all got infected within a few days of each other. It would make a lot of sense if these people are all multiply connected through social networks that are geographically dispersed.
But it does seem to me that even if this sort of tree-like hierarchy of network and sub-population interconnections has some geographic dispersal, that there should also be a cross-cutting set of geographically local connections. It might not be too hard to add extra connections across sub-populations that are geographically local, but maybe it is. It would be interesting to see how sensitive some of the big results are to the tree-like structure where sub-populations are grouped within a network, if instead we have sub-populations in a bit more complex of a network (like a small-world graph, for instance).
This also raises a deeper question. The idea of a population as a well-mixed and homogeneous network is obviously very unrealistic. But conversely, the idea of sub-populations as being mostly self-connected with only occasional connections within a larger network, and only very rare connections outside that network, also seems quite unrealistic. The latter model might well be a very good model of what happens during a period of very tight "lockdown" (whatever that means to people). I think you did one investigation that shows that with enough connections across these sub-populations, you end up with something that looks very much like the homogeneous model, but it would be interesting (and presumably extremely difficult) to see what happens between. And it also makes me wonder what happens if we model "lockdown" with a switch between the two types of dynamic.
Yes, modelling was the single key factor that converted the UK from a relatively Sweden-style relaxed approach to harsh China-style approach. Once Ferguson made his graph of a single giant curve totally swamping all hospitals, it was over because nobody in government seemed to be aware of his team's numerous prior failures, and suddenly all sorts of academics popped up claiming ICL were the "best epidemiologists in the world".
Lots of people in the general public caught on within months that the man was a ribald liar - despite claiming some sort of apocalypse was inevitable even if everyone did lock down, he was immediately caught breaking lockdown rules to go sleep with his lover, who not surprisingly was a far left extremist (extinction rebellion). Clearly he didn't actually believe his own predictions and the nation found out why later on, much to its own cost.
My understanding of the rationale behind lockdowns was to reduce the homogeneity of social structure, to give us contagion dynamics that look more like the waves and less like the big bell curve. I don't want to take responsibility for anyone else's messaging on that point, but that's how I understood from the start: reduce the number of edges on the graph, and reduce their strength, as much as possible, to reduce the degree to which infection was GEOGRAPHICALLY correlated (since that's also how medical resources are distributed).
Journalists may have made a hash of the point but who cares?
We saw really quite decent evidence for at least some of this population-structure based spreading in the last UK wave: it produced a very significant peak in the teenage population, and a corresponding smaller peak in the age group corresponding to their parents. The younger and older adults saw much smaller growth.
See the "Cases by specimen date age demographics" heatmap here (scroll down a bit): https://coronavirus.data.gov.uk/details/cases?areaType=nation&areaName=England
Spin-off question: Would it be theoretically possible for a variant to be, for all practical purposes, a covid vaccination? So it would be a virus that kills about 1 person in a million, but mostly causes no symptoms or very mild ones -- and confers immunity to covid for the next few months to all who have been infected. And if that's possible, mightn't it be possible to tweak the variant so as to make it so hugely transmissible that everybody gets it, even the animals who can be infected by covid? If that happened, could we truly and permanently get rid of all variants except the benign one?
I'm sure it's very unlikely that such a virus will pop up, but still would like to hear from knowledgeable people whether such a sequence of events could happen if such a virus did appear.
My guess is that it is way beyond existing biotechnology to create a variant that (1) is mostly asymptomatic or mildly symptomatic and (2) confers immunity to existing variants of SARS-CoV-2 and (3) is at least as transmissible as Delta. Being transmissible requires hijacking cellular machinery to replicate itself, and that's the kind of thing that triggers the immune system and usually provokes symptoms.
But there already are viruses that work this way. We call them "the common cold", though it's really many separate viruses.
It would probably require a fair amount of luck to have such a Covid-19 compatible virus appear, but it seems possible to me.
Disclaimer: I have no virology or related expertise at all.
Not a virologist, but it seems that you'd need to restrict infectivity to the upper respiratory tract (as opposed to systemic infectivity that COVID has) while keeping the spike protein structure largely intact, and the mechanism that you'd use for that need to be robust enough so that a few random mutations here and there would not cause it to become systemic again. Even if it were possible, it seems very easy to make matters worse while developing such a mechanism and having an accidental lab leak.
If there's "gain of function" research, why not "loss of function" research?
At that point why not just have vaccines? I guess the answer is (i) vaccines you have to go out and physically jab people, and (ii) some people don't like vaccines. On (i), fair enough but at this point the infrastructure required to physically jab people isn't the main bottleneck, and on (ii) I don't think the people who dislike the vaccine will be happier with "it's like a vaccine but it spreads through the air so now you CAN'T avoid it!"
I'd also be worried that it mutates into something deadlier, apparently that can happen a lot!
I think you'd still have to go physically jab people - as contagious as COVID is, it's nowhere near contagious enough to give the high % coverage you want by natural spread.
Maybe we need a more benign version of the Rage Virus in *28 Weeks Later" . . .
The Mild Annoyance Virus?
Yeah! Or the Peevishness Pox
It's sort of amazing that even at the low points of the vaccination program, there have been over 700,000 people getting vaccines per day in the United States, while the highest recorded number of infections was about 300,000 per day in the middle of last winter's wave. There are surely more actual infections than the recorded number, but probably since March or April, there has likely not been a single week where more people were infected than got vaccinated.
I think that at least some consequences of the virus (sneezing being a prime example) serve to enhance infectivity. So you wouldn't be able to both perfectly maximize infection and minimize damage. My guess is that the more harmful parts of COVID would somehow also enhance infectivity (since for example asymptomatic cases seem to have lower infectivity rates from what I last read, but that may be out of date), to the point that it would be difficult to make a low-damage competing strain.
This is sort of how the first vaccinations worked: using cowpox as a vaccine for smallpox. So like, it's *possible* for such a thing to happen by chance.
But we had to manually infect people with cowpox, it didn't outcompete smallpox on its own. And cowpox only had to be safer than smallpox, which is a much lower bar to clear.
Yes, it'll hopefully evolve into just another nasty cold within a number of years. no, deliberately exposing the population to infectious diseases is a bad idea, even if you think they're pretty innocuous.
That's the _Medical Police_ season finale!
That's pretty much what cowpox is for smallpox, and is why it's called a vaccination.
A couple of people have mentioned the original small pox vaccination. But a better match for this is the oral polio vaccine which is an attenuated virus that is capable of (at least in limited circumstances) of spreading from person to person. This helped increase vaccination coverage in difficult to reach areas. But it can also mutate to more severe forms (see vaccine derived polio, though note is still typically milder than wild type polio) and at one point such forms were the dominate circulating strains in certain areas. As polio has gotten closer to eradication this vaccine gets phased out in favor of the injected killed virus form.
+1 for not *completely* disregarding the "hugely transmissible" part of Eremolalos's comment, unlike nearly everybody else.
+2 for noticing.
That's the principle behind the cowpox->smallpox vaccine approach, so it seems at least feasible?
Theoretically, yes, practically, no.
Virulence is generally a side-effect of infectiousness. Cow pox and traditional attenuated/live vaccines are mild viruses. They injection causes enough of an infection to produce an immune response in the individual, but it doesn't reproduce enough to harm the individual or spread to other people. Often live vaccines evolve to spread faster, but when they do, they are harmful like the original form.
I have read that live polio vaccines rely on the target infecting other people who were not directly vaccinated. But I don't think it spread very far; rather it has R<1, which means it spread, but burns out.
It is plausible that omicron is less infectious and less virulent than delta, but with more immune escape, so that in a naive population it would be outcompeted, but in existing highly exposed populations it is spreading. This would be mildly bad for the people who were already exposed and get a second mild illness, but potentially good for the people who have so far gotten nothing to get omicron rather than delta.
There's evidence that something like this has happened naturally in the past. Despite its name, the 1889-1891 "Russian flu" pandemic may have been caused by a coronavirus, not an influenza virus. One candidate for its identity is a coronavirus that today is implicated in up to 30% of common colds. See: https://sfamjournals.onlinelibrary.wiley.com/doi/full/10.1111/1751-7915.13889 for more details.
Do we *know* the virus itself has become virulent? Or might it be that almost everybody catches it umpteen times while young but if someone somehow managed to catch it for the first time when they're 70 it'd be as bad for then as it would have been in 1890?
Sure, it could definitely happen. It's definitely occurred to me that the 'common cold' is actually the equivalent of [Carcinisation](https://en.wikipedia.org/wiki/Carcinisation) in viral evolution. There are over a hundred endemic viral infections with symptoms and contagiousness so similar that we lump them all into the same disease; it's clearly a pretty stable evolutionary equilibrium. It's entirely possible that COVID will also eventually create a 'common cold' version of itself, with all the advantages that brings.
However, it's worth noting that we'd still probably be quite a bit more worried about it than a normal cold. The mutational distance from the hypothetical 'COVID cold' variant to something legitimately deadly with systemic infection would be much smaller than for typical colds. We'd probably keep a high alert on monitoring it, and push vaccination for it into the mandatory vaccine schedule to prevent its inevitable return as a deadly disease.
(Of course, actually tweaking such a disease and deliberately increasing its transmissibility would be folly of the highest order, even if we could do it. You don't WANT to make a disease that's a short mutational distance from deadly, highly transmissible. Its eventual deadly offspring would likely retain that transmissibility and be horrifying.)
> Metaculus didn’t want to wade in to precise lethality statistics, so they just asked for a yes-or-no answer on whether it would be deadlier than Delta. Forecasters say there’s a 34% chance it will be.
Is that "deadlier on a per case basis", or "deadlier en masse"?
That is, if I catch Delta will I breathing a sigh of relief that at least I didn't get schwacked with something really horrible like Omicron, or is it that way more people will catch Omicron than Delta so Omicron will rack up a higher body count over all by increasing the sheer number of infections?
Because... I'm given to understand that that's how endemic diseases work. It starts off lethal, then millions of millions of generations later the virus finds a sweet spot of being super infectious but relatively chill about killing the host. If Omicron spreads like Gangnam Style but is, say, half as likely to kill you... isn't that a good thing, overall?
Somebody, anybody, feel free to correct me harshly if I'm way off base.
>Is that "deadlier on a per case basis", or "deadlier en masse"?
I'm pretty sure it means "deadlier on a per case basis", i.e., "probability of death, conditional on getting infected". But looking at the wording of the question, it's not very explicit.
> then millions of millions of generations later the virus finds a sweet spot of being super infectious but relatively chill about killing the host.
The virus doesn't really *care* if the host dies or not, except in the fact that it gets to spread.
For super serious diseases like Ebola, they would benefit from being less deadly, because when you get sick you know it and hide away from others.
There is lots of room for covid to be either more mild or more severe while still becoming more infectious.
Sure, I mostly agree, but I think there is still a selection force toward a less severe virus (subject to the actual constraints of chemistry and biology of course.) There are certainly symptoms that COVID would benefit from losing. Lost of smell/taste is an obvious one; if people have a harder time distinguishing COVID from the flu, it could spread more easily. Similarly, the general fatigue from COVID reduces the sociability of its host; a bad trait, evolutionarily. Same with a fever; easy identification of illness without an associated upside.
There's a reason that there are over a hundred cold viruses. 90% of the things that increase infectiveness are respiratory. Almost all the symptoms from a systemic infection like COVID don't contribute to contagiousness of the host. A variant of COVID that only infected the nose and throat (rather than the whole body) would probably be much more contagious on a host-behavioral level.
Is the deadliness question about whether it kills more people conditional on their being infected, or just that it kills more people altogether?
My reading of the more detailed explanation is "conditional on it being infected": https://www.metaculus.com/questions/8757/omicron-variant-deadlier-than-delta/
Do the prediction markets say anything about how likely it is that Omicron will evade current vaccine protection?
Can I ask when you pulled those graphs? Because the numbers on Metaculus look substantially different now. The median prediction on mortality is now 25%.
Is it reasonable to think:
1) The original covid strain is deadlier than most pathogens (and most coronaviruses)
2) Omicron has more mutations than other variants and is therefore the least like the original strain
3) We should expect reversion to the (less deadly) mean for Omicron which may be more like an average pathogen and less like the original covid variant
4) Therefore, all else being equal, we might expect Omicron to be less deadly?
Of course, all else is probably not equal. Notably, there's reason to believe it would have better vaccine escape, and if it's better at spreading that may be a sign that it has qualities that would make it deadlier. (Better at evading immune response? Better at replicating quickly?) But perhaps there's some reason to think these scary features may be offset by less deadliness otherwise? Maybe that's just wishful thinking, though.
I think you are wrong at least about 1). Look at SARS on wikipedia, the 2002-2004 outbreak of CoV-1 had some 8000 cases, with >700 deaths. I think this ratio is way over twice the lethality of COVID-19, and we were lucky that it did not spread as widely. (Note that I may be wrong about their relation, and also perhaps that initial rates of death might be higher due to the novelty, or perhaps we are doing better now due to advances in medicine. However I think we can agree that we did not get an uniquely lethal variant in ‘19. (MERS seems to have a cfr around 34.4?!))
Covid-19 is almost certainly the third or fourth most deadly coronavirus to humans that has affected people in the last century or so. It's far more deadly and dangerous than the majority of them. (We've known about four of them that cause "common colds" for a while, but two of them were only detected in the 1990s, and there just hasn't been a huge amount of virus sampling to determine whether or not there are many others.)
For the intrinsic deadliness that bortrand asks about, covid-19 is much less deadly than the common cold coronaviruses. The difference is that we get them as children and build immunity, whereas people are getting their first exposure to covid-19 as adults.
That actually does sound plausible, though I don't know how much evidence we have for that. The best evidence for this is the hypothesis that OC43 might have been the cause of the 1889 pandemic, but this is definitely far from conclusive.
I'm skeptical of the claim that we *all* get OC43 as children; there are too many varieties of the common cold for everyone to get them all, and even if OC43 is one of the more common strains, I'd expect a significant number of people to avoid it by luck (or maybe geography). In which case, there would be a significant number of people first contracting OC43 as adults, and enough of them becoming really sick or dead that this would be noticed.
But that needs more careful analysis than I expect anyone is likely to give it any time soon.
There are maybe 200 cold viruses, but the 4 coronaviruses are the common ones. 10-40% of colds are said to be from coronaviruses. In a sample of 105 people, no one tested negative for antibodies for OC43, 1 person for 229E, 2 for NL63, and 8 for HKU1:
https://pubmed.ncbi.nlm.nih.gov/20943876/
Metaculus says it will resolve the claim based on "the first credible systematic review that estimates this value," citing Liu and Rocklöv (2021) as an example of the sort of thing they consider a “credible review.” But the methodology of Liu and Rocklöv seems atrocious: they write that across five studies they have identified, "the basic reproductive number for Delta ranged from 3.2 to 8, with a mean of 5.08." This has the following problems:
(1) The five studies are one article in English about Guangdong, one article in Chinese about Guangdong, two white papers about the UK, and one MedRxiv preprint about “China”. The five studies all used different methodologies for estimating R0. Liu and Rocklov say nothing about the strengths and weaknesses of these methodologies. Only one of the five studies indicated an explicit confidence interval (95% CI: 2.0–4.8).
(2) The reported “mean of 5.08” is simply the arithmetic mean of the midpoints of the R0 estimates of the five studies. Given the disparity of methodologies, calculating such a mean in the first place seems meaningless, and reporting it to three significant figures is ludicrous. The fact that this went through to publication does not speak well of the peer review process at the “Journal of Travel Medicine”.
(3) It doesn’t even make sense in the first place to talk about “the” R0 for a virus or variant — R0 is a parameter used in certain models for describing the dynamics of a pathogen *in a particular epidemiological situation*. Liu and Roclov half admit this, writing: “Given that the reproductive number in the studies identified here was estimated at a time when most countries still enforce a variable extent of lockdown measures, there is a risk that the real reproductive number may be even higher than the estimated 5.08.” But they fail to realize that the same logic implies that there is no such thing as the “real reproductive number” of a virus: there is only a range of estimated R0 parameters used in various epidemiological models to fit the observed case data in particular times and places.
(Epistemic status: I am not an epidemiologist, and I will be happy if anyone who knows what they are talking about can tell me why I’m wrong.)
You're not wrong and there's a discussion up-thread about this exact problem.
> It doesn’t even make sense in the first place to talk about “the” R0 for a virus or variant — R0 is a parameter used in certain models for describing the dynamics of a pathogen *in a particular epidemiological situation*. Liu and Roclov half admit this, writing: “Given that the reproductive number in the studies identified here was estimated at a time when most countries still enforce a variable extent of lockdown measures, there is a risk that the real reproductive number may be even higher than the estimated 5.08.” But they fail to realize that the same logic implies that there is no such thing as the “real reproductive number” of a virus: there is only a range of estimated R0 parameters used in various epidemiological models to fit the observed case data in particular times and places.
I want to address only this point out of the good points you made. There is (obviously, IMO) some 'contagiousness' trait of a given disease. For instance, we know that measles spreads extremely rapidly through unvaccinated communities with little existing immunity, while a new flu (say, H1N1 when that came out a few years ago) can take weeks or months to cover the same community. So (to use Eliezer's map/territory metaphor) there's a real trait in the territory that is a property of a specific variant of disease which affects how quickly it would spread in a given community. The fact that the speed of spread will vary between *communities* for all viruses (due to population densities, precautions, behavioral differences, etc.) doesn't eradicate the fact that the speed of spread will vary between *viruses* for a given community. They both contribute.
R0 when used as a property of the virus is a crude estimate in our map of this real trait in the actual territory. It is definitely a meaningful thing to try to measure in spite of the fact that it's horribly confounded.
Right, I have no quibble with the idea that some variants of COVID are more contagious than others, and that the differences in contagiousness can be reasonably expressed in (ranges of) R0 values. My problem is with the framing by the Metaculus people: the idea that it is a sensible thing to bet on whether the R0 of omicron is 4.2 or 5.1 or 8.3. They get around this by saying they will resolve on the "first credible systematic review", but they seem to be unable to identify whether a published review is in fact credible or systematic. (Of course, Metaculus being unable to identify a credible review is somewhat less of a problem than the "Journal of Travel Medicine" being unable to identify a credible review...)
I am new to prediction markets as a tool. I understand that Metaculus is probably not Wikipedia, i.e. it's not like a few uninformed bu